基于改进BP神经网络砾石土-滤层渗透破坏判别研究

Study of Improved BP(Error Back-Propagation)-based Neutral Network Gravel Soil—Filter Seepage Failure Discrimination

针对BP人工神经网络具有易陷入局部极小、学习过程中常常发生振荡等缺陷,提出了在BP算法中引入动量因子,并采用自适应调整学习率的梯度下降算法,建立了砾石土渗透破坏判别的改进BP神经网络模型。根据砾石土渗透破坏的实测资料,分别对BP神经网络判别结果和改进的BP神经网络判别结果进行比较,结果表明后者比前者判别能力更佳。

In consideration of that BP artificial neutral network is apt to cause local minimization instead of whole minimization and to lower learning rate, etc., the paper has established the improved BP neutral network model for discrimination of gravel seepage failure by introducing momenturm factors into the BP algorithm and adopting the gradient method of self-adaptive regulating learning rate. A comparison of the results obtained using the BP neutral network and the improved BP neutral network, according to the measurement of gravel soil seepage failure, indicates that the latter is of higher capability of discrimination than the former.