Study On Fault Diagnosis Of Rolling Bearing Based On Deep Transfer Learning

Rolling bearings are the core components of rotating machinery,and their health is directly related to production.In actual production,accurate bearing fault diagnosis can guide the maintenance team to carry out maintenance work reasonably,reduce maintenance cost and time after failure,and avoid accidents caused by failure.With the rise of big data and artificial intelligence,the development of intelligent manufacturing has been promoted.In the current context,data-driven intelligent fault diagnosis of rolling bearings has received extensive attention.However,in actual industrial scenarios,intelligent diagnosis of rolling bearings still faces many challenges.Continuous advances in deep learning and transfer learning provide new opportunities in the field of intelligent diagnosis.This thesis takes rolling bearings as the research object,takes deep learning and transfer learning as the core theory,focuses on the key issues of intelligent diagnosis in practical industrial applications,and conducts research on small sample and cross-domain fault diagnosis.The main research contents of the thesis are as follows:Aiming at the problems of sparse labeled fault samples in actual industrial applications and low accuracy and poor generalization of deep network models under small sample conditions,a small sample fault diagnosis method based on Siamese networks was proposed.According to the spatio-temporal characteristics of the vibration signal,a feature extractor cascaded with a convolutional neural network and a bidirectional gated recurrent unit is designed to extract the feature of the vibration signal.Then,similarity measurement and fault classification are performed on the extracted features under the Siamese network architecture.Through the multi-task joint training mode,the knowledge sharing of multiple tasks is realized,and the similarity information and discriminative information of the fault data are fully utilized to reduce the dependence on a large number of labeled data.Using the bearing data collected in the laboratory,the algorithm is evaluated in the case of sparse fault samples,and the effectiveness of the proposed algorithm is verified.Aiming at the unsupervised cross-domain diagnosis problem in which the data distribution in two domains is different in practical,and the data in the target domain is not marked,the idea of transfer learning is introduced,and a fault diagnosis algorithm based on transferable capsule network is proposed.Convolutional networks and capsule networks are cascaded as the backbone network for extracting multi-layer features of two domains.In order to reduce the distribution difference of the two-domain data,a progressive domain alignment method is designed,which divides multiple alignment tasks according to the number of categories in the diagnosis problem,and executes them sequentially during the training process.Among them,the number of sub-domains in each task is continuously increasing,which avoids the negative transfer caused by wrong pseudo-labels when directly performing conditional distribution alignment,and effectively realizes fine-grained domain alignment.The algorithm is evaluated on the transfer tasks of cross-working conditions,cross-acquisition locations and transfer tasks from artificial damage to real damage,and the superiority of the proposed algorithm is verified.