Localisation in Wireless Sensor Networks

Transcription

1 Localisation in Wireless Sensor Networks Georg Gaderer, Patrick Loschmidt, Anetta Nagy, Reinhard Exel, Thilo Sauter Institute for Integrated Sensor Systems Austrian Academy of Sciences Wiener Neustadt, Austria {Georg.Gaderer, Patrick.Loschmidt, Anetta.Nagy, Reinhard.Exel, Abstract The traditional task of a sensor network is data collection and aggregation for further processing. In the case of mobile networks this is usually done via wireless technologies. However, with the establishment of mobility another problem occurs: The question of the actual location of the measured data. This paper proposes a localisation scheme for Wireless LAN, where an existing WLAN infrastructure can be used not only for the collection of sensor data but also for the localisation of the data source. In this investigation localisation in sensor networks is done via the Time Difference of Arrival scheme, which has the advantage that the sensor nodes can be left unmodified. I. INTRODUCTION The idea of combining multiple sensor values to gain additional precision, features or possibilities for calibration is well known and established under the term sensor fusion. This sensor fusion is, however, not limited to the combination of sensor values, it can be also done with a fusion of the sensor value and information of the network infrastructure. Doing so, mobile sensing devices can be enhanced with the actual value of the physical quantity and its exact location. This location information is something which has to be logically bound to the obtained sensor data, what can be done in several different ways. The probably most straight forward approach to obtain the location information is to add a second location measurement component on each and every sensing device. Such a component can be for example a distance measurement sensor which calculates the position via several measurements to reference points. Another very famous approach for outdoor mobile devices is GPS. This kind of system calculates the three dimensional position by observing signals from clocks which are located in satellites. As the sattelites clocks are very accurate, their postions are known, a simple GPS receiver can calculate the position. This approach has however several deficiencies, the biggest is of course the dependency on the signal reception, together with relatively expensive receivers. A second consideration in an localisation scheme is the need for a seamless integration with the sensor network. The most promising approach is to integrate this communication network with the localisation scheme. The advantage is obvious: Not only the reduced costs due to less components, also the dramatically simplified power awareness considerations are on the plus side for a fully integrated solution. This paper investigates such an approach for localisation of wireless sensors. The proposed approach is using the Time Difference of Arrival technique, which is described in the following state of the art section. After this investigation the paper describes a hardware, which has been designed for a proof of concept together with some preliminary results. II. STATE OF THE ART The localisation of mobile devices is not only a topic for sensor networks. Also sensor/actuator systems can benefit from location awareness. As for example shown in the flex WARE [1] project, localisation can be used for preparation of roaming mobile devices between access points. Thus, this section gives an overview on localisation technologies which are not necessarily bound to traditional sensor networks. A. Angle of Arrival Measurements The angle of arrival (AoA) based localisation is a technique where the location of a mobile station (MS) is determined by triangulation of bearing information collected at a number of access points (APs). In this scheme the mobile device is considered as broadcaster and the access points as sensors for the location determination. In AoA based localisation, the receivers angular separation plays an important role in determining the position of the emitter. In [2] the authors present the angular separation requirements for the APs in order to achieve the best mean square error localisation performance for arbitrary but fixed AP ranges. In this work, a characterization of optimal angles between access points and sensors for passive AoA based localisation problems is presented. According to the authors, the optimal access point placement is in general not unique and, in the case the receivers are equidistant from the MS, there may be optimal sensor configurations with non-uniform angular access points locations in addition to equiangular separation. The accuracy of AoA measurements is limited by the directivity of the antenna, by shadow fading and by multipath reflections. Since this technique relies on a direct line-of-sight path from the transmitter to the receiver, the multipath components may appear as signals from an entirely different direction and can lead to significant errors in the measurements [3]. Another technical issues according to the same source is the problem to overcome the variations in the signal strength of the transmitted signal. The receiver cannot differentiate between the signal strength variations due to the varying amplitude of the transmitted signal and the signal strength variation caused by the anisotropy in the reception pattern. As this method requires access points (APs) with directable antennas and usually the achieved positioning accuracy is lower than in the

