故障检测与隔离
计算机科学
超图
模式识别(心理学)
人工智能
目标检测
方位(导航)
算法
噪音(视频)
数据挖掘
噪声测量
计算机视觉
特征提取
异常检测
算法设计
信号处理
断层(地质)
变更检测
探测理论
图像分割
标识
DOI:10.1109/jsen.2026.3698433
摘要
Effective bearing health monitoring is critical for industrial predictive maintenance, yet it remains challenging due to the scarcity of labeled fault data and complex spatiotemporal dependencies in multi-channel sensor signals. To address these issues, this paper proposes an Unsupervised Anomaly-aware Spatiotemporal Hypergraph Model (AaSHM) for bearing fault detection. AaSHM learns normal operating patterns from healthy data and identifies potential faults through anomaly scoring during inference. The proposed model employs a dynamic local-global hypergraph structure to capture high-order spatial interactions among sensor channels and long-range temporal dependencies: local hyperedges are constructed using multi-source statistical features to characterize fine-grained sensor couplings, while global hyperedges model system-level correlations based on mutual information. Furthermore, an anomaly-gating mechanism is integrated into the hypergraph convolution process to adaptively modulate feature aggregation based on node-level anomaly characteristics, enhancing sensitivity to incipient faults. Extensive experiments on the CWRU and PU benchmark datasets demonstrate that AaSHM achieves superior comprehensive performance. Specifically, AaSHM consistently outperforms existing graph-based and deep learning baselines in F1-score, AUC, and AUPR across various operating conditions.
科研通智能强力驱动
Strongly Powered by AbleSci AI