Research On Rolling Bearing Fault Diagnosis Modeling Method Based On Convolutional Neural Network

Rolling bearings are widely used in industrial production of machinery and equipment.Monitoring and diagnosing the working process of rolling bearings and eliminating potential equipment faults in a timely,which are important prerequisites for ensuring efficient and safe production of enterprise machinery and equipment.Therefore,it is of great practical engineering significance to conduct research on the related fields of rolling bearing fault diagnosis.In recent years,artificial intelligence fault diagnosis led by large-scale data-driven and new generation deep learning has made great progress.Convolutional Neural Network(CNN)is developing rapidly in intelligent fault diagnosis with its advantages in fault feature extraction and pattern recognition.However,the structure and hyperparameters of fault diagnosis modeling based on CNN are selected based on experience,and there is a lack of uniform guidelines,which limits the application of CNN in fault diagnosis.In terms of fault diagnosis model generalization,the model generalization performance is poor due to insufficient fault samples and potential differences of training and testing samples,which lead to the fact that the diagnostic model constructed on one device cannot be well applied to other devices for diagnosis.In view of the above problems,this paper carries out the research on rolling bearing fault diagnosis modeling method based on CNN from model structure,model hyperparameters and model generalization respectively,and the details of the research are as follows.(1)At present,in the modeling research of fault diagnosis method based on CNN,there is no uniform standard for the selection of model structure,and most of them are selected by experience to build the corresponding fault diagnosis model,which cannot be modeled quickly and effectively for the actual fault diagnosis task,and the model diagnosis performance is low.To address this problem,the research on the influence of the structure of CNN on the diagnostic performance of the model is carried out,and the branch structure,connection mode,network depth and network width are the important factors affecting the diagnostic performance of the model by analyzing the structural characteristics of CNN.The bearing fault data set of Case Western Reserve University is used to test the structure model and the corresponding evaluation indexes are used to judge the diagnostic performance of the model,and finally the model structure selection strategy is given for rolling bearing fault diagnosis modeling.(2)Aiming at the problem of lack of unified guidance of hyperparameter configuration in fault diagnosis modeling based on CNN,a research on hyperparameter optimization of fault diagnosis model was carried out and applied to rolling bearing data set.Firstly,the correlation between hyperparameters of convolutional neural network and fault diagnosis is analyzed,and the adjustment range of hyperparameters optimization is derived;secondly,the corresponding benchmark model of fault diagnosis is constructed,and the law of hyperparameters affecting the diagnostic performance of the model is analyzed through optimization experiments;the intrinsic factors of hyperparameters affecting the diagnostic performance of the model are explained through the output of the visualization intermediate layer,and finally,the strategy of hyperparameters selection for rolling bearing fault diagnosis modeling is given.(3)In view of the problem that the fault diagnosis model constructed in the actual engineering application scenario cannot be well generalized to another equipment for diagnosis,a research on the generalization method of fault diagnosis model is carried out and applied to rolling bearing data set.Firstly,a parameter transfer-based fault diagnosis model generalization method is proposed,using the training data in the source domain to pre-train the rolling bearing fault diagnosis model and using the cosine similarity as a measure of the feature distribution between the source domain and the target domain,and realizing the adaptive transfer of the model parameters based on the parameter transfer strategy;secondly,a feature transfer-based fault diagnosis generalization method is proposed,by constructing a feature transfer network,combining the common fault modes of the equipment for unsupervised learning,mapping and sharing the association of the feature between the source and target domains to obtain domain invariant features,further improving the generalization performance of the model on different working conditions and different equipment,and achieving good diagnostic results on different bearing data sets.