Explainable Artificial Intelligence Applied to Predictive Maintenance: Comparison of Post-Hoc Explainability Techniques

Nowadays Predictive Maintenance (PdM) is widely used to maximize machine availability and minimize unnecessary maintenance activities. However, PdM systems often rely on complex machine learning models that cannot explain their actions, distancing users from the decision-making process. Accordingly, the purpose of the present paper is to explore and implement post-hoc explainability techniques that are capable of extracting global and/or local interpretations from a Random Forest model applied in a PdM dataset. To accomplish this, eight techniques are compared: impurity-based feature importance, permutation feature importance, partial dependence plot, Accumulated Local Effects (ALE), Shapley Additive Explanations (SHAP), local interpretable model-agnostic explanations, Anchor, and individual conditional expectation. When compared to one another, each technique has advantages and disadvantages, with the majority of them compensating for each other's weaknesses. Nevertheless, SHAP is the most practical as it includes global and local interpretations with a variety of plots, ALE is the most useful and reliable for global interpretations of the model’s features, while Anchor is best suited for simple and compact local explanations.