Hybrid Models for High-precision sEMG-based Joint Torque / Movement Prediciton for Wearable Robotics

为了使生物信号驱动的机器人身体支持系统（例如外骨骼）的发展能够不断适应用户的行为，必须解决基本的挑战。提出的项目的目标是通过引入混合扩展模型拓扑来提高肌电驱动的肌肉关节模型的预测精度。目前肌电信号驱动的基于模型的方法所面临的第一个挑战是预测扭矩的不准确性，这些扭矩主要发生在静态姿势、缓慢的关节旋转（低肌电信号活动时的机械漂移）和肢体动态运动的瞬态。为了弥补这些不准确性，可配置的滑动摩擦模型和关节几何自适应模型将被添加到已建立的肌肉-关节设置中。通过探索性地在子模型上并行添加灰盒模型，在迭代优化过程中进一步改进。预测关节运动的第二个挑战是肌电驱动的肌肉模型只产生力作为输出参数。这些力不能直接分配给关节力矩，噪声关节力矩不能用来确定绝对关节角。因此，仅仅观察肌电图并不能提供关节当前角度的信息，关节与肌肉相连。然而，在肌电图阵列的帮助下，可以估计肌肉神经支配区（IZ）的绝对位置，肌肉-肌腱交叉点以及肌肉纤维长度（例如，使用波束形成算法）。将这些信息与肌肉、肌腱和关节的子模型相结合，可以有效地估计肌肉和肌腱的长度以及关节的角度。基于肌电图测量，这可以消除位置漂移，而无需显式测量关节角度。由此产生的肌电驱动的肌肉关节模型也支持第三个挑战的解决方案，即减少矫形设备/可穿戴机器人和生物手臂（运动意图）产生的扭矩之间的延迟。由于扩展肌电驱动的肌肉-关节模型预测下一时刻的关节扭矩，因此将开发一种新的控制方法，该方法使用该扭矩预测和矫形器模型来预测相移。基于这种预测，控制器可以影响转矩产生，以适应相移和延迟。对支持水平的影响仍然取决于一个额外的支持水平控制器。模型在现实世界的适用性将基于一个最小驱动矫形器进行评估，该矫形器将在项目期间建造。为了参数化和评估，将使用运动和肌电数据，这些数据将在三个实验活动中记录。

To enable the development of biosignal-driven robotic body support systems (e.g., exoskeletons) which continuously adapt to the behaviour of the user, fundamental challenges have to be resolved. The goal of the proposed project is to improve the prediction accuracy of EMG-driven muscle-joint models by introducing hybrid extensions to the model topology. The first challenges in current EMG-driven, model-based approaches are prediction inaccuracies with respect to torque which occur mainly in static-postures, slow joint rotations (mechanical drift at low EMG-activity) and transients to dynamic limb movements. To compensate for these inaccuracies, configurable sliding friction models and adaptive models of the joint geometry will be added to the established muscle-joint-setup. A further improvement will be achieved in an iterative optimization process through explorative adding of grey box models in parallel to the submodels. A second challenge in the prediction of joint movements is the fact that EMG-driven muscle models produce solely forces as output parameters. These forces cannot directly be assigned to joint torques and noisy joint torques cannot be used to determine the absolute joint angle. Hence, the mere observation of EMGs does not provide information on the current angle of the joint, which is connected to the muscles. However, with the help of EMG-arrays, the absolute position of muscle innervation zones (IZ), muscle-tendon intersections and thus also of muscle fiber lengths can be estimated (e.g., with beamforming algorithms). Using this information in combination with the submodels of muscles, tendons, and joints, a valid estimate of muscle and tendon lengths as well as the joint angle will be achieved. Based on EMG-measurements, this allows the elimination of position drift without explicit measurement of the joint angle. The resulting EMG-driven muscle-joint models also support a solution for a third challenge which is the latency reduction between the torques generated by an orthotic device/wearable robot and the biological arm (movement intention). Since the extended EMG-driven muscle-joint model predicts the joint torque in the next instant of time, a new control approach will be developed which uses this torque prediction and a model of the orthotic device to predict the phase shift. Based on this prediction, the controller can influence the torque generation in order to adapt the phase shift and hence the latency. Influences on the level of support are still subject to an additional support-level-controller. The real world applicability of the models will be evaluated based on a minimally actuated orthosis which will be built during the project. For parameterization and evaluation, movement- and EMG-data will be used which will be recorded in three experimental campaigns.