Tutorial:
Machine Learning for Industrial Condition Monitoring – how to?

Condition monitoring (CM) of components and processes using machine learning (ML) is one of the central promises of Industry 4.0. Many successful examples have been demonstrated under laboratory conditions. However, the transfer to actual industrial application is proving difficult. The main challenge remaining is the data quality required for developing a meaningful and robust ML model: in industrial applications, most data represent the “good” condition, while samples for different fault scenarios are typically scarce. Furthermore, comprehensive training data are required covering all relevant circumstances to allow successful CM under changing environmental conditions and other causes of domain shift. Even if extensive data are available, most effort is spent on their organization to delete outliers, ensure correct labeling etc.

The tutorial will address these issues with two main approaches. The first is a checklist to guide users through the complete process of an ML project, starting with project, measurement, and data planning proceeding to data acquisition, checking and pre-processing up to finally building and validating the ML model. This checklist specifically supports users with little experience in ML to be successful. The second approach is classical process optimization based on insights gained using explainable machine learning methods.

Presenters:

Tizian Schneider, Centre for Mechatronics and Automation Technology

Andreas Schütze, Lab for Measurement Technology, Saarland University