본 논문은 다채널 근전도 센서를 이용한 손동작 인식에서 측정주파수, 윈도우 크기에 따른 분류 정확도와 센서 별 위치 중요도를 분석하였다. 근전도 측정은 연구실에서 자체 제각한 암밴드 모듈과 상용 근전도 센서를 이용하여 각각 측정하였으며, MAV,WL, SSC, WL값을 연산 후 ANN, SVM, DT, KNN에 적용하여 7가지 손동작을 분류하였다. 암밴드 모듈로 500Hz의 근전도를 측정 및 평가한 결과 정상인을 대상으로 약 96.62%의 분류 정확도를 절단자 대상으로 약 52.62%의 분류 정확도를 획득하였다. 정상인을 대상으로 5개의 특정근육에서 2,000Hz로 측정 한 결과, 최대 96.77%의 분류 정확도를 얻었으며, 절단자를 대상으로 3,000hz로 측정한 결과 최대 87.9%의 분류 정확도를 획득하였다. 일반적으로 절단자에게서 낮은 분류 정확도를 획득하였다. 측정 주파수와 윈도우 간의 연구에서는 주파수가 증가함에 따라, 윈도우 크기가 증가함에다라 분류 정확도가 증가함을 확인하였다. 높은 분류 정확도를 획득하기 위해서는 적어도 250Hz의 측정주파수와 200ms이상의 윈도우 크기 또는 500Hz의 측정주파소와 150ms의 윈도우 길이가 필요할 것으로 예상된다. 특정 근육 별 직접 부착하였을 때에, extensorcarpi ulnaris, flexor carpi ulnaris, flexor digitorium superficialis, flexor carpi radialis, brachioradilis 순으로 나타났으며 이는 암밴드의 중요도와 다르게나타났다. 암밴드와 센서 직접부착방식에서 근육 별 중요도 순서가 다른 이유는 암밴드의 경우 특정 근육 위치에 센서를 부착하기 어려운 단점이 존재하며, 특정 근육에 부착한 경우 암밴드에 비해 crosstalk이 감소한 것으로 보인다. 본 연구의 결과를 바탕으로 의수를 제작함에 있어서 특정 근육과 특정 주파수를 선정한다면 일상생활 의수제작에 도움이 될 것으로 예상된다.
In this study, an appropriate sensor arrangement and specifications for the design of a prosthetic socket were presented through hand gesture recognition using a multi-channel sEMG armband module and five commercial sEMG sensors. Twenty normal subjects and four upper limb amputees were recruited as subjects, half of the measured sEMG through the armband module, and the other half measured through five commercial sEMG sensors. Through the measured EMG signals, classification accuracies according to the sampling frequency and window length were evaluated, and the importance of each sensor location was researched. For this, feature values of MAV, WL, SSC, and ZC were calculated, and hand motion classification was performed through four classifiers: ANN, SVM, KNN, and DT. sEMG was measured using an armband module for seven hand movements of normal subjects, and the maximum classification accuracy of 96.6% was obtained, and the maximum classification accuracy of 72.4% was obtained for the amputee. On the other hand, when five commercial EMG sensors were attached to the specific muscles, the maximum classification accuracy of 97.2% was obtained for normal subjects, and he maximum classification accuracy of 87.9% was obtained for the amputee. In hand gesture recognition using a multi-channel EMG sensor, it was confirmed that the classification accuracy varies according to the sampling frequency and window length. The classification accuracy increased as the sampling frequency and window length increased, and there was a significant interaction between the two variables. Through this, it was confirmed that the number of samples in the window affects the classification accuracy. In order to obtain high classification accuracy of 90% or more, a sampling frequency of 250 Hz and a window length of 200ms or more, or a sampling frequency of 500 Hz and a window length of 150ms or more were required. The importance of each sEMG sensor was analyzed using the stepwise variable selection method in hand gestures of normal subjects. As a result, in order to obtain a high classification accuracy of 90% or more, four or more EMG sensors were required in the armband module. In order to recognize it with an accuracy of 90% or more by attaching the sEMG sensor directly to the skin surface of the muscle, it is proposed to attach at least three sensors to appropriate positions (Extensor carpi ulnaris, Flexor carpi ulnaris, Flexor digitorum superficialis).