CN110796206A - A data enhancement method and device for partial discharge map - Google Patents
A data enhancement method and device for partial discharge map Download PDFInfo
- Publication number
- CN110796206A CN110796206A CN201911074459.5A CN201911074459A CN110796206A CN 110796206 A CN110796206 A CN 110796206A CN 201911074459 A CN201911074459 A CN 201911074459A CN 110796206 A CN110796206 A CN 110796206A
- Authority
- CN
- China
- Prior art keywords
- data
- module
- noise
- weight
- weight matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000010168 coupling process Methods 0.000 claims abstract description 44
- 238000005859 coupling reaction Methods 0.000 claims abstract description 44
- 230000008878 coupling Effects 0.000 claims abstract description 41
- 238000012549 training Methods 0.000 claims abstract description 20
- NINIDFKCEFEMDL-UHFFFAOYSA-N Sulfur Chemical compound [S] NINIDFKCEFEMDL-UHFFFAOYSA-N 0.000 claims abstract description 8
- 229910052717 sulfur Inorganic materials 0.000 claims abstract description 8
- 239000011593 sulfur Substances 0.000 claims abstract description 8
- 239000011159 matrix material Substances 0.000 claims description 53
- 238000006243 chemical reaction Methods 0.000 claims description 12
- 238000005315 distribution function Methods 0.000 claims description 10
- 238000000605 extraction Methods 0.000 claims description 9
- 238000013135 deep learning Methods 0.000 claims description 6
- 239000000284 extract Substances 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims 2
- 238000004364 calculation method Methods 0.000 abstract description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 239000000049 pigment Substances 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Processing (AREA)
Abstract
本发明提出了一种针对局部放电图谱的数据增强方法和装置,在添加现场常见噪声如手机干扰、雷达干扰、微波硫灯干扰的基础上应用高斯模糊方法,对RGB像素进行随机扰动,用以训练泛化能力更强的网络。通过利用背景噪声耦合和高斯模糊等方法对局部放电图谱进行预处理,使一张图谱可以生成多张图谱,以较少的计算量生成大量带标签样本,解决了带标签样本获取成本高、训练数据不充分的问题;在原有数据样本的基础上考虑了现场的常见干扰,模拟真实现场数据,增加了训练模型的抗干扰能力;扩展了训练数据的多样性,以扩展后的训练数据训练模型,避免了模型的过拟合。
The invention proposes a data enhancement method and device for partial discharge atlas. On the basis of adding common noises such as mobile phone interference, radar interference, and microwave sulfur lamp interference, the Gaussian blurring method is applied to randomly perturb RGB pixels for the purpose of Train a network that generalizes better. By using background noise coupling and Gaussian blur to preprocess the partial discharge map, one map can generate multiple maps, and a large number of labeled samples can be generated with a small amount of calculation, which solves the problem of the high cost of obtaining labeled samples and the need for training. The problem of insufficient data; on the basis of the original data samples, the common disturbance on the site is considered, and the real field data is simulated to increase the anti-interference ability of the training model; the diversity of the training data is expanded, and the model is trained with the expanded training data. , to avoid overfitting of the model.
Description
技术领域technical field
本发明涉及局部放电检测领域,更具体地,本发明涉及一种针对局部放电图谱的数据增强方法。The present invention relates to the field of partial discharge detection, and more particularly, to a data enhancement method for partial discharge maps.
背景技术Background technique
随着智能电网的发展,对数据诊断的需求不断增加。机器学习诊断是一种替代人工诊断的解决方案。而在搭建机器学习模型对数据进行训练、验证时,如何快速、高效、便捷得获取大量带标签的训练样本是一个亟待解决的问题。With the development of smart grids, the demand for data diagnostics is increasing. Machine learning diagnosis is an alternative to manual diagnosis. When building a machine learning model to train and verify data, how to quickly, efficiently and conveniently obtain a large number of labeled training samples is an urgent problem to be solved.
