基于IMU与sEMG混合信号的实时手势分类算法研究

为了提高表面肌电信号(surface electromyography,sEMG)的手势分类准确率，通过惯性测量单元(inertial measurement unit,IMU)与采集姿态信号与sEMG的混合信号，提出了GRUBiLSTM双层网络的实时手势分类算法。第1层门控循环单元(gated recurrent unit,GRU)利用能量组合算子特征对混合信号进行突变点检测，定位运动态数据起始点；第2层双向长短时记忆循环神经网络(Bi-directional long short term memory,BiLSTM)使用能量核相图特征对运动态混合信号进行2个方向10种手势的分类。通过离线模型优化，分类算法识别时间低于40 ms，突变点检测精度88.7%以上，手势分类准确率为85%，信息传输率(informationtranslaterate, ITR)达到89.9 bits/min，与基于机器学习的分类算法相比，在准确率与计算效率上具有优势。

Research on Real-time Gesture Classification Algorithm Based on IMU and sEMG Mixed Signals

In order to improve the gesture classification accuracy of surface electromyography (sEMG), the mixed signal of attitude and sEMG is collected by inertial measurement unit (IMU) and EMG sensor, and a GRU-BiLSTM double-layer network real-time gesture classification algorithm is proposed. The first layer of gated recurrent unit (GRU) detects the mutation point of the initial mixed signal though energy combination operator feature and locates the starting point of the dynamic data. The second layer Bi-directional long short term memory (BiLSTM) classifies the motion state mixed signal into 10 gestures in two directions though energy kernel phase map feature. Through offline model optimization, the recognition time of the online GRU-BiLSTM double-layer classifier algorithm is less than 40 ms, the detection accuracy of mutation points is over 88.7%, the accuracy of gesture classification is 85%, and the information transmission rate (ITR) reaches 89.9 bits/min. Compared with machine learning-based classification algorithms, the algorithm has advantages in accuracy and computational efficiency.