근육 피로도 분석을 위한 CNN-BiLSTM 기반 인공지능 근육 수축 감지 모델 = CNN-BiLSTM-Based Artificial Intelligence Muscle Contraction Detection Model for Muscle Fatigue Analysis

다국어 초록 (Multilingual Abstract)

Recently, changes in lifestyle and environmental factors have been continuously increasing the incidence of stroke, leading to various sequelae. Foot drop, a representative stroke sequela, causes foot dragging or catching during walking due to the functional decline of ankle dorsiflexors. This leads to decreased walking stability and increased fall risk, lowering the patient's ability to perform activities of daily living and their quality of life.
While clinical practice widely applies Functional Electrical Stimulation (FES) for foot drop rehabilitation, repetitive and long-term electrical stimulation accumulates muscle fatigue. This accumulation of muscle fatigue can result in not only decreased stimulation efficiency but also reduced rehabilitation effects and the risk of tissue damage.
Therefore, the rehabilitation process requires quantitative evaluation and monitoring technologies for muscle fatigue as key elements to improve the safety and efficiency of electrical stimulation-based rehabilitation therapy.
Muscle fatigue assessment is a critical component for the quantitative analysis of muscle state changes during dynamic activities. For this purpose, Surface Electromyography(sEMG) signals, which allow for non-invasive and real-time measurement, are widely utilized. However, sEMG signals measured during actual dynamic activities such as rehabilitation, exercise, and physical labor simultaneously contain both muscle contraction and relaxation intervals. The relaxation interval consists primarily of low-frequency noise with negligible electrical muscle activation. When the analysis includes these relaxation intervals, frequency-based fatigue indices become distorted, leading to an inaccurate reflection of muscle fatigue.
To address this issue, previous studies have employed techniques that detect muscle contraction intervals by setting thresholds based on maximum voluntary isometric contraction (MVIC). However, this approach is highly dependent on individual muscle strength and requires repeated MVIC signal measurements, presenting practical limitations.
In this paper, to supplement these limitations, we proposed a CNN-BiLSTM-based muscle contraction detection model that automatically distinguishes muscle contraction and relaxation intervals from surface electromyography signals without separate MVIC measurements. The proposed model enables stable contraction interval detection even in dynamic environments by extracting local features of sEMG signals via Convolutional Neural Networks (CNN) and learning temporal contextual information using Bidirectional Long Short-Term Memory (BiLSTM) networks.
We evaluated the performance of the proposed model on 750 test data using metrics such as Accuracy, Precision, Recall, and F1-score. As a result of the experiment, the proposed model showed excellent performance in all metrics with Accuracy 0.9406, Precision 0.9397, Recall 0.9396, and F1-score 0.9395. Furthermore, by employing CNN-based preliminary feature extraction, we reduced memory usage by approximately 44% compared to a single BiLSTM structure, while ensuring computational efficiency and training stability.
These results indicate that the proposed model can serve as an effective alternative to existing MVIC-based muscle contraction detection techniques, successfully addressing their limitations regarding individual variability and practicality.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• Ⅱ. 이 론 6
• 2.1 근전도 6
• 2.2 근육 피로도와 근전도 신호 특성 8
• 2.3 기존의 근육 수축 감지 모듈 11
• 2.4 CNN-BiLSTM 17
• Ⅲ. 제안한 딥러닝 기반의 근육 수축 감지 모델 22
• 3.1 제안한 방법의 개요 22
• 3.2 데이터 세트 26
• 3.3 데이터 전처리 29
• 3.4 학습 데이터 라벨링 33
• 3.5 제안한 근육 수축 감지 모델 37
• 3.6 모델의 성능 지표 41
• Ⅳ. 실험 및 결과 44
• 4.1 구현 환경 44
• 4.2 성능 검증 및 결과 46
• 4.3 제안한 모델의 근육 수축 감지 성능 52
• 4.4 기존의 근육 수축 감지 모듈과 제안한 모델 비교 57
• 4.5 고찰 60
• Ⅴ. 결 론 62
• 참 고 문 헌 64
• 영 문 초 록 70