Windfarm Interference Mitigation using Compressive Sensing Techniques and Tilted-Wire Scatterer Model

Transcription

1 Windfarm Interference Mitigation using Compressive Sensing Techniques and Tilted-Wire Scatterer Model Ross Kyprianou, Krishna Venkataraman, Rocco Melino, Paul Berry and Hai-Tan Tran Defence Science and Technology Group, Edinburgh SA 5111, Australia. Abstract Surveillance and tracking radars, operating in the presence of high static and the dynamic clutter generated by the rotating blades of wind turbines in windfarms, have a huge task of mitigating such interference without compromising genuine target return signals. This has been of much interest in the radar research community in recent years. This poses a significant threat to the air traffic control operations and related safety aspects of civil and military aircraft. As part of an ongoing development of mitigation techniques worldwide against this interference we present a novel technique based on compressive sensing signal processing and the use of a tilted-wire scatterer model in the formation of a dictionary. We demonstrate with real data that mitigation of windfarm interference can be achieved without compromising components from targets of interest. I. I NTRODUCTION The renewable energy sector has grown strongly [1] and steadily in recent times, and specifically wind energy has seen widespread adoption. This has led to increased installation of wind farms and a continuing increase in the size of the turbines. While it is a boost for renewable energy, wind turbines have been potentially impacting the detection performance of surveillance radars, and Air Traffic Control (ATC) radars in particular. There is a significant possibility that wind turbine structures, if positioned with certain geometries and distances relative to ATC radars, may interfere with the ability of these surveillance radars to detect aircraft of interest, with adverse implications for safety-of-life and national security. This is further exacerbated with future proposed turbine heights reaching 700 ft above ground level. Possible mechanisms [2] for increased interference, contributing to further target loss, include electromagnetic shadowing effects, large clutter returns due to complex scattering characteristics of the rotating blades, tower and nacelle. Doppler processing, such as for Moving Target Indicator (MTI), of the turbine blade returns may result in the occurrence of moving clutter, ghost or false target detections in pulse-doppler radars. They could also mask genuine target detections, initiate false target tracks, or break genuine tracks. Moving target detection performance of ground-based and airborne surveillance radars may also be affected when in the vicinity of a wind farm, which are mitigated using clutter map CFAR, multi beam transmission, and other mitigating techniques. The mitigation strategies are mostly application oriented. In recent times, significant research has gone into characterising the signals scattered by the wind turbines and determining Fig. 1: A portion of the PPI display showing aircraft and wind farm interference and other returns (Extracted from BERR report: URN No. 08/618) the impact of these reflected signals on the detection ability of these radars to detect and track aircraft flying in the vicinity of wind farms. Additionally, there has been significant research are underway into developing various mitigation techniques in the radar receiver to suppress the windfarm returns while preserving aircraft target components. Some of these techniques were also based on the design and site installation of the wind turbines, but others include development of specific techniques for filtering these interfering signals at various detection stages [3] in the signal processing chain of the radar receiver. In this paper, we discuss the development of a novel technique based on compressive sensing signal processing and the use of the tilted-wire scatterer model in the formation of a dictionary [4] [5]. The proposed technique is demonstrated using real measured wind farm data with injected targets. II. BACKGROUND Radars that are mostly affected by wind farm interference are air traffic control radar systems, including both search and track functions for aircraft. When an aircraft flies directly over or in the vicinity of the wind farm, echoes from both aircraft and wind turbines are simultaneously detected, processed and displayed, which may result in serious target track ambiguity concerns. Wind turbine Radar Cross Section (RCS) can vary between 100m2 and 1000m2 contributing to the returned signal SNRs ranging between 50 db and 60 db [2]. The magnitude of strong reflections from wind turbines can cause radar receivers to be driven to saturation and raise receiver noise levels, resulting in the reduction of the probability of detection (Pd ). Windfarm reflections may interact with both main lobe and side lobes

