Abstract
Anomaly detection is essential for ensuring the safe and efficient operation of chemical plants. Although many deep-learning-based methods have been proposed in recent years, their evaluation remains largely limited to synthetic benchmarks such as the Tennessee Eastman Process (TEP) [1]. While these simulators enable controlled and reproducible comparisons, they fail to capture the noise characteristics, operational complexity, and irregular fault dynamics of real industrial plants, leaving the practical generalizability of many methods unclear. In this work, we extend our earlier ESCAPE study [2] beyond water-based systems to industrially relevant chemical processes. We analyze multivariate time-series data from two continuously operated pilot-plant scenarios at the Technical University of Munich, namely n-butanol/water heteroazeotropic distillation and poly(oxymethylene) ether purification, whose datasets were recently published at NeurIPS 2025 [3]. Using the open-source TimeSeAD library [4], we benchmark 30 anomaly detection methods, including 26 deep-learning-based and 4 classical approaches, under a unified preprocessing, model-selection, and evaluation pipeline. Performance is assessed using the F1-score and the area under the precision-recall curve (AUPRC). Our results show a substantial performance drop when moving from synthetic to real process data, with average scores far below those commonly reported for TEP. No single method performs consistently best across all datasets, and the ranking depends strongly on the chosen metric and process scenario. These findings highlight the limitations of synthetic benchmarks and underscore the need for more realistic industrial datasets, process-aware methods, and evaluation practices that better reflect real operating conditions.