中圖分類(lèi)號(hào)： TP391.41；TP181 文獻(xiàn)標(biāo)識(shí)碼： A DOI：10.16157/j.issn.0258-7998.212067 中文引用格式： 孔康，李德盈，孫中圣. 基于小波包組合特征和LMS-LSTM的表面肌電信號(hào)分類(lèi)[J].電子技術(shù)應(yīng)用，2022，48(10)：92-96. 英文引用格式： Kong Kang，Li Deying，Sun Zhongsheng. Classification of surface EMG signals based on wavelet packet combination and LMS-LSTM[J]. Application of Electronic Technique，2022，48(10)：92-96.
Classification of surface EMG signals based on wavelet packet combination and LMS-LSTM
Kong Kang，Li Deying，Sun Zhongsheng
School of Mechanical Engineering，Nanjing University of Science and Technology，Nanjing 210094，China
Abstract： To classify gestures using surface EMG signals, an innovative method is proposed to combine the time-domain and time-frequency features as characteristic parameters, namely the combined features of wavelet packet coefficients and variances. The classifier combined with least mean square and long and short time memory network(LMS-LSTM) is innovatively adopted. After the circuit filter is set for the first filtering, LMS is added to carry out the second filtering of the different features. The classification and recognition rate of the five gestures is 93.78%. Principal component analysis(PCA) is used to reduce the dimension, and the average recognition rate is 92.68%, and the optimization result is achieved. Experimental results show that LSTM classification results are higher than traditional linear discriminant and decision tree algorithms.
Key words : EMG signals；least mean square；long and short time memory network；wavelet packet combination characteristics
0 引言
表面肌電信號(hào)作為生物電信號(hào)的一種，由于能夠反映較多的生物運(yùn)動(dòng)特征，被廣泛應(yīng)用于康復(fù)訓(xùn)練裝置的設(shè)計(jì)和假肢控制等領(lǐng)域。肌電信號(hào)具有微弱性和突變性，在數(shù)據(jù)采集的過(guò)程中易受其他信號(hào)的干擾，給分類(lèi)的結(jié)果帶來(lái)較大影響。目前常見(jiàn)的分類(lèi)特征主要有時(shí)域、頻域和時(shí)頻域。于亞萍等人[1-2]利用多種母小波變換對(duì)表面肌電信號(hào)進(jìn)行識(shí)別；胡曉[3]等人利用小波包系數(shù)熵作為特征向量，但時(shí)頻信號(hào)存在高延遲性。本文在小波包系數(shù)特征的基礎(chǔ)上，添加了延遲性較低的方差特征，創(chuàng)新地采用將時(shí)域和時(shí)頻域組合的方式作為特征參數(shù)。常見(jiàn)的分類(lèi)器有支持向量機(jī)、隨機(jī)森林和線性判別等，但這些方法的識(shí)別率會(huì)隨輸入向量維度的增加而下降。而近年來(lái)國(guó)內(nèi)外對(duì)深度學(xué)習(xí)的研究越來(lái)越深入，此方法也被廣泛地應(yīng)用于信號(hào)處理領(lǐng)域。另外，自適應(yīng)濾波(Least Mean Square，LMS)作為一種檢測(cè)平穩(wěn)與非平穩(wěn)信號(hào)的濾波方式，被用于信號(hào)的去噪處理。本文參考長(zhǎng)短時(shí)記憶網(wǎng)絡(luò)(Long and Short Time Memory Network，LSTM)用于不同分類(lèi)的文獻(xiàn)[4-9]和陳景良等人[10]使用LMS對(duì)語(yǔ)音進(jìn)行降噪、石欣等人[11-12]利用LMS-隨機(jī)森林模型對(duì)下肢動(dòng)作進(jìn)行分類(lèi)后，綜合具有較高實(shí)時(shí)性的LMS和較高識(shí)別率的LSTM兩種算法的優(yōu)勢(shì)，采用兩種算法組合，與陳思佳等人[13]采用LSTM和卷積神經(jīng)網(wǎng)絡(luò)得到較高手勢(shì)動(dòng)作識(shí)別率相比提高了實(shí)時(shí)性。
