一句话：提出两阶段语义异常检测框架 Smart Brain，将运维时序异常检测从数值偏差转向可解释、可操作的语义异常决策。
大规模服务系统的运维自动化依赖时序异常检测，但现有方法通常把异常视为数据空间中的数值偏差，这与运维中判断某行为是否构成可操作异常事件的需求不匹配，导致阈值脆弱、误报过多且结果难以解释。本文提出两阶段语义异常检测框架 Smart Brain：第一阶段通过识别可解释的异常模式抽取结构化时序证据；第二阶段将证据与指标语义、运维人员编写的规则结合，生成语义异常决策和人类可读解释。作者在基准数据集和真实工业监控场景上评估该框架，表明其相比以异常分数为中心的流水线能提升具有运维意义的异常识别效果，并具有更好的可解释性和部署适用性。研究还总结了语义锚定、结构化证据以及对服务演化进行轻量适配等实践教训。
Time series anomaly detection is a core component of operational automation in large scale service systems. However, most existing methods model anomalies as numerical deviations in the data space, while real software operations require decisions over whether an observed behavior constitutes an actionable abnormal event. This mismatch often leads to brittle thresholds, excessive false alerts, and outputs that are difficult for operators to interpret and act upon. In this paper, we argue that practical anomaly detection should move from point-centric data anomaly detection to semantic anomaly detection. We present Smart Brain, a two stage framework for semantic anomaly detection in operational time series. Stage I extracts structured temporal evidence by recognizing interpretable anomaly patterns. Stage II combines the extracted evidence with metric semantics and operator authored rules to produce semantic anomaly decisions and human-readable explanations. Through evaluation on benchmark datasets and real industrial monitoring scenarios, we show that Smart Brain improves operationally meaningful anomaly identification while providing stronger interpretability and deployment suitability than score centric pipelines. Beyond the framework itself, our study offers practical lessons for building anomaly detection systems in production software environments, highlighting the importance of semantic grounding, structured evidence, and lightweight adaptation to evolving services.
@inproceedings{paperbot12369,
title = {Smart Brain: Semantic Anomaly Detection for Operational Time Series in Large Scale Service Systems},
author = {Hang Cui and Zexin Wang and Jingjing Li and Juncheng Hu and Haotian Si and Cenjie Hu and Quan Zhou and Yongchang Hu and Lei Han and Dan Pei and Changhua Pei and Gaogang Xie},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3837436}
}