Method and system for detecting defects of power transmission line

Abstract

The invention provides a power transmission line defect detection method and system, and the method comprises the steps: obtaining video data of a power transmission line, and carrying out the preprocessing of the video data of the power transmission line according to a preset data format rule, and obtaining an input item; inputting the input item into a preset detection model for insulator defect and bird nest intrusion detection to obtain a corresponding detection result; wherein the detection model identifies and marks the content of the input video data based on a pre-trained convolutional neural network, and determines insulator defects or bird nests existing in the video data; and when the insulator defect or bird nest invasion exists in the detection result, outputting an alarm signal and alarm information. According to the method, the parameter quantity and the calculation quantity can be reduced by reducing redundant calculation and memory access, the feature extraction effect is more remarkable, and the detection precision in the insulator defect and bird nest intrusion data set is obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission line defect detection, and in particular to a power transmission line defect detection method and system. Background Art

[0002] In order to ensure the safe and stable operation of the power grid, it is necessary to conduct regular inspections of the power grid, and timely locate and handle the components that need maintenance and replacement. The traditional inspection method mainly relies on manual methods, which has the problems of high cost, low efficiency, and easy to miss and misdetect. As an important equipment in the power transmission link, insulators play an important role in supporting and fixing the conductors and ensuring the insulation distance. Insulators are very easy to fail in the transmission lines due to their large number, cross-regional distribution, and long-term exposure to the air and the influence of the harsh natural environment; the presence of bird nests on overhead lines can easily cause the transmission line to trip or the insulator to be punctured. The nesting materials will contaminate the porcelain bottles, resulting in a decrease in the external insulation strength and even insulator flashover. The bird nests are mixed with metal objects such as iron wire and copper wire. When they are wetted by rain on rainy days, they come into contact with the wires below, which can easily cause equipment short circuits, power outages and other power grid accidents. In severe weather such as strong winds, it is very easy to cause line tripping. In recent years, thanks to the rapid development of computer vision technology and the increasing maturity of drone technology, research has been carried out on the use of intelligent inspection platforms to replace manual line inspection work, thereby further optimizing the inspection process and better ensuring the normal operation of the power system. The target detection algorithm can be used to analyze and identify power grid images, automatically detect various defects and hidden dangers in the power grid, and support the acquisition of high-resolution power grid inspection images from different distances, which can make the inspection work more efficient and provide algorithm and technical support for the identification of transmission line insulator defects and bird nest intrusion detection. However, in defect image detection, due to the high cost of data acquisition, difficulty in sample acquisition, and difficulty in data labeling, the scale and quality of the data set cannot reach the ideal situation. This also makes it impossible for deep learning models and target detection algorithms trained based on large-scale data sets to give full play to their own advantages. The transmission line images collected from a long distance contain a variety of target scales that span a large range. It is difficult to achieve balanced detection of multi-scale targets by directly using existing target detection algorithms. In addition, the deployment of target detection algorithms in inspection equipment also involves the problem of lightweight models. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for detecting defects in a power transmission line, so as to solve the technical problem of how to realize defect detection of power equipment.

[0004] On the one hand, a method for detecting defects in a power transmission line is provided, comprising:

[0005] Acquire video data of the power transmission line, and pre-process the video data of the power transmission line according to a preset data format rule to obtain an input item;

[0006] Inputting the input items into a preset detection model to perform insulator defect and bird nest intrusion detection to obtain corresponding detection results; wherein the detection model recognizes and annotates the content of the input video data based on a pre-trained convolutional neural network to determine the insulator defects or bird nests present in the video data;

[0007] When there are insulator defects or bird nest intrusions in the detection results, alarm signals and alarm information are output.

[0008] Preferably, a convolution module is provided in the detection model, and the convolution module is used to compress the input item in the channel, and pass the compressed input item through multiple convolution layers in sequence to obtain multiple feature maps;

[0009] All feature maps are concatenated in the channel direction to form a new feature map.

[0010] Preferably, a CA module is further provided in the detection model, and the CA module is used to perform pooling operations on each channel of the input image along the horizontal and vertical directions, and retain the position information of the target;

[0011] The height and width are obtained at a certain channel in the horizontal and vertical directions, and a feature map is output according to the obtained height and width.

[0012] Preferably, pooling operations are performed along each channel in the horizontal and vertical directions according to the following formula,

[0013]

[0014] in, and is the encoding value output by the cth channel, W is the upper limit of width, H is the upper limit of height, i is the x-axis coordinate sequence, j is the y-axis coordinate sequence, h is the height sequence, w is the width sequence, x c is the C channel input image.

