Research Papers 论文研究 12h ago Updated 1h ago 更新于 1小时前 41

Toward Reliable Railway-Bogie Response Prediction Using Multifidelity TDNN and Physics-Informed Residual Learning 基于多保真TDNN和物理信息残差学习的铁路转向架响应可靠预测

A multifidelity correction method merges low-fidelity multibody simulations with high-fidelity roller-rig measurements to predict railway-bogie responses across untested operating conditions A Time-Delay Neural Network (TDNN) captures condition-dependent simulation trends, while a residual-correction network learns reproducible discrepancies not explained by the baseline Physics-informed constraints via an effective dynamic-balance equation enforce consistency in inertia, damping, stiffness, and 提出多保真度铁路转向架响应校正方法,将多体仿真历史作为低保真信息、辊台测量作为高保真证据进行融合学习 使用时延神经网络(TDNN)建模条件依赖的仿真趋势,通过残差校正网络学习基线无法解释的可复现响应分量 引入有效动态平衡方程约束惯性、阻尼、刚度和外力差异,结合残差匹配、时间平滑性和位移-加速度一致性构建复合训练目标 在385 km/h未见过条件下验证,平均决定系数达0.8197,NRMSE为4.6055%,NMAE为1.9297%

52
Hot 热度
68
Quality 质量
55
Impact 影响力

Analysis 深度分析

TL;DR

  • A multifidelity correction method merges low-fidelity multibody simulations with high-fidelity roller-rig measurements to predict railway-bogie responses across untested operating conditions
  • A Time-Delay Neural Network (TDNN) captures condition-dependent simulation trends, while a residual-correction network learns reproducible discrepancies not explained by the baseline
  • Physics-informed constraints via an effective dynamic-balance equation enforce consistency in inertia, damping, stiffness, and external forcing differences between simulation and physical systems
  • The corrected model achieves a mean R² of 0.8197, NRMSE of 4.61%, and NMAE of 1.93%, with validated performance at a held-out 385 km/h condition
  • The training objective jointly optimizes physics constraints, residual matching, temporal smoothness, and displacement-acceleration consistency terms

Why It Matters

This work demonstrates a practical framework for bridging the gap between computational simulation and physical experimentation in safety-critical engineering domains, addressing the fundamental problem that model calibration at limited conditions does not generalize. For AI practitioners working in scientific machine learning, it showcases how physics-informed constraints can be integrated into neural architectures to improve generalization beyond training data—a growing priority as the field moves toward more reliable and trustworthy AI systems.

Technical Details

  • Multifidelity architecture: Multibody simulation histories serve as low-fidelity inputs, while roller-rig experimental measurements provide high-fidelity anchors, with experiment-anchored fidelity assignment determining the weight and contribution of each source
  • TDNN-based trend modeling: A Time-Delay Neural Network encodes condition-dependent simulation trends, and a development-fitted amplitude alignment establishes the low-fidelity baseline response
  • Residual-correction network: A dedicated neural network models the reproducible response component unexplained by the baseline, which is then added to the baseline to produce the corrected prediction
  • Physics-informed dynamic-balance constraint: An effective dynamic-balance equation constrains the learned discrepancy by explicitly representing differences in inertia, damping, stiffness, and external forcing between simulated and physical systems
  • Multi-objective training: The loss function combines the physics constraint, residual matching, temporal smoothness regularization, and selectively applied displacement-acceleration consistency terms to ensure physically plausible predictions

Industry Insight

  • Physics-informed neural architectures that explicitly encode domain equations (e.g., dynamic balance) offer a compelling path toward generalization beyond calibration conditions, particularly valuable in transportation and infrastructure sectors where exhaustive testing is impractical or unsafe
  • The multifidelity approach—leveraging inexpensive simulations as baselines corrected by sparse high-fidelity data—provides a cost-effective template for other engineering domains seeking to reduce reliance on expensive physical prototyping and testing
  • The demonstrated accuracy at a held-out 385 km/h condition suggests this methodology could enable predictive maintenance and safety certification at operating speeds that are currently difficult or impossible to test exhaustively, potentially reshaping validation workflows in railway engineering

TL;DR

  • 提出多保真度铁路转向架响应校正方法,将多体仿真历史作为低保真信息、辊台测量作为高保真证据进行融合学习
  • 使用时延神经网络(TDNN)建模条件依赖的仿真趋势,通过残差校正网络学习基线无法解释的可复现响应分量
  • 引入有效动态平衡方程约束惯性、阻尼、刚度和外力差异,结合残差匹配、时间平滑性和位移-加速度一致性构建复合训练目标
  • 在385 km/h未见过条件下验证,平均决定系数达0.8197,NRMSE为4.6055%,NMAE为1.9297%

为什么值得看

本文展示了物理信息机器学习在工业仿真校准中的前沿应用,为铁路工程提供了可泛化的响应预测方案。多保真度学习框架有效弥合了仿真与实测之间的差距,对需要高可靠性预测的工业场景具有重要参考价值。

技术解析

  • 多保真度校正框架:将多体动力学仿真历史作为低保真基线,辊台试验测量数据作为高保真锚点,通过实验锚定的保真度分配实现跨条件预测,解决有限校准条件无法保证全域精度的问题
  • TDNN残差学习架构:使用时延神经网络捕捉条件依赖的仿真趋势,开发拟合的振幅对齐定义低保真基线,残差校正网络学习基线无法解释的可复现响应分量并叠加至基线
  • 物理约束训练目标:引入有效动态平衡方程约束惯性、阻尼、刚度和外力差异,结合残差匹配、时间平滑性和选择性位移-加速度一致性项构建复合损失函数,确保物理一致性
  • 实验验证结果:在重构案例中达到平均决定系数0.8197、平均NRMSE 4.6055%、平均NMAE 1.9297%,并在385 km/h未见过条件下验证了预测准确性

行业启示

  • 物理信息机器学习正在成为工业仿真校准的关键技术路径,将领域知识嵌入神经网络可显著提升模型泛化能力和可解释性,降低对大规模实测数据的依赖
  • 多保真度学习框架为高成本物理试验与低成本仿真之间的数据鸿沟提供了有效解决方案,适用于航空航天、汽车工程、轨道交通等需要高可靠性预测的工业领域
  • 铁路行业对预测精度的高要求推动了AI与物理模型的深度融合,这种"仿真+实测+AI校正"的范式有望成为工业数字孪生的标准架构,推动预测性维护和智能运维的发展

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Training 训练 Dataset 数据集