Spatio-Temporal Graph Neural Network for Fault Diagnosis Modeling of Industrial Robot

摘要

Industrial robots play a critical role in modern industrial production. They are widely used in tasks requiring high precision and efficiency, such as manufacturing, assembly, and material handling. However, since industrial robots often operate in complex and changing environments, long-term use inevitably leads to issues such as component aging, wear, or other potential faults. These faults not only reduce production efficiency but may also cause equipment damage, production interruptions, or even safety risks. Industrial robot systems are typically composed of multiple highly interconnected components and sensors. Graph Neural Networks (GNNs), which can effectively model multivariable data, have been widely applied to modeling and analyzing such systems. However, different methods of graph construction vary significantly in their ability to capture system structures, dynamic relationships, and multivariable interactions. The impact of these differences on downstream fault diagnosis tasks remains an area that requires further research. To address these challenges, this study proposes a GNN-based fault diagnosis framework to demonstrate the practical effects of different graph construction methods. First, we transformed the state monitoring data of industrial robots into three different graph structures using KNN, radius, and path-based methods. Then, we used a graph attention network to capture the spatial dependencies among various variables. At the same time, a parallel encoder with diagonal masking self-attention (DMSA) was designed to model temporal dependencies. A spatiotemporal attention module was then applied to extract both spatial and temporal features. Finally, the type of fault present in the data was determined. Experimental studies based on real-world industrial robot datasets show that different graph construction methods significantly influence fault diagnosis accuracy. The proposed framework achieved diagnostic accuracies of 87.11%, 88.85%, and 92.68% under the three graph construction methods, respectively.