Multi-sparsity-based blind deconvolution and its application to wheelset bearing fault detection

Abstract

Fault detection of wheelset bearing is important to maintain the safety of high-speed train. The key in detecting the fault is to recover repetitive transients under heavy background noise from wheel/rail contacting. In this paper, a multi-sparsity-based blind deconvolution technique is proposed for vibration fault detection. Instead of maximizing impulsiveness nor cyclostationarity in finding an optimal filter, a comprehensive and compromised consideration of these two aspects is taken to increase the sparsity of time- and frequency-domain simultaneously. Based on this framework, smoothness index and Gini index are used in measuring the multi-sparsity. Then a knee point-driven strategy is incorporated to extend particle swarm optimization algorithm to solve the multi-objective blind deconvolution. The proposed approach is validated using a simulated signal and two cases of real wheelset bearing vibration signals. The comparative studies also demonstrate its robustness than some peer blind deconvolution techniques as well as envelope analysis methods.