Graph Embedded Patch-Sense Autoencoder with Prior Knowledge for Multi-Component System Anomaly Detection

For large-scale integrated systems, efficient anomaly detection is crucial in monitoring complex equipment. Interdependency among multiple components hinders the establishment of trust in system decision-making within multivariate time series. In this work, we propose a prior knowledge graph embedded patch-sense autoencoder (GEPAE), aiming to enable unsupervised anomaly detection in large-scale systems and provide a reference for anomaly localization. The proposed method learns from multi-source normal condition data to attain efficient and reliable anomaly detection. Diverging from pointwise reconstruction and evaluation, the proposed method employs unit-level patch embedding in the encoding module and patch correction in the decoding module, aiming to preserve data structure information by learning the global relationships among patches. System decision-making and component anomaly localization are subsequently conducted by Gaussian mixture density estimation of the patch embedding results. Meanwhile, prior knowledge of the physical entity structure and multi-sensor deployment is generalized and utilized for the model to obtain convincing decision-making. The proposed method is tested on two real-world data sets of liquid rocket engine systems and two fault simulation data sets of subway train transmission systems, and promising results demonstrate its validity and generality.