技术领域Technical field
本发明涉及一种基于云计算的用户配电房设计生成方法，属于业扩工程技术领域。The invention relates to a cloud computing-based user distribution room design and generation method, and belongs to the technical field of industrial expansion engineering.
背景技术Background technique
高压业扩工程一般分为临时业扩、高压新装和高压扩容三种业扩类型，其中高压扩容需要现场收集用户端已安装的高压配电设备资料，目前工作流程是由客户经理沟通好需求后，由设计人员完成现场查看工作，收集用户配电房和设备的资料，主要采用拍照和配电房测量绘制草图方法。现场收资完成后，将绘制的草图在CAD软件中重新绘制，然后再根据用户业扩需求将需要安装的设备设计与现有收资的设计结合起来，形成新的配电房设计图纸。这份图纸由客户经理与用户确认后，设计人员再绘制施工图纸，交给实施人员完成设备安装工作。High-voltage expansion projects are generally divided into three types of expansion: temporary expansion, high-voltage new installation, and high-voltage expansion. High-voltage expansion requires on-site collection of high-voltage power distribution equipment data installed at the user end. The current workflow is to communicate the needs with the customer manager. , the designer completes the on-site inspection work and collects information on the user's distribution room and equipment, mainly using the methods of taking photos and measuring and drawing sketches of the distribution room. After the on-site collection is completed, the sketch will be redrawn in CAD software, and then the design of the equipment to be installed will be combined with the design of the existing collection to form a new power distribution room design drawing based on the user's industry expansion needs. After this drawing is confirmed by the customer manager and the user, the designer will draw the construction drawing and hand it over to the implementation personnel to complete the equipment installation work.
目前采用的业扩设计方法，在一切顺利的情况下，需要消耗约1周时间才能完成设计工作。对电网、设计、施工和用户本身来说，该流程消耗了太多时间，所以需要引进新技术缩短该工作流程。The current industry expansion design method takes about one week to complete the design work if everything goes well. This process consumes too much time for the grid, design, construction and users themselves, so new technologies need to be introduced to shorten the workflow.
发明内容Contents of the invention
针对现有技术的缺陷，本发明的目的在于提供一种通过激光深度传感器数据收集配电房三维信息以及图片，并利用卷积神经网络，获取绘图所需的参数，并将参数传输到CAD模块；最终输出实际可用的配电房设计图纸的能够有效缩短业扩设计时间的提高业扩设计效率的基于云计算的用户配电房设计生成方法。In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a method for collecting three-dimensional information and pictures of a power distribution room through laser depth sensor data, and using a convolutional neural network to obtain the parameters required for drawing, and transmit the parameters to the CAD module ; A cloud computing-based user distribution room design generation method that can effectively shorten the industrial expansion design time and improve the industrial expansion design efficiency by finally outputting actual usable distribution room design drawings.
为实现上述目的，本发明的技术方案为：In order to achieve the above objects, the technical solution of the present invention is:
一种基于云计算的用户配电房设计生成方法，A cloud computing-based user distribution room design and generation method,
其包括以下步骤：It includes the following steps:
第一步，建立采用为python编程语音定义的Web服务器和Web应用程序和深度计算框架的接口程序WSGI；The first step is to establish the interface program WSGI using the web server and web application defined for Python programming language and the deep computing framework;
第二步，通过在配电房现场的工作人员，利用激光深度相机收集配电房三维信息；In the second step, the staff at the power distribution room site use the laser depth camera to collect the three-dimensional information of the power distribution room;
第三步，通过无线网络将三维重建信息与设备的激光深度照片传输到云端服务器的WSGI；The third step is to transmit the 3D reconstruction information and the laser depth photos of the device to the WSGI of the cloud server through the wireless network;
第四步，通过云端的WSGI接口将数据传输到基于GPU集群的卷积神经网络推理inference模块，获取绘图所需的参数，并将参数传输到CAD模块的pythonAPI接口；The fourth step is to transmit the data to the convolutional neural network inference module based on the GPU cluster through the WSGI interface of the cloud, obtain the parameters required for drawing, and transmit the parameters to the pythonAPI interface of the CAD module;
第五步，使用一种云端加速计算架构来加快现场传感器收集数据的空间计算量，最终输出配电房现场收资的标准图纸。The fifth step is to use a cloud accelerated computing architecture to accelerate the spatial calculation of data collected by on-site sensors, and finally output standard drawings for on-site collection of power distribution rooms.
