Research On Industrial Fault Diagnosis Method Based On Few Shot Learning

With the arrival of Industry 4.0,advanced technology technologies such as big data,cloud computing and 6G are gradually becoming the digital pillars of industry and infrastructure.Facing the complex and changing industrial scenes,fault diagnosis technology is also developing in the direction of informatization,digitalization and intelligence.Intelligent fault diagnosis technology based on deep learning algorithm has been widely used in industrial scenarios with great success,however,for limited fault data samples,deep learning algorithm seems to lack excellent results.How to achieve accurate fault diagnosis with small samples is an essential part of the industrial scenario.In this paper,we focus on the important technique of fault classification in intelligent fault diagnosis technology,and conduct further research on its algorithmic model and practical application based on the few shot learning algorithms that have been developed in full swing in recent years.The main research contents and results of this paper are as follows:(1)To address the problem that existing fault datasets are rich in classes but few samples of each type,which is not conducive to deep learning algorithms to effectively mine the implicit features in them,an intelligent fault diagnosis method based on semisupervised improved prototypical network(SSIPN)is proposed.Specifically,the proposed SSIPN first obtains the feature representations of the samples through the feature extraction module and inputs them into the weight module for processing,and then the sample prototypes computed by the prototype network are passed through the weight module and the prototype optimization module to obtain a weighted class prototype,respectively,where the weight module assigns different weights to the intraclass instances and the prototype optimization module is used to enhance the representativeness of the class prototypes.Then,SSIPN uses a scale scaling strategy learning method to scale the Euclidean distance between the prototype and the query samples by the distance scale module to maximize the inter-class differences while minimizing the intra-class differences.Finally,the scaled distances are processed with a softmax classifier to obtain the probabilities that represent the classes to which the samples belong.The proposed SSIPN achieves excellent results on both CWRU dataset and SQ dataset,not only with excellent fault recognition rate but also with excellent performance in the speed of fault recognition.(2)Since in many industrial scenarios,equipment often operates under different loads or operating conditions,which means it is difficult or even impossible to obtain training datasets with the same distribution as the test dataset before building a diagnostic model,traditional deep learning algorithms are difficult to solve the task of condition migration.A few shot intelligent fault diagnosis method based on improved meta relation network(IMRN)is proposed to address the cross-domain problem under small-sample conditions.The IMRN mainly consists of a multi-scale feature extractor and a metric learner.The multi-scale feature extractor consists of a two-channel onedimensional convolutional neural network,which maximizes the features of the original fault signal at different scales by setting different convolutional kernel sizes.Meanwhile,the weight parameters of this feature extractor are obtained by supervised learning strategy trained from the source domain dataset to make full use of the semantic information containing rich substance in the source domain.Second,the powerful learning capability of the metric learner is utilized to perform similarity metrics,and a label smoothing algorithm is used to alleviate the possible model overfitting problem of the metric learner.Through experimental validation on three public datasets(TE,PU,CWRU),IMRN is able to accurately identify fault classes with extremely small samples,which provides a feasible solution to address few shot cross-domain fault diagnosis.