Research On Fault Diagnosis Method Of Rotary Machine Based On Deep Transfer Convolutional Neural Network

With the rapid development of modern industry and the continuous upgrading of machine operating equipment,timely and effective fault diagnosis of rotary machine can avoid economic losses and safety accidents.Traditional fault diagnosis methods rely on expert experience and prior knowledge,which no longer meet the fault diagnosis requirements of complex mechanical equipment.It is necessary to continuously research and develop more efficient and intelligent fault diagnosis methods.The rise of deep learning and transfer learning provides new ideas for rotary machine fault diagnosis technology.Based on this,this thesis takes rotary machine as the research object,and uses convolutional neural network and transfer learning technology as the theoretical basis to carry out research on the fault diagnosis method of rotary machine.The main research focus of this thesis includes three aspects as follow:(1)Aiming at the problems of cumbersome manual feature extraction based on traditional machine learning fault diagnosis method and the impact of shallow representation on diagnosis performance,a fault diagnosis method based on time-frequency analysis and convolutional neural network is proposed.Firstly,raw vibration signals are converted into time-frequency images by continuous wavelet transform.Then the features that is beneficial to the fault classification task is automatically learned from the transformed images by using the feature learning ability of the convolutional neural network.Finally,the fault classification is completed after the Softmax.The effectiveness of this method was verified on the motor bearing dataset from Case Western Reserve University(CWRU).(2)In view of the insufficient number of labeled fault samples in the field of rotary machine fault diagnosis,which limits the depth of the network model,resulting in a series of problems such as poor generalization of the model and poor diagnosis effect,a deep transfer convolutional neural network(DTCNN)is proposed.This method combines convolutional neural networks with transfer learning,Res Net-50 architecture is used as the pre-trained DTCNN model.By fine-tuning the pre-trained deep network,the proposed DTCNN model has strong feature extraction ability in the fault diagnosis of rotary machine,so as to achieve high-precision fault classification.Experimental verification was carried out on bearing dataset and self-priming centrifugal pump dataset.The results show that the proposed method can achieve very high classification accuracy on the two datasets,and has good generalization ability.(3)In the fault classification task based on deep transfer Learning technology,the model with Softmax classifier exposes shortcomings such as insufficient generalization and insufficient classification speed.Aiming at this problem,a new fault diagnosis method combining DTCNN with extreme learning machine(ELM)is proposed.This method aims to use the powerful feature extraction capabilities of DTCNN and the effective classification capabilities of ELM to achieve more accurate and faster fault classification.The effectiveness of proposed method was verified on the CWRU bearing datasets and Paderborn University bearing dataset.The results show that the proposed method has certain advantages in classification accuracy and computation time,and the results verify the effectiveness and efficiency of the proposed methodBased on the convolutional neural network algorithm and transfer learning technology,a series of researches are carried out in this thesis,aiming to provide a new idea for the field of rotating machinery fault diagnosis,which has certain engineering application value and theoretical reference value.