基于AdaBoost集成网络的模拟电路单软故障诊断

AdaBoost based ensemble of neural networks in analog circuit fault diagnosis

基于神经网络的模拟电路故障诊断方法中会遇到网络训练误差很小甚至为零的情况下测试误差依然存在甚至很大的情况,这是由于神经网络的推广性能较差而造成的。针对模拟电路的神经网络故障分类器推广性能较差的问题,提出了将基于AdaBoost的神经网络集成应用于模拟电路故障诊断的新方法。通过对集成网络的诊断误差进行偏差-方差分析,说明可重复取样的AdaBoost技术可以降低集成网络中各成员网络的相关性以减少方差,从而使模拟电路故障诊断的集成网络系统的推广误差降低,使模拟电路故障诊断率得以提高。利用PSPICE的仿真数据和从实际电路获取的实测数据进行实验,实验结果均证明了该方法的有效性。

In neural network based analog circuit fault diagnosis methods, there exists big test error even when the training error reaches zero. It is owing to the bad generalization performance of neural network classifier. In order to improve the diagnosis system based on generalization performance of neural network in analog circuit fault isolation, an AdaBoost based ensemble of neural network （ENN） system is proposed. The bias-variance decomposition shows that AdaBoost resample technique can reduce the correlation of the component networks, cut down the variance and generalization error of the ensemble networks. Component neural networks in this ensemble are trained by their respective training sets created by AdaBoost resample technique. Data sets obtained from simulation and actual circuit were used to evaluate this ensemble system, and experimental results reveal the validity of this method and the improvement of fault diagnosis accuracy.