毕业生知识库

       随着我国人口老龄化趋势加剧，下肢运动障碍人群日益庞大，这对社会养老 与康复医疗系统造成了压力。外骨骼机器人作为可穿戴式助力设备，能够为人体 提供辅助力，从而减轻穿戴者运动负担，提升人体运动机能，是缓解社会养老压 力与康复医疗压力的有效途径之一。此外，下肢外骨骼机器人在军事、抢险救灾 和工业搬运等领域都能发挥重要作用。本文针对老年人及下肢运动障碍患者等 弱能人群研制了一款髋关节外骨骼助力机器人，并围绕下肢助力外骨骼人体运 动意图感知关键技术问题，从人体连续步态相位估计、人体三维动作预测、人体关节力矩估计等方面展开深入研究，旨在为外骨骼机器人智能感知人体运动意 图提供理论指导与技术支撑。本文主要创新点与贡献如下：

       一、针对在多种周期性运动模式下的人体运动意图感知不稳定的问题，提出 了一种基于自适应振荡器（Adaptive Frequency Oscillator，AFO）的连续步态相 位估计方法。首先，构建了一种基于长短时记忆网络（Long-Short Term Memory， LSTM）的方法，用于识别不同的步态模式。然后，设计了一种基于改进 AFO 模型 的人体连续步态相位估计方法，能够在多种运动模式频繁切换的情况下保持稳 定运行。最后设计了一种基于阈值的步态事件检测方法，用于同步 AFO 模型估 计的人体步态相位。此外，针对 AFO 模型提取到的人体步态相位没有考虑到个 体差异的问题，使用基于 LSTM 和动态时间规整（Dynamic Time Warping，DTW） 算法的连续步态相位估计方法，能够根据人体步态数据拟合得到个性化的步态 相位。实验数据表明，本章提出的基于改进 AFO 模型的方法与基于 LSTM+DTW 的方法都能有效地估计人体连续步态相位。

        二、针对非周期性运动下的人体运动意图感知容易误判以及基于物理信号 的传感器存在延迟的问题，提出了一种基于下肢姿态估计与动作预测的人体动 作意图感知方法，首次实现了仅使用稀疏的 IMU 数据预测人体下肢三维动作。 首先，构建了一种两阶段的网络模型（Motion Forecast Network，MoFCNet），用 于估计和预测人体下肢运动。其中阶段 I 根据历史 IMU 观测数据预测未来人体 关节位置、关节速度等中间变量，阶段 II 根据预测的中间变量估计人体未来关节 角度序列。然后，设计了基本构建块 Stack，包含分解单元、改进的线性单元等， 降低了网络的计算复杂度与内存消耗，能够用于网络的两个阶段。最后，利用基 于光学捕捉的高精度大型人体运动数据集，结合人体参数化模型 SMPL（Skinned Multi-Person Linear Model），生成了大量虚拟 IMU 运动数据，解决了基于 IMU 捕捉的人体运动数据不足的问题。大量实验数据表明，所提出的网络能够有效地 根据 IMU 数据估计和预测人体未来一段时间内的下肢动作。

       三、针对传统人体动力学估计算法获取人体关节力矩依赖测力板或足底压 力传感器，从而导致使用场景受限的问题，提出了一种基于下肢关节力矩估计的 人体运动意图感知方法。首先，设计了神经网络模块，使用挂载在肢体上的 IMU 数据估计人体下肢关节角度、速度、位置等运动学信息。然后，根据上述信息估 计地面反作用力与人体下肢关节力矩等动力学信息。最后，构建了一种新颖的基 于物理约束的网络模块，将人体运动过程中的摩擦力、关节位置等物理约束转换 为神经网络的损失函数，能够改善网络估计的人体运动学与动力学结果。实验数 据表明，所提出的仅使用 IMU 的人体动力学估计方法能够有效地估计地面反作 用力和人体下肢关节力矩。

       四、针对老年人及下肢运动障碍患者等弱能人群行动受限的问题，研发了一款轻量型髋关节下肢助力外骨骼机器人样机。通过对人体关节运动的分析，研发 了一款符合人体工程学的 4 自由度髋关节外骨骼机器人，其中两个主动自由度 位于人体矢状面上，由电机驱动，两个被动自由度位于人体冠状面上。本文分别 介绍了机器人机械系统、电气系统和软件系统设计方案，并给出了仿真和实际实 验结果。

