大气杂质气体神经网络识别的推广性研究

Neural Network Generalization Analysis of Recognition for Gas Impurity in Air

采用支持向量机(SVM)方法对小波变换压缩后的非线性荧光光谱数据进行识别,对学习样本和未学习过的样本进行测试,其正确识别率均为100%。为了全面比较推广性能的好坏,建立了一个模拟实际监测数据的模型,并采用这些数据对提出的SVM网络与概率神经网络(PNN)进行了比较研究。仿真结果表明,无论对实验室数据的推广能力,还是对监测条件变化的推广能力,SVM网络较PNN有更好的推广和容错性能。

In order to investigate the property of neural network generalization more roundly, a mode for the testing data is proposed and the generalization capability of the support vector machine (SVM) network of this work and the probabilistic neural network (PNN) of a previous work are compared with the data produced by the mode. The simulation results show that the SVM network provides better generalization capability than that of the PNN in terms of generalization for either laboratory data or changes in experimental conditions.