Research On Rolling Bearing Fault Diagnosis Method Based On Semi-supervised Learning

Rolling bearing is one of the important components of rotating machinery,and its good working condition is the premise of normal operation of equipment.Because it often operates in high temperature,high speed and heavy load and other harsh working conditions,rolling bearings will inevitably have a variety of failures,light will affect the working performance of the equipment,heavy will cause huge economic losses and accidents.Therefore,effective health monitoring and fault diagnosis of rolling bearings have positive research significance for ensuring the safe and stable operation of mechanical equipment.However,when rolling bearings fail,relying only on manual fault analysis requires rich prior knowledge of detection personnel,but the efficiency is insufficient.The existing diagnostic models have some problems,such as complex network,many training parameters and large model size.Due to the "black box" feature of deep learning,the diagnosis results of general diagnostic models cannot be visually displayed as the fault causes of signal analysis methods.When there is insufficient label sample data,the prediction performance is reduced due to the insufficient generalization performance of the diagnostic model.In view of the above problems,this paper carries out relevant research,and the main research contents are as follows:(1)Aiming at the large size of the diagnosis model,low training efficiency and "black box" problems of deep learning,a fault diagnosis model based on improved Mobile Net V3 is proposed.By replacing the point convolution function with BFT module,the model removes the redundant lifting and lifting operations in the classifier,optimizes the classifier structure,and reduces the model parameters and volume without affecting the model accuracy.At the same time,the fault characteristic frequency and Deep SHAP algorithm are combined in the diagnosis network to deduce the classification reason of the deep learning model.The feasibility of the algorithm is verified by experiments with multiple data sets and different fault stages.(2)Aiming at the problem that the predictive performance of the diagnostic model decreases when there are insufficient label samples,a CL-SSL fault diagnosis model based on semi-supervised learning is proposed.Based on the consistent regularization method,CL-SSL uses the idea of course learning and Bayesian inconsistent active learning to select unlabeled samples with high confidence from easy to difficult to add to the labeled training set,making full use of a large number of unlabeled samples to improve model performance.(3)In order to realize the identification of new compound fault types of rolling bearings without retraining the whole model,a compound fault diagnosis model based on semi-supervised learning ML-SSL is proposed.Firstly,a comparative learning method is used to obtain the highly differentiated surface features of all unlabeled samples unrelated to the classification task.Then Transformer is used as a multi-label classifier to fine-tune a small number of labeled samples.In the proposed method,multiple label classifiers are used to identify compound faults.New compound fault types can be identified by the combination of original labels without adding new fault categories and retraining.Since the feature extraction layer and classifier realize decoupling in training,when the existing tag combination cannot represent the new compound fault,adding a new tag type only needs to retrain the classifier,which takes very little time.