Multi-sensor Data Fusion Bearing Fault Diagnosis Method Based on MSCNN-CoordAtt-BiLSTM with Strong Noise-Resistant Performance

Abstract To address the limitations of single-sensor diagnostic information incompleteness and the poor noise resistance of existing multi-sensor fusion models, a multi-sensor data fusion bearing fault diagnosis method based on MSCNN-CoordAtt-BiLSTM is proposed. First, vibration signals from multiple sensors are segmented into subsequences and normalized. Then, a multi-scale convolutional neural network (MSCNN) layer is employed to extract key features from the noisy signals, followed by feature concatenation and fusion through a feature fusion layer. Subsequently, the fused features are fed into an improved one-dimensional coordinate attention (1D-CoordAtt) mechanism to capture effective fault feature information. Further feature extraction is performed using deep convolutional layers with adaptive batch normalization (AdaBN), combined with a bidirectional long short-term memory (BiLSTM) network to extract temporal feature information. Finally, fault classification is accomplished using an adaptive average pooling (AdaptiveAvgPool) layer and a Softmax layer. Ablation and comparative experiments on the Case Western Reserve University (CWRU) and Southeastern University (SEU) bearing datasets demonstrate that the proposed method achieves diagnostic accuracies of 99.37% and 99.51% under the signal-to-noise ratio of 0 dB, and 96.68% and 96.23% under -6 dB, respectively. These results surpass those of other comparative models, confirming the method's superior diagnostic accuracy and robust noise immunity.