ABSTRACT Accurate prediction of remaining useful life (RUL) is critical for effective predictive maintenance. While models like long short‐term memory (LSTM) are effective, they often lack interpretability, even when using explainable artificial intelligence (XAI) methods such as shapley additive explanations (SHAP). This is particularly true when these models are trained on high‐dimensional, redundant features. To tackle this issue, we introduce the interpretable divisive feature clustering (IDFC) algorithm. This unsupervised dimensionality reduction method combines the advantages of divisive clustering and ‐means‐like clustering to group highly correlated features into unidimensional clusters. Additionally, it selects representative features to maintain semantic meaning. By doing so, the reliability of post hoc explanation methods like SHAP is improved by reducing multicollinearity. When IDFC is combined with a one‐layer LSTM model on the C‐MAPSS dataset, it achieves competitive remaining useful life (RUL) prediction performance with significantly fewer features. This leads to a lower prediction error compared to principal component analysis (PCA)‐based approaches. Additionally, the quality of explanations provided by SHAP, assessed using several functionally grounded metrics such as coherence, stability, and acumen, is enhanced when IDFC is applied, as opposed to using all features. Finally, utilizing IDFC reduces explanation time by 41% compared to the baseline model that includes all features. These findings confirm that IDFC improves both the predictive accuracy and explanatory power of deep models in complex industrial environments.
Ndao et al. (Wed,) studied this question.