Feature Extraction and Recognition of Fault Vibration Signals Based on Principal Component Analysis and Neural Network: An Artificial Intelligence Approach

Abstract

In high speed train maintenance, the damage detection and fault recognition for local defects at early stage is very important. Because, damages generated in drive systems like axle box bearing and gear box bearing have characteristics to grow fast and which can cause mass disaster. In this study, the authors report a fault diagnosis method by the combination of advanced signal processing technique with pattern recognition methods. Sixty fault vibration signals gathered under strong background noise are taken as a case study. On the basis of wavelet de-noising and singularity analysis, features of them in three domains are all computed for a time-frequencyenergy representation. Improved principal component analysis has been conducted in the three spaces for independent analysis and rational integration. A neural network recognition model with three binary numbers as the output mode has been established so that a series of tests on four typical or restructured eigenvectors is performed accordingly to find the optimal eigenvector for these vibration signals. Using this carefully designed feature extraction and recognition approach, not only would the fault vibration signals be detected as expected, but location and types of them could also be recognized. Thus feasibility of this artificial intelligence approach is implied for further research as well as industrial applications in rail transit and other related engineering fields.