[0015] Preferably, the CA module is further used to embed the generated feature map by splicing the coordinate information according to the following formula: f = δ(F1([z h ,z w ]))

[0016] Among them, f is the intermediate feature map encoded in the horizontal and vertical directions, δ is the nonlinear activation function, F1 is the splicing coordinate information operation, [z h ,z w ] is a concat operation along the spatial dimension.

[0017] Preferably, the CA module is further used to perform dimensionality reduction according to the following formula:

[0018] g h =δ(F h (f h ))

[0019] g w =δ(F w (f w ))

[0020] Among them, g w To decompose the feature map horizontally, g h To decompose the feature map along the vertical direction, f h is the intermediate feature map encoded in the vertical direction, f w is the intermediate feature map encoded in the horizontal direction, F h is the vertical dimension reduction operation, F w It is a horizontal dimension reduction operation.

[0021] Preferably, the obtained tensors decomposed in the horizontal and vertical directions are fused with the input feature map according to the following formula to obtain the final output feature map:

[0022]

[0023] Among them, y c (i, j) is the final output feature map, x c (i, j) is the input feature map, is the tensor decomposed in the vertical direction, is the tensor decomposed horizontally.

[0024] Preferably, a loss function is also provided in the detection model, and the loss function is used to perform regression of the target bounding box.

[0025] Preferably, the loss function comprises:

[0026]

[0027] Among them, A and B are the areas of the predicted box and the real box respectively, α is the parameter for adjusting the gradient convergence speed of the loss function, IoU is the IoU loss function, and L A-SIoU is the A-SIoU loss function, L dis is the distance loss, L shape is the shape loss.

[0028] On the other hand, a power transmission line defect detection system is also provided, which is used to implement the power transmission line defect detection method, comprising:

[0029] A video acquisition module, used to obtain video data of the power transmission line, and pre-process the video data of the power transmission line according to a preset data format rule to obtain an input item;

[0030] A recognition module, used to input the input items into a preset detection model to perform insulator defect and bird nest intrusion detection to obtain corresponding detection results; wherein the detection model recognizes and annotates the content of the input video data based on a pre-trained convolutional neural network to determine the insulator defects or bird nests present in the video data;

[0031] The alarm module is used to output an alarm signal and alarm information when there is an insulator defect or a bird's nest intrusion in the detection results.

[0032] In summary, the implementation of the embodiments of the present invention has the following beneficial effects:

[0033] The power transmission line defect detection method and system provided by the present invention can reduce the number of parameters and calculations by reducing redundant calculations and memory accesses, and has a more significant effect on feature extraction, and significantly improves the detection accuracy in insulator defect and bird nest intrusion data sets. It achieves extremely high detection accuracy with a relatively low model volume, and still meets the real-time detection requirements, with high efficiency and excellence. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, without paying creative labor, other drawings obtained based on these drawings still belong to the scope of the present invention.

[0035] Figure 1 The figure is a schematic diagram of the main process of a method for detecting defects in a power transmission line according to an embodiment of the present invention.

[0036] Figure 2 This is a diagram of the YOLOv10s network structure in an embodiment of the present invention.

[0037] Figure 3 This is a structural diagram of the CA self-attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0039] like Figure 1FIG. 1 is a schematic diagram of an embodiment of a method for detecting defects in a power transmission line provided by the present invention. In this embodiment, the method comprises the following steps:

[0040] Step S1, obtaining the video data of the transmission line, and preprocessing the video data of the transmission line according to the preset data format rules to obtain input items; preliminary processing of the video data and making it conform to the preset standards is a commonly used setting in this field and will not be repeated here.

[0041] Step S2, inputting the input item into a preset detection model to perform insulator defect and bird nest intrusion detection, and obtaining corresponding detection results; wherein the detection model recognizes and annotates the content of the input video data based on a pre-trained convolutional neural network, and determines the insulator defects or bird nests present in the video data; understandably, the detection model provides YOLOv10 of different size models of N / S / M / B / L / X scales based on the scaling factor, such as Figure 2 As shown; at the same time, the backbone network is improved, the self-attention mechanism is introduced, and the loss function is improved.

