ATGCN: An Adaptive Temporal-Topological Graph Convolution Network With Nodal Attention for Robot Fault Diagnosis

The monitoring of wheeled robot health through intelligent fault diagnosis (IFD) is crucial for ensuring operational reliability. While traditional deep learning (DL) approaches to IFD are effective, the recent advancements in graph neural networks, such as graph convolution networks (GCNs), offer an enhanced ability to extract relational information among multisensor data, thus improving the accuracy of robot fault diagnosis. However, existing GCN-based IFD methods have limitations, particularly in their ability to concurrently process temporal dynamics and topological structures inherent in robot fault data. Furthermore, the construction of adjacency matrices in GCNs typically requires calculating the similarities from input data or using predefined mathematical models, which depend on domain expertise and lack adaptability across different robotic platforms. To address these challenges, an adaptive adjacency matrix-based temporal-topological graph convolution network with nodal attention (ATGCN) is proposed in this article, which contains a series of stacked temporal-topological graph convolution blocks designed to discern intricate features within multisensor data. In addition, to better capture the correlative information between disparate sensors, we design a nodal attention mechanism within the topological feature extraction process. Moreover, a graph learning module is proposed for constructing the adjacency matrix, which enhances model adaptability in diagnosing faults on different robot platforms. Through comparative experiments conducted on two real-world wheeled robots, the proposed ATGCN demonstrates notable improvements in terms of diagnostic accuracy and adaptability compared with the state-of-the-art methods.