Machine Learning Systems for Unsupervised Time Series Anomaly Detection

Abstract

Modern assets – from launched satellites to electric vehicles – output dense, multivariate time series data that must be monitored for deviations from “normal” behavior. This monitoring task is referred to as time series anomaly detection. The current state of the industry still depends on fixed or heuristic thresholds that often drown operators in false alarms, and can miss the subtle, context-dependent faults that matter most. This thesis addresses unsupervised time series anomaly detection as an end-to-end problem, asking how we can learn, evaluate, and deploy models that judiciously flag anomalies while remaining intuitive to the end user. This thesis provides contributions in the form of both algorithms and systems. First, it introduces three models that enlarge the design space of unsupervised time series anomaly detection: TadGAN, which leverages adversarial reconstruction; AER, which unifies predictive and reconstructive objectives in a single hybrid score; and MixedLSTM, which explicitly incorporates interdependencies to improve anomaly detection in multivariate time series. We propose two range-based evaluation metrics that quantify detection quality over temporal intervals. Second, it presents our system Orion, which abstracts anomaly detection pipelines as directed acyclic graphs of reusable primitives, providing user-friendly APIs and enabling interactive visual inspection. Building on this infrastructure, OrionBench performs periodic, fully reproducible benchmarks, producing leaderboards that align research innovations with the needs of end users. Third, the thesis explores a new paradigm – foundation models for unsupervised time series anomaly detection – by formulating SigLLM , which employs large language models and time series foundation models for zero-shot anomaly detection via prompting and forecasting. This paradigm indicates a promising path to developing scalable models for anomaly detection. Finally, beyond evaluating our systems on publicly available datasets, we provide extensive experiments on two industrial case studies that demonstrate improved detection accuracy and practical usability of our system.