A Low False Alarm Detection Method for Guide Wires Defects Based on Class-Balanced Sampling

Abstract

The present invention discloses a low false alarm detection method for ground wire defects based on class-balanced sampling. First, image data collected by a drone at the power transmission line site is obtained, and an image block of 1200*800 centered on the target is intercepted as the training set / validation set and imported into the training platform. The Cascade-RCNN target detection model is used as the basic defect detection model, and then the training set is redefined by the Class-balanced sampling strategy to train the defect detection model. The method of the present invention can detect defects on the ground wire under the conditions of relatively fast running speed, few false alarms and high discovery rate.

Description

Technical Field

[0001] The present invention belongs to the technical field of power line inspection, and particularly relates to a low false alarm detection method for conductor and ground wire defects based on class balance sampling. Background Art

[0002] The safe and stable operation of transmission lines is an important part of the power system. At the same time, ensuring the safety and stability of transmission lines is also an important cornerstone of the construction of the national power grid infrastructure. Conductors and ground wires are one of the most important components in overhead transmission lines. Power plants rely on conductors to transmit electricity to users to form a power grid.

[0003] However, due to the complex and changeable field environment where transmission lines are located, they are subject to the corrosion of rainwater and the action of micro-vibration for a long time, which easily causes various corrosion and wear on the conductors and ground wires. The superposition of corrosion and wear and stress concentration leads to radial cracks in the conductors at this place. Under the action of cyclic stress, the cracks gradually develop, and finally the conductors fatigue and break. Such cracks in the conductors and ground wires can easily affect the safe operation of the entire transmission line. In the lightest case, it will trip, and in the most serious case, the conductor will break or the tower will collapse, causing a large-scale power outage accident. Common defects of conductors and ground wires include broken strands, loose strands, damage, etc., as Figure 1 shown.

[0004] At present, the State Grid has used drones to inspect the transmission line towers of the power grid to achieve the detection of conductor and ground wire defects. By operating the drone to collect on-site data along a fixed cruise route for the transmission line, and then transmitting the image data back to the server, and finally processing the transmitted data by the target detection algorithm model (such as RetinaNet, YOLOv5, and Faster-RCNN, etc.) running on the server, the detection results of conductor and ground wire defects can be obtained.

[0005] However, due to the imbalance of conductor and ground wire defect categories and the extremely small size of some defects, a large number of false alarms or missed detections still exist in the results obtained by the algorithm model. Moreover, the front-line team members of the power grid need to review the results, which consumes a large amount of manpower and material resources. Therefore, a detection method with a relatively fast speed and high accuracy is required for conductor and ground wire defect detection, which can batch, intelligently, and quickly analyze various defects and positions on the conductors and ground wires in the visible light pictures taken by drone inspections, and minimize missed detections and false alarms. Summary of the Invention

[0006] Aiming at the disadvantages of the existing conductor and ground wire defect detection algorithms, such as many false alarms and occasional missed detections, the present invention aims to provide a detection method for various conductor and ground wire defects with a relatively fast operation speed, few false alarms, and a high detection rate.

[0007] The technical solution adopted by the present invention to solve its technical problems is: a low false alarm ground wire defect detection method based on class-balanced sampling, including the following steps:

[0008] Step S1, obtain the image data collected by the drone at the transmission line site, and intercept an image block of 1200*800 centered on the target as the training set / validation set and import it into the training platform;

[0009] Step S2, model training stage

[0010] Step S21, adopt the Cascade-RCNN detection model with higher accuracy in the academic community as the basic defect detection network;

[0011] Step S22, define the training data through the Class-balanced sampling technique: normalize the training set pictures. The frequency f(c) of the rarest category c contained in a certain picture is used as the ratio of the number of images containing the targets of this category to the total number of training set samples. Then the repetition factor where t is the repetition threshold. If the value of f(c) is less than t, then c is resampled. For the image X containing the category set L(X), the repetition factor is r(X) = max c∈L(X) r(c);

[0012] Step S3, model inference stage

[0013] Step S31, in order to improve the detection recall rate of small targets, use block detection to uniformly scale the segmented image blocks with a block size of 1200*800 and an overlapping area size of 100*100 to (2500, 1500) and send them into the basic defect detection network for inference. Finally, map the results generated by the image blocks belonging to the original same picture back to the original picture, and then perform non-maximum suppression (NMS) to obtain the final result of this stage;

