Scholarlink

<p>Engineering systems that experience dynamic events with amplitudes exceeding 100 times the force of gravity near the Earth’s surface, typically with durations less than 100 ms, are defined as <i>high-rate dynamic systems</i>. These thresholds are representative of conditions encountered in advanced aerospace structures and impact reduction systems. Such systems are characterized by large uncertainties in external loads, high levels of nonstationarity, and unmodeled dynamics arising from configuration changes. To ensure their safe and reliable operation, there is a critical need for forecasting near-future states at high rates. Such predictive capability enables timely feedback and control actions that mitigate risks and prevent catastrophic failures. This paper presents a framework for high-rate multi-step forecasting that combines recurrent neural networks (RNNs), topological data analysis (TDA), and conformal prediction. RNNs are employed to capture temporal dependencies in high-dimensional system dynamics, while TDA is used to extract structural features of the evolving states. To quantify uncertainty, we integrated conformal prediction to produce reliable, well-calibrated intervals that remain robust in high-rate dynamic environments. The effectiveness of the proposed conformal prediction method is benchmarked against Monte Carlo Dropout, neural network ensemble, and epistemic neural network using experimental data from the Dynamic Reproduction of Projectiles in Ballistic Environments for Advanced Research (DROPBEAR) testbed. Results demonstrate that the proposed framework produces accurate multi-step forecasts with well-calibrated uncertainty bounds, achieving 92.1% empirical coverage and outperforming other methods in calibration. Specifically, Monte Carlo Dropout achieved 34.1% coverage, neural network ensemble reached 74.3%, and the epistemic neural network yielded 52.3%. This highlights the superior reliability of the proposed approach for monitoring and decision-making in tight latency systems. This work makes three key contributions: (i) the first systematic benchmarking of conformal prediction for real-time multi-step forecasting in high-rate mechanical systems, (ii) a multi-horizon uncertainty calibration analysis under variable excitation conditions, and (iii) the integration of TDA-derived features with RNN-based forecasting to enable reliable uncertainty quantification suitable for digital-twin applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Conformalized multi-step forecasting for high-rate dynamic systems

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Yang Kang Chua,
  • Tingkai Li,
  • Sina Navidi,
  • Arman Razmarashooli,
  • Mohammad Mundiwala,
  • Daniel A. Salazar Martinez,
  • Metrid Okumu,
  • Chao Hu,
  • Simon Laflamme,
  • Paul T. Schrader

摘要

Engineering systems that experience dynamic events with amplitudes exceeding 100 times the force of gravity near the Earth’s surface, typically with durations less than 100 ms, are defined as high-rate dynamic systems. These thresholds are representative of conditions encountered in advanced aerospace structures and impact reduction systems. Such systems are characterized by large uncertainties in external loads, high levels of nonstationarity, and unmodeled dynamics arising from configuration changes. To ensure their safe and reliable operation, there is a critical need for forecasting near-future states at high rates. Such predictive capability enables timely feedback and control actions that mitigate risks and prevent catastrophic failures. This paper presents a framework for high-rate multi-step forecasting that combines recurrent neural networks (RNNs), topological data analysis (TDA), and conformal prediction. RNNs are employed to capture temporal dependencies in high-dimensional system dynamics, while TDA is used to extract structural features of the evolving states. To quantify uncertainty, we integrated conformal prediction to produce reliable, well-calibrated intervals that remain robust in high-rate dynamic environments. The effectiveness of the proposed conformal prediction method is benchmarked against Monte Carlo Dropout, neural network ensemble, and epistemic neural network using experimental data from the Dynamic Reproduction of Projectiles in Ballistic Environments for Advanced Research (DROPBEAR) testbed. Results demonstrate that the proposed framework produces accurate multi-step forecasts with well-calibrated uncertainty bounds, achieving 92.1% empirical coverage and outperforming other methods in calibration. Specifically, Monte Carlo Dropout achieved 34.1% coverage, neural network ensemble reached 74.3%, and the epistemic neural network yielded 52.3%. This highlights the superior reliability of the proposed approach for monitoring and decision-making in tight latency systems. This work makes three key contributions: (i) the first systematic benchmarking of conformal prediction for real-time multi-step forecasting in high-rate mechanical systems, (ii) a multi-horizon uncertainty calibration analysis under variable excitation conditions, and (iii) the integration of TDA-derived features with RNN-based forecasting to enable reliable uncertainty quantification suitable for digital-twin applications.