Scholarlink

<p>Simulating long-horizon state variable trajectories using neural network models is a challenging task due to error accumulation from auto-regressive predictions. Poor generalization exacerbates this issue, particularly under noisy and disturbed data conditions. This paper addresses these challenges by introducing a robust methodology that combines physics-based modeling with data-driven approaches to improve long-horizon simulation accuracy. Specifically, we propose a custom loss function designed to minimize auto-regressive prediction errors across extended time horizons, ensuring alignment between predicted trajectories and actual state measurements. Additionally, we leverage transfer learning by pretraining on simulated data generated from a dynamical model, where parameters are estimated using Bayesian inference, to incorporate prior physical knowledge into the learning process. Our methodology is demonstrated on the lateral dynamics of a vehicle using real-world noisy measurements, achieving competitive performance compared to traditional methods. This work highlights the effectiveness of combining custom loss design and transfer learning to enhance long-term simulation robustness.</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

Physics-guided approach with transfer learning in vehicle lateral dynamics

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • Fabien Lionti,
  • Nicolas Gutowski,
  • Sébastien Aubin,
  • Philippe Martinet

摘要

Simulating long-horizon state variable trajectories using neural network models is a challenging task due to error accumulation from auto-regressive predictions. Poor generalization exacerbates this issue, particularly under noisy and disturbed data conditions. This paper addresses these challenges by introducing a robust methodology that combines physics-based modeling with data-driven approaches to improve long-horizon simulation accuracy. Specifically, we propose a custom loss function designed to minimize auto-regressive prediction errors across extended time horizons, ensuring alignment between predicted trajectories and actual state measurements. Additionally, we leverage transfer learning by pretraining on simulated data generated from a dynamical model, where parameters are estimated using Bayesian inference, to incorporate prior physical knowledge into the learning process. Our methodology is demonstrated on the lateral dynamics of a vehicle using real-world noisy measurements, achieving competitive performance compared to traditional methods. This work highlights the effectiveness of combining custom loss design and transfer learning to enhance long-term simulation robustness.