This paper presents a robustness evaluation framework for deep learning (DL)-based predictive models for predictive modeling and health management (PHM) that utilize complex time-series data generated in cyber-physical systems (CPS) such as manufacturing and energy distribution. Existing robustness evaluations have limitations in that they focus on formal verification or adversarial interference and fail to sufficiently reflect the complexity of real CPS environments. In this paper, we propose a practical robustness definition based on distributional robustness and present a systematic framework to thoroughly analyze the robustness of predictive models on real CPS datasets by simulating realistic disturbances such as sensor drift, noise, and irregular sampling. We demonstrate the applicability and effectiveness of the proposed approach through empirical studies on various DL architectures (including recurrent, convolutional, attention-based, modular, and structured state-space models), and disclose robustness benchmarks to encourage further research and reproducibility.
