在火灾识别过程中,模式类的统计特性通常是未知的或者无法估计的,这类决策问题最好直接通过训练,生成所需的判别函数来处理.文中构建了高斯动态神经网络G-DNN结构,给出了G-DNN的理论分析和具体的算法步骤.对于网络学习的收敛速度非常重要网络参数σ,μ,ω,τ初值设定问题,充分利用参数与输入特征值、输出值和中间值分别相关的思想,引入参数预定义的方法,实验证明这种方法可以使网络学习较快收敛.%During fire recognition processing, pattern's statistical characteristic is often unknown or inestimable. The best method to solve this problem is direct training by producing criterion functions. In this article, Gauss Dynamic Nerve Network (DNN) is imported into fire recognition. At first, Gauss fuzzy dynamic nerve network (G-DNN) structure is set up and its theory and specific analyse steps are also present. In network, the parameters' initial value such as σ,μ,ω,τ is very important for the study convergence speed. The article make the best of idea that the parameters are separately correlative with the input eigenvalues, output values and middle values, it applies parameter predefined method and this method is confirmed to make the network have a better convergence speed capability.
展开▼
