System-level Calibration for Fusion-based Wireless Sensor Networks

1. System-level Calibration for Fusion-based Wireless Sensor Networks Rui Tan1Guoliang Xing1Zhaohui Yuan2 Xue Liu3Jianguo Yao4 1 Michigan State University, USA 2 Wuhan University, P.R. China 3 University of Nebraska-Lincoln, USA 4 Northwestern Polytechnical University, P.R. China

2. Surveillance Sensor Networks • Large-scale, dense sensing • Limited resources (power, computation…) • Uncertain data quality • Hardware biases, random noises, complex deployment terrain SensIT @ UWisc 29 Palms, CA, [Duarte 2004] VilgilNet @ UVa http://www.cs.virginia.edu/wsn/vigilnet

3. How can we ensure the fidelity of sensor data?

4. A Case Study up to 100% bias Vehicle detection experiment [Duarte 2004] • Sensing performance variation • Same sensor at different time and across different sensors

5. Existing Solutions • Collaborative signal processing, e.g., data fusion • Can mitigate impact of noise • Assume identical sensing model: [Clouqueur04TOC], [Sheng03IPSN], [Niu04FUSION], [Wang07IPSN], [Li04ICPPW], .... • Device-levelcalibration • Intractable for large-scale networks • Can’t handle post-deployment factors • System-level calibration • Centralized solutions, many samples and transmissions • Assume simplistic signal processing algorithms

6. A 2-Tier Calibration Approach reading reading • Local sensing model generation • Only need limited groundtruthinfomation (e.g.,positions) • System-level calibration • Only need few local sensing model parameters distance distance local model parameters local sensing model system sensing model cluster head

7. Agenda • Motivation • Preliminaries • Sensor measurement model • Data fusion model • 2-Tier Calibration Approach • Evaluation • Conclusion

8. Sensor Measurement Model • Reading of sensor i: yi = si + ni • Signal decay • Gaussian noise • S: target source energy • di: distance from target • ri: reference distance • ki: decay factor Acoustic noise [Duarte 2004]

9. Linear Calibration • Calibrated reading = ai yi • ai: calibration coefficient • Goal:system-level sensing model S’: common source energy r: common reference distance k: common decay factor

10. Data Fusion Model • System detection • Detection decision is probabilistic • False alarm rate • Detection probability

11. Agenda • Motivation • Preliminaries • 2-Tier Calibration Approach • Local sensing model generation • Optimal system-level calibration • Evaluation • Conclusion

12. Local Model Generation • Estimate noise mean μi and variance σi • Estimate decay model via linearization • Estimate si from multiple samples Linear fitting on real data } Q(x): Q-function of normal dist. unknown target energy local model: {μi, σi, ki, bi}

13. System-level Calibration Problem Given local models {μi, σi, ki, bi | i=1, …, N}, find1) common signal decay model {S’, k, r}2) calibration coefficients {a1, a2, …, aN} maximize detection prob. s.t. false alarm rate ≤ α Sensing performance of calibrated network is optimized!

14. Opt System-level Calibration • Unconstrained optimization • Efficient solution

15. Light Spot Detection bias • No-calibration approach • Device-level approach • Sensor w/ highest light-distance curve as ground truth ~45% TelosB Detect light spot on LCD

16. Trace-driven Simulations • Data traces collected from 23 acoustic sensors in vehicle detection experiment [Duarte 2004] ~50% improvement 10-fold reduction

17. Conclusions • 2-Tier system calibration approach • Local model generation: low comput. overhead • System-level calibration: low comm. overhead • Optimal system-level calibration • Maximize system detection performance • Extensive evaluation

18. Thank you!

19. Light Spot Detection more effective in low SNR cases