摘要
针对雷达目标信号的复杂性和实用雷达目标识别系统所应具备的稳健性、扩展性及通用性等要求,提出多种简单有效的BP神经网络算法改进。通过平衡训练样本数量、动态重置初始权值、评定网络规模等措施,解决了BP算法收敛速度慢、受初始样本分布影响大等缺陷,提高了识别算法的稳健性和泛化能力。结果已成功应用到不同型号雷达上的多套目标识别系统中。
According to the complexity of radar target signals and the robustness, extensibility and generality requirements of practical radar target recognition systems, several effective improvements are made for the BP neural network algorithm. By balancing the training samples, dynamically resetting initial weights and adaptively evaluating net scales, some shortcomings of the BP algorithm, such as low convergence speed and high dependency on initial sample distribution are overcome, and the algorithm has become more robust and generic. The method has been applied in several recognition systems of some models of radar. Test and application results show that proposed algorithm is effective and practical.
出处
《系统工程与电子技术》
EI
CSCD
北大核心
2005年第4期582-585,共4页
Systems Engineering and Electronics
基金
"十五"国防预研基金资助课题
