| Rolling bearing is an indispensable component of production system,and its health is of great significance in production.Finding the fault locations and the severity of the fault loss in time can not only minimize the economic loss,but also save valuable time and cost.However,the shallow model used in the traditional data-driven method is prone to problems such as poor self-learning ability and weak generalization ability.It is imperative to change the diagnostic model from ’shallow’ to ’deep’.Therefore,this article will summarize the previous research work and discuss three aspects:data feature extraction,parameter optimization,and fault diagnosis under variable conditions.Given the fact that rolling bearing in rotating machinery often operates in the variable loads and unstable environments.The deep learning model can extract the features of faults adaptively by using a multi-layer nonlinear mapping function.Therefore,a rolling bearing fault diagnosis method is proposed based on one dimensional convolutional neural network(1DCNN)and support vector machine(S VM).The general features of the original vibration signals are directly extracted as the input of the 1DCNN model to complete the CNN model training,and the traditional fully connection layer is replaced by the SVM classifier.This method fully combines the excellent small sample learning ability of the SVM and the powerful deep feature extraction ability of the CNN model.The test results of Case Western Reserve University dataset show that the proposed 1DCNN-SVM algorithm can effectively extract fault features and achieve higher classification accuracy compared with traditional fault diagnosis methods.Based on the fault feature classification of 1DCNN-SVM,an improved new particle swarm optimization algorithm INPSO is added to optimize the SVM parameters.The algorithm generates uniformly distributed initial particles through the traversal performance of Tent map and combines Levy flight to help particle swarm algorithm improve its global searching ability and premature phenomenon.Finally,the greedy algorithm is used to select the optimal particles in the particles,and then the SVM parameter value tuning is realized.Simulation experiments of several numerical functions show that INPSO algorithm has better search precision and speed.At the same time,bearing datasets from Case Western Reserve University and Jiangnan University are used for experimental verification of fault diagnosis.The results show that the 1DCNN-INPSO-SVM algorithm is efficient and accurate in the classification of bearing vibration signals.Under the actual working conditions,the traditional deep learning algorithm is not suitable for solving the problem of different data distribution of training data and test data.In view of this,the transfer learning is introduced into the field of fault diagnosis.The 1DSECNN-JDOT model is established based on the one dimensional squeeze-and-excitation network(1DSECNN)and joint distribution optimal transport(JDOT).Firstly,the algorithm uses the source data to train the one dimensional squeeze-and-excitation network to generate the pseudo-label of the target domain samples,and then uses the optimal transport algorithm to fine-tune the source domain data classification loss and the domain classification loss simultaneously.The model is tested on the rolling bearing dataset of Case Western Reserve University,and the results show that the algorithm can correctly classify the samples in the target domain and achieve high accuracy with less data. |
