基于迁移学习的快速多边响应法

Fast Multilateral Response Method Based on Transfer Learning

针对基于多边响应的快速时域仿真计算中,需要通过瞬态仿真得到所有边沿响应的时间花费问题,提出了基于迁移学习的多响应预测方法。首先,根据非线性链路的响应特点,将拖尾响应划为线性区和非线性区分别处理,以减少所需测量的响应长度;其次,以“0电平”响应作为源域数据构建预训练模型网络;然后,利用预训练模型的网络参数训练其他三种响应的迁移学习模型;最后,将所有非线性区响应绘制在一个单位间隔内,并将线性区的积累电压扩充至眼图轮廓。实验结果表明,该方法能以更少的响应数据,更全面且准确预测其余响应,从而提高快速多边响应法的计算速率。

In order to solve the time cost problem of obtaining all edge responses through transient simulation in fast time domain simulation based on multilateral response method,a multi response prediction method based on transfer learning was proposed.Firstly,based on the response characteristics of nonlinear links,the trailing response was divided into linear and nonlinear regions to be processed separately to reduce the required measurement response length;Secondly,a pre trained model network was constructed using the“0 level”response as the source domain data;Then,the network parameters of the pre-training model were used to train the other three response transfer learning models;Finally,all nonlinear region responses were plotted within a unit interval,and the accumulated voltage in the linear region was extended to the eye contour.The results indicated that this method predicted the remaining responses more comprehensively and accurately with fewer response data,thereby improving the calculation speed of the fast multilateral response method..