Fault Diagnosis Method Of Rolling Bearing Based On SVM And CNN

With the development of the modern industrial system,the safety of industrial equipment has become the focus of attention.Through fault diagnosis,engineers can judge the operating status and abnormal conditions of the system,and find potential safety hazards at an early stage,thereby avoiding risks.In recent decades,fault diagnosis research has attracted the attention of many domestic and foreign scholars,and a series of fault diagnosis methods have been proposed.With the rapid development of data acquisition and storage technology,current data generally have the characteristics of large capacity and high speed,so data-driven fault diagnosis methods have emerged.The data-driven fault diagnosis method is mainly embodied in the diagnosis method based on machine learning and the diagnosis method based on deep learning.At present,the diagnosis method based on machine learning is very mature,but for support vector machines(Support Vector Machines,SVM)in machine learning,it is difficult to extract features and hyperparameters are difficult to select.Signal feature extraction,dimensionality reduction processing,and parameter optimization can also be used.A combination of algorithms and other methods to solve the problem.But there is still a difficulty in that the random selection of the parameters of the optimization algorithm results in the performance of the optimization algorithm is not very stable,and the support vector machine can only solve the binary classification problem.Based on this,this article improves the related optimization algorithms to avoid random selection of parameters from affecting the performance of the optimization algorithm.At the same time,the optimization algorithm is combined with the support vector machine method of acyclic and acyclic graphs to accurately realize the multivariate classification of faults.Compared with machine learning,deep learning can adaptively extract signal features,but its disadvantage is that the model requires a large amount of data and a large amount of model parameters,so the model efficiency is low.Based on this,this article improves the Convolutional Neural Networks(CNN)method in deep learning,and combines the parameter optimization algorithm improved in this article.Experiments show that the improved model improves the accuracy and efficiency of fault diagnosis..The main research content and innovation points of this paper are as follows:(1)This paper first extracts the time-domain signal feature of the original vibration signal,and secondly uses Principal Component Analysis(PCA)to reduce the dimensionality of the signal feature.Then,in view of the defect that the SVM can only do binary classification in the existing algorithm,it adopts The Directed Acyclic Graph Support Vector Machines(DAG-SVM)model realizes the multivariate classification of fault diagnosis problems.Aiming at the problem of difficult selection of hyperparameters in SVM,combined with Improved Particle Swarm Optimization(IPSO)to automatically find suitable hyperparameters.IPSO improves the selection of learning factors 1,2 and inertia weighting factor ω in Particle Swarm Optimization(PSO)by introducing appropriate parameter setting functions,thereby automatically adjusting these three factors.The simulation results show that the proposed PCA-IPSO-DAGSVM algorithm has corresponding advantages in precision,recall,accuracy,and F metric value,and it has been processed by time-domain signal feature processing and PCA dimensionality reduction processing.The data effectively removes invalid features and has strong anti-interference performance.Therefore,the effectiveness and superiority of the proposed PCA-IPSO-DAGSVM algorithm are proved.(2)Although the PCA-IPSO-DAGSVM algorithm has achieved good diagnostic results,for the problem that the signal extraction process in machine learning is too cumbersome,some scholars have proposed an adaptive feature extraction CNN algorithm.This algorithm does not need to manually extract the features,and the data can be imported into the algorithm to adaptively extract the features and output the fault diagnosis results.But the difficulty of CNN is:1)It needs a lot of data support,so the efficiency of the convolutional neural network algorithm is insufficient,and the training speed is too slow;2)The selection of some parameters in the convolutional neural network has a great impact on the diagnosis result,so it needs to be suitable Optimize parameter method.In response to the above problems,this paper proposes an improved CNN fault diagnosis model based on IPSO optimization.IPSO is an improved algorithm for traditional PSO.For the traditional fully connected layer in the CNN model,global averaging technology is introduced to replace it.Compared with the traditional CNN,the improved model requires fewer training parameters,and the improved model is more efficient.Experimental results show that this method can ensure the efficiency of fault diagnosis and improve the accuracy of diagnosis under the condition of various types of faults.