基于高阶累积量和改进GRNN的CSI手臂行为识别

CSI arm activity recognition based on high order cumulants and improved GRNN

为了挖掘信道状态信息(channel state information, CSI)在手臂行为识别中的非线性深层特征,提高识别准确度,提出了一种基于高阶累积量和改进广义回归神经网络(generalized regression neural network, GRNN)的CSI手臂行为识别算法。离线阶段,将在不同手臂动作下采集的细粒度CSI幅度和相位差作为基信号,并利用平均绝对偏差改进的spearman rank相关系数选择敏感性强的子载波;针对CSI中的非线性非高斯信息,在所选子载波中提取高阶累积量特征;在灰狼算法(grey wolf optimizer, GWO)优化的GRNN神经网络中训练出能有效处理非线性问题的GWO-GRNN动作识别模型。在线阶段,利用训练好的识别模型对输入的CSI数据进行手臂动作的判别。通过仿真实验验证,该算法的手臂行为识别准确度为95.83%,高于目前相关算法所达到的准确度,具有明显的识别优势。

In order to deeply mine the nonlinear characteristics of channel state information(CSI) in arm behavior recognition to improve recognition accuracy, this paper proposes a CSI arm behavior recognition algorithm based on high-order cumulants and improved generalized regression neural network(GRNN). In the offline phase, firstly, the CSI amplitude and phase difference collected under different movements of the arm are used as the base signal, and the subcarriers with strong sensitivity are selected by the spearman rank correlation coefficient improved by the mean absolute deviation. The high-order cumulant features are extracted from selected subcarriers to obtain nonlinear non-Gaussian information in CSI. Finally, the action recognition model named GWO-GRNN that can effectively deal with nonlinear problems is trained in the GRNN optimized by the grey wolf optimizer(GWO). In the online phase, the input CSI data is used to distinguish arm movements through the trained recognition model. Through simulation experiments, the recognition accuracy of the algorithm is 95.83%, which is higher than the accuracy achieved by the current related algorithms, and has obvious recognition advantages.