本发明通过激光深度传感器数据收集配电房三维信息以及图片，并利用卷积神经网络，获取绘图所需的参数，并将参数传输到CAD模块；最终输出实际可用的配电房设计图纸；能够有效缩短业扩设计时间，提高业扩设计效率。This invention collects three-dimensional information and pictures of the power distribution room through laser depth sensor data, and uses a convolutional neural network to obtain parameters required for drawing, and transmits the parameters to the CAD module; finally outputs actual usable power distribution room design drawings; can Effectively shorten the industrial expansion design time and improve the industrial expansion design efficiency.
应用本发明的方案，能够将现有的高压业扩设计流程，从原有的5个工作日，降低到0.5个工作日，实现工作效率的革命性提升。本发明能够通过将供电服务、电力工程设计、电力工程施工、电力设备制造等有机融合，缩短设计周期，规范工程标准，增供扩销，提升用户满意度，最终不断推动企业价值的产生，形成良性循环发展的商业模式。By applying the solution of the present invention, the existing high-voltage industry expansion design process can be reduced from the original 5 working days to 0.5 working days, achieving a revolutionary improvement in work efficiency. This invention can shorten the design cycle, standardize engineering standards, increase supply and sales, and improve user satisfaction by organically integrating power supply services, electric power engineering design, electric power engineering construction, electric power equipment manufacturing, etc., and ultimately continuously promote the generation of enterprise value and form A business model of virtuous cycle development.
作为优选技术措施：As preferred technical measures:
卷积神经网络的输入数据为激光深度相机通过同步定位与建图技术SLAM收集到的配电房三维重建点云，点云应用偏微分算法，得出变化量较大的点作为特征点云，然后随机抽取其中若干坐标点作为卷积神经网络的一个输入源；采用激光深度相机拍摄的设备图像作为卷积神经网络的另一个输入。The input data of the convolutional neural network is the three-dimensional reconstructed point cloud of the distribution room collected by the laser depth camera through the simultaneous positioning and mapping technology SLAM. The point cloud is applied with the partial differential algorithm to obtain the points with large changes as the characteristic point cloud. Then several coordinate points are randomly selected as an input source of the convolutional neural network; the device image captured by the laser depth camera is used as another input of the convolutional neural network.
通过同步定位与建图技术实现三维数据的收集，可与现场三维数据收集相结合，也能够单独应用，进而能够有效提高数据收集效率，同时提高卷积神经网络的输入数据的多样性以及精准性。The collection of 3D data is realized through simultaneous positioning and mapping technology, which can be combined with on-site 3D data collection or used independently, thereby effectively improving the efficiency of data collection and improving the diversity and accuracy of the input data of the convolutional neural network. .
作为优选技术措施：As preferred technical measures:
卷积神经网络的训练通过模拟系统完成数据收集，采用机器人训练用三维引擎，gazebo，随机生成多个不同大小的配电房和设备；并通过在系统内采用机器人的虚拟激光深度相机作为传感器输入，虚拟激光深度传感器被安装到虚拟机器人的身上，在ROS(robotoperating system)系统内运行机器人SLAM技术用于机器人自主导航收集空间和设备数据，自动生成卷积神经网络训练用数据集，通过GPU集群生成预训练网络；最后，通过预训练网络输出设计勘察数据，其包括设备数量、设备类型和设备相对三维重建的位置坐标。The training of the convolutional neural network is completed through data collection through the simulation system. The three-dimensional engine for robot training, gazebo, is used to randomly generate multiple power distribution rooms and equipment of different sizes; and the robot's virtual laser depth camera is used as sensor input in the system. , the virtual laser depth sensor is installed on the virtual robot, and the robot SLAM technology is run in the ROS (robot operating system) system for autonomous navigation of the robot to collect space and equipment data, and automatically generates a data set for convolutional neural network training, through the GPU cluster Generate a pre-trained network; finally, output the design survey data through the pre-trained network, which includes the number of equipment, equipment type and the position coordinates of the equipment relative to the three-dimensional reconstruction.
