Towards spatial-temporal meta-hypergraph learning for multimodal few-shot fault diagnosis

Fault diagnosis is essential for maintaining equipment safety and reliability in smart industrial environments. Early identification of issues through intelligent maintenance systems helps prevent downtime, enhance productivity, and mitigate hazards. However, two major challenges exist: first, when machines exhibit faults, they are typically deactivated for safety, resulting in scarce fault data; second, existing methods disregard high-order relationships between working conditions, while failing to simultaneously consider signal heterogeneity and spatial–temporal correlations. To address these challenges, we propose a spatial–temporal meta-hypergraph learning for multimodal few-shot fault diagnosis (MetaSTH-FD) by integrating dynamic spatial–temporal hypergraph construction into meta-learning. The framework first decomposes vibration signals into multimodal features, then constructs hypergraphs to capture complex relationships. Our approach enables quick adaptation to new conditions with limited samples, while the hypergraph structure models complex relationships in multimodal signal data. Experimental results demonstrate significant performance improvements across various working conditions and noise levels, thereby providing new insights for intelligent maintenance in smart manufacturing. • Proposes a spatial–temporal meta-hypergraph for few-shot fault diagnosis. • Captures high-order relations in multimodal sensor data by hypergraph. • Enables fast adaptation under varying working conditions with few samples. • Demonstrates superior accuracy and robustness across noise and conditions. • Offers new insights into intelligent maintenance for smart manufacturing.