[0014] Step S32: Perform line detection on the geodesic line in the image through Hough transform. The polar coordinate formula of the line is r = xcosθ + ysinθ. Map the points in the original image's Cartesian coordinate system to the polar coordinate system in the parameter space. A certain pixel point (x, y) in the original image space is mapped to a sine curve in a (r, θ) polar coordinate system. The collinear points in the original image space intersect at a point (r’, θ’) on the corresponding sine curve in the (r, θ) parameter space. First, discretize the r and θ values into a finite number of equally spaced discrete values. After discretization, the parameter space is no longer continuous, that is, it is discretized into grids of equal size. Then, transform the coordinate values of each pixel point in the edge-detected image space into the parameter space. The obtained value will fall within a certain grid, and the accumulation counter of that grid will be incremented by one. After mapping all the pixel points in the image space to the corresponding unit grids through Hough transform, a threshold can be set. When the count value of a certain unit grid is greater than the set line threshold, it is considered a line;

[0015] Step S33: After detecting the geodesic line, calculate the perpendicular distance from the center point of each model output target box to the geodesic line where (x0, y0) is the center point of the target box, and A, B, and C are the coefficients of the line equation. When the distance from the target box to the geodesic line is greater than the threshold T, the target box is discarded as a false alarm, and the low false alarm geodesic line defect detection is completed.

[0016] Furthermore, the feature extraction network of the Cascade-RCNN detection model in step S21 consists of 13 convolutional layers, 13 BN+relu layers, and 4 pooling layers. The detection head of the Cascade-RCNN detection model has a class prediction (cls_pred) branch for classification calculation and a bounding box prediction (bbox_pred) branch for coordinate regression. The steps for the Cascade-RCNN detection model to perform the first-stage refinement on the candidate targets are as follows: The candidate targets that have been preliminarily refined and come out of the first detection head in series enter the next detection head. The second-stage detection head trained with a higher IOU threshold refines the candidate targets again to obtain candidate targets with higher quality;

[0017] Furthermore, the first 4 convolutional modules in the backbone network of the Cascade-RCNN detection model in step S21 are Res-DWConv modules. The Res-DWConv module consists of 1 depthwise convolution module and 3 pointwise convolution modules.

[0018] The beneficial effects of the present invention are:

[0019] The method of this patent is mainly based on cutting-edge artificial intelligence and computer vision technologies. It uses a multi-stage cascaded localization regression optimization network as the basic defect detection network, and adopts block detection and Class-balanced sampling technologies. Finally, spatial position constraints are carried out through line detection, so as to design a low false alarm ground wire defect detection method based on class-balanced sampling. In the inference stage of the present invention, block detection is adopted, that is, the divided image blocks with a block size of 1200*800 and an overlapping area size of 100*100 enter the inference module to output the ground wire defect area as candidate targets. Then, line detection is performed on the image through the Hough transform to extract the ground wire. After obtaining the ground wire area, the candidate defect targets not on the ground wire area are removed, that is, the false detections are eliminated, so as to finally obtain the low false alarm ground wire defect detection targets.

[0020] Compared with the existing power equipment defect detection methods, the method of the present invention can detect the defects on the ground wire under the conditions of faster operation speed, fewer false alarms and higher detection rate. Brief Description of the Drawings

[0021] Figure 1 Common ground wire defect pictures;

[0022] Figure 2 It is a schematic structural diagram of the Cascade-RCNN detection model of the present invention;

[0023] Figure 3 It is a schematic structural diagram of the Res-DWConv module of the present invention;

[0024] Figure 4 It is a schematic structural diagram of the block detection of the present invention;

[0025] Figure 5 It is a schematic diagram of the mapping principle from the image space to the polar coordinate space;

[0026] Figure 6 It is the overall flow chart of the ground wire defect detection of the present invention. Detailed Embodiments

[0027] The present invention will be further described in detail below with reference to the accompanying drawings.

[0028] The present invention discloses a low false alarm ground wire defect detection method based on class-balanced sampling, belonging to the technical fields of artificial intelligence and object detection, and involving deep convolutional neural network (CNN), computer vision object detection and spatial position constraints. The ground wire defect detection process in the present invention is mainly divided into a model training stage and a model inference stage. The convolutional neural network structures in the two stages are basically the same, but there are some differences in the data processing pipeline, which are described in detail as follows.

