WO2022036867A1 - 变电站三维模型重构方法 - Google Patents

变电站三维模型重构方法 Download PDF

Info

Publication number
WO2022036867A1
WO2022036867A1 PCT/CN2020/125909 CN2020125909W WO2022036867A1 WO 2022036867 A1 WO2022036867 A1 WO 2022036867A1 CN 2020125909 W CN2020125909 W CN 2020125909W WO 2022036867 A1 WO2022036867 A1 WO 2022036867A1
Authority
WO
WIPO (PCT)
Prior art keywords
substation
training set
neural network
deep convolutional
equipment
Prior art date
Application number
PCT/CN2020/125909
Other languages
English (en)
French (fr)
Inventor
邝振星
华耀
温爱辉
邱健文
朱红涛
林孝斌
欧冠华
李朝阳
罗欣礼
李存海
赖家文
何荣伟
何文滨
Original Assignee
广东电网有限责任公司清远供电局
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 广东电网有限责任公司清远供电局 filed Critical 广东电网有限责任公司清远供电局
Publication of WO2022036867A1 publication Critical patent/WO2022036867A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects

Definitions

  • the present disclosure relates to the field of smart substations, for example, to a method for reconstructing a three-dimensional model of a substation.
  • the daily operation and maintenance pressure of substations is increasing.
  • the construction of the substation model can assist the safety inspection of the staff and the visual training of the operation and maintenance operators, which has very important practical significance for the safe operation of the substation.
  • the technical shortcomings of the above four mainstream modeling methods are: the traditional close-range photogrammetry method is inefficient; the VRML modeling method has high requirements for modelers; the geometric modeling method relies too much on the original design data of the substation; ground lidar construction The mold method is expensive and time-consuming.
  • the present disclosure provides a method for reconstructing a 3D model of a substation to establish a 3D model of complex substation equipment.
  • the modeling process is simple and convenient, and the 3D modeling of the substation is more efficient.
  • a method for reconstructing a three-dimensional model of a substation including:
  • the substation is modeled according to the trained deep convolutional auto-encoding neural network, and the three-dimensional model of the substation is obtained.
  • obtain and label original images of substation equipment and determine a transfer learning training set including:
  • process and extract the images in the basic training set and the transfer learning training set to obtain a device feature map including:
  • the extracted feature information is decoded by a decoder, and a feature map of the substation equipment is obtained.
  • the pixel size of the two-dimensional image is 256 ⁇ 256
  • the size of the two-dimensional morphological operator is 16 ⁇ 16
  • the decoder obtains a 256 ⁇ 256 device feature map after decoding the feature information.
  • use the basic training set to train a deep convolutional autoencoder neural network including:
  • the Hinge loss function is used as the loss function of the deep convolutional auto-encoding neural network, and the training samples are randomly divided into mini-batches of size 64-256;
  • the stochastic gradient descent algorithm is used to train different sub-models on different mini-batches to complete the training of the entire deep convolutional auto-encoding neural network model.
  • use the transfer learning training set to further train the deep convolutional autoencoder neural network including:
  • the transfer learning training set is used as the input of the trained deep convolutional auto-encoding neural network model, the reprojection consistency function is used as the loss function, and the deep convolutional auto-encoding neural network model is reversely fine-tuned in combination with the supervision training label. Training.
  • the loss function adopts the following function:
  • modeling the substation according to the trained deep convolutional self-encoding neural network to obtain a three-dimensional model of the substation including: collecting images of field equipment for the substation to be modeled, and inputting the collected images into the depth of the training completed.
  • the output obtains the three-dimensional model of the substation equipment;
  • the three-dimensional models of each equipment in the substation are spliced to complete the three-dimensional modeling of the substation.
  • acquiring data from a public database of three-dimensional models to form a basic training set including: constructing a basic training set by using the engineering equipment images in the PASCAL3D+ or Pix3D public data set, wherein the multi-angle two-dimensional images of the equipment are used as the input of the basic training set, Its corresponding 3D model is the output.
  • the optional method includes: obtaining the three-dimensional modeling requirements of the substation and the internal equipment information, external structure information and geographical location information of the substation, as the original image acquisition scheme formulation and final substation equipment splicing reference.
  • the method for reconstructing a three-dimensional model of a substation realizes the establishment of a three-dimensional model of a substation by establishing a deep convolutional self-encoding neural network; directly extracts feature information from multiple two-dimensional images of different angles, and utilizes the feature The information directly generates the corresponding 3D model.
  • the established neural network is trained by using the database that has been published in the field of computer vision, and then the neural network is trained by migration learning combined with the field information of the substation to be modeled, which greatly shortens the training of the neural network. Time, improve the modeling efficiency and accuracy of the substation, and make the 3D modeling of the substation faster and more efficient.
  • FIG. 1 is a schematic flowchart of a method for reconstructing a three-dimensional model of a substation according to Embodiment 1 of the present invention
  • FIG. 2 is a schematic flowchart of a method for reconstructing a three-dimensional model of a substation according to Embodiment 2 of the present invention
