Bearing Fault Diagnosis Based on Auto-Encoder

Bearings are important components in the operation of mechanical equipment, but the current intelligent diagnosis methods have a low recognition rate in complex environments and variable working conditions. Considering the advantages of deep learning in data processing and feature extraction (no manual feature extraction, high recognition accuracy), a fault diagnosis method with improved auto-encoder incorporating attention mechanism is proposed. The method combines the superiority of convolutional neural network and attention mechanism, based on the auto-encoder structure, and adds fault classification branches, the obtained features can better express the information of the original data, which can be achieved better accuracy, generalization and robustness of fault diagnosis. In the experimental data, using the 1D time domain data of the bearing as the model input, the recognition accuracy reaches 99.68%; after adding white noise to the original data, it still has more than 97.5% accuracy; in the case of variable working conditions, it can also achieve more than 95% accuracy. The experimental results show that the proposed method can achieve fault diagnosis of bearings in complex environments and variable working conditions, and has good robustness and generalization.