基于神经网络免疫集成的非线性时间序列故障预报

针对神经网络集成对个体差异性的要求 ,提出了集成网络间的结构差异度的概念 .在此基础上设计了一种基于反向选择的免疫算法 ,该算法可以在减小集成网络各自训练误差的同时保持网络间的结构差异度 ,从而提高神经网络集成的泛化能力 .同时证明了该算法对最优个体的收敛性 .将该方法应用于受噪声污染的非线性时间序列故障预报 ,根据预测误差可以方便准确地检测系统的缓变故障和突变故障 ,实现对微小故障的快速故障预报 ,降低误检率 .仿真结果证明了该方法的有效性 .

Neural network immune ensemble based fault prediction for nonlinear time series

A concept named structure difference among neural networks in ensemble is propos ed, which is directed towards the request of individual discrepancy. Based on th is concept, a negative selection based immune algorithm is designed and the convergence of the proposed algorithm is proved. The algorithm can keep the structure difference d egree among networks while the training error of the network is reduced. Thus t he generalization performance of neural network ensemble is improved. By this me thod, the abrupt fault and incipient fault of system can be detected quickly and conveniently. And the tiny unknown fault in the nonlinear time series with nois e can be predicted conveniently and accurately based on the prediction error. Th e error rate of prediction is also reduced. Simulation results show the effectiv eness and feasibility of the method.