Stochastic optimization methods applied to BP network based fault diagnosis problems of rotating machinery

Abstract

BP network has been successfully used in the fault diagnosis of rotating machinery, however BP network's drawbacks, such as low convergence rate and its easy fall into local optima have restricted its wider applications, especially to those complex multimodal problems. Two of the recently proposed stochastic optimization methods: adaptive particle swarm optimization (APSO) and adaptive genetic algorithms (AGA) are discussed. And the way that BP network's initial weights and bias are optimized by those two methods is also carefully discussed. Compared with standard particle swarm optimization(SPSO), APSO solves the premature convergence problem better by giving particles a spatial extension and adaptive mutation. In this paper, firstly APSO and AGA are used to optimize the initial weights of BP network, then the APSO-BP and AGA-BP networks are used to diagnose the turbo-pump faults, and the experimental results show many advantages in convergence speed and accuracy. The comparison between AGA and APSO is also discussed.