2 other techniques, it has not been the focus of the investigation of this work. B. Received Signal Strength Technique This technique calculates the position of a localisation system based on received signal strength (RSS) measurements. These measurements are using a standard feature found in most wireless devices, the received signal strength indicator (RSSI). This approach requires no additional hardware, and is unlikely to significantly impact local power consumption, sensor size and thus cost. The signal-strength based localisation methods can be divided into manual and automatic methods. The manual group consists of training and an online phase. During the training phase, area-specific knowledge about the receivable APs and their corresponding signal strength is collected by manual walk-arounds. During the online phase, actual RSS readings are compared to the prior-knowledge by nearest neighbour search. The down side of this technique is that when the environment changes, the localisation accuracy decreases. Furthermore, systems based on radio maps can be categorized into systems using a deterministic or a probabilistic technique. In deterministic systems the signal strength of an AP at a location is represented by a scalar value (e. g. mean value) and nonprobabilistic approaches are used to estimate the location. A famous system of this category is called RADAR. The RADAR system is presented in [4] and belongs to the group of manual methods. It records and processes RSS information at different base stations in the range of the mobile node in order to have an overlapping coverage. It combines empirical measurements with signal propagation modelling using a single path model which is manually calibrated. The median resolution of the RADAR system is in the range of 2 to 3 m, when using the empirical measurements as training sequence and 4 to 5 m, when relying on the single path model only. The method used by the Horus system [5] is comparable to RADAR. It also uses signal strength measures from different involved APs, but stores information about the signal strength distributions from the APs. The location is then estimated by means of probabilistic techniques. It has been mainly designed to achieve a high accuracy and to have low computational requirements at the same time. Different wireless channel variations can be identified and addressed for high accuracy. A special location clustering results in reduced computational requirements for the algorithm. This approach achieves an average accuracy of less than 2 m. The approach in [6] employs online receiver signal strength observations with known or unknown locations, together with the original (previous) observed or estimated signal strength database to refine a radio map network environment. Online RSSI observations from client users may be compared with the original observed or estimated signal strength database and the radio map may be refined based on unsupervised training capabilities. As a summary, it can be stated that RSS based systems might be interesting for hybrid approaches with other methodologies, because none of the approaches achieves accuracies in the range below 1 m which, based on the targeted applications of sensor localisation. C. Time of Arrival (ToA) Techniques One-way propagation time measurements provide the difference between the sending time of a signal from the transmitter and the receiving time of the signal at the receiver. This localisation technique has several deficiencies: The position accuracy is usually poor under Non-Line-of- Sight (NLoS) conditions, where there are blockages between the transmitters and the receivers. Also, the clocks of APs and the nodes need to be accurately synchronised, which would imply a hardware modification of the nodes. A general problem in time-based methods (Time of Arrival (ToA), Round Trip Time (RTT) or Time Difference of Arrival (TDoA)) is to detect the leading edge of the arriving signal. The difference between the observed ToA of the first reflection path and actual ToA of direct path is called ranging error. As a mobile client roams in an indoor environment from one location to the other, the ranging error also fluctuates. Therefore, the analysis of the behaviour of the ranging error is mandatory in order to properly design indoor time-based localisation systems. 1) Roundtrip Propagation Time Technique: For localisation, roundtrip propagation time measurements have the advantage over the one-way propagation time that they do not require high-accuracy clock synchronization between the mobile devices and the AP. In this technique, the difference between the time, when a signal is sent by the AP and the time when the signal is returned by the mobile device and received at the AP, is measured. Since only one reference clock is used to compute the roundtrip propagation time, there is no synchronization problem. The major error source in roundtrip propagation time measurements is the delay required for handling the signal in the mobile station. This internal delay can be coarsely estimated by calibration or measurement, but might not be efficient for accurate localisation. One solution to improve the internal delay estimation could be done using differential, hardware timestamping in the mobile station. Still, the mobile station s oscillator directly influences the internal delay. Therefore, syntonisation of the oscillators is be advisable. Even if all these issues would be solved, it might happen that a malicious mobile station tries to report a false location by modifying the internal delay of the packet. To overcome this problem passive listeners (so-called witnesses) can help the verifiers (the entities which are determining the position) in order to locate a mobile station in an ad-hoc network. All witnesses monitor the packet exchange between the mobile node and the lead verifier and later report their respective inter-packet time measurements to the lead verifier for further processing. In [7] a positioning algorithm is presented, named Time of Arrival to time Difference of Arrival (ToAD). This algorithm derives Time Difference of Arrival (TDoA) measurements from the messages that two-way time of- rrival (basically RTT)