[0029] Step S1, obtain the image data collected by the drone at the transmission line site. Pictures of common conductor and ground wire defects are as shown in Figure 1 . Crop an image block of 1200*800 centered on the target as the training set / validation set and import it into the training platform.

[0030] Step S2, model training stage.

[0031] Step S21, to ensure the detection accuracy of defects, the present invention uses Cascade-RCNN as the basic defect detection network, as shown in Figure 2 . It can be seen from the figure that the feature extraction network of Cascade-RCNN is composed of 13 convolutional layers, 13 BN+relu layers and 4 pooling layers. The detection head has two branches: class prediction (cls_pred) and coordinate prediction (bbox_pred), which perform classification calculation and coordinate regression respectively, and conduct the first-stage refinement of the candidate targets: the candidate targets preliminarily refined from the first detection head in series enter the next detection head. The second-stage detection head trained with a higher IOU threshold further refines the candidate targets to obtain candidate targets with higher quality.

[0032] Considering the balance between accuracy and speed, in the present invention, 3 detection heads are connected in series, and each stage's detection head focuses on refining proposals (candidate targets) with higher IOU. Because the IOU of the proposals output by each detection head is generally greater than the input IOU, the refinement effect will be better and better.

[0033] In addition, to further reduce the computational amount and improve the operation speed, the present invention replaces the first 10 conventional convolutions in the backbone network of Cascade-RCNN with 4 Res-DWConv modules. The Res-DWConv module is as shown in Figure 3 . A standard Res-DWConv module is composed of 1 depthwise convolution module and 3 pointwise convolution modules.

[0034] Step S22, define the training strategy. Considering the correspondence with the block detection strategy in the inference stage, during training, the present invention uses the image block of 1200*800 centered on the target as the training set / validation set. In the data preprocessing stage, all training set pictures are uniformly scaled to the scale of (2500, 1500) for training, and random flipping is used as the online data augmentation method, and at the same time, the pictures are normalized.

[0035] To address the problem of unbalanced sample numbers of different types of conductor and ground wire defects, the present invention uses a Class-balanced sampling strategy for training. Based on this sampling strategy, in each training epoch, an image may appear multiple times according to its "repetition factor".

[0036] The repetition factor of an image is a function of the frequency of the rarest class contained in the image. The frequency f(c) of class c is equal to the ratio of the number of images containing the objects of this class to the total number of samples in the training set. The repetition factor calculation process is as follows:

[0037] i) For class c, calculate the proportion f(c) of the number of images containing this class, and calculate the repetition factor of this class where t is the repetition threshold, and if f(c) is less than this value, class c will be resampled;

[0038] ii) For image X and the set of classes L(X) it contains, calculate the repetition factor of this image as r(X) = max c∈L(X) r(c).

[0039] Step S3, model inference stage.

[0040] Step S31, block detection: Generally, the resolution of the conductor and ground wire images collected by drones is between 4000*3000 and 5000*4000, belonging to high-resolution images. And some conductor and ground wire defects such as loose strands and broken strands are about 100*100 in size, which are extremely small targets. If the image is directly sent into the network without scaling, it will consume a huge amount of video memory, resulting in insufficient video memory of the graphics card and a huge amount of computation; if the image is reduced before being sent into the network, those extremely small targets may disappear after downsampling, and the network will not be able to obtain and learn the features of the extremely small targets. Based on the above points, the present invention uses block detection technology. Consistent with the training stage, the block size is 1200*800, and the overlapping area size is the size of the small target, that is, 100*100, to ensure that the small target is not segmented, as Figure 4 shown. Then the segmented image blocks are uniformly scaled to (2500, 1500) and sent into the network for inference. Finally, the results generated by the image blocks belonging to the original same image are mapped back to the original image, and non-maximum suppression (NMS) is performed to obtain the final result of this stage.

[0041] Step S32, false alarm suppression through line detection: Since the defects of the conductor and ground wire only appear on the conductor and ground wire or within a certain distance from it, those results far from the conductor and ground wire can be removed based on this point to achieve the effect of false alarm suppression.

[0042] The ground wire basically appears as a straight line in the image, so the Hough transform can be used for line detection. The polar coordinate formula of a line is r = xcosθ + ysinθ, and the mapping of points in the rectangular coordinate system (i.e., the original image space) to the polar coordinate system (i.e., the parameter space) is as Figure 5 shown, that is, i) after transformation, a certain pixel point (x, y) in the image space is mapped to a sine curve in the (r, θ) polar coordinate system; ii) collinear points in the image space intersect at a point (r’, θ’) in the corresponding (r, θ) space.

