Process fault detection through quantitative analysis of learning in neural networks

In this paper we demonstrate the use of quantitative measures of learning in neural networks for detection of chronic disturbances. This is a first attempt to provide such a measure in the data-driven black-box engineering practice for detection of unknown non-cataleptic disturbances. Such events are characterized by the presence of unknown influences, but observable in measurable variables. Though such events do not necessarily have an immediate impact on a model’s validity for process measurements, a significant change in the model’s learning behavior can be detected. This is demonstrated by detecting discrepancies in some exemplary toy-problem data sets in approximation, prediction and identification. Though analysis of parameter dynamics appears successful in these problems, future research is required on sensitivity robustness and promptness considering measures of external and internal behavior. Keywords— neural networks, learning behavior, fault detection,