Reliable sEMG Motor Intent Perception for Upper Limb Rehabilitation Robots

For upper limb rehabilitation robots, surface electromyography (sEMG) is an essential signal source to assess motor-intent perception by the patients; however, the signals collected from patients are noisy and are different from person to person. Stable intent recognition is thus important with the use of reliable signal processing. This review poses the question of what denoising and information enhancement can do for sEMG-based motor-intent perception. It first outlines the origins of the instability of intent signals and the measures for denoising assessment, and then presents the performance and cost of three categories of methods, namely the classical denoising methods, the learning-based robustness enhancement methods and the multi-sensor information fusion methods. The three families are used to handle different stages of the problem and are not mutually exclusive—task-level metrics should be used to judge the methods, and they can be added in stages as needed. The comparison shows the amount of signal processing needed for a given task and the underlying evidence that supports translation from fixed filtering to a learned model or to additional sensing, for designers of upper-limb rehabilitation systems. The key questions that remain to be answered are cross-session generalization, cross-subject generalization, quantitative attribution of denoising gains to recognition performance, lightweight online deployment and long-term stability and individual weighting of multi-sensor fusion.