Generalized Force Estimation using Machine Learning and Deep Learning with EMG and Motion Data

Abstract

Human movements are accomplished by muscle contractions, resulting in force exertion. Accurate estimation of the generated force is desired in many applications such as robotics, prosthesis control, and sports. The surface electromyogram (EMG) has been widely used for estimating force since the 1950s. However, EMG-force relationship is nonlinear and dynamically changing, and thus the development of an accurate model is a challenging problem for static, and especially dynamic contractions. The objective of this research is to develop accurate force estimation models that are generalizable across individuals and experimental conditions, from static to fully dynamic cases. First, we study intra-subject force modelling during static conditions using fast orthogonal search and a novel channel selection method to reduce the dimensionality and correlation among recorded EMG channels. The results show a significant improvement in performance using the selected channels over the original dataset. To develop generalized models, classical machine learning is first used. Features from EMG are extracted, followed by feature selection using our proposed method, modified sequential feature selection. A bagged tree ensemble (BTE) is then used for force modelling. Next, deep learning is also explored, where convolutional neural networks (CNN) are developed for force estimation from EMG signals in both time and frequency domains. Results show that the CNN outperforms the BTE. To accurately model exerted force under quasi-dynamic and dynamic contractions, where the EMG-force relationship is highly non-linear and complex, a CNN ensemble is proposed to learn the EMG-force relationships from multi-domain and multi-modal inputs. The proposed model uses individual deep CNNs to extract robust representations from EMG in time and frequency domains, and from kinematic information recorded by an inertial measurement unit (IMU). The learned representations are then fused at the feature level and fed to a neural network to estimate force. Our CNN ensemble method achieved highly accurate force estimation results for different experimental conditions in both intra- and inter-subject schemes. Our results indicate the superiority of our approach over other methods in the literature, and further validate the need for including EMG in both domains as well as kinematic information for accurate and generalized force estimation.