局部放电图谱识别需要大量带标签的训练样本。带标签的样本的收集过程成本较高,对发生频率低的局部放电现象往往很难快速累积起足够的样本数量用以支撑训练。传统的局部放电数据增强方法中,一般以平移、旋转、缩放等简单处理方式为主。以这些方式处理的图谱数据:1)没有考虑到真实业务情况中数据可能受到的干扰;2)基于形状处理的数据增强方法没有对数据像素做扩展,训练出的模型仅适用于指定颜色、规格的图像,对不同来源的数据泛化性差。Partial discharge pattern recognition requires a large number of labeled training samples. The collection process of labeled samples is expensive, and it is often difficult to quickly accumulate enough samples to support training for the low-frequency partial discharge phenomenon. In traditional partial discharge data enhancement methods, simple processing methods such as translation, rotation, and scaling are generally used. The graph data processed in these ways: 1) The possible interference of the data in the real business situation is not considered; 2) The data enhancement method based on shape processing does not expand the data pixels, and the trained model is only suitable for the specified color and specification. images, which generalize poorly to data from different sources.
发明内容SUMMARY OF THE INVENTION
为了克服现有技术中的缺陷,本发明利用背景噪声耦合和高斯模糊等方法对局部放电图谱进行预处理,使一张图谱可以生成多张图谱。能够解决局部放电图谱识别中数据不足的方法,通过以一张图谱生成多张图谱的方式,增加样本数量,避免过拟合。In order to overcome the defects in the prior art, the present invention uses methods such as background noise coupling and Gaussian blur to preprocess the partial discharge atlas, so that one atlas can generate multiple atlases. The method that can solve the problem of insufficient data in the identification of partial discharge maps can increase the number of samples and avoid overfitting by generating multiple maps from one map.
本发明设计了一种基于噪声扰动和高斯模糊的数据增强方法和装置,在添加现场常见噪声如手机干扰、雷达干扰、微波硫灯干扰的基础上应用高斯模糊方法,对RGB像素进行随机扰动,用以训练泛化能力更强的网络。The invention designs a data enhancement method and device based on noise disturbance and Gaussian blurring. The Gaussian blurring method is applied on the basis of adding common noises such as mobile phone interference, radar interference, and microwave sulfur lamp interference to randomly disturb RGB pixels. It is used to train a network with stronger generalization ability.
具体地,本发明所述的针对局部放电图谱的数据增强方法,利用背景噪声耦合和高斯模糊方法对局部放电图谱进行预处理,使一张图谱可以生成多张图谱,包括以下步骤:Specifically, the data enhancement method for partial discharge atlas of the present invention utilizes background noise coupling and Gaussian blurring to preprocess the partial discharge atlas, so that one atlas can generate multiple atlas, including the following steps:
S1:对局部放电原始数据进行噪声耦合,生成图片文件;S1: Perform noise coupling on the original data of partial discharge to generate image files;
S2:对噪声耦合后的图片文件采用高斯模糊方法进一步处理;S2: The image file after noise coupling is further processed by the Gaussian blur method;
S3:合并经过噪声耦合、高斯模糊的样本和原样本,用于深度学习网络模型的训练。S3: Combine noise-coupling, Gaussian blurred samples and original samples for training of deep learning network models.
其中步骤一具体包括:The first step specifically includes:
S11:将采集前端采集到的局部放电原始数据作为输入数据;S11: Use the original partial discharge data collected by the collection front end as input data;
S12:将输入数据转成三维数据;S12: Convert the input data into three-dimensional data;
S13:按照噪声干扰的数据特征,分别生成各噪声对应的三维数据;S13: According to the data characteristics of noise interference, generate three-dimensional data corresponding to each noise respectively;
S14:对输入数据和噪声数据进行累加,得到噪声耦合后的数据;S14: Accumulate the input data and the noise data to obtain the noise-coupled data;
S15:将噪声耦合后的数据转换成图片文件。S15: Convert the noise-coupled data into a picture file.
进一步地,further,
所述三维数据以相位为x轴,周期为y轴,幅值为z轴。The three-dimensional data takes the phase as the x-axis, the period as the y-axis, and the amplitude as the z-axis.
进一步地,further,
所述噪声干扰包括雷达噪声、手机噪声、微波硫灯干扰。The noise interference includes radar noise, mobile phone noise, and microwave sulfur lamp interference.
