CAREER: Towards Automated and Reliable Time-Series Anomaly Detection

The ability to detect unexpected and potentially dangerous events in continuously collected data is a cornerstone of modern infrastructure and critical systems. From power grids and manufacturing plants to healthcare monitoring and financial systems, time-dependent data streams are essential for ensuring safety, reliability, and economic stability. Failures to identify anomalous patterns in these data can lead to cascading disruptions, large-scale financial losses, cybersecurity breaches, and threats to human life. Despite decades of research, existing anomaly detection approaches remain unreliable in practice, requiring significant human expertise and often failing under noisy, dynamic, and evolving conditions. This project addresses a critical need by advancing automated, accurate, and interpretable methods for anomaly detection in time-series data. By enabling timely and trustworthy identification of abnormal events, the project aims to strengthen the resilience and security of essential systems, enhance decision-making across data-driven industries, and contribute to education and workforce development in artificial intelligence and data science. This project develops Chameleon, a next-generation, open-source system that establishes a new paradigm for automated time-series anomaly detection by unifying foundation models, meta-learning, and probabilistic model ensembling into a single end-to-end framework. The research is organized around three synergistic components. First, foundation model–empowered detection introduces controllable synthetic data generation and large-scale pretraining to enable domain-agnostic, zero-shot anomaly detection, along with a multimodal architecture that localizes anomalies and generates interpretable natural-language explanations. Second, an automated model selection framework leverages meta-learning, out-of-distribution detection, and label-free verification to identify high-performing models without ground-truth annotations, addressing fundamental challenges in unsupervised settings. Third, an anomaly detection optimizer develops probabilistic ensembling and Bayesian thresholding methods to improve robustness, reduce computational cost, and produce calibrated anomaly decisions. The project is supported by strong collaborations with industry and academic partners, enabling validation across diverse, high-impact application domains, including large-scale infrastructure monitoring, manufacturing systems, energy networks, and healthcare diagnostics. By integrating these advances, the project delivers a scalable, self-tuning, and interpretable system that advances the state of the art in time-series analysis, automated machine learning, and trustworthy artificial intelligence. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.