[0042] In one embodiment, a convolution module is provided in the detection model, and the convolution module is used to compress the input items in the channel, and pass the compressed input items through multiple convolution layers in sequence to obtain multiple feature maps; all feature maps are spliced ​​in the channel direction to form a new feature map. It can be understood that Ghost-Net is a lightweight network, and Ghost convolution is a convolution module in the Ghost-Net network, which can reduce the computational complexity of the convolutional neural network and make the model more lightweight, so that it is easy to deploy terminal devices. The module is mainly composed of two parts. First, the input image is compressed in the number of channels using a general 1×1 convolution, and then more feature maps are obtained through layer-by-layer convolution. Finally, different feature maps are concat spliced ​​in the channel direction to form a new feature map.

[0043] Compared with the standard convolutional neural network, the Ghost convolution module can ensure that the overall parameters of the network are greatly reduced while the recognition effect is almost the same, and the computational cost of the network is reduced. The present invention introduces the Ghost convolution module in the network, and uses the Ghost convolution module to replace the traditional convolution for upsampling and downsampling. At the same time, the Ghost convolution is also added to the C2f module, which is embedded in front of the Concat layer, thereby reducing the number of model parameters while ensuring the detection accuracy requirements.

[0044] In one embodiment, a CA module is further provided in the detection model, and the CA module is used to perform pooling operations on each channel of the input image along the horizontal and vertical directions, and retain the position information of the target; obtain the height and width of a certain channel in the horizontal and vertical directions, and output a feature map according to the obtained height and width. Understandably, in the process of insulator defect and bird nest intrusion detection, since the proportion of pixels occupied by self-explosion defects is small, it is easy to miss the target after the convolution operation. In order to enhance the network's ability to detect targets in complex scenarios and strengthen the network's attention to global information, the CA self-attention mechanism is introduced, such as Figure 3 As shown in the figure. The CA module is a simple and efficient attention mechanism. Different from ordinary attention mechanisms, it comprehensively considers the correlation between different channels and the spatial position information of the feature map. This design enables the network to focus on key information while ignoring irrelevant background information. The CA self-attention mechanism module consists of two parts: coordinate information embedding and coordinate attention generation.

[0045] Specifically, pooling operations are performed along each channel in the horizontal and vertical directions according to the following formula:

[0046]

[0047] in, and is the encoding value output by the cth channel, W is the upper limit of width, H is the upper limit of height, i is the x-axis coordinate sequence, j is the y-axis coordinate sequence, h is the height sequence, w is the width sequence, x c is the input image of the C channel. In the coordinate information embedding stage, the input image passes through the CA module and performs pooling operations on each channel in the horizontal and vertical directions, which can better retain the location information of the target. The output feature map with a height of H and a width of W can be obtained in the c-th channel in the horizontal and vertical directions.

[0048] The CA module is also used to embed the generated feature map into the splicing coordinate information according to the following formula:

[0049] f=δ(F1([z h ,z w ]))

[0050] Among them, f is the intermediate feature map encoded in the horizontal and vertical directions, δ is the nonlinear activation function, F1 is the splicing coordinate information operation, [z h ,z w ] is a concat operation along the spatial dimension. In the coordinate attention generation stage, the concat operation is used to concatenate the coordinate information into the generated feature map, and then a 1×1 convolution is used to perform dimensionality reduction.

[0051] The CA module is also used to perform dimensionality reduction according to the following formula,

[0052] g h =δ(F h (f h ))

[0053] g w =δ(F w (f w ))

[0054] Among them, g w To decompose the feature map horizontally, g h To decompose the feature map along the vertical direction, f h is the intermediate feature map encoded in the vertical direction, f w is the intermediate feature map encoded in the horizontal direction, F h is the vertical dimension reduction operation, F w It is a horizontal dimension reduction operation. The feature map is decomposed into two separate tensors along the horizontal and vertical directions, and then two 1×1 convolutions are used to reduce f h ∈R C / r×H and f w ∈R C / r×W Perform dimensionality increase operation.

[0055] According to the following formula, the tensors decomposed in the horizontal and vertical directions are fused with the input feature map to obtain the final output feature map.

[0056]

[0057] Among them, y c (i, j) is the final output feature map, x c (i, j) is the input feature map, is the tensor decomposed in the vertical direction, is the tensor decomposed in the horizontal direction. Finally, the obtained g h and g w Fuse with the input feature map to obtain the final output feature map Y.

