Hierarchical Testing with Rabbit Optimization for Industrial Cyber-Physical Systems

Jinwei Hu, Zezhi Tang, Xin Jin, Benyuan Zhang, Yi Dong, Xiaowei Huang

Abstract

This paper presents HERO (Hierarchical Testing with Rabbit Optimization), a novel black-box adversarial testing framework for evaluating the robustness of deep learning-based Prognostics and Health Management (PHM) systems in Industrial Cyber-Physical Systems (ICPS). Leveraging Artificial Rabbit Optimization, HERO generates physically constrained adversarial examples that align with real-world data distributions via global and local perspectives. This study specifically focuses on the Proton Exchange Membrane Fuel Cell (PEMFC) system, chosen for its highly dynamic operational conditions and complex degradation mechanisms. Experimental results highlight HERO's ability to uncover vulnerabilities in state-of-the-art PHM models, demonstrating its potential to advance more resilient systems across ICPS domains.

Introduction

In industrial settings, Deep Learning (DL) is increasingly utilized for Prognostics and Health Management (PHM)—predicting the health and remaining useful life of critical components like hydrogen fuel cells. While these models are highly accurate, they often function as "black boxes" vulnerable to adversarial attacks: tiny, intentional changes to input data that cause catastrophic prediction errors. Existing adversarial testing tools, primarily designed for image recognition, are often unsuitable for ICPS due to the tabular, time-series nature of industrial data and the requirement that any "attack" must remain within physical operational limits to be a valid test case.

The HERO Framework

HERO (Hierarchical Testing with Rabbit Optimization) addresses these limitations through a four-tiered hierarchical framework designed to stress-test PHM models while ensuring physical realism.

The HERO framework workflow, illustrating the integration of global distribution approximation, local robustness indicators, and Artificial Rabbit Optimization (ARO) to generate physically constrained adversarial examples.

Methodology Components

  • Global Distribution Approximation: Uses a Long Short-Term Memory Variational Autoencoder (LSTMVAE) combined with Kernel Density Estimation (KDE) to capture the normal behavior and temporal dependencies of the system, identifying optimal locations for test seeds.
  • Local Robustness Indicator (LRI): Employs gradient norms to find sensitive spots in the data where small perturbations result in the largest prediction errors.
  • Artificial Rabbit Optimization (ARO): A meta-heuristic algorithm inspired by rabbit foraging behaviors (detour foraging vs. random hiding) that efficiently searches for effective adversarial perturbations.
  • Physical Constraints: A clamping function ensures that every generated adversarial example stays within the actual operational limits of the hardware, such as specific voltage and pressure bounds.

Target Model and Experimental Setup

The framework was evaluated using the IEEE PHM 2014 Data Challenge dataset, featuring over 127,000 data points from a 1kW PEMFC test bench. The dataset includes features such as stack voltage, gas temperatures, pressures, and flow rates. The primary target for testing was the state-of-the-art Temporal Scale Transformer (TSTransformer).

Architecture of the Temporal Scale Transformer (TSTransformer), the state-of-the-art model used as a target for HERO's adversarial testing.

Key Results and Insights

Experimental results underscore the vulnerability of even highly accurate models. Under a minor perturbation ($\epsilon = 0.03$), the TSTransformer's $R^2$ accuracy plummeted from 0.9951 to 0.4843, while its Root Mean Square Error (RMSE) increased tenfold, from 0.0012 to 0.0125.

HERO demonstrated several advantages over previous methods like HDA:

  • Efficiency: ARO proved computationally more efficient than standard Genetic Algorithms, with the speedup becoming more pronounced as sample sizes increased.
  • Realism: By incorporating physical constraints, HERO identifies "operational blind spots"—scenarios that could realistically occur due to sensor noise or cyber-tampering—rather than just mathematical edge cases.
  • Effectiveness: Combining global distribution analysis with local sensitivity proved far more effective than random sampling for uncovering critical model weaknesses.

Conclusion and Future Work

HERO provides a robust methodology for evaluating the resilience of PHM systems in industrial environments. However, its efficacy depends on the representativeness of training data and requires deep domain expertise to define physical constraints. Future research will explore more adaptive algorithms capable of handling larger-scale, high-frequency data and refining the framework to better mirror the highly dynamic nature of real-world industrial scenarios.

Summary updated

Abstract

This paper presents HERO (Hierarchical Testing with Rabbit Optimization), a novel black-box adversarial testing framework for evaluating the robustness of deep learning-based Prognostics and Health Management systems in Industrial Cyber-Physical Systems. Leveraging Artificial Rabbit Optimization, HERO generates physically constrained adversarial examples that align with real-world data distributions via global and local perspective. Its generalizability ensures applicability across diverse ICPS scenarios. This study specifically focuses on the Proton Exchange Membrane Fuel Cell system, chosen for its highly dynamic operational conditions, complex degradation mechanisms, and increasing integration into ICPS as a sustainable and efficient energy solution. Experimental results highlight HERO's ability to uncover vulnerabilities in even state-of-the-art PHM models, underscoring the critical need for enhanced robustness in real-world applications. By addressing these challenges, HERO demonstrates its potential to advance more resilient PHM systems across a wide range of ICPS domains.
BibTeX
@misc{hu2025hierarchical,
  author = {Jinwei Hu and Zezhi Tang and Xin Jin and Benyuan Zhang and Yi Dong and Xiaowei Huang},
  title = {{Hierarchical Testing with Rabbit Optimization for Industrial Cyber-Physical Systems}},
  year = {2025},
  eprint = {2507.04100},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.04100},
}
Figure