The Use of Multistatic Radar in Reducing the Impact of Wind Farm on Civilian Radar System

Transcription

1 The Use of Multistatic Radar in Reducing the Impact of Wind Farm on Civilian Radar System A thesis submitted to The University of Manchester for the degree of Doctor of Philosophy (PhD) In Faculty of Science and Engineering 2016 Waleed Al Mashhadani School of Electrical and Electronic Engineering

2 Table of Content List of Figures... 7 List of Tables... 9 Abbreviations Abstract DECLARATION COPYRIGHT STATEMENT Acknowledgement Chapter 1: Introduction Overview Motivation Impact of Wind Farms on Radar Impact on receivers Sidelobe detections Shadowing Multiple reflections Multistatic Radar Range-Only Detection within Wind Farm Environment Research Novelty Thesis outline Chapter Summary References Chapter 2: Bistatic and Multistatic Radar Systems Background Overview Monostatic, Bistatic and Multistatic Radar Systems Active Bistatic Radar Passive Coherent Location (PCL) Passive Bistatic Radar (PBR)

3 2.3 Bistatic Radar Systems Interest Cycles History of Bistatic and Multistatic Radar Geometry of Bistatic Radar Bistatic Radar Equation Bistatic Target Location Bistatic Doppler Frequencies Multistatic Radar Systems Classification of Bistatic and Multistatic Radar Complexity Level Cases Multistatic Radar System Degree of Spatial Coherence Multistatic Radar System Data Fusion Bistatic and Multistatic Advantages Bistatic Advantages Design advantages Operational Advantages Multistatic Radar Advantages Advantages Passive [4] Disadvantages of Bistatic Radar Systems Chapter Summary References Chapter 3: Multistatic Radar Simulation Tool and Wind Turbine RCS Modelling Introduction Simulation Tool Assumption and Operation Environment Simulation Tool Antenna Component Simulation Tool Bistatic Geometry Component Simulation Tool Power Received Power Received Simulation Tools Result Bistatic Propagation Factor

4 3.5.1 Simulation Tool Bistatic Propagation Factor Bistatic Signal-to-Noise (SNR) Simulation tool Directed Antenna Simulation Probability of Detection Wind Turbine RCS Modelling Wind Turbine RCS Modelling Tool Assumptions The ReMeRA Methodology Verification of RCS Modelling Tool Chapter Summary References Chapter 4: Multistatic Radar Target Detection and Tracking Introduction Multistatic Radar Range-Only Target Detection Ghost Target Detection Ghost Elimination by Multiple Results Analysis Approach (MRA) Target Localisation Algorithm Optimizing Localisation Algorithm Localisation Algorithm Limitation Particle Filter (PF) Implementation of PF Prediction step Sampling step Implementation Example PF Simulation Tool Results Interval Analysis (IA) IA Target Positioning Modelling IA Results IA Post Detection Processing Comparison between IA and PF Number of Real Targets Detection

5 4.5.2 Number of False Detection Target Detection Accuracy Detection History and Target Tracking Simulation Tool Results of History Tracking Chapter Summary References: Chapter 5: Wind Farm Modelling Results Introduction The Detection of Wind Turbine Post Detection Processing Multiple Transmitters Result Analysis (MTRA) The Detection of 5x5 Wind Farm PF Approach Particles Placement Approach PF Detection Result of the 5x5 Wind Farm IA Detection Result of the 5x5 Wind Farm Comparison between PF and IA Detection Results: Tracking Multiple Targets within Wind Farm Chapter Summary Chapter 6: Conclusion and Future Work Conclusion Multistatic Radar System Advantages Summation Tool Development Multiple Target Detection and Tracking Ghost Targets Elimination Target Localisation Algorithm Particle Filter (PF) Interval Analysis (IA) Comparison of IA vs. PF

6 Real targets detection rate: Ghost targets detection rate Detection Accuracy Future work PF and IA Hybrid Alternative Target Detection Methods Simulation tool implementation References APPENDIX 1: Calculating the Radar Returns A.1 Overview A.2 Radar Equation A.3 Receiver Noise Power A.4 Propagation Factor Own Publications Bibliography

7 List of Figures Figure 1.1: Monostatic radar Sidelobe detection near Wind Farm Figure 2.1: Bistatic Radar Geometry Figure 2.2: Bistatic and Monostatic iso-range contour difference Figure 2.3: Bistatic and Monostatic SNR Oval of Cassini Figure 2.4: Bistatic target range on Ellipsoid Figure 2.5: Bistatic Doppler geometry in 2D Figure 2.6: Classification of Bistatic and Multistatic radar systems Figure 2.7: Bistatic Accuracy illustration Figure 3.1: Simulation tool Antenna Component Figure 3.2: Bistatic geometry, Plane1 and Plane2 identification Figure 3.3: Optimising multistatic geometry illustration Figure 3.4: Flowchart of the simulation tool power received calculation Figure 3.5: Power Received with and without Propagation Factor (db), Target Range (km) Figure 3.6: Target movement illustration for power received plot Figure 3.7: Simulation tool algorithm of propagation factor regions Figure 3.8: The geometry of Bistatic propagation factor Figure 3.9: Propagation Factor (db) across Target Range (km) Figure 3.10: The correlation between propagation and power received Figure 3.11: Propagation Factor of two antennas at different elevations Figure 3.12: SNR for Bistatic Radar Figure 3.13: a) One Transmitter, Two Receivers SNR. b) One Transmitter, Three Receivers SNR Figure 3.14: a) Receiver bearing angle = 90 at the transmitter site. b) Bistatic pair of Receiver bearing angle = Figure 3.15: Two directional receivers (90 ) & one omnidirectional transmitter Figure 3.16: Bistatic pair of directional antennas Figure 3.17: SNR of bistatic Directional antennas, Horizontal view Figure 3.18: Individual Bistatic radar Pd plots Figure 3.19: The combined Multistatic radar Pd plots Figure 3.20: Pd comparison (Top three (Pd) approach Vs Old approach) Figure 3.21: General representation for wind turbine Dimension and Angles Figure 3.22: Wind turbine segments for RCS modelling Figure 3.23: SNR of multiple segments of a wind turbine structure Figure 4.1: Simulation tool Range-Only target detection Figure 4.2: Ghost target detection (in blue) Figure 4.3: MRA results, Real targets (in Red) and Ghost targets (in Blue) Figure 4.4: Comparison between one-off detection and MRA Figure 4.5: Targets vs ghost detections for 3 Receivers and 4 Receivers with MRA Figure 4.6: System geometry for three receivers Figure 4.7: Two solutions for one target detection Figure 4.8: Multi-thread process diagram Figure 4.9: PF steps visualisation

8 Figure 4.10: Simulation tool PF tracking one target Figure 4.11: PF multiple targets tracking and ghost elimination Figure 4.12: Separator s three output regions Figure 4.13: SIVIA algorithm routine Figure 4.14: 3D separator result of bistatic ellipsoid Figure 4.15: XY plane section from the 3D ellipsoid separator of range measurement Figure 4.16: Target detection algorithm routine Figure 4.17: The union of all measurements Separators Figure 4.18: Multistatic radar return from one target Figure 4.19: The intersection of all separators Figure 4.20: IA one target detection in XY plane Figure 4.21: IA one target detection in XZ plane Figure 4.22: The length of the estimated target elevation by IA Figure 4.23: The length of the estimated target (X/Y) by IA Figure 4.24: Post detection processing effects on false detection rate Figure 4.25: Post detection processing effects on detection accuracy level Figure 4.26: Multistatic sensors mapping with transmitters 1 and Figure 4.27: Number of detection from 20 targets return across 40 time steps Figure 4.28: Number of false detection from 20 targets return across 40 time steps Figure 4.29: Average detection accuracy of 20 targets return across 40 time steps Figure 4.30: PF Minimum-Average-Maximum detection accuracy Figure 4.31: IA Minimum-Average-Maximum detection accuracy Figure 4.32: Frequency of detection accuracy of IA and PF Figure 4.33: Target entry in history log list Figure 4.34: First process of History log target tracking Figure 4.35: Second process of History log target tracking Figure 4.36: Three pulses detection results Figure 5.1: Wind turbine Range-Only detection plot (XY plane) Figure 5.2: Wind turbine bistatic radar return at four receivers Figure 5.3: Wind turbine detection estimate by the IA approach Figure 5.4: Post detection processing of IA estimates for wind turbine (XY plane) Figure 5.5: Post detection processing of IA estimates for wind turbine (XZ plane) Figure 5.6: Comparison between (2Tx and 1Tx) detection for one turbine Figure 5.7: Four turbines detection (one transmitter and multiple receivers) Figure 5.8: Ghost targets elimination by MTRA technique Figure 5.9: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver a) Bistatic Range-Only measurements from Receiver b) Received1 radar return across the range line Figure 5.10: 5X5 windfarm detection Wind direction = a) Bistatic Range-Only measurements from Receiver1, wind direction = b) Received1 radar return across the range line Figure 5.11: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver a) Bistatic Range-Only measurements from Receiver b) Received2 radar return across the range line

9 Figure 5.12: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver a) Bistatic Range-Only measurements from Receiver b) Received3 radar return across the range line Figure 5.13: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver a) Bistatic Range-Only measurements from Receive b) Received4 radar return across the range line Figure 5.14: Trasmitter1 multistatic radar measurements of 5x5 Windfarm radar return Figure 5.15: Trasmitter2 multistatic radar measurements of 5x5 Windfarm radar return Figure 5.16: First iteration of particle placement process a) Grid generation step,(1km x 1 km) grid cell size. b) Cells elimination step Figure 5.17: Second iteration of particle placement process a) Grid generation step(500m x 500m) grid cell size. b) Cells elimination step Figure 5.18: Third iteration of particle placement process Figure 5.20: Fifth and final iteration of particle placement process Figure 5.21: a) Final step of particle placement process. b) randomly generated particles within the remaining grid cells Figure 5.22: PF detection of a 5x5 Wind Farm Figure 5.23: IA estimations of 5x5 wind farm (MTRA approach) Figure 5.24: IA targets estimation of 5x5 wind farm (1 transmitters and 4 receivers) Figure 5.25: Percentage of real targets detection of IA and PF of a 5x5 wind farm Figure 5.26: Number of false targets detection of IA and PF from a 5x5 wind farm Figure 5.27: PF accuracy of maximum, Average and minimum detection accuracy Figure 5.28: IA accuracy of maximum, Average and minimum detection accuracy Figure 5.29: Comparison between PF and IA detection accuracy Figure 5.30: IA output of XY plane of Ground-elevation multiple targets detection within 5x5 wind farm (Target elevation = 0) Figure 5.31: IA output of XZ plane of Ground-elevation multiple targets detection within 5x5 wind farm (Target elevation = 0m) Figure 5.32: IA output of XY plane of Mid-elevation multiple targets detection within 5x5 wind farm (Targets elevation = 100m) Figure 5.34: IA output of XZ plane of Mid-elevation multiple targets detection within 5x5 wind farm (Targets elevation = 100m) Figure 5.35: IA output of XY plane of High-elevation multiple targets detection within 5x5 wind farm (Targets elevation = 250m) Figure 5.36: XZ plane of High-elevation multiple targets detection within 5x5 wind farm (Targets elevation = 250m) List of Tables Table 2.1: Complexity levels of multistatic radar systems 39 Table 3.1: Single pulse SNR (db) in correlation to Pd and Pfa 66 Table 4.1: Numeric illustration of Five Particles sampling process.94 9

10 Abbreviations 2D 3D ABR ATC CW db dbms dbv I/O IA LOS MeMeRA MRA MTI MTRA Radar RCS SIR SIS SIVIA SNR PBR PCL Pd PF Pfa PHD PRF Two Dimensional Three Dimensional Active Bistatic Radar Air Traffic Control Continuous Wave Decibels Decibel Metre Squared Decibels Volts Input/Output Interval Analysis Line Of Sight Rectangular Meshing for RCS Approximation Multiple Results Analysis Moving Target Indicator Multiple Transmitters Results Analysis Radio Detection and Ranging Radar Cross Section Sampling/Important Resampling Sequential Important Sampling Set Inverter via Interval Analysis Signal to Noise Ratio Passive Bistatic Radar Passive Coherent Location Probability of detection Particle Filter probability of false alarm Probability Hypotheses Density Pulse Repetition Frequency 10

11 Abstract The effects of wind farm installation on the conventional monostatic radar operation have been investigated in previous studies. The interference on radar operation is due to the complex scattering characteristics from the wind turbine structure. This research considers alternative approach for studying and potentially mitigating these negative impacts by adapting the multistatic radar system technique. This radar principle is well known and it is attracting research interest recently, but has not been applied in modelling the wind farm interference on multistatic radar detection and tracking of multiple targets. The research proposes two areas of novelties. The first area includes the simulation tool development of multistatic radar operation near a wind farm environment. The second area includes the adaptation of Range-Only target detection approach based on mathematical and/or statistical methods for target detection and tracking, such as Interval Analysis and Particle Filter. These methods have not been applied against such complex detection scenario of large number of targets within a wind farm environment. Range-Only target detection approach is often considered to achieve flexibility in design and reduction in cost and complexity of the radar system. However, this approach may require advanced signal processing techniques to effectively associate measurements from multiple sensors to estimate targets positions. This issue proved to be more challenging for the complex detection environment of a wind farm due to the increase in number of measurements from the complex radar scattering of each turbine. The research conducts a comparison between Interval Analysis and Particle Filter. The comparison is based on the performance of the two methods according to three aspects; number of real targets detected, number of ghost targets detected and the accuracy of the estimated detections. Different detection scenarios are considered for this comparison, such as single target detection, wind farm detection, and ultimately multiple targets at various elevations within a wind farm environment. 11

12 DECLARATION No portion of the work referred to in the thesis has been submitted in support of an application for another degree or qualification of this or any other university or other institute of learning. COPYRIGHT STATEMENT i. The author of this thesis (including any appendices and/or schedules to this thesis) owns certain copyright or related rights in it (the Copyright ) and s/he has given The University of Manchester certain rights to use such Copyright, including for administrative purposes. ii. iii. iv. Copies of this thesis, either in full or in extracts and whether in hard or electronic copy, may be made only in accordance with the Copyright, Designs and Patents Act 1988 (as amended) and regulations issued under it or, where appropriate, in accordance with licensing agreements which the University has from time to time. This page must form part of any such copies made. The ownership of certain Copyright, patents, designs, trade marks and other intellectual property (the Intellectual Property ) and any reproductions of copyright works in the thesis, for example graphs and tables ( Reproductions ), which may be described in this thesis, may not be owned by the author and may be owned by third parties. Such Intellectual Property and Reproductions cannot and must not be made available for use without the prior written permission of the owner(s) of the relevant Intellectual Property and/or Reproductions. Further information on the conditions under which disclosure, publication and commercialisation of this thesis, the Copyright and any Intellectual Property University IP Policy (see in any relevant Thesis restriction declarations deposited in the University Library, The University Library s regulations (see and in The University s policy on Presentation of Theses. 12

13 Acknowledgement First and foremost, I would like to thank my supervisor, Professor Anthony K. Brown, for the opportunity to undertake this work. His continuing support and guidance from the start of this research has been invaluable. It has been a privilege to work under his supervision, and I am deeply grateful. I also owe thanks and my sincere gratitude to my co-supervisor Dr Laith Danoon. This research would have not been possible without his outstanding research contribution within the field of wind farms RCS modelling. I am thankful for his support throughout the four years of this research. I would like to thank Dr Alexandru Stancu for his guidance and collaboration work that helped me to understand the Interval Analysis theory. His experience and advice in Control Systems and Measurements Association proved to be invaluable for this work progress. On a personal note, I would like to thank my parents for all the love and encouragement. Without their support I would not have been able to be where I am today. My mother, thank you for the kind words and wise advice in all my pursuits. When I felt down and overwhelmed you were always there to comfort me and encourage me to keep going. My father Professor Khalid Hameed has been an inspiration to me in life and in his academic ethics and achievements. I aspire to do him proud. Thank you for your never ending love and support. I am thankful for the constant encouragement from my sisters. I am truly blessed to have been brought up in such a close family, full of people that I love. Last but not least, I would like to give my special acknowledgment to my beloved beautiful wife Myga the mother of our adored baby daughter Aya. I have been away from them for so long during this study. Their love and support kept me motivated. Thank you with all my love. 13

14 Chapter 1 Introduction 1.1 Overview Wind farms cover vast areas of land or sea, with extremely large wind turbines structures that can reach up to 180m height, particularly in offshore installations. These installations potentially affect nearby radar systems including civilian radar [1], such as Air Traffic Control (ATC), marine, coastal, weather radars as well as defence installations. Radar performance is compromised on the basis of both target position and radar accuracy [2]. The negative impact is caused by the high Radar Cross Section (RCS) from the turbine structure and/or the Doppler shift created by the turbine blades [2]. Different types of radar are affected differently by wind farms. For instance, marine radars are low cost systems which normally do not employ Doppler filtering and therefore are more likely to be affected by the high RCS of wind farms turbine structure [3]. On the other hand ATC radars are more sophisticated, incorporate Doppler techniques, so that the Doppler shift generated from the turbine blades is of importance since the blade tips move at speeds that might break through the moving target indicator (MTI) filters. The wind farm can cause the tracker to initiate a new false track or alter the track of an aircraft passing above the wind farm [1]. Section 1.3 will explain these negative impacts in more detail. This research aims to study and potentially mitigate the negative impact of wind farms on radar operation by considering the Multistatic radars approach. Multistatic radar systems can be considered as a network of Bistatic radar pairs of spatially separated antennas of transmitters and/or receivers [4]. This principle is well known and it is attracting research interest recently, but has not been applied in modelling the wind farm effects in the literature. This radar approach operates well in situations involving a relatively small 14

15 number of targets. The research extends multistatic radar for use within a large and complex wind farm environment, which could potentially generate hundreds of targets from the turbines in addition to a small number of wanted targets. This research maximizes on the advantages of multiple sensors and design flexibility of multistatic radar systems to mitigate the potential wind farm impact. The basic principles of multistatic radar are extended to use advanced target detection and tracking methods for the multiple sensors to allow the use of multistatic radar in high target densities. The approaches used include, Interval Analysis (IA) and/or the statistical approach of Particle Filter (PF) for target tracking. To critically evaluate these techniques a computer simulation tool for Multistatic radar system was developed during the course of this research. The models were used alongside integrated wind farm model which computes the radar returns from individual turbines. This made it feasible to model the impact of wind farms and compare different multistatic radar processing techniques in this complex radar environment. 1.2 Motivation The UK government has set a policy to meet 15% of UK energy demand from green renewable energy by 2020 [5], where electricity generated from wind farms is considered as a significant contributor toward that target. However, wind farms development needs to meet the regional agencies permission. Many wind farms site planning were objected on the basis of potential radar interference, as it might impose safety or security hazard by interfering with nearby radar systems [3]. Wind farms development started with Round 1 planning during the 1990s and The sites were mainly onshore or by the costal bases, as the number and size of wind turbines were small in comparison to the current wind farm technology. Even so many wind farm planning applications were refused on the basis of radar interference [6]. The current Round 2 of wind farms planning and development is characterised by much bigger wind turbines and much larger farms area because of the move toward remote offshore locations. However, wind farms planning are still facing objections due to 15

