事件检测概率神经网络模型的建立与验证

Development and simulation of PNN model for incident detection

在对概率神经网络（PNN）的分类机理、输入向量选取和网络设置进行分析的基础上，建立了用于识别两类事件模式（无事件模式和有事件模式）的事件检测PNN模型。采用高速公路路段I．880实地线圈数据集和事件数据集验证模型，通过比较PNN模型与多层前向神经网络（MLF）模型的结果，发现无论对于向北、向南或混合方向的高速公路事件检测，PNN模型的检测率（DR）比MLF模型高；平均检测时间（MTTD）比MLF模型短；但误报率（FAR）也较高。概率神经网络是高速公路事件检测的一种有效算法，其在理论基础、算法和学习速度等方面比多层前向神经网络具有优势。

This paper investigates the classification, input variables and settings of probabilistic neural network, including model development and simulation by I-880 database. Comparison of PNN and MLF simulation results show that DR and MTTD is achieved by PNN are better than MLF, FAR is inferior than MLF whether in north-toward, south-toward and two direction freeway incident detection. PNN is an effective algorithm in incident detection and is superior to MLF in theory, training numbers and learning speed.