论文部分内容阅读
本文提出用无先导卡尔曼滤波(UKF)法以取代一阶梯度法和二阶梯度法,用于训练通用学习网络(ULN),提高学习精度和收敛性。通用学习网络为多分支递归网络,结构复杂,其特点是多分支,互联及时间延迟。相比其他方法,UKF具有计算精度高,计算简便等优点。本文将UKF用于网络训练,进行CSTR过程建模和时间序列预测。结果表明UKF用于训练递归神经网络获得很好的效果,验证了所提出方法的可行性和有效性。
In this paper, we propose to use the UKF method instead of the first-order gradient method and the second-order gradient method to train the universal learning network (ULN) and improve learning accuracy and convergence. The general learning network is a multi-branch recursive network, which has a complicated structure and is characterized by multiple branches, interconnections and time delays. Compared with other methods, UKF has the advantages of high accuracy and easy calculation. In this paper, UKF is used for network training and CSTR process modeling and time series prediction are carried out. The results show that the UKF is very effective for training recurrent neural networks and validates the feasibility and effectiveness of the proposed method.