肌骨系统多体动力学仿真新进展:从数据驱动到数据-物理耦合驱动  

New Advances in Multibody Dynamics Simulation of the Musculoskeletal System:From Data-Driven to Data-Physics Hybrid Approaches

作　　者：陈文轩 任韦燕 姚杰[1] 蒲放[1] CHEN Wenxuan;REN Weiyan;YAO Jie;PU Fang(Key Laboratory of Biomechanics and Mechanobiology,Ministry of Education,Key Laboratory of Innovation and Transformation of Advanced Medical Devices,Ministry of Industry and Information Technology,National Medical Innovation Platform for Industry-Education Integration in Advanced Medical Devices,School of Biological Science and Medical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]生物力学与力生物学教育部重点实验室,高端医疗装备与器械创新及转化工业和信息化部重点实验室,国家医学攻关(医工结合方向)高端医疗装备与器械产教融合创新平台,北京航空航天大学生物与医学工程学院,北京100191

出　　处：《医用生物力学》2025年第2期255-262,共8页Journal of Medical Biomechanics

基　　金：国家重点研发计划项目(2023YFC3604503);国家自然科学基金项目(12072019)。

摘　　要：肌骨系统多体动力学仿真是解析人体运动生物力学机制的重要工具,其发展正从传统物理模型驱动向数据驱动、物理-数据混合驱动转变,本文对相关的新进展进行综述。传统的基于物理模型的肌骨系统多体动力学仿真已经在仿真度、优化方法、软件工具等方面发展非常充分,但由于依赖复杂实验数据与高计算量的微分方程求解,实际应用仍有很大限制。近年来,随着深度学习技术的快速发展,基于数据驱动的肌骨系统多体动力学仿真,已经在关节角度、位姿、地面反作用力、关节力矩、肌力的高效预测,以及外骨骼的控制算法等方面展现出了巨大优势,显著降低了参数测量复杂度。然而,纯数据驱动模型存在泛化性不足与生物力学规律失真的问题。为此,物理-数据耦合驱动模型通过嵌入牛顿-欧拉方程、肌肉本构方程等物理约束,结合物理信息神经网络等混合架构,在提升预测精度的同时保障解的生物合理性。当前研究仍面临跨尺度物理方程整合不足、多关节协同建模困难等挑战。未来需进一步优化模型架构,融合运动学无标记捕捉技术,并探索多尺度物理规律与个性化参数反演,以推动智能康复、运动评估等领域的应用转化。Multibody dynamics simulation of the musculoskeletal system is an essential tool for analyzing the biomechanical mechanisms underlying human motion.Recent research trends have shifted from traditional physics-based models toward data-driven or data-physics hybrid frameworks.This review presents the latest 255 developments in these areas.Physics-based multibody dynamics simulations have undergone significant progress in terms of simulation fidelity,optimization algorithms,and software tools.However,their practical implementation remains constrained by the need for complex experimental data and the computational expense of solving differential equations.Conversely,data-driven method bolstered by advancements in deep learning have demonstrated remarkable efficiency in predicting joint angles,postures,ground reaction forces,joint torques,and muscle forces,as well as developing control algorithms for exoskeletons.However,despite these advantages,data-driven approaches face challenges such as limited generalizability and potential violation of biomechanical principles.To address these limitations,data-physics hybrid approaches(e.g.,physics-informed neural network,PINN)which integrate physical constraints(e.g.,Newton-Euler equations,muscle constitutive laws)with data-driven architectures have been developed.This synergy enhances prediction accuracy while preserving the biological plausibility of solutions.Nevertheless,critical challenges persist,including the integration of multi-scale physical equations and the modeling of multi-joint coordination dynamics.Future research should prioritize:optimizing hybrid model architectures to balance computational efficiency and mechanistic accuracy,incorporating markerless motion capture techniques to improve real-world applicability,exploiting multi-scale physics and personalized parameter inversion to advance precision rehabilitation and motion analysis.These efforts will foster innovations in intelligent rehabilitation systems,clinical motion assessment,and related translationa

关 键 词：肌骨系统 多体动力学 数据驱动模型 物理信息神经网络 生物力学建模 

分 类 号：R318.01[医药卫生—生物医学工程]