  • FIG. 3 is a schematic diagram of a method for reconstructing a three-dimensional model of a substation provided by the present invention.
  • the modeling method of the related art has the following technical disadvantages: 1) The traditional close-range photogrammetry data collection can only be collected point by point, and cannot be obtained quickly, and the modeling work needs to deal with massive data and huge amount of data. It is not suitable for data collection of substations with small area, so there are disadvantages such as large workload, difficult point collection, low efficiency, and complicated procedures; 2) VRML modeling method adopts cube, cone, cylinder, sphere The combined construction of equipment models will cause the established substation model to lack realism, poor model accuracy and low precision, and this method requires operators to be familiar with VRML programming and the structure of electrical equipment, which requires high personnel quality and low modeling efficiency.
  • FIG. 1 is a schematic flowchart of a method for reconstructing a three-dimensional model of a substation according to Embodiment 1 of the present invention.
  • the model reconstruction method can be used for the three-dimensional modeling of the substation and the electrical equipment in the station, as shown in Figure 1, the method may include the following steps:
  • the public database data can be a general database in the field of computer vision.
  • This kind of database has a large amount of data, a variety of types, and high quality, which can well make up for the shortcomings of the small amount of data in the three-dimensional model of the substation equipment;
  • the basic training set is in After constructing a deep convolutional autoencoder neural network, the dataset for preliminary training of the neural network.
  • the basic training set can be constructed using the engineering equipment images in the public data set.
  • the multi-angle 2D images of the equipment can be used as the input of the basic training set, and the corresponding 3D model is the output.
  • the substation equipment is the equipment in the substation to be modeled;
  • the original image of the substation equipment is a multi-angle image of the equipment; marking the image includes: marking the acquired image and selecting the orthographic image therein as a follow-up image
  • the supervised training label of ; the transfer learning training set is the dataset used for further transfer learning training of the neural network.
  • a reasonable original image acquisition scheme can be determined, and then multi-angle photography of the substation equipment is carried out by using photography equipment such as drones and professional digital cameras, and then the acquired RGB format 2 Dimensional images are labeled, and the orthographic images are selected as the subsequent supervised training labels, and the other images constitute the transfer learning training set.
  • the transfer learning training set helps to perform customized transfer learning on the characteristics of the 3D model of the substation equipment in the deep convolutional neural network, which can further improve the accuracy of the generated 3D model.
  • the images in the basic training set and the transfer learning training set are cropped, sized and normalized, and then the preprocessed two-dimensional images are filtered by using multiple groups of different types of morphological operators, and finally , and extract the feature information of different dimensions of the substation equipment respectively to decode, and obtain the feature map of the substation equipment.
  • the pixel size of the two-dimensional image may be uniformly 256 ⁇ 256
  • the size of the two-dimensional morphological operator may be 16 ⁇ 16
  • the decoder obtains a 256 ⁇ 256 device feature map after decoding the feature information.
  • the structure of the deep convolutional self-encoding neural network can be: one input layer, six convolutional layers, one fully connected layer, five deconvolutional layers and one output layer, where the input layer can be a 256 ⁇ 256 feature map,
  • the convolution filter size of each convolutional layer is 64 ⁇ 64, 32 ⁇ 32, 16 ⁇ 16, 8 ⁇ 8, 4 ⁇ 4, 1 ⁇ 1, and the number of filters of each convolutional layer can be 64, 64, 128 , 128, 256, 512.
  • the fully connected layer node is 512.
  • the filter size of each 3D deconvolution layer is 8 ⁇ 8 ⁇ 8, 16 ⁇ 16 ⁇ 16, 32 ⁇ 32 ⁇ 32, 64 ⁇ 64 ⁇ 64, 128 ⁇ 128 ⁇ 128, and the number of filters in each layer is 64, 32, 32, 16, 1.
  • the output is 128 ⁇ 128 ⁇ 128 3D voxel data.
  • an appropriate loss function can be selected for the deep convolutional auto-encoding neural network, and then the training samples in the basic training set can be divided into mini-batches of a certain size, and then different sub-models can be trained using the stochastic gradient descent algorithm on different mini-batches. , and finally complete the training of the entire deep convolutional auto-encoding neural network model.
  • the transfer learning training set is used as the input of the trained deep convolutional auto-encoding neural network model; the re-projection consistency function is used as the loss function, and combined with the supervised training label data, the deep convolutional auto-encoding neural network model is reversely fine-tuned and trained.
  • the field equipment images are collected for the substation to be modeled, and the collected images are preprocessed and input into the trained morphological-deep convolutional auto-encoding neural network, and the three-dimensional model of each substation equipment is obtained as output. Then, the three-dimensional models of the equipment of the substation are spliced, and finally the three-dimensional modeling of the substation is completed.
  • the method for reconstructing a 3D model of a substation proposed in the present disclosure can directly extract feature information from multiple 2D images with different angles, and use the feature information to directly generate a corresponding 3D model.