2 Fig. 2: Top and front view of the wind turbine. creating false plots and also potential loss of targets in azimuth due to side lobe interaction. A portion of the PPI display shown in Fig.1 illustrates the complexity when aircraft flying over wind farms occupy the same or neighbouring range or azimuth cells as the wind turbines. This PPI display shows the target plots from the aircraft and the wind farm cluster nearby, the colours representing the intensity of the returns (yellow being stronger). While the radar tracker is locked to the aircraft returns, there is a very good possibility of the tracks being merged with those of the wind farm returns. Under such circumstances, the radar tracker would struggle to maintain track on the aircraft leading to safety issues. To qualitatively explain the interference effects, Fig.2 depicts the top and (part of) the front view of a typical wind turbine. The signals scattered from a farm of such rotating objects are complex micro-doppler signals [6] which depend on several factors, namely the number of turbines in the farm that are visible to the radar, for each blade, the rotation angle (γ) relative to a reference axis (as indicated in Fig.2), the angle ( θ yaw ) between long axis of the nacelle and the line of sight of the radar, the rotation rate (f r ) of the blades in Hz, and the relative distance (r) of the wind turbines from the radar. The instantaneous radial velocity component directed towards the radar is given by: V radial (r, γ, θ yaw ) = 2πf r r sin γ sin θ yaw. This equation implies, for a given distance between the radar and the wind farm, the maximum Doppler (velocity) component in one rotation occurs, when γ = θ yaw = π/2 i.e when the plane of rotation is orthogonal and the blade is vertical. Such an instance, referred to as a blade flash, is illustrated in the spectrogram generated from measured data in Fig.3. Fig.3 is a typical spectrogram of wind turbine data where the data is extracted from a single range bin. Here, 2.5 sec worth of data is shown to highlight the frequency response of the blades over time. Fig.4 displays the combined scattered returns of the wind farm and the aircraft flying in the vicinity of the wind farm [7]. This plot was generated from simulated data cubes of the scattered power from the wind farm using Monte Carlo simulation, using the parameters of the candidate ATC radar, wind turbine, and terrain. Note that the wind farm returns near zero Doppler represent the scattered returns from the tower and nacelle of the wind turbine. Our proposed mitigation technique for windfarm mitigation is based on the use of the tilted wire model as atoms in the effective algorithm of orthogonal matching pursuit (OMP). To illustrate the performance, returns from several realistic air targets are injected into the real data from a wind turbine, in such a way that the targets are well masked by the dynamic clutter. Injected target power densities are comparable to that of the rotating blades. We demonstrate that using this technique, the wind farm components can be cleanly separated from the rest of the signal which includes target returns, leaving the target signals available for conventional processing. For this paper, we demonstrate the technique for a single range bin assuming that the aircraft and wind turbine occupy this range bin during the measurement. For ATC, the Coherent Processing Interval (CPI) is in the order of ms, which means that at least one blade flash may occur during the capture time. The wind farm data used in our studies and experimentation correspond to the real time returns from an S-band radiator directed towards the wind farm and the Doppler contents of the signal vary depending upon various operating conditions such as the plane of rotation (θ yaw ), the angle of the blade to the horizontal plane (γ) and

