Research On Deep Few Shot Learning For Intelligent Fault Diagnosis Of Bearing

Advanced industrial equipment places extremely high demands on rotating machinery in terms of safety and reliability.Bearing is one of the important components of rotating machinery system.Once the failure occurs of bearing,it may cause the paralysis of the entire mechanical system,and finally lead to inestimable losses.Due to the suddenness and uncertainty of mechanical failures,the collected fault data is limited,so there is no large batch of data for training model in practical application.In order to effectively relieve the contradiction between the lack of fault samples and high-accuracy fault diagnosis,this paper studies the diagnosis method under small samples to provide important technical support for the intelligent fault diagnosis and the operation of equipment.Firstly,in order to effectively solve the problems of data scarcity and data distribution bias,this paper adopts the feature transfer method to achieve fault diagnosis across working conditions under small samples.Mainly through the combination of residual dense network and attention mechanism module to deeply extract fault feature information,and then the multi-kernel maximum mean discrepancy is introduced which self-adaptively solves the domain adaptation and extracts domain invariant feature signals to improve the performance of feature extraction.Through public data set,it’s proved that this method can achieve efficient and fast fault diagnosis.Secondly,based on the theory of meta-learning,a metric-based meta-learning model has discussed about fault diagnosis using a number of samples across working conditions.It mainly constructs five layers of convolutional neural networks with different kernel sizes to extract feature information,and the Squeeze-and-Excitation module is added to enhance the channel information of convolution features and assign different weights to different fault features.Then the metric network is used to convert the fault features into the metric features by mapping,and the cosine distance is adopt to achieve the matching of the metric features,finally the fault type is classified by high-accuracy.In order to find the optimal model parameters by analyzing the factors that affect the accuracy of fault diagnosis.Compared with other conventional methods,it’s confirmed that the proposed method has great fault diagnosis ability.Finally,a feature refinement model for intelligent fault diagnosis of few samples is studied based on generative adversarial network to solve the problem of the data set bias.Synthetic data is generated by the generative network,and then the synthetic data and real data are sent to the feature refinement model to produce the class correlation information.This module can integrate class semantic information and fault feature information into a generation model,and at the same time,in order to extract fully refined features,the feature refinement model is constrained by using double loss functions,and finally achieves high-accuracy recognition.