Research On Intelligent Fault Diagnosis Method Based On Hybrid Network Compression And Deep Transfer Learning For Rolling Bearings

Rolling element bearing is an important part of rotating machinery,and its sudden failure would learn to a serious of impacts on normal operations of the mechanical system.Therefore,it is critical to conduct accurate bearing fault diagnosis for the safe operation of the whole system.In recent years,intelligent fault diagnosis technology based on deep learning has significantly improved the accuracy of fault diagnosis models.However,there are still some problems of engineering application and unsatisfied reslts for intelligent fault diagnosis models.To solve these problems,this thesis studies network compression and transfer learning based on deep convolutional neural networks,and aims to provide a more accurate and smaller model for bearing intelligent fault diagnosis.The main research contents and innovations are summarized as follows:(1)In the process of knowledge distillation,a smaller student network can hardly follow the diagnostic capability of a larger teacher network.To solve this problem,an improved probability knowledge distillation(IPKD)method is proposed.In this method,the probability distribution of teacher network in feature space is taken as knowledge,and set as soft labels for training the student network.Meanwhile,hard labels are also used to guide the learning of the student network,both of which aim to improve the performance of knowledge distillation.The experimental results show that the student network after distillation has similar diagnosis accuracy to the teacher network,but its scale is smaller.(2)In the process of knowledge distillation,the performance of the student network is limited by its teacher network.When the distribution of actual test data and historical training data is quite different,the generalization ability of the teacher network is insufficient.To solve these problems,a batch normalization attention mechanism domainbased adversarial neural network(BNAM-DANN)is proposed.Firstly,a batch regularized channel and spatial attention mechanism are constructed to enhance the feature extraction capability of complex data.Then,they are embedded into the domain adversarial neural network in series,and the knowledge extraction and transfer of source domain are enhanced by the adversarial learning between source domain and target domain.The diagnostic performance of the proposed network is verified by using variable state and cross-equipment bearing data sets.The results indicate that the diagnosis accuracy of the teacher network can be significantly improved when having large drift between source and target domains is large and variable conditions.(3)Considering that the intelligent diagnosis model needs both high accuracy and applicability,a hybrid model based on quantitative distillation and deep transfer learning is constructed for intelligent fault diagnsos of bearings.Firstly,the BNAM-DANN is used to train the teacher network to improve its generalization ability.Then,by using the IPKD method,the knowledge learned from the teacher network is abailble for the learning of the student network.Moreover,the model quantization is added in the distillation process,which not only accelerates the distillation training process,but also further reduces the scale of the student network.Complex experimental bearing data sets and actual bearing test data set are used to set the transfer tasks,including variable speeds/variable fault modes and transfer from experimental bearings to wheelset bearings.The experimental results demonstrate that the proposed method makes a good balance between the requirements of higher accuracy and smaller network,and realizes bearing fault diagnosis with large domain shifts by using a student network with compact structure and small memory space.It can be a practical model and valuable reference for intelligent diagnosis in real applications.