sMSDCTrans: Multiscale Dilated Convolution and Transformer for Continuous Estimation of Cross-Subject Finger Motion from sEMG

• Published in: International Conference on Control, Automation and Robotics : proceedings pp. 470 - 476
• Main Authors: Guo, Jinghuan, Yang, Xiaotian
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.04.2025
• Subjects: Bidirectional control; continuous estimation; Convolution; cross-subjects; Encoding; Estimation; Feature extraction; hands kinematics; Long short term memory; Measurement; Root mean square; sEMG; Smoothing methods; Transformers
• ISSN: 2251-2454
• DOI: 10.1109/ICCAR64901.2025.11073056

Surface electromyography (sEMG) signals are widely recognized as highly suitable physiological feedback mechanisms for human-computer interaction (HCI) and have been extensively studied and applied in this domain. The individual variability of sEMG signals often necessitates methods tailored to specific individuals, resulting in highly targeted but limited applicability. Therefore, achieving cross-subject continuous estimation of finger motion from sEMG signals remains a significant challenge. This paper introduces sMSDCTrans, a novel approach that integrates multiscale dilated convolution and Transformer to facilitate cross-subject applications. First, to overcome the limitations of traditional convolutions in capturing local features, multiscale dilated convolutions are employed to extract local patterns, while Transformer architecture focuses on capturing global dependencies. Next, a feature fusion module is developed to integrate local and global features, optimizing their combined contribution to continuous estimation. Finally, a smoothing layer is incorporated before the joint angle output to mitigate fluctuations. 15 subjects from the Ninapro DB2 dataset was utilized to evaluate performance under both single-subject and cross-subject conditions. The evaluation employed Pearson Correlation Coefficient(CC), Normalized Root Mean Square Error(NRMSE), and Coefficient of determination (R2) metrics. Compared to LSTM, TCN, BERT, and CNN-Attention, sMSDCTrans achieved CC, NRMSE, and R2 values of 0.88, 0.09, and 0.76 for specific individuals, and 0.84,0.11 , and 0.68 for cross-subject scenarios. Relative to the best-performing models, sMSDCTrans demonstrated improvements in CC, NRMSE, and R2 by 0.06,0.02 , and 0.10 for specific individuals, and 0.08,0.03 , and 0.21 for cross-subject scenarios.