进一步地,步骤二具体包括:Further, step 2 specifically includes:
S21:将步骤一中噪声耦合后的图片文件作为输入数据;S21: The image file after noise coupling in step 1 is used as input data;
S22:提取输入图片文件的RGB(Red,Green,Blue)值;S22: Extract the RGB (Red, Green, Blue) values of the input image file;
S23:分别取图片文件中图片的中心点为零点,绘制横纵坐标轴;S23: respectively take the center point of the picture in the picture file as the zero point, and draw the horizontal and vertical axes;
S24:以二维高斯分布函数计算图片中各像素点的权重,形成权重矩阵;S24: Calculate the weight of each pixel in the image with a two-dimensional Gaussian distribution function to form a weight matrix;
S25:对权重矩阵进行归一化处理;S25: normalize the weight matrix;
S26:以归一化后的权重更新图片文件的RGB值;S26: Update the RGB value of the image file with the normalized weight;
S27:存储更新RGB值后的图片文件,作为新的样本。S27: Store the image file after updating the RGB value as a new sample.
进一步地,further,
步骤S24中所述的二维高斯分布函数如下:The two-dimensional Gaussian distribution function described in step S24 is as follows:
式中,x,y分别为图中各像素点距离零点的横纵坐标,G(x,y)为该点到零点的权重值,π为圆周率,e为自然常数,σ为正态分布的标准差,σ通常取1至3之间,取值越大图像越平滑。In the formula, x and y are the horizontal and vertical coordinates of each pixel in the figure from the zero point, G(x, y) is the weight value from the point to the zero point, π is the pi, e is the natural constant, and σ is the normal distribution. Standard deviation, σ usually ranges from 1 to 3, the larger the value, the smoother the image.
进一步地,further,
步骤S25中,对权重矩阵进行归一化处理包括:In step S25, normalizing the weight matrix includes:
计算权重矩阵中所有权重数值之和m,对权重矩阵中的每个权重值乘以1/m,得到归一化后的权重矩阵。Calculate the sum m of all weight values in the weight matrix, multiply each weight value in the weight matrix by 1/m, and obtain the normalized weight matrix.
进一步地,步骤S26中,Further, in step S26,
对每一个零点的权重矩阵,以各归一化后的权重矩阵的权重值乘以该权重值位置的像素值,并求和,作为该零点的新的像素值,以该种权重更新方法对RGB值分别更新,得到更新后的三色像素值。For the weight matrix of each zero point, the weight value of each normalized weight matrix is multiplied by the pixel value of the weight value position, and the sum is used as the new pixel value of the zero point. The RGB values are updated respectively to obtain the updated three-color pixel values.
本发明还提出一种针对局部放电图谱的数据增强装置,包括:The present invention also provides a data enhancement device for partial discharge atlas, including:
噪声耦合模块,高斯模糊模块和样本生成模块,其中噪声耦合模块的输出与高斯模糊模块输入相连,所述高斯模块的输出与样本生成模块相连,利用背景噪声耦合和高斯模糊方法对局部放电图谱进行预处理,使一张图谱可以生成多张图谱。A noise coupling module, a Gaussian blurring module and a sample generating module, wherein the output of the noise coupling module is connected with the input of the Gaussian blurring module, the output of the Gaussian module is connected with the sample generating module, and the partial discharge map is performed by using the background noise coupling and the Gaussian blurring method. Preprocessing enables one map to generate multiple maps.
进一步地,further,
所述噪声耦合模块包括数据输入模块、三维数据生成模块、数据耦合模块和图片转换模块,其中数据输入模块的输出与三维数据生成模块的输入相连,三维数据生成模块的输出与数据耦合模块的输入相连,数据耦合模块的输出与图片转换模块的输入相连。The noise coupling module includes a data input module, a 3D data generation module, a data coupling module and a picture conversion module, wherein the output of the data input module is connected with the input of the 3D data generation module, and the output of the 3D data generation module is connected with the input of the data coupling module. The output of the data coupling module is connected with the input of the picture conversion module.
进一步地,further,
所述数据输入模块,输入采集前端采集到的局部放电原始数据。The data input module inputs the original partial discharge data collected by the collection front end.
进一步地,further,
所述三维数据生成模块,将输入数据转成以相位为x轴,周期为y轴,幅值为z轴的三维数据;并且按照雷达噪声、手机噪声、微波硫灯干扰的数据特征,分别生成各噪声对应的三维数据。The three-dimensional data generation module converts the input data into three-dimensional data with the phase as the x-axis, the period as the y-axis, and the amplitude as the z-axis; and according to the data characteristics of radar noise, mobile phone noise, and microwave sulfur lamp interference, respectively 3D data corresponding to each noise.
进一步地,further,
所述数据耦合模块,对输入数据和噪声数据的三维数据进行累加,得到噪声耦合后的数据。The data coupling module accumulates the three-dimensional data of the input data and the noise data to obtain noise-coupled data.
