【作者】 钟义长；

【导师】 吴新开；

【作者基本信息】 湖南科技大学 ， 控制理论与控制工程， 2007， 硕士

【摘要】 随着电力电子设备复杂性的不断增强、功能的不断完善,自动化程度的不断提高,能否及时、准确地诊断出设备的故障,以保证其平稳可靠运行也就变得越来越重要。论文在绪论中详细介绍了为什么要进行故障检测与诊断,故障诊断的重要性,以及传统故障珍断所面临的难题等;神经网络的基本知识,其中的BP神经网络知识是论文介绍的重点、对于目前用得较多的故障检测与诊断的其它方法也进行了一个系统的比较:最后从硬件的角度上,讲述了DSP在数据采集中所发挥的作用,从而使神经网络故障检测与诊断达到了硬件的实现。基于神经网络的电力电子设备的故障诊断分为两部分,一是故障样本数据及检验数据的采集;二是故障诊断。在故障样本数据及检验数据的采集过程中,首先建立了故障系统的Matlab模型,人为地让系统发生各种故障,从而采集到各种故障数据,经过归一化处理后,作为训练神经网络的样本数据;与时同时,对检验数据也进行采集,以检验训练出的神经网络是否能够起到故障诊断的作用。在故障诊断过程中,采用BP网络结构,针对BP算法收敛速度慢、容易陷入局部极小的缺点,采用“成批处理”的学习方法,这种方法在训练神经网络的过程中,不受学习样本排序的影响,使收敛速度加快;采用了改进的BP算法,采用共扼梯度法,在迭代过程中增加了惯性量。从而达到了加快网络的收敛速度与提高网络的实时诊断精度。针对本论文所确定的研究对象,利用训练好的网络对其进行了故障诊断,从检测结果中可以看出,基于神经网络的故障诊断方法在实际应用过程中是完全可行的。更多还原

【Abstract】 With the development of power electronics’ complexity, function and automation, it’s more vital for fault detection and diagnosis in time and for steady power electronics operation. This paper studys the importance of fault diagnosis, why fault detection and diagnosis is carryied on,and the diffculty which consists in the traditional fault detection and diagnosis and so on in the prologue. Basic knowledge of neural network, especially the BP neural network, is the emphasis of the paper, and the ways of fault diagnosis also are discussed. At last, the author proposes the methods of neural network(NN) hardware implementation by using DSP(Digital Signal Processing).The fault diagnosis based on neural network consists of two parts, which one is the collection of sample data of fault and the other is fault diagnosis. Before the collection of sample data of fault and checking data, the paper establishs the Matlab model of fault diagnosis firstly. All kinds of fault are made artificially and collected by the Matlab model. After normalization, the sample data of fault is servered as the data of training neural network. At the same time, the checking data is used to examine whether the trained neural network have the function of fault detection and diagnosis. There are shortcoming of the slow constringency rate and disadvantage that error easily gets into partial minimal value in the process of diagnosis by BP algorithm in BP network. So it adopts the following solutions. Firstly, the learning method of "batch dealing" is used. During the training neural network, this method accelerates the speed of constringency while neglects the affection of studying sample orders. Secondly, the improving BP arithmetic is adopted, which is so called conjugated gradient method, which the inertia is iterative in the training process. The aim to accelect the speed of convergence and improve realtime diagnosis accuracy is achieved.For the research object, the data of the model’s electric currents and voltages are collected to train neural network and diagnosed. The result proves that the fault detection and diagnosis based on neural network can be used in reality.更多还原