多导联心电图识别的稳定步长ResNet深度网络

ResNet deep network with stable step for multi-lead electrocardiogram recognition

针对经典的ResNet深度神经网络对一维多导联心电图图像进行识别、分类时,因原始图像的维度较高导致提取到的深度特征维度高,造成全连接层训练出现收敛速度慢和过拟合的问题,在ResNet的全连接层提出一种稳定步长动量训练算法,通过引入近似二阶梯度信息增强动量法的寻优能力和加速收敛速度;利用连续2次迭代的参数变化量和梯度信息自适应调整步长,构造边界函数对步长的大小进行限制,以防止步长过大或过小而影响收敛稳定性,使用动量项对参数的更新方向进行修正。在CPSC2018心电图数据集上的实验结果表明:所提算法训练的ResNet取得的F 1分数、准确率、精确度分别达到0.859、97.4%、87.9%,收敛速度和整体分类指标值优于其他相比较的方法。

When the classical ResNet deep neural network is used to recognize and classify the one-dimensional multi-lead ECG image,the high dimension of the original image leads to the high dimension of the deep feature extracted,which leads to the problems of slow convergence and over-fitting in the full connection layer training.In order to deal with this problem,a stable step momentum training algorithm is proposed in the full connection layer of ResNet,which enhances the optimization ability and accelerates the convergence speed of the momentum method by introducing approximate second-order gradient information.Firstly,the step size is adaptively adjusted by using the parameter variation and gradient information of two consecutive iterations,and then the boundary function is constructed to limit the step size to prevent the step size from being too large or too small to affect the convergence stability.Finally,the momentum term is used to modify the updated direction of the parameters.The experimental results on the CPSC2018 ECG dataset show that the F 1 score,accuracy and accuracy of the ResNet trained by the proposed algorithm reach 0.859,97.4%and 87.9%,respectively,and the convergence speed and the overall classification index value are better than other comparative methods.