3 stations in sight exchange. The idea behind this method is that a mobile station is also listening to the message exchange between other mobile stations and the AP. It then calculates the time difference between the request and reply and uses this information in its own position calculation. The accuracy of the ToAD algorithm is simulated for both LoS and NLoS cases. The positions are estimated using the Gauss-Newton non-linear least squares algorithm in both ToA and ToAD methods. The results show that the new ToAD algorithm increases the root mean square error (RMSE) of the positioning error in LoS scenarios by 10 to 20 %. On the other hand, when NLoS scenarios are simulated, the RMSE of ToAD scheme is found to be at least 10 % lower than that achieved by ToA. This result is especially important since this latter scenario is the most common in practice. 2) Time Difference of Arrival Technique: The time difference of arrival (TDoA) method is based on the concept that a mobile station transmits a message, which is received by at least three APs (2D case) that are synchronized with each other. By knowing the position of the APs and the propagation delay differences, the location of the mobile station can be calculated. In [8] a localisation scheme is described which is based on TDoA for IEEE b WLAN. The authors have implemented a localisation system, which deals with more issues: AP synchronization, packet selection, leading edge detection, and diversity. The packet receive timing measurement was based on the cross-correlation of the code division multiple access (CDMA) chip code and the resolution was improved using high-rate signal sampling and interpolation. The system s performance has been experimentally evaluated in an actual storehouse. The results indicate an improvement of the localisation accuracy in the developed system of within 2.4 m, which meets the practical application requirements. The same approach is used in [9] where the position of a commercial-off-the-shelf (COTS) wireless sensor node is determined by evaluating differential signal propagation delays. The measurement is done by detecting the arrival time of a client signal at multiple APs. In this paper the IEEE 1588 clock synchronization protocol is used for synchronizing the APs. The crucial point in realizing accurate position determination in this scheme is the detection of the time instant when the packet from the node arrives at different APs. To gain a positioning accuracy below 1 m, the participating APs must be able to measure the same signal edge within a range of 3 ns assuming perfect synchronization between the APs. Additionally, the accuracy of the determined position can be improved by statistical methods. D. Selection of Localisation Schemes for Sensor Networks Summarising the state of the art, a comparison of the presented localisation techniques is needed. For this three aspects have to be considered: The accuracy and resolution of the obtained positional data, the spatial range of the localisation, and Accuracy Fig. 1. Repeat of Message, with Receive Timestamp Mobile Sensor Repeat of Message, with Receive Timestamp Principle of the Time Difference of Arrival approach Distance Measurement Fig. 2. RSS Power Awareness ToA GPS TDoA AoA Range Comparision of the different localisation approaches the ability to build power-aware nodes, which is from special interest for sensor networks due to the typical need for low power consumption. In order to give an overview, this considerations are taken as three axis in a coordinate system and the presented localisation techniques are classified as shown in figure 2. This figure has to be understood as a qualitative comparison rather than fixed numbers from existing solutions. However, it gives an idea about the capabilities of the proposed solutions. If the result of such a comparison is used together with the background of sensor networks two candidates have to be considered: Time Difference of Arrival and Time of Arrival. The first one has several advantages in terms of a lightweight sensor network. The strongest argument is the fact that for a TDoA scheme a node does not need to be modified, which is positive in terms of power consumption. The second is that the infrastructure itself can estimate the position of a data source.

4 GAIN Control DAC AD5623 HF-Transceiver with buffer amps Clock distribution/pll AD9511 AP1 Mobile Sensor AP2 FPGA for Configuration TX DAC AD9780 RX ADC AD9230 Fig. 5. Test setup for the one dimensional case Fig. 4. Implemented prototype III. LOCALISATION ENABLED RECEIVER AND FIRST RESULTS The remainder of this paper considers TDoA as the target approach for a sensor network, however the subsequently presented hardware can be in principle also used for the reverse approach ToA. A. Timestamping capable Access Point Hardware As the localisation requires highly accurate timestamps at the access points, it is mandatory to determine the arrival time of the incoming packet as soon as possible, i. e., directly after the transceiver. This timestamping is done as shown in figure 3. This approach was chosen in order to avoid re-implementation of a whole WLAN stack. The packets are uniquely identified by decoding and calculation of a checksum on the fly. This checksum can be later compared with the received data and the timestamps of other access points. This structure has been implemented in a two-tier architecture: The self-implemented board shown in figure 4 holds the mixer and a proper analogue to digital converter for