16 the radar interference, especially for marine radars used by shipping lines and fishing boats [6]. The push for more sites and bigger contribution from wind energy will continue if the possibility of meeting the 15% of the UK energy demand target is to be fulfilled. Yet those plans will face increasing objections based on radar interference affects. It is an increasingly acute issue and further research into understanding and reducing the impact of wind farms on radar performance is required. This research aims to address the above issue by studying the impacts of wind farms on the Multistatic radar system operation. Specifically, studying the impact on multiple targets detection and tracking within the wind farm environment. The developed simulation tool examine multiple methods (i.e. IA and PF) for targets position estimation and tracking, to better understand and possibly mitigate the negative impacts of wind farm. Therefore, the research motivation is to resolve the objections against new wind farms plans by enhancing the radar target detection tracking performance in such complex detection scenario. Data and measurements from field experiments for wind farm impact on radar operation tend to be high in cost and effort, especially for the case of multistatic radar systems. Multiple sensors at different location, configurations and antenna parameters increase the complexity and cost of the trials. Therefore, a computer simulation tool is needed, to reduce the cost and effort when predicting the impact of wind farms on different multistatic radar formations and sensors configurations. This research was conducted within the Microwave and Communication Systems research group at the University of Manchester where wind farm impact on radar operation is one of the principle research interests. The research presented in [2] aimed to investigate and model the impacts of wind farms by modelling the RCS of a wind turbine for Monostatic marine radars. This project builds on this previous work and introduces a new approach in modelling the impact of wind farms by using Multistatic radar techniques. 16

17 1.3 Impact of Wind Farms on Radar The negative impact of wind farms on the radar performance may arise due to the high RCS of the wind turbines structure and the constantly changing orientation of the turbine blade, which is affected by the rotation of the blades and the orientation of the turbine to face the wind direction. This causes pseudo-random radar return and Doppler shift effects that are difficult to predict [2]. For monostatic radar the high static RCS characteristic of the wind turbine can cause a turbine to be detected as an arc at a given range rather than a point, or the turbine might be detected in the antenna sidelobes causing multiple detections across the azimuth angle. In addition the possibility of multiple reflection of the radar signal from one turbine to another may cause the radar to detect ghost targets as false returns [3]. Wind farms effects ATC radars by causing twinkling effect, track seduction or shadowing the target [7]. The twinkling effect happen when the radar detects a turbine blade in one rotation and then detect a different turbine blade in the next radar rotation, which can cause the radar to initiate a track and lock on object. The track seduction is caused by detecting wind turbine twinkling effect while an airplane is nearby which can affect the track to be altered [1] Impact on receivers Radar receivers can be over saturated with the large reflection of the radar signal from the wind turbines due to the fact that shape, structure and materials of the turbine that creates high RSC effect. Over saturated receivers can cause smaller targets to be lost [2] Sidelobe detections The high RSC of wind turbine can cause the turbine to be detected through the sidelobes of the radar antenna at different azimuth angles. This can cause multiple false targets detection at the turbine range [2]. In the case of marine radars where radar is located near other masts and cables, the ships super-structure may cause a distortion to the antenna radiation pattern which will add to the severity of this effect. This impact of 17

18 signal distortion causes the return from each individual turbine to be merged on the radar display, which over crowds the display and makes it hard to keep track of targets [2]. Figure 1.1: Monostatic radar Sidelobe detection near Wind Farm [3] Shadowing A significant shadowing effect may arise due to the large size of the wind turbines structure and the high number of turbines within a wind farm site. This can cause the radar to fail to detect targets within the shadowing area as well as the wind turbines that are at the back of the wind farm orientation. There are many factors that will impact the severity of this effect such as the orientation of wind farm, the radar distance from the farm and the geometry of the wind turbine [2] Multiple reflections The radar signal can be reflected between the turbines before receiving it back at the receiver site which can cause false detections. This effect is particularly obvious when the radar is operating close to the wind farm. The signal multipath can affects the efficiency of the radar system by miscalculating the detected target position due to incorrect time different measurement [3]. 18

19 1.4 Multistatic Radar Range-Only Detection within Wind Farm Environment The radar return of a target at a sensor site is converted into range measurement that signifies the total distance of signal travel from transmitter to target then to receiver. The range measurement is estimated from the time different between the transmission and the reception times of the signal. The Range-Only detection approach is based on range measurements association from multiple sensors [9]. The measurement association is a process of correlating a set of Range-Only measurements from multiple sensors which corresponds to the same target. This is done by sending the registered detection information to a central data fusion unit to calculate targets positions. The detection of large number of targets implies higher complexity for the detection process. This is due to the increase in number of range measurements that require association, which will impose greater challenges on the performance and accuracy of the detection process [9]. According to the previous section of wind farm impacts, the detection of the wind farm exhibit a challenging task for the Range-Only approach. Each wind turbine can cause large and complex radar return potentially causing the detection process to identify one turbine as multiple targets detection (after association). Furthermore, a large number of turbines within a wind farm will result in a very large number of targets within a uniform and dense formation. Additionally, the multiple reflections of radar signal between turbines towers can cause false targets estimation [3]. In summary, the Range-Only approach present simplistic method for target detection by associating the data from multiple sensors. However, there are limitations and challenges of this approach when detecting large number targets. This research investigates these limitations by simulating the complex returns of a wind farm in addition to multiple small targets within and around the wind farm area. The models are then used to detect and track all targets within the region. 19

20 1.5 Research Novelty There are two main areas of novelty work in this research. The first area includes the simulation of multistatic radar operation near a wind farm environment. The second area includes the adaptation of Range-Only target detection approach based on mathematical and/or statistical methods for target detection and tracking. Novelties from the first area of the simulation tool: During the course of this research, a computer simulation tool of multistatic radar system was developed to be used as a tool for multistatic radar system design and operation. The originality of this simulation tool is the integration with the adapted wind farm scattering models to investigate target detection within the complex wind farm environment. Novelties from the second area of target detection and tracking include: Addressing the challenge of ghost targets elimination when processing large number of targets detection (large number of measurements from each sensor). Subsequently, this research proposes a new method of multiple detection results analysis to eliminate inconsistent results as ghost target. The multiple results are acquired by remapping the multistatic radar sensors into smaller subgroups as shown in Chapter 4, Section Critical analysis of a Range-Only localisation algorithm [10] using multiple targets detection and bistatic radar range resolution. A novel implementation of PF algorithm for Range-Only detection method within a complex wind farm environment. This includes new algorithms to alter the particles placement steps of PF. A new approach for implementing the IA method for Range-Only target detection in the three dimensional (3D) space. This includes the original implementation of the IA method within a wind farm environment. The research presents a comparison between PF and IA methods based on detection performance in complex environment. I.e. multiple targets and within a 20

21 wind farm. The comparison categories consider the accuracy, the number of false targets and the rate of successful detections of multiple targets. 1.6 Thesis outline The thesis is composed of six chapters. The first chapter introduces the potential impact of wind farm on the operation of radar systems. The research aim and objectives are also highlighted within this chapter along with the novelty and contribution of this research. The second chapter covers the background knowledge of bistatic radar and multistatic radar system. It will then cover the basic background of the bistatic radar equation for power received and illustrates the geometry components of a bistatic radar pair. The characterization of bistatic and multistatic radar systems are discussed according to system design and performance. Finally, an analytical view of the advantages and disadvantages of bistatic and multistatic radar system is presented. This was then used to understand the possible advantages that can be utilised for this research and to recognise the challenges of simulating multistatic radar system. The third chapter introduces the main components of the developed simulation tool for the assessment of multistatic radar system within a wind farm environment. The chapter starts with the simulation tool results of bistatic and multistatic radar key parameters, such as power received, bistatic propagation factors, signal to noise ratio, antenna patterns and probability of detection. The second part of the chapter concentrates on the wind turbine and wind farm modelling tool highlighting the main features and assumptions. Chapter four introduces the Range-Only detection concept for multistatic radar system and explains the concept of the ghost targets detection phenomena. The localisation algorithm [10] was explained and an analytical review of this algorithm s feasibility and limitation in multiple targets detection is presented. The chapter then discusses other mathematical and statistical methods for Range-Only detection, i.e. PF and IA method. A comparison between PF and IA is shown. Finally, the process of history tracking approach is explained. 21

22 Chapter five presents the results and challenges generated from one wind turbine detection using IA approach. The chapter then introduces a 5x5 wind farm scenario used to compare PF and IA. The chapter ends on comparison evaluation between the two methods detection performance of the given wind farm scenario. The final chapter summarises and concludes the findings of this research and highlights some of the key areas where this research can be extended in future work. 1.7 Chapter Summary The research considers multistatic radar approach for studying and potential mitigating the impacts of wind farms on radar operation. Multistatic radar system is defined as a network of Bistatic radar pairs of spatially separated antennas of transmitters and/or receivers. This radar approach is attracting research interest recently, but has not been applied in modelling the wind farm effects in the literature. The basic principles of multistatic radar are extended to use advanced target detection and tracking methods for the multiple sensors. These detection methods include, Interval Analysis (IA) and/or the statistical approach of Particle Filter (PF) for target tracking. The objective is to address the challenging detection process of high number of targets within a dense area, such as the wind farm installation. A computer simulation tool for multistatic radar system was developed during the course of this research, to critically evaluate the implemented detection techniques. The developed simulation tool was used alongside an integrated wind farm model to compute the radar returns from individual turbines. This made it feasible to model the impact of wind farms and compare different multistatic radar processing techniques for target detection and tracking. The negative impact of wind farms on the radar performance may arise due to the high RCS of the wind turbines structure and the constantly changing orientation of the turbine blade, which is affected by the rotation of the blades and the orientation of the turbine to face the wind direction. This causes pseudo-random radar return and Doppler shift effects that are difficult to predict. These unpredictable Doppler shift effects and the high reflectivity of turbine structure produces a complex targets for the radar to detect, which 22

23 can negatively impact the radar performance when it comes to multiple targets detection and tracking. The UK energy policy aims to meet 15% of UK energy demand from green renewable energy by 2020 [5], where wind farms are considered as significant contributor toward that target, as a green source of electricity generating. The motivation of this research is derived from the increasingly acute issue between the posh for more sites and bigger contribution from wind energy to meet that policy aim- and the increasing objections on new wind farm development based on radar interference effects. Therefore, the research aims to address this issue by better understanding the impact of wind farm on radar operation. Specifically, studying the impact on radar detection and tracking of multiple targets. In addition to the above motivation, data and measurements from field experiments for wind farm impact on radar operation tend to be high in cost and effort, especially for the case of multistatic radar systems. Therefore, a computer simulation tool is needed, to reduce the cost and effort when predicting the impact of wind farms on different multistatic radar formations and sensors configurations. The Range-Only approach for target positioning is considered by this research. This approach is based on range measurements association from multiple sensors [9]. The measurement association is a process of correlating a set of Range-Only measurements from multiple sensors that corresponds to the same target. This is done by sending the registered detection information to a central data fusion unit to calculate targets positions. This approach will be discussed in details within chapter 4. The novelty of this research can be characterised into two areas. The first area includes the simulation of multistatic radar operation near a wind farm environment. The second area includes the adaptation of Range-Only target detection approach based on mathematical and/or statistical methods for target detection and tracking, such as Interval Analysis and Particle filter. 23

24 1.8 References 1. Wind turbines and aviation interests European experience and practice. 2002, BWEA Aviation Group, ESTU W/14/00624/REP. 2. Laith S Rashid, Impact Modelling of Offshore Wind Farms on Marine Radars, The University of Manchester, MARICO, Investigation of Technical and Operational Effects on Marine Radar Close to Kentish Flats Offshore Wind Farm, 2007, BWEA Report. 4. N. Willis, Bistatic Radar, 1991, Artech House. 5. National Renewable Energy Action Plan for the United Kingdom Article 4 of the Renewable Energy Directive 2009/28/EC 6. Wind turbines and aviation interests European experience and practice. 2002, BWEA Aviation Group, ESTU W/14/00624/REP. 7. OPERA, Impact of Wind Turbines on Weather Radars, 2006, WP T. Lyul Song, D. Musichi, Range only Multistatic Tracking in Clutter, Department of E;ectronic systems Engineering, Hanyang University, Republic of Korea, W. Dou, P. Willett and Y. Bar-Shalom, Fusion of Range-Only Measurements from Multistatic Configurations for air Collision Warning, Department of Electrical and Computer Engineering, University of Cannecticut Storrs, CT , M. Malanowski, An Algorithm for 3D Target Localization from Passive Radar Measurements, SPIE Vol B-1, J Pinto, A. K. Brown, L. Rashid, Z. Moore, Requirements Capture Summary Report, Stealth Technology for Wind Turbines, TP/2/RT/6/I/10117 APPS2B, April A. Bishop, P. Pathirana, Localization of Emitters via the Intersection of Bearing Lines: A Ghost Elimination Approach Howard, M. and C. Brown, Results of the electromagnetic investigations and assessments of marine radar, communications and position fixing systems undertaken at the North Hoyle wind farm by QinetiQ and the Maritime and Coastguard Agency, 2004, MCA Report MNA 53/10/ I. A. Siradjuddin, M R. Widyanto, Particle Filter with Gaussian Weighting for Vehicle Tracking, Trunojoyo University, Indonesia, April 27,

25 Chapter 2 Bistatic and Multistatic Radar Systems Background 2.1 Overview This chapter aims to provide background knowledge of the radar technologies that are relevant to this research study. The chapter starts with defining the bistatic radar system and how it is different form the conventional monostatic radar system, all categories and classifications of such radar system are explained. After that, the chapter will showcase the resent research interest in bistatic radar systems, along with historical view of this radar approach. Afterward, the technical aspects of the bistatic radar system are discussed from the geometry setting, radar equation, target location and finally the Doppler shift. After establishing the understanding for bistatic radar as technical component, the chapter introduces the multistatic radar systems concept. The classification of multistatic radar systems is discussed according to the system complexity level, data fusion approach and level of coherent. Finally, an analytical view of the advantages and disadvantages of bistatic and multistatic radar systems is conducted; to understand the possible advantages that can be utilized for this research overall objectives, also to recognise the challenges of simulating multistatic radar system. 25

26 2.2 Monostatic, Bistatic and Multistatic Radar Systems Monostatic radar is the most commonly adopted architecture for radar surveillance. In its simplest form a pulse is transmitted from a single radar antenna, and the reflected return from targets received by the same antenna [1]. Timing the difference between when the pulse was transmitted and when it was received gives the range of the target. Bistatic radar systems operate on the basis of physical separation between transmitter and receiver [2]. However, Bistatic systems are further sub-divided into Active Bistatic Radar (ABR) or Passive Bistatic Radar (PBR) [3]. In ABR the transmitter is under the control of, and is integrated with, the bistatic system. In PBR the transmitted signal used is a signal of opportunity (that is one that already exists) such as a mobile communications signal [3]. The PBR then is simply a receive only system. This section provides definitions for both Active and Passive bistatic radars, along with a comparison between the Passive Coherent Location (PCL) and Passive bistatic Radar Active Bistatic Radar Active bistatic radar system is defined by IEEE as a radar system that uses antennas at different locations for transmission and reception [2]. In other words, the radar system separates the transmitter and receiver from each other by considerable distance (the baseline); in contrast to the Monostatic radar system which uses the same antenna for transmission and reception. The term Active means the transmitter is part of the radar system components so that the system has full control on the transmitter. Hence the location, modulation, frequency and transmit phase of the transmitter are all pre-set and known by the system Passive Coherent Location (PCL) The Passive Coherent Location is classified as bistatic system. Instead of using dedicated transmitter, the PCL takes the advantage of existing transmitter within the area as transmitter of opportunity. Such as TV broadcaster, GSM base station or even a close by Monostatic radar [4]. 26

27 2.2.3 Passive Bistatic Radar (PBR) Passive Bistatic Radar is a radar system that separates between the transmitter and receiver, knowing that the transmitter is a non-cooperative source of illumination, such as a broadcast or mobile communications signal. Therefore transmitter generation is not part of the system components [5]. The two definitions above (PCL and PBR) are quite similar. After all, both describe the transmitter antenna as non-cooperative and out of the system component. However, literatures denotes to PBR as a subset of bigger group as PCL [4]. The reason for that, the PCL system could be constructed from various sets of PBR. Beside, PCL system configuration works on utilizing transmitter of opportunity based on availability. In contrast to the PBR which agrees on the third party dedicated transmitter of opportunity [4]. 2.3 Bistatic Radar Systems Interest Cycles Before the invention of the duplexer component, radar systems from the early 1930s were operated as bistatic radar systems [1]. The transmitter was sited away from the receiver location to prevent interference with the receiver functionality; especially when early radar systems were continues wave (CW) type. After the development of radar antennas with duplexer capabilities (switching between transmission and reception), Monostatic radar system became more and more popular. Meanwhile, Bistatic radar system went through cycles of interest approximately every 15 years [6]. H. Griffiths portrayed those cycles of interest into three resurgences (see section 2.4). Some resurgence was based on interest in trade-off between bistatic and Monostatic radar systems, based on criteria of cost, maintenance or performance. And other resurgence was based on new target properties such as the anti-radiation missiles [5]. Currently the research interest in bistatic radar system is passing through the third resurgence, where the technology is at the peak of interest cycle [6] due to the advantages of bistatic radar. For instant, the detection of stealthy shaped targets where the reflected signal at different trajectories than the transmitter direction can be captured by the separated receiver at different site. Furthermore, the synchronisation 27

28 and geo-location difficulties in the past became less of a worry for the bistatic radar implementation, as the advancement in computing power provided the support for the complex synchronisation requirement and the GPS technology made the geo-location difficulties easier to manage. 2.4 History of Bistatic and Multistatic Radar Bistatic radar history started from the early experiments during the 1888 and the 1904 of transmitting electromagnetic wave, those experiences are considered as a foundation for the later on bistatic radar systems development. Sequentially, this section is going to walk through the history of bistatic radar from the aspects of first experiment, first documentation and first paten bistatic radar. After that, the section looks at examples of bistatic radar development based on countries of origin. Finally the three resurgences of interest in bistatic radar are explored along with the reasons that motivate those resurgences. N. Willis book Bistatic Radar dedicated full chapter (chapter 2)[2] about the history of bistatic, starting with the first experiments 1904 that are not necessary named as bistatic radar development experiments, yet those experiments were regarded as the fundamentals of electromagnetic wave transmission and reception [2]. Other literature [5] goes back to 1888 when H. Hertz proved by experiment the existence of electromagnetic waves. Hertz succeeded in transmitting electromagnetic wave between a spark-gap and a sensitive wire coils, and then he managed to illustrate how the electromagnetic wave could be reflected of a metal mirror from the transmitter toward the receiving coils. This experiment was not directly meant for bistatic Radar development, but it can be considered as a first demonstration of the concept [5]. The telemobilityscope is the name of the first radar developed by Christian Hulsmeyer in the He was successful in detecting a passing ship, and his radar model was based on separated transmitter and receiver to isolate the receiver away from the transmitter interference. The patent was issued in 1904, but the new invention failed to attract the necessary support from neither the traditional-bound naval authorities nor the public companies. For that reason the experiments failed to progress [2]. 28

29 Fourteen years after Hulsmeyer experiments, H. Gernsbach the editor of the Electrical Experimenter interviewed Nikola Tesla for the August 1917 edition. The outcome of that interview was regarded as the first documentation of bistatic radar concept [2]. Taylor and Young of the Naval Research Laboratory issued their first patent in 1934 for their earlier demonstration in 1922 of bistatic radar detection for ships. Consequently, Taylor and Young work effects of wave interference opened the opportunity for the development of radar technology into the Monostatic radar type in the late 1930s [3]. The development of radar systems throughout the early two resurgences was independent in many countries at the same period. Therefore, N. Willies illustrated the historical development of bistatic radar based on the country of development [2]. In the United Kingdom a secret memorandum titled Detection and location of aircraft by radio methods by R.A. Watson- Watt in 1935 was sent to the UK Air Ministry, which led to Deventry experiment for designing CW bistatic radar. The radar used the BBC Empire short-wave station at Deventry as a transmitter, and a receiver was located at Weedon in Northamptonshire. The radar successfully managed to detect the aircraft put for the test by displaying fluctuation in the receiving signal [5]. The success of this experiment was very important step toward the development of Home Chain radar system in 1939, which was the first bistatic radar system in the United Kingdom to proving air defence across the English Channel [5]. The airborne bistatic radar development was a subject of interest in the UK. In the late of 1936 a project was led by Dr E. G Bowen as the first trail of airborne bistatic radar concept. The receiver was fitted on Heyford bomber and the transmitter was at ground base. The detection range was mostly accurate when the target lies between the aircraft and the ground transmitter, which was a limitation from the system [5]. In Germany, The first hitchhiked bistatic radar was developed by Klein Heidelberg in 1944 during the Second World War. Klein radar utilized the British Chain Home radars transmitted signal to detect aircrafts across the English Channel [7]. The radar provided very good detection with only using silent receivers in Oostvorne, Holland [5]. 29