作为优选技术措施：As preferred technical measures:
SLAM技术目标是在没有任何先验知识的情况下，根据虚拟激光深度传感器数据实时构建周围环境地图，同时根据这个地图推测自身的定位，虚拟机器人携带虚拟激光深度传感器在未知环境中运动；The goal of SLAM technology is to construct a map of the surrounding environment in real time based on virtual laser depth sensor data without any prior knowledge, and at the same time infer its own positioning based on this map. The virtual robot carries the virtual laser depth sensor to move in an unknown environment;
为方便起见，把一段连续时间的运动变成离散时刻t＝1，…k，而在这些时刻，用x表示虚拟机器人的自身位置，则各时刻的位置就记为x1，x2…xk，它构成了虚拟机器人的轨迹；For the sake of convenience, the movement of a continuous period of time is turned into discrete moments t=1,...k, and at these moments, x is used to represent the virtual robot's own position, and the position at each moment is recorded as x1, x2...xk, which Constitutes the trajectory of the virtual robot;
地图方面，地图由许多个三维点坐标组成，而每个时刻，虚拟激光深度传感器会测量到一部分三维点坐标，得到它们的观测数据；In terms of maps, the map consists of many three-dimensional point coordinates, and at each moment, the virtual laser depth sensor will measure a part of the three-dimensional point coordinates and obtain their observation data;
三维点坐标共有N个，用y1，y2…yn表示；通过运动测量u和虚拟激光深度传感器读数z来求解定位问题x和建图问题y。There are N three-dimensional point coordinates, represented by y1, y2...yn; the positioning problem x and the mapping problem y are solved through motion measurement u and virtual laser depth sensor reading z.
作为优选技术措施：As preferred technical measures:
调用激光ToF传感器，由在配电房现场的工作人员完成配电房的三维点云收集，收集完成的激光点云会通过一个简单的数量判断；The laser ToF sensor is called, and the staff on site in the distribution room completes the three-dimensional point cloud collection of the distribution room. The collected laser point cloud will be judged by a simple quantity;
点数量超过50000的点云能够作为房间关键点坐标提取卷积神经网络的输入，不超过50000点需要工作人员重新收集三维点云；Point clouds with more than 50,000 points can be used as input to the convolutional neural network to extract room key point coordinates. If the number does not exceed 50,000 points, the staff will need to re-collect the three-dimensional point cloud;
点云输入卷积神经网络后，推理(Inference)房间基本形状关键点坐标并向量化。After the point cloud is input into the convolutional neural network, the key point coordinates of the basic shape of the room are inferred and quantized.
通过现场收集三维数据，可与同步定位与建图技术相结合，可与也能够单独应用，进而能够有效提高数据收集效率，同时提高卷积神经网络的输入数据的多样性以及精准性。By collecting three-dimensional data on site, it can be combined with simultaneous positioning and mapping technology, and can be used alone or independently, which can effectively improve the efficiency of data collection and improve the diversity and accuracy of the input data of the convolutional neural network.
作为优选技术措施：As preferred technical measures:
由现场工作人员对用户已有的高压配电设备，采用同一个激光ToF传感器拍摄深度图像，获取到的激光深度图像先要经过无效像素点判断流程；On-site staff use the same laser ToF sensor to capture depth images of the user's existing high-voltage power distribution equipment. The obtained laser depth images must first go through the invalid pixel point determination process;
如果无效像素点数量超过像素点总量的15％，则需要工作人员重新拍摄高压配电设备。无效像素点小于像素总量的15％的深度图像，通过双边滤波去噪算法处理后，输入预训练配网设备识别深度卷积神经网络。If the number of invalid pixels exceeds 15% of the total number of pixels, the staff will need to re-photograph the high-voltage power distribution equipment. Depth images with invalid pixels less than 15% of the total number of pixels are processed by the bilateral filtering denoising algorithm and then input into the pre-trained network equipment recognition deep convolutional neural network.
作为优选技术措施：As preferred technical measures:
在这个网络里完成设备分类和位置坐标提取的步骤，如输出中存在无法识别的模型，现场工作人员需要重新拍摄深度图像，重启流程；如所有模型均识别成功，则输出配电设备数量、类型、中心点坐标参数。Complete the steps of equipment classification and location coordinate extraction in this network. If there are unrecognizable models in the output, on-site staff need to re-take depth images and restart the process; if all models are successfully identified, the number and type of power distribution equipment will be output. , center point coordinate parameters.
作为优选技术措施：所述卷积神经网络是一种前馈型卷积神经网络，其CNNC结构分为五层：As the preferred technical measure: the convolutional neural network is a feed-forward convolutional neural network, and its CNNC structure is divided into five layers:
第一层输入图片，进行卷积(Convolution)操作，得到第二层深度为3的矩阵(Feature Map)；The first layer inputs the image and performs a convolution operation to obtain a matrix (Feature Map) with a depth of 3 in the second layer;
对第二层的矩阵进行池化(Pooling)操作，得到第三层深度为3的矩阵；Perform a pooling operation on the second layer matrix to obtain a third layer matrix with a depth of 3;
重复上述操作得到第五层深度为5的矩阵，最后将这5个矩阵，按行展开连接成向量，传入全连接(Fully Connected)层，全连接层就是一个BP卷积神经网络；每个矩阵都能够看成是排列成矩阵形式的神经元，与BP卷积神经网络中的神经元大同小异。Repeat the above operation to obtain a matrix with a depth of 5 in the fifth layer. Finally, expand and connect these 5 matrices by rows into vectors, and pass them into the fully connected layer. The fully connected layer is a BP convolutional neural network; each Matrices can be regarded as neurons arranged in a matrix form, which are similar to the neurons in BP convolutional neural networks.
