技术领域Technical field
本发明涉及计算机图像处理方法的技术领域，尤其涉及一种基于红外或紫外图像的电缆接头故障检测方法及系统。The present invention relates to the technical field of computer image processing methods, and in particular to a cable joint fault detection method and system based on infrared or ultraviolet images.
背景技术Background technique
由于电力电缆都是埋藏于地下，随着电缆广泛使用，长年累月运行过程中，其本身的质量问题、系统增容以及受到外界外力的作用发生破损，埋在土壤中，长时间受到泥土中潮气的侵蚀导致电缆附件等绝缘薄弱的环节出现缺陷，最终致使绝缘击穿产生故障。早期过程中由于绝缘老化会发生局部放电，长时间放电最终会导致永久性故障的产生。地下电力电缆一般都敷设于地下，发生故障时，对故障的定位和故障的诊断都具有较大的难度，将耗费大量的人力物力和财力寻找故障点。除此之外，电缆没有架空线路那样具有直接观测性，使得电缆的故障检测与诊断任务要更加艰巨而富有挑战性。由于电力电缆的故障定位和故障诊断是电力系统安全稳定运行的重要保障，因此快速准确定位故障点以及对于电缆故障和故障扰动的有效诊断和识别能够快速分析故障原因，尽快发现绝缘缺陷和隐患，避免故障进一步恶化，显著降低因故障造成的人力、物力、财力的损失。Since power cables are buried underground, as the cables are widely used, they are damaged due to their own quality problems, system capacity increase, and external forces during their long-term operation. They are buried in the soil and are exposed to moisture in the soil for a long time. Corrosion causes defects in weak insulation links such as cable accessories, which ultimately leads to insulation breakdown and failure. Partial discharge will occur due to insulation aging in the early process, and long-term discharge will eventually lead to permanent failure. Underground power cables are generally laid underground. When a fault occurs, it is difficult to locate and diagnose the fault. It will consume a lot of manpower, material and financial resources to find the fault point. In addition, cables are not as directly observable as overhead lines, making cable fault detection and diagnosis tasks more arduous and challenging. Since fault location and fault diagnosis of power cables are important guarantees for the safe and stable operation of the power system, rapid and accurate location of fault points and effective diagnosis and identification of cable faults and fault disturbances can quickly analyze the cause of the fault and discover insulation defects and hidden dangers as soon as possible. Avoid further deterioration of faults and significantly reduce losses in human, material and financial resources caused by faults.
目前，我国对于普通电缆的电力维护一直采用人工定期巡检的方式并进行预防性维护，然而电缆隧道内潮气对电缆接头的侵蚀与机械应力对电缆接头造成的损伤都是缓慢进行的，工作人员尽管定期对隧道内电缆进行巡检，还是无法预料电缆接头何时会发生故障。电缆接头出现故障的主要原因是绝缘老化劣化，从而导致温度上升以及局部放电，严重时甚至会引发电缆隧道内火灾事故。因此，如何在电缆故障之前对电缆状态进行预知和预警是该领域人员亟需解决的问题。At present, our country has always used manual regular inspections and preventive maintenance for the power maintenance of ordinary cables. However, the erosion of cable joints by moisture in the cable tunnel and the damage caused by mechanical stress to the cable joints are carried out slowly. The staff Despite regular inspections of cables in tunnels, it is impossible to predict when a cable joint will fail. The main reason for the failure of cable joints is the aging and deterioration of the insulation, which leads to temperature rise and partial discharge. In severe cases, it may even cause a fire accident in the cable tunnel. Therefore, how to predict and warn cable status before cable failure is an urgent problem that needs to be solved by personnel in this field.
目前，国内外对电缆接头状态检测的方法有：直流分量法、介质损耗角正切值检测法、局部放电监测法等。在实际应用中，电缆屏蔽层与大地之间存在杂散电流，将对直流分量的测量造成极大干扰，从而影响监测结果；在实际运行中电缆金属护层上的容性电流远大于阻性电流，故求得的电缆绝缘介质损耗角正切值很小，实际测量中设备往往难以得到准确的测量值；在实际应用中局放信号监测受传感器性能、现场噪音、电磁干扰、信号衰减等因素影响，采集难度大。At present, domestic and foreign methods for detecting the status of cable joints include: DC component method, dielectric loss angle tangent value detection method, partial discharge monitoring method, etc. In practical applications, there are stray currents between the cable shielding layer and the earth, which will cause great interference to the measurement of the DC component, thereby affecting the monitoring results; in actual operation, the capacitive current on the cable metal sheath is much larger than the resistive current. Current, so the obtained tangent value of cable insulation dielectric loss angle is very small. In actual measurement, it is often difficult for equipment to obtain accurate measurement values. In practical applications, partial discharge signal monitoring is affected by factors such as sensor performance, on-site noise, electromagnetic interference, and signal attenuation. The impact is difficult to collect.
为了准确测量和分析电缆接头在日常工作中存在的缺陷问题，国内外专家研究制造很多检测仪器(包括红外测温仪、紫外成像仪等)应用于电气设备的缺陷检测，对于巡检人员在巡检过程中排查设备的缺陷及诊断故障工作具有重大意义。In order to accurately measure and analyze the defects of cable joints in daily work, domestic and foreign experts have researched and manufactured many detection instruments (including infrared thermometers, ultraviolet imagers, etc.) to be used in defect detection of electrical equipment. For inspection personnel during inspections, It is of great significance to detect equipment defects and diagnose faults during the inspection process.
目前融合人工智能算法的故障诊断技术具备了相关学习能力和实时监测能力，从而实现了对电缆故障的在线诊断。YOLO系列算法是深度学习单阶段算法中的一种，其快速、高效、准确的检测性能受到了广泛关注和应用，能够适应不同尺度、姿态、遮挡等复杂情况下的目标检测任务。At present, fault diagnosis technology integrated with artificial intelligence algorithms has relevant learning capabilities and real-time monitoring capabilities, thus realizing online diagnosis of cable faults. The YOLO series algorithm is one of the deep learning single-stage algorithms. Its fast, efficient, and accurate detection performance has received widespread attention and application, and it can adapt to target detection tasks in complex situations such as different scales, postures, and occlusions.
发明内容Summary of the invention
有鉴于此，有必要提供一种基于红外或紫外图像的电缆接头故障检测方法，用以实时监控电缆接头的故障情况，通过改进YOLOv7算法对红外或紫外图像的电缆接头故障进行检测，提高了故障识别精度，并解决了红外测温易受环境湿度影响的缺点。In view of this, it is necessary to provide a cable joint fault detection method based on infrared or ultraviolet images to monitor the fault conditions of cable joints in real time. By improving the YOLOv7 algorithm, the cable joint faults in infrared or ultraviolet images are detected, the fault identification accuracy is improved, and the disadvantage that infrared temperature measurement is easily affected by environmental humidity is solved.
