An Unsupervised Condition Monitoring Method for Rolling Bearings Based on Compound Feature Selection and Multistep-Aware BiGRU-VAE

Establishing a health indicator (HI) is essential for monitoring the condition of rolling bearings and mitigating failure risks. Unsupervised learning, which eliminates the necessity for manual labeling, has gained considerable attention in constructing HIs. However, existing unsupervised methods mainly suffer from two limitations. First, redundancy analysis is generally disregarded in the feature selection process. Second, temporal dependencies between distinct instants are neglected due to single time-step samples. To address the issue, this article proposes an innovative unsupervised condition monitoring method based on compound feature selection (CFS) and a multistep-aware model. The major contributions are comprehensive feature selection and effective information extraction of continuous sampling instants. Specifically, in feature selection, a novel metric is designed to effectively capture abrupt changes in features proximate to failure. Utilizing this metric, a CFS algorithm is proposed to identify a superior feature subset with a high correlation to the degradation and low redundancy among features. Additionally, a bidirectional gated recurrent unit-based variational autoencoder (BiGRU-VAE) with multiple time-step inputs is constructed to leverage the information of different sampling instants. The BiGRU-VAE is trained exclusively on healthy features within the superior subset. The dissimilarity between the input sample and its reconstruction is employed as the HI to represent bearing degradation. Comparative experiments on two run-to-failure datasets validate the superiority of the HIs constructed by the proposed method in presenting the degradation process. Ablation experiments confirm the dominance of the CFS algorithm and the necessity of multiple time-step inputs.