Shapley-based Explainable AI for Clustering Applications in Fault Diagnosis and Prognosis

Shapley values can work with incomplete data: The paper demonstrates that SHAP-based explainability techniques can be adapted for semi-supervised learning scenarios with weakly labeled datasets, expanding the applicability of XAI beyond strictly supervised settings commonly used in industry.

Explainability improves clustering quality: When Shapley values were applied to clustering problems in fault diagnosis, the semi-supervised approach "significantly improves upon clustering quality compared to the fully unsupervised case," suggesting that explainability methods can simultaneously enhance model performance and interpretability.

Simple rules can explain complex failures: The research found that equipment failure clusters could be characterized by "high-precision decision rules in terms of original feature values, limited to 1-2 terms," meaning that even complex manufacturing faults can often be explained through straightforward, human-understandable conditions rather than black-box outputs.

Real-world manufacturing challenges require adapted XAI: The paper addresses practical obstacles in industrial AI adoption—specifically class imbalance and incomplete labeling in manufacturing data—showing that standard XAI techniques need modification to handle realistic, messy industrial datasets.

Validated across multiple manufacturing domains: The methodology was tested on both semiconductor heatmap data and equipment prognostics benchmarks, suggesting the approach has broader applicability across different types of manufacturing fault diagnosis problems.