Fault Diagnosis Method Based On Over-smoothing Relief Graph Convolutional Network And Its Application

Due to the complex working environment,long-time operation and other factors,faults occur inevitably in an industrial system,which may lead to the shutdown of the system,or threaten the life safety of operators.Therefore,the research on effective fault diagnosis methods is of great significance to ensure the safe and stable operation of the system.With the improvement of data perception,processing and storage capacity,data-based fault diagnosis method has become a research hotspot in the field.Industrial data is generally divided into Euclidean structure data and non-Euclidean structure data.The traditional Euclidean structure data-based methods are difficult to directly use non-Euclidean structure data for fault diagnosis,which restricts the improvement of fault diagnosis ability.Graph convolutional network is a common model to deal with this kind of data,but the model has the problems of over-smoothing and weak interpretability.Therefore,based on the widely existing non-Euclidean structure data in industrial systems,this thesis carries out the research on fault diagnosis method based on over-smoothing relief graph convolutional network.The main innovative work is as follows:(1)For the over-smoothing problem of graph convolutional network,an over-smoothing relief graph convolutional network based on the influence of adjacent nodes is proposed.The network introduces the influence coefficient of adjacent nodes,which reflects the influence of all adjacent nodes in the process of data aggregation in graph convolutional network.A coefficient optimization method based on optimization algorithm is proposed,the dynamic adjustment of data aggregation process is realized,and the problem of over-smoothing is alleviated.The effectiveness of the proposed method is verified by simulation experiments.(2)For the problem of weak interpretability of graph convolutional network,a class activation mapping method-based interpretability analysis method of graph convolutional network is proposed.This method constructs the data classification module to complete the classification task of the input data,and constructs the class activation mapping method to analyze the interpretability of the classification results.The effectiveness of the proposed method is verified by simulation experiments.(3)For the problem that the traditional data-based fault diagnosis method is difficult to directly deal with the non-Euclidean structure data of industrial systems,an over-smoothing relief graph convolutional network-based fault diagnosis method is proposed.Using structural analysis,this method constructs the association graph integrating system structural knowledge,realizes the effective fusion of data and knowledge,and designs the over-smoothing relief graph convolutional network-based fault diagnosis module and the interpretability analysis module of the diagnosis results.The effectiveness of the proposed method in fault diagnosis of the high-speed train traction drive system is verified by a hardware-in-the-loop simulation platform.