发表机构
Drew University（德鲁大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
ARGUS是一个配置驱动的探测器健康监测框架，通过分离通用与特定部分，利用四种独立监视器实现异常检测，在合成数据上验证了高准确率与适应性。
AI 中文摘要
每个粒子物理实验都会构建探测器健康监测系统，而许多实验最终不得不重建它，因为在调试压力下编写的首个系统往往过于僵化而难以维护。ARGUS（自动异常检测、运行质量与通用统一监视）是一个配置驱动的框架，它将此问题中的通用部分（健康监视器、其验证和报告）与特定部分（数据模式和阈值）分离。为新探测器部署它需要一个数据适配器和一个配置文件，而非重写。探测器健康由四个独立的监视器评判：专家截断，具有权威性；机器学习第二意见，具有附加性且默认关闭；趋势预测，在阈值被越过之前提供早期预警；以及带有可插拔统计检验的参考直方图比较。这些监视器相互比较但从不合并，警报层将其标志和警告转化为去重通知。我们在一个带有植入病理的合成参考探测器上展示了完整链路，并通过故障注入活动制造了真实标签：截断恢复了所有植入病理且无误报，机器学习监视器在注入故障上达到0.896的ROC AUC，并将覆盖范围扩展到未编写截断的故障类别，预测在截断触发之前标记了漂移通道，直方图比较捕获了所有植入的形状畸变，同时在干净对上保持了配置的误报率。第二个演示展示了框架适应重新组织的文件布局，无需更改代码或配置。本文中的每个结果和图表仅从代码重新生成。该框架已在两个粒子物理实验中实现，目前正在评估中；采用决定由合作组做出。
英文摘要
Every particle-physics experiment builds detector-health monitoring, and many end up rebuilding it, because a first system written under commissioning pressure often proves too rigid to maintain. ARGUS (Automated Anomaly-detection, Run-quality and General Unified Surveillance) is a configuration-driven framework that separates what is generic in this problem (the health monitors, their validation, and the reporting) from what is not (the data schema and the thresholds). Standing it up for a new detector requires a data adapter and a configuration file, not a rewrite. Detector health is judged by four independent monitors: expert cuts, which are authoritative; a machine-learning second opinion that is additive and off by default; a trend forecast that gives early warning before a threshold is crossed; and a reference-histogram comparison with pluggable statistical tests. The monitors are compared and never merged, and an alerting layer turns their flags and warnings into deduplicated notifications. We demonstrate the full chain on a synthetic reference detector with planted pathologies and a fault-injection campaign that manufactures ground truth: the cuts recover every planted pathology with no false positives, the machine-learning monitor reaches a ROC AUC of 0.896 on injected faults and extends coverage to fault classes no cut was written for, the forecast flags a drifting channel before its cut fires, and the histogram comparison catches every planted shape distortion while holding its configured false-positive rate on clean pairs. A second demonstration shows the framework adapting to a reorganized file layout with no change to code or configuration. Every result and figure in this paper regenerates from the code alone. The framework has been implemented for two particle-physics experiments, where it is under evaluation; adoption decisions rest with the collaborations.
Comments15 pages, 6 figures, 4 tables
