Subject-independent Continuous Estimation of sEMG-based Joint Angles using both Multisource Domain Adaptation and BP Neural Network

• Published in: IEEE transactions on instrumentation and measurement Vol. 72; p. 1
• Main Authors: Li, He, Guo, Shuxiang, Wang, Hanze, Bu, Dongdong
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation; Adaptation models; Artificial neural networks; Back propagation; Back propagation networks; BP neural network; Continuous angle estimation; Domains; Elbow; Elbow (anatomy); Electromyography; Estimation; Exoskeletons; Inter-subject variability; Invariants; Limbs; Multisource domain adaptation; Muscles; Neural networks; Rehabilitation; Statistical methods; Surface electromyography (sEMG); Training
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2022.3225015

Continuous angle estimation from surface electromyography (sEMG) is crucial for robot-assisted upper limb rehabilitation. The sEMG-based control provides an optimal way to achieve harmonic interactions between subjects and upper limb rehabilitation exoskeletons. And for upper limb exoskeleton systems with sEMG as the control signal, accurate identification of elbow angles from sEMG is essential. However, sEMG signals have a subject-specific nature, causing the estimation model with sEMG signals as input to have poor generalization across multiple subjects. Aiming at the above problem of inter-subject variability on sEMG, multisource domain adaptation (MDA) is combined into the estimation of continuous joint movements to obtain subject-invariant features of sEMG. And feature distribution of the training set and test set is evaluated using the kernel density estimation method. Further, the subject-invariant features obtained through MDA are the input of the back propagation neural network (BPNN). Different evaluation indicators and the statistical method are used to compare the estimation results between original features and subject-invariant features, which proves the better generalization ability of the model based on subject-invariant features. And the estimation angle error calculated by using subject-invariant features as the input of BPNN is controlled within 10°, which shows the effectiveness of the combination of MDA and shallow neural network for the accurate subject-independent estimation of elbow joint continuous movements.