Research On Bearing Fault Diagnosis Algorithm Based On Long-short Term Memory Networks

Rolling bearings are one of the most commonly used mechanical components.Effective and reliable real-time monitoring systems for bearing faults detection are of great significance for modern industrial development.Most existing fault diagnosis algorithms use a combination of manual feature extraction and classifiers to implement fault identification.These models are often complicated and it is easy to lose key information and cannot be generalized.This paper takes the rolling bearing as the research object and proposes a bearing fault diagnosis model based on the LSTM,i.e.the longterm and short-term memory network.It can automatically extract features and accomplish fault identification based on the characteristics of the network.This paper first establishes the LSTM-based bearing fault diagnosis model,maps the data to the linear network layer,trains the parameters through the long-term and shortterm memory network,and then inputs them into the softmax output layer to obtain the probability distribution for each classification category.The model was applied to the bearing data from the Case Western Storage Bearing Failure Test Platform and was performed on the bearings with pitting failures in the inner ring,outer ring or rolling body.The results show that the model can effectively identify the bearing fault locations and failure levels.The comparisons with the SVM model and with the LSTM model using the energy features extracted from wavelet packet show the proposed model produces more accurate and stable results.The bearing fault diagnosis model based on the LSTM has reached a high accuracy.However,the training speed is slow and it is difficult to meet the requirements of realtime monitoring.The convolutional neural network CNN with the convolutional layer and the pooling layer can fully extract data features and reduce the data dimensions and training time.Thus,a LSTM fault diagnosis model combined with the one-dimensional convolution is proposed.The convolution layer and the pooling layer replace the linear input layer in the LSTM model.Simulation results show that for the same dataset,the CNN-LSTM model improves the accuracy of the LSTM model and greatly speeds up the training.In order to improve the generalization ability of the model for bearing faults detection under different working conditions and to enhance the robustness of the model,an improved CNN-LSTM fault diagnosis model is proposed.The datasets with different types of bearings and loads were reconstructed by data enhancement.The output of the LSTM network layer was dropped out,the linear output layer was batch-standardized,and the probability distribution of the predicted values was obtained by the softmax activation function.Experimental results show that this model can solve the overfitting problems more effectively than the previous CNN-LSTM model and therefore improves the diagnostic performances.Finally,the t-SNE visualization technology was used to verify the rationality of the network model design.