摘要
在典型舰船RCS的FEKO仿真数据基础上,提取均值、标准差、变异系数、平滑系数4种RCS数字特征作为BP神经网络的分类特征向量,进行目标分类仿真实验。实验结果显示,由于BP神经网络采用梯度下降法,其初始权值、阀值的随机设置会导致BP神经网络易陷入局部极小,为此,本文研究采用遗传算法对BP网络节点权值和阀值进行优化选择,仿真实验结果显示采用遗传算法优化后的BP神经网络分类识别性能稳定,不易陷入局部极小。
On the basis of FEKO simulation data of typical ship RCS,the classification simulation experiment of BP neural network was carried out by means of 4 kinds of RCS feature vectors,the mean,standard deviation,coefficient of variation,and smoothing coefficient. The simulation results show that random sets of the initial weights and threshold will led to BP neural network into a local minimum,because of the gradient descent method. Therefore,this paper uses genetic algorithm to optimize BP network weights and threshold. The results of simulation experiments show that the BP neural network classifier optimized by genetic algorithm is stable,and is not easy to fall into local minimum.
出处
《舰船科学技术》
北大核心
2016年第2期125-130,共6页
Ship Science and Technology
基金
国家自然科学基金资助项目(61401493)
国家部委基金资助项目(9140A01060113JB11012)