本发明提供一种基于红外或紫外图像的电缆接头故障检测方法，包括如下步骤：The invention provides a cable joint fault detection method based on infrared or ultraviolet images, which includes the following steps:
(1)通过红外或紫外成像仪采集电缆接头的红外或紫外图像数据作为数据集；(1) Collect infrared or ultraviolet image data of the cable joint as a data set by using an infrared or ultraviolet imager;
(2)对采集的数据集进行基于限制对比度自适应直方图均衡化的增强，利用图像标注软件LabelImg对原始图像进行标注，并按3:1随机划分训练集和测试集；(2) The collected data set is enhanced based on contrast-constrained adaptive histogram equalization, the original image is annotated using the image annotation software LabelImg, and the training set and test set are randomly divided into 3:1;
(3)将训练集输入到改进的YOLOv7模型中进行故障诊断训练，得到训练后的模型；(3) Input the training set into the improved YOLOv7 model for fault diagnosis training to obtain the trained model;
(4)使用训练后的改进YOLOv7模型对测试样本库中的红外或紫外图像进行效果测试和故障诊断，通过比较两类图像的故障重叠区域，进而判断故障的严重程度。(4) Use the improved YOLOv7 model after training to perform effect testing and fault diagnosis on infrared or ultraviolet images in the test sample library, and then determine the severity of the fault by comparing the fault overlapping areas of the two types of images.
进一步的，步骤(2)中对采集的数据集进行基于限制对比度自适应直方图均衡化的增强包括：Further, in step (2), the enhancement of the collected data set based on restricted contrast adaptive histogram equalization includes:
(2.1)分块：将输入图像划分为大小相等的不重叠子块rk，其中k＝0，1，…，L＝1，L为子块数量；(2.1) Blocking: Divide the input image into non-overlapping sub-blocks rk of equal size, where k = 0, 1, ..., L = 1, and L is the number of sub-blocks;
(2.2)计算子块直方图h(rk)；(2.2) Calculate the sub-block histogram h(rk );
(2.3)用剪切阈值剪辑每个子块的直方图：剪切规则为(2.3) Use a clipping threshold to clip the histogram of each sub-block: the clipping rule is
式中，h'(rk)为每个子块的剪辑直方图；Nclip为定义的实际剪切阈值；Navg为要重新分配到每个直方图的像素的平均值；In the formula, h'(rk ) is the clipping histogram of each sub-block;Nclip is the actual clipping threshold defined;Navg is the average value of the pixels to be redistributed to each histogram;
(2.4)像素点重新分配：对于每个子块，使用步骤(2.3)中多余的像素重新分配，直到所有被剪切的像素点分配完毕；(2.4) Pixel reallocation: For each sub-block, use the excess pixels in step (2.3) to reallocate until all the cut pixels are allocated;
(2.5)对经过上述步骤后的图像每个子区域的灰度直方图均衡化处理，使非均匀分布的直方图变成均匀分布；(2.5) Equalize the grayscale histogram of each sub-region of the image after the above steps to make the non-uniformly distributed histogram into a uniform distribution;
(2.6)采用双线性插值方法对像素点灰度值进行重构，以得到的各子块中心像素点的灰度值为参考点，计算最终输出图像中各点的灰度值。(2.6) Use the bilinear interpolation method to reconstruct the gray value of the pixel point, and use the obtained gray value of the center pixel of each sub-block as a reference point to calculate the gray value of each point in the final output image.
进一步的，所述改进的YOLOv7模型对YOLOv7模型的改进具体包括：Further, the improvements of the improved YOLOv7 model to the YOLOv7 model specifically include:
使用轻量级网络MobileOne作为YOLOv7的主干网络；Use lightweight network MobileOne as the backbone network of YOLOv7;
在模型颈部添加全局注意力机制GAM来获取更丰富的跨通道信息，提高模型的特征提取能力；Add a global attention mechanism GAM to the neck of the model to obtain richer cross-channel information and improve the feature extraction capability of the model;
引入Focal-EIoU Loss损失函数，增加算法收敛速率。The Focal-EIoU Loss loss function is introduced to increase the convergence rate of the algorithm.
进一步的，所述轻量级网络MobileOne的核心模块基于MobileNetV1设计，结构与MobileNetV1基本一致，区别是把MobileNet中的深度可分离卷积替换为了神经网络结构块，其左侧部分构成了MobileOne的一个完整结构块，由上下两部分构成，其中上面部分基于深度卷积，下面部分基于点卷积，Act.表示激活函数；深度卷积模块由三条分支构成，最左侧分支是1×1卷积；中间分支是过参数化的3×3卷积，即k个3×3卷积；右侧部分是一个包含BN层的跳跃连接；深度卷积本质上是分组卷积，分组数与通道数相同，这里的1×1卷积和3×3卷积都是深度卷积；点卷积模块由两条分支构成，左侧分支是过参数化的1×1卷积，由k个1×1卷积构成，右侧部分是一个包含BN层的跳跃连接。Furthermore, the core module of the lightweight network MobileOne is designed based on MobileNetV1, and its structure is basically the same as MobileNetV1. The difference is that the depth-separable convolution in MobileNet is replaced by a neural network structural block, and the left part of it constitutes one of MobileOne The complete structure block is composed of upper and lower parts. The upper part is based on depth convolution, and the lower part is based on point convolution. Act. represents the activation function; the depth convolution module consists of three branches, and the leftmost branch is 1×1 convolution. ; The middle branch is an over-parameterized 3×3 convolution, that is, k 3×3 convolutions; the right part is a skip connection containing a BN layer; the depth convolution is essentially a grouped convolution, and the number of groups and the number of channels Same, the 1×1 convolution and 3×3 convolution here are both depth convolutions; the point convolution module consists of two branches, the left branch is an over-parameterized 1×1 convolution, consisting of k 1× 1 convolution, the right part is a skip connection containing a BN layer.
