Interpreting NAS-Optimized Transformer Models for Remaining Useful Life Prediction Using Gradient Explainer

Abstract

Remaining Useful Life (RUL) estimation of complex machinery is critical for optimizing maintenance schedules and preventing unexpected failures in safety-critical systems. While Transformer architecture has recently achieved state-of-the-art performance on RUL benchmarks, their design often relies on expert tuning or costly Neural Architecture Search (NAS), and their predictions remain opaque to end users. In this work, we integrate a Transformer whose hyperparameters were discovered via evolutionary NAS with a gradient-based explainability method to deliver both high accuracy and transparent, perprediction insights. Specifically, we adapt the Gradient Explainer algorithm to produce global and local importance scores for each sensor in the C-MAPSS FD001 turbofan dataset. Our analysis shows that the sensors identified as most influential, such as key temperature and pressure measurements, match domain-expert expectations. By illuminating the int ernal decision process of a complex, NAS-derived model, this study paves the way for trustworthy adoption of advanced deep-learning prognostics in industrial settings.