CNN-based Controller for Multi-DoF Prosthetic Wrist using sEMG Data during Activities of Daily Living

Controllers based on surface EMG (sEMG) data and pattern recognition are widely investigated methods for prosthetic arms with multiple Degrees of Freedom (DoF). Most of these controllers have been trained with movements that are artificially performed for training the machine learning models. In real-life scenarios, the output of the models could be poor-performed due to various arm postures, duration of the movement, and range of motion. Thus, the suggested work introduces a framework to generate a controller for a multi-DoF prosthetic wrist, trained by raw sEMG data collected during Activities of Daily Living (ADL) tasks. During ADL tasks, a motion capture system is used to label the kinematic data of the subject's wrist motion, which is trained for a deep Convolution Neural Networks (CNN) model. The paper focuses on two major functional wrist movements: Pronation-Supination and Dart throwing movement (DTM). Further, a wrist controller design based on multiple CNN models is proposed, which would directly map the sEMG signals to the joint velocities of the wrist. The prosthetic wrist controller is designed based on the data of eight participants and its performance is evaluated in Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Pearson Correlation. A novel kinematic approach to calculating the multi-plane DTM angles is also presented. Further, the model proposes a framework for integrating classification and regression that is based on real-world ADL data. The presented work also includes a robust case study on investigating the effect of data from different heights of ADL tasks.