作为优选技术措施：As preferred technical measures:
根据需要设定补零的层数；补零层称为Zero Padding，是一个能够设置的超参数，但要根据卷积核的大小，步幅，输入矩阵的大小进行调整，以使得卷积核恰好滑动到边缘；Set the number of zero-padding layers as needed; the zero-padding layer is called Zero Padding, which is a hyperparameter that can be set, but it must be adjusted according to the size, stride, and input matrix size of the convolution kernel so that the convolution kernel Slides right to the edge;
一般情况下，输入的图片矩阵以及后面的卷积核，特征图矩阵都是方阵，这里输入矩阵大小为w，卷积核大小为k，步幅为s，补零层数为p，则卷积后产生的特征图大小计算公式为：Under normal circumstances, the input image matrix and subsequent convolution kernel and feature map matrix are all square matrices. Here, the input matrix size is w, the convolution kernel size is k, the stride is s, and the number of zero-padding layers is p, then The calculation formula for the size of the feature map generated after convolution is:
作为优选技术措施：As preferred technical measures:
为了提取更多的特征，采用多个卷积核分别进行卷积，这样便能够得到多个特征图；有时，对于一张三通道彩色图片，输入的是一组矩阵，这时卷积核也不再是一层的，而要变成相应的深度。In order to extract more features, multiple convolution kernels are used to perform convolution respectively, so that multiple feature maps can be obtained; sometimes, for a three-channel color image, the input is a set of matrices, and then the convolution kernel also It is no longer one layer, but has to become a corresponding depth.
与现有技术相比，本发明具有以下有益效果：Compared with the prior art, the present invention has the following beneficial effects:
本发明通过激光深度传感器数据收集配电房三维信息以及图片，并利用卷积神经网络，获取绘图所需的参数，并将参数传输到CAD模块；最终输出实际可用的配电房设计图纸；能够有效缩短业扩设计时间，提高业扩设计效率。This invention collects three-dimensional information and pictures of the power distribution room through laser depth sensor data, and uses a convolutional neural network to obtain parameters required for drawing, and transmits the parameters to the CAD module; finally outputs actual usable power distribution room design drawings; can Effectively shorten the industrial expansion design time and improve the industrial expansion design efficiency.
应用本发明的方案，能够将现有的高压业扩设计流程，从原有的5个工作日，降低到0.5个工作日，实现工作效率的革命性提升。本发明能够通过将供电服务、电力工程设计、电力工程施工、电力设备制造等有机融合，缩短设计周期，规范工程标准，增供扩销，提升用户满意度，最终不断推动企业价值的产生，形成良性循环发展的商业模式。By applying the solution of the present invention, the existing high-voltage industry expansion design process can be reduced from the original 5 working days to 0.5 working days, achieving a revolutionary improvement in work efficiency. This invention can shorten the design cycle, standardize engineering standards, increase supply and sales, and improve user satisfaction by organically integrating power supply services, electric power engineering design, electric power engineering construction, electric power equipment manufacturing, etc., and ultimately continuously promote the generation of enterprise value and form A business model of virtuous cycle development.
附图说明Description of the drawings
图1为本发明工作流程图；Figure 1 is a work flow chart of the present invention;
图2为本发明输出设备数量、设备类型和位置坐标的工作流程图。Figure 2 is a workflow diagram for outputting the number of devices, device types and location coordinates of the present invention.
具体实施方式Detailed ways
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention and are not intended to limit the present invention.
相反，本发明涵盖任何由权利要求定义的在本发明的精髓和范围上做的替代、修改、等效方法以及方案。进一步，为了使公众对本发明有更好的了解，在下文对本发明的细节描述中，详尽描述了一些特定的细节部分。对本领域技术人员来说没有这些细节部分的描述也能够完全理解本发明。On the contrary, the invention covers any alternatives, modifications, equivalent methods and solutions that fall within the spirit and scope of the invention as defined by the claims. Furthermore, in order to enable the public to have a better understanding of the present invention, some specific details are described in detail in the following detailed description of the present invention. A person skilled in the art will be able to fully understand the present invention without these detailed descriptions.
