Self-Supervised Unified Representation Learning with Structural Dependency Modeling for Early Risk Detection in Distributed Systems

Abstract

This study addresses the challenge of accurately modeling system states in distributed systems under highly dynamic workloads, multi-indicator coupling, and continuously evolving structures. A self-supervised unified representation framework is proposed for capturing multi-indicator association features. The method first constructs a high-dimensional temporal input matrixfrom multi-sourcemonitoring data and uses a nonlinear encoder to extract multi-scale features, including local trends, short-term disturbances, and long-term dependencies. A dependency graph based on structural consistency is then used to describe functional relations and potential propagation paths between nodes. A structural propagation mechanism integrates multi-indicator coupling relationships into a unified latent representation space. Predictive, contrastive, and reconstruction-based self-supervised tasks are further designed to enable the model to learn stable structural patterns and statistical regularities of normal operation without labels. Early risk signals are identified by detecting shifts in the latent space, providing efficient and reliable support for system safety. Experimental results show that the framework exhibits strong robustness and adaptability under conditions such as structural disturbances, node fault injection, and data noise. It can capture hidden early anomaly features and significantly enhance early risk detection. The method integrates temporal features, structural dependencies, and self-supervised signals in a unified modeling manner and offers a new technical pathway for understanding complex association behaviors in large-scale distributed systems.