Research On Few-shot Bearing Fault Diagnosis Based On Graph Convolutional Network Metrics And Self-Training Data Enhancement

As one of the most common and easily damaged components of rotating machinery,it is particularly important to detect and diagnose faults in rolling bearings.Fault diagnosis methods based on deep neural networks usually require a large number of samples to train the network,and in actual industrial production,the fault data that can be collected is very limited,resulting in the performance of such methods in practical applications.In this paper,the theory of few-shot learning is introduced,and the intelligent fault diagnosis method for rolling bearings with few samples is investigated from the perspectives of metric-based learning and data augmentation,respectively.Firstly,for the problem of few labeled samples and complex and variable mechanical equipment working conditions,this paper introduces metric-based meta-learning methods into the fault diagnosis process and investigates a few-shot fault diagnosis method based on graph network metrics.The method uses a convolutional neural network to build a feature extraction network and a graph convolutional network to build a metric network,which solves the problem of low generalization of traditional metric-based methods using fixed metric functions.With the help of the auxiliary data set to learn rich prior knowledge,the network model of the studied method can achieve high accuracy in identifying rolling bearing fault samples with different working conditions under few samples.Secondly,to address the problem that most of the current few-shot fault diagnosis methods do not make sufficient use of unlabelled samples,this paper investigates a small sample fault diagnosis method based on self-training data enhancement.The method is based on the self-training method to fully exploit the hidden information of unlabelled fault samples,and incorporates a pseudo-label reliability assessment mechanism for the pseudolabel selection problem in the self-training process to avoid the negative impact of unreliable pseudo-labels on the model training.The network model of the studied method is pre-trained using the auxiliary data set and self-trained using the test data set,and can achieve bearing fault diagnosis in few sample scenarios.The few-shot fault diagnosis method investigated in this paper is experimentally validated on several bearing datasets and the influence of key parameters on the experimental results is investigated.After comparison with other classical methods,the effectiveness of the research method in this paper is shown.