MuscleNET: mapping electromyography to kinematic and dynamic biomechanical variables by machine learning

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