(207g) A Knowledge-Guided Framework for the Effective Application of Machine Learning Models in the Development of Condition Monitoring Systems.

Machine learning models have suitable applications in the development of condition monitoring systems that ensure the quality, safety, and overall performance of a process. This preference stems from the black-box nature of machine learning models, where it is not necessary to have a well-established relationship between the predictions and the input variables, and all that is required are sufficient amounts of data. Often, this approach to developing condition monitoring systems can yield poor performances in detecting and diagnosing faults, and strategies for improvement are often unclear. Furthermore, the exclusion of process knowledge from the development process is counterproductive and could lead to misleading predictions.

In this work, we introduce a framework for the development of condition monitoring systems, where machine learning models are utilized based on knowledge of the process. This framework was applied to a pharmaceutical dry granulation line and was shown to be more effective than the traditional approach. Moreover, these advantages were quantified via metrics commonly used in machine learning, as well as proposed indices to quantify performance such as the ability of a condition monitoring system to classify novel faults.