本研究架構於先前研究所設計之七個主動電極的一腕套系統,研究之目的是找尋一較佳的表面肌電訊號前置處理流程,對上手臂的表面肌電圖(EMG)作一定性之分類,並大幅提升手部細微動作之辨識能力。研究中利用Daubechies 2階穩態小波轉換,將所擷取出來的表面肌電訊號做多解析分析,並擷取肌電封包能量與肌電封包,來代表各小波轉換層之訊號振幅與頻率之特徵,作為取代傳統的演算。在資料縮減方面,本研究採用主成分分析法(PCA)來挑選出整個特徵化樣本之主成分,以縮減樣本大小;並與使用獨立成分分析法(ICA)來挑選肌電通道之獨立成分,比較資料降低量與類神經網路辨識結果。 實驗結果發現,新的演算流程在整體辨識率與網路訓練週期上皆明顯的優於傳統演算法,其成對t檢定結果分別為0.001與0.003(p < 0.05)。在比較三種不同小波轉換層數選擇時,發現利用第2、3、4階的小波轉換階數能以較少的運算,達到所需的辨識效果與網路效能,並與4層及5層轉換階數無明顯統計上之差異。在資料縮減方面,本研究所使用的非監督式Fast ICA演算流程,將7個頻道之肌電圖縮減為4個頻道,然而其辨識結果並不理想。而在使用PCA縮減樣本時,則能有效的縮減網路架構為原來的30%,同時辨識率與網路訓練週期與未縮減前比較在統計方面亦無差異,其t檢定值各為0.78與0.85(p > 0.05)。所以本研究提出以新的演算流程,並配合PCA以縮減樣本大小,作為表面肌電訊號的前置處理流程,能有效降低網路架構並提升手部動作辨識效果。
This study bases on the result of previous research that designed a wrist-band electromyographic(EMG)acquisition system with seven active electrodes arranged in a ring faction . The purpose of this study is to establish a superior pre-processing procedure for surface EMG to achieve qualitative identification. Thus, we can increase the discrimination rate for different ranges of hand motions. The 2 order Daubechies stationary wavelet transform was used in order to extract the characteristics from surface EMG to achieve multi-resolution analysis. Additionally, the EMG envelope energy and EMG envelope were extracted to represent the amplitude and frequency characteristic of the EMG signals in every wavelet layer. We use this method to substitute for the traditional method. For data reduction, this study adopts the principal component analysis(PCA)method to choose the principal components from all characteristic sample to reduce the size of sample. Furthermore, we use independent component analysis(ICA)method to choose the channels of independent component. Lastly, we compare the size of data reduction and the result of discrimination using neural network. The results demonstrated that the discrimination rate and neural network training epochs using new algorithm method is superior to the traditional method. The results of paired t-test are 0.001 and 0.003 respectively (p value < 0.05). In addition, comparison was made between different wavelet layers. While using wavelet transform layer of the second, the third, and the forth level, it is found that better discriminative rate and less neural network training epochs are achieved. Additionally, there is no significance difference between using four levels and five levels of wavelet transform. At the same time, in the part of data reduction, by using the unsupervised fast ICA algorithm, the numbers of channel were reduced from seven to four channels. However, the result of discrimination rate is significantly worst than traditional method. On the other hand, we can efficiently reduce 70% of the neural network size when we use the PCA method. Moreover, the results of discrimination rate and neural network training epochs of PCA have no significant difference as compared to the results before PCA reduction. The values of t-test are 0.78 and 0.85 respectively (p value < 0.05). In conclusion, this study demonstrated a new procedure and using PCA method to reduce the sample size. The result of this study demonstrats that using wavelet and PCA are effective pre-processing for surface EMG analysis. It can efficiently reduce the size of neural network size and increase the discrimination rate for different hand motions.