30 Japan developed their air defence system in the 1940s as Type A Radar CW bistatic radar system. The longest baseline record of Type A radar was between Taiwan and Shanghai, hence the distance was above the 400 miles [5]. H. Griffiths portrayed those cycles of interest into three resurgences. The first resurgence was around the 1950s, which was mainly characterised by the development and deployment of bistatic radar as forward-scatter fences and as precision test range instrumentation, in addition to satellite tracking systems [7]. During this resurgence, the bistatic radar terminology was established in 1952 [2]. Although, the first time the term bistatic was used by Seigel in 1955 to describe the scattered radar energy [8]. The second resurgence was in the 1970s and 1980s, which was influenced by the new development of anti-radiation missile in addition to retrodiractive jammer threats. Here the geometry of bistatic radar became of an interest to counter those threats [2]. And finally, the current research and development interest in bistatic radar is conceder as the third resurgence [2]. 2.5 Geometry of Bistatic Radar Figure 2.1 illustrates the basic geometry of bistatic radar. The transmitter (T), target (Tr) and receiver (R) form a triangle that is referred to as the bistatic radar triangle, and the distance (L) between the transmitter and receiver is called the baseline. When target is detected, the signal travels from transmitter-to-target on the distance Rt, then it reflects of the target toward the receiver traveling the distance Rr. The total distance of transmitter-target-receiver (Rt + Rr) represent the total indirect path (Scatter path or Range Measurement). On the other hand, the direct path is the distance from the transmitter-to-receiver across the baseline which represents the shortest path between the two antennas. The bistatic angle (β) is the top angle of the bistatic radar triangle at the scattering point of the target. When a target is above or below the baseline, the bistatic angle will be less than 180 (β<180 ); in this case, the indirect path is longer than the baseline. In the event of (β = 180 ), then the target will be between the transmitter and receiver along the baseline. This case represents the forward scatter geometry. And finally the target could 30

31 be behind the antennas along the extended line of the baseline. This follows quasibistatic characteristics [2]. The angle (φ) represents the transmitter azimuth angle, and (θ) is the receiver antenna azimuth angle. The last angle is important to know (or estimate) for the calculation of target location, as it signifies the Angle of Arrival. Figure 2.1: Bistatic Radar Geometry Figure 2.2 below shows the difference between the Monostatic and Bistatic iso-range contours [9]. The Monostatic radar target range forms a circle around the radar sensors, while the Bistatic radar target range (Rt + Rr) forms an elliptical shape around the baseline. The two antennas are located at the two focal points of the ellipse, where the target could be at any point on that ellipse. Figure 2.2: Bistatic and Monostatic iso-range contour difference [9]. 31

32 2.6 Bistatic Radar Equation The power received equation of bistatic radar is similar in concept to the Monostatic equation. In fact, this section will derive the power received equation based on the Monostatic equation. According to the geometry of bistatic radar, this radar system have two different target-to-antenna ranges, two different antenna gains, two different propagation factors and antenna losses [1]. Therefore, bistatic equation could be divided into two sections, Transmitter-to-Target and Target-to-Receiver calculation. The first section of the equation is to find power flux density from the transmitter to target as shown in equation 2.1. Where (Pt) is power transmitted, (Gt) is transmitter antenna gain, (Rt) is transmitter-to-target range and (Ft) is pattern propagation factor (Transmitter-to-target) Trasmitter- to- Target PG t t F = 4 R 2 t 2 t (2.1) The second section is to measure the scattered power flux density from the target toward the receiver. Therefore, bistatic RCS (σ b ) is included along with the spherical area at the receiver site. Where (R r ) is the distance between target-to-receiver equation 2.2. Target - to- Receiver PG t rf b 1 = (2.2) 4 R 4 R 2 r 2 t 2 r The final step is to add the power received part at the receiver site by an antenna of effective aperture (Grλ 2 /4π). Where (Gr) is the receiver antenna gain and (λ) is the wavelength, In addition to the propagation factor over the target-to-receiver path (Fr). As shown in equation 2.3. P r P G G F F λ t t r t r b (2.3) R t R r The above equation is simplified to illustrate the basic concept of bistatic equation and how it could be driven from Monostatic equation. In fact, when substitute (Gt, Gr) with one antenna gain (G 2 ), (Ft and Fr) with one propagation factor (F 2 ) and (Rt and Rr) are the same distance as R 2, then the result will be a Monostatic radar equation. Additional 32

33 elements could be added such as antenna loss for transmitter and receiver and noise temperature. The SNR ratio plot of bistatic radar takes the shape of Oval of Cassini in contrast to Monostatic radar where the SNR plot is circular sphere. The Oval of Cassini shape is defined as the locus of the vertex of a triangle when the product of the sides adjacent to the vertex is constant and the length of the opposite side is fixed [2]. To link this definition with figure 2.3 below, the bistatic baseline represents the length of the fixed opposite side, and (Rt, Rr) are at the sides adjacent to the vertex, finally the vertex is where the target is located. Figure 2.3: Bistatic and Monostatic SNR Oval of Cassini [12]. 2.7 Bistatic Target Location The basic requirement from a Bistatic radar system is to fulfil target detection and decide the three basic detection elements, firstly the presence of the target, secondly the location of that target and finally the Doppler as a component of velocity [3]. The range measurement (Rt + Rr) is calculated in two different methods, as direct or indirect method. For the direct method, the two antennas are within the Line of Sight (LOS) of each other, thereby the baseline distance is known, in addition to the time different between receiving the direct signal and the indirect signal ( T tr ). Therefore, the range measurement is calculated as (Rt + Rr) = c T tr + L, where (c) is speed of light and (L) is baseline range [1]. 33

34 On the other hand the indirect method is used in case of no LOS between the two antennas; thereby the baseline distance is unknown and no direct signal from the transmitter to receiver. Instead, the two antennas are synchronised by stable clocks so the receiving site can estimate the time different ( T tt ). It then calculates the range measurement as (Rt + Rr) = c T tt [1]. In the three dimensional (3D) presentation, a target range measurement crates the bistatic ellipsoid (Prolate spheroid) [3] around the two focal points where the transmitter and receiver are located, as shown in Figure 2.4. To locate the target at that elliptical range, the intersection point between the ellipsoid surface and the vector from receiver to target of angle of arrival is taken. Target range from the receiver is calculated from equation 2.4 below. Where S is the range sum and (ψ e ) is angle of arrival for the two dimensional elevation plane. R r = S 2 L 2 2(S L cos ψ e ) (2.4) Figure 2.4: Bistatic target range on Ellipsoid. In the case of target range (Rt + Rr) is equal to the range between transmitter and receiver, then the target is undetectable with this method, and the target could be located anywhere along the baseline, hence the ellipsoid collapses over the baseline range. 34

35 2.8 Bistatic Doppler Frequencies The Doppler shift in signal frequency occurs due to the change in total path length travelled by that signal. Thereby, the Bistatic radar Doppler shift (f D ) is measured by calculating the different in target range measurements, to the ratio of the radar wavelength (λ) [1], as shown in equation 2.5. f D = 1 λ d dt (R T + R R ) (2.5) In the case of moving target with velocity component (V) and velocity angle (α), they are incorporated into Doppler calculation, where the Doppler shift equations will be written as equation 2.6 and 2.7 [1]. The angle (Φ) represents the angle between transmitter to target range and baseline, and (θ) represents the angle of arrival at the receiving site, as shown in figure 2.5. f D = V λ (cos(θ α)) + (cos(φ α)) (2.6) Or f D = 2V λ θ+ Φ cos (α ) cos (θ Φ ) (2.7) 2 2 Figure 2.5: Bistatic Doppler geometry in 2D [1]. The format of the Doppler shift in equation 2.7 insinuates three Doppler parts. The first part is ( 2V ) which signify the maximum Doppler frequency term, the second part is ( λ cos (α θ+ Φ 2 ) ) which illustrate the dependence on the target s trajectory, and the third part is (cos ( θ Φ ) ) which includes the bistatic angle [5]. 2 35

36 2.9 Multistatic Radar Systems Multistatic radar is a radar system which consists of multiple transmitter/s and receiver/s that are separated on wide area [3][9][10]. This radar system employs variant types of antennas according to application type. Therefore, Multistatic radar could include multiple sets of bistatic and/or Monostatic radars [6]. Multistatic radar has a number of different terminologies used by literatures; for example, netted radar, radar network, multisite radar, distributed radar and Multiple Input Multiple Output (MIMO) Radar [10]. Multistatic radars can be visualized as multiple connected nodes of radar network. Each node could have one out of three functional component options, transmitter, receiver or transmitter and receiver at the same location. This aspect provides the capability of having different parameters setting for each node, during the design stage of the radar system. As for a receiver type node, the radar designer is able to make a choice in frequency band, bandwidth and polarisation for every individual receiver. Similarly, at the transmitter nodes, the system could be optimised by customising the carrier frequency, pulse length, power transmitter, bandwidth, pulse repetition frequency (PRF) and the polarisation of transmitter [10]. The radar nodes of separated transmitters and/or receivers are connected to form a synchronised radar network thus the term netter radar or radar network- with defined communication bandwidth among them. The level of synchronisation is referred to as coherence/incoherence [9], and the data fusion method can be centralised or decentralised which classifies the complexity level as well as the operational performance of the radar system [9]. The separation layout and lengths of the baselines between the radar network nodes determines the form, function and performance of the system [10]. 36

37 2.9.1 Classification of Bistatic and Multistatic Radar Radar systems are classified based on the transmission and reception antenna sites, as well as the transmitter operational setting. Figure 2.6 illustrates the classification titles of radar systems along with their hierarchy. The first classification level based on transmitter and receiver location, in the case of the two antennas are co-sited, then the radar system will be classified as Monostatic. On the other hand, radars with separated antennas are classified as Bistatic and Multistatic radar. Furthermore, the second level of hierarchy is based on the transmitter operational setting. The cooperative transmitter radar considers the transmitter as part of the system configuration. However, Non-Cooperative transmitter radar exploits transmitter of opportunity which is considered as third-party component. Finally, within the third level of classification, the non-cooperative radar system could be divided into hitchhiker radar which utilizes other Monostatic radar signal, and the rest of non-cooperative radar systems that exploit third-party transmitters such as TV broadcast, GSM and Radio station [7][10]. Figure 2.6: Classification of Bistatic and Multistatic radar systems [7]. 37

38 2.9.2 Complexity Level Cases Table 2.1 by C. Backer [10][7] showcase the six cases of complexity of multistatic radar based on the number of system components, location of radar nodes, data fusion level and coherency model. The colour code show the level of complexity in green is fairly simple, amber is challenging and red is complex [10]. Case 1: This case signifies multistatic radar systems of a fixed location network of Monostatic radars as transmitters and/or receivers nodes. Each node is capable to process data and generates tracks. At this complexity level, this multistatic system is decentralised and requires little data communication between nodes. This implies simplicity in design. Case 2: Similar to the above case, the multistatic radar system is decentralised and the level of complexity is low. However, this case uses multiple pairs of bistatic radars, instead of using Monostatic radars. Case 3: This case represents the scenario of one transmitter and many receivers Multistatic radar system. Each pair communicates detection data to central processing node. Hence the system called semi-centralised processing. At this level, the system is considered more complex than the above cases due to increase in communication bandwidth and processing requirement. Case 4 and 5: The two cases include multistatic radar system with high demand for coherence and accurate synchronisation between nodes. Thus, such systems are highly centralised processing with demanding communication level. Case 6: In this case, multistatic system sensors are located at moving platforms. Wider communication bandwidth is required due to adding the station location. Additionally, the synchronisation complexity is increased. 38

39 Table 2.1 Complexity levels of multistatic radar systems [10]. Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Location Fixed Fixed Fixed Fixed Fixed and moving platforms Nodes on moving platforms Data Level Tracks Tracks Detections Detection s Raw Raw Coherence Incoherent Incoherent Incoherent Coherent Coherent Coherent Operational N Tx, N Rx 1 Tx, N Rx 1 Tx, N Rx 1 Tx, N Rx 1 Tx, N Rx N Tx, N Rx mode Monostatic multistatic multistatic multistatic multistatic multistatic Distribution Decentralis ed Decentralise d Semi Decentralise d Centralise d Centralaise d Centralise d Assessment Straight forward Multiple bistatic challenging Complex Very complex Extremely complex Multistatic Radar System Degree of Spatial Coherence Multistatic radar systems could be classified based in the degree of spatial coherence between the bistatic pair [7]. The term spatial coherence is the correlation between the radiated fields at points in a given spatial region [9]. After all, a multistatic system with many pairs of transmitters and receivers could have different level of spatial coherence among each pair. That is to say, there are three degrees of spatial coherence that will define the data fusion approach of the system: Fully-known spatial coherence: From the name, the knowledge of phase and frequency in addition to the location of the antenna phase centre are fully known and preserved throughout the transmission across the whole system. That is why; any 39

40 received signal could be fully referenced across the system. At this level of coherence, the communication level is demanding as well as higher integration is required. Short-term spatial coherence: This coherence type is similar to the one above, in a since that frequency and phase are fully known and maintained, along with phase location. But it is based on short periods. During these periods the measurements are made across the system. Multistatic radar systems with such short-term coherence require high stability in frequency between sub bistatic pairs. Incoherence: at this degree the system cannot utilize phase information and its changes over time, which imposes higher level of complexity in system operation. Consequently, the cost of incoherence multistatic system is higher than coherence systems Multistatic Radar System Data Fusion Multistatic radar systems are categorized based on data processing method, where it could be centralized or decentralized data processing. The centralised data fusion model gathers all data into a central fusion location where all information is processed. This method depends on high level of communication between network nodes. Thereby, this model provides better sensitivity, which leads for more accuracy. On the other hand, the data grouping process as decentralised data fusion method, utilizes the system nodes to initialise tracker and only part of the information will be sent to the central data fusion unit, where the final decisions are made. The decentralized approach provides the advantage system robustness, as the system is less dependent on the central processing unit. Hence, it would not be total system failure in case of partial system failure [9][10]. 40

41 2.10 Bistatic and Multistatic Advantages Bistatic Advantages Advantages of Bistatic radar system can be divided into two categories, namely as advantages from the design aspect and advantages from the operational aspect [5] Design advantages Those advantages are mainly influenced by the geometry of bistatic radar; as the transmitter and receiver are sited at different locations. That is, transmitter and receivers can be optimized differently. Thereby, bistatic radar technique provides design gains over the Monostatic radar, particularly in flexibility, reliability and Doppler components enhancement [5]. However these advantages come with the price of higher system complexity and processing. Where that will be further explore in the disadvantages section [4]. As transmitter and receiver are separated it is not necessary to coexist on the same platform. Therefore, platforms with less power and processing capability can operate the receiving antenna only, while for instance utilising transmitter of opportunity. In addition, bistatic radars do not employ the considerably heavy Duplexer component [11]. This advantage made the bistatic systems as an attractive radar system for the unmanned air vehicles (UAVs) [7], to keep the UAV light weight. Location wise, in some cases it is prohibited to deploy transmission antenna nearby flammable containers. This is a clear limitation of the Monostatic radar. Instead, the bistatic radar approach provides the flexibility of locating the receiver site at that location and the transmitter could be deployed away from the flammable container [11]. The separation of the receiver makes it difficult to detect and safe from attacks by anti-radiation missiles or highly directional jamming. Therefore, bistatic radar systems provide an effective solution against electronic warfare threats [11]. 41

42 The bistatic Doppler component is measured along the tangent to confocal hyperbola between the transmitter and receiver. This provides a new Doppler component dimension, which could be combined with the Monostatic radar Doppler components, toward both the transmitter and receiver, to give a two-dimensional velocity vectors [11] Operational Advantages Operational advantages of bistatic radar systems are based on the system power usage, RCS forward-scattering and flexibility in utilizing existing transmitters. The flexibility of bistatic setting can provide the advantage of using less transmitting power than Monostatic radar to detect a target at the same range [17], hence the spatially separated transmitter and receiver will provide wider area coverage of SNR level. In addition, it is feasible to operate the bistatic radar on high pulse repetition frequency (PRF) [11]. The bistatic radar can be operated on jamming sources as well as utilizing transmitter of opportunity, which present the advantage of making use of incident signals that exists within their operating band width [5]. This aspect can be regarded as supplementary roles for ESM (electronic support measures) [4]. Bistatic systems are susceptible to external jamming due to the fact that the transmitter and receiver are separated. However, the receiver performance can be affected when Omni-directional jamming is used with an increased power source [7]. The forward scatter and back scatter RCS from a target located on the baseline of the bistatic radar is much greater than the normal Monostatic RCS. Where the forward scatter RCS increases when frequency increases, while the back scatter RCS remains basically constant [3]. 42

43 Multistatic Radar Advantages Multistatic radar systems advantages were featured with all the above bistatic operational and design advantages. However, extra gains are particular for multistatic radar systems because of the additional system components (such as multiple transmitter and/or receivers), beside the data fusion and communication among those components (network nodes). The synchronisation concept between the multistatic radar system nodes can be implemented for old air traffic radars to operate as multistatic radar system, which offers the opportunity of extending the operational life of conventional radars [5]. Multistatic systems are reliable and residual radar systems. That is, the separation between the transmitters and receivers provides reliability advantage as if a node experience malfunction, the other receivers remain in operation and the overall system is not jeopardised [5]. Multistatic radar systems could provide better coverage, based on the following three factors. o The geometry is widely spread throughout the surveillance area with overlapping cell style coverage [9]. Also, multiple receivers at separated sites capture more reflected energy from the target. That is way, small transmitted power would still be sufficient to detect targets [12]. o The flexibility in sites setup could provide the opportunity to reduce shadowing effects [9]. For instance, Multistatic radar approach could be adapted as gap filler at a wind farm site to provide extra coverage from surrounding sites [7]. o Finally, more information is available to combine (data fusion process) from the multiple transmitter-receiver pairs. That will enhance detection accuracy, target classification and imaging [9, 12]. Multistatic radar of multiple receivers can increase the probability of detection as an overall system. That is, multiple receivers at different sites will not be all affected by jamming or RCS fading at the same time [11]. 43

44 Using high pulse repetition PRF is feasible for multistatic radar [11]. Given that the multistatic radar do not suffer from range blindness [12], as the transmitters and/or receivers are remain switched on all the time. Multistatic radar present high resolution by using multiple spatially diverse downrange profiles. As shown in the graph below figure 2.7. The two targets are detected despite the fact that both of them are at the same range from the transmitter and within the main beam width. Using separated receiver (at node 2) provides the capability to distinguish between those two targets, because they are positioned at different range from the receiver point of view. Figure 2.7: Bistatic Accuracy illustration [9] Advantages Passive [4] In case of passive bistatic radar, the system would require lower maintenance and procurement. No need for frequency clearance when using transmitter of opportunity approach. Monostatic radar suffers from the congestion of electromagnetic spectrum. Contradictory, passive radar systems see that as advantage. Passive bistatic radar systems can utilise VHF or UHF frequencies. 44

