SIA OpenIR  > 工业控制网络与系统研究室
基于深度学习的变电站设备故障诊断算法研究
Alternative TitleResearch on Substation Equipment Fault Diagnosis Algorithm Based on Deep Learning
徐志远
Department工业控制网络与系统研究室
Thesis Advisor王忠锋
Keyword红外监测 故障诊断 深度学习 压缩与加速 图像分类
Pages74页
Degree Discipline模式识别与智能系统
Degree Name硕士
2020-05-26
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract红外监测技术是一种无接触式的测量技术。在测量过程中不与被测设备直接相连,所以不会给系统造成额外负担。因此,红外监测技术越来越受到重视,被应用于各行各业的监测任务中。电力行业引进红外监测技术主要是为了完成一些电力巡检和设备状态监控任务。变电站作为电网系统中最重要的一个环节之一,维护变电站系统稳定、安全运行对电网系统有着十分重要的意义。在变电站设备故障诊断的工作中融入红外监测技术可以在不停电的情况下完成在线监测任务,不仅如此,变电站系统运行在高电压高电流的条件下,被测设备周围可能会存在比较强的电磁场,这就导致一些传统的测量传感器出现不灵敏甚至是失效的现象,而红外监测技术在采集红外热图像时则不会受到电磁场的干扰,测量出的数据更接近设备的真实状态。传统处理红外热图像的方法是对图像进行分割,提取目标区域,这种方法需要人为设定分割算法的参数,对于拍摄角度、距离、图像质量的前提要求太苛刻。近十几年来,人工智能技术得到了飞速发展,深度学习已经成为图像识别领域主流的计算模型。所以本文采用深度学习的方法,利用深度卷积神经网络提取红外图像的特征然后对这些特征进行故障的识别和分类。这种方法可以省去人工提取图像特征的工作,而且有着比较高的识别精度。本文的第一个创新点是图像预处理方案,图像预处理就是在提取红外图像特征之前,去除红外图像的噪声和干扰,便于接下来的图像识别工作。传统的预处理方法是使用一些常规的滤波器如中值滤波器、高斯滤波器。这种方法的缺点是,在去除噪声和干扰的同时丢失了图像的边缘或纹理等细节信息。本文采用的预处理方案是首先利用图像通道重组的方法增强特征区域的对比度,有利于特征提取。然后使用双边滤波器对通道重组后的图像进行滤波。双边滤波器在去除图像的噪声的同时能保留图像中的细节信息,有利用于故障和设备的识别。本文的第二个创新点是针对深度网络模型计算时需要占用大量的存储空间和计算资源的问题设计了一种计算模型轻量化的方法,有效减少了深度网络的计算量和存储空间的占用。深度网络计算模型通常部署在云端,计算时首先将本地数据上传至云端,然后由云端处理数据,最后云端返回给本地计算结果。但是对于变电站设备故障,由于故障发生周期很短,基于云端的故障诊断方案存在一定的传输时延,如果故障发生在数据传输的过程中,那么故障诊断将是毫无意义的。而且变电站中的数据是保密的,上传至云端还存在数据安全的问题。本文设计了一种基于梯度的深度网络剪枝算法——GP算法,通过修剪网络中冗余的参数,使模型获得加速计算的效果并且有效减少了对存储空间的占用。计算模型轻量化的目的是使深度网络能够部署在计算和存储资源相对受限的嵌入式设备中,在本地嵌入式设备中处理数据能够同时解决传输时延和数据安全这两个问题。最后针对变电站设备故障诊断的工作中的问题,在上述研究的基础上,本文设计了变电站故障诊断的方案和软件平台,通过对绝缘子和刀闸开关故障诊断的实验,结果表明故障诊断的正确率可达90%,证明该方案是可行的。
Other AbstractInfrared monitoring technology is a non-contact measurement technology. It is not directly connected with the device under test during the measurement process, so it will not cause additional burden to the system. Therefore, infrared monitoring technology has received more and more attention, and it has been applied to monitoring tasks in various industries. The introduction of infrared monitoring technology in the power industry is mainly to complete some power inspection and monitoring tasks. The substation is one of the most important links in the power grid system. Maintaining the long-term, stable and safe operation of the substation system is of great significance to the power grid system. Incorporating infrared monitoring technology into the fault diagnosis of substation equipment can complete online monitoring tasks without power outages. Not only that, the substation system runs in a high voltage and high current environment, and there must be a relatively strong electromagnetic field around the equipment under test. This has caused some traditional sensors to be insensitive or even ineffective. Infrared monitoring technology will not be disturbed by electromagnetic fields when collecting infrared thermal images, and the measured data is closer to the true state of the device. The traditional method of processing the infrared thermal image of the checked device is to segment the image and extract the target area. This method requires artificially setting the parameters of the segmentation algorithm, and the prerequisites for the capturing angle, distance, and image quality are too harsh. In recent years, with the development of artificial intelligence technology, deep learning has become the mainstream computing model in the field of image recognition. Therefore, this paper uses deep learning methods, using deep convolution networks to extract features of infrared images, and then identify and classify these features. This method can save the work of manually extracting image features. The first innovation of this paper is the image pre-processing scheme. Image pre-processing is to remove noise and interference from the infrared image before extracting the infrared image features, which is easy for the next image recognition work the pre-processing method adopted in this paper is that the first uses the method of image channel reconstruction to enhance infrared features, which is helpful for fault diagnosis. Then use bilateral filters to filter the image after channel reconstruction to remove noise. The bilateral filter can remove the Gaussian noise of the image while maintaining detailed information such as edges in the image, which is beneficial for the identification of faulty equipment. The second innovation of this paper is to design a lightweight method for the calculation model for the problem that the deep network model requires a large amount of storage space and computing resources, which effectively reduces the calculation amount and storage space of the deep network. the conventional method is that the first deploy the deep network model on a cloud platform, the second upload the data to the cloud for calculation, and finally return the calculation results. However, in the case of substation equipment faults, the fault occurrence cycle is very short. There is a certain transmission delay in fault diagnosis based on the cloud platform. If the fault occurs during the data transmission, then the fault diagnosis will be meaningless. In addition, that the data in the substation is uploaded to the cloud is also a problem of data security. Aiming at the limitations of the deep network model, this paper has done the work of light-weighting the computational model. In this paper, Gradient-based deep network pruning algorithm is designed. By pruning the redundant weights in the network, the model can obtain the effect of accelerated calculation and occupy less storage space. The purpose of lightweight computing models is to enable deep networks to be deployed in embedded devices with limited computing and storage resources for local fault diagnosis, and to solve transmission delay and data security issues. Finally, aiming at the problems in the substation equipment fault diagnosis work, on the basis of the above research, this paper designs the substation fault diagnosis scheme and software platform. In the insulator and knife switch fault diagnosis, the correct rate of fault diagnosis can reach 90%., and proves that the scheme is feasible.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/27139
Collection工业控制网络与系统研究室
Affiliation中国科学院沈阳自动化研究所
Recommended Citation
GB/T 7714
徐志远. 基于深度学习的变电站设备故障诊断算法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2020.
Files in This Item:
File Name/Size DocType Version Access License
基于深度学习的变电站设备故障诊断算法研究(3001KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[徐志远]'s Articles
Baidu academic
Similar articles in Baidu academic
[徐志远]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[徐志远]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.