전완부 근전도 신호를 활용한 ResNet 기반 손 안 조작 및 엄지 촉진 분류

• Published in: Journal of biomedical engineering research Vol. 46; no. 1; pp. 21 - 28
• Main Authors: 조현승, 안해원, 박형순, Hyeonseung Cho, Haewon An, Hyung-Soon Park
• Format: Journal Article
• Language: Korean
• Published: 대한의용생체공학회 01.02.2025
• Subjects: 의공학; Palpation; Electromyography (EMG); Human-machine interface (HMI); In-hand manipulation; ResNet
• ISSN: 1229-0807; 2288-9396

Accurate recognition of user intentions in Human-Machine Interfaces (HMIs) for assistive robots, such as prosthetic hands, is crucial for user-friendly control. Surface electromyography (sEMG)-based HMIs offer great potential due to their non-invasive nature and ability to classify gestures with high accuracy using artificial intelligence-based pattern recognition. While considerable research has been conducted on classifying grasping postures, studies focusing on in-hand manipulation, the ability to rearrange objects within the hand, and thumb palpation, the process to identify object properties through thumb movements, remain limited. Prior studies have attempted to classify in-hand manipulation and thumb movements using sEMG. However, these studies either utilized both extrinsic forearm muscles and intrinsic hand muscles or employed high-density sEMG instead of commercial sEMG, making them unsuitable as generalizable methods for intention recognition in prosthetic control for amputees. This study proposes a ResNet-based model to classify 10 primitives for in-hand manipulation and thumb palpation using eight-channel sEMG from extrinsic forearm muscles. The model achieved an accuracy of over 87% for seven subjects and maintained an accuracy of over 75% through fine-tuning based user-specific calibration. These results demonstrate that the model can maintain high accuracy under electrode reattachment through fine-tuning. In conclusion, the proposed model utilized signals exclusively from forearm muscles through commercially available sEMG and demonstrated high classification accuracy exceeding 87.26% through user-specific calibration. This confirms its potential as a generalizable intention recognition system for in-hand manipulation.