Study On Fault Diagnosis Method Of Industrial Process Based On Generalized Zero-shot Learning

With the further advance of intelligent manufacturing,industrial processes are becoming more complex.At this time,the fault types of processes and equipment will become more diversified,and higher requirements for diagnosis methods are put forward.Traditional fault diagnosis methods based on supervised learning rely on a large number of labeled fault samples,which is difficult to meet in industrial processes,especially some zero-sample but know faults(unseen faults),which cannot be handled by supervised learning.Zero-shot Learning(ZSL)method has been proved to classify unseen faults,that it only has seen fault(fault types with samples participating in model training)samples in the training stage.However,ZSL fault diagnosis method can only classify unseen faults when it is in the testing stage,and it cannot classify the cases that both seen faults and unseen faults may occur,which seriously limits its practical application.To solve the above problems,it is necessary to extend the ZSL method,which is generalized zero-shot learning(GZSL),so that it can classify both seen and unseen faults in the testing stage.In this thesis,introduce GZSL into the field of industrial process fault diagnosis.By introducing semantic description information,GZSL can establish the connection between the disjoint seen faults and unseen faults,so that it can classify both seen faults and unseen faults in the testing stage.However,the fault diagnosis method for GZSL still faces key technical difficulties such as domain bias,semantic bias,and end-to-end difficulty.This thesis is devoted to solving the above difficulties,and the main contents are as follows:1.To solve the domain bias problem that unseen fault samples are incorrectly classified into seen faults in the testing stage,a method of generating pseudo samples through generate model is proposed to transform the domain discrimination of seen faults and unseen faults into a binary classification problem.The samples identified as seen faults are classified by supervised learning.The samples identified as unseen faults are classified by ZSL.The GZSL problem is transformed into supervised learning problem and ZSL problem.2.To solve the semantic bias problem that the artificial definition of semantic description information subjectivity is too strong.a method to correct the semantic description information of seen faults by predicting the error rate of attributes,to reduce the influence of human subjectivity on the results and improve the accuracy of model classification.3.To solve the problem of too many modules needed for GZSL,an end-to-end fault diagnosis framework for GZSL is proposed.In this framework,GZSL can be realized only with one type of binary classification algorithm.In addition,the end-to-end model is divided into several stages by some known rules,and all stages are completed by a binary classification algorithm,so that the endto-end model has a strong interpretability,and the method is more suitable for the field of fault diagnosis.