3 Fig. 4: Combined Range - Doppler plot of windfarm and aircraft Fig. 3: Spectrogram of rotating blades from a single wind turbine the angular velocity of the blade (centre to the tip). Similarly, the target Doppler depends on the radial speed of the aircraft along the line-of-sight to the candidate radar. Following the verification of the techniques, the model is proposed to be validated using the acquired real time data with the aircraft flying in the desired locations in the vicinity of the candidate wind farm. III. SIGNAL MODEL A closed-form solution, derived in [8] for the backscattered signal, is reproduced for a uniform straight wire oriented arbitrarily in 3D space and rotating as a rigid body around some fixed axis. The wire has length L, a constant rotational velocity Ω and is considered to be a continuous distribution of point scatterers of uniform reflectivity along its length. It is assumed that the transmit signal is narrowband with wavelength λ. g(t; r c, ŵ, L) =σl exp{iω(t 2R0 c )} exp { 2iωr c c cos γ cos(ωt + ψ c θ) } sinc{ ωl c (sin γ sin φ w + cos γ cos φ w cos(ωt + ψ c + α θ))}. (1) Fig. 5 illustrates the system geometry, where θ = π/2 (without loss of generality, we can set the y-direction such that the radar is in the zy plane). In the figure, the tilted wire is denoted by AB and the projection of AB onto the xy plane is CD. Equation (1) is the general solution for the return from an arbitrarily-oriented wire in 3D space in the radar s far field. There are 6 geometric parameters of interest, namely the radar look angle γ, the position of the centre of the wire (r c, ψ c ) in polar coordinates, the tilt of the wire in elevation and azimuth, φ w and α, respectively, and the length of the wire L. This solution will be used to model the radar scattering from the wind turbines blades. In the special case of the wire being within the plane of rotation (refer to Fig. 6), φ w = 0, therefore Equation (1) simplifies to g(t; r c, ψ c, α, L) =σl sinc { } ωl c sin(ωt + ψ c + α) exp{ 2iωr c c sin(ωt + ψ c )}, (2) with L = L cos γ and r c = r c cos γ, ignoring both the constant phase shift term and the high frequency term by assuming a frequency shift to baseband [9]. IV. ANALYSIS The Orthogonal Matched Pursuit (OMP) technique [9], [10] is applied to the scattered signals in order to separate them into two components: the modulated returns from the blades and the remnant returns. The remnant is expected to include any targets of interest, noting that the targets may possibly be obscured by the blade returns. The component comprising the blade scattering s(t), can be approximated as the return from a sum of tilted wires of various radii, lengths and angles as described in Section III s(t) = K c k g(t; r k, ψ k, α k, L k ), (3) k=1 where K is the number of tilted-wire atoms and r k, ψ k, α k and L k are the four parameters corresponding to each atom. This model of the target as a relatively small number of scattering elements conforms to a sparse signal problem [10] [12]. The aim of OMP, which belongs to the class of greedy techniques, is to decompose the signal s(t) into its component atoms. The technique proceeds by (i) constructing a uniform grid on the 4-dimensional parameter space so that each grid point corresponds to an atom; (ii) identifying the atom whose signal response correlates most strongly with the remnant of the measured signal; (iii) applying least squares minimisation to estimate the complex amplitudes of all of the previously identified atoms so as to minimise the remnant signal; (iv) re-compute the remnant signal with the newlyidentified atom and repeat until an error threshold has been reached. The 4-dimensional grid of points of representative values for (r k, ψ k, α k, L k ) was created for the wind farm application as follows. Specifically, r k were 51 equidistant points ranging between 1 m to 31 m; α k were the specific values of -2, 0 and 2 degrees; and the L k were 21 equidistant points between 5 m

4 Fig. 5: 3D geometry of radar and tilted wire in the (î, ĵ, ˆk) coordinate system. and 32 m. These values were used to create the dictionary of atoms required by the OMP algorithm. V. EXPERIMENTAL SETUP Our aim is to demonstrate that by using the proposed algorithm, interference from wind turbines can be efficiently mitigated without compromising the information content of targets of interest. The real data was collected on 19 May 2012 using the Defence Science Technology (DST) Group Wandana II pulse- Doppler radar, measuring various targets at the Starfish Windfarm at Cape Jervis, South Australia. The measurements pertain to the main beam of the radar illuminating a single turbine. The wind turbine supports three blades of length 32 m, rotating at 17 RPM. The radar was positioned in the plane of rotation so that the 2D case of the model (described above) was applicable. Also, the radar operated at a carrier frequency of 3.1 GHz (Sband) using a PRF of 4 khz with horizontal polarisation on transmit and receive. A single sample per pulse was collected at the range bin where the turbine was located. Since real aircraft targets were absent during the experiment, three synthetic targets were added to the real wind turbine data, in the time domain, to produce a composite signal (Fig. 8). Target 1, a chirp starting at 180 Hz simulates a target accelerating towards the radar. The remaining 2 targets simulate targets approaching the radar with constant velocity where Targets 2 and 3 are at 360 Hz and 900 Hz respectively. To test the ability of the proposed algorithm to discern between blade returns and target returns, the targets were created with relatively lower amplitude than the blades and three of the targets were intentionally placed in frequency bins with the greatest blade response. VI. RESULTS We applied the proposed technique to the data as described in Section V. A high pass filter with cut-off frequency of Fig. 6: Tilted wire geometry in 2D. Radar receiver is located in the positive y-direction. 