SSENet: A Deep Neural Network Featuring Stacked Attention mechanism for Enhanced Bearing Fault Diagnosis

This paper presents a deep learning network model, SSENet, tailored to address the performance degradation of bearing fault diagnosis models due to significant data distribution variations across different operating conditions. SSENet innovatively incorporates two core modules: a dual-branch feature extraction module and a joint optimization module. The dual-branch feature extraction module comprises two distinct network architectures, each specialized in processing data from specific operating conditions to capture the unique characteristic information of each. Conversely, the joint optimization module integrates the loss functions of both branches to achieve global optimization, ensuring robust model performance across diverse environmental conditions. The model introduces a Squeeze-andExcitation (SE) module with regularization and incorporates a transformer block capable of extracting long-term dependency information. This design aims to bolster the model’s representational capacity by dynamically adjusting the significance of feature channels and effectively mitigating the risk of overfitting. Consequently, this architecture significantly enhances the SSENet model’s performance in bearing fault classification under complex and fluctuating operating conditions. Furthermore, we augmented the original vibration signal with the first percentile signal ‘q1’ and the third percentile signal ‘q3’ to extract additional useful features for signal classification, thereby further improving performance.