Abstract
Reliable fault detection in complex industrial systems requires capturing structured dependencies among process variables rather than relying solely on individual signal patterns. However, most existing approaches primarily depend on reconstruction or prediction errors, which may overlook relational interactions and lead to unstable detection performance under varying operating conditions. To address this limitation, this paper proposes a framework named causality-inspired graph representation with adaptive thresholding (CGRA). CGRA combines temporal causal representation to capture structured temporal dependencies from multivariate time series and graph attention representation to encode these relations into discriminative features for anomaly scoring. An adaptive thresholding strategy is further employed to obtain statistically reliable decision boundaries. Experimental results on multiple industrial datasets demonstrate consistent performance improvements over representative existing methods.
Funding Information
  • Open Project of the Key Laboratory of Intelligent Operation and Maintenance of High-parameter Lifts, State Administration for Market Regulation of China (JSTJ-IOMHL-202501)
  • Natural Science Foundation of the Jiangsu Higher Education Institutions of China (25KJD460006)
  • National Natural Science Foundation of China (52505610)
  • Natural Science Foundation of Jiangsu Province (BK20251073)

This publication has 39 references indexed in Scilit: