The Application Of The Improved Particle Filter Algorithm On Fault Diagnosis Of Nonlinear System

With the rapid development of computer technology,artificial intelligence and information fusion technology, the fault diagnosis method based on filter has become one of the key technology in the field of hybrid intelligent fault diagnosis in recent years, in which the performance of filtering algorithm is directly related to the sensitivity of the fault diagnosis method and directly affect the diagnosis accuracy. In recent years, the fault diagnosis technology of linear dynamic system has developed very well, but the actual dynamic systems usually have nonlinear features to some degree, some are even strongly nonlinear system. For this reason, the deep study of nonlinear filtering algorithm is the key task for the fault diagnosis method based on the filter. Particle filter algorithm, as the most effective algorithm to solve such problems, is gradually becoming the mainstream method of studying the nonlinear system fault diagnosis based on filter method. For there are still some shortcomings in current particle filter method, the particle filter method is analysed and studied comprehensively in this paper, and two kinds of improved particle filter algorithm is put forward, and the improved particle filter algorithm is applied to the nonlinear systems fault diagnosis. The main work in this paper is as follows:(1) For the standard particle filter is unable to make full use of the latest measured values,with the problem of weight degradation, the estimation precision is not high. For this reason, the reasonable proposal distribution design and resampling techniques are studied in this paper. Firstly, choose Unscented Kalman Filter(UKF) to generate proposal distribution function which is more close to real posterior distribution. Secondly, use the invasive weeds bionics algorithm to optimize the particle set during the resampling process, and use new weights jitter diffusion modes to guide the particle’s dynamic development in the process of optimizing the particle set, which accelerate the particles movement toward the high likelihood domain and improve the estimation precision.(2) For the problem that system state estimation accuracy is not high when noise properties are unknown, UT transform and h-infinity filter method is introduced to overcome the bondage of application environment. Generate proposal distribution with the use of UT transform and h-infinity filtering firstly, and then choose the Sigma points flexibly using Gaussian distribution, and use the invasive weeds bionics algorithm with weights jitter to optimize the particle set, which improve the estimation precision when noise properties are unknown.(3) In view of the low fault diagnosis precision of the nonlinear system when the disturbance characteristics is unknown,this paper produces proposal distribution function with the use of UT transformation and H∞ filter selected by Gaussian Sigma point, and then in the improved particle filter framework, combined with the residual smoothing detection method,it validates two groups of nonlinear system model chosen by this thesis —the low dimensional single variable nonlinear system and the three-tank DTS200 system.