receiving and, optionally, sending. Besides this basic functionality the possibility for bypassing of some signal flows exists. It is for example possible to intercept the baseband signal directly via SMA connectors to perform direct measurements with an oscilloscope. However, this is only used for this experimental platform. In the final implementation the signals from the ADC will directly lead to an FPGA board. This board is not shown in 4, as it is attached via connectors to the bottom of this PCB. For efficiency reasons a commercial FPGA development board equipped with an Altera Stratix II GX device has been used. This board also features a Ethernet connector, which is used as later on described for data-collection. B. One Dimensional Test-Case The geometric simplest case of localisation is the one dimensional one consisting of two receivers. The thorough investigation has been conducted for this case since the results are not blurred by non-linear effects (e. g. geometric dilution of precision) which are caused by the intersection of hyperbolas in the two and three dimensional case. The test setup as shown in figure 5 consists of 2 receivers and a PC calculating the time difference. The two receiving access points AP1 and AP2 are syntonized via Synchronous Ethernet [10] and forward all the frames received on the wireless channel with the corresponding information via UDP to the PC. The access point is configured to broadcast beacons every 10 ms which are received by the nodes. The PC runs a C program opening a socket for each node and listens to the incoming datagrams. Each datagram contains not only the WLAN frame, but also additional information like the timestamp, the gain of the transceiver IC, the frame CRC, and the originating MAC address. These values are put into a separate ring-buffer-like table for each node. A search routine tries to find entries in the both tables that match between the nodes (i. e. CRC are equal and sources MAC addresses match). Since there is just syntonization and no synchronization between the nodes, the offset due to unsynchronized initialization must be compensated once. This is accomplished by positioning the sensor node equally spaced between the nodes and retrieving the current offset value. For every future TDoA calculation this value is subtracted and the calibration location is said to be the reference location. The conversion from oscillator ticks including sub-symbol timing to nanoseconds is done multiplication with 1000/44/256 since every 44 MHz clock cycle is evenly fractioned into 256 slices, i. e. the timestamp resolution is ps. After the initial calibration the sensor node can be moved and the TDoA can be monitored and stored on the PC. In the one dimensional case a TDoA of 1 ns is equivalent to about 15 cm given that the speed in the media is almost c 0. The measurement clearly shows that even a small displacement of the sensor node like 20 cm can be measured. Interestingly the calculated displacement was always greater than the real sensor node dislocation. In order to reduce the relative jitter of the timestamps, the nodes were placed in a distance of 6 m. After the calibration of the system with the AP in den middle, the AP was moved towards AP1 and then towards AP 2. If the node was in the vicinity of a node, the calculated position was far outside the range between the nodes. This behaviour can be caused by two different errors:

5 Baseband Altera FIR Sqrt Raised Cosine Filter Timing Recovery Phase Recovery Barker / CCK Decoder M-DPSK Demodulator Descrambler Frame started Receiver Control Data Fig. 3. Structure of the receiver The timestamps are erroneous The scale of timestamps is wrong Based to the measured TDoA it looked like the propagation speed in the media is only around c 0 /3 which is of course contradictory to the free air propagation. The reason for the discrepancy between measurement and simulation could be localised in the transceiver IC. Its receive amplifiers delay the baseband signal depending on the amplification. The higher the gain of the amplifier, the higher the group delay of the signal. The varying group delay explains the measured TDoA values. If the node was close to AP 1 and away from AP 2, the path was virtually elongated by the transceiver IC so that the node seemed to be outside of the connecting line between the AP. C. Influence of the Transceiver Amplification on the Group Delay Measurements reveal that the transceivers on the prototype board show a non-linear, amplification dependent group delay. The magnitude of this delay is in the order of several nanoseconds and has to be numerically compensated. The correlation between group delay and selected gain was measured using the Azimuth WLAN measurement system. The setup is based on a 3 db-splitter and a variable attenuation matrix. The reference signal is fed via the splitter and an additional 20 db attenuation into the reference AP 2, the other signal is directed into the attenuation matrix and its output into the node AP1. At the start of the measurement, the offset of both counters is calibrated and the attenuation matrix is set to the initial attenuation value. Using a script the attenuation is increased by 1 db/min until the signal is too weak for correct decoding. Note that the following measurements were performed with the HF transceivers pre-amplifiers in low gain mode. The high gain mode would increase the total amplification by 30 db, but will presumably cause different group delay. For every db-value around 6000 TDoA values are recorded together with the gain setting of AP1 and AP2 and stored in a file. The recorded data is analysed using Matlab. The raw unfiltered group delay versus attenuation is shown in figure 6. With increasing attenuation the SNR drops