基於深度學習之跨受試者肌電波模型用於氣 動式外骨骼輔助力道決策;Cross-Subject sEMG-Based Deep Learning Model for Assistive Force Determination in Pneumatic Exoskeletons

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/99412

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/99412

Title:	基於深度學習之跨受試者肌電波模型用於氣動式外骨骼輔助力道決策;Cross-Subject sEMG-Based Deep Learning Model for Assistive Force Determination in Pneumatic Exoskeletons
Authors:	吳松育;Wu, Song-Yu
Contributors:	電機工程學系
Keywords:	模糊神經網路;肌電訊號;深度學習;氣動肌肉外骨骼;Fuzzy Neural Network;Electromyography;Deep Learning;Pneumatic Artificial Muscle
Date:	2025-12-08
Issue Date:	2026-03-06 18:56:21 (UTC+8)
Publisher:	國立中央大學
Abstract:	外骨骼機器人逐漸成為復健輔助的重要技術，能提供高重複性與規律性的訓練，以促進運動功能恢復。其中，表面肌電訊號（EMG）可即時反映肌力狀態並提升使用者的參與度，但其個體間差異降低了模型泛化能力，易導致預測不準確。另一方面，氣動人工肌肉（PAM）具高功率重量比與良好順應性，適合安全柔順的人機互動，惟其非線性與遲滯效應亦增加控制挑戰針對上述問題，本研究提出一套以 EMG 驅動的上肢外骨骼系統，並採用多層次設計架構：首先，於力量預測使用長短期記憶網路（LSTM），結合 EMG 的時序與頻譜特徵，並加入跨受試者正規化，以提升泛化能力並克服個體差異；其次，在控制上採用雙路模糊神經網路（FNN），補償 PAM 在充放氣過程的非線性與遲滯，並具備自適應調整能力，以確保動態操作下的穩定與精準；最後，提出動態輔助比調整策略，根據即時預測肌力與期望值之差異自動調整輔助比，當肌力不足時即時補償，當肌力充足時降低輔助，以促進使用者的自主參與，達到靈活且個別化的輔助效果。實驗結果顯示，有輔助條件下 EMG 呈現「時域下降、頻域上升」趨勢，反映外骨骼有效分擔出力並延緩疲勞；無輔助時則出現「時域上升、頻域下降」變化，顯示肌肉需承擔更高收縮強度並累積疲勞。由此可見，本研究系統可即時補償力量不足，展現延緩疲勞、提升耐力與強化訓練持續性的效益，驗證其應用於個別化復健輔助之可行性。;Exoskeleton robots have gradually become an important technology for rehabilitation assistance, providing highly repetitive and consistent training to promote motor function recovery. Surface electromyography (EMG) can reflect muscle strength status in real time and enhance user engagement; however, inter individual variability reduces the model’s generalization ability, leading to inaccurate predictions. On the other hand, pneumatic artificial muscles (PAMs) possess a high power-to-weight ratio and excellent compliance, making them suitable for safe and natural human–robot interaction, though their nonlinear and hysteretic characteristics pose control challenges. To address these issues, this study proposes an EMG-driven upper-limb exoskeleton system featuring a multi-layered architecture. First, muscle force prediction is achieved using a long short-term memory (LSTM) network that integrates temporal and spectral EMG features and applies cross-subject normalization to enhance generalization and overcome individual differences. Second, a dual-channel fuzzy neural network (FNN) controller compensates for PAM nonlinearity and hysteresis during inflation and deflation, while its adaptive capability ensures stability and precision under dynamic conditions. Finally, a dynamic assist-ratio adjustment strategy is introduced, which automatically regulates the assist level based on the difference between the predicted and desired muscle forces—providing additional support when strength is insufficient and reducing assistance when strength is adequate—to encourage active participation and achieve flexible, individualized assistance. Experimental results show that under assisted conditions, EMG exhibits a “decrease in time-domain features and increase in frequency-domain features,” indicating that the exoskeleton effectively shares the load and delays fatigue; in contrast, unassisted conditions show “increased time-domain and decreased frequency-domain features,” suggesting greater muscle effort and accumulated fatigue. These findings demonstrate that the proposed system can compensate for insufficient muscle strength in real time, effectively delay fatigue, enhance endurance, and improve training continuity, thereby confirming its feasibility for personalized rehabilitation assistance.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	105	View/Open

社群 sharing

Loading...