摘要详情

ID / 提交时间

213 / 2025-06-14 14:56:57

标题

Fault Diagnosis Using Graph Neural Networks with Dynamically Updated Edge Weights via Metric Function Fusion

关键字

fault diagnosis, graph neural networks (GNN), multimodal attention

主题及专题

4、System Monitoring, Fault Diagnosis and Prediction

状态

全文待审

作者

Lin Lin / Beijing University of Chemical Technology

Huaqing Wang / Beijing university of chemical technology

Wang Siyuan / Beijing University of Chemical Technology

Zesheng Lin / Beijing University of Chemical Technology

Liuyang song / Beijing university of chemical technology

摘要

Rotating machinery equipment exhibits diverse failure modes under prolonged operating conditions, posing significant challenges for fault diagnosis. While Graph Neural Networks (GNNs) have demonstrated notable success in addressing complex fault problems, their edge weight calculations often rely on a single metric function, limiting their ability to capture multi-physics coupling characteristics. Consequently, computing appropriate edge weights has garnered increasing attention. Fusing multiple metric functions for edge weight calculation enables the capture of multi-dimensional relationships between nodes and enhances the robustness of the weighting process. Furthermore, incorporating a dynamic computation mechanism based on node features allows for adaptive edge weight updates, thereby improving the model's expressiveness and adaptability. This paper proposes a GNN-based fault diagnosis method for rotating machinery that dynamically updates edge weights through metric function fusion (DMF-GNN). First, datasets are constructed using vibration signals collected from rotating machinery. During the creation of input samples for the GNN, edge weights are dynamically computed and updated by fusing multiple metric functions and introducing a node-feature-based dynamic mechanism. Subsequently, propose the Multi-Layer Perceptron Enhanced Edge Aggregated Graph Attention Networks (ME-EGAT) method. EGAT integrated with (MLP) modules is employed for deep feature extraction, ultimately classifying the operational states of the rotating machinery. Finally, the proposed method is empirically validated on the Paderborn University bearing dataset. Experimental results demonstrate that the method effectively implements dynamic edge weight updates based on metric function fusion within a GNN framework, leading to enhanced diagnostic performance. It provides an efficient solution integrating multi-source information with stronger adaptability for diagnosing complex faults characterized by difficult-to-capture multi-physics coupling effects.