An Adaptive Tri-Domain Spatiotemporal Graph Fusion Framework for Few-Shot Multisensor Fault Diagnosis of Rotating Machinery

Abstract

Few-shot fault diagnosis of rotating machinery is challenging due to limited labeled data and the underutilization of multidomain information from multisensor systems. To address this issue, we propose an adaptive tri-domain spatiotemporal graph fusion (TTGF) framework. The proposed method constructs spatiotemporal graphs from the time, frequency, and time–frequency domains, enabling effective modeling of spatial correlations among multisensor signals and parallel extraction of discriminative spatiotemporal features. Before graph construction, a cross-domain attention enhancement module enriches each domain’s node representations by selectively incorporating complementary information from the other two domains, enabling cross-domain-aware graph topologies while preserving domain-specific structural diversity. By leveraging graph convolution enhanced with learnable spatiotemporal embeddings, together with an adaptive multiclassifier gating fusion mechanism to mitigate overfitting, the framework is particularly suitable for few-shot fault diagnosis. Extensive experiments on benchmark datasets and a real industrial ball mill drive system further validate the effectiveness and engineering applicability of the proposed method.