[0058] In one embodiment, a loss function is also set in the detection model, and the loss function is used to regress the target bounding box. The loss function includes:

[0059]

[0060] Among them, A and B are the areas of the predicted box and the real box respectively, α is the parameter for adjusting the gradient convergence speed of the loss function, IoU is the IoU loss function, and L A-SIoU is the A-SIoU loss function, L disis the distance loss, L shape It is the shape loss. A-SIoU is used

[16] The A-SIoU loss function replaces the CIoU loss function. The A-SIoU loss function is an improvement on the SIoU loss function. It introduces an angle penalty metric to guide the bounding box to move faster to the axis closest to the target box. In addition, the A-SIoU loss function also considers three factors that may affect the model evaluation: shape, distance, and angle, so that the model can converge faster. The A-SIoU loss function is mainly composed of the angle loss L angle , distance loss L dis , shape loss L shape And IoU loss consists of four parts.

[0061] The angle loss L angle Such as:

[0062]

[0063] In the formula, z h is the height difference between the center point of the predicted box and the real box, σ is the distance between the center point of the predicted box and the real box, and is the center point coordinate of the real frame, and is the center point coordinate of the prediction box.

[0064] Distance loss L dis Such as:

[0065]

[0066] In the formula, c h , c w , are the height and width of the minimum bounding rectangle of the predicted box and the true box respectively.

[0067] Shape loss L shape Such as:

[0068]

[0069] In the formula, h p , w p is the height and width of the prediction box, h gt , w gt is the height and width of the real frame, θ is the coefficient for controlling the degree of attention to shape loss, and in this embodiment, the value is 4.

[0070] The overall A-SIoU loss function is as follows:

[0071]

[0072] In the formula, A and B are the areas of the predicted box and the real box respectively, and α is a parameter that adjusts the gradient convergence speed of the loss function.

[0073] Step S3: when there is an insulator defect or a bird's nest intrusion in the detection result, an alarm signal and alarm information are output.

[0074] Experimental environment and evaluation indicators,The operating system used in the experiment is Ubuntu 20.04.1, the GPU is NVIDIAGeForce RTX 3090Ti, the video memory is 24G, the CUDA version is 11.6, the deep learning framework is Pytorch, the Pytorch version is 1.13.1, and the Python version is 3.7. During training, the image size is set to 640×640, the training round epoch is 200 rounds, the batch size is 8, and the initial learning rate of the network is 0.01.

[0075] When analyzing the experimental results, the accuracy indicators of the present invention are precision, recall, and mean average precision (mAP). The model indicators are model parameter quantity (Params), model inference floating point number (GFLOPs), and detection speed (FPS). The relevant formulas are as follows.

[0076]

[0077]

[0078] In the formula, is the average precision of each category, and C is the total number of categories.

[0079] Among them, (True Positives) means that the prediction is a positive example and the label value is a positive example, (True Negatives) means that the prediction is a negative example and the label value is a negative example, (False Positives) means that the prediction is a positive example but the label value is a negative example, and (False Negatives) means that the prediction is a negative example but the label value is a positive example.

[0080] Dataset and data enhancement experiment, the data set used in the present invention is obtained from intelligent inspection shooting. This experiment uses Labelimg for manual annotation and saves it in PascalVOC format. The image resolution in the data set is high, and most of the image sizes exceed 1280*1280px. In order to adapt to the characteristics of this high-resolution image, the present invention uses slicing-aided hyper inference technology (SAHI) to slice the training images. The image size after slicing is adjusted to 640*640px. SAHI technology can better handle the training needs of high-resolution images, significantly improve the training speed and verification accuracy of the network, and expand and enhance the data to achieve the effect of enhancing the generalization ability of the model. There are 500 data sets in this experiment, including 350 training sets, 100 verification sets, and 50 test sets. The training set is expanded to 1050 samples, totaling 1200 pictures. The data set is divided into two types of defects, namely insulator-drope and nest invasion.

[0081] In order to analyze the relationship between training samples and network performance, a comparative experiment on the number of training samples was conducted. While keeping the number of parameters basically the same, comparative experiments were conducted on the insulator defect and bird nest intrusion data training sets with the number of amplifications being 350, 700, and 1050, respectively. The experimental results are shown in Table 1.

[0082] Table 1 The impact of training samples on network performance

[0083]

[0084] As can be seen from Table 1, the network performance continues to increase with the increasing number of training samples. When the number of training samples reaches 1050, the recognition rate of the test set reaches a maximum of 88.5%. Therefore, during the network training process, a sufficient number of training samples should be maintained as much as possible so that the network can learn enough information.

