Objective. This paper proposes machine learning models for mapping surface electromyography (sEMG) signals to regression of joint angle, joint velocity, joint acceleration, joint torque, and activation torque. Approach. The regression models, collectively known as MuscleNET, take one of four forms: ANN (Forward Artificial Neural Network), RNN (Recurrent Neural Network), CNN (Convolutional Neural Network), and RCNN (Recurrent Convolutional Neural Network). Inspired by conventional biomechanical muscle models, delayed kinematic signals were used along with sEMG signals as the machine learning model's input; specifically, the CNN and RCNN were modeled with novel configurations for these input conditions. The models' inputs contain either raw or filtered sEMG signals, which allowed evaluation of the filtering capabilities of the models. The models were trained using human experimental data and evaluated with different individual data. Main results. Results were compared in terms of regression error (using the root-mean-square) and model computation delay. The results indicate that the RNN (with filtered sEMG signals) and RCNN (with raw sEMG signals) models, both with delayed kinematic data, can extract underlying motor control information (such as joint activation torque or joint angle) from sEMG signals in pick-and-place tasks. The CNNs and RCNNs were able to filter raw sEMG signals. Significance. All forms of MuscleNET were found to map sEMG signals within 2 ms, fast enough for real-time applications such as the control of exoskeletons or active prostheses. The RNN model with filtered sEMG and delayed kinematic signals is particularly appropriate for applications in musculoskeletal simulation and biomechatronic device control.

Myoelectric signals from the human motor control system can improve the real-time control and neural-machine interface of robotic leg prostheses and exoskeletons for different locomotor activities (e.g., walking, sitting down, stair ascent, and non-rhythmic movements). Here we review the latest advances in myoelectric control designs and propose future directions for research and innovation. We review the different wearable sensor technologies, actuators, signal processing, and pattern recognition algorithms used for myoelectric locomotor control and intent recognition, with an emphasis on the hierarchical architectures of volitional control systems. Common mechanisms within the control architecture include 1) open-loop proportional control with fixed gains, 2) active-reactive control, 3) joint mechanical impedance control, 4) manual-tuning torque control, 5) adaptive control with varying gains, and 6) closed-loop servo actuator control. Based on our review, we recommend that future research consider using musculoskeletal modeling and machine learning algorithms to map myoelectric signals from surface electromyography (EMG) to actuator joint torques, thereby improving the automation and efficiency of next-generation EMG controllers and neural interfaces for robotic leg prostheses and exoskeletons. We also propose an example model-based adaptive impedance EMG controller including muscle and multibody system dynamics. Ongoing advances in the engineering design of myoelectric control systems have implications for both locomotor assistance and rehabilitation.

To physically assist workers in reducing musculoskeletal strain or to develop motor skills for patients with neuromuscular disabilities, recent research has focused on exoskeletons. Designing exoskeletons is challenging due to the complex human geometric structure, the human-exoskeleton wrench interaction, the kinematic constraints, and the selection of power source characteristics. This study concentrates on modeling a 3D multibody upper-limb human-exoskeleton, developing a procedure of analyzing optimal assistive torque profiles, and optimizing the passive mechanism features for desired tasks. The optimization objective is minimizing the human joint torques. Differential-algebraic equations (DAE)s of motion have been generated and solved to simulate the complex closed-loop multibody dynamics. Three different tasks have been considered, which are common in industrial environments: object manipulation, over-head work, and static pointing. The resulting assistive exoskeleton's elevation joint torque profile decreases the specific task's human shoulder torque. The exoskeleton is not versatile or optimal for different dynamic tasks since the passive mechanism produces a specific torque for a given elevation angle. We concluded that designing a passive exoskeleton for a wide range of dynamic applications is impossible.

InverseMuscleNET, a machine learning model, is proposed as an alternative to static optimization for resolving the redundancy issue in inverse muscle models. A recurrent neural network (RNN) was optimally configured, trained, and tested to estimate the pattern of muscle activation signals. Five biomechanical variables (joint angle, joint velocity, joint acceleration, joint torque, and activation torque) were used as inputs to the RNN. A set of surface electromyography (EMG) signals, experimentally measured around the shoulder joint for flexion/extension, were used to train and validate the RNN model. The obtained machine learning model yields a normalized regression in the range of 88–91% between experimental data and estimated muscle activation. A sequential backward selection algorithm was used as a sensitivity analysis to discover the less dominant inputs. The order of most essential signals to least dominant ones was as follows: joint angle, activation torque, joint torque, joint velocity, and joint acceleration. The RNN model required 0.06 s of the previous biomechanical input signals and 0.01 s of the predicted feedback EMG signals, demonstrating the dynamic temporal relationships of the muscle activation profiles. The proposed approach permits a fast and direct estimation ability instead of iterative solutions for the inverse muscle model. It raises the possibility of integrating such a model in a real-time device for functional rehabilitation and sports evaluation devices with real-time estimation and tracking. This method provides clinicians with a means of estimating EMG activity without an invasive electrode setup.

No abstract