Rolling Bearing Fault Diagnosis Based On Deep Learning Combination Model

Rolling bearings are the core components widely used in the field of rotating machinery and equipment.Once a fault occurs,it will pose a serious threat to the mechanical system,the life safety of workers and the national economy.Therefore,the accurate diagnosis of rolling bearings’ faults has far-reaching practical significance.For rolling bearing fault diagnosis problems,combined with the advantages of deep learning algorithm,we propose the use of deep learning to establish a composite model to realize intelligent fault diagnosis of rolling bearings.The main research contents include:Starting from the structural composition and failure mechanism of rolling bearings,respectively time domain,frequency domain and time domain features extraction were analyzed.On this basis,Introduced a representative basic artificial neural network model:Convolutional Neural Network(CNN),Back Propagation(BP)and Deep Belief Net(DBN),It provides theoretical and experimental verification basis for the follow-up study of rolling bearing fault diagnosis based on deep learning in this paper.A combined model of rolling bearing fault diagnosis based on CNN-LSTM is proposed,which improves the diagnosis performance of a single neural network model.Use Long Short Memory Network(LSTM)to extract global features of rolling bearing fault vibration signals to make up for the defect that CNN can only extract short-term local features,and input the extracted feature vectors into the fully connected layer and the Softmax classification layer.The Adam optimization algorithm is used to dynamically adjust the parameters of the entire model to achieve rapid and accurate diagnosis of rolling bearing faults.Experimental results on the Bearing Data Set(CWRU)of Case Western Reserve University in the United States show that the proposed model is more accurate than the DBN and LSTM fault diagnosis models,and solves the problem of poor accuracy and time-consuming LSTM fault diagnosis models.Based on the unsupervised learning method,the SSAE-SVM rolling bearing fault diagnosis combined model is proposed,which solves the problem that the existing algorithms rely excessively on labeled fault data.Using Stacked Sparse Autoencoder(SSAE)for unsupervised deep learning to obtain high-dimensional deep features of rolling bearing faults,a 5-layer SSAE adaptive learning network is constructed.The layer-by-layer greedy algorithm training and the reverse fine-tuning algorithm are used to improve it,and finally the deep feature vector is output to the Support Vector Machine(SVM)supervised learning classifier to realize the rolling bearing fault classification.The SSAE-Softmax classifier,the BP neural network model based on supervised learning,and the CNN-LSTM rolling bearing fault diagnosis combined model proposed in Chapter 3 are used for comparative experiments on the data sets of the driving end and the fan end of the CWRU rolling bearing.Accuracy,training time,and performance under different iteration times all show better superiority.Aiming at the fault diagnosis of rolling bearing with limited training data samples,a FewShot Learning(FSL)model based on twin CNN is proposed to ensure the good diagnostic performance of the model in the case of fewer samples.A twin neural network is formed by two CNNs with the same network structure and a first-layer wide convolution kernel.When the model is trained,two samples of the same or different categories are input,and the output uses a metric learning method to measure the similarity between the input signals To judge whether it is regarded as the same fault,and finally input the test set samples into the trained model to realize fault category diagnosis.According to the principle of FSL,two models of One-shot learning and Five-shot learning are established,and a CNN fault diagnosis model based on a wide convolution kernel is constructed for comparison and verification.On the CWRU rolling bearing drive end data set,comparative experiments were carried out on the impact of different training sample sizes and different noise levels on the fault diagnosis performance of the built model.The results show that the proposed learning method with fewer samples is compared with the limited amount of data available.More effective than comparing models.