Early fault diagnosis based on reinforcement learning optimized-SVM model with vibration-monitored signals

Effective fault diagnosis maximizes economic benefits by ensuring the stability of machinery systems. Detecting the faults of the key components in machinery, such as rolling bearings, at an early stage, helps to avoid accidents to optimize the maintenance efficiency. It is well known that faulty bearings always deliver a message through their abnormal vibration variation, which can be captured by vibration acceleration sensors in order to facilitate the deteriorating status assessment. However, a clue for an early fault is so ambiguous and sometimes masked by ambient noise, which makes the early fault diagnosis a challenging problem. To tackle the problem, we propose a vibration signal-based data-driven early fault diagnosis approach based on the reinforcement learning (RL) optimized support vector machine (SVM) model. The exploration of the hyperparameter optimization using RL to improve SVM performance motivates this research. Firstly, the corresponding features in the time domain, frequency domain and time-frequency domain are extracted from the obtained vibration signals of the key components under certain working conditions. Subsequently, to better recognize the pattern of an early fault, linear discriminant analysis (LDA) is employed in fuzing the multi-domain early fault features. Finally, the fused features are fed into the RL optimized-SVM model for fault diagnosis. Experimental validation was performed with a public dataset of rolling bearings, and the results confirmed the effectiveness and superiority of the approach compared with other methods.