Una red neuronal convolucional temporal gráfica mejorada por representación bajo patrones de datos faltantes...

Extracting effective anomaly features and mitigating interference from missing data is crucial for multivariate time series anomaly detection in equipment. However, existing studies tend to focus on modeling under complete data scenarios and fail to adequately account for the challenges posed by complex missing data patterns, ultimately compromising the reliability of anomaly detection. To this end, this work proposes a representation-enhanced graph temporal convolutional network (REGTCN) under a complex missing pattern for equipment anomaly detection. This method is designed as a jointly optimized framework integrating reconstruction-based and prediction-based paradigms to enhance the representation of system health status. In the reconstruction module, we develop a missing-tolerant masked graph attention (MGAT) network to mitigate the adverse effects of missing patterns. In the prediction module, we propose an adaptive multi-scale temporal convolutional interaction network (AMTCIN) to capture sufficient temporal features. Finally, Extensive experiments are conducted under various missing-data scenarios. Experimental results demonstrate that our method outperforms all baseline models.