Data-driven bearing fault identification using improved hidden Markov model and self-organizing map

Abstract Efficient condition monitoring and fault diagnosis of bearings are of great practical significance since bearings are key elements in most rotating manufacturing machineries. In this study, a condition monitoring index of bearings is developed based on self-organizing map (SOM) in order to detect incipient bearing faults quickly. It requires low computation cost and is robust to the change of load level and motor speed, hence is quite suitable for online condition monitoring of bearings. Furthermore, a novel hybrid algorithm combining diversified gradient descent (DGD) method and Bayesian model selection (BMS) called DGD-BMS for the optimization of discrete hidden Markov model (DHMM) parameters is formulated under a general Bayesian framework. The flexibility of the DGD-BMS consists in that the algorithm can increase the diversity of the searching paths generated for DHMM parameters so that the true underlying parameters are more likely to be found out. Thus it provides an effective way to avoid trapping in one local maximum. Both simulation and industrial case study are presented to validate the proposed approach. Results show that the monitoring index can detect incipient bearing faults efficiently with 100% accuracy even under varying load levels, and the DGD-BMS method achieves on average the classification rate of 99.58%. The proposed method exhibits excellent performance compared to the conventional gradient descent (GD) and Baum-Welch (BW) methods.