Multi-scale Graph-level Anomaly Detection of Complex Equipment via a Subgraph Augmented Contrastive Self-supervised Network

Anomaly detection (AD) of complex equipment is critical to improving operational safety and reliability. Currently developed system-level intelligence methods neglect structural information between components and the multi-scale composition of anomalies, leading to frequent missed detections and false alarms. Towards this end, this paper proposes a subgraph augmented self-supervised network that represents multivariate time series (MTS) data in a non-Euclidean space to realize multi-scale graph-level AD on complex equipment. First, we present a subgraph contrastive self-supervised framework that emphasizes the acquisition of context-scale anomaly information in MTS data, resulting in accelerated training speed and improved fault detection rate (FDR). Furthermore, responding to the lack of scale, a subgraph self-learning strategy is proposed to capture patch-scale information, leading to an improved FDR. Meanwhile, we design a graph augmentation technique to alleviate the scarcity of graph-level labeled samples, increasing the robustness and scalability of the network and further reducing the false alarm rate (FAR). To assess the efficacy, we perform uni-modal, multi-modal, and cross-device experiments on various MTS datasets of liquid rocket engines. Compared to the state-of-the-art method, the proposed approach increases the FDR by 2% and reduces the FAR by half to 0.08%, demonstrating the superiority of the method.