Compressive Data Gathering for Large- Scale Wireless Sensor Networks Chong Luo Feng Wu Shanghai Jiao Tong University Microsoft Research Asia Jun Sun Chang Wen Chen Shanghai Jiao Tong University SUNY at Buffalo, NY , USA MobiCom 2009, Sep
Outline Compression techniques on sensor networks – Compression with explicit communication – Distributed source coding – Compressive Sensing(sampling) Proposed Compressive Data Gathering – Data gathering diagram – Compressive sensing Simulation Conclusions
Compression Techniques on Sensor Networks Compression with explicit communication Cristescu et al. (2006) proposed a joint entropy coding approach 12 X1X1 H(X 2 |X 1 ) X 1, H(X 2 |X 1 ) EZLMS Link:
Distributed Wavelet Transform Assumptions: piecewise smooth data – Ciancio et al. (2006) and A’cimovi’c et al. (2005) (1)Even nodes first broadcast their readings. (2)Upon receiving the readings from both sides, odd nodes compute the high pass coefficients h(·) (3)Then, odd nodes transmit h(·) back and even nodes compute the low pass coefficients l(·) (4) After the transform, nodes transmit significant coefficients to the sink
Distributed Source Coding -- Slepian-Wolf coding D. Slepian and J. K. Wolf (1973) EZLMS Link:
Compressive Sensing Measurement matrix
Compressive Sensing transform basiscoefficient
Compressive Sensing transform basiscoefficient
G. Quer et al. (2009) x 11 x 12 x 13 x 14 x 21 x 22 x 23 x 24 …….. … …… X Example of the considered multi-hop topology. Irregular network setting [4] (1)Graph wavelet (2)Diffusion wavelet Network Scenario Setting
Measurement matrix Built on routing path Routing path …………………… …… …… …… ……………………
Proposed Compressive Data Gathering -- Measurement Matrix
Goal: (1)Reduce global communication cost. (2)Load balance
Proposed Compressive Data Gathering -- Measurement Matrix
Proposed Compressive Data Gathering -- Data Recovery Conditions: (1) (2) Incoherence: correlation between and
Reconstruction: optimization Linear programming Orthogonal matching pursuit (OMP)
Recover Data with Abnormal Readings
Proposed Solution Normal reading Deviated values of abnormal readings New basis
NS-2 Simulation Topology: – Chain vs. Grid Data sparsity is assumed to be 5%. – For example, when N = 1000, K = 50, and M = 200
Capacity -- Chain topology N=1000 The distance between adjacent nodes are 10 meters
Capacity -- Grid topology N= rows x 33 cols The distance between adjacent nodes is 14 meters
Packet Loss Rate -- Grid topology
Experiments on Real Data Sets -- CTD Data from Ocean K=40M=100
Experiments on Real Data Sets -- CTD Data from Ocean
Experiments on Real Data Sets -- Temperature in Data Center
Low spatial correlation : not sparse
Experiments on Real Data Sets -- Temperature in Data Center Sort d i in ascending order according to their sensing values at a particular moment t 0 – The resulting readings are piece-wise smooth. – server temperatures do not change violently, sensor readings collected within a relatively short time period can also be regard as piece-wise smooth if organized in the same order. N=498
Experiments on Real Data Sets -- Temperature in Data Center
Conclusions This paper proposed a novel scheme for energy efficient data gathering in large scale wireless sensor networks based on compressive sampling theory. – Convert compress-then-transmit process into compress-with-transmission process We have shown that CDG can achieve a capacity gain of N/M over baseline transmission.