进一步的，所述全局注意力机制GAM由通道注意力子模块和空间注意力子模块构成，该机制在CBAM中的顺序通道-空间注意机制的基础上，对其子模块进行了优化设计；Further, the global attention mechanism GAM is composed of a channel attention sub-module and a spatial attention sub-module. This mechanism optimizes the design of its sub-modules based on the sequential channel-spatial attention mechanism in CBAM;
其中，输入特征由F1表示，将GAM注意力机制中的一系列中间操作定义为中间状态F2，输出状态定义为F3，则三者之间的关系如下式所述：Among them, the input feature is represented by F1 , a series of intermediate operations in the GAM attention mechanism are defined as the intermediate state F2 , and the output state is defined as F3. The relationship between the three is as follows:
进一步的，所述通道注意力子模块的原始输入特征F1维度为C×W×H，通道注意力子模块首先对其进行三维通道置换，将信息保存为W×H×C的形式；利用一个两层的MLP，第一层进行编码操作将通道数C缩减至C/R，第二层进行解码操作以获取与输入特征具有相同通道数的结果。最终，通过对结果进行Sigmoid激活函数处理，得到权重系数Mc，可以有效地扩大跨维通道空间的依赖性。Further, the original input feature F1 dimension of the channel attention sub-module is C×W×H. The channel attention sub-module first performs a three-dimensional channel replacement on it and saves the information in the form of W×H×C; using A two-layer MLP, the first layer performs an encoding operation to reduce the number of channels C to C/R, and the second layer performs a decoding operation to obtain results with the same number of channels as the input features. Finally, by performing Sigmoid activation function processing on the results, the weight coefficient Mc is obtained, which can effectively expand the dependence of the cross-dimensional channel space.
进一步的，所述空间注意力子模块的输入特征F2采用了双重卷积，每个卷积层都使用了7*7的卷积核，以达到空间信息融合的效果；通道注意力子模块根据特征的重要性进行缩减并得到缩放后的新特征；最后，对于这个特征的权重系数Ms，采用激活函数Sigmoid进行处理，以获得更加准确的权重值。Furthermore, the input feature F2 of the spatial attention sub-module adopts double convolution, and each convolution layer uses a 7*7 convolution kernel to achieve the effect of spatial information fusion; the channel attention sub-module The feature is reduced according to its importance and the scaled new feature is obtained; finally, the weight coefficient Ms of this feature is processed using the activation function Sigmoid to obtain a more accurate weight value.
进一步的，所述引入Focal-EIoU Loss损失函数，具体包括：Furthermore, the introduction of the Focal-EIoU Loss loss function specifically includes:
EIoU Loss相较于原始YOLOv7使用的CIoU Loss所做出的改进是将损失函数拆分为重叠面积损失、中心点距离损失及宽高比损失3部分，并且对CIoU Loss中的a和v进行修改，Cw和Ch为包含真实框、预测框的最小框的宽度和高度，EIoU Loss利用和直接计算包围框宽高的真实值，解决CIoU Loss使用宽高比比例所造成的阻碍优化问题以及训练过程中的发散问题；考虑到预测框的回归中存在训练样本不平衡的问题，在输入图像中回归误差小的优质锚框的数量远少于误差大的低质量数量，质量较差的样本会产生过大的梯度影响参数的优化；通过整合EIoU损失函数和Focal损失函数，得到最终的Focal-EIoU Loss的表达式，如式(5)所示：The improvement of EIoU Loss compared to the CIoU Loss used in the original YOLOv7 is to split the loss function into three parts: overlap area loss, center point distance loss and aspect ratio loss, and modify a and v in CIoU Loss. , Cw and Ch are the width and height of the minimum box including the real box and predicted box, EIoU Loss uses and Directly calculate the true value of the width and height of the bounding box to solve the obstacle optimization problem caused by CIoU Loss's use of aspect ratio and the divergence problem during the training process; considering the imbalance of training samples in the regression of the prediction box, in the input image The number of high-quality anchor boxes with small regression errors is much less than the number of low-quality anchor boxes with large errors. Poor-quality samples will produce excessive gradients that affect the optimization of parameters; by integrating the EIoU loss function and the Focal loss function, the final Focal is obtained -The expression of EIoU Loss, as shown in equation (5):
LFocal-EIoU＝IoUγLEIoU (5)LFocal-EIoU = IoUγ LEIoU (5)
式中LIoU、Ldis、Lasp分别表示重叠面积损失、中心点距离损失及宽高比损失。In the formula, LIoU , Ldis , and Lasp represent the overlapping area loss, center point distance loss, and aspect ratio loss respectively.
进一步的，步骤(4)中的故障诊断包括：Further, the fault diagnosis in step (4) includes:
通过计算已识别的电缆接头区域与发热区域以及光斑区域的重叠度，来诊断待检测的电缆接头是否存在故障；重叠度指的是YOLOv7模型产生的电缆接头目标窗口与发热区域目标窗口以及光斑区域目标窗口的交叠率，即二者检测框选区域的交集比上它们的并集，即为故障诊断的准确率重叠度，对于故障诊断而言，当重叠度低于一定值X时，可认为电力设备存在异常发热以及异常放电的可能性较小，此时通常需要进行二次人工诊断；当重叠度为0时，可基本判断电缆接头上不存在异常区域，此时无需进行二次诊断；当重叠度大于一定值是，可判断电缆接头存在异常区域。Diagnose whether there is a fault in the cable joint to be detected by calculating the overlap between the identified cable joint area, the heating area and the light spot area; the overlap refers to the cable joint target window, the heating area target window and the light spot area generated by the YOLOv7 model The overlap rate of the target window, that is, the intersection of the two detection frame selection areas is greater than their union, is the accuracy of fault diagnosis. For fault diagnosis, when the overlap is lower than a certain value It is considered that the possibility of abnormal heating and abnormal discharge in power equipment is small, and a second manual diagnosis is usually required at this time; when the overlap degree is 0, it can be basically judged that there is no abnormal area on the cable joint, and there is no need to perform a second diagnosis at this time. ; When the overlap is greater than a certain value, it can be determined that there is an abnormal area in the cable joint.
一种基于红外或紫外图像的电缆接头故障检测系统，包括处理器以及存储器，存储器上存储有计算机程序，计算机程序被处理器执行时，实现所述的基于红外或紫外图像的电缆接头故障检测方法。A cable joint fault detection system based on infrared or ultraviolet images comprises a processor and a memory. A computer program is stored in the memory. When the computer program is executed by the processor, the cable joint fault detection method based on infrared or ultraviolet images is implemented.
