Distributed System Anomaly Detection via Collaborative Discrimination of Reconstruction Residuals and Latent Distribution Shifts

Abstract

To address the challenges of concealed anomalies, complex state evolutions, high-dimensional and heterogeneous monitoring data, and the difficulty of traditional methods in simultaneously addressing representation learning and anomaly deection during distributed system operation, this paper proposes a distributed system anomaly detection method based on VAE reconstruction. This method takes a multivariate time series composed of system operation indicators as input, organizes local contextual information through a sliding time window, and compresses redundant perturbations in the original observations using a feature projection mechanism. Then, it learns the latent probability distribution of normal states using a variational autoencoder structure. Based on this, the model not only reconstructs the input features but also jointly characterizes the reconstruction residuals and latent distribution shifts, fusing the two types of anomaly evidence into a unified anomaly score, thereby enhancing the ability to identify complex anomaly states. Simultaneously, a time consistency constraint is introduced to reduce the interference of short-term random fluctuations on the discrimination results, making anomaly detection more consistent with the continuous change characteristics of real distributed system operation. Results show that the proposed method can more effectively distinguish between normal and abnormal states, exhibiting good performance in terms of accuracy, stability, and overall discrimination capabilities. This study provides an implementation path for anomaly detection in distributed systems that combines generative modeling and discriminant analysis, and also provides a methodological foundation for intelligent monitoring and risk identification in complex operating environments.