Feature Extraction and Recognition of Fault Vibration Signals Based on KPCA and QPSO-BP.

Abstract

Drive system is the central of rail vehicle and large machine, such as cardan shaft and gear box of the MU train. Identification of early vibration anomalies can effectively prevent further expansion of the fault. Taking the vibration signals of rolling bearing in strong noise background as an example, an automatic method of feature extraction and recognition based on Kernel-Based Principal Component Analysis (KPCA) and BP network optimized by Quantum-behaved Particle Swarm Optimization (QPSO-BP) is presented. Firstly 16 characteristics are calculated in timefrequency domain of denoised signals. After optimizing kernel function from Polynomial, Gaussian Radial Basis (GRB) and Multilayer Perceptron (MLP), the KPCA method is used to reduce the dimension of characteristics, and the eigenvector of different fault types is constructed as the input of BP network. From the view of quantum mechanics, the global search ability of traditional particle swarm optimization (PSO) is improved. In order to avoid falling into local minimum in network training and improve recognition accuracy, the initial weights and thresholds of BP network is optimized based on QPSO algorithm. The recognition accuracy of the traditional BP network is 87.1% while that of the improved QPSO-BP network model can reach 94.1%, which proves that this method can effectively diagnose the early fault of the drive system.