When contexts change : hallucination effects in process monitoring and fault diagnosis models

This work addresses the unpredictable degradation of Process Monitoring and Fault Diagnosis models when they encounter new, but normal, operational contexts. We address a common, yet overlooked, misconception that training data comprehensively represents the entire operational domain of a process. In reality, training datasets are limited to specific conditions, making models vulnerable to context shifts where the process transitions to a new, but normal, operating state. This shift directly impacts model reliability by causing `hallucinations,' an increase in false positives and false negatives. While this is a known challenge in machine learning theory, it is often forgotten in real-world industrial applications. Unlike Large Language Models, which are trained on vast, multi-context datasets, Industrial AI models are typically trained on limited data, making them particularly susceptible to context shift. We use the Tennessee Eastman Process as a case study to illustrate how these unmodeled contexts can lead to a degradation of model performance. The objective is not to propose a new detection method but to clarify the fundamental relationship between a model’s training scope and its reliability in a dynamic industrial environment. Our contribution is to provide an understanding of why models fail when confronted with new normal conditions, laying the groundwork for more robust and trustworthy industrial AI systems