Structure-Aware Multi-Scale Behavioral Learning for Distributed System Observability and Anomaly Identification

Abstract

To address the challenges of complex anomaly behaviors, diverse propagation paths, and the intertwining of local disturbances and global imbalances in distributed system environments, this paper proposes an anomaly detection method based on multi-scale behavioral pattern learning. This method leverages the characteristics of frequent node interactions, tight service dependencies, and dynamic evolution of operational states in distributed systems. It organizes system monitoring data into a unified dynamic behavioral representation and constructs a multi-layered modeling framework for anomaly identification based on this representation. In terms of method design, firstly, the node states, edge relationships, and multi-source operational information in the system are structurally represented to form a dynamic graph representation that reflects service call relationships and behavioral coupling characteristics. Then, focusing on the behavioral change patterns over different time periods, a multi-scale temporal pattern extraction mechanism is introduced to jointly characterize short-term fluctuations and long-term drifts, thereby enhancing the model's ability to perceive complex anomaly evolution processes. Furthermore, by incorporating spatial dependencies between services, the propagation characteristics of anomaly signals along call chains and interaction paths are modeled, enabling local anomalies and associated disturbances to be co-represented in a unified representation space. Finally, by fusing temporal behavioral features and spatial dependency features, an anomaly scoring mechanism for deviations from normal state prototypes is constructed to achieve effective identification of anomaly samples. The proposed method effectively adapts to the diversity, hierarchy, and dynamism of anomaly patterns in distributed systems. It balances representational and discriminative capabilities during detection, enhancing the modeling depth for complex operational states. Related research indicates that this method significantly improves the accuracy and stability of anomaly detection in distributed systems, providing valuable technical support for the monitoring, anomaly identification, and intelligent operation and maintenanceof complexsoftwaresystems.