  • the established database and related research results combined with the on-site information of the substation to be modeled, carried out targeted transfer learning, which greatly shortened the training time of the neural network, improved the modeling efficiency and accuracy of the substation, and reduced the sampling work in the modeling. It makes the 3D modeling of substations faster and more efficient, and has a wider range of applications.
  • FIG. 2 is another method for reconstructing a three-dimensional model of a substation provided by an embodiment of the present invention, and the method includes:
  • the basic training set can be constructed mainly through the engineering equipment images in the public data sets such as PASCAL3D+ and Pix3D. Among them, the multi-angle 2D images of the equipment are marked as the independent variables of the basic training set, and the corresponding 3D model voxel data are marked as dependent variable.
  • a reasonable original image acquisition scheme is formulated. Then, a drone equipped with a high-definition camera can be used to take multi-angle photography of the substation equipment one by one, and the obtained two-dimensional images in RGB format can be marked, and the orthographic image is selected as the follow-up supervision training label data.
  • the images constitute the transfer learning training dataset.
  • a morphological operator feature extractor can be constructed to perform multi-dimensional feature extraction on the acquired image data, including the following steps: positioning, cropping, scaling and other operations on the two-dimensional RGB format image data obtained from the Internet and collected in the field, so as to unify the image size is 256 ⁇ 256, and then normalized; four groups of morphological operators with different kernels with a size of 16 ⁇ 16 are constructed to segment and filter the preprocessed two-dimensional image, filter the irrelevant background information of the image, and then separately Extract feature information of four different dimensions such as contour, depth, surface texture and surface normal of substation equipment; use the decoder to decode the contour, depth, surface texture and surface normal feature information of substation equipment, and obtain four sets of 256 ⁇ 256 characteristic map of substation equipment.
  • a deep convolutional autoencoder neural network can include: one input layer, six convolutional layers, one fully connected layer, five deconvolutional layers, and one output layer.
  • the input layer can be a 256 ⁇ 256 feature map
  • the output can be 128 ⁇ 128 ⁇ 128 three-dimensional voxel data.
  • Using the basic training set to train a deep convolutional auto-encoding neural network includes: using the Hinge loss function as the loss function of the deep convolutional auto-encoding neural network, and randomly dividing the training samples into mini-batches with a size of 64-256; The stochastic gradient descent algorithm is used to train different sub-models on different mini-batches to complete the training of the entire deep convolutional auto-encoding neural network model.
  • Using the transfer learning training set to further train the deep convolutional auto-encoding neural network including: using the transfer learning training set as the input of the trained deep convolutional auto-encoding neural network model; using the reprojection consistency function as the loss function ; Perform reverse fine-tuning training on the deep convolutional auto-encoding neural network model in combination with the supervised training label data.
  • the functional expression of the loss function is:
  • the image of the field equipment can be collected, and the collected image can be input into the trained morphological-deep convolutional auto-encoding neural network, and the three-dimensional model of each equipment in the substation can be obtained as output.
  • the three-dimensional models of the equipment of the substation are spliced, and the three-dimensional modeling of the substation is finally completed.
  • the method for reconstructing a 3D model of a substation performs multi-angle image acquisition of the substation according to information such as modeling requirements, internal equipment, external structure, and geographic location, and then extracts feature information, and uses the feature information to generate a corresponding 3D model.
  • the image acquisition plan is formulated using information such as the three-dimensional modeling requirements of the substation, and the information such as the three-dimensional modeling requirements of the substation is used as the final substation equipment model splicing reference, which increases the accuracy of image acquisition and greatly improves the The efficiency of model splicing greatly improves the modeling efficiency of substations, and the 3D modeling of substations is more accurate, fast and efficient.
  • FIG. 3 is a schematic diagram of another three-dimensional model reconstruction process of a substation provided by an embodiment of the present invention, and the method can be divided into offline training and online application.
  • two-dimensional images of substation equipment collected by UAV and images of general engineering equipment in public data sets such as PASCAL3D+ and Pix3D are preprocessed in the morphological feature extractor to obtain basic training sets and transfer learning training.
  • the two-dimensional image of the substation to be modeled acquired by the UAV is preprocessed, the feature map of the equipment is extracted by the morphological feature extractor, and the feature map of the equipment of the substation to be modeled is input into the deep convolutional auto-encoding neural network.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Business, Economics & Management (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Geometry (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Economics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Public Health (AREA)
  • Human Resources & Organizations (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Graphics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Computer Hardware Design (AREA)
  • Water Supply & Treatment (AREA)
  • Medical Informatics (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Image Analysis (AREA)

