Generalized EMG-based isometric contact force estimation using a deep learning approach

Abstract EMG-based force estimation is generally done in a subject specific manner. In this paper, we explore force estimation in a manner generalizable across individuals, where the EMG signals are recorded from the long head and the short head of biceps brachii, and the brachioradialis, under isometric elbow flexion, while the contact force is measured at the wrist. Deep convolutional neural networks (CNN), which utilize feature-level fusion of representations learned from high-density (HD) EMG in the time and frequency domains, are developed. The performance of the proposed solution (CNN-FLF) is compared to a number of baselines including CNNs with input-level fusion of HD-EMG data in the time and frequency domains, CNNs which estimate force from time or frequency domain EMG data separately, and a number of classical machine learning methods, which use hand-crafted features extracted from the EMG signals. Results show that the CNN-FLF, with optimized hyper-parameters, outperformed all other methods, giving a normalized mean squared error for estimated force of 1.6±3.69% (mean±SD). In visualization of the extracted features for the different CNN models, it is apparent that the final features of the CNN-FLF enable finding an accurate regressive relationship with output force levels.