Research On Rolling Bearing Fault Diagnosis Based On Optimized Capsule Network

With the development and application of intelligence in the industrial field,the high efficiency work of rotating machinery and equipment has brought great benefits to industrial manufacturing production.And rolling bearings as an important part of the equipment,when its components fail,it will cause the equipment to stop running,thus affecting production efficiency,economic benefits decline,etc.Therefore,real-time monitoring of the operating status of bearings and effective fault identification and diagnosis of key components is of great significance.In practice,the operating condition of the bearing is not only under a single operating condition,but more than not the bearing is under changing conditions,making the collected data complex and diverse,which can easily cause incomplete feature extraction when extracting the fault characteristics,which will lead to the data features classification recognition effect is not obvious,can not make an accurate judgment on the bearing fault category.To this end,this paper focuses on how to improve the classification performance and diagnostic accuracy of the model in the case of single and variable operating conditions,and the main work is as follows:(1)To address the problem of incomplete feature extraction of traditional convolutional neural networks,a rolling bearing fault diagnosis method based on Bayesian optimization Capsule Network(BO-Caps Net)is proposed.A large convolutional kernel is designed in the capsule network to initially extract the signal features to increase the perceptual field of the model,and a Bayesian optimization algorithm is used to quickly find the optimal solution of the network to improve the diagnostic efficiency of the Capsule Network.In this paper,experimental simulations and performance evaluation of the model and algorithm are carried out on bearing data from Case Western Reserve University,and the diagnostic results are compared with those of Wide Kernel Convolutional Networks,Capsule Networks and Adam’s optimisation algorithm.The experimental results show that the method achieves an accuracy of97.8% for bearing fault diagnosis and has better classification and identification ability than other methods.(2)To address the problem of low accuracy of fault feature extraction under variable working conditions,a fault diagnosis method based on Self-Calibrated Convolutional Capsule Network(SC-Caps Net)for variable working conditions of bearings is proposed.SelfCalibrating convolutional blocks are added to the capsule network to extract richer and more closely related features,and Local Maximum Mean Difference is used as the functional loss term to adjust the distribution of sub-domains of the same class to make the model more effective in classification.Experimental analysis of accuracy and generalisation on a dataset of bearing faults under time-varying speed conditions at the University of Ottawa has shown that the proposed method is more stable in dealing with different operating conditions,has at least5% better accuracy than the SC-Res Net,Res Net and Cap-Res Net models,and has better performance in terms of generalisation capability.(3)To address the problem that the parameters are not assigned weights in the SelfCalibrating convolutional block,which leads to the absence of key information of the features,the Split Attention(SA)mechanism is introduced into the above SC-Caps Net model to assign different weights for the different features extracted by the model to further improve the fault identification accuracy of the model under variable working conditions.The proposed model SC-SA-Caps Net achieves 96.5%,96.4%,95.7% and 96.2% diagnostic accuracy under four sets of different working conditions respectively.