进一步地,further,
所述图片转换模块,将噪声耦合后的数据转换成图片文件。The picture conversion module converts the noise-coupled data into a picture file.
进一步地,further,
所述高斯模糊模块包括像素提取模块、权值矩阵生成模块、权值更新模块和文件生成模块,其中所述图片转换模块的输出连接像素提取模块的输入,像素提取模块的输出连接权值矩阵生成模块的输入,权值矩阵生成模块的输出连接权值更新模块的输入,权值更新模块的输出连接文件生成模块的输入,用以对噪声耦合后的图片进行高斯模糊处理。The Gaussian blurring module includes a pixel extraction module, a weight matrix generation module, a weight update module and a file generation module, wherein the output of the picture conversion module is connected to the input of the pixel extraction module, and the output of the pixel extraction module is connected to the weight matrix to generate The input of the module and the output of the weight matrix generation module are connected to the input of the weight update module, and the output of the weight update module is connected to the input of the file generation module to perform Gaussian blurring on the noise-coupled picture.
进一步地,所述像素提取模块,输入图片转换模块转换后的图片文件,提取所述的图片文件的RGB(Red,Green,Blue)值。Further, the pixel extraction module inputs a picture file converted by the picture conversion module, and extracts RGB (Red, Green, Blue) values of the picture file.
进一步地,further,
所述权值矩阵生成模块,分别取图片文件中图片的中心点为零点,绘制横纵坐标轴;The weight matrix generation module takes the center point of the picture in the picture file as the zero point, respectively, and draws the horizontal and vertical axes;
以二维高斯分布函数计算图片中各像素点的权重,形成权重矩阵,并进一步得到归一化的权重矩阵。Calculate the weight of each pixel in the picture with a two-dimensional Gaussian distribution function, form a weight matrix, and further obtain a normalized weight matrix.
进一步地,further,
所述二维高斯分布函数如下:The two-dimensional Gaussian distribution function is as follows:
式中,x,y分别为图中各像素点距离零点的横纵坐标,G(x,y)为该点到零点的权重值,π为圆周率,e为自然常数,σ为正态分布的标准差,σ通常取1至3之间,取值越大图像越平滑。In the formula, x and y are the horizontal and vertical coordinates of each pixel in the figure from the zero point, G(x, y) is the weight value from the point to the zero point, π is the pi, e is the natural constant, and σ is the normal distribution. Standard deviation, σ usually ranges from 1 to 3, the larger the value, the smoother the image.
进一步地,further,
得到归一化的权重矩阵具体包括:The normalized weight matrix specifically includes:
计算权重矩阵中所有权重数值之和m,对权重矩阵中的每个权重值乘以1/m,得到归一化后的权重矩阵。Calculate the sum m of all weight values in the weight matrix, multiply each weight value in the weight matrix by 1/m, and obtain the normalized weight matrix.
进一步地,further,
所述权值更新模块,以权重更新图片文件的RGB值。The weight updating module updates the RGB value of the picture file with the weight.
进一步地,further,
以权重更新图片的RGB值具体包括:Updating the RGB values of the image with weights specifically includes:
对每一个零点的权重矩阵,以各归一化后的权重矩阵的权重值乘以该权重值位置的像素值,并求和,作为该零点的新的像素值,以该种权重更新方法对RGB值分别更新,得到更新后的三色像素值。For the weight matrix of each zero point, the weight value of each normalized weight matrix is multiplied by the pixel value of the weight value position, and the sum is used as the new pixel value of the zero point. The RGB values are updated respectively to obtain the updated three-color pixel values.
进一步地,further,
所述文件生成模块,存储更新RGB值后的图片文件,作为新的样本。The file generation module stores the image file after updating the RGB value as a new sample.
进一步地,further,
所述样本生成模块,合并经过噪声耦合、高斯模糊的样本和原样本,用于深度学习网络模型的训练。The sample generation module combines the noise-coupling, Gaussian blurred samples and the original samples for training of the deep learning network model.
