표면 근전도 기반 미세 움직임 해독을 위한 특수 자기지도 학습 프레임워크

📝 Abstract

Decoding fine-grained movement from non-invasive surface Electromyography (sEMG) is a challenge for prosthetic control due to signal non-stationarity and low signal-tonoise ratios. Generic self-supervised learning (SSL) frameworks often yield suboptimal results on sEMG as they attempt to reconstruct noisy raw signals and lack the inductive bias to model the cylindrical topology of electrode arrays. To overcome these limitations, we introduce SPEC-TRE, a domain-specific SSL framework. SPECTRE features two primary contributions: a physiologically-grounded pre-training task and a novel positional encoding. The pretraining involves masked prediction of discrete pseudo-labels from clustered Short-Time Fourier Transform (STFT) representations, compelling the model to learn robust, physiologically relevant frequency patterns. Additionally, our Cylindrical Rotary Position Embedding (CyRoPE) factorizes embeddings along linear temporal and annular spatial dimensions, explicitly modeling the forearm sensor topology to capture muscle synergies. Evaluations on multiple datasets, including challenging data from individuals with amputation, demonstrate that SPECTRE establishes a new state-of-the-art for movement decoding, significantly outperforming both supervised baselines and generic SSL approaches. Ablation studies validate the critical roles of both spectral pre-training and CyRoPE. SPECTRE provides a robust foundation for practical myoelectric interfaces capa- * Equal contributions

💡 Analysis

Decoding fine-grained movement from non-invasive surface Electromyography (sEMG) is a challenge for prosthetic control due to signal non-stationarity and low signal-tonoise ratios. Generic self-supervised learning (SSL) frameworks often yield suboptimal results on sEMG as they attempt to reconstruct noisy raw signals and lack the inductive bias to model the cylindrical topology of electrode arrays. To overcome these limitations, we introduce SPEC-TRE, a domain-specific SSL framework. SPECTRE features two primary contributions: a physiologically-grounded pre-training task and a novel positional encoding. The pretraining involves masked prediction of discrete pseudo-labels from clustered Short-Time Fourier Transform (STFT) representations, compelling the model to learn robust, physiologically relevant frequency patterns. Additionally, our Cylindrical Rotary Position Embedding (CyRoPE) factorizes embeddings along linear temporal and annular spatial dimensions, explicitly modeling the forearm sensor topology to capture muscle synergies. Evaluations on multiple datasets, including challenging data from individuals with amputation, demonstrate that SPECTRE establishes a new state-of-the-art for movement decoding, significantly outperforming both supervised baselines and generic SSL approaches. Ablation studies validate the critical roles of both spectral pre-training and CyRoPE. SPECTRE provides a robust foundation for practical myoelectric interfaces capa- * Equal contributions

📄 Content

SPECTRE: Spectral Pre-training Embeddings with Cylindrical Temporal Rotary Position Encoding for Fine-Grained sEMG-Based Movement Decoding Zihan Weng∗1,2, Chanlin Yi*1,2, Pouya Bashivan3, Jing Lu1,2, Fali Li1,2,4,5, Dezhong Yao1,2,4,6, Jingming Hou†7, Yangsong Zhang†8,9, and Peng Xu†1,2,4,10,11 1Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, University of Electronic Science and Technology of China, Chengdu, China 2School of Life Science and Technology, Center for Information in Medicine, University of Electronic Science and Technology of China, Chengdu, China 3Department of Physiology, McGill University, Montreal, Canada 4Research Unit of NeuroInformation, Chinese Academy of Medical Sciences, Chengdu, China 5Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Macau, China 6School of Electrical Engineering, Zhengzhou University, Zhengzhou, China 7Department of Rehabilitation, Southwest Hospital, Army Medical University, Chongqing, China 8Laboratory for Brain Science and Artificial Intelligence, School of Medicine, Southwest University of Science and Technology, Mianyang, China 9School of Computer Science and Technology, Southwest University of Science and Technology, Mianyang, China 10Radiation Oncology Key Laboratory of Sichuan Province, Chengdu, China 11Rehabilitation Center, Qilu Hospital of Shandong University, Jinan, China Abstract Decoding fine-grained movement from non-invasive surface Electromyography (sEMG) is a challenge for prosthetic control due to signal non-stationarity and low signal-to- noise ratios. Generic self-supervised learning (SSL) frame- works often yield suboptimal results on sEMG as they at- tempt to reconstruct noisy raw signals and lack the induc- tive bias to model the cylindrical topology of electrode ar- rays. To overcome these limitations, we introduce SPEC- TRE, a domain-specific SSL framework. SPECTRE fea- tures two primary contributions: a physiologically-grounded pre-training task and a novel positional encoding. The pre- training involves masked prediction of discrete pseudo-labels from clustered Short-Time Fourier Transform (STFT) rep- resentations, compelling the model to learn robust, phys- iologically relevant frequency patterns. Additionally, our Cylindrical Rotary Position Embedding (CyRoPE) factor- izes embeddings along linear temporal and annular spatial dimensions, explicitly modeling the forearm sensor topol- ogy to capture muscle synergies. Evaluations on multiple datasets, including challenging data from individuals with amputation, demonstrate that SPECTRE establishes a new state-of-the-art for movement decoding, significantly out- performing both supervised baselines and generic SSL ap- proaches. Ablation studies validate the critical roles of both spectral pre-training and CyRoPE. SPECTRE provides a robust foundation for practical myoelectric interfaces capa- ∗Equal contributions †Corresponding author: jingminghou@hotmail.com, zhangysacademy@gmail.com, xupeng@uestc.edu.cn ble of handling real-world sEMG complexities. 1 Introduction The human hand, with its remarkable dexterity, enables in- tricate interactions with the physical world, forming the cor- nerstone of countless daily activities and complex skills [29]. Capturing the neural intent behind these fine-grained fin- ger movements through non-invasive means is a critical pur- suit in fields like advanced prosthetics [21], neurorehabilita- tion [49], and intuitive human-computer interfaces (HCIs) [26]. Among various biosignals, surface Electromyography (sEMG), which measures the electrical activity produced by skeletal muscles, offers a promising window into motor in- tent, originating from neural commands and reflecting neu- romuscular activation patterns [13]. Its non-invasive nature and potential for real-time decoding make it particularly attractive. Despite its potential, decoding continuous, fine finger movements from sEMG presents substantial challenges. sEMG signals are notoriously non-stationary and suscep- tible to various noise sources, including muscle fatigue, elec- trode shift, skin impedance variations, motion artifacts, and environmental interference [12]. Furthermore, signifi- cant inter-subject and even intra-subject variability exists due to differences in anatomy, muscle activation strategies, and recording conditions. Traditional machine learning ap- proaches [6] often struggle with the complexity and variabil- ity of sEMG for continuous, multi-degree-of-freedom (DoF) decoding. Deep learning, particularly Convolutional Neural Net- 1 arXiv:2512.22481v1 [cs.HC] 27 Dec 2025 works (CNNs) and Recurrent Neural Networks (RNNs), has shown promise in improving sEMG-based gesture recogni- tion and movement decoding [16]. However, these super- vised methods typically require large amounts of accurately labeled data, where continuous finger kinematics are pre- cisely synchronized with multi-channel

This content is AI-processed based on ArXiv data.