如图1-2所示，一种基于云计算的用户配电房设计生成方法，As shown in Figure 1-2, a cloud computing-based user distribution room design and generation method,
其包括以下步骤：It includes the following steps:
第一步，建立采用为python编程语音定义的Web服务器和Web应用程序和深度计算框架的接口程序WSGI；The first step is to establish the interface program WSGI using the web server and web application defined for Python programming language and the deep computing framework;
第二步，通过在配电房现场的工作人员，利用激光深度相机收集配电房三维信息；In the second step, the staff at the power distribution room site use the laser depth camera to collect the three-dimensional information of the power distribution room;
第三步，通过无线网络将三维重建信息与设备的激光深度照片传输到云端服务器的WSGI；The third step is to transmit the 3D reconstruction information and the laser depth photos of the device to the WSGI of the cloud server through the wireless network;
第四步，通过云端的WSGI接口将数据传输到基于GPU集群的卷积神经网络推理inference模块，获取绘图所需的参数，并将参数传输到CAD模块的pythonAPI接口；The fourth step is to transmit the data to the convolutional neural network inference module based on the GPU cluster through the WSGI interface of the cloud, obtain the parameters required for drawing, and transmit the parameters to the pythonAPI interface of the CAD module;
第五步，使用一种云端加速计算架构来加快现场传感器收集数据的空间计算量，最终输出配电房现场收资的标准图纸。The fifth step is to use a cloud accelerated computing architecture to accelerate the spatial calculation of data collected by on-site sensors, and finally output standard drawings for on-site collection of power distribution rooms.
本发明通过激光深度传感器数据收集配电房三维信息以及图片，并利用卷积神经网络，获取绘图所需的参数，并将参数传输到CAD模块；最终输出实际可用的配电房设计图纸；能够有效缩短业扩设计时间，提高业扩设计效率。This invention collects three-dimensional information and pictures of the power distribution room through laser depth sensor data, and uses a convolutional neural network to obtain parameters required for drawing, and transmits the parameters to the CAD module; finally outputs actual usable power distribution room design drawings; can Effectively shorten the industrial expansion design time and improve the industrial expansion design efficiency.
应用本发明的方案，能够将现有的高压业扩设计流程，从原有的5个工作日，降低到0.5个工作日，实现工作效率的革命性提升。本发明能够通过将供电服务、电力工程设计、电力工程施工、电力设备制造等有机融合，缩短设计周期，规范工程标准，增供扩销，提升用户满意度，最终不断推动企业价值的产生，形成良性循环发展的商业模式。By applying the solution of the present invention, the existing high-voltage industry expansion design process can be reduced from the original 5 working days to 0.5 working days, achieving a revolutionary improvement in work efficiency. This invention can shorten the design cycle, standardize engineering standards, increase supply and sales, and improve user satisfaction by organically integrating power supply services, electric power engineering design, electric power engineering construction, electric power equipment manufacturing, etc., and ultimately continuously promote the generation of enterprise value and form A business model of virtuous cycle development.
本发明卷积神经网络输入源一种具体实施例：A specific embodiment of the input source of the convolutional neural network of the present invention:
卷积神经网络的输入数据为激光深度相机通过同步定位与建图技术SLAM收集到的配电房三维重建点云，点云应用偏微分算法，得出变化量较大的点作为特征点云，然后随机抽取其中若干坐标点作为卷积神经网络的一个输入源；采用激光深度相机拍摄的设备图像作为卷积神经网络的另一个输入。The input data of the convolutional neural network is the three-dimensional reconstructed point cloud of the distribution room collected by the laser depth camera through the simultaneous positioning and mapping technology SLAM. The point cloud is applied with the partial differential algorithm to obtain the points with large changes as the characteristic point cloud. Then several coordinate points are randomly selected as an input source of the convolutional neural network; the device image captured by the laser depth camera is used as another input of the convolutional neural network.
