The Research On Fault Diagnosis Based On Particle Filter And CUDA

At present,the fault diagnosis of linear dynamic system has formed a more mature theory and method.However,the actual system has more or less nonlinear characteristics,and the study of fault diagnosis for nonlinear system is of great significance.As a nonlinear,non-Gaussian filtering method based on Monte Carlo idea,the particle filter algorithm is made into a powerful tool for fault diagnosis by the capability of state estimation under nonlinear state and non-Gaussian noise,but there still exist many problems such as poor real-time performance.CUDA is a GPU-based parallel computing architecture proposed by NVIDIA,it enables GPUs to increase the speed of execution of algorithms.Therefore,this paper is based on the analysis of the parallelism of particle filter algorithm,and studies the fault diagnosis method of particle filter based on CUDA.The main research contents are as follows:(1)A parallel algorithm for optimizing particle filtering based on CUDA is preposed.Firstly,analyze the parallelism of particle swarm optimization(PSO)algorithm and particle filter algorithm,designe and implemente a CUDA-based parallel algorithm of particle group optimization particle filtering algorithm.Further,in order to avoid non-merging access to the GPU global memory problem,an improved reject resampling parallel algorithm is proposed,so that the threads in the same warp are enabled to resample the particles stored in the corresponding memory segment,so as to improve the efficiency of the resampling algorithm.The simulation of sone-dimensional nonlinear model shows that the proposed method has higher estimation accuracy and real-time performance.(2)Taking the variable pitch system and doubly fed induction generator(DFIG)of wind turbine as the research object,establish a fault diagnosis model according to the physical model of the variable pitch system,and realize fault diagnosis by modifying particle swarm optimization parallel particle filter algorithm combined with the smoothed residual.Firstly,obtain the estimated measured values of the system by using the improved particle swarm optimization particle filter parallel algorithm,compare the measured values with the actual measured values and produce residuals,then use the residual as the basis of fault detection to judge whether the system has a fault,which shows that the improved method can effectively improve the fault diagnosis performance of variable pitch system.