Bearing Fault Diagnosis and Life Prediction Based on Attention Mechanism and Long Short-term Memory Network under Multi-source Sensor Data

To solve the problem of the poor diagnosis effect of rolling bearing faults in noisy environments, we propose a new bearing fault diagnosis method based on an attention mechanism and a long short-term memory (LSTM) network using multisource sensor data. First, we normalize the one-dimensional (1D) data collected by multisource sensors and then construct a double-channel twin convolutional network with adaptive batch normalization technology to extract effective features and perform data fusion. Second, the fusion data are input into the improved 1D coordinate attention (CoordAtt) which is capable of considering the relationship between channels and position information simultaneously. Third, we use the LSTM layer to extract time features and evaluate the diagnosis effect by the loss function after label smoothing regularization. We use the new optimizer Adan for optimization. Finally, the diagnosis model is applied to a test set, and the diagnosis results of fault categories are output. We compare the diagnostic accuracy of the model under different test set ratios, determine that the optimal ratio is 0.KG-*53, and test it in a noisy environment. The experimental results show that the proposed method can better resist the influence of noisy environments. The validity of the CoordAtt-LSTM model in life prediction is verified by experimental results on the C-MAPSS dataset.