Fault Diagnosis Method Of Rotating Machinery Based On Graph Convolutional Network

Fault diagnosis is a key technology for ensuring the dependable,efficient and sustainable operation of industrial systems.As industrial informatization and intelligence develops,fault diagnosis methods are gradually moving from signal processing methods to deep learning methods.Graph convolutional neural networks(GCN)is a class of method developed based on deep learning to process graph data,which have better performance and interpretability.In this thesis,the collected sensor signals are modelled into fault sample graphs and three fault diagnosis methods for rotating machinery are proposed based on graph convolutional neural networks,the main work of that is as follows.(1)Unsupervised fault diagnosis method for motor rotors based on graph neural networks.The method firstly uses the Fourier transform to extract the frequency domain features of each fault sample of the motor rotor;secondly uses the K-nearest neighbour(KNN)method to construct a fault sample map based on the frequency domain features;uses the fault sample map as input,uses the graph sample and aggregate(Graph SAGE)algorithm to obtain the embedding of each fault sample,and finally uses the K-means clustering algorithm to obtain the class labels of each fault sample.The method enables fault diagnosis of motor rotors without labelled samples,thus helping experts to gain an initial understanding of the large amount of fault data and to tag the samples.In addition,the method constructs a fault sample map by mining the relationships between fault samples,thus providing additional features for fault diagnosis.(2)A semi-supervised learning motor rotor fault diagnosis method based on GCN under multiple distance measures.The method uses the KNN method based on multiple distance measures to construct a fault sample graph for motor rotor fault samples;using this fault sample graph as input,the semi-supervised learning algorithm of graph convolutional neural network is used to achieve the acquisition of labels for the fault samples under different scales of labelled samples.The semi-supervised fault diagnosis method not only avoids the problem of insufficient accuracy in unsupervised fault diagnosis,but also overcomes the problem of lack of labeled samples in the actual fault diagnosis.(3)An End-to-End semi-supervised learning method for rolling bearing fault diagnosis.The method firstly uses Pearson correlation coefficients to construct a fault sample graph for rolling bearing fault samples;secondly,a joint model based on convolutional neural network(CNN)and GCN is constructed to achieve classification of fault samples;finally,an End-to-End semi-supervised learning fault diagnosis framework is constructed.The method does not require feature extraction of the fault samples but directly constructs the fault sample graph,so the fault diagnosis results do not depend on the results of feature selection.The CNN is trained together with the GCN in this method to achieve high fault diagnosis accuracy with few labelled samples.