“…To prove the effectiveness of the proposed ASG-HOMGATbased fault diagnosis method, the standard graph convolutional neural network (GCN) [33], the multiple-receptive field graph convolutional neural network (MRF-GCN) [24], the graph attention network (GATv2) [41], the GraphSAGE [42], the three kinds of Chebyshev graph convolutional networks with receptive fields k of 1,2,3, respectively [43], the GC-ResCNN [44], the ASG-HOGCN (not using multi-head attention mechanism network) for comparative experiments, these models are able to capture the structural relationship between samples and achieve effective fault diagnosis, GCN can effectively aggregate the information of neighboring nodes in the graph structure through spectral convolution, which in turn improves the ability of fault feature extraction, MRF-GCN helps the model to analyze the fault features in a more comprehensive way by converting the data samples into a weighted graph and learning the feature representations from multiple neighborhoods, GATv2 uses the attention mechanism combined with assigning different weights to neighboring node in the graph-structured data, GraphSAGE generates new feature representations for each node by learning an aggregation function to aggregate the neighboring features of the nodes, Chebynet by using chebyshev polynomial expansions of different orders, ChebyNet can flexibly handle graphs with different sensory wild sizes of the data and enables efficient graph convolution operations through approximate computation.…”