Explainable Artificial Intelligence for a high dimensional condition monitoring application using the SHAP Method

Raphael Wallsberger, Tim Dieter Eberhardt, Paul-Albert Bartlau, Maurice Lucas Dörnte, Tim Lukas Schröter, Stephan Matzka

Abstract: In this paper, a new visualization-based method for understanding industrial machine learning models trained on high dimensional data is proposed. For a view that includes all dimensions, a dimensionality reduction towards a 2D projection, the UMAP method, is used. With the TreeSHAP algorithm the most important features for each machine condition are identified, visualized and evaluated. A closer look at different data points of the most important features provides more information about the behavior of the model. In addition, this knowledge is used to derive a class-optimized 2D visualization to increase trustworthiness of individual classification results for domain experts.