and therefore the accuracy of the timestamps. In an over-determined localisation system (i. e. if more receivers are present than necessary), this additional information could be used to improve the position estimation. Timestamps of nodes with low SNR and therefore a high time variance could have a lower weight in the Fig. 6. Delay Distribution as a function of the attenuation Fig. 7. Deviation of the delay distribution position determination algorithm than those with high SNR. The standard deviation of the raw unfiltered timestamps is depicted in figure 7. The standard deviation up to 60 db is almost constant at about 0.4 ns. From 70 db and upwards the standard deviation increases significantly and random packet errors occur. Passing the data through a 2nd order IIR Butterworth filter in both directions (to make sure it has zero phase delay), allows to derive the mean group delay for every attenuation

6 setting. This plot can be used to derive a compensation value for each frame received. Linear interpolation can be used in the range of up to 65 db where the group delay can be compensated by ns/db. A better solution is to store supporting points for every 3 db value and perform linear interpolation for the table lookup. Moreover, it should be noted, that the compensation curve is slightly different (up to 2 ns) for every IC. For a correct compensation the curve of every transceiver has to be measured separately against a reference receiver. Furthermore, the delay is also temperature dependent, but since there is no temperature sensor in the IC or nearby, no compensation for this error source can be done. The resulting bias of the timestamps has therefore a major impact on the localisation. Whereas the variance can be reduced by averaging, the remaining offset after the compensation cannot be minimized or compensated by any means. IV. CONCLUSION AND OUTLOOK One very vital and yet still missing feature for wireless sensor networks is the localisation of the data origin. This localisation can be done by determining the actual position of a mobile node. There are several methods available to fulfil this task. One is to utilize dedicated hardware such as distance measurements or 3D scanners. This has the disadvantage that dedicated hardware modules are needed. The present work first discusses the different ways to get the position directly from the network infrastructure. In that approach mainly three different strategies crystallize out: Approaches where the node measures the time of arrival, algorithms which are based on the signal strength, and finally approaches where the node is in terms of localisation fully passive and only the access points calculate the time difference of arrival. To quantify these different methods a metric has been proposed, which considers, the localisation accuracy, scalability, and the potential impact on energy awareness. The ability to work with a limited amount of energy is of special importance for sensor networks. As a conclusion the time difference of arrival method was chosen as the most promising candidate for a sensor network, as the nodes do not need to be modified for the localisation measurements. The presented work implemented a proof of concept of an access point, which has to be modified, enabling the network to determine the position of an unmodified sensor node. The first results of an one-dimensional localisation are discussed. It is shown that the main problem is an amplification dependent group delay of the receiver which has to be compensated via calibration. of Lower Austria and the European Regional Development Fund. REFERENCES [1] G. Gaderer, P. Loschmidt, and A. Mahmood, A novel Approach for Flexible Wireless Automation in Real-Time Environments, in Proceedings of the 2008 IEEE International Workshop on Factory Communication Systems, G. Cena and F. Simonot-Lion, Eds., IEEE. IEEE, May 2008, pp [2] K. Dogancay and H. Hmam, Optimal angular sensor separation for aoa localization, Signal Processing, vol. 88, pp , [3] G. Mao, B. Fidan, and B. O. Anderson, Wireless sensor network localization techniques, Computer Networks: The International Journal of Computer and Telecommunications Networking, vol. 10, pp , [4] Mao, B. Fidan, and B. D. O. Anderson, Wireless sensor network localization techniques, Computer Networks: The International Journal of Computer and Telecommunications Networking, vol. 51, no. 10, pp , July [5] Youssef and A. Agrawala, The horus location determination system, Wireless Networks, vol. 14, no. 3, pp , [6] M. Tao and K. N. Lau, Location determination and location tracking in wireless networks, US patent US , April [7] I. Martin-Escalona and F. Barcelo-Arroyo, A New Time-Based Algorithm for Positioning Mobile Terminals in Wireless Networks, EURASIP Journal on Advances in Signal Processing, vol. 2008, p. 10, [8] R. Yamasaki, A. Ogino, T. Tamaki, T. Uta, N. Matsuzawa, and T. Kato, TDOA location system for IEEE b WLAN, in Proc. IEEE Wireless Communications and Networking Conference, vol. 4, 2005, pp Vol. 4. [9] P. Loschmidt, G. Gaderer, and T. Sauter, Location based Services for IEEE a/b/g Nodes, in Proceedings of the 6th Intl Workshop on Real-Time Networks RTN2007, Pisa / Italy, July 2007, pp [10] S. Rodrigues, IEEE-1588 and Synchronous Ethernet in Telecom, in Proc. IEEE International Symposium on Precision Clock Synchronization for Measurement, Control and Communication ISPCS 2007, 2007, pp ACKNOWLEDGMENT The authors would like to thank Miss Spreitzhofer and Miss Lisa for their marvellous support writing this paper. The work for this paper was co-founded by the FIT-IT Project ε-wifi Embedded Position Determination and Security in Wireless Fidelity Networks. Grant Number and the EU Project flex WARE under grant number as well as the province