- Conference Article
- 10.1109/rams50514.2026.11424550
Explainable AI for Obsolescence Using Weak Labels
- Jan 26, 2026
- Aya Mrabah + 3 more +3
SUMMARY & CONCLUSIONS This work presents an explainable machine learning approach for assessing obsolescence risk in electronic components, targeting long-lifecycle systems in sectors such as aerospace, rail, and defense. Unlike traditional approaches that focus solely on predicting end-of-life dates or binary obsolescence status, our method incorporates multiple dimensions of risk: lifecycle, compliance, inventory, and multisourcing, reflecting the complexity of real-world operational environments.Since most machine learning models rely on labeled data (i.e., annotations specifying risk categories) to learn accurate predictions, its absence poses a significant challenge in industrial contexts. To address that common absence of labeled data in industrial datasets, we applied a weak-label supervision strategy. These weak labels are approximate but structured risk classifications derived from expert-defined business rules, grounded in domain knowledge.Each risk dimension was evaluated separately and weighed using a multi-criteria decision-making approach, ensuring that the final global risk score aligned with strategic priorities in obsolescence management. To ensure model transparency, we employed explainable AI techniques specifically SHapley Additive exPlanations (SHAP) method, which allow engineers to understand and trust how features contribute to each prediction. Each feature or risk factor is given a relevance value by SHAP for a specific prediction.This study demonstrates that multi-dimensional risk prediction is feasible even without labeled training data. The results provide a strong foundation for supporting proactive obsolescence management strategies that go beyond binary obsolescence status and align more closely with industrial decision-making needs.
Read more