Explainable AI for Industrial Predictive Maintenance in the UNDERPIN Manufacturing Dataspace

Predictive maintenance in manufacturing relies heavily on machine learning models to detect anomalies and predict failures before they occur. However, the complexity of these models often limits interpretability, which reduces trust and hinders adoption in industrial contexts. This has raised the demand for explainable AI (XAI) techniques. This paper presents a real-world implementation of Effector, an explainable AI package for tabular data, in an industrial predictive maintenance system within a Dataspace environment, to interpret predictions from a CatBoost model trained on time series manufacturing data for anomaly detection. Unlike traditional ”black box” AI solutions, Effector provides transparent decision-making through global and regional effect plots that reveal how each feature influences model predictions. We demonstrate how Effector’s regional explanations reduce heterogeneity in feature effects, improving interpretability and providing actionable insights for maintenance decision-makers. The integration of Effector XAI within a manufacturing Dataspace [1] highlights the potential of semantic technologies to enhance transparency and trust in industrial AI systems.