MSGAFN: Multi-scale graph attention fusion network for machine fault diagnosis

Abstract

Fault diagnosis is essential in various fields, such as industrial manufacturing and engineering maintenance. Graph neural networks that take graph data as input can explore the relationships between data, which has strong feature expression capabilities. Currently, traditional fault diagnostic methods based on graph neural networks (GNN) are challenging to capture local and global feature information of data effectively. Most GNN models fail to consider the inherent differences between adjacent nodes, and they perform poorly in processing vibration signals in real-world industrial scenarios that commonly have strong noise. To resolve these concerns, this paper proposes a method for fault diagnosis based on a multi-scale graph attention fusion network (MSGAFN). In MSGAFN, data samples are constructed as clan graphs with multiple information scales to effectively represent local and global information of the graph structure data. Additionally, MSGAFN designed a new multi-scale feature fusion layer (MSFFL) to automatically learn the weights of adjacent nodes to represent their importance to the central node and reflect the differences between different adjacent nodes. This method is fully validated on bearing and gear datasets. The experimental results demonstrate that the proposed method exhibits excellent performance under conditions of imbalanced datasets and strong noise, providing a promising approach for bearings and gears fault diagnosis in real-world industrial scenarios with strong noise.