Research on Multi-Sensor Fusion Motor Bearing Fault Diagnosis Based on Improved CNN-BiLSTM Hybrid Model with Attention Mechanism

Currently, most multi-source sensor fusion schemes often focus on the integration of information at a single level, and fail to fully consider the interdependence characteristics among different signal pathways during the integration process. To alleviate the impact of class imbalance on deep networks, this study designs a CNN-BiLSTM hybrid architecture based on attention mechanism optimization, and builds a multi-level fusion network for motor bearing fault identification around this core. In terms of methods, first, a feature extraction backbone combining pyramid structure and attention operation is established; then, a bidirectional feature pyramid module is introduced to further enhance the ability to extract multi-scale features; on this basis, a dual-branch deep neural network is constructed, which is composed of an attention-enhanced branch and a CNN-BiLSTM branch connected together. Specifically, the attention-enhanced branch is responsible for capturing multi-level convolutional features from grayscale time-frequency images and embedding channel attention mechanisms to enhance the discrimination ability of spatial expression; the CNN-BiLSTM branch uses the collaboration of convolutional modules and bidirectional long short-term memory networks to depict the dependency pattern of acoustic signals in the time dimension. Finally, fault classification is performed at the output of the fully connected layer to identify the motor bearing fault categories under the multi-sensor fusion scenario. It should be emphasized that this study focuses on the fault classification task rather than regression prediction. In addition, the model visualization method and interpretability analysis are introduced in the paper to improve the interpretability of the behaviour of the deep model.