In human-machine interfaces (HMI), deep learning (DL) techniques such as convolutional neural networks (CNN), long-short term memory networks (LSTM) and the hybrid CNN-LSTM framework have been exploited for hand kinematics estimation using surface electromyography (sEMG). However, these DL techniques only capture the relationship between sEMG and hand kinematics, but ignores the prior knowledge of the system. By contrast, Kalman filter (KF) can apply Kalman gain to combine the internal transition model and the observation model effectively. To this end, we propose a novel architecture named deep Kalman filter network (DKFN), in which we utilize CNN to extract high-level features from sEMG and employ a LSTM-based Kalman filter process (LSTM-KF) to conduct sequential regression. In particular, LSTM-KF adopts the computational graph of KF but estimates parameters of the transition/observation model and the Kalman gain from data using LSTM modules. With this process, the advantages of KF and LSTM can be exploited jointly. Experimental results demonstrate that the proposed DKFN can outperform CNN and CNN-LSTM in the sequential regression for wrist/fingers kinematics estimation.

在人机界面（HMI）中，深度学习（DL）技术（例如卷积神经网络（CNN），长期短期记忆网络（LSTM）和混合CNN-LSTM框架）已被用于使用表面肌电图进行手运动学估计（sEMG）。但是，这些DL技术仅捕获sEMG和手运动学之间的关系，而忽略了系统的先验知识。相比之下，卡尔曼滤波器（KF）可以应用卡尔曼增益将内部转换模型和观测模型有效地结合起来。为此，我们提出了一种称为深度卡尔曼滤波网络（DKFN）的新颖架构，其中我们利用CNN从sEMG中提取高级特征，并采用基于LSTM的卡尔曼滤波过程（LSTM-KF）进行顺序回归。特别是，LSTM-KF采用了KF的计算图，但使用LSTM模块从数据中估计了转换/观测模型的参数和卡尔曼增益。通过此过程，可以共同利用KF和LSTM的优势。实验结果表明，所提出的DKFN在顺序回归中可优于CNN和CNN-LSTM进行手腕/手指运动学估计。