45 2.11 Disadvantages of Bistatic Radar Systems The geometry of bistatic radar for target location and the synchronisation between the transmitting and receiving sites are more complicated [11]. Some bistatic radar systems require expensive multi-beam receivers [11] for higher radar accuracy. The cost of allocating two sites for bistatic radar could be higher than one site for Monostatic radar [11]. For target detection, both transmitter and receiver should be able to see that target. In the scenario of low elevated target, a terrain can cause failure to see that target by one of the antennas [11]. In bistatic radar the sidelobes of one antenna can illuminate the main beam of the other. This will cause clutter path of that main beam [11]. Passive bistatic radar could be more complex, because of the lack of control over the transmitter which would make it harder to deploy [8]. Synchronisation: bistatic receiver need to know the transmitter waveform timing and frequency, in addition to antenna orientation; so that an accurate calculation of time different and Doppler Shift can be achieved. Therefore, the system complexity and demand for computing power increases. This could be seen as a disadvantage, which depends on the development requirement [4, 11]. 45

46 2.12 Chapter Summary This chapter provided background knowledge of radar technologies that are relevant to this research study. The chapter highlights the difference in principle between the bistatic and monostatic radar systems. In addition, all categories and classifications of bistatic and Multistatic radar system were explained. The bistatic radar systems operate on the basis of physical separation between transmitter and receiver [2], unlike the conventional monostatic radar systems where the transmitter and receiver are located at the same site. Bistatic systems are further subdivided into ABR or PBR [3]. In ABR the transmitter is integrated with and under the control of the radar system, Hence location, modulation, frequency and transmit phase of the transmitter are all pre-set and known by the system. On the other hand, the PBR utilises a transmitter of opportunity (that is one that already exists) such as a mobile communications signal or FM radio transmitter [3]. Multistatic radar system configuration includes multiple pairs of bistatic radars and/or monostatic radar system. The separated antennas are mapped to form a synchronised radar network with defined communication bandwidth among them. The level of synchronisation is referred to as coherence/incoherence [9], and the data fusion method can be centralised or decentralised. The coherence level and data fusion approach defines the classification and the complexity level of the radar system, as well as the operational performance [9]. The collaboration of such network of radar system provides the design and operational advantages that were listed in section The early development of CW radar system separates between transmitter and receiver antennas, to prevent over saturating the receiver by the transmitted signal. Such radar setting in concept signifies the bistatic radar design. After the development of radar antennas with duplexer component capabilities (switching between transmission and reception), Monostatic radar system became more and more popular. Meanwhile, Bistatic radar system went through cycles of interest approximately every 15 years [6]. H. Griffiths portrayed those cycles of interest into three resurgences. The first interest cycle was driven by attentiveness for trade-off studies between bistatic and Monostatic radar systems, based on criteria of cost, maintenance or performance. The second wave 46

47 of interest was influenced by new target properties such as the anti-radiation missiles [5]. Finally it has been stated that the research interest in bistatic radar system is passing through the third resurgence, where the technology is at the peak of interest cycle [6] due to the advancement in technologies. For instant, the advancement in computing power provided the support for the complex process of synchronisation and communication among the multiple sensors, and the GPS technology made the geolocation difficulties easier to manage References 1. M, SKOLNIK, Radar Handbook, second edition, 1990, Boston, Mass: McGraw-Hill. 2. N. Willis, Bistatic Radar, 1991, Artech House. 3. M. SKOLNIK, An Analysis of Bistatic Radar, March 1961, IRE Transaction on Aerospace and Navigation Electronic. 4. H. Griffiths, C. Baker, Passive coherent location radar systems. Part 1: Performance prediction, April 2005, IEE Proceedings online no M. Dunsmore, Chapter 11: Bistatic radar, Advance Radar Techniques and Systems, 1993, Peter Peregrinus Ltd. ISBN x. 6. C. Baker, H. Griffiths, Bistatic and Multistatic Radar Sensors for Homeland Security, 2006, Advances in Sensing with Security Applications, H. Griffiths, Multistatic MIMO and Networked Radar: the Future of Radar Sensors?, September 2010, Proceedings of the 7th European Radar Conference, B. Chan, Receiver Site Optimisation for Passive Coherent Location (PCL) Radar System, October 2008, University of Cape Town. 9. S. Doughty, Development and Performance valuation of a Multistatic Radar System, 2008, University College London. 10. C. Baker. An Introduction to Multistatic Radar, College of Engineering and Computer Science, ANU, RTO-EN-SET S. Kingsely, S. Quegan, Understanding radar Systems, 1992, MsGRAW-HILL, ISBN Ching-Wei (W. Chang), System Level Investigations of Television Based Bistatic Radar, December 2005, University of Cape Town. 47

48 Chapter 3 Multistatic Radar Simulation Tool and Wind Turbine RCS Modelling 3.1 Introduction A computer simulation tool was developed using Java programming language due to its object-oriented features and for its numerous advantages such as, portability, multithread, robustness and architecturally neutral [1]. This is in contrast to MATLAB, where object-oriented programming is supported however it is less robust and stable than in java [1]. The simulation tool employs the polymorphism programming concept that is well supported by Java, whereas in MATLAB is less practical. Java memory allocation is done automatically, it is unnecessary to be concerned about disposing objects and reallocating memory. Ultimately, the adapted RCS modelling tool of wind turbine structure was developed in Java as well, so the two computer tools shares the same programming platform and it is easy to integrate. The aim of this chapter is to introduce the main components of the simulation tool and explain their algorithmic steps and logic. The first section discusses the antenna component from the aspect of its flexibilities and adaptabilities for changes. It will then go on to the second section which talks about the geometry component as a central component and further explains the single geometry plane approach that simplifies the bistatic geometry implementation as well as saving on processing time. The third section discusses the power received component then followed by the section of propagation 48

49 factor component. The final section will introduce the adapted RCS modelling tool for the wind farm structure, in addition to that model assumptions and verification. The Multistatic Radar Simulation implements the equations from the Computer-Aided Radar Performance Evaluation Tool (CARPET V 2.0) [2] equations for the implementation of the above concepts (such as power received, propagation, reflection coefficient and radar geometry), see Appendix 1. All simulation tool results of these radar parameters were verified against CARPET Simulation Tool Assumption and Operation Environment Results and plots illustrated within this chapter are produced by using the developed simulation tool. The following antenna configuration and operational environment setting are assumed to produce this chapter results and plots. These parameters are used for the radar equations in Appendix 1 Target RCS = 100 m² Transmitter antenna setting: Device gain = 5 db Antenna type: Uniform Diameter = 4 m Default bearing angel = 90 Power transmitted = w Wavelength (λ) = 3 cm Receiver antenna setting: Device gain = 5 db Antenna type: Uniform Diameter = 4 m Default bearing angel = 90 Operational environment variables: Land Land or Water Temperature = 15 c Air temperature = k Pressure =

50 Humidity = 0.7 Water in Soil ratio = 60% Ground Type = 1.0 Hsd land = 0.1 Sea state = 3 Salinity = 35. Earth effective radius (Re) = x (4.0/3.0) m 3.2 Simulation Tool Antenna Component The simulation tool provides the feature of design and configuration of multistatic radar system. Therefore, the simulation tool should have the flexibility in adapting different types of sensors of different configurations. This was considered when developing the simulation component of antenna class. Figure 3.1 below illustrates the inheritance relationship between the generic level class as Antenna and the device level class as Transmitter or Receiver. The implementation was based on the Polymorphism technique in object-oriented programming. The Antenna class was set to include all generic variables that a radar antenna can have, and the device level class (i.e. Transmitter or Receiver) will inherit all antenna properties and variables from the Antenna class. Besides, the device level class can have different variables setting. Genaric level Antenna: Position (X, Y, Z) Devive Gain. Gain Variable. Type. Diameter. Antenna Tilt. Bearing Angle. Device level Transmitter Receiver Figure 3.1: Simulation tool Antenna Component 50

51 To showcase this flexibility in design and configuration, the simulation tool can design a scenario of multistatic radar that utilizes multiple Monostatic radars. Parts of these radars are set to act as transmitters and others as receivers. In this regard, with the current simulation, It would be feasible to change the functionality of one node from transmission to reception by simply redefining that sensor as Receiver object. Implementing this type of class hierarchy provides the simulation with the capability to adapt to a new antenna types by simply adding new classes in the device level side. After all, the simulation needs to be capable of future integration for new devices. 3.3 Simulation Tool Bistatic Geometry Component Bistatic radar geometry has numerous range variables and various signal angles. Consequently, the development of the simulation tool had to dedicate a geometry component to calculate those variables and associate them specifically to each bistatic pair. This component is triggered by power received, propagation factors as well as target positioning components. Therefore, it should be optimised and flexible for necessary modifications, to suit different multistatic radar design layout. Figure 3.2 represent a simplified drawing for bistatic radar pair geometry with one target above the baseline. Noticeably, when dividing the geometry into two planes (as plane1 and plane2) at the target site, each plane can be accounted for as monostatic radar geometry. The calculation of range variables and angles for both planes can be performed by the same set of equations. Figure 3.2: Bistatic geometry, Plane1 and Plane2 identification 51

52 The two planes approach minimizes code duplication and optimises the calculation process by dedicating geometry component for a generic single plane calculation. For example, when simulating multistatic radar of one transmitter and N receivers, the geometry component calculates the angles and variables of Plane1 only onetime and then calculate plane2 variables N number of times for each receiver, as illustrated in figure 3.3. In this case, by treating the geometry as two separate planes, the computation process was cut in half. Figure 3.3: Optimising multistatic geometry illustration. 3.4 Simulation Tool Power Received The power received equation was adapted from CARPET V 2.0. However, CARPET equations are meant to be for Monostatic radar; therefore it had to be modified to suit the bistatic calculation. Overall, the bistatic power received equation can be driven from the Monostatic radar equation as it was explained in chapter 2. Bistatic Power Receiver Equation (from chapter 2, Equation 2.6): P r = P t G t G r F t 2 F r 2 σ b λ 2 (4π) 3 R t 2 R r 2 (3.1) The wavelength (λ), transmitted power (Pt) and the bistatic RCS (σ) are prefixed values. However, the transmitter and the receiver gain (G t, G r ) accordingly are a combination of the peak in antenna gain of the azimuth gain and the elevation gain radiation pattern. Thus, the last two gain values are generated from the antenna azimuth and elevation angles. Furthermore, the transmitter-to-target and target-to-receiver (Rt, Rr) range variables are calculated by the geometry component. Lastly, the two propagation factors 52

53 are generated by the simulation propagation component which will be explained in the next section. Figure 3.4 illustrates the interaction among the simulation tool components during the calculation process of power received. The Run-Routine component acts as a middleware layer between the geometry, propagation and power received components, which ultimately interfaces with the simulation I/O (input/output). The power received calculation steps starts with initiating antenna location and the preset target location. Then, the run-routine populates the two geometry planes objects along with their calculated variables, and sends those planes objects to the propagation factor component to calculate the signal propagation factor for each plane accordingly. Ultimately, the range (Rt, Rr) and antenna gain (azimuth/elevation) variables are sent to Power received component which dose the final calculation for power received at that target location. Figure 3.4: Flowchart of the simulation tool power received calculation. 53

54 3.4.1 Power Received Simulation Tools Result Figure 3.5 shows the power received in (dbw) based on target range in (Km). For this simulation run, the transmitter and receiver were deliberately located at the same site to emulate the behaviour of Monostatic radar; thus results can be compared to CARPET Manual results. In order to produce the plot in figure 3.5, a moving target was set to start from the antennas location and continue to propagate away from the antennas by incrementing the longitude (X) while latitude (Y), and elevation (Z) stays constant as shown in figure 3.6. In figure 3.5, the red curve shows the values of power received only, whereas the blue curve represents values of power received with propagation factors effect. The fluctuation in power values of the blue line is due to the phase difference between the direct signal and multipath signal (either adds to the direct signal phase or cancels when the phase is opposite). The blue curve (power + propagation) witnesses a dramatic decline in power values after the 50km range. This decline in power received values is due to the fact that the target range starts to exceed the horizon range; enters the intermediate region and continues towards the diffraction region of the propagation factor. The propagation regions will be discussed in the next section. 54

55 Figure 3.5: Power Received with and without Propagation Factor (db), Target Range (km) Figure 3.6: Target movement illustration for power received plot 55

56 3.5 Bistatic Propagation Factor The bistatic propagation factor is one of the external factors that affect the radar operation. As for radar microwave signal does not travel in a direct pre-set path. Instead, the signal could be reflected of other surfaces either on the way towards the target or on the way back towards the receiver. This behaviour is referred to as signal multipath. The difference in distance between direct and indirect signal can alter the signal phase at the receiving site, which can increase or decrease the power received value. For instance, when two phases are opposite, the two signals will cancel each other causing the case of detection null (blind spot). Signal propagation factor affects Monostatic and bistatic radars, yet the effect of bistatic signal propagation can occur in the first bistatic plane (from Transmitter-to-Target) and/or in the second bistatic plane (From Target-to-Receiver). As a result, there are two separate bistatic propagation factors of different values. However, the method for calculating those factors has no significant difference to the Monostatic propagation factor method. The of bistatic signal propagation factor for curved earth was implemented for the simulation tool, which included the three propagation regions of optical-interference region, intermediate region and diffraction region. As such, each region required different approach for calculating the propagation factor as a function of range. The mathematical equations for calculating the propagation factor based on a given range region are listed in (Appendix 1, section A3) The optical interference region begins from the antenna location and it ends at roughly about the radio horizon. Mathematically, the maximum range of the optical interference region is defined by the larger range value of two measurements Rs or R1 4 λ, where Rs represents the range value where the spherical earth divergence factor (D) becomes invalid [2], hence the divergence factor depends on the grazing angle where (D = 0) when grazing angel (ψz <= 0). The second range occurs when the difference between the direct and the reflected signals is equal to one-quarter the wavelength. 56

57 The diffraction region starts beyond the intermediate region where the target range is equal or greater than diffraction range (Rd). Figure 3.7 explains the logic flow for the propagation factor component of the simulation tool. The main objective was to optimise the propagation routine by utilizing the propagation range regional conditions. Figure 3.7: Simulation tool algorithm of propagation factor regions As mentioned before, despite the fact of having two propagation factors for each of the two bistatic planes (plane1 and plane2), each plane represents similar geometry that of Monostatic radar (appendix 1, figure A1) does, see figure 3.8. The propagation factor calculation starts with finding the target ground range (R) to decide on which propagation region to calculate. Then it continues to calculate the ground range from antenna to the rays reflection point against the earth (R1), in addition to the ground range from reflection point to the target ground range (R2). To simplify the 57

58 rest of the calculations, a tangent plane is defined (to work with straight lines and avoid ground curve). Therefore, the new slant range is calculated to find the difference between the direct signal and reflected signal. And finally, the range difference is substituted into the propagation equation. These steps are summarised and can be varied based on the region. Figure 3.8: The geometry of Bistatic propagation factor Simulation Tool Bistatic Propagation Factor The result of propagation factor shown in Figure 3.9 is based on the same simulation run that was illustrated in figure 3.6. In this case the two propagation factors are identical for both geometry planes (transmitter-target and target-receiver) knowing that the target is moving at constant elevation and the transmitter and receiver are at the same (x, y, z) coordination. The optical interference region starts from the antenna position at range 0km and extends up to the (Rs) range value of 50km, which is marked by the green line. Beyond this range, the propagation factor value starts to drop dramatically within the intermediate region between the green and yellow lines. The reason of this decline in values is that the target range crosses the radio horizon range whereas the grazing angle of the propagated signal devolves to zero. Finally, the diffraction region starts at the (Rd) range, marked with yellow line, where the propagation factor continues the dramatic decline in value. 58

59 Figure 3.9: Propagation Factor (db) across Target Range (km) Within the optical interference region, the propagated signal fluctuates in value. That is, at 30Km range the propagated signal reaches to 2.4dB up, then it goes down to -7.5dB around the 32km range. This fluctuation in value can eventually alter the phase and value of power received. In fact, the two plots in figures 3.6 and 3.9 can be closely examined to illustrate the propagation factor effects on power received. For this reason, figure 3.10 below was compiled from a snap cut from the late figures. The two arrows below at 30km range indicate the correlation between the increase in propagation factor and power received. Overall, this explains the value fluctuation in power received when incorporating the propagation factors. Figure 3.10: The correlation between propagation and power received 59

60 The previous propagation result was based on setting the transmitter and receiver at the same location and equal elevations. However, when the elevation of the transmitter is set to be different than the receiver, the red propagation curve associated with the transmitter to target propagation plane will shift to the left as the antenna elevation increases or shift to the right as the elevation decreases. For example, Figure 3.11 shows the shift to the left in the transmitter to target curve (red curve) when changing the transmitter elevation from 65m to 60m. Figure 3.11: Propagation Factor of two antennas at different elevations 60

61 3.6 Bistatic Signal-to-Noise (SNR) Simulation tool The plot in figure 3.12 shows the colour-scale values of SNR for a bistatic pair of omnidirectional transmitter and receiver. The SNR is calculated by measuring the power received and dividing onto the system temperature noise. In which, the process to produce this plot was based on scanning a target across the longitude and latitude of 12x 12 km area with accurate increment of 20m for each calculation point. For every increment, the power received was calculated. This scenario implies that the calculation is repeated 360,000 times and so does the simulation produces the result in few seconds. The colour-scale plot is developed as a part of the simulation graphical interface, which is considered as a handy tool to simulate power return based on area of illumination. The main observation from the plot below is the Oval of Cassini shape around the baseline. This result validates the simulation power received component, where the generated Oval of Cassini plot matches that of the bistatic SNR theory (see chapter 2, section 2.6). Figure 3.12: SNR for Bistatic Radar 61

62 Additional pairs of bistatic radar can be added to the plot as shown in figure 3.13.a and b. The SNR plot for multistatic radar takes the merging shape of each bistatic pair Oval of Cassini. The plot illustrates that SNR tends to be higher within the close range from the baseline, thus, a multiple receivers located around the transmitter increase the illuminated area of higher SNR. Figure 3.13: a) One Transmitter, Two Receivers SNR. b) One Transmitter, Three Receivers SNR. 62

63 3.7 Directed Antenna Simulation This section demonstrates the simulation integration of directional antenna configuration. Besides, results from SNR plots for directed multiple receivers. For multistatic radar system, the antenna direction can be configured individually by setting the bearing angle variable. For example, Figure 3.14.a shows the SNR plot of omnidirectional transmitter and one receiver with (90 ) bearing angle where (30 db) threshold is set for the purpose of plot clarity. In the other hand, in Figure 3.14.b, the transmitter was moved to the other end of the plot, this created wider coverage by the Mainbeam of the receiver and higher SNR around the far end transmitter area. Figure 3.14: a) Receiver bearing angle = 90 at the transmitter site. b) Bistatic pair of Receiver bearing angle =

64 The figure 3.15 demonstrates the simulation flexibility of incorporating new receiver and manipulating the bearing angle to provide wider area of elimination which is clearly illustrated in the figure below, as a wider coverage can be provided by additional receiver beam directed towards the area of interest. Figure 3.15: Two directional receivers (90 ) & one omnidirectional transmitter The directional antenna configuration can be set for the transmitter antenna type as shown in figure This feature provides the flexibility to set-up each bistatic pair bearing angles to concentrate their main beams for illuminating specific area and gaining high SNR. This flexibility in design shows the advantage of multistatic radar system, where coverage area can be divided into sub areas that are managed by subgroups of multiple bistatic pair of directional configured. Figure 3.16: Bistatic pair of directional antennas 64

