Fault diagnosis method based on multi-scale adaptive super-graph convolutional neural networks

Abstract

In the research on fault diagnosis using deep learning, there is a lack of effective cross-scale feature association modeling, which overlooks the potential correlations between signals at different scales, resulting in weak generalization capabilities. This paper proposes a multi-scale feature construction space Supergraph integrated with physical prior knowledge, and based on this method, introduces a Multi-Scale Adaptive Supergraph Convolutional Neural Network (MS-ASGCN). The vibration signals are decomposed into different frequency bands, and features at various scales are extracted while incorporating prior knowledge. Local topological graphs are constructed using various weighted metrics, and these local graphs together form a spatial Supergraph, which serves as the input for the model. A dual-channel Graph Convolutional Network (GCN) is employed to extract features, and an attention mechanism is introduced to adaptively assign weights to different channels, achieving deep feature fusion. Experiments on two benchmark datasets demonstrate that MS-ASGCN effectively improves model accuracy and exhibits good stability and generalization capabilities.