深度学习在局部放电模式识别中的应用

发布时间：2018-05-24 08:45

本文选题：局部放电 + 模式识别　；参考：《华北电力大学(北京)》2016年硕士论文

【摘要】：局部放电是电力设备运行中的常见现象,它不仅表征着设备的绝缘强度、运行情况等重要信息,而且长期的放电也会扩大缺陷导致击穿。因此,对局部放电进行正确有效的种类识别,对于设备监测和故障诊断有着非常重要的意义。目前,在局放识别中经常使用的BP神经网络算法,可以从一定程度上实现有效判别,但是由于其训练机制上的限制,其识别能力有限,改进的空间并不大。而近年来在模式识别领域出现的深度学习方法,从一定程度上克服了BP神经网络的上述缺点,已经越来越广泛地应用到各类识别领域,并得到了较好的效果。为了改进神经网络在局放模式识别的应用现状,本文主要研究引入深度学习机制来实现局部放电模式识别。具体来说,为了研究各类局部放电的特性,建立用于识别研究的样本库,本文搭建了基于脉冲电流法的局放测试平台。针对局部放电中常见的内部放电、沿面放电和电晕放电三个基本类别,设计实验模型并进行了油纸绝缘、合成绝缘材料气隙、空气中沿面、SF6中沿面、空气中针板、SF6中针板6个细分组别的测试。分别采集了各组实验的放电发展期特征数据,整理成每个组别的典型样本集,用于训练和测试。为了优化网络性能,避免输入维度过高带来的过拟合,本文又提出了基于参量分布的蒙特卡洛方法,作为对局放实验的补充。通过对放电量和放电相位两个维度进行统计分布研究求取二维随机变量的联合概率分布。并设计了模拟实验,可以实现参数可控、统计分散的样本扩容。基于以上样本,本文对局部放电三个类别、6个组别分别进行了识别。研究对比了BP神经网络和基于自动编码机的深度学习网络的识别性能差异。结果表明,自编码网络的识别效果整体优于BP神经网络,其总体识别率更高,识别率的输出更为稳定。但相应的训练时间也有所增长。因此,本文认为基于自动编码机的深度学习网络,因其训练机制上的改进,识别效果优于传统BP神经网络。这也为利用深度学习方法处理局放模式识别的进一步研究提供了参考依据。
[Abstract]:Partial discharge (PD) is a common phenomenon in the operation of power equipment. It not only represents the important information of insulation strength and operation of the equipment, but also extends the defects and leads to breakdown. Therefore, it is very important for equipment monitoring and fault diagnosis to identify PD correctly and effectively. At present, BP neural network algorithm, which is often used in PD recognition, can realize effective discrimination to a certain extent, but because of its limited training mechanism, its recognition ability is limited, and the improvement space is not large. In recent years, the deep learning method in the field of pattern recognition has overcome the shortcomings of BP neural network to a certain extent, and has been more and more widely used in all kinds of recognition fields, and got better results. In order to improve the application of neural network in PD pattern recognition, the depth learning mechanism is introduced to realize PD pattern recognition. Specifically, in order to study the characteristics of various partial discharges and establish a sample library for identification research, a partial discharge test platform based on pulse current method is built in this paper. Aiming at the three basic types of internal discharge, surface discharge and corona discharge, the paper designs the experimental model and carries on the oil-paper insulation, the synthetic insulation material air gap, the surface discharge in the air and the middle side of SF _ 6 in the air. Test of 6 subdivision groups of needle board in SF 6 in air. The characteristic data of the discharge development period of each group were collected, and the typical samples of each group were collected for training and testing. In order to optimize the network performance and avoid the over-fitting caused by the high input dimension, a Monte Carlo method based on parameter distribution is proposed as a supplement to the partial discharge experiment. The joint probability distribution of two-dimensional random variables is obtained by statistical distribution of discharge quantity and discharge phase. The simulation experiment is designed, which can realize the expansion of the sample with controllable parameters and statistical dispersion. Based on the above samples, three types of PD are identified, and six groups are identified. The difference of recognition performance between BP neural network and depth learning network based on automatic coding machine is studied and compared. The results show that the recognition effect of the self-coding network is better than that of the BP neural network, and the overall recognition rate is higher and the output of the recognition rate is more stable. But the corresponding training time has also increased. Therefore, this paper considers that the recognition effect of the deep learning network based on the automatic coding machine is better than that of the traditional BP neural network because of the improvement of its training mechanism. It also provides a reference for the further research of PD pattern recognition by using the deep learning method.
【学位授予单位】：华北电力大学(北京)
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TM855

【相似文献】