本发明卷积神经网络训练模拟一种具体实施例：A specific embodiment of the convolutional neural network training simulation of the present invention:
卷积神经网络的训练通过模拟系统完成数据收集，采用机器人训练用三维引擎，gazebo，随机生成多个不同大小的配电房和设备；并通过在系统内采用机器人的虚拟激光深度相机作为传感器输入，虚拟激光深度传感器被安装到虚拟机器人的身上，在ROS(robotoperating system)系统内运行机器人SLAM技术用于机器人自主导航收集空间和设备数据，自动生成卷积神经网络训练用数据集，通过GPU集群生成预训练网络；最后，通过预训练网络输出设计勘察数据，其包括设备数量、设备类型和设备相对三维重建的位置坐标。The training of the convolutional neural network is completed through data collection through the simulation system. The three-dimensional engine for robot training, gazebo, is used to randomly generate multiple power distribution rooms and equipment of different sizes; and the robot's virtual laser depth camera is used as sensor input in the system. , the virtual laser depth sensor is installed on the virtual robot, and the robot SLAM technology is run in the ROS (robot operating system) system for autonomous navigation of the robot to collect space and equipment data, and automatically generates a data set for convolutional neural network training, through the GPU cluster Generate a pre-trained network; finally, output the design survey data through the pre-trained network, which includes the number of equipment, equipment type and the position coordinates of the equipment relative to the three-dimensional reconstruction.
本发明SLAM技术一种具体实施例：A specific embodiment of the SLAM technology of the present invention:
SLAM技术目标是在没有任何先验知识的情况下，根据虚拟激光深度传感器数据实时构建周围环境地图，同时根据这个地图推测自身的定位，虚拟机器人携带虚拟激光深度传感器在未知环境中运动；The goal of SLAM technology is to construct a map of the surrounding environment in real time based on virtual laser depth sensor data without any prior knowledge, and at the same time infer its own positioning based on this map. The virtual robot carries the virtual laser depth sensor to move in an unknown environment;
为方便起见，把一段连续时间的运动变成离散时刻t＝1，…k，而在这些时刻，用x表示虚拟机器人的自身位置，则各时刻的位置就记为x1，x2…xk，它构成了虚拟机器人的轨迹；For the sake of convenience, the movement of a continuous period of time is turned into discrete moments t=1,...k, and at these moments, x is used to represent the virtual robot's own position, and the position at each moment is recorded as x1, x2...xk, which Constitutes the trajectory of the virtual robot;
地图方面，地图由许多个三维点坐标组成，而每个时刻，虚拟激光深度传感器会测量到一部分三维点坐标，得到它们的观测数据；In terms of maps, the map consists of many three-dimensional point coordinates, and at each moment, the virtual laser depth sensor will measure a part of the three-dimensional point coordinates and obtain their observation data;
三维点坐标共有N个，用y1，y2…yn表示；通过运动测量u和虚拟激光深度传感器读数z来求解定位问题x和建图问题y。There are N three-dimensional point coordinates, represented by y1, y2...yn; the positioning problem x and the mapping problem y are solved through motion measurement u and virtual laser depth sensor reading z.
本发明卷积神经网络输入源另一种具体实施例：Another specific embodiment of the input source of the convolutional neural network of the present invention:
调用激光ToF传感器，由在配电房现场的工作人员完成配电房的三维点云收集，收集完成的激光点云会通过一个简单的数量判断；The laser ToF sensor is called, and the staff on site in the distribution room completes the three-dimensional point cloud collection of the distribution room. The collected laser point cloud will be judged by a simple quantity;
点数量超过50000的点云能够作为房间关键点坐标提取卷积神经网络的输入，不超过50000点需要工作人员重新收集三维点云；Point clouds with more than 50,000 points can be used as input to the convolutional neural network to extract room key point coordinates. If the number does not exceed 50,000 points, the staff will need to re-collect the three-dimensional point cloud;
点云输入卷积神经网络后，推理(Inference)房间基本形状关键点坐标并向量化。After the point cloud is input into the convolutional neural network, the key point coordinates of the basic shape of the room are inferred and quantized.
本发明无效像素点判断的一种具体实施例：A specific embodiment of invalid pixel judgment according to the present invention:
由现场工作人员对用户已有的高压配电设备，采用同一个激光ToF传感器拍摄深度图像，获取到的激光深度图像先要经过无效像素点判断流程；On-site staff use the same laser ToF sensor to capture depth images of the user's existing high-voltage power distribution equipment. The obtained laser depth images must first go through the invalid pixel point determination process;
如果无效像素点数量超过像素点总量的15％，则需要工作人员重新拍摄高压配电设备。无效像素点小于像素总量的15％的深度图像，通过双边滤波去噪算法处理后，输入预训练配网设备识别深度卷积神经网络。If the number of invalid pixels exceeds 15% of the total number of pixels, the staff will need to re-photograph the high-voltage power distribution equipment. Depth images with invalid pixels less than 15% of the total number of pixels are processed by the bilateral filtering denoising algorithm and then input into the pre-trained network equipment recognition deep convolutional neural network.
