Interpretable artificial Intelligence for COndition Monitoring (IICOM)

Important results from the project

The goals were met.

Expected long term effects

We have advanced in the understanding of how to use machine learning models for condition monitoring. By using uncertainty quantification, counterfactual explanations and mutual information, we understand better when to trust the model prediction, better explain the outcome of black-box ML models and can create skeleton causal graphs. The results show how to use ML effectively for predictive maintenance, as well as the outcomes serve to set requirements to the industry to enable it.

Approach and implementation

In this project, we have developed a set of complementary tools which addresses some of the challenges of machine learning for condition monitoring. We used uncertainty quantification to tackle the robustness and adaptability challenges, counterfactual explanation for improving interpretability of black box models as well as causality for ML models to make the models more adaptable, robust and interpretable.