A Refined Identification Method for the Hidden Dangers of External Damage in Transmission Lines Based on the Generation of a Vanishing Point-Driven Effective Region

Abstract

As the carrier of electric energy transmission, transmission lines undertake the important task of electric energy distribution and transfer. However, with the increasing frequency of construction using large machinery such as tower cranes and excavators under the transmission channels, transmission line accidents occur frequently. Therefore, this paper proposes a refined identification method for the hidden dangers of external damage in transmission lines based on the generation of effective regions driven by vanishing points. The comprehensive and accurate perception of external damage targets through the perception model of scene elements based on slicing-aided hyperinference was realized. Secondly, the accuracy and robustness of the calculation of the transmission line’s vanishing point were improved based on Canny edge detection and Hough linear detection. The effective region on the visual images was generated by combining the vanishing point and the bottom of transmission tower coordinates. Finally, the relative position relationship between areas with hidden dangers of external damage and the effective warning regions were compared, and the refined identification of hidden dangers was realized. The experimental data show that the proposed method realized a perception accuracy of 82.9% in identifying hidden dangers of external damage caused by ground- and aerial-moving targets, which shows better detection performance and practical value compared with the existing method.

1. Introduction

China’s electric power sector has been growing rapidly in recent years, with transmission lines taking on the major responsibility of moving massive amounts of electricity from power plants to distribution centers or urban substations. The transmission system has become more complex as a result of the expansion of transmission lines. Once a power outage occurs, it will cause serious consequences such as reduced industrial production, traffic jams, and medical failure.

There are many causes of transmission line damage. The authors of [1] analyzed the effects of lightning, typhoons, wildfires, foreign objects, and external damages. According to statistics, external damages have become the largest risk to the safe operation of transmission lines in addition to lightning damage. Illegal construction operations, illegal buildings, foreign object intrusion, and sawing and dismantling are greater threats to the safe and stable operation of transmission lines. Thus, the environment in which the lines are located is critical to the safe operation of transmission lines [2,3]. Ensuring the safe operation of the lines and reducing unnatural factors such as external damages through artificial intelligence methods are of great significance for improving the stability of the power system [4]. In particular, after the 2010s, departments paid more attention to the application of non-engineering measures in the planning of external damage prevention, that is, through the early warning of external damages, risk assessment, the establishment of vandalism insurance, and other means, as shown in Figure 1 [5]. Combined with the existing various engineering measures, including manual marking of dangerous areas, promoting safety knowledge, and offline notification, to reduce the impact of external damages, there are still some problems at the technical level that have not been addressed at a large scale.

This study adopts a camera-based online monitoring method to continuously observe the transmission lines with their surrounding environments, which avoids the weather and geographic constraints of helicopter inspections [6] and the endurance issues and line collision risks caused by drone inspections [7,8]. In the construction site around the transmission corridors, there are more personnel and construction types of machinery, making the monitoring environment complex, and infrared, laser, radar, and other instruments are susceptible to interference, most of which can only detect a single type of external damage [9,10]. Most electric companies are now moving in the direction of visual imagery, supplemented by regular drone inspections. In [11], the authors proposed a deep learning-based external force damage detection method for transmission lines, which is robust against changes in illumination and climate. Nevertheless, the current target detection algorithms are significantly less accurate when researchers directly use high-resolution externally damaged target datasets captured from the surveillance cameras, which have the problem of leakage detection and are particularly ineffective in detecting targets with small image sizes. In [12], high AP performance was achieved through new multi-scale prediction improvements; however, it has comparatively worse performance on medium and larger size objects. The authors of [13] improved the YOLOX-S model for the detection of external damage targets in transmission line corridors through the incorporation of a global context module and the introduction of an attention mechanism to enhance the recognition of objects with random features among external damage targets. In [14], an ECA-Net channel attention mechanism network based on FPN was proposed to enhance the model’s ability to extract features in the fusion layer, which has some adaptability for detecting small targets. In [15], the YOLOv5s object detection algorithm was improved by replacing the original convolutional network and introducing a new attention mechanism and regression loss function, which improved the model’s generalization ability and the competence to detect small target features.

In addition, most of the previous studies focused on improving algorithms to increase the accuracy of the identification of external damage targets in transmission lines or conducted post-analysis of accidents that have occurred and alarm systems for accidents that are occurring in order to allow the operation and maintenance personnel to analyze the cause of the accidents and deal with them in time for lines that have been or are being damaged, thus overlooking the problem of external damages themselves. If a movable mechanical vehicle has a wide range of active regions in the image, image distortion can affect the detection results due to the distance between the vehicle and the camera. Moreover, the scale of targets of the same type varies greatly, and for detecting distant objects or small-sized objects in the image, in particular, the detection effect is not good. In fact, some instruments may be far away from the wires, and they are only identified by the algorithm in the image, but they will not cause hidden dangers or substantial damage to the lines. The existing algorithms mostly issue alarm signals immediately after extracting external damage information and therefore are deficient in the identification of hidden dangers of external damages through image processing after detecting external damage targets and actual safety behaviors. They rely on manual experience to determine the effectiveness of detecting hidden dangers of external damage, which leads to poor effectiveness. There are few studies on the delineation of external damage regions in transmission corridors. With the dramatic increase in the mileage of power transmission, an automatic, efficient, and intelligent target identification technology for external damage is urgently needed to ensure the safety of transmission lines [16]. If the identified external damage targets obviously do not cause damage to the lines or towers but are still identified as hidden dangers, they may lead to the generation of false alarms, which are a waste of human and material resources; thus, it can be seen that simply detecting external damage targets is not enough. In [17], a mathematical model for path planning considering tower risk probability and objective function was established, which provides a more comprehensive and targeted approach to transmission line inspection path planning. The authors of [18] used the remote sensing satellite “Gaofen-1” to obtain regional remote sensing images and used deep learning methods to classify the bare land in the images, thereby obtaining the regions with hidden dangers of external damage to transmission lines. In [19], the authors used an edge feature detector to generate candidate target sets and then suppressed background noise based on a convolutional neural network and completed transmission lines’ data extraction through the Hough transform.

The main contributions of this paper are summarized as follows:

(1)

We proposed an effective early warning region generation method based on the vanishing point driving method, which achieved precise and reliable identification of hidden dangers of external damage in transmission lines.

(2)

A scene element perception model for transmission corridors based on slicing-aided hyperinference was proposed to achieve the comprehensive and accurate detection of external damage targets around the scene.

(3)

An image vanishing point calculation method based on Canny edge detection and Hough line detection was proposed, which improved the accuracy and robustness of vanishing point calculation.

2. Methodology

According to the complexity analysis of external damage scenes in transmission corridors, we proposed a scene perception model based on slicing-aided hyperinference (SAHI) [20] to realize the comprehensive and refined detection of external damage targets in the surrounding scenes of transmission corridors. We also proposed an effective warning region generation method based on the vanishing point driving technique to realize the effective screening of external damage and hidden dangers. Through the intelligent assessment of intrusion behaviors and the threat degree of external damage targets, the effective identification of hidden dangers of external damage in transmission lines is realized.

2.1. Complexity Analysis of External Damage Scenarios of Transmission Lines

Transmission lines are located in mountainous regions, rivers, cities, and wild areas with many scene elements and complex geographic environments, in which illegal construction or encroachment of transmission corridors by construction machinery such as excavators, cranes, and cement pumps trucks are the main cause of hidden dangers of external damages, as shown in Figure 2.

External damages are caused by objects that threaten the stable operation of the lines in the transmission corridors. There are various background elements in transmission line scenes that are prone to overlap or occlusion with external damage targets. Due to the installation position of the cameras, the collected data sources mostly belong to high-resolution long-view images, which easily leads to limitations in detecting external damage when the target size is reduced, poor identification of significant features, and other phenomena. The instantaneity of accidents caused by external damage necessitates their identification and prevention in the hazard stage [21]. For this reason, how to efficiently and accurately perceive external damage targets and prevent hidden dangers of external damage in a refined and reliable way is an urgent problem, which is solved in this study.

2.2. A Scene Perception Model Based on Slicing Aided Hyper Inference

We adopted YOLOv8, which has the characteristics of lightweight structure and low inference delay, as the basis of the scene perception model and introduced the slicing-aided hyperinference method to ensure that all external damage targets in the transmission corridors were sensed as much as possible. Hence, we achieved a comprehensive and refined perception of external damage targets in transmission corridors, thus facilitating the prevention of external damage accidents that cause their failure through detection in a timely manner.

(1)

YOLOv8 model principle

As shown in Figure 3 [22], YOLOv8 has been improved based on YOLOv5. By adding the C2f module to replace the C3 module, it can utilize advanced features and context information in external damage images to improve the detection accuracy of scene elements.

An anchor frame freedom strategy was adopted to enable each branch to focus on its task, thus improving the overall prediction accuracy of the model. Meanwhile, YOLOv8 uses CIoU and DFL loss functions to calculate bounding box losses, enabling the model to focus more quickly on the specific positions of scene elements around the transmission corridors and perceive the distribution of adjacent areas, thereby achieving a fast and accurate perception of targets in the scenes [23].

(2)

Slicing-aided hyperinference

Transmission corridors’ scenes may have huge losses due to external damage targets not being sensed in time. YOLOv8 belongs to one-stage models and still has shortcomings in thoroughly detecting external damage targets. To this end, we introduced the slicing-aided hyperinference method based on YOLOv8, which reduced the model’s under-reporting rate of small or distant targets in the scenes by slicing the input images into small-image sections with overlaps. The strategy involves the parallel perception of full graph inference and slice-assisted inference and using the perceived results as the final output through post-processing operations such as superposition. The slicing-aided hyperinference model copies the original image and divides it into m × n subgraphs like p₁, p₂, … p_l. The segmentation box size and the overlap rate will produce different subgraphs and perceptual results.

Figure 4 shows two example diagrams of transmission line scenes’ segmentation with 5280 × 2992 pixels, setting the slicing frame pixels to be 1000 × 1000. Since there are less than 1000 pixels on the far right and bottom of the images, they need to be overlapped forward, as shown in Figure 4a. In that case, the final slice is 4 × 7 subgraphs. A 20% overlap rate was reserved during normal slicing to avoid missing detection at adjacent edges of the subgraphs, as the overlap rate should not be set too high; otherwise, the detection accuracy of slightly larger pixel targets will be reduced. It can be seen in Figure 4b that the adjacent subgraphs after slicing cannot fully fit subsequent concatenation. However, in this way, the repeated detection of the same object occurs between adjacent subgraphs and original graphs. It is necessary to adopt post-processing to eliminate redundant detection boxes in the output results and retain the best detection results, with higher values than the setting threshold. The simulation shows that the target detection results based on SAHI inference are better than the original target detection results, significantly reducing the under-reporting rate of external damage target detection and diminishing the misdetection rate.

2.3. An Effective Early Warning Region Generation Method Based on Vanishing Point Driving Technique

In order to realize the refined perception of hidden dangers of external damage, we proposed a method for generating effective early warning regions of transmission channels based on the vanishing point driving method and stipulated that when external damage targets enter the effective warning regions, the probability of external damage accidents are the highest. Therefore, the intelligent refined recognition of hidden dangers of external damage was realized by detecting the relative position relationship between the external damage targets and the effective warning regions.

(1)

Image vanishing point detection algorithm

The acquisition of transmission corridors’ scenes with cameras is a mapping process from 3D to 2D space. The acquired images lack the ability to represent the direction coordinates of the depth axis, causing the distortion of scene elements parallel to the depth axis in the image and the convergence of their extension lines at a point in the image, which is the vanishing point [24]. A monitoring camera for external damages to the transmission corridors was installed at the front view of the towers, with the transmission lines parallel to the depth axis, to facilitate the localization of the vanishing point to the towers’ portion of the image after mapping, as shown in Figure 5.

The authors of [25] used a deep multitask learning method to detect vanishing points in railways and determined the alarm region through the left and right tracks to detect foreign object intrusions. However, this method may result in missed detections for foreign object intrusions with certain heights. In addition, transmission towers and lines are higher than railway tracks; thus, high-altitude targets must be considered. Consequently, we proposed a reliable target region segmentation method based on detecting vanishing points. There are numerous scene elements parallel to the depth axis in the transmission corridors, and the number of vanishing points that converge in the image also increases accordingly, leading to a large number of invalid vanishing points, which is not conducive to the subsequent generation of effective warning regions based on the vanishing point driving method. For this reason, we focused on finding the image vanishing point, using the Canny edge detection and Hough straight-line detection algorithms to realize the feature extraction of the images. We proposed an optimal selection strategy to ensure the accuracy and reliability of image vanishing point selection [26].

The Canny edge detection algorithm converts the transmission corridors’ scene image into a binary graph, that is, with pixel values of only 0 and 255, and preserves only the contour features of the scene elements. Thus, the feature search range of Hough line detection is narrowed, and the interference of background elements to image vanishing points is eliminated. The Hough line detection algorithm maps the lines in the image to the parameter space based on the principle of point curve duality, and its polar coordinate expression is as follows [27]:

$$\lambda =x\cos \theta +y\sin \theta$$

(1)

where $\lambda$ is the vertical distance from the origin to the straight line, and $\theta$ is the angle between the vertical and the X-axis.

According to the same line of $\lambda$, $\theta$ is a fixed value, and when an image of different coordinates (x, y) is input into the $\lambda -\theta$ coordinate system, the corresponding number of curves are generated, where the intersection of the curves corresponds to the number of points through the straight line. We implemented the detection of image line parameters based on this principle. The image of the external damage scenes captured in practical applications may not intersect perfectly at one point despite the distortion; more often than not, multiple parallel lines in 3D will intersect at several points in the 2D image. For this purpose, it is necessary to calculate and determine the optimal point as the vanishing point of the image from these intersections, which is required to best reflect the degree of distortion in the image.

Figure 6 shows the method for calculating a vanishing point. Parameters for the vanishing point detection line can be obtained by Canny edge detection. Firstly, we calculate the intersection of the two lines L₁ and L₂ that are parallel in reality but intersect in the image and use it as a temporary vanishing point (x_VP, y_VP), which is shown in Equation (2). Then, a vertical line L_⊥ perpendicular to the third line L₃ is generated through the point (x_VP, y_VP). The parameters of the vertical line equation are shown in Equation (3). Subsequently, the intersection point (x₃, y₃) of L₃ and L_⊥ is calculated. Its calculation method is shown in Equation (4). Finally, the distance between the temporary vanishing point and the third line L₃ is calculated, that is, the distance between two points (x_VP, y_VP) and (x₃, y₃). The calculation method is shown in Equation (5).

$$\left\{\begin{array}{l}{L}_1:y=k_{1}x+b_1\\ L_2:y=k_{2}x+b_2\end{array}\right.\Rightarrow \left\{\begin{array}{l}{x}_{\mathrm{VP}}={\displaystyle \frac{b_2-b_1}{k_1-k_2}}\\ y_{vp}={\displaystyle \frac{k_1{b}_2-k_2{b}_1}{k_1-k_2}}\end{array}\right.$$

(2)

$$k_{\perp }=-{\displaystyle \frac{1}{k_3}} , b_{\perp }=y_{\mathrm{VP}}+{\displaystyle \frac{1}{k_3}}{x}_{\mathrm{VP}}={\displaystyle \frac{k_1{b}_2-k_2{b}_1}{k_1-k_2}}+{\displaystyle \frac{1}{k_3}}{\displaystyle \frac{b_2-b_1}{k_1-k_2}}$$

(3)

$$\left\{\begin{array}{l}{L}_3:y=k_{3}x+b_3\\ L_{\perp }:y=k_{\perp }x+b_{\perp }\end{array}\right.\Rightarrow \left\{\begin{array}{l}{x}_3={\displaystyle \frac{b_{\perp }-b_3}{k_3-k_{\perp }}}\\ y_3={\displaystyle \frac{k_3{b}_{\perp }-k_{\perp }{b}_3}{k_3-k_{\perp }}}\end{array}\right.$$

(4)

$$distance=\sqrt{{(x_{\mathrm{VP}}-x_3)}^2+{(y_{\mathrm{VP}}-y_3)}^2}$$

(5)

where k₁, k₂, k₃, and k_⊥ are the slope of lines L₁, L₂, L₃, L_⊥; b₁, b₂, b₃, and b_⊥ are the intercept of lines L₁, L₂, L₃, and L_⊥; and distance is the length from the temporary vanishing point to L₃.

The sum of the squares of the distance from the temporary vanishing point to more lines is calculated, and the square root is determined as the calculation error to compare with the current minimum error value. If the new calculation error is smaller, the vanishing point coordinates are redetermined, and the point with the smallest total square distance from all lines is ultimately selected as the vanishing point.

(2)

Effective warning region generation method

Considering that the transmission towers are perpendicular to the depth axis, there is only size scaling with no obvious position distortion in the imaging process. The experiment also verified that most of the image vanishing points were located inside the tower detection frame. As a result, we defined the lower triangle area generated by the vanishing point, which was obtained based on the scene perception model, and the two rays extending from the bottom vertex of the transmission tower detection box were used to determine the effective warning region for external damage and hidden dangers. This effective region plays an important role in standardizing the external damage targets among moving locomotives such as bulldozers, compactors, and dumpers, while it is not effective in standardizing the external damage targets with great height such as tower cranes. Thus, we had to expand the scope of the effective region. Based on the triangle region, the two rays were extended in reverse, and a tower detection box was added to combine the three regions in the two-dimensional image as an effective region for the detection of transmission lines’ external damage and hidden dangers in practical application scenarios. The final result is shown in Figure 7.

The lower triangle region was used to control the ground-moving external damage targets, while the upper triangle region was mainly used to control the aerial-moving external damage targets such as tower cranes, foreign objects in the air, and crane mechanical arms. By comparing the relative position of the external damage targets and the effective warning regions, the mechanism of the potential sensing of external damage in the transmission corridors was set up, which can generate alarm signals regarding the external damage targets in the regions and control external damages outside the regions.

3. Results and Discussion

3.1. Data Description and Experimental Environment

In this study, the video images collected by the external damage monitoring cameras of the transmission corridors were used as a data source. Considering the little feature difference between the front and back frames, a picture was extracted every 1 min, and invalid pictures such as blurred and severely blocked images were eliminated. Subsequently, they were annotated, and a dataset of external damage targets was constructed. The dataset was divided into training and testing sets in an 8:2 ratio, with 2755 images as the training set and 700 images as the testing set. It mainly consisted of overhead lines and included the following types of external damage targets: crane, tower crane, cement pump truck, bulldozer, soil compactor, excavator, dumper, and smoke. We used Chinese Pinyin as the label name to distinguish external damage targets effectively. Taking into account the need for the effective screening of external damage accidents, the transmission towers were also marked and tested together with the external damage targets to facilitate the subsequent division of the effective regions of hidden danger targets combined with the transmission towers.

In order to verify the performance of the refined perception method for identifying external damage accidents in transmission lines, we completed model environment construction, training, and performance testing in the server; its configuration is shown in

Table 1. Experimental environment configuration.

Configuration Name	Specific Information
Server type	DELL Precision T5820 GPU server
CPU	i9-10980XE, 18 cores, 3.0 GHz
GPU	2*RTX3090, 24 GB
RAM	128 GB
Hard disk	10 T, solid state drive

.

3.2. Comprehensive and Accurate Perception of External Damage Targets in Transmission Lines

(1)

Ablation experiment

The prediction mechanism of YOLOv8 first divides the input image into multiple grid cells, predicts multiple potential bounding boxes within each grid cell, and assigns corresponding confidence scores and category probabilities to these bounding boxes. This grid-based prediction strategy enables YOLOv8 to focus on multiple potential targets in the image simultaneously, ensuring fast processing speed and detailed and error-free identification of complex targets in the image, significantly improving the efficiency of target detection. By comparing the detection effects of YOLOv8n, YOLOv8m, and YOLOv8x in actual transmission corridors, we selected the detection framework with the best accuracy for the perception of external damage targets in transmission corridors. The detection results are shown in Figure 8.

It can be seen from Figure 8 that the n model in the left figure has problems such as missed detection and some redundant detection boxes. Compared with the n model, the m model significantly improved regarding these issues. However, the confidence scores in the right figure for the external damage targets are lower than those of the x model. Therefore, we used YOLOv8x as the detection framework of the perception model of transmission corridors’ scenes. In order to cover as much area as possible in the transmission line regions in the external damage scenes, the volume of low-resolution data collected by the cameras was not large, and the area of transmission scene region detected using low-resolution images was small, which made it difficult to conduct effective screening research on external damage and hidden danger targets based on the detection results. Thus, we tried to improve the recognition effect of external damage targets from the algorithm level.

(2)

Analysis of YOLOv8x with SAHI model detection performance

We proposed a method to improve the refined perception ability of model scene elements based on the slicing-aided hyperinference module and verified the degree of influence of the slicing frame 1000 × 1000, with an overlap rate of 20%; the slicing frame 800 × 800, with an overlap rate of 20%; the slicing frame 500 × 500, with an overlap rate of 20%; the slicing frame 1000 × 1000, with an overlap rate of 30%; and the slicing frame 800 × 800, with an overlap rate of 30% on the performance of the YOLOv8x model. The detection results are shown in Figure 9. It can be seen that the YOLOv8x model performed best when the slicing frame was 1000 × 1000, and the overlap was 20%. The model with SAHI significantly reduced the under-reporting rate.

In Figure 9b, the tower crane target is not detected in the left image, and the slender weed is mistakenly detected as the crane target in the right image. In Figure 9c, the tower crane target is not detected in the left image, and there are some redundant boxes in the right image. In Figure 9d, the tower crane target is detected in the left image with lower detection reliability compared to Figure 9a, and there are a few redundant boxes in the right image. Figure 9e has the same detection result as Figure 9b. In summary, the YOLOv8x model performed best when the slicing frame was 1000 × 1000, and the overlap rate was 20%. After a number of experiments performed for verification, it was found that a too-small slicing frame and too much overlap do not assist YOLOv8 inference well, whereas 20% of the overlap rate can achieve a better result.

To verify the performance difference between the proposed method and the same type of scene perception model, we compared the detection accuracy of YOLOv7 for external damage targets. The detection results are shown in

Table 2. Comparison results of different scenes with YOLOv7 model detection.

Target Type	mAP (%)
	YOLOv7	Detection Methods in This Paper
Crane	65.5	81.5
Tower crane	72.4	83.4
Cement pump truck	66.7	79.9
Bulldozer	81.5	88.2
Soil compactor	65.5	78.6
Excavator	82.5	86.6
Dumper	75.1	83.0
Transmission tower	81.2	88.9
Smoke	73.3	76.3
Total	73.7	82.9

.

As illustrated in Table 2, the proposed method demonstrated a detection accuracy that was 9.2% higher than YOLOv7. Except for the detection accuracy for the cement pump truck and soil compactor, the accuracy for other external damage targets exceeded 80%. The overall detection accuracy of the proposed model was 82.9%, outperforming the comparative models and thereby confirming the effectiveness of the YOLOv8x integrated with the SHAI model presented in this study. The recognition accuracy for transmission towers was the highest, which can be attributed to the abundance of tower targets in our dataset. A single image may contain multiple towers and tower cranes. However, due to the substantial size of the tower cranes, a complete high-resolution image may not capture all the features of the tower cranes, leading to missed detections and consequently lower identification accuracy for the tower cranes. Among all models, the recognition accuracy for smoke was the lowest, primarily because the dataset contained the least number of smoke instances.

3.3. Refined and Reliable Identification of External Damage Accidents in Transmission Lines

(1)

Image vanishing point detection

For the purpose of verifying the effectiveness of the vanishing point accurate location method based on the vanishing point straight lines’ recognition and the optimal vanishing point selection strategy combined with the Canny edge detection and Hough line detection algorithm, the selection angle of the Hough line detection algorithm was limited to [−85°, −5°] ∪ [5°, 85°] to prevent the interference of approximate horizontal and vertical lines in images. The comparison results of detection with the vanishing point obtained without this process are shown in Figure 10.

The angle screening function of Hough detection is not used in Figure 10a, leading to the detection of not only transmission line features but also features of other linear scene elements, which negatively affects the accuracy of the optimal vanishing point selection strategy for image vanishing point calculation. Its specific manifestation in Figure 10a is that the vanishing point seriously deviates from the transmission corridors, which is not conducive to the task of generating an effective warning region based on the vanishing point driving method. Figure 10b shows the method used in this paper, which filters out the approximate horizontal and approximate vertical lines by the angle screening function of Hough detection and determines the specific location of their vanishing points based on this feature. As shown in Figure 10b, the vanishing point is located in the center of the transmission line extension lines and falls on the tower targets in most cases. Therefore, the proposed method lays a foundation for generating an effective early warning region based on the vanishing point and the bottom vertex of the towers’ detection frame.

(2)

Generation of effective early warning regions driven by the vanishing point

We verified the viability of the effective region generation method based on the detection boxes of transmission corridors’ vanishing points and the towers obtained from previous experiments. The effective warning region generation results are shown in Figure 11.

As shown in Figure 11, two rays were obtained from the vanishing point, and the two bottom vertices of the transmission tower detection box were used to construct an effective warning region. The direction of the rays started from the vanishing point and faced the bottom vertex of the detection box. The generation strategy for effective warning regions involved creating a lower triangular area between the vanishing point and the bottom vertices of the tower detection box. The upper triangular area was formed by the reverse extension of two rays, which required the inclusion of all transmission lines as much as possible. If the transmission lines could not be included, the line itself detected by the Hough transform was used as the upper triangular boundary. The bounding box section was used to connect the two areas. By connecting the above regions, an effective early warning region based on the vanishing point was constructed.

(3)

Identification of hidden dangers of external damage to transmission lines

To verify the utility of the refined and reliable identification method for detecting external damage and hidden dangers in transmission corridors, the scene of external damage targets invading effective warning regions was selected. The detection results are shown in Figure 12.

Figure 12 shows, from left to right, the detection results of external damage targets with vanishing points, the effective region division method, and the visualization of effective regions. Scenario one shows the excavator target. By comparing its position with the effective warning region, it can be seen that the two do not overlap. Thus, under the current state, the model does not consider the data sufficient to evolve into a model that considers the hidden danger of external damage. The system should focus on monitoring its operating status without warning. When it continues to drive forward and cross the effective region, the monitoring device sends an alarm signal to the driver and the monitoring center, reminding the driver to steer carefully. The monitoring center manually determines that there is no actual risk and eliminates the alarm signal. The excavator target in scenario 2 is judged not to intersect by comparing its position with the effective warning region. The model believes that it does not belong to a target with hidden dangers of external damage. However, due to its close distance from the effective region, the transmission lines should be monitored continuously even if no alarm signal is sent. In scenario 3, the tower crane detection frame intersects with the effective region and is very close to the transmission corridors, which is highly likely to touch the wire and seriously threaten the safety of the transmission lines. There is an urgent need to send an alarm signal to the driver and the monitoring center to remind the driver to operate with caution. The monitoring center, after manually determining that there is indeed a great risk, also needs to dispatch staff to deal with the situation in a timely manner to inform the tower crane driver of safety issues and sign a safety notification letter.

3.4. Discussion

According to the coordinates of the vanishing points and the transmission tower detection frame, the example shows the method of generating the effective region driven by the vanishing points and then judging the geometric relationship between the detected external damage targets and the effective regions. We analyzed the alarm methods for three different scenarios: penetrating into and out of the effective region, not in the effective region, and in the effective region. In summary, we verified the effectiveness of the proposed method through actual scenarios and proved that it can achieve the refined and reliable identification of the hidden dangers of external damages.

4. Conclusions

In this paper, we first proposed the YOLOv8 transmission line external damage monitoring method with fusion slicing-aided inference for the problem of small-target leakage detection in high-resolution images. SAHI was used to improve the YOLOv8 method for extracting small-target features and using the non-maximum suppression method to filter out redundant target detection boxes. The detection results of external damage targets on transmission lines with low under-reporting rates were obtained. The method enhanced the detection capability of the identification model of transmission lines’ external damage for small-sized targets in high-resolution images, minimized the external damage missed detection phenomena, and consequently reduced the frequency of transmission line outages. Additionally, to address the problem of the poor effectiveness of external damage target identification method for transmission lines, we proposed an effective screening method for external damage accidents based on vanishing point-driven effective region generation. Images were analyzed from the perspective of gray scaling, Gaussian filtering, edge detection method, and the application of Hough transform theory combined with mathematical computation method, to obtain the coordinates of the images’ vanishing points. Then, the effective region with hidden dangers of external damage in the two-dimensional plane was constructed by the lines between the image vanishing point and the boundary points of the transmission tower detection frame. By comparing the pixel relationship between the detection frame of the external damage targets and the effective region to determine the invasive behavior and the threat level of the hidden dangers of external damage, we obtained improved identification results for detecting the hidden dangers of external damage with a prominent increase in the effective alarm rate. The research results of this paper indicated that these methods improved the effective identification and early warning ability of the online monitoring approach based on cameras for external damage accidents and reduced the number of false alarms.

Although our research has achieved some results, there are still some limitations and room for development. Firstly, due to the scarcity of externally broken segmentation datasets, we mainly used target detection methods for recognition. In the future, the development of SAM (Segment Anything Model) and large language models can be used to realize the annotation of large-scale segmentation datasets, so as to improve the structure and training strategy of the cascade external breakage recognition model and improve the model performance and accuracy. Secondly, the current model mainly recognizes large construction machinery and does not adequately identify secondary hidden targets. The robustness and generalization ability of the model in complex scenarios should be enhanced by expanding the dataset and combining it with safety risk assessment specifications while accurately assessing the risk level and formulating targeted measures. Finally, by combining the spatial–temporal architecture and using BEV features collected by cameras and multimodal information such as millimeter-wave radar, the effect of 3D image distortion can be reduced to achieve the fine identification of normalized external breach at the distance level.