Evaluating Advanced Anomaly Detection Models for Time Series Data of Refrigeration Systems

Reliable cold chain monitoring in supermarkets is essential, yet current anomaly detection often relies on fixed expert-defined thresholds and raw data, limiting flexibility. This study evaluates different advanced anomaly detection on real-world refrigeration time series data, focusing on generalization across similar refrigeration units with varying configurations. It examines model performance on both seen and unseen units and explores calibration methods, including data scaling and individual anomaly threshold determination. Results show that Transformer-based models and state-space models outperform traditional baselines and generalize well, even across differing systems. Tailored thresholds and calibration improve detection, offering a scalable approach for industrial deployment.