图  1  根据官方指导说明, 将Myo臂环固定在小臂处; 初始位置: 带有Logo指示灯的通道与中指对齐

Fig.  1  The Myo around the lower arm according to official guides. The position where the logo channel and the middle finger are on one line as the initial position

图  2  实验中8种常用手部动作: 休息、握拳、展掌、剪刀手、捏中指、捏食指、内翻、外翻

Fig.  2  Eight common gestures in the experiment: rest, grasp, open hand, peace sign, pinch thumb finger, pinch middle finger, carpal varus, carpal valgus

图  3  基于滑动窗的特征提取, 其中窗长为LW, 滑动为LI

Fig.  3  Feature extraction by a sliding window technique with incremental length (LI) and window length (LW)

图  4  电极循环单位为1时, 由模拟法实验范式产生的电极偏移下表面肌电信号

Fig.  4  Electrode-shift sEMG by a simulation paradigm with a circulation step$\tau= 1 $

图  5  基于极坐标系的电极偏移估计与自适应校正的表面肌电手势识别方法与传统识别框架的对比.

所采集带标签的数据集$ X_L $, 分割成训练集$X_L^{\rm{train}}$ 与测试集$X_L^{\rm{test}}$; 电极偏移干扰新样本集$ X_N $, 按照原分类器得到的预测结果$ \hat{y}_N $; 或将新样本集$ X_N $训练独立的分类器, 得到训练集与测试集结果$y_N^{\rm{train}}$ 和$y_N^{\rm{test} }$ . 点划线表示路径, 首先估计电极偏移前后的偏移角度$ \delta\alpha $, 以此建立两者之间的关系$ T_{n\times n} $, 最终经原分类器得到预测结果$ y_{p\dot N} $

Fig.  5  The proposed electrode shifts estimation and adaptive correction method for sEMG-based gesture recognition and its traditional comparisons sEMG-based frameworks.

Here, the labeled datasets $ X_L $ were split into the training part ($X_L^{\rm{train}}$, $y_L^{\rm{train}}$) and the testing part ($X_L^{\rm{test}}$, $y_L^{\rm{test}}$). The dataset $ X_N $ was interfered by electrode shifts, whose classification results were $ \hat{y}_N $ by the pre-trained classifier. Or, $ X_N $ was used to train an individual classifier, resulting in $y_N^{\rm{train}}$ and $y_N^{\rm{test}}$. As the dot-dash line, the shift $ \Delta \alpha $ was first estimated, and then $ T_{n\times n} $ was established. Finally, $ y_{p\dot N} $ was the estimated results of the interfered by the pre-trained classifier

图  6  横截面图和散点图(a) 截面视角, 包括Myo臂环和小臂各肌肉与骨骼的截面分布; (b) 散点图: 不同通道上的样本点绘制在不同极角的极轴上, 点划线为所提出并定义的活跃极角

Fig.  6  The cross-section diagram and scatter diagram (a) the cross-section view of proposed polar coordinate with Myo armband and the lower arm's muscles and bones; (b) the scatter plot with every channel samples on different polar angle axis, and the proposed APA with dot-dash line

图  7  偏移范围$ \Delta \alpha $为$ [{0^\circ}, {45^\circ }) $时, 所有通道电极相对于初始位置偏移情况

Fig.  7  Every electrode shifts $ \Delta \alpha $ in a counter-clockwise direction within the range of $ [{0^\circ}, {45^\circ }) $

图  8  各电极通道MAV特征值映射变换与偏移角度间关系

Fig.  8  The mapping relationship between original one-channel MAV and shifted MAV on the shift angle

图  9  模拟电极偏移情况下, 极坐标内的肌电信号样本点的散点图与虚线所代表的活跃极角

Fig.  9  Scatter plots in the polar coordinate of sEMG samples from simulated electrode shifts and the proposed activation polar angle (APA) in dash line

图  10  不同偏移程度下的手势识别精度对比: 偏移估计与自适应校正的精度(实线)和未处理电极偏移干扰下的精度(虚线)

Fig.  10  Gestures recognition comparison with different shifts: proposed shifts estimation and adaptively corrected accuracies (solid line) v.s. un-corrected accuracies (dotted line)

图  11  真实的电极−肌肉偏移情况下, 极坐标系内肌电信号样本点的散点图与虚线所代表的活跃极角

Fig.  11  Scatter plots in the polar coordinate of sEMG samples from real electrode shifts and the proposed activation polar angle (APA) in the dash line

图  12  在9次电极−肌肉偏移过程中, 对活跃极角的偏移角度估计

Fig.  12  The estimations of activation polar angle for nine shifts in $ [0^{\circ},\, 360^{\circ}) $

图  13  不同真实电极−肌肉偏移程度下的手势识别精度对比: 偏移估计与自适应校正的精度(实线)和未处理电极偏移干扰下的精度(虚线)

Fig.  13  Gesture recognition accuracy comparison with different electrode-muscle shifts: proposed shifts estimation and adaptively corrected accuracies (solid line) v.s. un-corrected accuracies (dotted line)