Hybrid Convolutional Neural Network–Bidirectional Gated Recurrent Unit with Dual Attention and Reinforcement Learning for Rolling Element Bearing Fault Diagnosis

Abstract

The diagnosis of rolling element bearing faults is of most importance for the reliable and safe operation of rotating machinery. An adaptive selection of fault features based on a hybrid convolutional neural network (CNN)-bidirectional gated recurrent unit (BiGRU) framework with dual attention and reinforcement learning (RL) is investigated. First, the CNN learns the local vibration features from raw one-dimensional signals. Second, the BiGRU learns the bidirectional temporal dependencies from the learned features. Then, the channel attention and temporal attention further refine the learned features to be more informative feature channels and diagnostically significant temporal states. Finally, an RL-based probabilistic masking policy adaptively suppresses redundant latent features through sparse feature selection. Therefore, the proposed framework combines adaptive latent-feature sparsification, bidirectional temporal modeling, dual-level attention refinement and local feature extraction in a unified end-to-end architecture. In the experimental study on a laboratory vibration dataset with five bearing conditions, the proposed framework achieved 99.31% accuracy under the primary evaluation protocol. As a separate and more reliable assessment of generalization across independent original recordings, a five-fold source-level cross-validation procedure was performed, with a mean test accuracy of 90.38% ± 6.27%.The generalization ability of the proposed approach was also validated by the accuracy of 99.57% obtained in the cross data validation using the benchmark dataset of Case Western Reserve University (CWRU). The results demonstrate the diagnostic capability of the proposed framework while maintaining practical computational characteristics and incorporating adaptive feature selection.