Theory-Guided Data Science for cross-application enhancement of data-driven diagnostic and prognostic methods in Prognostics and Health Management.

• Applicants: Professor Dr.-Ing. Marco Huber; Professor Dr.-Ing. Peter Zeiler
• Subject Area: Mechanics ; Engineering Design, Machine Elements, Product Development ; Production Automation and Assembly Technology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 514247199

The use of data-driven methods for condition diagnosis and prognosis for technical systems has a broad spread in Prognostics and Health Management (PHM). These methods are based on statistical modeling of interrelations within the training data provided. Therefore, they are not suitable for extrapolation to domains without training data. Furthermore, their prediction may take implausible values that are not consistent with governing physical laws and other constraints. In PHM, these drawbacks are of considerable relevance due to the high cost and time involved in generating comprehensive training data. To mitigate these drawbacks, there are approaches in machine learning that address the incorporation of various forms of knowledge about the system under consideration. These are referred to among other things as Theory-Guided Data Science (TGDS). Although it is usually not possible to model degradation processes with sufficient physical detail, basic knowledge about the system under consideration and the inherent regularities of its degradation process is usually available. Therefore, the research project aims to reduce the drawbacks of data-driven methods in PHM by integrating knowledge on regularities occurring across applications. Within the project, the regularities that frequently apply in PHM are first identified and their scope analyzed. Subsequently, a selection of the most relevant regularities is made. The main focus of the project is to investigate how these regularities can be used in the context of TGDS to improve data-driven methods and how the improvement evolves in relation to the amount of available data. The research results are validated using experimental data from two mutually heterogeneous PHM use cases. In PHM, various studies exist on combining data-driven and physical models using hybrid methods. The present project, however, considers the integration of cross-application regularities occurring in PHM, which are not sufficient for a physical modeling of the degradation process. This sets it significantly apart from the state of research. On the one hand, the methods of TGDS have not yet been specifically addressed in PHM. On the other hand, the literature on PHM so far lacks any cross-application consideration of the subject.

• DFG Programme: Research Grants