在这个网络里完成设备分类和位置坐标提取的步骤，如输出中存在无法识别的模型，现场工作人员需要重新拍摄深度图像，重启流程；如所有模型均识别成功，则输出配电设备数量、类型、中心点坐标参数。Complete the steps of equipment classification and location coordinate extraction in this network. If there are unrecognizable models in the output, on-site staff need to re-take depth images and restart the process; if all models are successfully identified, the number and type of power distribution equipment will be output. , center point coordinate parameters.
本发明卷积神经网络架构层次的一种具体实施例：所述卷积神经网络是一种前馈型卷积神经网络，其CNNC结构分为五层：A specific embodiment of the architecture level of the convolutional neural network of the present invention: the convolutional neural network is a feed-forward convolutional neural network, and its CNNC structure is divided into five layers:
第一层输入图片，进行卷积(Convolution)操作，得到第二层深度为3的矩阵(Feature Map)；The first layer inputs the image and performs a convolution operation to obtain a matrix (Feature Map) with a depth of 3 in the second layer;
对第二层的矩阵进行池化(Pooling)操作，得到第三层深度为3的矩阵；Perform a pooling operation on the second layer matrix to obtain a third layer matrix with a depth of 3;
重复上述操作得到第五层深度为5的矩阵，最后将这5个矩阵，按行展开连接成向量，传入全连接(Fully Connected)层，全连接层就是一个BP卷积神经网络；每个矩阵都能够看成是排列成矩阵形式的神经元，与BP卷积神经网络中的神经元大同小异。Repeat the above operation to obtain a matrix with a depth of 5 in the fifth layer. Finally, expand and connect these 5 matrices by rows into vectors, and pass them into the fully connected layer. The fully connected layer is a BP convolutional neural network; each Matrices can be regarded as neurons arranged in a matrix form, which are similar to the neurons in BP convolutional neural networks.
根据需要设定补零的层数；补零层称为Zero Padding，是一个能够设置的超参数，但要根据卷积核的大小，步幅，输入矩阵的大小进行调整，以使得卷积核恰好滑动到边缘；Set the number of zero-padding layers as needed; the zero-padding layer is called Zero Padding, which is a hyperparameter that can be set, but it must be adjusted according to the size, stride, and input matrix size of the convolution kernel so that the convolution kernel Slides right to the edge;
一般情况下，输入的图片矩阵以及后面的卷积核，特征图矩阵都是方阵，这里输入矩阵大小为w，卷积核大小为k，步幅为s，补零层数为p，则卷积后产生的特征图大小计算公式为：Under normal circumstances, the input image matrix and subsequent convolution kernel and feature map matrix are all square matrices. Here, the input matrix size is w, the convolution kernel size is k, the stride is s, and the number of zero-padding layers is p, then The calculation formula for the size of the feature map generated after convolution is:
为了提取更多的特征，采用多个卷积核分别进行卷积，这样便能够得到多个特征图；有时，对于一张三通道彩色图片，输入的是一组矩阵，这时卷积核也不再是一层的，而要变成相应的深度。In order to extract more features, multiple convolution kernels are used to perform convolution respectively, so that multiple feature maps can be obtained; sometimes, for a three-channel color image, the input is a set of matrices, and then the convolution kernel also It is no longer one layer, but has to become a corresponding depth.
以上所述仅为本发明的较佳实施例而已，并不用以限制本发明，凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911341997.6ACN111192363B (en) | 2019-12-23 | 2019-12-23 | User power distribution room design generation method based on cloud computing |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911341997.6ACN111192363B (en) | 2019-12-23 | 2019-12-23 | User power distribution room design generation method based on cloud computing |
| Publication Number | Publication Date |
|---|---|
| CN111192363A CN111192363A (en) | 2020-05-22 |
| CN111192363Btrue CN111192363B (en) | 2023-09-29 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911341997.6AActiveCN111192363B (en) | 2019-12-23 | 2019-12-23 | User power distribution room design generation method based on cloud computing |
| Country | Link |
|---|---|
| CN (1) | CN111192363B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111832542B (en)* | 2020-08-15 | 2024-04-16 | 武汉易思达科技有限公司 | Tri-vision identifying and positioning device |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103400415A (en)* | 2013-08-07 | 2013-11-20 | 赞奇科技发展有限公司 | Method and system for automatic generation and online interaction of three-dimensional digital model of planar space structure |
| CN105027190A (en)* | 2013-01-03 | 2015-11-04 | 美达公司 | Projective spatial imaging digital glasses for virtual or augmented mediated vision |
| CN108171796A (en)* | 2017-12-25 | 2018-06-15 | 燕山大学 | A kind of inspection machine human visual system and control method based on three-dimensional point cloud |
| CN108365994A (en)* | 2018-03-13 | 2018-08-03 | 山东超越数控电子股份有限公司 | A kind of cloud security management platform for cloud computing security and unity management |
