【摘要】 针对目前采用肌电信号的手指关节角度连续解码误差较大，导致肌电假肢手运动效果较差的情况，提出了一种应用表面肌电信号、深度回归森林模型和人工神经网络相结合的手指关节角度连续精确感知方法。首先，应用基于滑动时间窗的特征提取器从前臂8个通道的肌电信号中各提取7种肌电信号特征(肌电信号的平均绝对值、积分肌电值、均方根、波形长度、对数特征、过零点数、斜率符号变化数),输入深度森林回归模型得到具有较大波动的掌指关节估计角度；然后，采用人工神经网络对这些掌指关节估计角度进行优化，以创建一种深度森林回归模型与人工神经网络相结合的综合回归模型；最后，利用该综合回归模型对采集到的表面肌电信号进行连续精确解码，得到肌电假手掌指关节角度控制量，其余手指关节角度可通过比例控制得到。采用所提方法进行实验验证，结果表明：所提方法的平均轨迹跟踪精度比传统高斯过程方法提高了42%,达到82.12%,证明所提方法对基于肌电信号的手指关节角度估计具有非常优良的效果。更多还原

【Abstract】 To tackle the problem of poor movement of EMG-controlled prosthetic hand resulting from large continuous decoding errors of finger joint angles, this paper proposes a continuous and accurate sensing method that combines the deep regression forest model and artificial neural network based on EMG signals. Firstly, the feature extractor based on sliding time window is used to extract 7 kinds of EMG features(mean absolute value, integral EMG value, root mean square, waveform length, logarithmic feature, zero crossing points and slope symbol change) from the EMG signals of 8 channels of the forearm. These features are then input into the deep forest regression model to obtain the estimated angles of metacarpophalangeal joints with large fluctuation; then, the artificial neural network is used to optimize these estimated angles of metacarpophalangeal joints, so as to create a comprehensive regression model combining the deep forest regression model and artificial neural network; finally, the comprehensive regression model is used to continuously and accurately decode the collected surface EMG signals to obtain the control quantities of finger joint angles of the prosthetic hand, and other finger joint angles were obtained through proportional control. The experimental results show that the average trajectory tracking accuracy of the proposed method is 42% higher than that of the traditional Gaussian process method, reaching 82.12%. This proves that the proposed method can achieve excellent effects on finger joint angle estimation based on EMG signals.更多还原