神经网络在变压器油中溶解气体微机在线监测中的应用研究

【作者】 林俊；

【导师】 章兢；

【作者基本信息】 湖南大学 ， 控制理论与控制工程， 2001， 硕士

【摘要】 本论文研究应用变压器油中溶解气体分析方法诊断变压器故障，针对气体传感器之间存在的“交叉敏感”特性提出了一种基于人工神经网络的智能嗅敏系统。研究了采用各气敏传感器组成的气敏元件阵列与人工神经网络来实现混合气体的组分与浓度识别的原理、方法和构成，提出了一种改进型误差反向传播网络训练方法，它具有良好的泛化能力。由于B-P网络具有极强的学习能力，能够充分逼近任意复杂的非线性关系，所以网络的学习过程是将隐含在训练样本中的气体之间复杂的非线性关系分布到各连接权上，形成所谓的“知识”或“经验”，这些“知识”或“经验”能较精确地反映复杂的气体信号，从而能构成对气体的较准确的分析测试。仿真试验显示了这一方法的可行性与优越性。结合变压器油中溶解气体微机在线监测项目讨论了系统硬件以及软件的具体实现。混合气体识别的传统方法——色谱法，采用将各混合气体逐次分离后再进行识别，此法装置复杂、费用昂贵；多变量分析模式识别法需要建立大量的气体响应方程，工作量非常大，使其实用范围受到影响。采用神经网络进行识别则克服了上述种种不便，而且模拟运行结果表明该法识别速度快、识别率较高，因其不再过分依赖传感器的选择性，是实现智能化检测混合气体一种很有前途的方法。更多还原

【Abstract】 In this paper the dissolved gas analysis is applied to diagnose the transformer fault,and an intelligent olfaction system based on artificial neural network is also presented toavoid the cross sensitivity among gas sensors. The principle, method, and constitution ofmixed gases identification of elements and concentration are studied by using gas sensorarray and artificial neural network. The improved error back propagation that has goodgeneralization ability is adopted to train the neural network. As B-P network has theability of study, it can approximate any complex non-linear relation. The knowledge andexperience which is distributed in the weights among the neural nodes is obtained bytraining the neural network, which can reflect the non-linear relation among the gassignals and analyze the mixed gases more accurately. Simulation experiment shows thefeasibility and advantage of this method. The hardware and software of microcomputeron-line detection of dissolved gas in transformer oil is also discussed how to realize.Chromatogram, the traditional, complex, expensive method, identifies the mixed gases byseparating them. Multi-variable analysis pattern recognition is narrowly employed becauseabundant gas response equations are needed. The method of artificial neural network notonly overcomes the above shortcoming but also identifies rapidly and accurately.Intelligent instrument of detecting mixed gases has a great future because the selectivity ofsensors is not so necessary.更多还原