| CN108921893A (en)* | 2018-04-24 | 2018-11-30 | 华南理工大学 | A kind of image cloud computing method and system based on online deep learning SLAM |
| CN109801358A (en)* | 2018-12-06 | 2019-05-24 | 宁波市电力设计院有限公司 | A kind of substation's three-dimensional investigation method scanning and put cloud visual fusion based on SLAM |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105027190A (en)* | 2013-01-03 | 2015-11-04 | 美达公司 | Projective spatial imaging digital glasses for virtual or augmented mediated vision |
| CN103400415A (en)* | 2013-08-07 | 2013-11-20 | 赞奇科技发展有限公司 | Method and system for automatic generation and online interaction of three-dimensional digital model of planar space structure |
| CN108171796A (en)* | 2017-12-25 | 2018-06-15 | 燕山大学 | A kind of inspection machine human visual system and control method based on three-dimensional point cloud |
| CN108365994A (en)* | 2018-03-13 | 2018-08-03 | 山东超越数控电子股份有限公司 | A kind of cloud security management platform for cloud computing security and unity management |
| CN108921893A (en)* | 2018-04-24 | 2018-11-30 | 华南理工大学 | A kind of image cloud computing method and system based on online deep learning SLAM |
| CN109801358A (en)* | 2018-12-06 | 2019-05-24 | 宁波市电力设计院有限公司 | A kind of substation's three-dimensional investigation method scanning and put cloud visual fusion based on SLAM |
| Title |
|---|
| 张向荣等.《模式识别》.西安电子科技大学出版社,2019,第161-164页.* |
| 张彬等.《图像复原优化算法》.国防工业出版社,2019,第262页.* |
| 王权等.基于ROS与Kinect的移动机器人同时定位与地图构建.《计算机应用研究》.2016,第34卷(第10期),第3184-3187页.* |
| 黑马程序员.《Python程序开发案例教程》.中国铁道出版社有限公司,2019,第262页.* |
| Publication number | Publication date |
|---|---|
| CN111192363A (en) | 2020-05-22 |
| Publication | Publication Date | Title |
|---|---|---|
| US20240257462A1 (en) | Method, apparatus, and storage medium for three-dimensional reconstruction of buildings based on missing point cloud data | |
| CN113822982B (en) | Human body three-dimensional model construction method and device, electronic equipment and storage medium | |
| CN108647639B (en) | A real-time human skeleton joint detection method | |
| WO2021175050A1 (en) | Three-dimensional reconstruction method and three-dimensional reconstruction device | |
| CN108537191B (en) | A three-dimensional face recognition method based on structured light camera | |
| CN112818925B (en) | Urban building and crown identification method | |
| CN111598998A (en) | Three-dimensional virtual model reconstruction method and device, computer equipment and storage medium | |
| WO2022001236A1 (en) | Three-dimensional model generation method and apparatus, and computer device and storage medium | |
| CN113077545B (en) | Method for reconstructing clothing human body model from image based on graph convolution | |
| CA3090747C (en) | Automatic rig creation process | |
| JP6863596B2 (en) | Data processing device and data processing method | |
| US12175698B2 (en) | Method and apparatus with object pose estimation | |
| CN109325995B (en) | Low-resolution multi-view hand reconstruction method based on hand parameter model | |
| CN111028335B (en) | A deep learning-based patch reconstruction method for point cloud data | |
| CN113313176B (en) | A point cloud analysis method based on dynamic graph convolutional neural network | |
| CN116310219B (en) | A three-dimensional foot shape generation method based on conditional diffusion model | |
| CN115018992B (en) | Method, device, electronic device and storage medium for generating hairstyle model | |
| CN118820557B (en) | Cable line fine monitoring method and system based on three-dimensional digital model | |
| CN109583377A (en) | A kind of control method, device and host computer that pipeline model is rebuild | |
| CN111192363B (en) | User power distribution room design generation method based on cloud computing | |
| CN111047703B (en) | User high-voltage distribution equipment identification and space reconstruction method | |
| CN105321205B (en) | A kind of parameterized human body model method for reconstructing based on sparse key point | |
| CN116363329B (en) | Three-dimensional image generation method and system based on CGAN and LeNet-5 | |
| CN117576213A (en) | Non-cooperative target pose estimation method based on global feature vector field voting network | |
| CN117253012A (en) | Method for restoring plane building free-form surface grid structure to three-dimensional space |
| 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 |