与直接使用样本作为训练数据相比,本发明所述的方法体现了如下优点:Compared with directly using samples as training data, the method described in the present invention has the following advantages:
(1)以较少的计算量生成大量带标签样本,解决了带标签样本获取成本高、训练数据不充分的问题;(1) Generate a large number of labeled samples with less computation, which solves the problems of high cost of obtaining labeled samples and insufficient training data;
(2)在原有数据样本的基础上考虑了现场的常见干扰,模拟真实现场数据,增加了训练模型的抗干扰能力;(2) On the basis of the original data samples, the common disturbances in the field are considered, the real field data is simulated, and the anti-interference ability of the training model is increased;
(3)扩展了训练数据的多样性,从基础样本扩展到多像素、多形状的样本。以扩展后的训练数据训练模型,可以避免模型的过拟合。(3) Expand the diversity of training data, from basic samples to multi-pixel and multi-shape samples. Training the model with the expanded training data can avoid overfitting of the model.
附图说明Description of drawings
图1是本发明的一种基于噪声扰动和高斯模糊的数据增强方法的流程图。Fig. 1 is a flow chart of a data enhancement method based on noise disturbance and Gaussian blur of the present invention.
图2是本发明的一种基于噪声扰动和高斯模糊的数据增强装置的示意图。FIG. 2 is a schematic diagram of a data enhancement device based on noise disturbance and Gaussian blur according to the present invention.
图3是未经处理的原始数据三维图像。Figure 3 is an unprocessed 3D image of the raw data.
图4是噪声叠加后的图像文件。Figure 4 is the image file after the noise is superimposed.
图5是叠加高斯模糊后的图像文件。Figure 5 is the image file after superimposed Gaussian blur.
具体实施方式Detailed ways
为使本发明实施例的目的、技术方案和优点更加清楚,下面将结合本发明实施例的附图,对本发明实施例的技术方案进行清楚、完整地描述。显然,所描述的实施例是本发明的一部分实施例,而不是全部的实施例。基于所描述的本发明的实施例,本领域普通技术人员所获得的所有其他实施例,都属于本发明保护的范围。In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art fall within the protection scope of the present invention.
参照附图1所示的流程图,本发明采取的技术方案是,一种针对局部放电图谱基于噪声扰动和高斯模糊的数据增强方法,包括以下步骤:Referring to the flowchart shown in FIG. 1 , the technical solution adopted in the present invention is a data enhancement method based on noise disturbance and Gaussian blurring for partial discharge maps, comprising the following steps:
(1)噪声耦合(1) Noise coupling
①输入数据为采集前端采集到的局部放电原始数据;①The input data is the original partial discharge data collected by the front-end;
②将输入数据转成以相位为x轴,周期为y轴,幅值为z轴的三维数据;② Convert the input data into three-dimensional data with the phase as the x-axis, the period as the y-axis, and the amplitude as the z-axis;
③按照雷达噪声、手机噪声、微波硫灯干扰的数据特征,分别生成各噪声对应的三维数据;③ According to the data characteristics of radar noise, mobile phone noise, and microwave sulfur lamp interference, generate three-dimensional data corresponding to each noise respectively;
④对输入数据和噪声数据的三维数据进行累加,得到噪声耦合后的数据;④ Accumulate the three-dimensional data of the input data and the noise data to obtain the data after noise coupling;
⑤将噪声耦合后的数据转换成图片文件。⑤ Convert the noise-coupled data into image files.
(2)高斯模糊(2) Gaussian blur
①输入数据为噪声耦合后的图片文件;①The input data is the image file after noise coupling;
②提取输入图片文件的RGB(Red,Green,Blue)值;② Extract the RGB (Red, Green, Blue) values of the input image file;
③分别取图片文件中图片的中心点为零点,绘制横纵坐标轴;以二维高斯分布函数计算图片中各像素点的权重矩阵,二维高斯分布如下:③ Take the center point of the picture in the picture file as the zero point, and draw the horizontal and vertical axes; calculate the weight matrix of each pixel in the picture with the two-dimensional Gaussian distribution function, and the two-dimensional Gaussian distribution is as follows:
式中,x,y分别为图中各像素点距离零点的横纵坐标,G(x,y)为该点到零点的权重值,π为圆周率,e为自然常数,σ为正态分布的标准差,σ通常取1至3之间,取值越大图像越平滑;In the formula, x and y are the horizontal and vertical coordinates of each pixel in the figure from the zero point, G(x, y) is the weight value from the point to the zero point, π is the pi, e is the natural constant, and σ is the normal distribution. Standard deviation, σ usually takes between 1 and 3, the larger the value, the smoother the image;
以3*3像素的图像为例,以图像中点为零点时,各点的(x,y)取值如下表所示:Taking an image of 3*3 pixels as an example, when the point in the image is zero, the (x, y) values of each point are shown in the following table:
取σ为1.5,对各组(x,y)分别计算权重,结果如下表所示:Take σ as 1.5, and calculate the weights for each group (x, y). The results are shown in the following table:
④对权重矩阵归一化:计算权重矩阵中所有权重数值之和m,对权重矩阵中的每个权重值乘以1/m,得到归一化后的权重矩阵;④ Normalize the weight matrix: calculate the sum m of all weight values in the weight matrix, multiply each weight value in the weight matrix by 1/m, and obtain the normalized weight matrix;
⑤以权重更新图片文件的RGB值:⑤Update the RGB value of the image file with the weight:
对每一个零点的权重矩阵,以各归一化后的权重矩阵的权重值乘以该权重值位置的像素值,并求和,作为该零点的新的像素值,以该种权重更新方法对RGB值分别更新,得到更新后的三色像素值;For the weight matrix of each zero point, the weight value of each normalized weight matrix is multiplied by the pixel value of the weight value position, and the sum is used as the new pixel value of the zero point. The RGB values are updated respectively to obtain the updated three-color pixel values;
⑥存储更新RGB值后的图片文件,作为新的样本。⑥ Store the image file after updating the RGB value as a new sample.