与现有技术相比，本发明的有益效果包括：Compared with the prior art, the beneficial effects of the present invention include:
由于红外或紫外成像仪的成像效果同时受到拍摄对象材料发射率、背景与目标相似度以及探测距离等因素影响，因此相较于可见光图像，红外或紫外图像质量较低，主要表现在对比度较低、细节分辨率较差以及信噪比较低，因此，有必要先对红外或紫外图像进行基于限制对比度自适应直方图均衡化的图像增强，有效提高了图像中物体轮廓的清晰度，方便后续的数据集标注和物体识别精确率；YOLOv7主干网络的通道数设置较大，其模型复杂度往往较高，为了使模型复杂度降低，使算法更适合完成电缆接头的图像检测任务，本发明使用了轻量级网络MobileOne作为YOLOv7的主干网络以加快识别速度达到实时检测的要求；为了减少网络信息缩减并放大全局维度交互特征，在模型颈部添加了GAM注意力机制来获取更丰富的跨通道信息，提高模型的特征提取能力；为了增加算法收敛速率，提高原算法性能和检测精度，采用Focal-EIoU Loss损失函数来代替CIoU Loss损失函数。GAM注意力机制和Focal-EIoU Loss损失函数进一步增加了目标识别准确度，MobileOne主干网络则提高了检测速度，三者相互配合，使综合性能相较于原算法更加优越，并由此提出了一种轻量级YOLOv7网络。Since the imaging effect of an infrared or ultraviolet imager is also affected by factors such as the emissivity of the object material, the similarity between the background and the target, and the detection distance, compared with visible light images, the quality of infrared or ultraviolet images is lower, mainly due to lower contrast. , poor detail resolution and low signal-to-noise ratio. Therefore, it is necessary to first perform image enhancement based on limited contrast adaptive histogram equalization on infrared or ultraviolet images, which effectively improves the clarity of object contours in the image and facilitates follow-up. Data set annotation and object recognition accuracy; the channel number of the YOLOv7 backbone network is set to be larger, and its model complexity is often higher. In order to reduce the model complexity and make the algorithm more suitable for completing the image detection task of cable joints, the present invention uses The lightweight network MobileOne is used as the backbone network of YOLOv7 to speed up the recognition speed and meet the requirements of real-time detection; in order to reduce the reduction of network information and amplify the global dimension interaction characteristics, a GAM attention mechanism is added to the neck of the model to obtain richer cross-channel information. information to improve the feature extraction capability of the model; in order to increase the convergence rate of the algorithm and improve the performance and detection accuracy of the original algorithm, the Focal-EIoU Loss loss function is used to replace the CIoU Loss loss function. The GAM attention mechanism and Focal-EIoU Loss loss function further increase the accuracy of target recognition, and the MobileOne backbone network improves the detection speed. The three cooperate with each other to make the comprehensive performance more superior than the original algorithm, and thus propose a A lightweight YOLOv7 network.
附图说明Description of drawings
图1为本发明提供的基于红外或紫外图像的电缆接头故障检测方法流程框图；Figure 1 is a flow chart of a cable joint fault detection method based on infrared or ultraviolet images provided by the present invention;
图2为本发明提供的MobileOne神经网络结构块示意图；FIG2 is a schematic diagram of a MobileOne neural network structure block provided by the present invention;
图3为本发明提供的GAM注意力机制示意图；Figure 3 is a schematic diagram of the GAM attention mechanism provided by the present invention;
图4为本发明提供的通道注意力示意图；FIG4 is a schematic diagram of channel attention provided by the present invention;
图5为本发明提供的空间注意力示意图；Figure 5 is a schematic diagram of spatial attention provided by the present invention;
图6为本发明提供的故障诊断流程图。Figure 6 is a fault diagnosis flow chart provided by the present invention.
具体实施方式Detailed ways
下面结合具体实施方式对本发明进行阐述。The present invention will be described below in conjunction with specific embodiments.
本发明实施例提供了一种基于红外或紫外图像的电缆接头故障检测方法，结合图1来看，图1为本发明提供的基于红外或紫外图像的电缆接头故障检测方法流程框图，包括步骤(1)至步骤(4)。Embodiments of the present invention provide a cable joint fault detection method based on infrared or ultraviolet images. Viewed in conjunction with Figure 1, Figure 1 is a flow chart of a cable joint fault detection method based on infrared or ultraviolet images provided by the present invention, including steps ( 1) Go to step (4).
(1)通过红外或紫外成像仪采集电缆接头的红外或紫外图像数据；(1) Collect infrared or ultraviolet image data of cable joints through infrared or ultraviolet imagers;
(2)对采集的数据集进行基于限制对比度自适应直方图均衡化的增强，利用图像标注软件LabelImg对原始图像进行标注，并按3:1随机划分训练集和测试集；(2) Enhance the collected data set based on limited contrast adaptive histogram equalization, use the image annotation software LabelImg to annotate the original images, and randomly divide the training set and the test set at a ratio of 3:1;
基于限制对比度自适应直方图均衡化的图像增强的步骤包括：The steps of image enhancement based on restricted contrast adaptive histogram equalization include:
分块：将输入图像划分为大小相等的不重叠子块rk(k＝0，1，…，L＝1，L为子块数量)；Blocking: Divide the input image into non-overlapping sub-blocks rk of equal size (k=0, 1,..., L=1, L is the number of sub-blocks);
计算子块直方图h(rk)；Calculate the sub-block histogram h(rk );
用剪切阈值剪辑每个子块的直方图：剪切规则为Clip the histogram of each sub-block with a clipping threshold: the clipping rule is
式中，h'(rk)为每个子块的剪辑直方图；Nclip为定义的实际剪切阈值；Navg为要重新分配到每个直方图的像素的平均值；In the formula, h'(rk ) is the clipping histogram of each sub-block;Nclip is the actual clipping threshold defined;Navg is the average value of the pixels to be redistributed to each histogram;
像素点重新分配：对于每个子块，使用步骤(3)中多余的像素重新分配，直到所有被剪切的像素点分配完毕；Pixel reallocation: For each sub-block, use the excess pixels in step (3) to reallocate until all clipped pixels are allocated;
对经过上述步骤后的图像每个子区域的灰度直方图均衡化处理，即使非均匀分布的直方图变成均匀分布；Equalize the grayscale histogram of each sub-region of the image after the above steps, so that the non-uniformly distributed histogram becomes a uniform distribution;
采用双线性插值方法对像素点灰度值进行重构，以得到的各子块中心像素点的灰度值为参考点，计算最终输出图像中各点的灰度值。The bilinear interpolation method is used to reconstruct the gray value of the pixel point, and the obtained gray value of the center pixel of each sub-block is used as the reference point to calculate the gray value of each point in the final output image.
与一般的直方图均衡化相比，自适应直方图均衡化先计算多个直方图，每一个直方图对应图像的一个部分，然后使用它们重新分配图像亮度来改善图像质量，但其往往会因为图像近恒定区域中的直方图高度集中而放大图像近恒定区域中的对比度，导致噪声在近恒定区域中被放大，而限制对比度自适应直方图均衡化会对对比度放大进行限制，从而减小了噪声放大的问题。因此，使用限制对比度自适应直方图均衡化对图像进行增强，可以提高图像中物体轮廓的清晰度，方便数据集标注。Compared with general histogram equalization, adaptive histogram equalization first calculates multiple histograms, each of which corresponds to a part of the image, and then uses them to redistribute the image brightness to improve image quality. However, it often amplifies the contrast in the near-constant area of the image because the histogram in the near-constant area is highly concentrated, causing the noise to be amplified in the near-constant area. Contrast-limited adaptive histogram equalization limits the contrast amplification, thereby reducing the problem of noise amplification. Therefore, using contrast-limited adaptive histogram equalization to enhance the image can improve the clarity of the object contours in the image and facilitate data set annotation.
