Bearing Fault Detection Method Based On Deep Clustering Ensemble Key Technology

Bearings are one of the most critical components in most mechanical equipment.Effective bearing failure detection is very important to maintain the safety and proper operation of mechanical equipment.Mechanical big data is characterized by large size,diversity,and high speed.Also current fault detection methods usually require collecting data under fault conditions and training them before they can be applied to real industrial environments for fault detection.The problem with this approach is that the performance of the model may be affected if the detected faults are different from the fault conditions under which the model was previously trained.In addition,data collection and training are time and resource intensive.In order to solve this problem,this paper draws on the mainstream algorithms of clustering analysis,clustering ensemble and deep clustering at home and abroad.Based on the relevant theories in clustering analysis and deep learning,and taking mechanical bearings as the research object,we carry out the research on the bearing fault detection method based on the key technology of deep clustering ensemble,mainly focusing on the fault detection research of unsupervised method from the selection of the number of clusters of clustering members and the pre-training of deep clustering model,which is mainly centered on the situation that the data labels are sometimes unavailable in the detection of bearing faults.It is hoped to provide an efficient and accurate unsupervised method for bearing fault detection.This will help reduce the time and resource costs of data acquisition and training,and improve the robustness and adaptability of the fault detection model.The main contributions and innovations of this paper are as follows.(1)The choice of the number of clusters k has a significant effect on the results of clustering ensemble.Different choices of k produce completely different results.In order to determine the most appropriate strategy for selecting the number of clusters for clustering ensemble,a variety of commonly used selection methods were compared.The selection of five commonly used methods for determining the number of clusters of clustering members showed that the best results were obtained when the number of clusters was equal to the true number of categories k*.In order to further investigate whether there exists a better range of choices for the number of clusters,groups of comparative experiments were conducted,choosing six shorter intervals from k* to 2k*and comparing the results of clustering ensemble with those obtained using k*.The results of the experiments show that clustering ensemble works best when the number of clusters is k*,which provides a basis for k value selection in clustering ensemble studies.(2)Deep clustering algorithms have gained breakthroughs in clustering performance by jointly optimizing deep embedding feature representations.This paper uses contrast learning to pre-train the encoder by automatically constructing similar and dissimilar instances so that similar instances are closer together in the feature space and dissimilar instances are farther apart,thus allowing the encoder to generate parameters for clustering preferences.The trained encoder is migrated to a deep learning task to jointly learn clustering-oriented features and optimize the assignment of clustering labels by simultaneously minimizing clustering loss and reconstruction loss.The clustering performance of the proposed method and multiple algorithms are compared on six image datasets.Experimental results show that deep clustering with the addition of contrast learning pre-training is satisfactory compared to popular deep clustering algorithms.Also,this method was validated on the bearing dataset.