(3)将经过噪声耦合、高斯模糊的样本和原样本合并为新的样本集,用于深度学习网络模型的训练。(3) Combine the noise coupling, Gaussian blurred samples and the original samples into a new sample set for training the deep learning network model.
参照附图2所示,本发明还提出一种基于噪声扰动和高斯模糊的数据增强装置,包括:噪声耦合模块,高斯模糊模块和样本生成模块。Referring to FIG. 2, the present invention also proposes a data enhancement device based on noise disturbance and Gaussian blurring, including: a noise coupling module, a Gaussian blurring module and a sample generation module.
其中噪声耦合模块包括数据输入模块、三维数据生成模块、数据耦合模块和图片转换模块。The noise coupling module includes a data input module, a three-dimensional data generation module, a data coupling module and a picture conversion module.
数据输入模块,输入采集前端采集到的局部放电原始数据;Data input module, input the original partial discharge data collected by the front-end;
三维数据生成模块,将输入数据转成以相位为x轴,周期为y轴,幅值为z轴的三维数据;并且按照雷达噪声、手机噪声、微波硫灯干扰的数据特征,分别生成各噪声对应的三维数据;The three-dimensional data generation module converts the input data into three-dimensional data with the phase as the x-axis, the period as the y-axis, and the amplitude as the z-axis; and according to the data characteristics of radar noise, mobile phone noise, and microwave sulfur lamp interference, each noise is generated separately. Corresponding 3D data;
数据耦合模块,对输入数据和噪声数据进行累加,得到噪声耦合后的数据;The data coupling module accumulates the input data and the noise data to obtain the data after noise coupling;
图片转换模块,将噪声耦合后的数据转换成图片文件。The image conversion module converts the noise-coupled data into image files.
高斯模糊模块包括:The Gaussian Blur module includes:
像素提取模块,输入噪声耦合后的图片文件,提取输入图片文件的RGB(Red,Green, Blue)值;Pixel extraction module, input the image file after noise coupling, and extract the RGB (Red, Green, Blue) value of the input image file;
权重矩阵生成模块,分别取图片文件中图片的中心点为零点,绘制横纵坐标轴;以二维高斯分布函数计算图片中各像素点的权重,形成权重矩阵,并进一步得到归一化的权重矩阵。The weight matrix generation module takes the center point of the picture in the picture file as the zero point, and draws the horizontal and vertical axes; calculates the weight of each pixel in the picture with a two-dimensional Gaussian distribution function, forms a weight matrix, and further obtains the normalized weight matrix.
权值更新模块,以权重更新图片文件的RGB值;The weight update module updates the RGB value of the image file with the weight;
对每一个零点的权重矩阵,以各归一化后的权重矩阵的权重值乘以该权重值位置的像素值,并求和,作为该零点的新的像素值,以该种权重更新方法对RGB值分别更新,得到更新后的三色像素值。For the weight matrix of each zero point, the weight value of each normalized weight matrix is multiplied by the pixel value of the weight value position, and the sum is used as the new pixel value of the zero point. The RGB values are updated respectively to obtain the updated three-color pixel values.