(3)将训练集输入到改进的YOLOv7模型中进行故障诊断训练，得到训练后的模型；(3) Input the training set into the improved YOLOv7 model for fault diagnosis training to obtain the trained model;
(4)使用训练后的改进YOLOv7模型对测试样本库中的红外或紫外图像进行效果测试和故障诊断，通过比较两类图像的故障重叠区域，进而判断故障的严重程度。(4) Use the improved YOLOv7 model after training to perform effect testing and fault diagnosis on infrared or ultraviolet images in the test sample library, and then determine the severity of the fault by comparing the fault overlapping areas of the two types of images.
步骤(3)、(4)中YOLOv7模型的改进点包括：The improvements of the YOLOv7 model in steps (3) and (4) include:
使用轻量级网络MobileOne作为YOLOv7的主干网络；Use lightweight network MobileOne as the backbone network of YOLOv7;
在模型颈部添加GAM注意力机制来获取更丰富的跨通道信息，提高模型的特征提取能力；Add the GAM attention mechanism at the neck of the model to obtain richer cross-channel information and improve the model's feature extraction capability;
引入Focal-EIoU Loss损失函数，增加算法收敛速率。The Focal-EIoU Loss loss function is introduced to increase the convergence rate of the algorithm.
结合图2来看，轻量级网络MobileOne包括：Combined with Figure 2, the lightweight network MobileOne includes:
MobileOne的核心模块基于MobileNetV1设计，结构与MobileNetV1基本一致，区别是把MobileNet中的深度可分离卷积替换为了神经网络结构块，其左侧部分构成了MobileOne的一个完整结构块，由上下两部分构成，其中上面部分基于深度卷积，下面部分基于点卷积，Act.表示激活函数；深度卷积模块由三条分支构成，最左侧分支是1×1卷积；中间分支是过参数化的3×3卷积，即k个3×3卷积；右侧部分是一个包含BN层的跳跃连接；深度卷积本质上是分组卷积，分组数与通道数相同，这里的1×1卷积和3×3卷积都是深度卷积；点卷积模块由两条分支构成，左侧分支是过参数化的1×1卷积，由k个1×1卷积构成，右侧部分是一个包含BN层的跳跃连接。The core module of MobileOne is designed based on MobileNetV1, and its structure is basically the same as MobileNetV1. The difference is that the depth-separable convolution in MobileNet is replaced by a neural network structural block. The left part constitutes a complete structural block of MobileOne, consisting of upper and lower parts. , the upper part is based on depth convolution, the lower part is based on point convolution, Act. represents the activation function; the depth convolution module consists of three branches, the leftmost branch is 1×1 convolution; the middle branch is over-parameterized 3 ×3 convolution, that is, k 3×3 convolutions; the right part is a skip connection containing a BN layer; depth convolution is essentially a grouped convolution, and the number of groups is the same as the number of channels. The 1×1 convolution here and 3×3 convolution are both depth convolutions; the point convolution module consists of two branches. The left branch is an over-parameterized 1×1 convolution, consisting of k 1×1 convolutions, and the right part is A skip connection containing BN layers.
结合图3来看，全局注意力机制GAM包括：Combined with Figure 3, the global attention mechanism GAM includes:
全局注意力机制(Global Attention Mechanism，GAM)，可以起到减少网络信息缩减并放大全局维度交互特征的作用。全局注意力机制由通道注意力子模块和空间注意力子模块构成。该机制在CBAM中的顺序通道-空间注意机制的基础上，对其子模块进行了优化设计；The Global Attention Mechanism (GAM) can reduce the reduction of network information and amplify the global dimension interaction features. The global attention mechanism consists of channel attention sub-module and spatial attention sub-module. This mechanism is based on the sequential channel-spatial attention mechanism in CBAM and optimizes the design of its sub-modules;
其中，输入特征由F1表示，将GAM注意力机制中的一系列中间操作定义为中间状态F2，输出状态定义为F3，则三者之间的关系如下式所述：Among them, the input feature is represented byF1 , a series of intermediate operations in the GAM attention mechanism is defined as the intermediate stateF2 , and the output state is defined asF3 . The relationship between the three is as follows:
结合图4来看，通道注意力子模块包括：Combined with Figure 4, the channel attention sub-module includes:
原始输入特征F1维度为C×W×H，通道注意力子模块首先对其进行三维通道置换，将信息保存为W×H×C的形式；利用一个两层的MLP，第一层进行编码操作将通道数C缩减至C/R，第二层进行解码操作以获取与输入特征具有相同通道数的结果。最终，通过对结果进行Sigmoid激活函数处理，得到权重系数Mc，可以有效地扩大跨维通道空间的依赖性。The original input featureF1 has a dimension of C×W×H. The channel attention submodule first performs a three-dimensional channel replacement on it and saves the information in the form of W×H×C; using a two-layer MLP, the first layer performs encoding The operation reduces the channel number C to C/R, and the second layer performs a decoding operation to obtain the result with the same channel number as the input feature. Finally, by performing Sigmoid activation function processing on the results, the weight coefficient Mc is obtained, which can effectively expand the dependence of the cross-dimensional channel space.
结合图5来看，空间注意力子模块包括：Combined with Figure 5, the spatial attention submodule includes:
输入特征F2采用了双重卷积，每个卷积层都使用了7*7的卷积核，以达到空间信息融合的效果。通道注意力子模块根据特征的重要性进行缩减并得到缩放后的新特征。最后，对于这个特征的权重系数Ms，采用激活函数Sigmoid进行处理，以获得更加准确的权重值。The input feature F2 uses double convolution, and each convolution layer uses a 7*7 convolution kernel to achieve the effect of spatial information fusion. The channel attention sub-module reduces the features according to their importance and obtains scaled new features. Finally, the weight coefficient Ms of this feature is processed using the activation function Sigmoid to obtain a more accurate weight value.
Focal-EIoU Loss损失函数包括：Focal-EIoU Loss loss functions include:
原始YOLOv7使用CIoU作为坐标损失函数，CIoU损失考虑了重叠面积、中心点距离已经纵横比这3个几何因素。然而在目标识别和检测过程中，边界框是决定其精确度的重要因素之一，原始YOLOv7算法的Loss函数并未考虑到，因此引入Focal-EIoU Loss函数来代替原始损失函数。The original YOLOv7 uses CIoU as the coordinate loss function. CIoU loss takes into account the three geometric factors of overlapping area, center point distance and aspect ratio. However, in the process of target recognition and detection, the bounding box is one of the important factors that determine its accuracy. The Loss function of the original YOLOv7 algorithm was not taken into account, so the Focal-EIoU Loss function was introduced to replace the original loss function.
