Bearing Fault Diagnosis Based on the Improved Residual Network

Aiming at the problem that the samples of fault data in bearing fault diagnosis account for a relatively small proportion in the dataset, which leads to a low correct rate of fault diagnosis, a bearing fault diagnosis method based on CNN-LSTM with attention mechanism (SE-1DCNN-LSTM) and multi-channel residual one-dimensional convolutional neural network (RES-M1DCNN) is proposed. First, the training set of one-dimensional data of bearings is input into the network, and the combined network of SE-CNN-LSTM with attention mechanism and multi-channel RES-M1DCNN is utilized for feature extraction. SE-CNN-LSTM first extracts the more important temporal signals of the fault data by using the attention mechanism, and then inputs the extracted signals into the multi-channel residual network. The multi-channel network structure improves the feature extraction capability and adaptability of the network by increasing the width of the network and expanding the receptive field. The residual connection retains the shallow features while extracting the deep features, which are integrated in the multi-channel output and input to the Softmax layer for classification, and then classify the bearing vibration timing signals. This paper compares different fault diagnosis models and proves that the fault diagnosis model combining SE-CNN-LSTM and RES-M1DCNN has improved the correct rate in bearing fault diagnosis, which verifies the effectiveness of this model.