毕业生知识库

人-机自然交互与协作是机器人领域中重要的科学问题。人机接口(Human Machine Interface, HMI)是其中的核心环节之一，是机器人感知人体运动意图的载体。及时、准确的运动意图识别是该领域的研究热点。然而，现有人机接口技术在非理想环境中人体运动意图实时识别的鲁棒性与可靠性依然存在不足。针对此问题，本研究提出了基于生物阻抗断层扫描（Electrical Impedance Tomography, EIT）信号和肌肉骨骼模型的上肢意图识别方法，从HMI信号源和算法两个方面展开研究。一方面，EIT技术可以通过非侵入的方式实现对扫描截面内深层肌肉活动的检测，可从新的维度为算法提供更多信息，为提升算法鲁棒性提供可能；另一方面，肌肉骨骼模型满足生物力学原理，可减轻对训练数据的依赖，提高算法的可靠性。

本文以EIT信号作为信号源，结合机器学习和肌肉骨骼模型的方法，重点解决上肢运动参数的识别与估计问题。首先，我们基于并行采样方法，研发了高速EIT信号采样系统，提出了基于EIT空间电导率信息的腕关节运动自适应解码算法，解决了重新穿戴场景下实时意图识别准确率低的问题。其次，本研究提出了基于运动分析的EIT信号特征提取与肌肉辨识方法，并建立了腕关节肌肉骨骼运动学模型，实现了腕关节2自由度的角度估计。第三，本文进一步地探究了人体上肢肌肉空间、关节空间和笛卡尔空间的刚度关系，提出了基于EIT和上肢肌骨模型的手臂末端刚度估计模型。最后，通过菲茨定律测试和机器人实时控制等实验，检验了基于EIT信号的人机接口以及上肢运动识别算法的有效性。

实验结果表明：第一，本研究提出的EIT信号采样系统帧率可达104fps，相比现有用于运动意图识别的穿戴式EIT系统，帧率提升了3~10倍，提高了系统的时域响应能力。第二，该HMI系统结合腕关节运动自适应解码算法，在标准菲茨定律测试中，吞吐量为1.03 ± 0.10bits/s，与现有基于肌肉信号的HMI技术相比处于相同水平；在存在重新穿戴干扰的条件下，测试吞吐量为1.01 ± 0.10bits/s，优于现有技术，验证了该人机接口的实时性和方法的有效性。第三，基于EIT信号和腕关节肌肉骨骼模型的方法，以较低的训练成本实现了不同握力等级下的关节角度有效估计，有利于模型的应用部署。第四，基于EIT信号和上肢肌肉骨骼模型的末端刚度估计模型与领域内先进方法相比，可实现较低的估计误差（13.85% ± 3.21%）。

本研究系统性地探究了基于EIT驱动的肌骨模型，提出了基于EIT的HMI及上肢运动参数识别与估计的方法，进一步解决了HMI从理论走向实际应用的共性问题，为人机协作任务提供了更多有效交互信息，有助于实现柔顺自然的机器人控制。

Intuitive human-robot interaction and cooperation is an important scientific problem in the field of robotics. Human Machine Interface (HMI) is one of the core parts, which  serves as the bridge between the human sensorimotor system and the robot system. One critical issue in developing the HMI is to produce as high accuracy (low errors) as possible with the fastest time
response. However, the robustness and reliability of the existing HMI technology for real-time human motion intention recognition in non-ideal environments are still insufficient. To solve this problem, this study proposes an upper limb intention recognition method based on EIT and musculoskeletal model, in which HMI signal source and algorithms are studied. On the one hand, EIT signals can monitor the deep muscle activity in the cross-section in a non-invasive way, which can provide more information  from a new dimension and provide the possibility to improve the robustness of the algorithm. On the other hand, the musculoskeletal model satisfies biomechanical principles, which can reduce the dependence on the training data and improve the reliability of the algorithms. 

In this study, EIT signal is used as the HMI signal source, and machine learning and musculoskeletal model methods are combined to identify and estimate the motion parameters of the human upper limb. Firstly, a high-framerate parallel EIT signal acquisition system and an adaptive calibration mechanism of model parameters are designed and implemented to resolve the problem of real-time intention recognition in sensor re-wearing scenarios. Secondly, a novel angle estimation algorithm based on the EIT signal and wrist musculoskeletal kinematics model is proposed to estimate 2-DoF angle of wrist. Thirdly, we further explore the stiffness transformation among muscle space, joint space and cartesian space of human upper limb. A novel method based on EIT signals and upper limb musculoskeletal models is proposed to estimate the human arm endpoint stiffness. Finally, the effectiveness of the EIT-based HMI and the recognition algorithms of upper limb motion are tested by online Fitts' law tests and robot  control experiments.

According to the experimental results, the following conclusions can be drawn. Firstly, The framerate of the proposed EIT system is 104 fps. The time response of the proposed EIT system achieved a significant improvement for forearm motion recognition based on EIT, which was around 3~10 times faster in the framerate than other related studies. Secondly, The average throughputs (TPs) of Fitts' law tests are 1.03 ± 0.10 bits/s and 1.01 ± 0.10 bits/s without
and with sensor re-donning, respectively, which were comparable to the TPs of sEMG-based studies. The results showed the promise of the EIT-based interface on real-time human motion
intent recognition. Thirdly, the proposed muscoloskeletal model based angle estimation algorithm realizes the effective estimation of joint angles under different grip force levels with less training data, reduces the training cost, and is conducive to the application and deployment of the model. Lastly, compared to related stiffness estimation methods, the proposed model-based algorithm can effectively estimate the arm endpoint stiffness with lower error (13.85% ± 3.21%).

This study systematically explored the EIT-based musculoskeletal model, and proposed the identification and estimation methods of upper limb motion parameters based on EIT-based HMI. This study further solves the common problem of HMI from theory to practice and can provide more effective interaction information for human-robot collaborative tasks to achieve compliant and intuitive robot control.