How to find the needle in the haystack? Outlier detection and prediction in high-throughput industrial production processes

Angela Bitto-Nemling, Johannes Kepler Universität

Abstract: In high-throughput industrial processes, failures are rare yet costly events that can cause severe production downtime or equipment damage. This work addresses the challenge of detecting such events in multivariate time series data under strong operational constraints: limited access to labeled failures, proprietary data, and strict runtime requirements. We evaluate a set of lightweight, interpretable anomaly detection methods that operate locally on sensor time series, without relying on deep learning architectures. In particular, we explore the use of adaptive Bayesian change point models for identifying early warning signals. Our findings offer practical guidance for deploying real-time monitoring solutions in sensitive, large-scale production environments.