65 Figure 3.17 portraits similar result to the previous plot, yet the transmitter and receiver bearing angles were set to look horizontally across the plot, that is, to show the extent of the overlapped main beams across the range line, with high SNR coverage. Figure 3.17: SNR of bistatic Directional antennas, Horizontal view. 3.8 Probability of Detection Probability of detection Pd is defined by the likelihood of the scattered pulse from the target to exceed the threshold level of the noise received by the radar system. Therefore, Pd correlates with the SNR positively. However, probability of false alarm Pfa represents the probability of a false return pulse that exceeds the noise threshold level. Thus, a lower level of threshold of SNR will increase the probability of false detection [3]. Based on the tolerated level of false alarm for the system as Pfa and the SNR figure, the probability of detection can be accurately approximated from equation 3.2 [11]. Where (3.2) (3.3) 65

66 That is, Table 3.1 shows a list of SNR values in correspond to the Pfa, for each Pd level. Those values were generated by using the above equation 3.2. For instance, when allowing 1 in 1000 detections as false alarm; then the minimum threshold of SNR is going to be as low as 4 dbi, and the probability of detection is going to be 0.1. Table 3.1: Single pulse SNR (db) in correlation to Pd and Pfa [3]. The simulation tool approximates the probability of detection figure for every bistatic radar pair individually, and then it combines all approximations into one overall approximation figure for the multistatic radar system. The calculation was based on the lookup table 3.1 in the manner of establishing the false alarm figure for that bistatic pair first, then calculating the SNR at target positing, and finally looking up the probability of detection values that correspond to that SNR. A cubic spline function was used to estimate the value of Pd when the SNR was not the exact value to those in table 3.1. The individuality of calculating the probability of detection for each bistatic pair provides the advantage of tolerating a higher figure of false alarm per bistatic link; because the overall figure of false alarm for the multistatic system is going to be the multiple probabilities of those individual false alarm figures. Consequently, the SNR threshold is going to be low for each link and altimetry higher probability of detection. For instance, figures 3.18 and 3.19 show results of Pd in form of a colour scale plot. The results were based on simulating one transmitter and four receivers multistatic radar 66

67 system. Probability of false alarm for each bistatic pair was set to (P fa = 10 3 ). Accordingly, results from each bistatic Pd is shown in figure 3.18 (a, b, c and d) separately, and then the final combine result is plotted in figure Figure 3.18: Individual Bistatic radar Pd plots Figure 3.19: The combined Multistatic radar Pd plots. 67

68 To increase the probability of detection figure in multistatic radar system, the simulation tool combines the top three individual probabilities of detections only in order to approximate the overall system Pd. Knowing that, only three pairs of bistatic radar is required to detect a target. In this way, less probability figures to multiply and the end result will be improved. Figure 3.20 shows the improvement in probability of detection figures based on area of coverage, between the new approach of top three bistatic pairs Pd and the old approach of multiplying all bistatic Pd. Figure 3.20: Pd comparison (Top three (Pd) approach Vs Old approach) 68

69 3.9 Wind Turbine RCS Modelling This research adapts the Rectangular Meshing for RCS Approximation (ReMeRA) modelling methodology of the wind farm RCS modelling tool that was developed at the University of Manchester [4]. The adapted modelling tool devised a methodology that gives an accurate geometrical representation of the wind turbine for RCS and Doppler modelling. This accuracy in the representation of wind turbine geometry is important when modelling the RCS of a turbine structure, because simplifying the geometry can significantly alter the RCS at radar frequencies [5]. Therefore, this model implements the standardised coordinate system and angels, as shown in figure 3.21, that are commonly used by the wind farms industries, to accurately define the turbine orientation [22]. Figure 3.21: General representation for wind turbine Dimension and Angles [4] The RCS modelling is achieved though segmenting the turbine into multiple smaller targets, each with its own RCS and Doppler signature, as it is illustrated in figure The modelling tool was designed for Monostatic radar operation, however this tool can be adjusted to operate based on multistatic radar principles. The separation between the radar sensors at different locations can be addressed by modifying the configuration parameters of the RCS modelling tool. The transmission point and reception point were setup at different locations, instead of collocated sensors points for the Monostatic radar scenario. 69

70 Figure 3.22: Wind turbine segments for RCS modelling. There are three main components of a wind turbine based on their functionality, the tower, blades and the nacelle. The tower is the largest part of the turbine structure that is shaped as cylinder and made of rolled steals and concrete for foundation. This makes the tower the largest reflecting object of the turbine. Where RCS modelling tools showed 85% of the radar return can be of the tower alone. The RCS of some turbines tower can reach up to (10 6 m 2 ) which is much larger than a typical RCS of a vessel ( m 2 ). RCS modelling for the blades components proved to be more challenging process. The blades have more complex geometry, and tend to be unpredictable when it comes to modelling its orientation and title. The complex aerodynamic profile of the blade makes the RCS modelling highly variable with time depending on the incidence angle of the radar single. It is important to accurately represent the blade geometry for better RCS modelling, because small change in the profile at a radar frequency will result into major change in the RCS profile of the blade. In comparison to the tower return, the nacelle component has much smaller return due to its smaller size. The radar return from the nacelle can be significant when the direct incident of radar signal front illuminate the flat plates from the side of the nacelle. This can increase the RCS. Nacelle with round design tends to have higher RCS profile then the rectangular design [15]. 70

71 3.9.1 Wind Turbine RCS Modelling Tool Assumptions The simulation of large number of turbines is a demanding computing process. To reduce the computing time of simulating large wind farm installation, this model established a set of assumptions that does not significantly affect the RCS modelling accuracy. To reduce the complexity of the wind farm, the model assumes all turbines within a wind farm are of the same type and size, and the signal multiple reflections of a turbine components was not considered. Additionally, the wind direction was set to be constant across the wind farm formation, so all turbines have the same orientation facing the wind direction. Blades start rotation angel is randomly set between (0-120 ) for each turbine individually. This modelling tool calculates the RCS of a turbine individually. That was based on the assumption of a turbine s RCS calculation is not affected by the surrounding turbines within the wind farm formation, because the power summation at the receiver is non-coherent [4]. The separation distance between turbines is far enough for this assumption to be valid. The separation distance is based on 3-4 rotation diameters in the crosswind direction and 5-10 rotation diameters in the downwind direction, with respect to the prevailing wind direction [4]. Turbine components can shadow each other for short period. This is referred to as shelfshadowing. The adapted RCS model dose not account for self-shadowing because it showed no significant impact on the RCS modelling based on its short period and small shadowing area. The self-shadowing is referred to by RCS chopping as well [9]. The only geometry assumption and simplification was for the nacelle component. The approximation of the geometry was based on general descriptions from the manufacturer and from proportion diminutions in respect to the tower size and blades. Wind loading and solar heating affects the title and bend of the turbine tower and blades. The tower component can bend up to 0.4m at the top sections under maximum wind load [4]. Also the differential expansion of the tower due to the solar heating of one side can cause similar effect on the tower [4]. This bending change in tower posture is not accounted for. It might effect on the RCS modelling when signal reflects of the bended part away from radar sensors. 71

72 For the blades component, the RCS profile of a blade can be altered according to the bend effect that is cause by the wind loading [4]. This impact on RCS profile depends on the angel of illumination, the stiffness of the blade and its typical bending behaviour. With these impact factors in mind, the current model validates the assumption of the blade to be straight and bend free. This research focuses on Range-Only detection and no Doppler processing is required. Therefore, the Doppler signature cause by the blades is a subject for further development The ReMeRA Methodology The RCS of a turbine can be very large based on the large size of the turbine structure and depending on the turbine orientation with respect to the radar. The RCS modelling for this type of structure is a complex parameter and depends on incident wave, frequency, material, illumination pattern and distance from the source [6]. Additionally, the complexity in wind turbine geometry and the extension over hundreds or thousands of wavelengths at radar frequencies proved to be a challenging task [4]. Accordingly, the RCS modelling process of a wind farm is a computationally demanding process, with long run time in order to predict the scattered signals to an acceptable accuracy for the large number of turbines of the farm. However, the RCS modelling tool defines a new methodology (ReMeRA) to model the returns from the wind turbines blades and tower in a computationally efficient manner that can be used on standard desktop computers. The model accounts for the farfield and nearfield RCS modelling [4]. This aspect is very important when it comes to multistatic radar RCS modeling of a wind farm. The radar system includes multiple receivers that are located at different sites. Some sensors can be at the farfield and others can be located within the nearfield. Therefore this modeling tool is adequately adjustable form multiple sensors scenario. As it has been mentioned above, the modelling tool divides the turbine tower and blades into segments of small sections along their length. A single blade is typically divided into number of sections between The number of sections for the tower is approximately twice the number of segments used in the blade. However segments number and size can vary for different applications [4]. The model computes the RCS for 72

73 each segment and considers it as a potential targets that creates its own radar signature. This radar return is modeled by accounting for the antenna pattern and other propagation factors. The simulation tool of multistatic radar system tests this modeling method according to the bistatic radar equation. The source of the radar signal (i.e. transmitter) was separated from the multiple receivers nearby a wind turbine. Figure 3.23 shows the color scale plot of SNR for each segment of a wind turbine. This result was produces from the simulation tool to illustrate the calculated SNR of each segment according to its RCS modeling. The plot in figure 3.23 does not represent the actual radar imaging of a turbine. The RCS modeling tool was integrated with the bistatic power received and geometry components of the simulation tool. The same methodology was used to model the RCS for each segment, but the geometry and power equation was different than the monostatic radar. It is clear that for multistatic radar system, the tower remains the largest source of radar return. On the other hand, the blades reflect much less signals. The illustration in figure 3.23 shows that each segment can be treated as small target, where the radar return can vary from one segment to another as such the far down blade have multiple degrees of SNR across the blade segments. Figure 3.23: SNR of multiple segments of a wind turbine structure. 73

74 3.9.3 Verification of RCS Modelling Tool The proposed ReMeRA methodology for RCS modelling was based on Physical Optics (PO) approximation and does not account for the diffracted fields [4]. This implies, when testing this methodology according to a prism shaped target, the predicted RCS will be lower than the actual measurements around the edge if the prism. However this issue is not of a significant within the context of wind turbine, as the total RCS of the turbine is dominated by surface scattering [21]. The ReMeRA model was extensively tested against measured RCS data of canonical shapes with flat and curved surfaces [4]. To further test the ReMeRA model ability to model turbine components, it was compared against FEKO [17]. The ReMeRA model shows an excellent correlation around the specular region in both cases. The model accurately predicts angles and the levels of the specular reflections. The levels of the peak values are within 0.1 db of the measured data. As the incidence angle moves away from the specular angle the correlation remains good although loses some accuracy when the scattering through diffracted waves from the edges becomes more apparent on the total RCS. The ReMeRA model s ability against modelling using the Method of Moments (MoM) was tested. Additionally, The ReMeRA was also compared against the RCS modelling results obtained from FEKO for a full wind turbine blade [4]. The good agreement with the measurements and results from other models gives confidence that the ReMeRA model can predict the RCS of turbine components to a good level of accuracy. 74

75 3.10 Chapter Summary The chapter objective was to showcase the developed software components of the multistatic radar simulation tool during the course of this research. In addition, the novelty of integrating the computer tool for wind turbine RCS modelling was discussed. The simulation tool implements the radar return equations from the Computer-Aided Radar Performance Evaluation Tool (CARPET V 2.0) [2], such as power received, propagation, reflection coefficient, SNR and radar geometry (see Appendix 1). The simulation tool supports the feature of design and configuration of multistatic radar system. The antenna component was developed to be adaptable to different types of sensors with different configurations. This provided the capability to adapt to a new antenna type by simply adding new classes in the device level side. The simulation tool supports the feature of color scale plot of signal to noise ratio (SNR) and probability of detection (Pd) for bistatic and multistatic radar configurations. This functionality can be utilised as a tool when designing the radar system. The plot shows the SNR according to antennas placements, antenna patterns, system noise, and power transmitted. Therefore, according to the SNR plot the radar design and configuration can be further modified to suit the radar operational objective by achieving the required SNR level within the area of interest. Similar radar design practice can be performed when examining the Pd color scalp pot for a multistatic radar system, so further design alteration can be performed to achieve the desire Pd level across the area of interest. This research adapts the Rectangular Meshing for RCS Approximation (ReMeRA) modelling methodology of the wind farm RCS modelling tool that was developed at the University of Manchester [4]. The modelling tool was designed for Monostatic radar operation, however this tool was adjusted to operate based on multistatic radar principle. The separation between the radar sensors at different locations was addressed by modifying the configuration parameters of the RCS modelling tool. The transmission point and reception point for that modelling tool were setup at different locations. The ReMeRA model was extensively tested against measured RCS data of canonical shapes with flat and curved surfaces [4]. In addition, the ReMeRA was compared against 75

76 FEKO [17] to test the RCS modelling of turbine structure. The ReMeRA model shows an excellent correlation around the specular region in both cases. The ReMeRA model s ability against modelling using the Method of Moments (MoM) was tested. Additionally, The ReMeRA was also compared against the RCS modelling results obtained from FEKO for a full wind turbine blade. The good agreement with the measurements and results from other models gives confidence that the ReMeRA model can predict the RCS of turbine components to a good level of accuracy References 1. Pradnya Choudhari, Java Advantages & Disadvantages, ArizonaCommunity.com. 2. Albert G. Huizing, Arne Theil, Computer-Aided Radar Performance Evaluation Tool (CARPET Version 2.0) User Manual, TNO Physics and Electronics Laboratory. 3. M, SKOLNIK, Radar Handbook, second edition, 1990, Boston, Mass: McGraw-Hill. 4. Laith S Rashid, Impact Modelling of Offshore Wind Farms on Marine Radars, The University of Manchester, Jago, N. Taylor, Wind Turbines and Aviation Interests - European Experience and Practice, ETSU W/14/00624/REP, DTI PUB URN No. 03/515, G. J. Poupart, Wind Farms Impact on Radar Aviation Interests, BWEA Radar Aviation Interests Report, DTI report number W/14/00614/00/REP, September F. Costen J. P. Brenger, A. K. Brown. Comparison of FDTD Hard Source With FDTD Soft Source and Accuracy Assessment in Debye Media IEEE Transactions on Antennas and Propagation, vol. 57, pp , July D. B. Davidson Computational Electromagnetics for RF and Microwave Engineering, Cambridge University Press, Butler, M.M. and D.A. Johnson, Feasibility of Mitigating the Effects of Wind farms on Primary Radar. 2003, Alenia Marconi Systems Limited. 10. Simpson, S.H.W., P. Galloway, and M. Harman, Applications of Epsilon A Radar Signature Prediction and Analysis Tool, in International Radar Symposium IRS : Munich, Germany. 11. D. P. Allen, et al., MITRE RCS Calculation Capability. 1994, MITRE Corp 76

77 12. Kent, B.M., et al., Dynamic Radar Cross Section and Radar Doppler Measurements of Commercial General Electric Windmill Power Turbines Part 1: Predicted and Measured Radar Signatures. Antennas and Propagation Magazine, IEEE, (2): p Knott, E.F.a.S., J. F. and Tuley, M. T., Radar Cross Section 2nd ed. 1993: Boston: Artech House 14. Jenn, D.C. and M.F. Chatzigeorgiadis, A MatLab physical Optics RCS Prediction Code, in IEEE Antennas and Propagation Magazine. August Youssef, N.N., Radar cross section of complex targets. Proceedings of the IEEE, (5): p Crispin, J.W., Jr. and A.L. Maffett, Radar cross-section estimation for complex shapes. Proceedings of the IEEE, (8): p Davidson, D.B., et al. "Recent progress on the antenna simulation program FEKO". in South African Symposium on Communications and Signal Processing South Africa. 18. Escot-Bocanegra, D., et al., New Benchmark Radar Targets for Scattering Analysis and Electromagnetic Software Validation. 2008, Instituto Nacional de T ecnica Aeroespacial: Madrid. p D. G. Falconer, Extrapolation of Near-Field RCS Measurements to the Far Zone, IEEE Transactions on Antennas and Propagation, vol. 36, pp , JUNE Supergen. Supergen Wind Energy Technologies Consortium (Supergen Wind) 2009 July 2009 [cited December]; Available from: J. Pinto, J. C. Matthews, and C. Sarno. Radar signature reduction of wind turbines through the application of stealth technology. Antennas and Propagation Conference EuCAP Pp March David Jenn, Cuong Ton, Wind Turbine Radar Cross Section, International Journal of Antennas and Propagation, Volume 2012, Article ID , 14 pages, December

78 Chapter 4 Multistatic Radar Target Detection and Tracking 4.1 Introduction Multiple targets detection and tracking by multiple sensors system such as the multistatic radar system is a challenging process with regard to measurements association and false detections elimination [1]. Therefore, a robust mathematical approach is required to predict the unknown system status from a list of input measurements with uncertainty of error margins. Range-Only target detection approach was adapted for the simulation tool detection component. That is, the positioning of a target depends on the bistatic range measurement from multiple bistatic pairs which eliminate the need for angle of arrival. This approach requires association technique of multiple measurements and elimination process of potential ghost targets detection. Target localisation method was implemented to calculate the intersection point among multiple associated range measurements. The method provided detection results but it had some performance limitation and disadvantages when detecting multiple targets. This chapter will discuss this method and consider its limitations. Additionally, alternative methods for target detection were investigated to overcome such limitations. Many research publications investigate multistatic radar target detection and tracking methods based on Probability Density Function (PDF) according to all available measurements. Such methods attempt to find the posterior of system status given the 78

79 prior measurement [1][2][3]. Such methods include Kalman filter [1], PF [3] or Probability Hypotheses Density (PHD) filter [2]. In these methods the system status is predicted based on the likelihood of sample measurements matching the true measurements. However, as an alternative mathematical technique the IA showed promising potential within the field of target detection and tracking [4]. This approach was first reported in the 1931 [5]. However, only recently there has been an increase in popularity for adapting IA computing model within various scientific fields due to the recent expansion on IA - credited to R. E. Moore [6] - and the development of computational packages [5]. Solving a problem with degree of uncertainty is adequately represented by interval vectors, which qualifies the IA implementation within the field of state estimation of a system. The real solution value is assured within the interval and bounded by margin of error, which is defined by the lower bound and upper bound of the interval. L. Jaulin [7] investigates the use of IA for multistatic radar system target tracker. The tracker uses Range-Only measurements to find target position. However, the solution was presented in the two dimensions 2D (i.e. Horizontal plane) of X and Y axis only of the Cartesian coordination. In addition, the scope of that implementation was to track small number of targets. However, this research investigates the detection solution on IA in the three dimensions 3D of X, Y and Z axis, and the objective is to detect and track large number of targets within complex detection environments such as wind farms. This chapter aims to introduce PF and IA detection methods along with presenting the simulation tool implementation of the two methods. Ultimately, comparison analysis of the two methods is undertaken while considering the target detection rate, false detection rate and target detection accuracy. 79

80 4.2 Multistatic Radar Range-Only Target Detection The bistatic radar estimation of a target location depends on the reflected radar return of that target in addition to the angle of arrival at the receiving site. Target range measurement is derived from the received signal according to the time difference between the transmission and reception of that signal [8]. This range measurement signifies the total travel distance of the received signal from transmitter to target Rt and from target to receiver Rr (See Chapter 2, Section 2.7). The developed simulation tool of multistatic radar system utilises the Range-Only detection method as an alternative approach for estimating target position, which does not require the knowledge of angle of arrival at the receiving site. Target location is estimated by associating multiple measurements from multiple bistatic radar pairs to find the intersection point among them. For instance, figure 4.1 shows Range-Only detection of a target by the simulated multistatic radar system of three receivers and one transmitter. Each received signal is translated into a range measurement that forms an elliptical curve around the bistatic pairs, where target location can be at any point on these ellipses. Therefore, the intersection point between the three associated measurements will represent the common solution for all three curves, which represent target position. Figure 4.1: Simulation tool Range-Only target detection 80

