Estimating parameter space limits for industrial processes with in-distribution data

Abstract Machine learning has been finding its way into industrial systems with use cases like predictive maintenance, process optimization, anomaly detection, and many more. Low data coverage limits the usability of these machine learning models to only a small fraction of the overall parameter space. To make predictive machine learning models feasible for industrial processes, it is important to collect relevant data to reduce epistemic uncertainty in all areas of the parameter space. Approaches from experimental design aim to patch areas with low data coverage and provide useful information for previously unknown process settings. These methods assume that the factor limits in the parameter space are known. However, the high-dimensional parameter spaces in industrial processes with complex factor interdependencies make it difficult to find the dynamic physical limits of the system. Therefore, it is hard to define a set of experiments that are physically possible to reach. We provide a novel method to explore these high-dimensional parameter spaces. Through in-distribution data projection and out-of-distribution predictions with autoencoders, we estimate the physical limits of the parameter space of industrial processes.