Multi sensor fusion motor bearing lightweight fault diagnosis based on IAOO-CNN-MA-BiLSTM model

To address the difficulties in multi-sensor signal feature extraction, high model complexity, and uncertain hyperparameter determination in motor bearing fault diagnosis, this study proposes a lightweight multi-sensor fusion fault diagnosis method based on an improved Animated Oat Optimization algorithm (IAOO) for optimizing a convolutional neural network (CNN) and a multi-scale attention-enhanced bidirectional long short-term memory network (MA-BiLSTM). First, IAOO is used to adaptively optimize the key hyperparameters of CNN and MA-BiLSTM. By introducing adaptive inertia weights, chaotic initialization and perturbation, elite opposition-based learning, and Lévy flight mechanisms, the global search capability and convergence accuracy of the AOO algorithm are improved. Second, CNN is employed to extract local features from multi-sensor vibration signals, while MA-BiLSTM captures bidirectional temporal dependencies. A multi-scale attention mechanism is introduced to focus on key information at different temporal scales. Meanwhile, gated convolution and depthwise separable convolution are incorporated to achieve model lightweighting. Finally, experimental validation is conducted on the self-built NU-Bearing-4States dataset. The proposed method is compared with mainstream fault diagnosis algorithms, including SVM, Random Forest, 1D-CNN, LSTM, BiLSTM, CNN-LSTM, Transformer, and ResNet-18, and time–frequency analysis is also performed. The experimental results show that the proposed method achieves a test accuracy of 99.32%, precision of 99.28%, recall of 99.25%, F1 score of 99.26%, inference time of 12.96 s, and parameter size of 3.74 M, outperforming all comparison algorithms across all metrics. The time–frequency analysis further reveals the differences in vibration signal characteristics under different fault states. The core innovations of this study are as follows: (1) an IAOO algorithm is proposed, which effectively overcomes the slow convergence and tendency to fall into local optima of the original AOO through the synergistic mechanism of adaptive inertia weights, chaotic initialization and perturbation, elite opposition-based learning, and Lévy flight; (2) an MA-BiLSTM module is designed to adaptively fuse fault features at different temporal scales using a multi-scale attention mechanism; and (3) gated convolution and depthwise separable convolution are introduced to achieve model lightweighting while maintaining diagnostic accuracy, with only 3.74 M parameters. The experiments demonstrate that the proposed method achieves an accuracy of 99.32% on the test set and exhibits strong noise robustness and real-time performance, providing an effective solution for intelligent early warning of motor bearings.