Abstract

一种变电站三维模型重构方法，所述方法包括：获取三维模型公开数据库数据构成基础训练集（110）；获取变电站设备原始图像进行标记并确定迁移学习训练集（120）；对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图（130），基于所述设备特征图构造深度卷积自编码神经网络（140）；使用所述基础训练集训练深度卷积自编码神经网络（150），使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练（160）；根据训练完成的深度卷积自编码神经网络对变电站进行建模，得到变电站三维模型（170）。

Description

变电站三维模型重构方法
本申请要求在2020年08月21日提交中国专利局、申请号为202010850897.2的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本公开涉及智能变电站领域，例如涉及一种变电站三维模型重构方法。
背景技术
随着科技的进步，用电设备的不断增加，变电站的日常运维压力越来越大。构建变电站模型可以辅助工作人员的安全巡视和运维操作人员的可视化培训，对变电站的安全运行具有非常重要的现实意义。
国内主流的变电站建模方法有以下四种：近景摄影测法，通过专业摄影工具获取变电站设备的资料，配合专业后期软件实现变电站的建模；虚拟现实建模语言(Virtual Reality Modeling Language，VRML)建模，通过编程语言以圆柱、立方体等规则立体图形为原始对象构造电气设备模型；几何造型模型，依据变电站图片和设备图纸，采用立体图形建立电气设备的三维模型；地面激光雷达建模法，利用大量的扫描数据和图片数据融合，再对得到的数据进行分类重构得到模型。
但上述四种主流建模方法的技术缺点是：传统的近景摄影测量法效率低；VRML建模方法模型对建模人员要求高；几何造型建模方法过于依赖变电站原始设计资料；地面激光雷达建模法成本高，耗时长。
发明内容
为解决相关技术问题，本公开提供一种基于变电站三维模型重构方法，对复杂的变电站设备进行三维模型的建立，建模过程简洁方便，使得变电站的三维建模更高效。
提供一种变电站三维模型重构方法，包括：
获取三维模型公开数据库数据构成基础训练集；
获取变电站设备原始图像并进行标记并确定迁移学习训练集；
对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图，基于所述设备特征图构造深度卷积自编码神经网络；
使用所述基础训练集训练深度卷积自编码神经网络，并使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练；
根据训练完成的深度卷积自编码神经网络对变电站进行建模，得到变电站三维模型。
可选地，获取变电站设备原始图像并进行标记并确定迁移学习训练集，包括：
获取变电站设备多角度摄影图像，对获取的图像进行标记，并选取所标记的图像中的正投影图像作为监督训练标签，其他图像构成迁移学习训练集。
可选地，对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图，包括：
对获取的所有二维图像数据进行裁剪、统一尺寸以及归一化处理；
使用多组不同类型形态学算子对预处理后的二维图像进行滤波，分别提取变电站设备不同维度的特征信息；
使用解码器对所提取的特征信息进行解码，得到变电站设备特征图。
可选地，所述二维图像的像素大小为256×256，所述二维形态学算子尺寸为16×16，所述解码器对特征信息解码后获得256×256的设备特征图。
可选地，使用所述基础训练集训练深度卷积自编码神经网络，包括：
采用Hinge损失函数作为深度卷积自动编码神经网络的损失函数，将训练样本随机划分为大小为64～256的mini-batch；
在不同的mini-batch上利用随机梯度下降算法训练不同的子模型，完成对整个深度卷积自动编码神经网络模型的训练。
可选地，使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练，包括：
以所述迁移学习训练集作为训练好的深度卷积自动编码神经网络模型输入，以重投影一致性函数为损失函数，结合所述监督训练标签对深度卷积自动编码神经网络模型进行反向微调训练。
可选地，所述损失函数采用如下函数：
Figure PCTCN2020125909-appb-000001
可选地，根据训练完成的深度卷积自编码神经网络对变电站进行建模，得 到变电站三维模型，包括：针对待建模变电站进行现场设备图像采集，并将采集所得的图像输入训练完成的深度卷积自编码神经网络中，输出得到变电站设备的三维模型；
根据所获取的变电站基本建模信息，对变电站各设备的三维模型进行拼接，完成对变电站的三维建模。
可选地，获取三维模型公开数据库数据构成基础训练集，包括：通过PASCAL3D+或Pix3D公开数据集中的工程设备类图像构建基础训练集，其中以设备的多角度二维图像作为基础训练集的输入，其对应的三维模型为输出。
可选的方法包括：获取变电站三维建模需求以及变电站的内部设备信息、外部结构信息以及所处地理位置信息，作为原始图像采集方案制定以及最终变电设备拼接参考。
