Research On Fault Diagnosis Algorithm Based On Deep Transfer Learning

With the rapid development of science and technology and the substantial improvement of productivity,rotating machinery plays an important role in engineering applications,but at the same time,its internal structure is complex,and various failures will inevitably occur,leading to interruption of production processes and even catastrophic accidents.occurrence.Intelligent fault diagnosis is to use machine learning algorithms to detect and judge the operating status of mechanical equipment to ensure the safe and reliable operation of equipment.However,there are still some problems in fault diagnosis at this stage.First,the traditional intelligent diagnosis algorithm relies on a large amount of labeled data to train the model,and manual fault labeling is a huge and impractical project.Second,the feature extraction method based on deep learning is not well adapted to the fault diagnosis task,Secondly,the feature extraction method based on deep learning is not well adapted to the fault diagnosis task,only uses it to extract the general features of the fault diagnosis signal but fails to mine the characteristics of the data set.This thesis improves the intelligent diagnosis algorithm for the above problems,the main research work is as follows:（1）In view of the problem that a large amount of labeled data is required for model training,this thesis introduces the idea of transfer learning,which migrates labeled data in related fields to fault diagnosis in the unlabeled field,and proposes a Discriminative Joint Matching（DJM）algorithm.The DJM algorithm combines the two methods of transfer learning domain adaptation,feature matching and instance reweighting,when constructing the feature matrix to narrow the difference between the two domains as much as possible.In addition,DJM redivides the domain into the same class domain and different class domain according to the category,improves the transferability by reducing the distance of the same class domain,and expands the distance between different class domains to improve the discrimination.The experimental results show that the algorithm proposed in this thesis has good performance in both transfer learning general image datasets and fault diagnosis datasets.（2）Aiming at the problem of poor generalization ability of deep learning models in fault diagnosis tasks,this thesis proposes a fault diagnosis model based on Multi-Layer Subspace Domain Adaptation（MLSDA）.First,the original vibration signal samples are preprocessed,and the grayscale image is used for imaging;then,the pre-trained network is used to extract the general features of the image,and the gradient features of the pre-trained network parameters are calculated;finally,embedded in multiple high layers of the network Improved domain adaptation module to obtain more local information.In this thesis,experiments are carried out on the fault diagnosis dataset of variable working conditions.The experimental results show that the proposed model has good diagnostic ability.