A Multi-Metric Trade-Off Framework for AI Model Selection for Diagnosis and Prognosis in Predictive Maintenance

Predictive maintenance is a strategy designed to trigger maintenance actions only when necessary for human-made assets. It relies on advanced systems capable of detecting early-stage faults and assessing degradation to estimate the Remaining Useful Life (RUL) of the component or system under consideration. Given its importance for both academia and industry, predictive maintenance has attracted increasing attention over the past decade. However, developing effective predictive maintenance systems remains a major challenge. Although numerous models support diagnostic and prognostic tasks, their selection is often guided by practitioner experience and evaluated primarily through precision. Relying on a single performance metric is insufficient for a comprehensive tradeoff analysis. This study proposes a holistic trade-off evaluation procedure for predictive maintenance models based on a scoring system that incorporates multiple performance indicators. This evaluation focuses on the four most relevant metrics in the literature: precision, computational efficiency, generalization, and robustness. To validate the proposed trade-off methodology, a case study on Li-Ion Battery RUL estimation is considered in which two deep learning models are compared: Physics-Informed Neural Networks (PINNs) and Attention-based Convolutional Neural Networks. The results highlight the decisive value of the proposed evaluation framework, as it accounts for a wider range of performance criteria beyond precision alone. This work provides a solid foundation for advancing future research on the selection of predictive maintenance models.