一种基于鲁棒自联想神经网络的传感器故障诊断方法_李欢欢.pdf

第32 卷 第14 期 中 国 电 机 工 程 学 报 Vol.32 No.14 May 15, 2012 116 2012 年5 月15 日 Proceedings of the CSEE ©2012 Chin.Soc.for Elec.Eng. 文章编号：0258-8013 (2012) 14-0116-06 中图分类号：TP 206.3 文献标志码：A 学科分类号：470·20 一种基于鲁棒自联想神经网络的 传感器故障诊断方法 李欢欢，司风琪，徐治皋 ( 东南大学能源与环境学院，江苏省 南京市 210096) A Sensor Fault Diagnosis Method Based on Robust Auto-associative Neural Network LI Huanhuan, SI Fengqi, XU Zhigao (School of Energy and Environment, Southeast University, Nanjing 210096, Jiangsu Province, China) ABSTRACT: A new robust auto-associative neural network 0 引言 was presented for the modeling of nonlinear system. Distinguished from conventional auto-associative neural 随着火电机组自动化程度的不断提高，机组运 network, the proposed network was separately trained by two 行越来越依赖于各种控制系统，而过程参数的测量 sub networks, the mapping network and the de-mapping 准确性则是这些系统可靠运行的重要保证，对整个 network, such that higher convergence rate was achieved. The objective function with weights restriction term was used to 机组运行的安全性和经济性也有重要影响，因此一 train the network for improving the accuracy and the 方面需要努力提高测量系统的准确性和可靠性，另 robustness of the model by restricting the abnormal adjustment 一方面需要发展传感器故障诊断与数据检验方法。 of the parameters. For the training of mapping network, some 针对机组热力系统的非线性特点，研究者们先后提 perturbation noisy data sets were added to expand the training samples based on training results of the de-mapping network, 出了基于机理模型的解析冗余方法[1-2]和基于过程 such that the robustness of whole network was distinctly 数据的数据驱动(data-driven)方法[3-6] 。同时，专家 improved. Based on the proposed neural network model, a new