Adaptive Representation Learning and Temporal State Modeling for Robust Anomaly Detection in Distributed Systems

Abstract

To address the complexity of anomaly patterns, the heterogeneity of multi-source observations, and the continuous evolution of system states in distributed environments, this paper proposes a learning enhanced approach for anomaly perception and detection in distributed systems. The method starts from holistic system state modeling. Logs, metrics, and traces are mapped into a unified latent representation space. Temporal state modeling is used to characterize normal operating trajectories. Prediction consistency is employed to quantify anomalous deviations. An online update mechanism is further introduced. The model can continuously adjust its understanding of normal behavior during system operation. This reduces the impact of environmental changes and distribution drift on detection stability. The framework does not rely on static rules or fixed threshold assumptions. It automatically learns system operating structures in a data-driven manner. Anomaly detection is thus shifted from local discrimination to holistic perception of global state evolution. Comparative analysis under unified data and evaluation settings shows that the proposed method achieves more robust anomaly discrimination and result consistency. It more effectively distinguishes normal operational fluctuations from true anomalous behavior. This work provides a dynamic environment-oriented modeling perspective for distributed system anomaly detection and has important implications for enhancing system observability and supporting intelligent operations.