A Prior-Aware Spatio-Temporal Graph Neural Network for Fault Diagnosis of Mechanical Equipment With Multisensor Data

Abstract

Fully leveraging spatio-temporal correlations in multisensor data is crucial for achieving accurate fault diagnosis of mechanical equipment. Although spatio-temporal graphs (STGs) have demonstrated potential in modeling such correlations, existing methods still suffer from limitations, such as a single-perspective feature extraction paradigm and a lack of prior knowledge guidance, which consequently restrict their performance. To address these challenges, this article proposes a novel prior-aware spatio-temporal graph neural network (PA-STGNN). First, an adaptive multidomain feature extractor is designed to fuse complementary information from multiple domains, thereby generating more informative node representations. Subsequently, a prior-aware STG construction strategy is proposed. Specifically, an initial STG is constructed to capture both spatial correlations via a Gaussian kernel k-nearest neighbors approach and temporal correlations using a multihead attention mechanism. Furthermore, a prior matrix is designed to inject physical prior knowledge about temporal decay and sensor autocorrelation into the graph to optimize its topology. Finally, Chebyshev graph convolution is employed to aggregate information from multihop neighborhoods, enabling the effective use of complex spatio-temporal correlations. Comprehensive experiments on a public dataset and a real-world coal mill dataset demonstrate that PA-STGNN achieves superior diagnostic accuracy, consistently outperforming state-of-the-art baselines, particularly in few-shot scenarios.