    With the increasing trend of aging population in China, the population of individuals with lower limb mobility impairments is growing rapidly. This places significant pressure on the social care and medical rehabilitation systems. Exoskeleton robots, when worn externally on the limbs and torso of individuals, serve as assistive devices capable of providing supplementary support to the human body. They alleviate the physical burden on the wearer and enhance their motor functions, thus representing an effective approach to alleviating the pressures on elderly care and medical rehabilitation systems. Additionally, lower limb exoskeleton robots can play an important role in areas such as military, disaster relief, and industrial transportation. This paper focuses on the development of a hip exoskeleton assistive robot for vulnerable populations such as the elderly and individuals with lower limb mobility impairments. It addresses key technological challenges related to lower limb exoskeleton perception. The research delves into various aspects, including continuous gait phase estimation, human motion prediction, and estimation of human joint torques. The primary innovations and contributions are as follows:

    1. A continuous gait phase estimation method based on Adaptive Frequency Oscillator (AFO) is proposed to address the issue of unstable perception of human motion intentions in various periodic motion patterns. Firstly, a method based on Long-Short Term Memory (LSTM) is introduced for recognizing different gait patterns. Then, a continuous human gait phase estimation method based on an improved AFO model is presented, which maintains stable operation in situations where there are frequent transitions between multiple motion patterns. Finally, a threshold-based gait event detection method is proposed for synchronizing the human gait phases obtained from the AFO model. In addition, a continuous gait phase estimation method based on LSTM and Dynamic Time Warping (DTW) algorithm is introduced to address the issue that the gait phases extracted by the AFO model do not account for individual differences. This method allows for the personalized fitting of gait phases based on individual gait data. Experimental data demonstrates that both the DAFO-based method and the LSTM+DTW-based method proposed in this chapter effectively estimate the continuous human gait phase.

    2. A method for estimating lower limb posture and motion prediction in human movements is proposed to address the challenges associated with the perception of human motion intentions in non-periodic movements, as well as issues arising from the delay in sensor data based on physical signals. For the first time, the proposed method only uses sparse IMU data to predict 3D human lower limb movements. Firstly, a two stage network called the Motion Forecast Network (MoFCNet) is introduced for estimating and forecasting human lower limb movements. Stage I utilizes historical IMU observation data to forecast future human joint positions, joint velocities, and other intermediate variables. Stage II estimates future sequences of human joint angles based on the predicted intermediate variables. Then, a fundamental building block called Stack is proposed, incorporating decomposition modules and improved linear modules. This building block reduces computational complexity and memory consumption, making them applicable to both stages of the network. Finally, high-precision, large-scale human motion data captured through optical sensors, in conjunction with the parametric human model (SMPL), is employed to generate a vast amount of virtual IMU motion data. This approach addresses the issue of limited human motion data based on IMU sensors. Extensive experimental data demonstrates the effectiveness of the proposed network in accurately estimating and forecasting lower limb movements in the future.

    3. A method for estimating lower limb joint torques using only wearable IMUs is proposed to address the limitations of traditional biomechanical estimation algorithms that rely on force plates or plantar pressure sensors for estimating joint torques of humans. Firstly, IMU data is used to estimate kinematic information such as human lower limb joint angles, velocities, and positions. Then, based on this information, ground reaction forces and human lower limb joint torques are estimated. Finally, a physics constrained optimization module is introduced, converting physical constraints, such as friction forces during human movement, into loss functions for training the network. This optimization module enhances the accuracy of the estimations of human biomechanics, both kinematics and dynamics. Experimental data demonstrates that the proposed IMU-only method is effective in accurately estimating ground reaction forces and human lower limb joint torques, eliminating the need for traditional force measurement equipment.

    4. A prototype of lightweight hip assisted exoskeleton robot has been developed to address mobility limitations in elderly individuals and patients with lower limb movement disorders. The robot features a 4-degree-of-freedom hip exoskeleton designed according to ergonomic principles by analyzing human joint motion. Two active degrees of freedom are situated on the sagittal plane of the human body and are driven by an electric motor, while two passive degrees of freedom are positioned on the coronal plane. This article delineates design solutions for the mechanical, electrical, and software systems of the robot, presenting simulation alongside actual experimental result.