Research On Fault Diagnosis Method Based On Transfer Learning And Subspace Learning

Fault diagnosis research is of great significance to engineering fields such as machinery manufacturing,petrochemical industry,and wind power generation.As a key link in modern automation systems,it plays a role in ensuring the normal and stable operation of equipment,and at the same time,it also helps to improve the efficiency and product quality of industrial systems and prevent safety accidents.Transfer learning,as an important branch of machine learning methods,reduces the requirement that training samples and test samples must be independent and equally distributed.Based on the similarity of data or tasks,knowledge learned in one or more fields（source fields）can be transferred and applied to new scenes（target fields）.Subspace learning is a branch of fault diagnosis field,which solves the problem that traditional methods consider the whole relationship of data,but ignore the local information between samples.With transfer learning and subspace learning as the technical means,rolling bearing datasets and Tennessee Eastman chemical process as the research objects,this thesis mainly from the field adaptation,feature extraction and other aspects of research.The main contents are as follows:1.A rolling bearing fault diagnosis method based on deep residual neural network and transfer learning was proposed to solve the problem that the fault diagnosis effect of rolling bearing was not good when effective data samples were insufficient in variable working conditions.Firstly,the primary characteristic information of the signal is obtained by Laplace wavelet transform.Secondly,deep residual neural network is used to extract advanced feature information from primary feature information and construct residual neural network.Finally,transfer learning is introduced and the joint maximum discrepancy（JMMD）is used to measure the difference between the two distributions.2.Aiming at the problem of fault migration under variable working conditions,a fault diagnosis method based on conditional adversarial network with improved classifier is proposed.The domain discriminator is used to adjust the weight of the classifier samples to help distinguish similar samples,and the loss function of the network is rewritten.3.The proposed method improves feature extraction and classification,mined local hidden information on data.Combining fuzzy clustering and partial F test to estimate the contribution of samples to fault classification and divide the samples into smaller classes,the weight setting during classification is optimized.Experiments are carried out on four rolling bearing data sets and TEP data set to verify the validity and feasibility of the proposed methods based on transfer learning and subspace learning.