230 Hz was applied to the data to filter out the low-frequency terms corresponding to the tower, nacelle and rotating nose cone. The CPI was set to 0.6 sec to capture a single blade return and the processing was applied to this truncated dataset. The aim was to extract from the original signal the tilted-wirelike components that are returned from the blades, leaving a remnant signal which includes most of the energy from targets. The results of the processing are presented in Fig.7 9. Fig.7 is the spectrum of the original unfiltered data showing the returns from the wind turbine tower, nacelle, turbine blades and land clutter. Superimposed on this are three synthetic targets (located in Doppler bins 180, 360 and 900 Hz). The two constant velocity targets appear as spikes and a third, corresponding to an approaching accelerating target, appears as a chirp. A composite signal is generated by adding the three targets to the original signal with the spectrum shown in Fig.8. After OMP processing, the composite signal is separated into two components: the extracted signal comprising the returns from the blades and the remnant signal (Fig.9). The spectrograms of the composite, extracted and remnant signals are shown in Fig.10. The top spectrogram (based on the short-time Fourier transform) shows the composite signal (original wind turbine data with the three synthetic targets Fig. 7: Spectrum of the wind turbine: tower, nacelle and blade with the three targets superimposed.

5 Fig. 8: Spectrum of the composite signal: wind turbine return with targets added. A narrow notch filter around 0 Hz has been applied. added). The centre plot shows the remnant signal containing the targets at their expected Doppler bins. The bottom is the spectrogram of the extracted signal showing the blade flash. Fig. 11 and 12 show the time-domain and spectra of the composite and extracted signal respectively. The extracted blade flash response in the time-domain closely matches the original time data. In the spectrum, the blade frequency components are extracted, with target returns not appearing in this plot. Fig.13 shows the spectra of the composite and remnant signal. The spectrum shows the remaining energy after extraction highlighting the targets of interest. Note also that clutter and remnant blade components contribute to this spectrum. Closer inspection of the target returns in the remnant signal shows that the two constant Doppler targets are not extracted by the OMP processing. They appear as peaks in the spectrum. However, Fig.14 shows the remnant signal spectrum zoomed around the Doppler region occupied by Target 1. The spectrum of Target 1 only is superimposed as a reference. The induced chirped response of the remnant signal is degraded compared to the original signal response. On inspection of the remnant spectrogram shown in Fig.10(b), the slight degradation can be seen, where frequency components between the times of 0.2 and 0.4 sec have been extracted as part of the OMP processing. Hence these terms which contribute to the chirp are missing resulting in a rabbit ear like response for Target 1, possibly degrading detection performance. It must be noted that the location, in Doppler, of this Fig. 10: Spectrograms of the a) composite signal before OMP processing, b) remnant signal, and c) extracted blade. synthetic target has been deliberately placed to yield a worst case scenario, where the target Doppler occupies the same region as the high intensity blade response. Conventional processing such as thresholding can then be used to detect the targets. VII. DISCUSSION AND CONCLUDING REMARKS The key factors that enable the excellent performance of the proposed technique are: 1) the signal model of the tilted wire elements, which provide sufficient flexibility to describe the scattering behaviour of a turbine blade, and 2) the capability of the OMP technique to obtain an efficient sparse representation of a signal scattered from such a rotating objects. Fig. 9: Spectrum of the extracted signal showing the turbine blade response and remnant signal showing clearly the three target returns. Fig. 11: Time domain plot of composite signal and extracted blade signal.