81 Range-Only detection approach provides the advantage of flexibility in multistatic radar design. That is, the detection process does not depend on angle of arrival, thus target detection process is independent from antenna configuration. Furthermore, Range-Only detection capitalise on the advantage of multistatic radar systems where information from multiple sensors are associated for better target detection and accuracy (see chapter 2, section ). However, this detection approach requires a minimum of three pairs of bistatic radars for target positioning, which can be seen as a potential limitation Ghost Target Detection Ghost target detection occurs from false association between range measurements during the process of Range-Only detection for multiple targets (does not occur for single target positioning). For example, figure 4.2 demonstrates the detection result of two targets in red, and one ghost detection in blue. The detection of ghost target was due to the intersection between three ellipses that are not generated from a radar return of an existing target; however they presented a mathematically valid intersection point and the detection process will include this point as possible target detection. The increment in number of target and/or bistatic pairs will increase the possibility of intersection among ellipses, which might increase likelihood of ghost detections. Also, in the case of multistatic radar, every three ellipses can have one intersection point in the horizontal of XY plane, because of geometry layout where all ellipses share one of the focal points at the transmitter site figure 4.2. The number of ellipses increases by the multiple of three with every additional target, thus for T number of targets, the minimum number of ellipses M = T X 3. This multiple increment in number of ellipses increases the number of possible intersections between M numbers of ellipses, which is given by the tetrahedral number formula as shown in equation 4.1, where N is number of receivers. Number of Intersection = M (M-2) (N-1) / 6 (4.1) For example, when simulating three receivers (N=3) multistatic radar to detect (T =10) targets, then the minimum number of ellipses is M = 30, and according to equation 4.1 the number of possible intersection is 290, which is a considerably high number of 81

82 potential ghost detections for 10 targets detections. Therefore, a devised approach for ghost elimination is required. Figure 4.2: Ghost target detection (in blue) Ghost Elimination by Multiple Results Analysis Approach (MRA) Bishop A. and Pathirana P. paper [9] discuss a similar issue of ghost targets detection when using Range-Only detection approach. Their approach for eliminating ghost detections was to relocate each mobile sensor multiple times. Thereby, the system captures multiple detection results from different viewpoints. Subsequently, results are analysed to eliminate inconsistent detection outcomes among every reading as ghost detections. The final results showed noticeable improvement for reducing the number of ghost detections. However, higher number of targets required higher number of runs to eliminate ghost detection. The simulation tool adapts the above approach for more accurate detection and to eliminate ghost targets. However, the simulated multistatic radar sensors are stationary antennas at fixed location. As an alternative, the concept of acquiring multiple results to analyse was implemented by reconfiguring the mapping of the multistatic radar sensors into subgroups of a minimum three receivers. Each subgroup provides an isolated 82

83 detection result. For example, the simulated multistatic radar of one transmitter and four receivers would have four possible subgroups of receivers as shown in figure 4.3. Each subgroup will act as a multistatic radar system to detect the same set of targets. At the end, the simulation will compare and analyse all results from every subgroup to compile the end detection list. Figure 4.4 shows the dramatic difference in the number of ghost targets detection between one-off detection run based on all four receivers, and the multiple results analysis run based on multiple subgroups results comparison. In this scenario, the filter managed to eliminate all ghost targets and successfully detect all real targets. Figure 4.3: MRA results, Real targets (in Red) and Ghost targets (in Blue). 83

84 Figure 4.4: Comparison between one-off detection and MRA. To further illustrate the effect of ghost target detection a test run was undertaken using different number of real targets. Figure 4.5 demonstrates the correlated increase between number of targets and ghost detection. The X axis signifies number of real targets and Y axis represents number of ghost detection. The red curve shows result from multistatic radar system of three receivers, as the dramatic increase in number of ghost detection correlates with the increase in number of real targets. On the other hand, the blue curve represents result from multistatic radar system of four receivers, where multiple subgroups of receivers were utilized to supply multiple results to be analysed. It is clear that the number ghost detection is significantly lower than that in the red curve, when applying the adapted filter of MRA. Figure 4.5: Targets vs ghost detections for 3 Receivers and 4 Receivers with MRA. 84

85 4.2.3 Target Localisation Algorithm The localisation algorithm by M. Malanowski [10] was primarily implemented for the target position component of the developed simulation tool. The choice of the algorithm was based on three attractive aspects of this technique; the algorithm is based on Range- Only detection, the technique uses closed-forms of equations and the capability of accommodating subjective number of transmitter-receiver pairs. This algorithm aims to find the intersection point between a given set of range measurements (i.e. the intersection points between multiple ellipses).the minimum number of measurements within one set is three measurements. During the period of one pulse detection, all captured range measurements are compiled into a range matrix. After that, the algorithm loops throughout all possible combinations of measurements sets from each bistatic pair, where each combination set is tested for intersection based on the mathematical solution proposed by [10]. To explain the localisation algorithm steps figure 4.6 is used to illustrate the geometry of three receivers and one transmitter. For simplicity, the transmitter position is set at the original coordination as (0, 0, 0) and the rest of radar nodes are altered accordingly. Figure 4.6: System geometry for three receivers [10] 85

86 The algorithm s objective is to calculate target position (x t, y t, z t ) from the given range measurement s equations by finding the common root. In this case, there are three pairs of bistatic radar; therefore the objective is to find the common root of the three range measurements equations. Equation 4.4 represents the range measurement equation which is equal to the range sum of transmitter to target range in equation 4.2 and the target to receiver range in equation 4.3, where Ri is the range measurement for the ( i th ) bistatic pair. (4.2) (4.3) (4.4) The next step is to rearrange equation 4.4 into the format of equation 4.5. Where all variables of the equation are known except x t, y t and z t (target coordinates). However, equation 4.5 can be re-written in Matrix vector form as shown in equation 4.6. (4.5) (4.6) Where X t is the target position matrix, S Matrix contains receivers position, Z represents the constant matrix and finally r is the range measurements matrix for the given input of range measurement combination set. Solving for X t, the last equation 4.6 transforms into equation

87 (4.7) At this stage, new notations (as a and b) are added to simplify the written form of equation 4.7 as equation 4.8 below or explicitly as equation 4.9 from the two equations: (4.8) (4.9) That is to note, R t is still unknown value within the two equations above 4.8 and 4.9. Thereby solving for R t, equation 4.9 is substituted into equation 4.2. The result is the quadratic equation Equation 4.11 represents the solution of the quadratic equation with respect to R t. (4.10) (4.11) Finally, target position X t is calculated after substituting R t into equation 4.9. When solving for the square root in equation 4.11 two mathematical considerations must be taken into account for the final solution for target position X t. The first is based on the (±) term, which implies that there are two possible solutions for R t. Consequently, target position X t will have two mathematically valid solutions as location points. For example, figure 4.7 show the possible two solutions for one target detection based on three measurements intersection. That is, the three ellipses intersects at two different points 87

88 (Tr1 and Tr2), in different orientations (tilt and rotation of the ellipses). This indicates that the two solution points have the same x t and y t coordinate and different z t coordinates. Figure 4.7: Two solutions for one target detection. The second mathematical aspect to be considered is within the square root notation in equation When the term has a negative value, then there is no possible real number solution for R t, which implies no solution for X t and no target detection based on the given input of that combination of measurements set. This aspect is used to test measurements association, where the negative value indicate that no intersection point between measurements. However, all programming languages have a finite number of bits to represent numbers. Therefore, when solving for intersections using the above method, numeric fluctuation between two very large numbers can be very small and will be misrepresented by the programming language s accuracy issue. This can affect the sign of the term within the square-root notation which may cause failure in detecting real targets. To explain this scenario, the first part of the square root in equation 4.10 is denoted as (A) and the second part as (B) as shown below. 88

89 For instance, when part A = X and part B = 1 x 10 20, then difference between these two large number is barely significant, yet (-1) causes no solution. This numeric problem caused an issue during the simulation testing for target detection. Better performance is to be achieved by rewriting the root term into simpler form, in addition to dividing calculation into three steps Optimizing Localisation Algorithm The localisation algorithm exploits every possible solution from every received pulse at the radar sensors. As it has been noted earlier, range measurements matrix is compiled and loops through all combinations. The number of combinations grows exponentially as the number of targets or sensors increases. For example, three receivers system and 10 targets to detect will generate a range matrix of (3 x 10) with (10x10x10 = 1000) of subrange combinations sets to loop through and check for solutions. In terms of adding one target to this scenario, the range matrix will become (3 x 11), thus it will exponentially increase the sub-range combinations sets number to (11 x 11 x 11 = 1331). As such, a range matrix of (N x M) will generate M N combinations. To address the issue of exponential growth of sub-range combinations sets number, the simulation tool utilizes multiple threads technique. The large numbers of sub-range combinations are divided into subgroups where each group is processed by one thread to look through for solutions. The simulation tool performance was enhanced by 75% faster when processing 1 million combinations. However, the number of threads to trigger simultaneously is limited on the simulation machine CPU type and specification. In some cases, excessive number of threads can result in over saturating the machine processor, which results in slower simulation performance. Figure 4.8 shows the steps for this process. 89

90 Figure 4.8: Multi-thread process diagram Localisation Algorithm Limitation The localisation algorithm [10] was primarily implemented to provide detection results based on Range-Only measurements. However, the algorithm had several limitations when considering a large numbers of targets within a dense environment such as a wind farm. The limitations are; the low performance for high number of targets detection, the inadequate accountability for measurements error (or radar range resolution), two solution aspect in elevation coordinate and the fluctuation in numeric values. The previous section addresses the optimization method for enhancing the algorithm performance. However, the objective of this research is to detect very large number of targets within a complex wind farm environment. Thus, the exponential increase in number of combinations sets with the increase in number of targets was evidently a significant limitation. When deriving range measurement from radar return, the range is accompanied with uncertainty or a margin of error according to the range resolution for that bistatic radar pair. That is why, the presentation of measurements as single elliptical curve around the bistatic pair does not accurately reflect the margin of error for that range resolution. To address this limitation, each combination set of measurements needs to be evaluated for possible intersection over multiple increments on each range according to the margin of 90

91 error for it. This will increase the computing time and severely impact the performance of the simulation tool especially when detection high number of targets. The final two limitations of this algorithm are the two solution aspect in elevation coordinate and the fluctuation in numeric values, as they have been discussed in section Due to the above limitations of localisation algorithm it was deemed that the use of the algorithm at its current state provides limited use for the detection of large number of targets within a wind farm environment. Therefore, alternative positioning and detection methods were investigated and implemented, such as PF or IA. 4.3 Particle Filter (PF) The objective of target detection and tracking is to estimate the change in target state over time, which can be referred to as a dynamic target tracking. In order to adapt to this dynamic estimation, two models are required, i.e. system model and measurement model [11]. The system model signifies the change in target state over time, whereas, the measurement model represents the system noise, namely the range resolution figure for a bistatic radar pair. When integrating these two models, they are represented by the Bayesian approach, where the objective is to construct the Probability Density Function (PDF) based on all available measurements [3]. Thereby, the tracker finds the likelihood of the posterior state of (A) given the measurement (B). As shown in the Bayesian equation P(A B) = P(B A)P(A) P(B) (4.12) The PF evolved from the Bayesian approach [3]. By principle, this filter is based on approximating the discrete posterior distribution by utilising sequential importance sampling framework [11]. There are two main stages in PF method, prediction stage and sampling stage. First stage includes the generation of random set of particles within the area of interest. Each particle represents state prediction with associated weight, which indicates the likelihood of that particle measurement to the true measurement that is registered by the system. The larger number of generated particles provides higher likelihood of particles with high weight association, which result in higher accuracy in target positioning/tracking. 91

92 The sampling stage redistributes particles that have low weight association figures to areas where particles have high weights. That is, the sampling step is a sequential selection with replacement from particles based on weight value. Consequently, particles with high weight are more likely to be selected more frequently than particles with low weight. In other words, particles with closer associated measurements to the true target measurements are going to survive and regenerate for the next time set. The number of sequential selection is equal to the number of generated particles. Figure 4.9 visually illustrates the main steps of PF. The prediction stage includes the first step (I) where particles are randomly generated with initial uniformed weight distribution for all particles. After that, weights values are updated according to the captured measurements during step (I), as it is visually demonstrated in figure 4.9, where bigger circles signifies larger weight values. The sampling stage is represented in step (II), which is according to weight updates process in the same step (II). Particles with low weight values disappear and repopulate around the area of measurements where particles with high weight association are located. Finally, step (III) resets particles weights uniformly and generate birth particles for the next association process with the new true measurements (of step VI) to update the newly distributed particles [11]. Figure 4.9: PF steps visualisation [11] 92

93 PF is frequently used to tackle multiple objects tracking in dynamic environments because of the accuracy, robustness and sensor variety [12]. However, the sampling stage in PF can be computational demanding and several versions of this filter were developed to optimize computational overhead. Such as, Sampling/Important Resampling (SIR) approach, Sequential Important Sampling (SIS), Exact particle filter and Auxiliary particle filter [11]. The question of which filter to use is down to the complexity of the implementation and the computing capability of the system. Nevertheless, the main advantage of PF is the feasibility of customizing the system model and the measurement model to suit the challenges of object tracking. For this research s implementation of PF the system model can be adapted to express the window of detection for the simulated multistatic radar system. The number of sampling iterations is set according to number of radar pulses within the window of detection. The measurement model can be customized according to radar measurements margin of error, which is set according to the range resolution of individual bistatic pairs of the radar system Implementation of PF The classic implementation of a SIR PF was adapted for multiple targets tracking within this research. The approximations of measurements weights were based on Gaussian distribution where the true measurement is the mean of the distribution and the variance is equal to range resolution (or measurement margin of error). In this stage, system range and velocity noises were assumed to be ±10m Prediction step At the final step of particle generation process of (N) number of particles, each particle is given a location Pr (Xp,Yp,Zp), velocity vector Vp(Vxp, Vyp, Vzp) and uniform weight (W) value, where W = 1/N. The next step is to update every particle measurement and weight according to its position from every sensor (i.e. the multistatic receivers). Each particle will be associated with multiple range measurements from each bistatic pair. The weight approximation of a particle is the product of the individual likelihoods of the particle range measurements to the true measurements of a target. 93

94 Sampling step The sampling step is based on multiple selections with replacement from the set of the generated particles. This process aims to exclude particles of low weight and regenerate particles of high weight in the regions of interest. The number of iteration is equal to number of particles (N) Implementation Example To showcase the previous steps, an example of PF process based on system measurements from a multistatic radar system with one transmitter and three receivers is used. The aim is to track one target with (N = 5) number of particles. Each particle is updated with three range measurements from each receiver accordingly and then the weight is approximated according to those measurements based on Gaussian distribution shown in equation The mean (ϻ) is the true target range measurement for a bistatic pair, (σ) is range noise, and (i) indicate the index of the bistatic pairs. n W = i Gaulssian(ϻ, σ, particle range measurement (i) ) (4.13) To demonstrate the sampling step numerically, the table 4.1 below shows the list of five particles, along with their assumed weights figures. By normalizing weights, the probability of not selected is found from equation Probability Not select (Pr 1 ) = (P(Pr 2 ) + P(Pr 3 ) + + P( Pr N )) N (4.14) Table 4.1: Numeric illustration of the sampling process for five particles. Particle Weight Normalised Weight Probability of NOT Select Probability of Select Pr / 6.0 = Pr / 6.0 = Pr / 6.0 = Pr / 6.0 = Pr / 6.0 = Total = 6.0 Total = 1 94

95 It is clear that particles with high weight such as (Pr3) have very low probability of not being selected during the sampling process. On the other hand, (Pr1 and Pr5) holds the highest probability of not being selected as they have the lowest figures of weights among the five particles. Consequently, five iterations (N = 5) of selections-withreplacement will result into selecting (Pr3 and/or Pr2) multiple times, and (Pr1, Pr4 and Pr5) are eliminated PF Simulation Tool Results Figure 4.10 demonstrates the result from the simulation tool for tracking one target based on the PF method. Figure 4.10.a represent the prediction stage of PF, where (N = 1000) particles were randomly generated around the target, and figure 4.10.b shows the next step after sampling all particles. It is clear that the PF initiates a track on the target after completing the sampling step successfully. This can be seen as all particles are repopulated around target position as shown in Figure 4.10.b. The velocity measurement was integrated for particle weight approximation. Thereby, the sampled particles will have the highest likelihood of associated range and velocity measurements to the true target range and velocity measurements. Alternatively, weight association based on range measurements only can result into qualifying the closest particle from the true target, but the velocity vector might be in different orientation. This can result into sampling particles that follows different velocity orientation than the real target. Figure 4.10: Simulation tool PF tracking one target. 95

96 In the case of false target detection, such as the illustrated detection scenario in figure 4.11, particles are regenerated around the false detection at first in step 1 and 2 of figure 4.11, but when this ghost target is no longer detected during the subsequent time frames in figure 4.11 steps 3 and 4, then these set particles are going to have lower weights association to the true measurements and they will fade away from that area to repopulate elsewhere. This feature of PF is considered as an advantage for utilising multiple pluses ghost elimination technique. Figure 4.11: PF multiple targets tracking and ghost elimination 96

97 4.4 Interval Analysis (IA) IA utilises set of numerical interval variable, instead of set of number points. Each interval is represented with lower bound (A ) and upper bound(ā), where all numerical values within those two boundaries belongs to the interval variable, as defined by the expression shown in equation 4.15 [5]. [ A, Ā ] = {A R, A A Ā}. (4.15) All basic mathematical operations apply to the interval variables. Thereby, equations can be rewritten by replacing single numeric variable with interval variables [5]. The constraint sets represent the input functions (or parameters) for the system. For example, each function that represents the range measurement is characterized as constraint input for IA process. The separator method calculates all possible constraint outputs within the solution search interval [4]. In other words, separator method divides the wide solution interval into three areas. As shown in figure 4.12, i.e. separators of a ring function, the first area in blue where the solution is outside the constraint, the second area in red where the solution is inside the constraint and finally in yellow is the third area where the solution is undetermined. Figure 4.12: Separator s three output regions 97

98 SIVIA: The Set Inverter via IA, solves set inversion problem using intervals [5]. Based on separator sets, SIVIA iterate throughout the wide search interval for solutions that satisfy the constraint values set from each separator. The desired solution is obtained through the iterative process of bisecting the search interval into left and right boxes [5]. Then check if the inverse of the box in hand reflects the constraint values. The bisection continues until the box size reaches the minimum dimensions that are defined by the system resolution. Visually, the iterative effect of SIVIA process is shown in figure 4.12, where multiple bisected boxes appear in different sizes. The SIVIA algorithm takes three inputs, the first input is the search interval [X](0), the second input is the separator (S) for each constraint (Y) and finally the third input is system resolution (ε) or the margin of error for the system. Steps of algorithm are listed in the Figure Figure 4.13: SIVIA algorithm routine [4] The final result from SIVIA represents the separator three solution areas. The first set includes all boxes that are within the constraint solution. The second set includes all the boxes that are outside the constraint solution. Finally the third set contains all boundary boxes where the solution area is extended based on system resolution. 98

