A Multimodal Fusion Model for Estimating Human Hand Force: Comparing surface electromyography and ultrasound signals

Abstract

Biomimetic robots have received significant attention in recent years. Among them, the wearable exoskeleton, which imitates the functions of the musculoskeletal system to assist humans, is a typical biomimetic robot. Given that safe human–robot interaction plays a critical role in the successful application of wearable exoskeletons, this work studies the clinical readiness of a multimodal fusion model that estimates hand force based on the surface electromyography (sEMG) and A-mode ultrasound signals of the forearm muscles. The proposed multimodal fusion model affords the biomimetic hand exoskeleton assisting the elderly in completing daily tasks or quantitatively assessing the recovery level of poststroke patients. The suggested fusion model is called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Optimization of Latent Representation for the Self-Attention Convolutional Neural Network</i> ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">OLR-SACNN</i> ), which utilizes a common component extraction module (CCEM) and a complementary component retention module (CCRM) to optimize latent representation of the multiple modalities. Then the optimized latent representations are fused with the self-attention mechanism. The experiments conducted on a self-collected multimodal data set verify performance of the proposed OLR-SACNN model. Specifically, compared to solely employing sEMG or A-mode ultrasound signals, the force estimation’s normalized mean-square error (NMSE) based on the multiple modalities decreases by 97.7 and 38.92%, respectively. Furthermore, the OLR-SACNN model has been used to estimate the hand force of some poststroke patients and attained the desired performance.