6 chirped components of those targets; however, the degrading effects are small, and we will characterise these cases in more detail, and discuss applications to more practical scenarios, in a future publication. Fig. 12: Spectrum of composite signal and extracted blade signal. Fig. 13: Spectrum of composite signal and remnant signal. Fig. 14: Zoomed in spectrum of Target 1 pre-omp processing and the remnant signal showing degraded Target 1 response. REFERENCES [1] L. Norin and G. Haase, Doppler weather radars and wind turbines, in Doppler Radar Observations-Weather Radar, Wind Profiler, Ionospheric Radar, and Other Advanced Applications. InTech, [2] J. J. Lemmon, J. E. Carroll, F. H. Sanders, and D. Turner, Assessment of the effects of wind turbines on air traffic control radars, National Telecommunications and Information Administration Technical Report, no , [3] L. Sergey, O. Hubbard, Z. Ding, H. Ghadaki, J. Wang, and T. Ponsford, Advanced mitigating techniques to remove the effects of wind turbines and wind farms on primary surveillance radars, in Radar Conference, RADAR 08. IEEE. IEEE, 2008, pp [4] F. Nai, R. D. Palmer, and S. M. Torres, Range-doppler domain signal processing to mitigate wind turbine clutter, in Radar Conference (RADAR), 2011 IEEE. IEEE, 2011, pp [5] S. Allabakash, P. Yasodha, S. V. Reddy, and P. Srinivasulu, Wavelet transform-based methods for removal of ground clutter and denoising the radar wind profiler data, IET Signal Processing, vol. 9, no. 5, pp , [6] R. Wu, J. Mao, X. Wang, and Q. Jia, Simulation of wind farm radar echoes with high fidelity, in Signal Processing (ICSP), 2012 IEEE 11th International Conference on, vol. 3. IEEE, 2012, pp [7] R. R. Ohs, G. J. Skidmore, and G. Bedrosian, Modeling the effects of wind turbines on radar returns, in military communications conference, 2010-milcom IEEE, 2010, pp [8] S. Kodituwakku, R. Melino, P. E. Berry, and H.-T. Tran, Tilted-wire scatterer model for narrowband radar imaging of rotating blades, IET Radar, Sonar & Navigation, vol. 11, no. 4, pp , [9] H.-T. Tran, R. Melino, P. E. Berry, and D. Yau, Microwave radar imaging of rotating blades, in Radar (Radar), 2013 International Conference on. IEEE, 2013, pp [10] J. A. Tropp and A. C. Gilbert, Signal recovery from random measurements via orthogonal matching pursuit, IEEE Transactions on information theory, vol. 53, no. 12, pp , [11] D. Needell and R. Vershynin, Signal recovery from incomplete and inaccurate measurements via regularized orthogonal matching pursuit, IEEE Journal of selected topics in signal processing, vol. 4, no. 2, pp , [12] M. Elad, From exact to approximate solutions, in Sparse and Redundant Representations. Springer, 2010, pp We wish to emphasise that this preliminary work is a concept demonstration, that the novel tilted-wire model combined with the standard technique of OMP has a great potential as signal processing method to mitigate wind farm interference in pulse-doppler air traffic radars especially when the radial velocities of aircraft targets fall totally inside the Doppler extents of the dynamic clutter caused by a wind turbine. The demonstration uses CPIs and PRFs that may not be typical values found in operational ATC radars; however, our results are based on real wind farm data at the typical S-band, and similar advantages of the proposed techniques should be expected for real operational systems. We found that the technique works best with for aircraft targets with constant velocities; for accelerating and decelerating targets, the turbine blade extraction processing of the technique may extract also a small portion of the RF energy from the