99 4.4.1 IA Target Positioning Target position is represented by the interval vector [[X], [Y], [Z]] where [X] is the interval of [ε-x, ε+x] and so forth for [Y] and [Z], where (ε) denotes range resolution of each bistatic pair. Visually, the interval vector is represented by a 3D box. Transmitter and receiver locations are denoted by interval vectors {[Xt], [Yt], [Zt]} and {[Xr], [Yr], [Zr]} respectively. The bistatic range measurement (m i,j ) of a target is represented by the following constraint equation [m i,j ] = R t + R ri (4.16) Where: [m i,j ] is the measurement with uncertainty. i {1 N}, j {1 M}, N is number of receivers, M is number of measurements by each receiver; Rt = ([X] [Xt]) 2 + ([Y] [Yt]) 2 + ([Z] [Zt]) 2 (4.17) Rr = ([X] [Xr]) 2 + ([Y] [Yr]) 2 + ([Z] [Zr]) 2 (4.18) The above equations 4.17 and 4.18 provide the constraint input for the IA method. When solving for separator of one constraint the result will form an ellipsoid shape around the bistatic pair. The ellipsoid surface contains all possible solution for target position. Noticeably, each segment of the ellipsoid contains a list of boxes where the solution is assured (red), in addition to a list of boundary boxes where the solution is possible based on range resolution (yellow). Figure 4.14.a shows the guaranteed solution sections, which is enclosed by the boundary boxes in yellow of figure 4.14.b. For further illustration, a cut along the XY plane is shown in figure Figure 4.14: 3D separator result of bistatic ellipsoid. 99

100 Figure 4.15: XY plane section from the 3D ellipsoid separator of range measurement. When detecting multiple targets, the bistatic pair registers multiple range measurements. Each range measurement represents a constraint and solved by a separator. Therefore, the union of all separators will define the total detections for a given bistatic pair. For a multistatic radar system the target position is estimated by finding the intersection between all separators groups for every bistatic pair as denoted in equation The steps of positioning algorithm are listed in figure The final step of the algorithm indirectly associates target measurements from multiple receivers. That is, instead of searching for associated measurements to find intersection among them, the algorithm takes all the possible solutions of bistatic radar range measurements and intersect them with other bistatic pair measurements simultaneously. However, incorrect associations can occur during the intersection step which may lead to increase in number of false detections (or ghost targets). Figure 4.16: Target detection algorithm routine 100

101 The solution set for all the measurements from all receivers is: N M S = S i,j i=1 j=1 (4.19) Where the solution set for one measurement is defined as: S i,j = {x R, y R [m i,j ] = R t + R ri } (4.20) To locate a target multiple bistatic pairs are required. Three pairs are considered to be the minimum number; because two measurements are more likely to have multiple intersections at separate locations. Whereas, three associated measurements intersect around one location Modelling IA Results The IA component was developed in python using PyIbex [13] interval library. There are Ibex IA library available in C++ and Matlab as well. However, the Python library was chosen for its ease of integration and compatibility with the java packages of the simulation tool. The solution was graphically displayed using VIBes viewer API, which plots IA figures in 2D to show inbox, outbox and maybe boxes solution. To showcase the simulation tool implementation of IA algorithm, a scenario is considered based on multistatic radar system of one transmitter at (2000, 2000, 10) and four receivers at R1 (2000, 6000, 10), R2 (3500, 6000, 10), R3 (6000, 3500, 10) and R4 (6000, 2000, 10) as shown in figure Within the presented modelling case, it was assumed that all of the Txs and Rs use omnidirectional antennas and they are synchronized in time. The collected range measurements from all four bistatic pairs are processed within a centralized data fusion center. Number of measurements per receiver at a time step could vary. For example, when one out of the four receivers registers two measurements while the rest have one measurement each. 101

102 Power Received / db Figure 4.17: The union of all measurements Separators. Figure 4.18, shows power received (in dbs) from a target with 20m 2 RCS by each receiver, across a range line (in meters) that represent the bistatic range of the target. The threshold of power received is set to (-100 db) and range resolution (ε) is set to [-5m, +5m]. The position of each peak represents the range measurement for that particular bistatic pair. At each time step, range measurements are extracted from the range line at each peak (Figure 4.18: (m1, m2, m3 and m4)) to compile into one range measurements matrix (N x M). Where N is number of receivers and M is the maximum number of measurements m1 m2 m3 m4 Receiver1 Receiver2 Receiver3 Receiver Range Measurement / Km Figure 4.18: Multistatic radar return from one target. Each entry from the range matrix is defined by its constraint based on equation 4.16, and then the separator result is processes as shown in figure Noticeably, each separator set forms elliptical shape in XY plane, and all four associated ellipses pass through the target position where the intersection point is. Figure 4.19, shows the estimated 2D target 102

103 location from the intersection of the four separators at the surface of the XY plane where Z is assumed to be zero. Figure 4.19: The intersection of all separators To illustrate the estimated 3D target position, figure 4.20 and figure 4.21 show the intersection result of all separators at the XY plane and the XZ plane respectively. In XY plane, the predicted target position is accurate and within the interval vector [[X± ε ], [Y± ε ]] with some exception for the case of boundary yellow boxes outside that vector. However, in XZ or YZ planes the predicted target elevation extends far beyond the real elevation interval [Z ± ε ]. In this case, the real target elevation was set to (100m) and the predicted target elevation was within the interval of [0, 240] meter. Figure 4.20: IA one target detection in XY plane. 103

104 Figure 4.21: IA one target detection in XZ plane. The extended range in elevation prediction is due to the nature of geometry of multistatic radar network. Range measurements generated from a multistatic radar system of multiple receivers and one transmitter will share the same elliptical focal point at the transmitter site. Consequently, the intersection area among the associated measurements is more likely to form an arc along the elevation dimension. Furthermore, in practice the height of multistatic sensors tend to not vary in big margins. Thus the difference in elevation tilt among ellipsoids may not be significant, which will increase the intersection surface between ellipsoids in elevation. To investigate the effect of target height on the predicted elevation interval, the target height was varied. Figure 4.22, shows the result of predicted elevation interval length as the target height was changing from 0m to 1000m. The layout of multistatic radar sensors was as same as the layout in figure At each run, target height was increased by 10m, while the X and Y coordinates remain fixed. The detected elevation accuracy was calculated from subtracting the minimum from the maximum elevation intervals. 104

105 Detected (X/Y) Interval Lenght (m) Detected Elevation Interval Lenght / m Z Guaranteed Z Boundary Target Elevation Z (m x 10) Figure 4.22: The length of the estimated target elevation by IA Noticeably, at a lower elevation the predicted elevation interval is significantly high for both the guaranteed solution boxes and the boundary solution boxes. However, the elevation accuracy improved considerably once the target height is above 250m due to the geometrical setup of the bistatic pairs. It was also noted that the accuracy of the XY prediction for the target location was not affected by varying its location along the XY plane. Further test for X and Y coordinates was done in the similar context as the above test, as show in figure The results show better accuracy and consistency for X and Y interval detection length than Z dimension. For example, the red and green curves represent the detection of guaranteed boxes for X and Y accordingly, where the interval length remains between 5m to 10m. This result was accurate according to the assumed range resolution of (±5m) X InBoxes X MaybeBoxes Y InBoxes Y MaybeBoxes Target X/Y (m x 10) Figure 4.23: The length of the estimated target (X/Y) by IA. 105

106 Number of False Detection IA Post Detection Processing IA state estimation of a target is represented by multiple interval vectors of adjacent boxes as shown in prior plots figure 4.20 and 4.21 where the real target position could be within one of the vectors. An estimated interval vector (or box) is labelled as ghost detection when a real target position is not included within that interval vector. Consequently, for multiple adjacent vectors of target estimation only one vector will include target position and the rest are labelled as ghost detections. However, all the surrounding boxes represent a valid estimation results for that target. Therefore, further results processing were needed to merge adjacent boxes and/or combine nearby boxes within one box estimation. This process aims to consolidate target estimations into one result to reduce number of unnecessary labelling as ghost detection. Figure 4.24 illustrates the progressive drop in ghost targets detection counts when further post detection processing was done. In figure 4.24, C1 to C2 indicates first processing of combining/merging boxes, then (From C2 to C3) is the second processing step as further boxes were combined/merged by lowering accuracy level -so more nearby boxes will qualify to be combined. The post detection processing can be done in one step by establishing a threshold of desired accuracy level as the limit. Nevertheless, multiple processing steps were considered for illustrating the drop in false targets detection figures based on deferent accuracy levels C1 C2 C3 C4 C Time Steps Figure 4.24: Post detection processing effects on false detection rate The by-product of the post detection processing is the decline in detection accuracy. Figure 4.25 shows the impact and the compromises of multiple post detection processing 106

107 on the accuracy of target detection. When merging nearby or combining adjacent results, the interval dimensions increases which results into lower detection accuracy. Accuracy / m First output C1 C2 C3 C4 C Time Steps Figure 4.25: Post detection processing effects on detection accuracy level 4.5 Comparison between IA and PF This section compares the performance of PF and IA in multiple targets detection and tracking based on Range-Only detention approach. The comparison will be according to three performance criteria, number of real targets, number of false targets and the detection accuracy. For this comparison, the simulated multistatic radar includes two transmitters and four receivers as shown in figure Each mapped subgroup of one transmitter and four receivers generates an independent detection result. The intersection between the two results will represent the final detection output of the system. Figure 4.26: Multistatic sensors mapping with transmitters 1 and 2 107

108 The comparison data were compiled from the simulated scenario of randomly generated 20 moving targets for the duration of 40 time step, where each time step represent one pulse detection. Each target propagates according to the velocity vector assigned to at the first step. The velocity vector for each axis was randomly modified throughout the consecutive time steps by a value between (±5m/s) to simulate the change in speed and direction of a target. Targets were densely positioned within the surrounding area of radar coverage to emulate complex detection scenario for both tracking methods. Thus, the difference among the captured Range-Only measurements could be close and the likelihood of false measurement association is high, which can increase the possibility of ghost targets detection Number of Real Targets Detection The IA approach was considerably successful in detecting the 20 targets during the 40 time steps, as the detection counts were within 17 to 20 targets. However, PF approach shows lower success rate. Results were between 6 to 15 targets throughout the 40 time steps test. As shown in figure 4.27 IA approach searches for all possible solutions according to the given measurements, therefore, it is more likely to include real target detection. There are two possible reasons for not detecting a target. The first reason, when the distance between two targets is less than the range resolution for the system, then only one target will be detected. The second reason is due to MRA, whereby, the target could be missing from one result set and the analysis process will exclude that target from all other result sets. The low counts of real target detection by PF were due to particle nomination approach and ghost elimination method. The nomination approach is based on selecting particles of the highest weight within each range cell. Knowing that PF is a probability estimation approach, the results will vary for each detection process of the same set of targets. The nominated particles for each run will be at the same range cells but could be at different coordinates. When considering MRA for eliminating ghost target detection, the comparison between the multiple results will be difficult to correlate due to the 108

109 Number of Detected Targets aforementioned fluctuation in estimated positions and might result in the loss of real targets. This issue impacted the analysis process of ghost elimination approach and real targets recognition Interval Particle Time Steps Figure 4.27: Number of detection from 20 targets return across 40 time steps Number of False Detection The comparison based on number of false detection between PF and IA shows the particle approach produces less false detection than the interval approach, as shown in figure The lower false detection rate in particle approach because of the sampling steps, whereby false detection might had been recorded during earlier steps but further sampling and updates steps can repopulate the falsely associated particles toward a rightfully associated measurements of a real target. Interval approach detects all possible associations among all measurements which offers high detection rate of real targets. On the other hand, the possibility of false association between measurements is high, which can results in higher false detection rate. Particularly in the presented scenario, where targets are densely positioned, which increases the possibility of measurements intersecting elsewhere than at a target position. 109

110 Number of False Detection Interval Particle Time Steps Figure 4.28: Number of false detection from 20 targets return across 40 time steps Target Detection Accuracy The third comparison criterion between the two methods is based on the accuracy of the estimated position in comparison to the target s real (exact) position. The comparison indicates better accuracy results with PF method than the IA method, as shown in figure The compared accuracy values represent the average of all detection accuracy results for each time step (40 time steps). That is, the average accuracy figure could be averaged for 20 targets detection at a time step or less than 20 targets at other time steps where it depends on number of detections (as discussed earlier). The accuracy for the PF estimation was calculated based on the distance between target position and the position of the nominated particle within that target range cell. If multiple particles were nominated, then the closest particle is considered. On the other hand, the estimated detection results from the IA method are represented as interval vectors of a rectangular shapes in X and Y dimensions. All possible combinations of x and y coordinates within that interval vector are valid to be the estimated detection. Therefore, target detection accuracy of the interval method is taken as the diagonal distance of the interval vector that includes the real target position, which represents the worst case scenario. PF shows better detection accuracy because the nominated particle can be no farther than the system range resolution. However the interval method accuracy level is lower 110

111 because of the post process which merges adjacent boxes and/or combining boxes that are located within a distance less than the system range resolution. Accuracy / m Interval Particle Time Steps Figure 4.29: Average detection accuracy of 20 targets return across 40 time steps A closer look at the accuracy of PF, the plot in figure 4.30 shows the minimum, maximum and average accuracy values for the 20 targets detection, over 40 time steps. The average figures of accuracy are around the 15 meters line, where the accuracy minimum figures for an individual target is below the 10 meters mark and the maximum accuracy could reach the 50 meters. On the other hand, the plot in figure 4.31 of the interval method minimum, maximum and average accuracy values shows higher line of average accuracy around 23 meters mark. The minimum accuracy for individual target detection sits higher than 5 meter because the minimum estimated box diminutions are set by the range resolution of (±5m). Accuracy / m Avg Poly. (Avg) Time Steps Figure 4.30: PF Minimum-Average-Maximum detection accuracy. 111

112 Accuracy / m Frequency Avg Poly. (Avg) Time Steps Figure 4.31: IA Minimum-Average-Maximum detection accuracy. From the above two plots, figure 4.32 illustrates the frequency of the detection accuracy values within the range interval of 0 to 40 meters for PF and IA. The plot concludes that PF is more frequent to hit lower detection range accuracy than the interval method. However, the variant for both curves is around 10 meters range Frequ. Interval Frequ. Particle Accuracy / m Figure 4.32: Frequency of detection accuracy of IA and PF. 4.6 Detection History and Target Tracking The modelling tool tracking algorithm is based on sustaining a history log of the detected targets along with predicting the next step estimations. The two main objectives of this algorithm are to track multiple targets and to eliminate ghost detections. History log includes a list of targets entries, whereby each entry stores target position, velocity vector component and three flags variables as shown in figure Lost_flag variable is raised if a target has not been detected by the current detection result. 112

113 Track_flag indicates the initiation of track lock of target based on multiple detections across a set number of pulses. Finally Found_flag is used by the algorithm to indicate that a target is found within a set number of pulses after being lost. Figure 4.33: Target entry in history log list. History log is kept up to date during every time cycle (pulse) of radar detection, which includes two processes. The first process figure 4.34 is to loop throughout the history log list and compare against the new detection result for matching entities. This process aims to identify missing targets that was in history log but not in current result. This process also removes lost targets that have been undetected for a set number of detection cycles (i. e. Window of detection). Figure 4.34: First process of History log target tracking 113

114 The second process figure 4.35 is to loop throughout the new detection results to compare it against the history log. If new and unmatched detections are found, they are then added as new entries onto history log. In addition, matching entries are flagged as tracked targets and all variables of those history log entries are modified according the new results, such as the target s new location and velocity vector. Figure 4.35: Second process of History log target tracking. Window of detection is set to three pulses so that when target remains undetected for three consecutive pulses it will be eliminated from the history log with no further tracking. This approach helps in eliminating ghost targets based on multiple pulses detection, which depends on the history log to compare results from multiple consecutive pulses. For example, if ghost target is detected at time (t0), but then at times of (t1, t2 and t3) it remains undetected, then this target will be eliminated as a ghost target. Consequently, the simulation tool compiles all results during the detection window of three pulses then displays the final history log. The three pluses detection window was arbitrarily set for this research. Longer detection windows can be can be considered which may improve the overall detection performance. Given that Multistatic radar systems can operate at higher pulse repetition frequencies, using more pulses can still provide rapid tracker updates (faster than the monostatic, since it often depends on the mechanical radar rotations). 114

115 4.7 Simulation Tool Results of History Tracking To illustrate the result from history log tracking processes figure 4.36 shows three consecutive screen shots from the simulation tool. At the first detection (pulse at T0), two ghost targets were detected. However, during the second detection cycle (pulse at T1), real targets positions shifted according to their velocity vector and the old ghost targets are not visible anymore. Yet, new ghost targets appear elsewhere. Detection at (T2) shows two deferent ghosts detection in this scenario. In conclusion, History log tracker filters out ghost targets by finding the consistency in detection results across multiple detection cycles. In figure 4.36 the tracker overlays the three result images on top of each other, and any mismatched detection will be disregarded from the final displayed result. Figure 4.36: Three pulses detection results 115

116 4.8 Chapter Summary It has been established within this chapter that detecting and tracking larger number of targets by multiple sensors system such as the multistatic radar is a challenging process. The difficulty of such detection approach was due to large number of measurements to associate as well as the task of false detections elimination. To address this challenging process, the research examines multiple detection methods to predict the unknown system status from a list of input measurements with uncertainty of error margins, such as Particle Filter and Interval Analysis. The simulation tool utilises Range-Only target detection approach. This approach requires multiple range measurements from different antennas to be associated for finding the intersection point among them. Thereby, the estimation of angle of arrival at the receiving site was not required for target positioning. The by-product of Range-Only approach is the potential of ghost targets detection due to false association between range measurements during the process of target detection. This disadvantage tends to be acute when detecting very large number of targets because of the increase in number of measurements to be associated. This will result into higher probability of false association. The simulation tool implemented new approach (Multiple Results Analysis MRA) for eliminating ghost detections based on one pulse detection results from multiple subgroups of multistatic radar systems. The multiple detections results were acquired by reconfiguring the mapping of the multistatic radar sensors into subgroups of a minimum three receivers. Each subgroup provides an isolated detection result. Multiple results are analysed to eliminate inconsistent detections as ghost targets. This approach proved to be effective in eliminating large number of ghost detections as it was shown in section The localisation algorithm by M. Malanowski [10] was primarily implemented for the target position component of the simulation tool. This algorithm aims to find the intersection point between a given set of range measurements (i.e. the intersection points between multiple ellipses). The critical analysis of this algorithm shows several limitations when processing large numbers of targets within a dense area. The limitations 116

117 were; the inadequate accountability for measurements error (or radar range resolution), two solution aspect in elevation coordinate and the fluctuation in numeric values. Additionally, the algorithm required large memory allocation when processing very large number of targets. Therefore, an alternative detection method was considered such as PF and IA. PF approach is a statistical model for system state estimation based on the likelihood of true measurements matching the sample measurements of a particle. This approach is widely used for the scenario of multiple sensors and measurements association, which suits the application of multistatic radar target detection based on range only approach. The chapter explains the basic steps of PF from particle generating step, to weight update step and then sampling step according to input measurements. This approach was implemented by the simulation tool; as results were illustrated in section and section PF provided the advantage of ghost targets elimination during multiple consecutive pulses. That is, when ghost target is no longer detected during the consecutive time steps of radar window of detection, particles around that ghost target will repopulate elsewhere around real targets detections during later time steps. See figure IA utilises set of numerical interval variable, instead of set of number points. Each interval is represented with lower bound and upper bound, where all numerical values within those two boundaries belong to the interval variable. This mathematical approach was utilised for target positioning estimation of multistatic radar system in 2D (horizontal plane). This chapter illustrates the stated novelty of this research where the implementation of IA method for target positioning was expended in 3D (horizontal and vertical planes). According to results of one target detection, the IA approach showed an accurate estimation for target position in the horizontal plane (XY Cartesian coordinates). However, in XZ or YZ planes the estimated target elevation extended far beyond the real elevation. This limitation in elevation accuracy was due to the nature of geometry of multistatic radar network. Range measurements generated from a multistatic radar system of multiple receivers and one transmitter will share the same elliptical focal point at the transmitter site. Consequently, the intersection area among the associated measurements is more likely to form an arc along the elevation dimension. 117

