Self-Supervised Anomaly Detection with Knowledge-Enhanced Representation Learning for Distributed System Environments

Abstract

To address the challenges of anomaly detection in distributed system environments, such as complex anomaly types, high annotation costs, and limited anomaly sample quantities, this paper proposes a unified modeling method for self-supervised detection and knowledge enhancement, specifically for scenarios with scarce anomalies. This method is based on multi-source operational observations, organizing logs, metrics, and dependencies into a unified representation object. Building upon this, it combines structural modeling and temporal context encoding to learn a latent representation that characterizes the inherent patterns of the system state. To alleviate the limitations of scarce anomaly samples on supervised training, this paper constructs a self-supervised learning objective through mask reconstruction and contrastive constraints, enabling the model to extract discriminative state features from a large amount of unlabeled operational data. Simultaneously, addressing the issue that data-driven representations alone are insufficient to fully express operational semantics in complex distributed environments, this paper further introduces knowledge memory and adaptive fusion mechanisms. Prior knowledge of system structure and fault-related information is injected into the representation learning process, thereby enhancing the model's ability to identify anomaly states and its semantic consistency. Finally, this paper constructs a unified framework for anomaly detection in distributed systems, organically combining multi-source information modeling, self-supervised representation learning, knowledge enhancement constraints, and anomaly scoring. Comparative analysis shows that the proposed method can more effectively improve anomaly detection performance, providing a new research approach for intelligent operation and maintenance analysis in complex digital infrastructure scenarios.