[0043] First, the values of r and θ are discretized into a finite number of equally spaced discrete values. After discretization, the parameter space is no longer continuous, that is, it is discretized into grids of equal size. Then, the coordinate values of each pixel point in the image (which must first undergo edge detection) space are transformed into the parameter space, and the obtained values will fall within a certain grid, and the accumulation counter of that grid is incremented by one. After all pixel points in the image space are mapped to the corresponding unit grids through the Hough transform, a threshold can be set. When the count value of a certain unit grid is greater than the set line threshold, it is considered a straight line.

[0044] After detecting the ground wire, calculate the perpendicular distance from the center point of each model output target box to the ground wire. The calculation method is where (x0, y0) is the center point of the target box, and A, B, and C are the coefficients of the line equation. When the distance from the target box to the ground wire is greater than the threshold T, the target box is discarded as a false alarm. The overall process of ground wire defect detection in the present invention is as Figure 6 shown.

[0045] The above embodiments only illustrate the principle and efficacy of the present invention and some applied embodiments. For those of ordinary skill in the art, without departing from the creative concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention.

Claims

1. A low false alarm detection method for guide wire defects based on class balance sampling, characterized in that It includes the following steps: Step S1: Obtain the image data collected by the drone at the transmission line site, and intercept an image block of 1200*800 centered on the target as the training set and the validation set, and import them into the training platform; Step S2: Model training Step S21: Adopt the Cascade-RCNN detection model as the basic defect detection network; Step S22, redefine the training data through Class-balanced sampling: Normalize the training set images. The frequency f(c) of the rarest class c contained in a certain image is used as the ratio of the number of images containing the targets of this class to the total number of training set samples. Then the repetition factor where t is the repetition threshold. If the value of f(c) is less than t, then resample c. For the image X containing the class set L(X), the repetition factor is r(X) = max c∈L(X) r(c); Step S3: Model inference Step S31: Use block detection to uniformly scale the segmented image blocks with a block size of 1200*800 and an overlapping area size of 100*100 to (2500, 1500), then send them into the basic defect detection network for inference. Finally, map the results generated by the image blocks belonging to the same original image back to the original image, and then perform non-maximum suppression together to obtain the final result of this stage; Step S32: Perform straight line detection on the ground wire on the image through the Hough transform. The polar coordinate formula of the straight line is r = xcosθ + ysinθ, and map the points in the original image space rectangular coordinate system to the parameter space polar coordinate system. A certain pixel point (x, y) in the original image space is mapped to a sine curve in the (r,θ) polar coordinate system. The collinear points in the original image space intersect at a point (r’,θ’) in the corresponding (r,θ) parameter space sine curve. First, discretize the r and θ values into a finite number of equally spaced discrete values to obtain several grids of equal size, and then transform the coordinate values of each pixel point in the edge-detected image space into the parameter space. After all the pixel points in the image space are mapped to the corresponding unit grids through the Hough transform, set a threshold. When the count value of a certain unit grid is greater than the set straight line threshold, it is considered a straight line; Step S33, after detecting the ground wire, calculate the perpendicular distance from the center point of the target box output by each model to the ground wire Among them, (x0, y0) is the center point of the target box, A, B, and C are the coefficients of the straight-line equation. When the distance from the target box to the ground wire is greater than the threshold T, the target box is discarded as a false alarm, and the low-false-alarm ground-wire defect detection is completed.

2. The low false alarm guide wire defect detection method based on class balance sampling according to claim 1, wherein In the step S21, the feature extraction network of the Cascade-RCNN detection model consists of 13 convolutional layers, 13 BN+relu layers and 4 pooling layers. The detection head of the Cascade-RCNN detection model has a class prediction branch for classification calculation and a coordinate prediction branch for coordinate regression. The steps for the Cascade-RCNN detection model to refine the candidate targets are as follows: The candidate targets preliminarily refined by the first detection head enter the second-stage detection head trained with a higher IOU threshold, and the candidate targets are refined again to obtain candidate targets with higher quality.

3. A method for detecting low false alarm guide wire defects based on class balance sampling according to claim 2, characterized in that In the step S21, the first 4 convolutional modules in the backbone network of the Cascade-RCNN detection model are Res-DWConv modules, and the Res-DWConv module consists of 1 depthwise convolution module and 3 pointwise convolution modules.

Images