Spatiotemporal graph neural diagnostics for rotating machinery: a single-sensor multidimensional fusion framework

In industrial equipment condition monitoring, utilizing tri-axial vibration data acquired from a single sensor poses a critical challenge due to hardware and space constraints. Traditional methods often treat multiaxis signals as independent channels or simple concatenations, neglecting the underlying dynamic coupling mechanisms and failing to fully mine latent fault features. To address this, this article proposes a novel fault diagnosis method for rotating machinery based on space-time decoupling graph construction and graph convolutional networks (GCNs). Addressing the topological constraints of single-sensor scenarios, the present article proposes a time-point-based graph construction strategy that shifts from the conventional sensor-as-node paradigm to the reconstruction of high-dimensional dynamic graphs from time-series data. Adhering to the space-time decoupling principle, the method employs local sparse connections to constrain temporal evolution and utilizes Euclidean distance within the tri-axial feature space to quantify state differences; this mechanism facilitates the adaptive transformation of fault-induced transient impulses into high-weight graph edges. By leveraging a GCN model to aggregate nonlinear spatiotemporal features, the proposed method demonstrates superior performance, yielding an average recognition accuracy of over 99% on standard bearing fault datasets. These results validate that the proposed graph construction strategy can effectively capture dynamic correlations and weak features without relying on multisensor arrays, offering a promising solution for equipment maintenance under hardware constraints.