DELPHI Diagnosis by ExpLoiting PHysical Insights in neural network models

Important results from the project

The project had as objective to develop methods for generating data-driven residuals using structural information and time series data. It was investigated how the models can be trained and how to handle when the training data does not represent the fault cases we want to diagnose. In case studies together with industry, it was possible to see that the method works and is an opportunity to increase performance compared to the diagnostic solutions of today.

Expected long term effects

In the long term, we believe that the project can increase the performance and lower the cost of vehicle-related diagnosis by making it data-driven, automated and self-learning. Parts of the results from the projects are expected to be further developed in future research projects, while other parts are expected to be further developed by the industry.

Approach and implementation

The project work was divided into four work packages, two of which were led by the university party and conducted research aimed at various research questions. The other two were led by the industry party and focused on developing test infrastructure and conducting case studies.