Abstract
Anomaly detection is essential for keeping chemical plants safe and running efficiently. Although many deep-learning methods have been proposed, most are still tested mainly on synthetic benchmarks such as the Tennessee Eastman Process (TEP). While these simulators enable fair comparisons, they do not reflect the noise, complexity, and irregular fault behavior of real industrial plants. As a result, it remains unclear how well these models generalize in practice. In this work, we extend our earlier ESCAPE study and move beyond water systems to industrially relevant chemical processes. We analyze data from two continuously operated pilot plant scenarios at the Technical University of Munich: n-butanol/water heteroazeotropic distillation and poly(oxymethylene) ether purification. We published these datasets for the first time at NeurIPS 2025. In this work, 30 anomaly detection methods, including 26 deep-learning and 4 classical approaches, are benchmarked using the open-source TimeSeAD library to ensure consistent preprocessing and evaluation. Performance is measured using F1-score and AUPRC. Our results show a clear performance drop when moving from synthetic to real-world data, with average scores far below those reported for TEP. No single method performs well across all datasets, and simple techniques such as PCA often perform on par with, or better than, more complex models. Together, these results highlight the need for more realistic benchmarks, process aware methods, and evaluation practices that better reflect real industrial conditions.