118 The two implemented methods for target position estimation (i.e. PF and IA) were further analysed by comparing them according to three performance criteria, number of real targets, number of false targets and the detection accuracy. IA showed better rate in number of real targets detected than PF, because IA approach searches for all possible solutions according to the given measurements, therefore, it was more likely to include real target detection. Consequently, solving for all possible intersections will produce larger number of ghost detection, which is why the comparison results showed higher ghost detection in IA approach than PF. Finally the estimation of target positing by the IA method was lower than PF due to the post results processing of interval vectors as explained in section The final section of this chapter discussed the detection history log and target tracking algorithm. The main two objectives of this algorithm were to maintain log of all detection results throughout multiple time steps of radar detections and initiate target tracking based on that log. The algorithm was set according to the defined window of detection for the radar system. The inconsistent detection results of radar returns during the period of window detection were labelled as ghost detections. see section

119 4.9 References: 1. RONALD P. S. MAHLER, Statistics 101 for Multi sensor, Multi target Data Fusion, IEEE A&E SYSTEMS MAGAZINE VOL. 19, NO. 1, JANUARY M. Tobias, Probability Hypothesis Densities for Multi target, Multi sensor Tracking with Application to Passive Radar, School of Electrical and Computer Engineering, Georgia Institute of Technology, May I. A. Siradjuddin, M R. Widyanto, Particle Filter with Gaussian Weighting for Vehicle Tracking, Trunojoyo University, Indonesia, April 27, L. Jaulin, M. Kieffer, O. Didrit, and E. Walter, Applied Interval Analysis. New York: Springer, R. B. Kearfott, Interval Computations: Introduction, Uses, and Resources, University of Louisiana at Lafayette, Department of Mathematics, Lafayette, United States. 6. R. E. Moore, R. Backer Kearfott, M. J. Cloud, Introduction to Interval Analysis, the Society for Industrial and Applied Mathematics (siam), G. L. Soares, A. Arnold-Bos, L. Jaulin, C. A. Maia_, J. A. Vasconcelos, An Interval- Based Target Tracking Approach for Range-Only Multistatic Radar, IEEE TRANSACTIONS ON MAGNETICS, VOL. 44, NO. 6, JUNE T. L. Song, D. Muˇsicki, Range Only Multistatic Tracking in Clutter, Tyrrhenian International Workshop on Digital Communications, 2011, ISBN: A. Bishop, P. Pathirana, Localization of Emitters via the Intersection of Bearing Lines: A Ghost Elimination Approach M. Malanowski, An Algorithm for 3D Target Localization from Passive Radar Measurements, 2009, SPIE Vol B Maheswaran Subramaniam, Multitarget Tracking Using Multistatic Sensors, Phd thesis, the school of graduate studies of Mcmaster University September M. Johannes, N. Polson, Particle filtering, University of Chicago, R. B. Kearfott, Interval Computations: Introduction, Uses, and Resources, University of Louisiana at Lafayette, Department of Mathematics, Lafayette, United States. 119

120 14. G. Soysal, M Efe, Data Fusion in a Multistatic Radar network Using Covariance Intersection and Particle Filtering, Ankara University, Ankara, Turkey, IEEE, M. B. Guldogan, D. Lindgren, F. Gustafsson, H. Habberstad, U. Orguner, Multi- Target Tracking with PHD Filter using Doppler-Only Measurements, Digital signal processing, W. Shenshen, C. Wanfang, F. Jinfu, W. Fangnian, B. Yun, Data Association Algorithm for Bistatic Radar Network, International Workshop on Information and Electronics Engineering (IWIEE), T. Zajic, R. Mahier, A particle-systems implementation of the PHD multi target tracking filter, Signal Processing, Sensor Fusion, and Target Recognition XII, Proceedings of SPIE Vol. 5096,

121 Chapter 5 Wind Farm Modelling Results 5.1 Introduction This chapter aims to demonstrate detection results of IA and PF methods based on complex detection scenarios involving a large wind farm and multiple targets. The configuration of the simulated multistatic radar system is similar to chapter 4 of two transmitters and four receives. Ultimately, a comparison was conducted between the two methods of PF and IA based on detection accuracy, false alarm rate and turbine detection rate. The research introduced the Multiple Results Analysis (MRA) technique for ghost targets elimination, in chapter 4 section Within this chapter an extra transmitter is introduced into the multistatic system configuration giving a total of two transmitter and four receivers. The MRA technique was modified to allocate the subgroups based on having each transmitter paired with the four receivers. This adjusted approach for the MRA technique is referred to as Multiple Transmitter Results Analysis (MTRA) technique throughout this Chapter. The chapter concludes by presenting results of three detection scenarios of multiple targets within a wind farm environment based on IA algorithm. The detection scenarios included ground level targets, mid-evaluation targets and high-elevation targets. 121

122 5.2 The Detection of Wind Turbine The multistatic radar system shown in figure 5.1 detects the wind turbine by associating the detection results from its four receivers. All measurements registered by each bistatic pair are sent to central processing unit for data fusion by measurements association to estimate turbine position. In the case of wind turbine detection, each visible section (from the bistatic point of view) of the turbine tower or blades can potentially generate different radar return, thus multiple measurements can be observed of one turbine. The range measurement was estimated by the direct detection approach as the two antennas are within the LOS of each other (see chapter 2). The bistatic radar return of a turbine depends on the blades rotation angle and turbine orientation according to wind direction. For instance, the orientation of the detected turbine in figure 5.1 was set to face north (upward) and blades rotation angle was at 30. According to the plot of bistatic power received in (dbw) across the range line in figure 5.2, receivers 1 and 2 registered one peak in power received at one range point. On the other hand, receivers 3 and 4 showed multiple peaks in power received because more sections from the blades were visible at different range cells from the receiver point of view. Further explanation of wind direction effect on wind turbine radar returns will be discussed in section (5.5) Figure 5.1: Wind turbine Range-Only detection plot (XY plane) 122

123 Power Received (db w) Detection Range to Receiver (m) Receiver1 Receiver2 Receiver3 Receiver Figure 5.2: Wind turbine bistatic radar return at four receivers To detect the turbine location from the radar returns shown in figure 5.2, the range measurements are extracted at each peak point in power received to compile the measurements matrix. The detection process was based on the Rage-Only approach to find the intersection point (i.e. the common solution point among measurements from each bistatic pair). This step was achieved by utilising state estimation algorithms such as PF or IA methods (as discussed in the previous chapter 4). The estimated result by IA is shown in figure 5.3. The plot shows the three dimensional planes XY, XZ, and YZ for the wind turbine detection. The blue boxes represent the position of turbine sections and the red boxes are the estimated interval vectors of IA. The top view of (XY) plane shows the interval vectors are located around the tower and the right hand side where the two blades are located (as shown in (XZ) plane, figure 5.3). The side plot of (YZ) plane shows the estimated detection was accurate around the width of the turbine. The limitation of the IA in elevation accuracy still occurs in this scenario. The XZ and YZ planes in figure 5.3 shows the highest section of the turbine was at (140m), but the estimated interval vector extends to (160m). Figure 5.3: Wind turbine detection estimate by the IA approach 123

124 5.3 Post Detection Processing The estimated outcomes of IA are presented in multiple interval vectors. Some of the vectors are adjacent or located nearby each other. From the radar detection prospective, adjacent vectors are merged and the nearby vectors are combined as long as they collocated within the same range resolution cell of the radar system. This process was discussed in the chapter 4 (section 4.4.3). Figuer 5.4 shows the result of post detection processing for the IA estimation of one turbine detection. The process merged and combined the first group of intevale vectors around the tower sections and the second group of interval vector around the blade sections separetly as the two groups are located at diffiernt range cells. The end detection result were three estimated interval vectors as shown in figure 5.4.b and figure 5.5.b. Figure 5.4: Post detection processing of IA estimates for wind turbine (XY plane) Figure 5.5: Post detection processing of IA estimates for wind turbine (XZ plane) 124

125 5.4 Multiple Transmitters Result Analysis (MTRA) Multistatic radar system of multiple transmitters and receivers offers the advantage of further result processing by comparing multiple detection results from each transmitted signal to eliminate ghost detection. When considering one turbine detection figure 5.6 shows the interval estimations in (XZ) plane of one transmitter on the right and the result from utilising two transmitters on the left. As the two transmitters are located at different sites, then each bistatic pair will have different view point of the turbine orientation. Consequently, the results generated from each transmitted signal will be different. For example, the first mapped group of one transmitter and multiple receivers can detected more sections than the second group of the other transmitter. That is why; the intersection between the two sets of detections will produce less detection results. Figure 5.6 illustrate this outcome where the detection based on one transmitter estimated the location of more sections from the right blade than result based on two transmitters. Nonetheless, the detection accuracy improved with the intersection of the two results, which will provide a less cluttered map for postprocessing. Figure 5.6: Comparison between (2Tx and 1Tx) detection for one turbine. MTRA technique shows promising advantage in improving detection results in the case of larger number of targets detection. For example, a small wind farm of four wind turbines in figure 5.7 was detected by a multistatic radar system of one transmitter and four receivers configuration. The result shows estimations around the four turbines, in addition to multiple false detections as ghost targets at various locations, as it is visible from the enlarged section around the third turbine (Tu3) of figure

126 Figure 5.7: Four turbines detection (one transmitter and multiple receivers) Evidently, by utilising extra transmitter to detect the late four turbines, the result in Figure 5.8 shows how ghost detections were eliminated around (Tu3) in addition to the improvement in elevation accuracy. Such technique of MTRA increases the processing time of target detection; however it is highly effective in eliminating ghost targets, particularly when handling large number of detections such as large wind farms of 25 turbines or more (see section 5.7). Figure 5.8: Ghost targets elimination by MTRA technique 5.5 The Detection of 5x5 Wind Farm The simulation results of this section were produced based on a four receivers configuration similar to the configuration in the previous section. However, two transmitters were utilised to produce the final detection result based on the MTRA technique. To analyse the detection result of the 5X5 wind farm formation, a breakdown 126

127 illustrations for each bistatic pair radar returns (of one multistatic radar subgroup) was considered as shown in figures 5.9, 5.11, 5.12 and 5.13 Figure 5.9.a shows the elliptical curves that represent the constant range measurements that were derived from each peak in power received at a range cell. These ellipses were based on the power received shown in figure 5.9.b, which illustrates the bistatic radar return from the wind farm for Receiver1. Wind direction was set to face north (the angle of wind direction = 0 ) therefore the wind turbines were facing the baseline of the bistatic pair as shown in figure 5.9. In this case, the turbine sections were co-located within the same range cell resolution from the bistatic pair s prospective. Consequently, the bistatic pair will observe one peak in power received at the turbine bistatic range. The radar return was mainly of the tower in addition to some reflections of the blades, which were combined to the overall radar return at that range cell. Further illustration of turbine orientation effects on power received for the same bistatic pair is shown in figure In this scenario, the angle of wind direction was set to 75. The change in turbine orientation altered the bistatic view point of the turbine. Tower and blades sections were observed at different range cells as shown in the enlarged section of figure 5.10.a, which illustrates the turbine s new orientation. The largest value of power received was from the tower; however, parts of the blades were visible at different range cells with smaller peaks in power received. This change in power received is shown in figure 5.10.b where for each turbine multiple ellipses are introduced due to the multiple peaks in the power received. Going back to the scenario discussed earlier where the wind direction is at 0 and illustrated in Figure 5.9, figures 5.11, 5.12 and 5.13 shows the remaining breakdown of the detection results for the 5x5 wind farm for the bistatic pairs of receiver2, receiver3, and receiver4 respectively. Different baseline orientations of different bistatic pairs will alter the power received plot. It can be noted that for receiver3 the bistatic pair (the baseline) is located at an orientation that will see more sections of the turbine at different range cells for this particular wind turbine orientation. This is also the case for receiver4 where the turbine returns appear at multiple range-cells. 127

128 Figure 5.9: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver1 a) Bistatic Range-Only measurements from Receiver1. b) Received1 radar return across the range line. 128

129 Figure 5.10: 5X5 windfarm detection Wind direction = 75 a) Bistatic Range-Only measurements from Receiver1, wind direction = 75 b) Received1 radar return across the range line. 129

130 Figure 5.11: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver2 a) Bistatic Range-Only measurements from Receiver2.. b) Received2 radar return across the range line. 130

131 Figure 5.12: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver3 a) Bistatic Range-Only measurements from Receiver3 b) Received3 radar return across the range line 131

132 Figure 5.13: 5X5 windfarm detection by the bistatic pair of trasmitter1 and receiver4 a) Bistatic Range-Only measurements from Receive4. b) Received4 radar return across the range line 132

133 To illustrate the high complexity of target detection, measurements association and Ghost targets elimination for a 5x5 wind farm. The previous plots of the four bistatic pairs were combined in figures 5.14 and 5.15 for transmitter1 and trasmitter2 respectively. Target detection requires four associated measurements between all bistatic pair to find the intersection area among the elliptical curves. Conventionally, searching for associated measurements for solution involves evaluating every possible combination of measurements from each bistatic pair. This detection approach is feasible for low number of targets. However, the number of possible combinations will increase exponentially with the increase in number of targets and/or number of sensors. As shown earlier, each wind turbine from the 5x5 wind farm can be detected as multiple targets. This can translate into very large number of targets to detect within a dense area with multiple measurements. Thereby, the conventional search throughout all combinations of associations is not feasible and a robust statistical or analytical approach such as PF or IA is needed to solve this challenge. For instance, the IA approach computes all possible measurements solutions for each bistatic pair individually and then find the intersection values. Therefore, this association approach handles all the registered measurements by a bistatic pair as one group to associate, rather than the computationally costly process of searching each individual measurement for association. The possibility of ghost detection increases with the increase in measurements, especially for the case of large number of targets within dense area such as the current wind farm detection. Figures 5.14 and 5.15 illustrate the high possibility of ghost detection within the wind farm area where all measurements from the four bistatic pairs pass through. Thereby, ghost targets can appear within the wind farm because of the possibility of random intersection between four measurements not at a turbine location is high. Ghost elimination was achieved by utilising the MTRA approach. The two multistatic radar systems will detect multiple ghost targets separately. However, the matched final detection results shows less ghost detection. This result will be further discussed in section 5.7 with results comparison between one transmitter and two transmitters detections based on IA approach. 133

134 The next sections of this chapter will show the detection results of the current windfarm senior based on PF first then IA, and finally a compression between the two approaches Figure 5.14: Trasmitter1 multistatic radar measurements of 5x5 Windfarm radar return. Figure 5.15: Trasmitter2 multistatic radar measurements of 5x5 Windfarm radar return. 134

135 5.6 PF Approach Particles Placement Approach The initial step of PF is particle generation process where particles are spread randomly to cover the detection area. The higher number of particles provides better detection and higher accuracy. As large number of particle gives higher probability of particle placement near a real target position, and higher numbers of particles can be associated with that target during later steps of weights association and sampling of the PF. However, there are trade-offs between number of particles and tracking performance. The ideal implementation of PF is to minimize the number of generated particles to adequately satisfy the performance of target positioning and tracking. As such, the lower number of particles delivers faster algorithm performance. Within this research, the surveillance area is large (24km X 24km) and the wind farm covers vast area with complex wind turbine structures in addition to multiple targets around and within the wind farm. This complexity across a large area requires very large number of particles to randomly cover all the area of interest. This can slowdown the simulation performance and cause inaccurate targets detection. To overcome this limitation, the key is to strategically place particles around the recorded measurements instead of the random placement. That is, segments from surveillance area with no predicted radar return will be excluded from particle placement. To achieve the strategic placement of particles, the first step of particle generation was replaced by an iterative process whereby each step includes particle generation into grid formation. Then, weight association is performed for each grid cell according to range measurements. After that, the grid cell elimination is based on weight threshold while the remaining cells are bisected into smaller grid for the next iteration. The iterative process stops when the grid cell size is equal to the largest bistatic range resolution from the multistatic radar system (to account for the overall radar resolution). The weight association step was based on the same concept as PF step of weight updates. For each particle the weight is updated based on the likelihood function of that particle measurement matching a true measurement. 135

136 The threshold level for one bistatic pair is set according to the lowest likelihood of a particle range measurement (M p ) to match a true range measurement (M bi ) with a margin of error (±ε) associated to (i.e. M p >= M bi + ε or M p >= M bi ε ). In other words, the particle is located at the boundary of the true measurement s margin of error. The multiplication of the weight thresholds of every bistatic pair presents the overall threshold of the multistatic radar system. The 5x5 wind farm detection scenario was considered to illustrate the devised approach of particle placement. Figure 5.16.a shows the captured range measurements for each bistatic pair in blue elliptical curves, and the initial grid cells placement across (12km, 12km) detection area. Each cell of (1km x 1km) has a generated particle placed at the centre shown in red dot. The next step of updating the associated weight values of all grid cells and nominating grid cells with weights higher than the threshold is shown in figure 5.16.b. Noticeably, the remaining cells are surrounding the area where multiple measurements pass through. On the other hand, the top-left and bottom-right grid cells were eliminated as their associated weight was under the set threshold. For Example, the grid cell of (1km, 12km) is only associated with receiver4 measurements therefore the product of weight update for the rest of receives measurements will result into lower overall weight association, which was below the threshold. Figure 5.16: First iteration of particle placement process. a) Grid generation step,(1km x 1 km) grid cell size. b) Cells elimination step 136

137 Figure 5.17.a shows the next iteration of further dividing each grid cell into four cells with generated particle at the centre of each new cell. Then figure 5.17.b shows the weight updates and cells elimination step as further nominated grid cells are selected around the area where it is the densest with measurements curves. Figure 5.17: Second iteration of particle placement process. a) Grid generation step(500m x 500m) grid cell size. b) Cells elimination step The above two steps are iterated multiple times until reaching the grid cell dimensions that is equal to the overall system resolution of detection as illustrated in the following figures 5.18, 5.19 and Figure 5.18: Third iteration of particle placement process. 137

138 Figure 5.19: Fourth iteration of particle placement process. Figure 5.20: Fifth and final iteration of particle placement process. The final step of particle placement approach is to randomly generate a set number of particles within each remaining grid cell. For this example, number of particles was set to 5 randomly generated particles per grid cell. As shown in Figure Figure 5.21: a) Final step of particle placement process. b) randomly generated particles within the remaining grid cells. 138

139 The selective particle placement approach provides the advantage of generating smaller number of particles than the conventional approach of randomly generating particles over the whole surveillance area. The by-product of the devised approach is the multiple iterations of particle generation before resuming the remaining steps of PF. However, the advantage of reducing number of particles for the update and sampling steps surpasses the aforementioned disadvantage according to the overall performance for such complex multiple targets detection PF Detection Result of the 5x5 Wind Farm The previous section explains the process of particle placement. This process was used within the overall PF approach to detect a 5x5 wind farm. The result in figure 5.22 shows a screen capture from the simulation tool of PF over the proposed wind farm. The red dots represent the wind turbines structures and the white dots represent particles after three time steps of weight update and sampling. The yellow dot shown in the enlarged section of figure 5.22 represents the nominated particle as a potential target within a range cell. These nominated particles were based on the maximum weight value among all particles within that range cell. According to this result, a potential challenges and limitations of PF method can be noted when operating in such complex detection scenario. The first challenge is to do with particle placement and sampling steps while the second challenge is the limitations of particle nomination method when applying the MTRA technique. Figure 5.22: PF detection of a 5x5 Wind Farm 139

140 140