【作者】 杨涛；

【导师】 赵曙光；

【作者基本信息】 东华大学 ， 电力电子与电力传动， 2021， 硕士

【摘要】 随着自动化设备与人工智能日新月异的发展,我国电网建设也发展到了智能化的新阶段。在变电站领域,根据国家电网制定的发展规划要求,在传统变电站的运行模式满足不了当前需求的情况下,智能变电站的应用普及已势在必行。在智能变电站的建设和运营中,日常检测等过程中需要处理大量的电力设备红外图像,如何高效处理这些海量数据已成为关键性问题和瓶颈。本文以研究和解决这一问题为目标,首先研究了红外图像的滤波方法,以尽量消除红外图像的噪声;然后研究了电力设备故障区域的图像分割方法,为后续的故障诊断奠定基础;最后,研究了随机森林与深度学习相结合的电力设备故障分类方法;对上述方法均进行了实验验证并获得了较好的效果。本文的具体研究内容和结果主要包括:(1)在红外图像滤波消噪方面,本文在分析对比常用的有关方法后,选用了自适应中值滤波方法。它可以根据像素点是否是噪声,自动地改变滤波窗口的大小,并使用最适当的滤波方法,从而能既有效去噪又保留图像的细节部分,获得明显优于常规中值滤波器的效果。(2)为了能够快速提取红外图像中电力设备故障区域,本文提出了一种基于OTSU(最大类间方差法)和PCNN(Pulse Coupled Neural Network,脉冲耦合神经网络)的提取方法,并通过实验证明其能够准确、完整地提取电力设备的故障区域。(3)本文将红外图像与机器学习方法结合起来对电力设备故障进行分类。首先针对传统Alex Net网络训练时间长且出现过拟合的问题,对网络进行了改进,减少了卷积层数量,加快了训练速度。在本文算法中,一个Alex Net网络提取电力设备原始图像的深度特征,另一个Alex Net网络提取电力设备故障区域图像的有效特征,这两种特征经过融合后输入随机森林算法中进行训练,随机森林通过随机选取样本特征子集和样本特征子向量构建多棵决策树,然后利用决策树构建的随机森林进行投票从而得到电力设备的故障类型。按照上述方法,先利用训练数据集对其模型进行训练,再利用训练好的模型对实际数据进行故障分类,实现电力设备故障诊断,实验中获得较好的效果。更多还原

【Abstract】 With the rapid development of automation equipment and artificial intelligence,China’s power grid construction has also developed to a new stage of intelligence.In the field of power transformation,according to the development planning requirements formulated by the State Grid,the application and popularization of Smart Substation is imperative under the situation that the operation mode of traditional substation can’t meet the current demand.In the construction and operation of intelligent substation,a large number of infrared images of power equipment need to be processed in the process of daily detection.How to efficiently process these massive data has become a key problem and bottleneck.In order to study and solve this problem,firstly,the filtering method of infrared image is studied in thesis to eliminate the noise of infrared image;secondly,the image segmentation method of power equipment fault area is studied to lay the foundation for subsequent fault diagnosis;finally,the fault classification method of power equipment based on random forest and deep learning is studied It is verified by experiments and good results are obtained.The main contents and results of this thesis are as follows:(1)In the aspect of infrared image denoising,the adaptive median filtering method is selected after analyzing and comparing the commonly used methods.It can automatically change the size of the filtering window according to whether the pixel is noise or not,and uses the most appropriate filtering method,which can not only effectively denoise but also retain the details of the image,and obtain a better effect than the conventional median filter.(2)In order to quickly extract the fault area of power equipment from infrared image,this thesis proposes an extraction method based on OTSU(maximum inter class variance)and PCNN(pulse coupled neural network).The experiment shows that it can extract the fault area of power equipment accurately and completely.(3)In this thesis,infrared image and machine learning method are combined to classify power equipment faults.Firstly,aiming at the problem of long training time and over fitting of traditional Alexnet network,the network is improved to reduce the number of convolution layers and accelerate the training speed.In this algorithm,one Alexnet network extracts the depth features of the original image of power equipment,and the other extracts the effective features of the fault area image of power equipment.After fusion,the two features are input into the random forest algorithm for training.The random forest builds multiple decision trees by randomly selecting the sample feature subset and sample feature sub vector,and then uses the decision tree The random forest constructed by policy tree can vote to get the fault type of power equipment.When using the above method,the training data set should be used to train the model,and then the trained model should be used to classify the actual data to realize the fault diagnosis of power equipment.更多还原