Research on Representation Learning Methods for Fault Detection in Complex Mechanical Systems Based on Graph Neural Networks

Abstract

With the increasing complexity and intelligence of modern mechanical systems, traditional fault detection methods face significant challenges in capturing nonlinear relationships, multi-source information interactions, and complex degradation patterns. This study proposes a graph neural network (GNN)-based representation learning framework for fault detection in complex mechanical systems. By constructing graph-based representations of mechanical components and their dynamic interactions, the proposed method effectively integrates heterogeneous sensor signals and extracts high-level fault-related features. A graph convolutional network architecture is developed to learn spatial and temporal dependencies among system components, enabling accurate identification of abnormal operating conditions. Furthermore, an adaptive feature aggregation strategy is introduced to enhance the representation capability of the model under varying working conditions. Experimental results demonstrate that the proposed GNN-based approach achieves superior fault detection performance compared with conventional machine learning and deep learning methods, particularly in scenarios involving complex coupling relationships and limited fault samples. The proposed framework provides an effective solution for intelligent health monitoring and predictive maintenance of advanced mechanical systems..