Predicting the Level of Co-Activation of One Muscle Head from the Other Muscle Head of the Biceps Brachii Muscle by Linear Regression and Shallow Feedforward Neural Networks

One of the challenges in close-to-body robotics is the intuitive control of exoskeletal devices which requires lag-free responses of its actuated joints. A frequently used signal domain to satisfy the required control properties is surface electromyography (sEMG). By using a Hill-type model of the muscle mainly responsible for the movement of a biological joint, which is excited by the corresponding sEMG of this muscle, the joint movement can be pre-calculated. If the muscle internal delays are used, this information can be used for an intuitive and lag-free control. So far, biomechanical limb and joint models including Hill-type muscle submodel were used. In current studies, state-of-the-art machine learning models are evaluated for this problem. Both types, classical and machine learning models, depend on the measured sEMG signals of all muscle heads of a relevant muscle and on their respective signal quality. This work introduces a method to train a virtual sEMG-sensor as a replac ement for the real sEMG signal of a muscle head, thus reducing the number of real sensor electrodes on a given muscle. The virtual sensor is trained based on data from the remaining sensor. This method allows to compare the measured sEMG signal with the virtual sensor output to assess the measured signal. Furthermore, this study explains the training process and evaluates the use of the virtual sensor in a biomechanical limb model.