文件生成模块,存储更新RGB值后的图片文件,作为新的样本。The file generation module stores the image file after updating the RGB value as a new sample.
样本生成模块,合并经过噪声耦合、高斯模糊的样本和原样本,用于深度学习网络模型的训练。The sample generation module combines noise-coupling, Gaussian blurred samples and original samples for training of deep learning network models.
实施例Example
以下参照图3-5给出实际应用数据进一步说明本发明所述的方法。The method of the present invention is further described below with reference to FIGS. 3-5 given practical application data.
本发明所述的针对局部放大图谱的基于噪声扰动和高斯模糊的数据增强方法,包括:The data enhancement method based on noise disturbance and Gaussian blurring for a partially enlarged atlas according to the present invention includes:
采集一条局放原始数据,其未经处理的三维图像如图3所示。A piece of PD raw data is collected, and its unprocessed 3D image is shown in Figure 3.
对原始样本数据分别叠加灯光、雷达、手机噪声干扰,并对叠加干扰后的数据做三维化处理,得到噪声叠加后的图像文件,如图4所示。The original sample data is superimposed on the noise interference of lights, radar and mobile phones, and the data after the superimposed interference is three-dimensionally processed to obtain the image file after the noise is superimposed, as shown in Figure 4.
取σ为1.5,对噪声叠加后的图像文件的RGB三色素分别计算高斯权重矩阵,并叠加高斯模糊,得到图片文件如图5。Take σ as 1.5, calculate the Gaussian weight matrix for the RGB three pigments of the image file after noise superposition, and superimpose the Gaussian blur to obtain the image file as shown in Figure 5.
将高斯模糊后的图片文件和原始文件共同纳入样本库中,完成一次基于噪声叠加和高斯模糊的数据增强。Combine the Gaussian blurred image file and the original file into the sample library to complete a data enhancement based on noise superposition and Gaussian blurring.
可见,与直接使用样本作为训练数据相比,本发明可以以较少的计算量生成大量带标签样本,并在原有数据样本的基础上考虑了现场的常见干扰,能够解决局部放电图谱识别中数据不足的方法,通过以一张图谱生成多张图谱的方式,增加样本数量,避免过拟合。It can be seen that, compared with directly using samples as training data, the present invention can generate a large number of labeled samples with less calculation amount, and on the basis of the original data samples, common disturbances in the field are considered, and the data in the identification of partial discharge maps can be solved. The insufficient method, by generating multiple maps from one map, increases the number of samples and avoids overfitting.
申请人结合说明书附图对本发明的实施例做了详细的说明与描述,但是本领域技术人员应该理解,以上实施例仅为本发明的优选实施方案,详尽的说明只是为了帮助读者更好地理解本发明精神,而并非对本发明保护范围的限制,相反,任何基于本发明的发明精神所作的任何改进或修饰都应当落在本发明的保护范围之内。The applicant has described and described the embodiments of the present invention in detail with reference to the accompanying drawings, but those skilled in the art should understand that the above embodiments are only preferred embodiments of the present invention, and the detailed description is only to help readers better understand The spirit of the present invention is not intended to limit the protection scope of the present invention. On the contrary, any improvement or modification made based on the spirit of the present invention should fall within the protection scope of the present invention.
