Optimized GRU with Self-Attention for Bearing Fault Diagnosis Using Bayesian Hyperparameter Tuning

Rolling bearing failures cause significant production downtime and economic losses. Traditional diagnostic methods suffer from low efficiency, suboptimal accuracy, and susceptibility to human subjectivity. To address these limitations, this paper proposes a novel bearing fault diagnosis (BFD) approach leveraging a Gated Recurrent Unit (GRU) network. Key contributions include: (1) Employing Bayesian optimization to automate the search for the optimal GRU architecture (layers, hidden units) and hyperparameters (learning rate, batch size, epochs), significantly enhancing diagnostic performance (achieving 97.9% accuracy). (2) Integrating a self-attention mechanism to further improve the GRU’s feature extraction capability from vibration signals, boosting accuracy to 99.6%. (3) Demonstrating the robustness of the optimized GRU with self-attention across varying motor speeds (1772 rpm, 1750 rpm, 1730 rpm), consistently maintaining diagnostic accuracy above 97%. Comparative studies with Bayesian-optimized Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models confirm the superior accuracy (97.9% vs. 95.1% and 90.0%) and faster inference speed (0.27 s) of the proposed GRU-based method. The results validate that the combination of Bayesian optimization, GRU, and self-attention provides an efficient, accurate, and robust intelligent solution for automated BFD.