Rolling Bearing Fault Diagnosis Based on VMD-DWT and HADS-CNN-BiLSTM Hybrid Model

Abstract

This study proposes a hybrid framework for rolling bearing fault diagnosis by integrating a Variational Mode Decomposition–Discrete Wavelet Transform (VMD-DWT) with a Hybrid Attention-Based Depthwise Separable CNN-BiLSTM (HADS-CNN-BiLSTM) to address noise interference and low diagnostic accuracy under complex conditions. The vibration signals are first reconstructed using a genetic algorithm (GA)-optimized VMD and particle swarm optimization (PSO)-optimized DWT for noise suppression. Subsequently, the denoised signals undergo multimodal feature fusion through depthwise separable convolution, triple attention mechanisms, and BiLSTM temporal modeling. The hybrid model incorporates dynamic learning rate scheduling and a two-stage progressive training strategy to accelerate convergence. The experimental results on the Case Western Reserve University (CWRU) dataset demonstrate 99.58% fault diagnosis accuracy in precision, recall, and the F1 Score, while achieving 100% accuracy on the Xi’an Jiaotong University (XJTU-SY) dataset, confirming superior generalization and robustness under varying signal-to-noise ratios. The framework provides an effective solution for enhancing rolling bearing fault diagnosis technologies.