Research On Imbalanced Fault Data Classification Algorithm Based On Graph Neural Network

Most traditional classification models are based on the relatively balanced distribution of samples,but in practical applications such as fault diagnosis,the sample distribution is not balanced due to sampling bias and other objective reasons.In addition,with the development of deep learning,graphical data structures representing complex relationships and dependencies between objects are emerging in more and more applications.However,the traditional classification model doesn’t take into account the special structure of the graph,so directly using the traditional method to deal with the graph data structure doesn’t have much improvement effect.Graph Neural Network(GNN)has powerful representation ability.At present,the node classification task based on GNN mainly uses the nodes with known label information to predict the nodes with unknown label information for classification.The key to this process is to distinguish the value of different categories of samples,but most models cannot recognize the value of scarce categories.Due to the special structure of the graph,in addition to the imbalanced number of marked nodes,it also includes the imbalanced problem caused by the asymmetric distribution of topological attributes.Topological imbalance will affect the migration of decision surface,and then affect the classification effect of classifier.Based on the above problems,this paper proposes an imbalanced classification method based on GNN from two aspects: the imbalanced number of labeled nodes and the imbalanced topology.The main work of this paper is as follows:(1)From the perspective of the imbalanced number of annotation nodes,an attention network and cost-sensitive GNN algorithm(ACS-GNN)is proposed.By introducing the distance embedding vector and fusing it with the feature information of neighborhood nodes,the correlation between nodes is obtained.Secondly,the twolayer attention network is used to parameterize the correlation representation and the upper feature information of the unlabeled node to obtain the attention factor,and then the individualized weight is given to different types of nodes according to the attention factor.At the same time,cost-sensitive learning is used in the loss function to improve the misclassification cost of the minority class and improve the classification effect.The model is compared with different classification algorithms on imbalanced fault dataset as well as the generic dataset,and the results show that ACS-GNN has good classification performance.(2)An adaptive weighted GNN algorithm(TAW-GNN)based on topology perception is proposed from the perspective of topological imbalance.The soft labels of unlabeled nodes are obtained by label propagation and the topological influence matrix is calculated.Then the structure of the graph is analyzed,and the heterogeneity between the node pairs is calculated according to the connection of the nodes.Finally,cosine annealing mechanism is used to assign different training weights according to the heterogeneity of the nodes.TAW-GNN fully considers both topological imbalance and quantitative imbalance.TAW-GNN also demonstrates its effectiveness after comparing with graph-specific imbalance methods and general domain imbalance methods.