Research Of Data Blind Calibration Method In Wireless Sensor Networks

In the large-scale deployment of wireless sensor networks,the goundtruth data is usually not available to be the reference of calibrating,so the traditional calibration methods are difficult to implement.Blind calibration method can calibrate each sensor in the network only based on the collected observation data.But most of the existing blind calibration methods assume that the observed signals of the sensor are time-invariant and part of them have the problem of slow convergence.Aiming at the scenes of multiple time-varying observation signals exsiting in wireless sensor networks,a blind calibration method based on hybrid state space model is proposed to improve the convergence speed of the calibration method and the accuracy of estimating sensor calibration parameters.The main work of this paper is as follows:1.Aiming at the scenes of multiple time-varying observation signals in wireless sensor networks,the sensor data blind calibration problem is modeled as a hybrid state space model with both continuous variables and discrete variables exsiting,and combining the parameter S_n?t? and Z_n?t? in the model and the Dirichlet process for achieving sensor clustering;2.Aiming at the nonlinear observation signal parameter ?_k?1:T? in the hybrid state space model,a unscented transform-forward and backward algorithm is proposed.The algorithm can reduce the characteristics of the nonlinear function as far as possible by using the idea of unscented transform,and the posterior distribution of the signal state ?_k?t? at time t is more accurately estimated from the observations before and after time t in the time series.So this algorithm can improve the accuracy of the calibration parameter estimation;3.The Markov chain Monte Carlo sampling algorithm is improved in this paper by adding a hidden variable which determines the direction of next sampling step in the posterior distribution of parameters.In this way,we can improve the convergence speed of the sampling algorithm while ensuring the accuracy of parameter estimation.After sampling a set of samples,the algorithm discards the inaccurate samples sampled before reaching the stable distribution according to the proprety of the Markov chain:the distribution always become stable.So the accuracy of the calibration parameter estimation will be further improved.In this paper,the four data blind calibration methods of wireless sensor networks including the proposed method are simulated and analyzed first,and then verified in the real dataset provided by IBRL.The results show that compared with the best performance in the comparison methods:in the average square error value of sensor gain and offset estimation,the proposed method is reduced by 1.742×10^-3and 0.1729 respectively in the simulation analysis,and reduced by 2.193×10^-3 and 0.1145 respectively in the real dataset verification;in the convergence speed of calibration methods,the proposed method in this paper improved by 1.7792s?about 47%?in the simulation analysis,and 0.7609s?about 33%?in the real dataset verification.