State-Space Temporal Modeling and Feature Representation Learning for Anomaly Detection in Backend Microservice Systems

Abstract

This paper addresses the challenge of reliably identifying end-to-end latency anomalies in backend microservice architectures by proposing a link-level anomaly detection method based on distributed tracing and multi-source observable signals. This method uses the request-level call chain as the basic object, unifying end-to-end latency, service segment latency, and link structure information. It aligns contextual signals from metrics and logs within a time window to form a joint feature representation usable for detection. To reduce interference from noise, scale differences, and extreme values, robust normalization and lightweight smoothing are employed to normalize the input. Subsequently, neighborhood aggregation is performed based on the service call graph to characterize the consistency relationship between node states and their upstream and downstream components. Residual deviation is used to measure local anomalies and propagational disturbances. Furthermore, end-to-end deviation and node residuals along the path are fused to construct a unified anomaly score, and detection results are output through threshold rules, while retaining link contribution decomposition to support interpretable analysis. Comparative experiments, conducted under a unified evaluation index system, validate the advantages of this method in detection accuracy, recall, overall performance, and discriminative ability, demonstrating that the collaborative modeling of the end-to-end structure and time-series statistics can effectively improve the stability and reliability of microservice latency anomaly detection.