层次分析的神经网络集成方法

Method for Neural Network Ensemble Based on Analytical Hierarchy Process

提出了基于层次分析法的神经网络集成方法;介绍了四种不同类型的神经网络,并用这四种不同类型的神经网络构造了五个独立并行处理的模式识别子系统,对每个子系统进行独立训练并分别进行模式识别,统计每一输入模式被各子系统识别成的模式及次数,按识别次数进行两两成对比较;计算各输入模式的识别结果与典型故障模式识别结果间的期望与方差的大小,进行两两成对比较并进行一致性检验。选取层次分析法中准则层对目标层的权重,综合各神经网络子系统的识别情况与输入模式和目标模式的差异大小计算输入模式对各典型故障模式的整体权值并以此作出决策。使用matlab软件进行仿真计算,得到的故障正确识别率。层次分析法神经网络集成方法取得了比相对多数集成方法好的识别效果,还可集成其他模式的识别方法。

This paper presents a method based on analytical hierarchy process （AHP） for neural network ensemble. In the method, five subsystems are constructed. Those neural networks compute parallel which have been trained solely. The decisions are made. Correct classification rates are obtained from the mat lab simulation. In experiments, the proposed methods have been successfully evaluated using thirteen different datasets, it is more effective than the relative majority voting scheme, it can be applicable to other method of pattern recognition.