本发明实施例提供的变电站三维模型重构方法，通过建立深度卷积自编码神经网络，实现对变电站三维模型的建立；从多张不同角度的二维图像中直接提取特征信息，并利用该特征信息直接生成相应的三维模型，该过程中利用计算机视觉领域已经公开的数据库对建立的神经网络进行训练，再结合待建模变电站现场信息对神经网络进行迁移学习训练，大大缩短了神经网络的训练时间，提高了变电站的建模效率与精度，使得变电站的三维建模更快速高效。
附图说明
下面将对本发明实施例描述中所需要使用的附图作简单的介绍。
图1为本发明实施例一提供的一种变电站三维模型重构方法的流程示意图；
图2为本发明实施例二提供的一种变电站三维模型重构方法的流程示意图；
图3为本发明提供的一种变电站三维模型重构方法的示意图。
具体实施方式
下面将结合附图对本发明实施例的技术方案进行描述，所描述的实施例仅仅是本公开的一部分实施例，而不是全部的实施例。
正如背景技术中所述，相关技术的建模方法存在如下技术缺点：1)传统的近景摄影测量法数据采集只能逐点采集，不能快速获取，且建模工作中需要处理海量的数据以及庞大的计算量，不适合区域面积小的变电站的数据采集，因此存在工作量大、采点困难、效率低、程序复杂等缺点；2)VRML建模方法采 用立方体、圆锥体、圆柱体、球体的组合构建设备模型，会造成建立的变电站模型缺乏真实感，模型精确度差、精度低，且该方法需要作业人员熟悉掌握VRML编程和电气设备的构造，对人员素质要求高，建模效率较低；3)几何造型建模方法极度依赖变电站原始设计资料，但实际运行中变电站原始设计资料完整度不高，且局部改建造成的电气设备数量变化以及带电设备长期运行造成的电气设备尺寸变化均很难单从设计资料获取，这也造成该方法建模精度较低，耗时长，效率低；4)地面激光雷达建模则极度依赖于专业设备以及专业软件，设备成本较高，且对海量的三维点云数据的处理工作量大，计算负荷重，耗时长，效率偏低，且最后生成的模型无法形象反映设备表面纹理，精度仍有所欠缺。
实施例一
图1为本发明实施例一提供的变电站三维模型重构方法的流程示意图。该模型重构方法可用于对变电站及站内电气设备的三维建模，参见图1所示，该方法可以包括如下步骤：
110：获取三维模型公开数据库数据构成基础训练集。
其中，公开数据库数据可以是计算机视觉领域通用数据库，这类数据库数据量较大、种类较多、质量较高，能很好弥补变电设备三维模型数据量较少的缺点；基础训练集是在构造深度卷积自编码神经网络之后，对该神经网络进行初步训练的数据集。
在该过程中，可以使用公开数据集中的工程设备类图像构建基础训练集，在该数据集中，设备的多角度二维图像可以作为基础训练集的输入，其对应的三维模型为输出。
120：获取变电站设备原始图像进行标记并确定迁移学习训练集。
其中，所述变电站设备是待建模的变电站中的设备；所述变电站设备原始图像为设备的多角度图像；对图像进行标记包括：对获取的图像进行标记和选取其中的正投影图像作为后续的监督训练标签；迁移学习训练集是用于神经网络进一步迁移学习训练的数据集。
根据所建模变电站的外部结构和地理位置，可以确定合理的原始图像采集方案，再通过无人机、专业数码相机等摄影设备对变电站设备进行多角度摄影，然后对获取的所获取RGB格式二维图像进行标记，并选取其中的正投影图像作为后续的监督训练标签，其他图像则构成迁移学习训练集。迁移学习训练集有助于在深度卷积神经网络对变电设备三维模型特征进行定制化迁移学习，可进一步提高生成三维模型的精度。
130：对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图。
例如对所述基础训练集和所述迁移学习训练集中的图像进行裁剪、统一尺寸和归一化处理，然后，使用多组不同类型形态学算子对预处理后的二维图像进行滤波，最后，分别提取变电站设备不同维度的特征信息进行解码，得到变电设备特征图。可选地，二维图像像素大小可以统一为256×256，所述二维形态学算子尺寸可以为16×16，解码器对特征信息解码后获得256×256的设备特征图。
140：基于所述设备特征图构造深度卷积自编码神经网络。
深度卷积自编码神经网络的结构可以为：一个输入层、六个卷积层、一个全连接层、五个反卷积层和一个输出层，其中输入层可以为256×256的特征图，各卷积层的卷积滤波器尺寸为64×64、32×32、16×16、8×8、4×4、1×1，各卷积层的滤波器数目可以为64、64、128、128、256、512。全连接层结点为512。各三维反卷积层的滤波器尺寸为8×8×8、16×16×16、32×32×32、64×64×64、128×128×128，其中各层滤波器数目为64、32、32、16、1。输出为128×128×128的三维体素数据。
150：使用所述基础训练集训练深度卷积自编码神经网络。
首先可以为深度卷积自动编码神经网络选用适当的损失函数，再将基础训练集中的训练样本划分成一定大小的mini-batch，然后在不同mini-batch上利用随机梯度下降算法训练不同的子模型，最终完成对整个深度卷积自动编码神经网络模型的训练。
160：使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练。
以迁移学习训练集为训练好的深度卷积自动编码神经网络模型输入；以重投影一致性函数为损失函数，再结合监督训练标签数据对深度卷积自动编码神经网络模型进行反向微调训练。
170：根据训练完成的深度卷积自编码神经网络对变电站进行建模，得到变电站三维模型。
针对待建模变电站进行现场设备图像采集，并将采集所得图像预处理后输入训练完毕的形态学-深度卷积自编码神经网络中，输出得到各变电站设备的三维模型。然后，对变电站各设备的三维模型进行拼接，最终完成对变电站的三维建模。