EIoU Loss相较于CIoU Loss所做出的改进是将损失函数拆分为重叠面积损失、中心点距离损失及宽高比损失3部分，并且对CIoU Loss中的a和v进行了修改，Cw和Ch为可包含真实框、预测框的最小框的宽度和高度，EIoU Loss利用和/>直接计算包围框宽高的真实值，这样便解决了CIoU Loss使用宽高比比例所造成的阻碍优化问题，还能解决训练过程中的发散问题。此外，考虑到预测框的回归中存在训练样本不平衡的问题，即在输入图像中回归误差小的优质锚框的数量远少于误差大的低质量数量，质量较差的样本会产生过大的梯度影响参数的优化，又借鉴了解决正负样本不平衡的Focal Loss，将偏差大的地方设置更大的梯度优化，以便于关注对难样本的检测，降低质量差的样本对算法性能的影响。通过整合EIoU损失函数和Focal损失函数，得到最终的Focal-EIoU Loss的表达式，如式(5)所示：The improvement made by EIoU Loss compared to CIoU Loss is to split the loss function into three parts: overlap area loss, center point distance loss and aspect ratio loss, and modify a and v in CIoU Loss, Cw and Ch are the width and height of the minimum box that can include the real box and the predicted box. EIoU Loss uses and/> Directly calculating the true value of the width and height of the bounding box solves the problem of hindering optimization caused by CIoU Loss's use of aspect ratio, and also solves the divergence problem during the training process. In addition, considering the problem of training sample imbalance in the regression of prediction boxes, that is, the number of high-quality anchor boxes with small regression errors in the input image is far less than the number of low-quality anchor boxes with large errors, and poor-quality samples will generate excessive The optimization of gradient influence parameters also draws on Focal Loss, which solves the imbalance of positive and negative samples, and sets larger gradient optimization in places with large deviations, so as to focus on the detection of difficult samples and reduce the impact of poor quality samples on algorithm performance. Influence. By integrating the EIoU loss function and the Focal loss function, the final expression of Focal-EIoU Loss is obtained, as shown in Equation (5):
LFocal-EIoU＝IoUγLEIoU (5)LFocal-EIoU =IoUγ LEIoU (5)
式中LIoU、Ldis、Lasp分别表示重叠面积损失、中心点距离损失及宽高比损失。Where LIoU , Ldis , andLasp represent the overlap area loss, center point distance loss, and aspect ratio loss, respectively.
Focal-EIoU Loss函数不仅包含了CIoU的有益特性，还聚焦于高质量的边界框，提高模型检测精度同时，还可以加快模型收敛。The Focal-EIoU Loss function not only includes the beneficial characteristics of CIoU, but also focuses on high-quality bounding boxes, improving model detection accuracy and speeding up model convergence.
结合图6来看，步骤(4)中故障诊断包括：In conjunction with FIG6 , the fault diagnosis in step (4) includes:
首先提取目标检测框的坐标，将电缆接头和发热区域、光斑区域每一行的坐标信息分三类保存到矩阵中；再提取三类矩阵的坐标求重叠度，并对应每一类电缆接头故障类型和正常情况进行坐标文件的自动归类，最终得到故障诊断效果。First, extract the coordinates of the target detection frame, and save the coordinate information of each row of the cable joint, heating area, and light spot area into three categories into a matrix; then extract the coordinates of the three types of matrices to calculate the degree of overlap, and correspond to each type of cable joint fault type Automatically classify coordinate files according to normal conditions, and finally obtain fault diagnosis results.
本实施例中数据集的具体情况如下：The specific situation of the data set in this embodiment is as follows:
本文收集了2240张220kV某隧道的电缆接头红外或紫外图像，将整理收集的图片批量做去噪处理，对采集的数据集进行基于限制对比度自适应直方图均衡化的增强，利用图像标注软件LabelImg对原始图像进行标注，并按3:1随机划分训练集和测试集。This paper collects 2240 infrared or ultraviolet images of cable joints in a 220kV tunnel, performs batch denoising on the collected images, enhances the collected data set based on contrast-limited adaptive histogram equalization, annotates the original images using image annotation software LabelImg, and randomly divides the training set and test set into 3:1 ratios.
训练参数及评价指标：Training parameters and evaluation indicators:
对于训练参数，初始学习率设置为0.005，动量项为0.9，权重衰减正则项为0.0005，Batch size设置为8。For training parameters, the initial learning rate is set to 0.005, the momentum term is 0.9, the weight decay regularization term is 0.0005, and the batch size is set to 8.
为了验证基于改进YOLOv7的电缆接头目标检测框架的实际识别效果，在测试集上通过准确率、召回率以及平均精度指标。AP(averageprecision)指标。计算公式如下：In order to verify the actual recognition effect of the cable joint target detection framework based on the improved YOLOv7, the accuracy, recall and average precision indicators were passed on the test set. AP (average precision) indicator. Calculated as follows:
式中，P为网络对正样本预测的准确程度，R为网络对正样本的检索效果，TP为电缆接头被判别准确的数量，FP是非电缆接头误判为电缆接头的数量，FN则是将电缆接头误判为非电缆接头的数量。Where P is the accuracy of the network's prediction of positive samples, R is the network's retrieval effect on positive samples, TP is the number of cable joints accurately identified, FP is the number of non-cable joints misidentified as cable joints, and FN is the number of cable joints misidentified as non-cable joints.
采用消融实验法则对改进的结构在模型性能中的权重占比进行探究。上述实验结果如表1所示。The ablation experimental rules are used to explore the weight proportion of the improved structure in model performance. The above experimental results are shown in Table 1.
