Research On Rolling Bearing Fault Diagnosis Method Based On Deep Auto-encoder Neural Network

With the rapid development of industrialization,normal and orderly operation of machinery,especially large rotating machinery,plays a vital role in industrial production and life.Rolling bearing is the key component of many large rotating machinery.The long-term safe operation of rotating machinery is closely related to the stability of rolling bearing.Therefore,it is of great significance to study the fault diagnosis method of rolling bearing.Since the era of big data,traditional fault diagnosis methods need a lot of manual processing,which can not meet the development requirements of the era of big data.Due to the advantages of automatic feature extraction from original data,deep learning has been more and more popular in the field of mechanical fault diagnosis.However,deep learning exposes the difficulties of parameter selection,training and long training time,so it is difficult to train a deep learning model with good generalization and high accuracy.In order to solve these problems,this thesis proposes a fault diagnosis model based on improved deep stack auto-encoder neural network(DSAE).In order to study the influence of the super parameter setting of DSAE network model on the model performance,the vibration simulation signal is constructed,and the comparative experiments of various Super-parameter settings in the network are carried out.The performance is measured by the learning rate,training time and error convergence curve.In view of the need of DSAE network for large samples and the long processing time of traditional time-frequency preprocessing method,a new data preprocessing scheme is proposed,which can extract the composite information of data in a short time.The basic framework of fault diagnosis model based on DSAE is determined.The performance of neural network depends largely on its model structure and corresponding learning optimization algorithm.In order to solve the problem of slow gradient transfer caused by internal covariate migration in the fine-tuning process of deep stack autoencoder neural network(DSAE),an adaptive balance processing layer is added in the fine-tuning phase of DSAE network based on Batch Normalization(BN)algorithm to regularize the input of encoder layer,so as to adaptively correct the feature space and improve the training effect.In DSAE network,the fixed learning rate based on gradient algorithm leads to slow updating of network parameters and insufficient fitness.This thesis compares various gradient based optimization algorithms and selects Adam algorithm with adaptive learning rate to learn the network.Aiming at the problem of setting exponential decay rate of moment estimation in Adam algorithm,the adaptive mechanism is used to improve Adam algorithm,so that the new super parameter can realize adaptive change.Then,the improved Adam algorithm is used to optimize the weight distribution of BN-DSAE network.In order to test the effectiveness of the proposed method,the improved BN-DSAE network,the improved Adam algorithm with adaptive mechanism and the final BNDSAE fault diagnosis framework are verified on the motor rolling bearing data set on the MATLAB platform,and compared with DSAE and common learning classification algorithms BP and SVM,Experimental results show that the proposed method can effectively improve the classification accuracy and loss value of DSAE network,and the optimized DSAE framework is better than other deep learning methods and traditional machine learning methods.