本公开所提出的变电站三维模型重构方法能够从多张不同角度的二维图像中直接提取特征信息，并利用该特征信息直接生成相应的三维模型，建模过程 中充分利用了计算机视觉领域已建立的数据库及相关研究成果，并结合待建模变电站现场信息进行了针对性迁移学习，大大缩短了神经网络的训练时间，提高了变电站的建模效率与精度，减少了建模中采样的工作量，使得变电站的三维建模更快速高效，适用范围更广。
实施例二
参考图2，图2是本发明实施例提供的另一种变电站三维模型重构方法，该方法包括：
210：获取变电站三维建模需求以及变电站的内部设备信息、外部结构信息以及所处地理位置信息。
获取变电站三维建模需求以及变电站的内部设备信息、外部结构信息以及所处地理位置等相关信息，主要用于后续原始图像采集方案制定以及最终变电设备拼接参考。
220：获取三维模型公开数据库数据构成基础训练集。
可以主要通过PASCAL3D+、Pix3D等公开数据集中的工程设备类图像构建基础训练集，其中，并把设备的多角度二维图像标记为基础训练集的自变量，其对应的三维模型体素数据标记为因变量。
230：获取变电站设备原始图像进行标记并确定迁移学习训练集。
根据步骤210所得变电站三维建模需求以及变电站的内部设备信息、外部结构信息以及所处地理位置等相关信息，制定合理的原始图像采集方案。进而可以利用配备有高清摄像头的无人机对变电站设备逐一进行多角度的摄影，并对获得的RGB格式的二维图像进行标记，选取其中的正投影图像作为后续的监督训练标签数据，其他角度的图像构成迁移学习训练数据集。
240：对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图。
可以构造形态学算子特征提取器对所获取图像数据进行多维度特征提取，包括以下步骤：对互联网获取以及实地采集的二维RGB格式图像数据进行定位、裁剪、缩放等操作，将图像尺寸统一为256×256，然后进行归一化处理；构建尺寸为16×16的四组不同内核的形态学算子对预处理后的二维图像进行分割滤波操作，过滤图像的无关背景信息，然后分别提取变电站设备的外形轮廓、深度、表面纹理以及表面法向等四种不同维度的特征信息；使用解码器对变电站设备的外形轮廓、深度、表面纹理以及表面法向特征信息进行解码，得到四组256×256的变电设备特征图。
250：基于所述设备特征图构造深度卷积自编码神经网络。
深度卷积自编码神经网络可以包括：一个输入层、六个卷积层、一个全连接层、五个反卷积层和一个输出层。其中输入层可以为256×256的特征图，输出可以为128×128×128的三维体素数据。
260：使用所述基础训练集训练深度卷积自编码神经网络。
使用所述基础训练集训练深度卷积自编码神经网络，包括：采用Hinge损失函数作为深度卷积自动编码神经网络的损失函数，将训练样本随机划分为大小为64～256的mini-batch；在不同的mini-batch上利用随机梯度下降算法训练不同的子模型，完成对整个深度卷积自动编码神经网络模型的训练。
270：使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练。
使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练，包括：以所述迁移学习训练集作为训练好的深度卷积自动编码神经网络模型输入；以重投影一致性函数为损失函数；结合所述监督训练标签数据对深度卷积自动编码神经网络模型进行反向微调训练。其中，所述损失函数的函数表达式为：
Figure PCTCN2020125909-appb-000002
280：根据训练完成的深度卷积自编码神经网络对变电站设备进行建模，得到变电站各设备三维模型。
针对待建模变电站可以进行现场设备图像采集，并将采集所得图像输入训练完成的形态学-深度卷积自编码神经网络中，输出得到变电站中各个设备的三维模型。
290：根据变电站三维建模需求等信息对变电站各设备的三维模型进行拼接，得到变电站的三维模型。
结合步骤210所获取的变电站三维建模需求和变电站的内部设备信息、外部结构信息以及所处地理位置信息，对变电站各设备的三维模型进行拼接，最终完成对变电站的三维建模。
本公开所提出的变电站三维模型重构方法根据建模需求、内部设备、外部结构以及所处地理位置等信息进行变电站多角度图像采集进而提取特征信息，利用该特征信息生成相应三维模型。在建模过程中，利用变电站的三维建模需求等信息制定图像采集方案，并将变电站的三维建模需求等信息作为最终的变电设备模型拼接参考，增加了图像采集的精度，大大提高了模型拼接的效率， 从而使得变电站的建模效率大幅提高，变电站的三维建模更加精确、快速、高效。
参考图3，图3是本发明实施例提供的另一种变电站三维模型重构过程的示意图，该方法可以分为离线训练和在线应用。
在离线训练阶段：对无人机采集变电站设备的二维图像以及对PASCAL3D+、Pix3D等公开数据集中的通用工程设备类图像，在形态学特征提取器进行预处理，得到基础训练集和迁移学习训练集；基于变电站设备特征图构造深度卷积自编码神经网络；基于随机梯度下降训练法，采用基础训练集对深度卷积自编码神经网络进行训练；采用迁移学习训练集对深度卷积自编码神经网络进行进一步训练。
在线应用阶段，将无人机获取的待建模变电站二维图像进行预处理，采用形态学形态学特征提取器提取设备特征图，将待建模变电站的设备特征图输入深度卷积自编码神经网络，得到各个变电设备三维模型；根据变电站三维建模需求等信息对变电站设备三维模型进行拼接，得到变电站三维模型。