表1改进YOLOv7算法的消融实验结果Table 1 Ablation experimental results of the improved YOLOv7 algorithm
从上表可知，改进后的YOLOv7算法的准确率AP为94.66％，与YOLOv7原始模型相比提高了8.05％。并且，每个模块的添加都起着正向的效果。由此可见，本发明能够达到较为理想的检测效果。As can be seen from the table above, the accuracy AP of the improved YOLOv7 algorithm is 94.66%, which is 8.05% higher than the original YOLOv7 model. Moreover, the addition of each module has a positive effect. It can be seen that the present invention can achieve a relatively ideal detection effect.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311713768.9ACN117809083B (en) | 2023-12-14 | 2023-12-14 | Cable joint fault detection method and system based on infrared or ultraviolet images |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311713768.9ACN117809083B (en) | 2023-12-14 | 2023-12-14 | Cable joint fault detection method and system based on infrared or ultraviolet images |
| Publication Number | Publication Date |
|---|---|
| CN117809083Atrue CN117809083A (en) | 2024-04-02 |
| CN117809083B CN117809083B (en) | 2024-08-30 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311713768.9AActiveCN117809083B (en) | 2023-12-14 | 2023-12-14 | Cable joint fault detection method and system based on infrared or ultraviolet images |
| Country | Link |
|---|---|
| CN (1) | CN117809083B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118657917A (en)* | 2024-08-19 | 2024-09-17 | 山东济矿鲁能煤电股份有限公司阳城煤矿 | Image enhancement method for low voltage leakage protection in coal mines |
| CN119846403A (en)* | 2024-12-19 | 2025-04-18 | 北京松岛菱电电力工程有限公司 | Power cable diagnosis method, equipment, medium and product |
| CN120275789A (en)* | 2025-06-10 | 2025-07-08 | 中国南方电网有限责任公司超高压输电公司电力科研院 | Infrared-ultraviolet bimodal partial discharge detection method based on lightweight network |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105425123A (en)* | 2015-11-20 | 2016-03-23 | 国网福建省电力有限公司泉州供电公司 | Method and system for collaboratively detecting power equipment failure through ultraviolet imaging and infrared imaging |
| CN109785289A (en)* | 2018-12-18 | 2019-05-21 | 中国科学院深圳先进技术研究院 | A kind of transmission line of electricity defect inspection method, system and electronic equipment |
| CN110795991A (en)* | 2019-09-11 | 2020-02-14 | 西安科技大学 | Mining locomotive pedestrian detection method based on multi-information fusion |
| CN111948501A (en)* | 2020-08-05 | 2020-11-17 | 广州市赛皓达智能科技有限公司 | Automatic inspection equipment for power grid operation |
| CN112184601A (en)* | 2020-09-09 | 2021-01-05 | 中国计量大学 | A method of enhancing vein images under near-infrared light source by using improved CLAHE algorithm |
| CN112233073A (en)* | 2020-09-30 | 2021-01-15 | 国网山西省电力公司大同供电公司 | A real-time detection method for abnormal infrared thermal imaging of substation equipment |
| US20210020360A1 (en)* | 2019-07-15 | 2021-01-21 | Wuhan University | Internal thermal fault diagnosis method of oil-immersed transformer based on deep convolutional neural network and image segmentation |
| CN112380952A (en)* | 2020-11-10 | 2021-02-19 | 广西大学 | Power equipment infrared image real-time detection and identification method based on artificial intelligence |
| CN112381784A (en)* | 2020-11-12 | 2021-02-19 | 国网浙江省电力有限公司信息通信分公司 | Equipment detecting system based on multispectral image |
| CN112562255A (en)* | 2020-12-03 | 2021-03-26 | 国家电网有限公司 | Intelligent image detection method for cable channel smoke and fire condition in low-light-level environment |
| US20210174149A1 (en)* | 2018-11-20 | 2021-06-10 | Xidian University | Feature fusion and dense connection-based method for infrared plane object detection |
| WO2021139069A1 (en)* | 2020-01-09 | 2021-07-15 | 南京信息工程大学 | General target detection method for adaptive attention guidance mechanism |
| CN113159334A (en)* | 2021-02-24 | 2021-07-23 | 广西大学 | Electrical equipment infrared image real-time detection and diagnosis method based on light-weight deep learning |
| US11194997B1 (en)* | 2020-08-04 | 2021-12-07 | Nanjing Huatu Information Technology Co., Ltd. | Method and system for thermal infrared facial recognition |
| CN114264915A (en)* | 2021-12-13 | 2022-04-01 | 国网甘肃省电力公司临夏供电公司 | Power distribution network cable joint operation condition assessment early warning device and method |
| CN115546565A (en)* | 2022-11-09 | 2022-12-30 | 上海朗驰佰特智能技术有限公司 | Oil Leakage Detection Method Based on Pipeline Oil Leakage in Key Areas of YOLOCBF Power Plant |
| US20230039196A1 (en)* | 2021-08-09 | 2023-02-09 | The United States Of America, As Represented By The Secretary Of The Navy | Small unmanned aerial systems detection and classification using multi-modal deep neural networks |
| CN115909093A (en)* | 2022-10-21 | 2023-04-04 | 前郭富汇风能有限公司 | Power equipment fault detection method based on unmanned aerial vehicle inspection and infrared image semantic segmentation |
| US20230186439A1 (en)* | 2021-06-28 | 2023-06-15 | Zhejiang Gongshang University | Lane detection method integratedly using image enhancement and deep convolutional neural network |
| CN116434119A (en)* | 2023-04-21 | 2023-07-14 | 河北工程大学 | A method and system for detecting targets in mine tunnels |
| CN116758393A (en)* | 2023-05-24 | 2023-09-15 | 淮阴工学院 | A flame detection method based on MPGD-YOLO network |
| CN116824341A (en)* | 2023-07-03 | 2023-09-29 | 海南电网有限责任公司电力科学研究院 | An improved insulator abnormal temperature rise detection method based on YOLOv7 |
| CN116883873A (en)* | 2023-07-12 | 2023-10-13 | 哈尔滨工业大学 | An infrared small target detection method for air and ground applications |
| CN117078695A (en)* | 2023-08-18 | 2023-11-17 | 内蒙古工业大学 | A deep learning-based carotid artery plaque ultrasound image recognition and segmentation method |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105425123A (en)* | 2015-11-20 | 2016-03-23 | 国网福建省电力有限公司泉州供电公司 | Method and system for collaboratively detecting power equipment failure through ultraviolet imaging and infrared imaging |
| US20210174149A1 (en)* | 2018-11-20 | 2021-06-10 | Xidian University | Feature fusion and dense connection-based method for infrared plane object detection |
| CN109785289A (en)* | 2018-12-18 | 2019-05-21 | 中国科学院深圳先进技术研究院 | A kind of transmission line of electricity defect inspection method, system and electronic equipment |
| US20210020360A1 (en)* | 2019-07-15 | 2021-01-21 | Wuhan University | Internal thermal fault diagnosis method of oil-immersed transformer based on deep convolutional neural network and image segmentation |
| CN110795991A (en)* | 2019-09-11 | 2020-02-14 | 西安科技大学 | Mining locomotive pedestrian detection method based on multi-information fusion |
| WO2021139069A1 (en)* | 2020-01-09 | 2021-07-15 | 南京信息工程大学 | General target detection method for adaptive attention guidance mechanism |
