Genane Youness (Laboratoire LINEACT CESI, Département de Recherche et Innovation)
Artificial Intelligence (AI) is increasingly used in predictive maintenance to support fault diagnosis and Remaining Useful Life (RUL) prediction. However, the high performance of deep learning models often comes at the cost of interpretability, limiting their adoption in safety-critical industrial applications. Explainable Artificial Intelligence (XAI) has emerged as a promising approach to improve the transparency of AI models and support trustworthy decisions. In this talk, I introduce the main concepts of XAI and illustrate the role of post-hoc explainability in predictive maintenance through our research on RUL prediction. An explainable framework combining feature clustering, Long Short-Term Memory (LSTM) networks, and SHAP explanations is first presented to improve both prediction performance and model interpretability. The talk then addresses a major challenge in XAI: how can the quality of AI explanations be objectively evaluated? Although post-hoc methods may generate different explanations for the same prediction, their evaluation remains an open issue. A property-based evaluation framework is presented that links quantitative evaluation metrics to key explainability properties, providing a structured approach for comparing post-hoc XAI methods.
Poster