Claims (10)

  1. 一种变电站三维模型重构方法，包括：
    获取三维模型公开数据库数据构成基础训练集；
    获取变电站设备原始图像并进行标记并确定迁移学习训练集；
    对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图，基于所述设备特征图构造深度卷积自编码神经网络；
    使用所述基础训练集训练深度卷积自编码神经网络，并使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练；
    根据训练完成的深度卷积自编码神经网络对变电站进行建模，得到变电站三维模型。
  2. 根据权利要求1所述的方法，其中，获取变电站设备原始图像并进行标记并确定迁移学习训练集，包括：
    获取变电站设备多角度摄影图像，对获取的图像进行标记，并选取所标记的图像中的正投影图像作为监督训练标签，其他图像构成迁移学习训练集。
  3. 根据权利要求1所述的方法，其中，对所述基础训练集和所述迁移学习训练集中的图像进行处理并提取得到设备特征图，包括：
    对获取的所有二维图像数据进行裁剪、统一尺寸以及归一化处理；
    使用多组不同类型形态学算子对预处理后的二维图像进行滤波，分别提取变电站设备不同维度的特征信息；
    使用解码器对所提取的特征信息进行解码，得到所述变电站设备特征图。
  4. 根据权利要求3所述的方法，其中，所述二维图像的像素大小为256×256，二维形态学算子尺寸为16×16，所述解码器对特征信息解码后获得256×256的设备特征图。
  5. 根据权利要求2所述的方法，其中，使用所述基础训练集训练深度卷积自编码神经网络，包括：
    采用Hinge损失函数作为深度卷积自动编码神经网络的损失函数，将训练样本随机划分为大小为64～256的mini-batch；
    在不同的mini-batch上利用随机梯度下降算法训练不同的子模型，完成对整个深度卷积自动编码神经网络模型的训练。
  6. 根据权利要求2所述的方法，其中，使用所述迁移学习训练集对深度卷积自编码神经网络进一步训练，包括：
    以所述迁移学习训练集作为训练好的深度卷积自动编码神经网络模型输入，以重投影一致性函数为损失函数，结合所述监督训练标签对深度卷积自动编码神经网络模型进行反向微调训练。
  7. 根据权利要求6所述的方法，其中，所述损失函数采用如下函数：
    Figure PCTCN2020125909-appb-100001
  8. 根据权利要求6所述的方法，其中，根据训练完成的深度卷积自编码神经网络对变电站进行建模，得到变电站三维模型，包括：针对待建模变电站进行现场设备图像采集，并将采集所得的图像输入训练完成的深度卷积自编码神经网络中，输出得到变电站设备的三维模型；
    根据所获取的变电站基本建模信息，对变电站各设备的三维模型进行拼接，完成对变电站的三维建模。
  9. 根据权利要求1所述的方法，其中，获取三维模型公开数据库数据构成基础训练集，包括：通过PASCAL3D+或Pix3D公开数据集中的工程设备类图像构建基础训练集，其中以设备的多角度二维图像作为基础训练集的输入，其对应的三维模型为输出。
  10. 根据权利要求1所述的方法，还包括：获取变电站三维建模需求以及变电站的内部设备信息、外部结构信息以及所处地理位置信息，作为原始图像采集方案制定以及最终变电设备拼接参考。
PCT/CN2020/125909 2020-08-21 2020-11-02 变电站三维模型重构方法 WO2022036867A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010850897.2 2020-08-21
CN202010850897.2A CN111985161A (zh) 2020-08-21 2020-08-21 一种变电站三维模型重构方法

Publications (1)

Publication Number Publication Date
WO2022036867A1 true WO2022036867A1 (zh) 2022-02-24

Family

ID=73442433

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/125909 WO2022036867A1 (zh) 2020-08-21 2020-11-02 变电站三维模型重构方法

Country Status (2)

Country Link
CN (1) CN111985161A (zh)
WO (1) WO2022036867A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115082770A (zh) * 2022-07-04 2022-09-20 青岛科技大学 一种基于机器学习的图像中心线结构提取方法
CN115082695A (zh) * 2022-05-31 2022-09-20 中国科学院沈阳自动化研究所 一种基于改进Yolov5的变电站绝缘子串建模及检测方法
CN115953486A (zh) * 2022-12-30 2023-04-11 国网电力空间技术有限公司 一种直流t型直线塔部件巡检图像自动编码方法
CN116996926A (zh) * 2023-09-26 2023-11-03 香港中文大学（深圳） 一种数据驱动的基站功耗建模方法
CN117494356A (zh) * 2023-10-31 2024-02-02 成都建工第九建筑工程有限公司 一种基于bim技术的装配式施工方法

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112634215B (zh) * 2020-12-16 2022-06-07 华虹半导体(无锡)有限公司 一种提高多类型图片深度学习模型训练效率的方法
CN112733676A (zh) * 2020-12-31 2021-04-30 青岛海纳云科技控股有限公司 一种基于深度学习的电梯内垃圾检测识别方法
CN113074718B (zh) * 2021-04-27 2024-03-29 广东电网有限责任公司清远供电局 一种定位方法、装置、设备及存储介质
CN113284112B (zh) * 2021-05-27 2023-11-10 中国科学院国家空间科学中心 一种基于深度神经网络的熔滴图像轮廓提取方法及系统
CN113781541B (zh) * 2021-09-15 2024-03-26 平安科技（深圳）有限公司 基于神经网络的三维图像处理方法、装置及电子设备
CN115482248B (zh) * 2022-09-22 2023-12-08 推想医疗科技股份有限公司 图像分割方法、装置、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107122809A (zh) * 2017-04-24 2017-09-01 北京工业大学 基于图像自编码的神经网络特征学习方法
CN107451661A (zh) * 2017-06-29 2017-12-08 西安电子科技大学 一种基于虚拟图像数据集的神经网络迁移学习方法
CN108875076A (zh) * 2018-07-10 2018-11-23 重庆大学 一种基于Attention机制和卷积神经网络的快速商标图像检索方法
US20200012923A1 (en) * 2016-10-06 2020-01-09 Siemens Aktiengesellschaft Computer device for training a deep neural network
CN111127360A (zh) * 2019-12-20 2020-05-08 东南大学 一种基于自动编码器的灰度图像迁移学习方法