Claims (23)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911074459.5A CN110796206B (en) | 2019-11-06 | 2019-11-06 | Data enhancement method and device for partial discharge map |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911074459.5A CN110796206B (en) | 2019-11-06 | 2019-11-06 | Data enhancement method and device for partial discharge map |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110796206A true CN110796206A (en) | 2020-02-14 |
CN110796206B CN110796206B (en) | 2022-08-30 |
Family
ID=69442926
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911074459.5A Active CN110796206B (en) | 2019-11-06 | 2019-11-06 | Data enhancement method and device for partial discharge map |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110796206B (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112101294A (en) * | 2020-09-29 | 2020-12-18 | 支付宝(杭州)信息技术有限公司 | Enhanced training method and device for image recognition model |
CN117975201A (en) * | 2024-03-29 | 2024-05-03 | 苏州元脑智能科技有限公司 | Training data generation method, device, computer equipment and storage medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130179100A1 (en) * | 2012-01-11 | 2013-07-11 | Utilx Corporation | System for analyzing and locating partial discharges |
CN108470139A (en) * | 2018-01-25 | 2018-08-31 | 天津大学 | A kind of small sample radar image human action sorting technique based on data enhancing |
CN110208660A (en) * | 2019-06-05 | 2019-09-06 | 国网江苏省电力有限公司电力科学研究院 | A kind of training method and device for power equipment shelf depreciation defect diagonsis |
CN110703057A (en) * | 2019-11-04 | 2020-01-17 | 国网山东省电力公司电力科学研究院 | Partial discharge diagnosis method of power equipment based on data augmentation and neural network |
-
2019
- 2019-11-06 CN CN201911074459.5A patent/CN110796206B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130179100A1 (en) * | 2012-01-11 | 2013-07-11 | Utilx Corporation | System for analyzing and locating partial discharges |
CN108470139A (en) * | 2018-01-25 | 2018-08-31 | 天津大学 | A kind of small sample radar image human action sorting technique based on data enhancing |
CN110208660A (en) * | 2019-06-05 | 2019-09-06 | 国网江苏省电力有限公司电力科学研究院 | A kind of training method and device for power equipment shelf depreciation defect diagonsis |
CN110703057A (en) * | 2019-11-04 | 2020-01-17 | 国网山东省电力公司电力科学研究院 | Partial discharge diagnosis method of power equipment based on data augmentation and neural network |
Non-Patent Citations (2)
Title |
---|
在路上DI蜗牛: "图像处理:高斯模糊,https://blog.csdn.net/qinghuaci666/article/details/81870277", 《网页文档》 * |
李沅箐: "基于深度学习的声目标识别方法研究", 《万方数据》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112101294A (en) * | 2020-09-29 | 2020-12-18 | 支付宝(杭州)信息技术有限公司 | Enhanced training method and device for image recognition model |
US11403487B2 (en) | 2020-09-29 | 2022-08-02 | Alipay (Hangzhou) Information Technology Co., Ltd. | Enhanced training method and apparatus for image recognition model |
CN117975201A (en) * | 2024-03-29 | 2024-05-03 | 苏州元脑智能科技有限公司 | Training data generation method, device, computer equipment and storage medium |
CN117975201B (en) * | 2024-03-29 | 2024-06-25 | 苏州元脑智能科技有限公司 | Training data generation method, device, computer equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN110796206B (en) | 2022-08-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108520504A (en) | An End-to-End Blind Restoration Method for Blurred Images Based on Generative Adversarial Networks | |
CN109191392B (en) | Image super-resolution reconstruction method driven by semantic segmentation | |
CN108537191B (en) | A three-dimensional face recognition method based on structured light camera | |
CN110543878A (en) | A Neural Network-Based Recognition Method of Pointer Meter Readings | |
CN107680077A (en) | A kind of non-reference picture quality appraisement method based on multistage Gradient Features | |
CN111260591B (en) | Image self-adaptive denoising method based on attention mechanism | |
CN114529946A (en) | Pedestrian re-identification method, device, equipment and storage medium based on self-supervision learning | |
CN110796206B (en) | Data enhancement method and device for partial discharge map | |
CN118211268A (en) | Heterogeneous federal learning privacy protection method and system based on diffusion model | |
CN114972323B (en) | Defect picture generation method and system for model training | |
CN117576755A (en) | A hyperspectral face fusion and recognition method, electronic device and storage medium | |
CN114581318A (en) | Method and system for image enhancement with low illumination | |
CN114331937A (en) | Multi-source image fusion method under low light conditions based on feedback iterative adjustment | |
CN115760630A (en) | Low-illumination image enhancement method | |
CN110717960A (en) | A method for generating remote sensing image samples of construction waste | |
CN116363011A (en) | Multi-branch low-illumination image enhancement method based on frequency domain frequency division | |
CN105791635A (en) | GPU-based enhanced video denoising method and apparatus | |
LU500193B1 (en) | Low-illumination image enhancement method and system based on multi-expression fusion | |
CN105426847A (en) | Nonlinear enhancing method for low-quality natural light iris images | |
CN112802009B (en) | Similarity calculation method and device for product detection data set | |
CN111127587B (en) | Reference-free image quality map generation method based on countermeasure generation network | |
CN115965555B (en) | Noise estimation image enhancement method based on transducer network | |
CN117635592A (en) | Denim defect detection method and device based on transform and discrete Fourier transform | |
CN118397391A (en) | Training method and device for deep learning model, electronic equipment and storage medium | |
CN106875101B (en) | Energy management system control method and control device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |