CN101833673A - Image Recognition System of Power Switchgear Switching State - Google Patents

Abstract

The invention relates to an electric power switchgear switch state image recognition system comprising an off-line training module and an on-line detecting module. The off-line training module collects and trains a large number of switch samples to form a plurality of strong classifiers, and the strong classifiers are loaded in the on-line detecting module. The on-line detecting module transforms the collected switch images to switch gray images and denoises, utilizes the strong classifiers to respectively scan and search the gray images, and obtains the positions and state information of the switches. The invention can realize that the image collection, processing and recognition processes are completed in the on-line detecting module, and overcomes the dependence to communication bandwidth and equipment because videos collected by the traditional video monitoring systems are transmitted to remote background systems through a wired or wireless mode. The invention can automatically determine the positions and the states of the switches in the images, does not need manual operation, has high degree of automation, and is suitable for state recognition of various types of electric power switches.

Description

Translated fromChinese

电力开关柜开关状态图像识别系统Image Recognition System of Power Switchgear Switching State

技术领域technical field

本发明涉及电力配网监测领域基于图像识别的电力开关柜遥信系统，具体涉及对电力开关柜开关状态的实时监控的图像处理、识别系统。The invention relates to an image recognition-based power switch cabinet remote signaling system in the field of power distribution network monitoring, in particular to an image processing and recognition system for real-time monitoring of the switch state of the power switch cabinet.

背景技术Background technique

在电力系统中，开关柜应用非常广泛，能够实时指示电力线路运行状态，对排除电力故障有一定的帮助。但由于电力系统分布区域广阔，在出现故障时无法在第一时间报告工作人员在哪一处发生故障，只能由人工一处一处检查电力机房开关柜状态，非常耗费时间与人力，效率低下。In the power system, the switchgear is widely used, and it can indicate the operating status of the power line in real time, which is helpful to eliminate power failures. However, due to the vast distribution area of the power system, when a fault occurs, it is impossible to report where the fault occurred to the staff at the first time. It is only possible to manually check the status of the switchgear in the power room one by one, which is very time-consuming and manpower-inefficient. .

目前如果要远程读取开关柜的状态就要将开关电路通过现场总线技术连接到处理器或者控制器单元，用处理器或者控制器来读取电力线路的状态，并通过有线或者无线的方式来把读取到的状态发送到远程的控制中心。这种实现方式在可靠性和灵活性上都存在一些不足。而现有的视频监控系统，终端通常仅起到采集视频的作用，图像还需通过有线或无线的方式传输到远端控制中心，进一步进行后台处理识别。这样需要较高通信带宽，对设备成本要求极高。At present, if you want to remotely read the state of the switch cabinet, you need to connect the switch circuit to the processor or controller unit through field bus technology, use the processor or controller to read the state of the power line, and use wired or wireless methods to Send the read status to the remote control center. This implementation has some shortcomings in reliability and flexibility. In the existing video monitoring system, the terminal usually only plays the role of collecting video, and the image needs to be transmitted to the remote control center through wired or wireless means, and further background processing and identification are carried out. This requires a higher communication bandwidth and requires extremely high equipment costs.

现已有的电力开关柜开关开关状态实时图像识别装置采用灰度行程长度统计法实现状态识别，此算法简单，但需人工进行目标加框定位，自动化程度较低，且不易扩展适用于其他类型电力开关。The existing real-time image recognition device for switch state of electric switchgear adopts the statistical method of gray stroke length to realize state recognition. This algorithm is simple, but it needs manual positioning of the target frame, the degree of automation is low, and it is not easy to expand and apply to other types power switch.

发明内容Contents of the invention

本发明所要解决的技术问题是针对目前流行的电力故障监控系统的不足，提出了一种电力开关柜开关状态实时图像采集处理识别系统，它可以实现对电力开关柜开关状态的远程实时自动监控，无需人工进行目标定位，自动化程度高，适用于各种类型电力开关。The technical problem to be solved by the present invention is to address the deficiencies of the current popular power failure monitoring system, and propose a real-time image acquisition and processing recognition system for the switch status of the power switch cabinet, which can realize remote real-time automatic monitoring of the switch status of the power switch cabinet, No manual target positioning is required, the degree of automation is high, and it is suitable for various types of power switches.

为解决上述技术问题，本发明提出了一种电力开关柜开关状态实时图像识别系统，所述图像识别系统包括离线训练模块及在线识别模块，应用于电力开关柜开关状态的实时图像处理识别装置中；所述离线训练模块，基于Adaboost算法对大量包含开关状态的真样本和假样本进行学习，形成一系列的弱分类器，然后依据权重把这些弱分类器级联成若干强分类器，所述强分类器分别代表不同的开关状态；In order to solve the above technical problems, the present invention proposes a real-time image recognition system for the switch status of the power switch cabinet, the image recognition system includes an offline training module and an online recognition module, which is applied to a real-time image processing and recognition device for the switch status of the power switch cabinet ; The off-line training module, based on the Adaboost algorithm, learns a large number of true samples and false samples that include switch states to form a series of weak classifiers, and then cascades these weak classifiers into several strong classifiers according to the weights. The strong classifiers represent different switch states respectively;

所述在线检测模块包括图像初步处理模块和状态识别模块；The online detection module includes an image preliminary processing module and a state recognition module;

所述图像初步处理模块,用于将采集到的开关图像转换为开关灰度图并去噪，以便状态识别模块进行处理；The image preliminary processing module is used to convert the collected switch image into a switch grayscale image and remove noise, so that the state recognition module can process it;

所述状态识别模块,装载了所述强分类器，并利用所述强分类器分别对所述开关灰度图进行扫描搜索，得到开关的位置和状态信息。The state identification module is loaded with the strong classifier, and uses the strong classifier to scan and search the switch grayscale image to obtain the position and state information of the switch.

对采集到的图像数据实时处理，采用基于Adaboost的目标检测算法，分别用三种开关状态的分类器对图像进行检测，得出当前图像中包含的开关状态结论。The collected image data is processed in real time, and the target detection algorithm based on Adaboost is used to detect the image with three switch state classifiers respectively, and the conclusion of the switch state contained in the current image is obtained.

优选的，在所述离线训练模块中，每个特定弱分类器所使用的特征用形状、感兴趣区域中的位置以及比例系数来定义；提取矩形特征时，用积分图像方法来减小计算量。Preferably, in the offline training module, the features used by each specific weak classifier are defined by the shape, the position in the region of interest and the scale factor; when extracting the rectangular feature, the integral image method is used to reduce the amount of calculation .

优选的，所述图像初步处理模块包括转化为灰度图单元和图像去噪单元。所述转化灰度图单元将采集到的彩色开关图像转化为灰度图像，所述图像去噪单元对灰度图像进行中值滤波，将灰度中存在的噪音点去除，避免了噪声点对图像识别带来的干扰。Preferably, the image preliminary processing module includes a conversion into a grayscale image unit and an image denoising unit. The converted grayscale image unit converts the collected color switch image into a grayscale image, and the image denoising unit performs median filtering on the grayscale image to remove noise points existing in the grayscale, thereby avoiding the noise point Interference caused by image recognition.

优选的，状态识别模块采用基于Adaboost的目标检测算法，以不同比例大小的扫描窗口对图像进行搜索，能够通过强分类器的即为一个目标，识别单元通过该目标的区域位置和强分类器的种类以确定某个开关的状态。Preferably, the state recognition module uses an Adaboost-based target detection algorithm to search the image with scan windows of different scales, and the object that can pass the strong classifier is a target, and the recognition unit passes the regional position of the target and the strong classifier. Kind to determine the state of a switch.

采用本发明，通过视频监控和图像识别的方式来读取开关状态，并可配合视频监控系统将识别处理后的开关状态信息通过配网通信网络实时传送到后台系统，达到远程监控的目的。本发明可自行判断图像中开关的位置与状态，无需人工操作，自动化程度较高；也可与已有的人工甄选方法结合判断复杂开关的状态。本发明仅将计算量大的离线训练模块在远端控制中心完成，而所有处理识别过程都在终端装置中完成，仅需将处理识别后的开关状态信息发送到远端工作人员手中，也可根据后台系统需要传送实时视频。克服了现有视频监控系统，终端仅起到采集视频并通过有线或无线的方式传输到远端控制中心，进一步进行后台处理识别而产生的对通信带宽和设备成本的依赖。By adopting the present invention, the switch status is read through video monitoring and image recognition, and the switch status information after identification and processing can be transmitted to the background system in real time through the distribution network communication network in cooperation with the video monitoring system, so as to achieve the purpose of remote monitoring. The invention can judge the position and state of the switch in the image by itself, without manual operation, and has a high degree of automation; it can also be combined with the existing manual selection method to judge the state of complex switches. In the present invention, only the off-line training module with a large amount of calculation is completed in the remote control center, and all the processing and identification processes are completed in the terminal device. Send real-time video according to the needs of the background system. It overcomes the dependence on communication bandwidth and equipment cost caused by the existing video surveillance system, where the terminal only collects video and transmits it to the remote control center through wired or wireless methods, and further performs background processing and identification.

同时，由于监控采用图像识别方式，所用摄像头在开关柜远处对其进行监控，和电力设备距离比较远，之间没有任何连接，从而能够保证系统可靠性。另外监控系统可以单独建立，位置可以自由选择，设备安装和维护都很灵活方便，和电力设备之间不会产生相互影响，并且安装和维护人员的安全也可以得到充分的保障。At the same time, since the monitoring adopts the image recognition method, the camera used is used to monitor the switch cabinet at a distance, and the distance from the power equipment is relatively far away, and there is no connection between them, so that the reliability of the system can be guaranteed. In addition, the monitoring system can be established independently, the location can be freely selected, the equipment installation and maintenance are very flexible and convenient, and there will be no interaction with the power equipment, and the safety of the installation and maintenance personnel can also be fully guaranteed.

附图说明Description of drawings

下面结合附图和具体实施方式对本发明的技术方案作进一步具体说明。The technical solutions of the present invention will be further specifically described below in conjunction with the accompanying drawings and specific embodiments.

图1是电力配网系统开关柜开关状态图像识别系统的结构框图。Fig. 1 is a structural block diagram of an image recognition system for switchgear switch status in a power distribution network system.

图2是开关柜圆盘旋转型开关三种开关状态原始图，导通状态（a），接地状态（b），悬空状态（c）。Figure 2 is the original diagram of the three switching states of the disc rotary switch in the switch cabinet, the conduction state (a), the grounding state (b), and the floating state (c).

图3是分类器特征矩形示意图。Fig. 3 is a schematic diagram of a feature rectangle of a classifier.

图4是计算积分图像的示意图。Fig. 4 is a schematic diagram of calculating an integral image.

图5是截取开关柜某一时刻的开关状态，取图像灰度图。Figure 5 is to intercept the switch state of the switch cabinet at a certain moment, and take the grayscale image of the image.

图6是应用本发明算法的开关状态检测装置得到的结果图。 Fig. 6 is a result diagram obtained by applying the switch state detection device of the algorithm of the present invention. the

具体实施方式Detailed ways

下面以电力配网系统开关柜圆盘旋转型开关为例，说明本发明的具体实施方式。The specific implementation of the present invention will be described below by taking the disc rotary switch of the switchgear of the power distribution network system as an example.

如图1为电力开关柜开关状态图像识别系统的结构框图。总体可为离线训练模块1和在线检测模块3。离线训练模块1在远端后台系统中训练形成分别包含三种开关状态特征的强分类器2，存储在前台终端中，供状态识别模块5使用。采集到的开关柜上开关状态图像送往图像初步处理单元4，经灰度图转化和去噪后，状态识别模块5利用级联分类器2分别对图像进行搜索，判断开关的状态输出。Figure 1 is a structural block diagram of the image recognition system for the switch state of the power switch cabinet. Overall, it can be an offline training module 1 and an online detection module 3. The off-line training module 1 trains in the remote background system to form astrong classifier 2 containing three switch state features, which is stored in the foreground terminal for use by the state recognition module 5 . The collected switch state images on the switchgear are sent to the image preliminary processing unit 4. After the grayscale image conversion and denoising, the state recognition module 5 uses thecascade classifier 2 to search the images respectively, and judge the state output of the switch.

如图2所示的电力开关柜开关实物图片，圆盘旋转型开关在不同工作状态会呈现出三种不同开关状态：导通状态（a），接地状态（b），悬空状态（c）。As shown in Figure 2, the physical picture of the switch in the power switch cabinet, the disc rotary switch will show three different switching states in different working states: conduction state (a), grounding state (b), and floating state (c).

离线训练模块是通过对大量真样本和假样本的学习，形成一系列的弱分类器，然后依据权重把这些弱分类器级联成强分类器。理论证明，只要每个弱分类器分类能力比随机猜测好，当弱分类器个数趋向于无穷时，强分类器的错误率将趋于0。训练方法是，给定一个训练集(x1,y1),…,(xL,yL)，其中，xi是输入的训练样本,yi是分类的类别标志即真假样本。在初始化时，对所有训练样本均赋以一个相同的权重，然后用该弱学习算法对训练样本集进行T轮训练。在每一轮训练结束后，从若干个简单分类器中选择最小误差的那个，作为一个弱分类器hi,并对训练失败的样本赋以较大的权重，以便让学习算法在后来的学习中主要对比较难的训练样本进行学习。这样，就可以得到一个弱分类器序列(h1,h2,…,ht)，其中，分类效果比较好的权重较大。最终的分类函数f(x)采用一种有权重的投票方式产生，即将多个弱分类器通过一定的方法叠加起来组合成一个强分类器，即The offline training module forms a series of weak classifiers by learning a large number of real samples and fake samples, and then cascades these weak classifiers into strong classifiers according to weights. Theory proves that as long as the classification ability of each weak classifier is better than random guessing, when the number of weak classifiers tends to infinity, the error rate of strong classifiers will tend to 0. The training method is to give a training set (x1, y1),..., (xL, yL), where xi is the input training sample, and yi is the category mark of the classification, that is, the true and false samples. At the time of initialization, all training samples are assigned the same weight, and then the weak learning algorithm is used to perform T rounds of training on the training sample set. After each round of training, the one with the smallest error is selected from several simple classifiers as a weak classifier hi, and a larger weight is assigned to the samples that fail to train, so that the learning algorithm can be used in subsequent learning. It mainly studies the more difficult training samples. In this way, a sequence of weak classifiers (h1, h2, ..., ht) can be obtained, in which the weight of the better classification effect is greater. The final classification function f(x) is generated by a weighted voting method, that is, multiple weak classifiers are superimposed and combined into a strong classifier by a certain method, that is,

每个特定分类器所使用的特征用形状、感兴趣区域中的位置以及比例系数来定义。矩形特征的值是指图像上两个或多个矩形内部所有像素灰度值之和的差值。如图3所示，其中3(a)是训练用的正样本图片，3(b)是提取的一种开关状态的一个矩形特征,它代表的是黑色矩形区域的像素灰度要比其上下白色区域的灰度大，对应于开关圆盘及上下标示白线部分。The features used by each particular classifier are defined by shape, location in the region of interest, and scale factor. The value of the rectangle feature refers to the difference between the sum of the gray values of all pixels inside two or more rectangles on the image. As shown in Figure 3, 3(a) is a positive sample image for training, and 3(b) is a rectangular feature extracted from a switch state, which represents that the pixel grayscale of the black rectangular area is higher than its upper and lower The gray scale of the white area is large, corresponding to the switch disc and the white line marked up and down.

在提取矩形特征的过程中，可以使用积分图像的方法来减小计算量。在提取矩形特征的过程中，可以使用积分图像的方法减小计算量。如图4，点x,y处的积分图像值是其上部和左边所有像素的和。这样通过积分图像法，任意矩形中的像素和可以通过其4个顶点的值计算出。因此整个训练过程只需扫描原图一遍。In the process of extracting rectangular features, the method of integral image can be used to reduce the amount of calculation. In the process of extracting rectangular features, the method of integrating images can be used to reduce the amount of calculation. As shown in Figure 4, the integral image value at point x, y is the sum of all pixels above and to the left of it. In this way, through the integral image method, the sum of pixels in any rectangle can be calculated from the values of its four vertices. Therefore, the entire training process only needs to scan the original image once.

针对开关的三种不同状态，分别训练三个分类器，当前状态之外的另两种状态都应作为假样本参与训练，这样能提高算法的识别率。For the three different states of the switch, three classifiers are trained respectively, and the other two states other than the current state should be used as fake samples to participate in the training, which can improve the recognition rate of the algorithm.

在对采集到的图像进行识别之前，需要对图像进行预处理。预处理效果的好坏是影响整个测试系统性能的主要因素，主要是转化为灰度图和去噪。首先由转化为灰度图单元将截取的图像转化为灰度图，如图5所示，以便于后面的图像处理识别。Before recognizing the collected images, the images need to be preprocessed. The quality of the preprocessing effect is the main factor affecting the performance of the entire test system, mainly converting to grayscale image and denoising. First, the intercepted image is converted into a grayscale image by the converting to grayscale image unit, as shown in Figure 5, so as to facilitate subsequent image processing and identification.

然后，由状态识别单元对转化为灰度图的图像进行状态识别。状态识别采用已训练好的分别含有不同状态的开关特征的三个分类器在图像中找到包含目标的矩形区域，并通过相应的分类器判断开关状态。识别单元以不同比例大小的扫描窗口对图像进行几次搜索。每次扫描中，当分析的矩形框全部通过级联分类器每一层时返回正值，代表这是一个候选目标。在处理和收集到候选的方框之后，接着对这些区域进行组合并且返回一系列个数足够大的组合中的平均矩形作为检测到的目标区域。以此可得到开关在图像中的位置信息，并通过相应的分类器判断开关状态，即若检测到开关的分类器是导通状态的，则该开关应处于导通状态。Then, the state recognition unit performs state recognition on the image converted into a grayscale image. State recognition uses three trained classifiers that contain switch features of different states to find the rectangular area containing the target in the image, and judge the switch state through the corresponding classifier. The recognition unit searches the image several times with scan windows of different scale sizes. In each scan, when all the analyzed rectangles pass through each layer of the cascade classifier, a positive value is returned, indicating that this is a candidate object. After processing and collecting candidate boxes, these regions are then combined and the average rectangle in a series of sufficiently large combinations is returned as the detected object region. In this way, the position information of the switch in the image can be obtained, and the switch state can be judged by the corresponding classifier, that is, if the classifier that detects the switch is in the conduction state, then the switch should be in the conduction state.

得到的开关状态信息可通过通信装置传到后台系统中，以供进一步分析。如图6是应用本发明算法的开关状态检测装置得到的结果图，其训练和检查阶段都使用的是模拟开关模型。The obtained switch status information can be transmitted to the background system through the communication device for further analysis. Fig. 6 is a result diagram obtained by applying the switch state detection device of the algorithm of the present invention, and both the training and inspection stages use an analog switch model.

最后所应说明的是，以上实施例仅用以说明本发明的技术方案而非限制，尽管参照较佳实施例对本发明进行了详细说明，本领域的普通技术人员应当理解，可以对本发明的技术方案进行修改或者等同替换，而不脱离本发明技术方案的精神和范围，其均应涵盖在本发明的权利要求范围当中。Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention without limitation. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements of the technical solutions without departing from the spirit and scope of the technical solutions of the present invention shall be covered by the scope of the claims of the present invention.

Claims (4)

Translated fromChinese

1.一种电力开关柜开关状态图像识别系统，其特征在于，所述图像识别系统包括离线训练模块和在线检测模块；1. a power switch cabinet switch state image recognition system, is characterized in that, described image recognition system comprises off-line training module and on-line detection module;所述离线训练模块，基于Adaboost算法对大量包含开关状态的真样本和假样本进行学习，形成一系列的弱分类器，然后依据权重把这些弱分类器级联成若干强分类器，所述强分类器分别代表不同的开关状态；The offline training module, based on the Adaboost algorithm, learns a large number of real samples and fake samples containing switch states to form a series of weak classifiers, and then cascades these weak classifiers into several strong classifiers according to the weight. The classifiers represent different switch states respectively;所述在线检测模块包括图像初步处理模块和状态识别模块；The online detection module includes an image preliminary processing module and a state recognition module;所述图像初步处理模块,用于将采集到的开关图像转换为开关灰度图并去噪，以便状态识别模块进行处理；The image preliminary processing module is used to convert the collected switch image into a switch grayscale image and remove noise, so that the state recognition module can process it;所述状态识别模块,装载了所述强分类器，并利用所述强分类器分别对所述开关灰度图进行扫描搜索，得到开关的位置和状态信息。The state identification module is loaded with the strong classifier, and uses the strong classifier to scan and search the switch grayscale image to obtain the position and state information of the switch.2.根据权利要求1所述的电力开关柜开关状态图像识别系统，其特征在于，在所述离线训练模块中，每个所述特定弱分类器所使用的特征用形状、感兴趣区域中的位置以及比例系数来定义；提取矩形特征时，用积分图像方法来减小计算量。2. the power switchgear switch state image recognition system according to claim 1, is characterized in that, in described off-line training module, the feature used by each described specific weak classifier uses shape, region of interest Position and scale factor to define; when extracting rectangular features, use the integral image method to reduce the amount of calculation.3.根据权利要求1或2所述的电力开关柜开关状态图像识别系统，其特征在于，所述图像初步处理模块包括转化为灰度图单元和图像去噪单元，所述转化灰度图单元将采集到的彩色开关图像转化为灰度图像，所述图像去噪单元对灰度图像进行中值滤波，将灰度中存在的噪音点去除，避免了噪声点对图像识别带来的干扰。3. The power switchgear switch state image recognition system according to claim 1 or 2, wherein the image preliminary processing module includes a grayscale image unit and an image denoising unit, and the grayscale image conversion unit The collected color switch image is converted into a grayscale image, and the image denoising unit performs median filtering on the grayscale image to remove noise points existing in the grayscale, thereby avoiding the interference of noise points on image recognition.4.根据权利要求3所述的电力开关柜开关状态图像识别系统，其特征在于，状态识别模块采用基于Adaboost的目标检测算法，以不同比例大小的扫描窗口对图像进行搜索，能够通过强分类器的即为一个目标，识别单元通过该目标的区域位置和强分类器的种类以确定某个开关的状态。4. the power switchgear switch state image recognition system according to claim 3, is characterized in that, state recognition module adopts the target detection algorithm based on Adaboost, searches for image with the scanning window of different scale size, can pass strong classifier is a target, and the recognition unit determines the state of a certain switch through the area position of the target and the type of the strong classifier.

Images