[0085] To verify the effectiveness of the improved model structures, YOLOV10-n is used as the baseline network, and ablation experiments are conducted on PConv, Ghost, CA self-attention mechanisms and loss functions, and verified on insulator defect and bird nest datasets. As shown in Table 2, each improvement improves the detection performance of the network to varying degrees.

[0086] After adding the lightweight model GhostConv and the designed C2f-Ghost, mAP@0.5 increased by 0.3%, while the model parameters and calculation amount decreased by 33% and 14.2% respectively, and the detection effect of the baseline network was improved, indicating that the simple structure of the C2f-Ghost module does not affect the feature extraction effect of the network, and can balance the network calculation and detection effect; the introduction of the CA module can make the network pay more attention to the key feature information of the target, thereby enhancing the detection performance of the model. Without increasing the model calculation amount and parameter amount, mAP@0.5 and mAP@0.5:0.95 increased by 0.2% and 0.1% respectively. The ASIoU loss function adds different loss weights to strengthen the network's learning ability for shallow features, which helps the network detect small targets, making the network improve mAP@0.5 and mAP@0.5:0.95 by 0.6% and 0.3% respectively. Experimental results show that the improved algorithm model can reduce the number of parameters and calculations by reducing redundant calculations and memory accesses, and has a more significant effect on feature extraction. The detection accuracy is significantly improved in the insulator defect and bird nest intrusion data sets.

[0087] Table 2 Ablation experiment

[0088]

[0089] In order to further verify the performance of the algorithm, and considering the real-time requirements of the intelligent inspection task of power transmission lines, this experiment compares the YOLOv10-GCA algorithm with some current mainstream single-stage target detection algorithms. The results are shown in Table 3. Compared with the YOLOv7-tiny and YOLOv8 algorithms, YOLOv10-GCA has a lower GFLOPs and weight file, and the precision (P) is increased by 9.1 and 5.8 percentage points, the recall rate (R) is increased by 7.1 and 2.8 percentage points, and the mAP@50 is increased by 7.9 and 3.4 percentage points.

[0090] Compared with the latest real-time target detection algorithms YOLOv9 and YOLOv10, YOLOv10-GCA still has the highest detection performance. Compared with the YOLOv9 algorithm, YOLOv10-GCA has increased the precision (P), recall (R) and mAP@50 by 4.9, 3.0 and 2.9 percentage points respectively. Compared with the YOLOv10 algorithm, YOLOv10-GCA has increased the precision (P), recall (R) and mAP@50 by 2.6, 1.9 and 1.1 percentage points respectively. Experimental results show that compared with other single-stage target detection algorithms, the YOLOv10-GCA algorithm achieves extremely high detection accuracy with a lower model size, and still meets the real-time detection requirements, with high efficiency and excellence.

[0091] Table 3 Comparison results of different algorithms

[0092]

[0093] The comparison results show that YOLOv10-GCA can efficiently and accurately complete target detection tasks in different scenarios, proving that the improved YOLOv10-GCA algorithm has strong robustness and accuracy.

[0094] The present invention proposes an efficient and lightweight YOLOv10-GCA network model to detect and identify transmission line insulator defects and bird nest intrusions. The GhostConv module and the designed new module C2fGhost are introduced to reduce the number of model parameters and model size; the CA attention module is added to improve the feature fusion ability of the model; and the ASIoU is used as the loss function to improve the bounding box regression performance of the network model. The improved YOLOv10-GCA network model improves the accuracy and mAP@0.5 by 2.6% and 1.1% respectively compared with the original network model, and the number of network parameters and the amount of calculation are reduced by 29.6% and 13% respectively compared with the original network model. In the future, the network model will continue to be studied and improved based on this model, so that the model has a higher detection accuracy and a faster detection speed. Considering the practical application value of this application direction in power transmission management, the model will also be transplanted and verified and improved on the edge mobile platform in the future, making the model smaller and easier to deploy.

[0095] An embodiment of the present invention further provides a system for detecting defects in a power transmission line, which is used to implement the method for detecting defects in a power transmission line, comprising:

[0096] A video acquisition module, used to obtain video data of the power transmission line, and pre-process the video data of the power transmission line according to a preset data format rule to obtain an input item;

[0097] A recognition module, used to input the input items into a preset detection model to perform insulator defect and bird nest intrusion detection to obtain corresponding detection results; wherein the detection model recognizes and annotates the content of the input video data based on a pre-trained convolutional neural network to determine the insulator defects or bird nests present in the video data;

[0098] The alarm module is used to output an alarm signal and alarm information when there is an insulator defect or a bird's nest intrusion in the detection results.