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107798349B (zh) * 2017-11-03 2020-07-14 合肥工业大学 一种基于深度稀疏自编码机的迁移学习方法
CN109543745B (zh) * 2018-11-20 2021-08-24 江南大学 基于条件对抗自编码网络的特征学习方法及图像识别方法
CN110544297B (zh) * 2019-08-06 2023-06-20 北京工业大学 一种单幅图像的三维模型重建方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200012923A1 (en) * 2016-10-06 2020-01-09 Siemens Aktiengesellschaft Computer device for training a deep neural network
CN107122809A (zh) * 2017-04-24 2017-09-01 北京工业大学 基于图像自编码的神经网络特征学习方法
CN107451661A (zh) * 2017-06-29 2017-12-08 西安电子科技大学 一种基于虚拟图像数据集的神经网络迁移学习方法
CN108875076A (zh) * 2018-07-10 2018-11-23 重庆大学 一种基于Attention机制和卷积神经网络的快速商标图像检索方法
CN111127360A (zh) * 2019-12-20 2020-05-08 东南大学 一种基于自动编码器的灰度图像迁移学习方法

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115082695A (zh) * 2022-05-31 2022-09-20 中国科学院沈阳自动化研究所 一种基于改进Yolov5的变电站绝缘子串建模及检测方法
CN115082770A (zh) * 2022-07-04 2022-09-20 青岛科技大学 一种基于机器学习的图像中心线结构提取方法
CN115082770B (zh) * 2022-07-04 2024-02-23 青岛科技大学 一种基于机器学习的图像中心线结构提取方法
CN115953486A (zh) * 2022-12-30 2023-04-11 国网电力空间技术有限公司 一种直流t型直线塔部件巡检图像自动编码方法
CN115953486B (zh) * 2022-12-30 2024-04-12 国网电力空间技术有限公司 一种直流t型直线塔部件巡检图像自动编码方法
CN116996926A (zh) * 2023-09-26 2023-11-03 香港中文大学（深圳） 一种数据驱动的基站功耗建模方法
CN116996926B (zh) * 2023-09-26 2023-11-28 香港中文大学（深圳） 一种数据驱动的基站功耗建模方法
CN117494356A (zh) * 2023-10-31 2024-02-02 成都建工第九建筑工程有限公司 一种基于bim技术的装配式施工方法

Also Published As

Publication number Publication date
CN111985161A (zh) 2020-11-24

Similar Documents

Publication Publication Date Title
WO2022036867A1 (zh) 变电站三维模型重构方法
WO2021115124A1 (zh) 一种耕地现场边云协同的三维重建方法
CN109410321A (zh) 基于卷积神经网络的三维重建方法
CN110765528A (zh) 一种基于虚拟仿真技术的三维重构变电站实现方法
CN111127538B (zh) 一种基于卷积循环编码-解码结构的多视影像三维重建方法
CN108537191A (zh) 一种基于结构光摄像头的三维人脸识别方法
CN103606151B (zh) 基于影像点云的大范围虚拟地理场景自动构建方法
CN104134234A (zh) 一种全自动的基于单幅图像的三维场景构建方法
CN111951368B (zh) 一种点云、体素和多视图融合的深度学习方法
CN110660125B (zh) 用于电力配网系统的三维建模装置
CN111640116B (zh) 基于深层卷积残差网络的航拍图建筑物分割方法及装置
CN105719277A (zh) 一种基于测绘与二维图像的变电站三维建模方法及系统
CN115082254A (zh) 一种变电站精益管控数字孪生系统
CN115471634B (zh) 一种城市绿植孪生的建模方法及装置
CN111598803B (zh) 一种基于变分辨率体素格网与稀疏卷积的点云滤波方法
CN109598782A (zh) 基于三维建模技术的建筑文物修复方法、存储介质
CN110717978B (zh) 基于单张图像的三维头部重建方法
CN113626917A (zh) 基于全景技术和bim模型的构件工程量算量方法及系统
Li et al. Deep-learning-based 3D reconstruction: a review and applications
CN112966698A (zh) 基于轻量化卷积网络的淡水鱼图像实时识别方法
CN116977586A (zh) 一种基于电网gis平台的实景三维建模方法
CN116071504A (zh) 一种面向高分辨率图像的多视图立体重建方法
CN115527082A (zh) 一种基于图像多重预处理的深度学习小目标检测方法
CN113192204B (zh) 单幅倾斜遥感图像中建筑物的三维重建方法
KR102440916B1 (ko) 데이터 결합을 통한 3차원 문화재의 데이터베이스 구축방법 및 시스템

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20950070

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20950070

Country of ref document: EP

Kind code of ref document: A1