Enhancing Remaining Useful Life Prediction Against Adversarial Attacks: An Active Learning Approach

Abstract

Accurate assessment of the Remaining Useful Life (RUL) is critical for preventing catastrophic equipment failures, but the machine learning models used for these predictions are highly vulnerable to adversarial attacks that can corrupt sensor data and produce dangerously incorrect outputs. To address this critical security gap, this paper introduces a novel active learning framework designed to bolster the adversarial robustness of RUL models. By strategically selecting the most informative and uncertain data points for labeling, our method enhances defensive capabilities while significantly reducing the reliance on large, pre-labeled datasets. Experiments conducted on the NASA C-MAPSS FD001 dataset against CW, FGSM, and JSMA attacks demonstrate that this approach yields significant gains. Specifically, the Random Forest model paired with active learning achieves a superior accuracy of 89%. This robust performance is achieved while requiring 42% fewer labeled instances compared to a traditional supervised approach, at the cost of a 28% increase in iterative training time. This work provides a practical pathway toward developing more secure, efficient, and resilient predictive maintenance systems for safety-critical applications.