[0099] It should be noted that the system described in the above embodiment corresponds to the method described in the above embodiment. Therefore, the undetailed parts of the system described in the above embodiment can be obtained by referring to the contents of the method described in the above embodiment, and will not be repeated here.

[0100] In summary, the implementation of the embodiments of the present invention has the following beneficial effects:

[0101] The power transmission line defect detection method and system provided by the present invention can reduce the number of parameters and calculations by reducing redundant calculations and memory accesses, and has a more significant effect on feature extraction, and significantly improves the detection accuracy in insulator defect and bird nest intrusion data sets. It achieves extremely high detection accuracy with a relatively low model volume, and still meets the real-time detection requirements, with high efficiency and excellence.

[0102] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for detecting defects in a power transmission line, characterized in that: include: Acquire video data of the power transmission line, and pre-process the video data of the power transmission line according to a preset data format rule to obtain an input item; Inputting the input items into a preset detection model to perform insulator defect and bird nest intrusion detection to obtain corresponding detection results; wherein the detection model recognizes and annotates the content of the input video data based on a pre-trained convolutional neural network to determine the insulator defects or bird nests present in the video data; When there are insulator defects or bird nest intrusions in the detection results, alarm signals and alarm information are output.

2. The method according to claim 1, characterized in that A convolution module is provided in the detection model, and the convolution module is used to compress the input item in the channel, and pass the compressed input item through multiple convolution layers in sequence to obtain multiple feature maps; All feature maps are concatenated in the channel direction to form a new feature map.

3. The method according to claim 2, characterized in that A CA module is also provided in the detection model, and the CA module is used to perform a pooling operation on each channel of the input image along the horizontal and vertical directions, and retain the position information of the target; The height and width are obtained at a certain channel in the horizontal and vertical directions, and a feature map is output according to the obtained height and width.

4. The method according to claim 3, characterized in that Pooling operations are performed along each channel in the horizontal and vertical directions according to the following formula, in, and is the encoding value output by the cth channel, W is the upper limit of width, H is the upper limit of height, u is the x-axis coordinate sequence, j is the y-axis coordinate sequence, h is the height sequence, w is the width sequence, and x c is the C channel input image.

5. The method according to claim 3, characterized in that The CA module is also used to embed the generated feature map into the splicing coordinate information according to the following formula: f=δ(F1([z h ,z w ])) Among them, f is the intermediate feature map encoded in the horizontal and vertical directions, δ is the nonlinear activation function, F1 is the splicing coordinate information operation, [z h ,z w ] is a concat operation along the spatial dimension.

6. The method according to claim 5, characterized in that The CA module is also used to perform dimensionality reduction according to the following formula, g h =δ(F h (f h )) g w =δ(F w (f w )) Among them, g w To decompose the feature map horizontally, g h To decompose the feature map along the vertical direction, f h is the intermediate feature map encoded in the vertical direction, f w is the intermediate feature map encoded in the horizontal direction, F h is the vertical dimension reduction operation, F w It is a horizontal dimension reduction operation.

7. The method according to claim 6, characterized in that According to the following formula, the tensors decomposed in the horizontal and vertical directions are fused with the input feature map to obtain the final output feature map. Among them, y c (i, j) is the final output feature map, x c (i, j) is the input feature map, is the tensor decomposed in the vertical direction, is the tensor decomposed horizontally.

8. The method according to claim 1, characterized in that A loss function is also set in the detection model, and the loss function is used to regress the target bounding box.

9. The method according to claim 8, characterized in that The loss function includes: Among them, A and B are the areas of the predicted box and the real box respectively, α is the parameter for adjusting the gradient convergence speed of the loss function, IoU is the IoU loss function, and L A-SIoU is the A-SIoU loss function, L dis is the distance loss, L shape is the shape loss.

10. A power transmission line defect detection system, used to implement the method according to any one of claims 1 to 9, characterized in that: include, A video acquisition module, used to obtain video data of the power transmission line, and pre-process the video data of the power transmission line according to a preset data format rule to obtain an input item; A recognition module, used to input the input items into a preset detection model to perform insulator defect and bird nest intrusion detection to obtain corresponding detection results; wherein the detection model recognizes and annotates the content of the input video data based on a pre-trained convolutional neural network to determine the insulator defects or bird nests present in the video data; The alarm module is used to output an alarm signal and alarm information when there is an insulator defect or a bird's nest intrusion in the detection results.

Images