| In recent years,with the promotion of emerging intelligent sensing technology,multi-sensor networked systems(MSNSs)with intelligence,distribution and informatization have been developed rapidly.In the data processing of MSNSs,the traditional centralized data processing technology is not suitable for real-time data processing.Distributed data processing technology is widely used in target tracking,environmental monitoring,intelligent agriculture,medical care and other fields due to its reliability,scalability,flexibility and rapidity.However,there are many uncertain factors in MSNSs,such as limited sensor power,noise interferences,data communication delays and losses,fading measurements and nonlinear dynamics.To reduce the impact of these factors on the state estimation performance of MSNSs,considering the difference in estimation accuracy of different sensor nodes in MSNSs,this paper proposes a distributed filter structure with different consensus gains under the framework of Kalman filter,and then in linear unbiased minimum variance criterion,the distributed estimation problems of multi-sensor networked linear and nonlinear discrete stochastic systems are studied respectively.The main research contents are as follows:1.For the MSNSs with random sensor activation and channel noises,each sensor in the network uses the random activation method to measurement the target state,so as to achieve the purpose of saving sensor power,and meanwhile the data communication among sensors may suffer from communication noise interferences.Considering the above factors,under the distributed filter structure with different consensus gains,the optimal distributed filter algorithm in linear unbiased minimum variance criterion is designed,and the recursive equation of the cross-covariance matrices among different sensors is derived.Besides,to avoid the calculation of the cross-covariance matrices among sensors,a suboptimal distributed filter based on a set of free scalar parameters is proposed by minimizing an upper bound of the local filtering error covariance matrix.By further minimizing the upper bound of the filtering error covariance matrix of the suboptimal filter,a set of coupled nonlinear equations satisfied by the optimal scalar parameters and the optimal gain matrices are obtained,whose approximate optimal numerical solutions can be obtained by using the nonlinear optimization method.Under the condition that the whole network is detectable and stable,the stability and steady-state characteristics of the above two filters are analyzed.The gain and covariance matrices of the proposed filters can be calculated offline,which is convenient for real-time applications.2.For the MSNSs with random missing measurements and random missing estimates of neighbor nodes,when sensor knows whether data is missing or not,the online and offline distributed filters dependent on whether data is missing or not are proposed respectively,where the difference between the above two filters is that the gain and covariance matrices of the latter filter depend on probabilities of missing data and have steady-state characteristics.When sensor does not know whether data is missing or not,a distributed filter that completely depends on probabilities of missing data is proposed,i.e.,the filter and its gain and covariance matrices are all depend on probabilities of missing data.The stability of the above three distributed filters is proved under the condition of weak conservation.Finally,in terms of estimation accuracy and calculation amount,the proposed distributed estimation algorithms are compared and analyzed.3.For the MSNSs with random communication delays and packet dropouts,firstly,a random sensor activation mechanism is adopted to save the energy consumption of sensors and the prediction compensation strategy is used to compensate the data of time delays and losses.A distributed filter with minimizing the upper bound of the filtering error covariance matrix is proposed,the optimal multi-gain matrices and a set of optimal scalar parameters are obtained,the boundedness of the minimum upper bound of the filtering error covariance matrix is analyzed.Besides,considering the problem of fading measurements of sensors,a distributed filter with fading measurements,random communication delays and packet dropouts is proposed,and the boundedness of the minimum upper bound of its error covariance matrix is proved.Due to the existence of random delays,the proposed filter gain matrices require real-time calculation,which have a large online calculation burden.To avoid real-time calculation of gain matrices,using the steady-state values obtained by the distributed steady-state filter with a largest time delay,a distributed filter that avoids online optimization of scalar parameters and a distributed filter that avoids online calculation of gain matrices and scalar parameters are proposed respectively,which have reduce online calculation burden.4.For multi-sensor networked discrete stochastic nonlinear systems,considering the high-order infinitesimal term in the Taylor series expansion,a distributed extended Kalman filter is proposed to avoid computing the cross-covariance matrices.By minimizing the upper bound of the filtering error covariance matrix and nonlinear optimization method,the optimal multi-gain matrices with a set of optimal scalar parameters are obtained,and then the minimum upper bound of the filtering error covariance matrix is obtained.It is proved that the mean square exponential boundedness of the filtering error is bounded.In addition,a distributed extended Kalman filter is designed for multi-sensor networked nonlinear systems with random sensor activation,random communication delays and packet dropouts.Finally,an experimental platform of two-dimensional target tracking and positioning system is built,and the proposed distributed extended Kalman filtering algorithms are applied to the experimental platform.The experimental results verify the effectiveness of the proposed distributed estimation algorithms in practical scenario applications. |