| US11194997B1 (en)* | 2020-08-04 | 2021-12-07 | Nanjing Huatu Information Technology Co., Ltd. | Method and system for thermal infrared facial recognition |
| CN111948501A (en)* | 2020-08-05 | 2020-11-17 | 广州市赛皓达智能科技有限公司 | Automatic inspection equipment for power grid operation |
| CN112184601A (en)* | 2020-09-09 | 2021-01-05 | 中国计量大学 | A method of enhancing vein images under near-infrared light source by using improved CLAHE algorithm |
| CN112233073A (en)* | 2020-09-30 | 2021-01-15 | 国网山西省电力公司大同供电公司 | A real-time detection method for abnormal infrared thermal imaging of substation equipment |
| CN112380952A (en)* | 2020-11-10 | 2021-02-19 | 广西大学 | Power equipment infrared image real-time detection and identification method based on artificial intelligence |
| CN112381784A (en)* | 2020-11-12 | 2021-02-19 | 国网浙江省电力有限公司信息通信分公司 | Equipment detecting system based on multispectral image |
| CN112562255A (en)* | 2020-12-03 | 2021-03-26 | 国家电网有限公司 | Intelligent image detection method for cable channel smoke and fire condition in low-light-level environment |
| CN113159334A (en)* | 2021-02-24 | 2021-07-23 | 广西大学 | Electrical equipment infrared image real-time detection and diagnosis method based on light-weight deep learning |
| US20230186439A1 (en)* | 2021-06-28 | 2023-06-15 | Zhejiang Gongshang University | Lane detection method integratedly using image enhancement and deep convolutional neural network |
| US20230039196A1 (en)* | 2021-08-09 | 2023-02-09 | The United States Of America, As Represented By The Secretary Of The Navy | Small unmanned aerial systems detection and classification using multi-modal deep neural networks |
| CN114264915A (en)* | 2021-12-13 | 2022-04-01 | 国网甘肃省电力公司临夏供电公司 | Power distribution network cable joint operation condition assessment early warning device and method |
| CN115909093A (en)* | 2022-10-21 | 2023-04-04 | 前郭富汇风能有限公司 | Power equipment fault detection method based on unmanned aerial vehicle inspection and infrared image semantic segmentation |
| CN115546565A (en)* | 2022-11-09 | 2022-12-30 | 上海朗驰佰特智能技术有限公司 | Oil Leakage Detection Method Based on Pipeline Oil Leakage in Key Areas of YOLOCBF Power Plant |
| CN116434119A (en)* | 2023-04-21 | 2023-07-14 | 河北工程大学 | A method and system for detecting targets in mine tunnels |
| CN116758393A (en)* | 2023-05-24 | 2023-09-15 | 淮阴工学院 | A flame detection method based on MPGD-YOLO network |
| CN116824341A (en)* | 2023-07-03 | 2023-09-29 | 海南电网有限责任公司电力科学研究院 | An improved insulator abnormal temperature rise detection method based on YOLOv7 |
| CN116883873A (en)* | 2023-07-12 | 2023-10-13 | 哈尔滨工业大学 | An infrared small target detection method for air and ground applications |
| CN117078695A (en)* | 2023-08-18 | 2023-11-17 | 内蒙古工业大学 | A deep learning-based carotid artery plaque ultrasound image recognition and segmentation method |
| Title |
|---|
| 李庆忠;李宜兵;牛炯;: "基于改进YOLO和迁移学习的水下鱼类目标实时检测", 模式识别与人工智能, no. 03, 15 March 2019 (2019-03-15)* |
| 武建华;梁利辉;纪欣欣;刘云鹏;裴少通;: "基于YOLOv3算法的绝缘子红外图像故障检测方法", 广东电力, no. 09, 30 September 2020 (2020-09-30)* |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118657917A (en)* | 2024-08-19 | 2024-09-17 | 山东济矿鲁能煤电股份有限公司阳城煤矿 | Image enhancement method for low voltage leakage protection in coal mines |
| CN118657917B (en)* | 2024-08-19 | 2024-11-08 | 山东济矿鲁能煤电股份有限公司阳城煤矿 | Image enhancement method for low voltage leakage protection in coal mines |
| CN119846403A (en)* | 2024-12-19 | 2025-04-18 | 北京松岛菱电电力工程有限公司 | Power cable diagnosis method, equipment, medium and product |
| CN120275789A (en)* | 2025-06-10 | 2025-07-08 | 中国南方电网有限责任公司超高压输电公司电力科研院 | Infrared-ultraviolet bimodal partial discharge detection method based on lightweight network |
| CN120275789B (en)* | 2025-06-10 | 2025-09-05 | 中国南方电网有限责任公司超高压输电公司电力科研院 | Infrared-ultraviolet bimodal partial discharge detection method based on lightweight network |
| Publication number | Publication date |
|---|---|
| CN117809083B (en) | 2024-08-30 |
| Publication | Publication Date | Title |
|---|---|---|
| CN117809083A (en) | Cable joint fault detection method and system based on infrared or ultraviolet images | |
| CN108711148B (en) | An intelligent detection method for tire defects based on deep learning | |
| CN109872303B (en) | Surface defect visual inspection method, device and electronic equipment | |
| CN113569756B (en) | Abnormal behavior detection and location method, system, terminal equipment and readable storage medium | |
| CN117541535A (en) | A transmission line inspection image detection method based on deep convolutional neural network | |
| CN116152227A (en) | Deep learning-based commutator inner side image defect target detection method | |
| CN120088240B (en) | Improved steel surface defect detection method and device based on YOLO11n | |
| CN114897855B (en) | Method for determining defect type based on gray value distribution of X-ray images | |
| CN118052760B (en) | An automated method for detecting the thickness of ice-covered transmission lines | |
| CN115359064A (en) | Industrial defect detection method and device | |
| CN117274212A (en) | Bridge underwater structure crack detection method | |
| CN118212546A (en) | Power tower remote sensing target detection method based on large kernel selection feature fusion network | |
| CN116934696A (en) | Industrial PCB defect detection method and device based on YOLOv7-Tiny model improvement | |
| CN116824341A (en) | An improved insulator abnormal temperature rise detection method based on YOLOv7 | |
| Bai et al. | A vision-based nondestructive detection network for rail surface defects | |
| Ramlal et al. | Toward automated utility pole condition monitoring: A deep learning approach | |
| CN119090861A (en) | A defect target detection method for substation equipment based on multi-scale features of generated images | |
| CN118710990A (en) | A classification and processing method for power line defect images based on drone images | |
| CN113256613A (en) | Pipeline damage data feature recognition model and method | |
| CN118521568A (en) | Nut appearance defect detection method and device based on neural network | |
| CN115830012B (en) | Method for detecting and analyzing clue damage data of overhead contact system | |
| CN116773678A (en) | Intelligent maintenance method and system for subway rail | |
| CN112070730A (en) | Anti-vibration hammer falling detection method based on power transmission line inspection image | |
| CN118096689A (en) | Inhaul cable surface image defect positioning and self-adaptive classifying method | |
| Li et al. | ADMA-YOLO: A Lightweight Strip Steel Surface Defect Detection Network via Adaptive Dual-Path and Multi-Branch Aggregation |
| 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 |