A real-time fault detection method for transmission lines based on edge computing

Abstract

A real-time detection method for power transmission line faults based on edge computing includes the following steps: Step 1: Acquire a patrol image dataset containing images of insulator self-explosion, shock-absorbing hammer detachment, and bird's nests; Step 2: Expand the patrol images obtained in Step 1; Step 3: Label the expanded dataset using the LabelImg annotation tool; Step 4: Improve the convolution operation process for acquiring data features in the YOLOv5s basic model; Step 5: Adjust the target box positioning loss function; Step 6: Set the image size, momentum term, learning rate, batch size, number of training times, and weight decay regularization term, and train the improved YOLOv5s model; Step 7: Implant the trained model into the edge device, perform fault detection on the patrol blurred images, and obtain a performance evaluation value. The present invention provides a real-time detection method for power transmission line faults based on edge computing, which improves the fault detection speed and detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission line fault detection, and in particular to a real-time power transmission line fault detection method based on edge computing. Background Art

[0002] State Grid Corporation of China currently uses drones to detect critical faults in transmission lines, such as insulator explosions, falling shock-absorbing hammers, and bird nests. Currently, image data collected by drone inspections is primarily imported into offline servers for intelligent analysis using convolutional neural networks or manual analysis. This delays the feedback of inspection results, resulting in delays in timely on-site line restoration and further damage.

[0003] In recent years, with the rapid development of edge computing and the power Internet of Things (IoT), services are trending towards edge nodes. When edge computing devices are combined with inspection drones, real-time defect detection can be performed at the edge, improving inspection efficiency. However, in the actual application of edge computing combined with drones for transmission line inspections, fault detection algorithms have exposed shortcomings such as slow detection speed and susceptibility to image blur.

[0004] Currently, the detection algorithms used in edge computing modules are primarily single-stage object detection algorithms, which are known for their speed. Representative algorithms include YOLOv3, YOLOv4, YOLOv5, and SSD. These algorithms eliminate the need for pre-extraction of candidate regions and instead directly train the network end-to-end. Compared to two-stage object detection algorithms, they offer advantages such as a simpler network structure and better real-time performance. A modified SSD algorithm has been proposed for detecting insulator explosions and bird's nests, achieving high accuracy. However, it neglects the need for model compression, which hinders real-time detection performance when used directly on edge devices with limited computing power. Using YOLO as a base algorithm, some have implemented model compression improvements to detect insulator failures, bird's nests, anti-vibration hammers, and pole towers, achieving good real-time detection results. However, this algorithm only works when inspection images are clear and lacks effective detection methods for blurred image data. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a real-time detection method for transmission line faults based on edge computing to improve the fault detection speed and detection accuracy.

[0006] To solve the above technical problems, the technical solution adopted by the present invention is: a real-time detection method for transmission line faults based on edge computing, comprising the following steps:

[0007] Step 1: Obtain an inspection image dataset, which includes images of insulator explosions, shock-absorbing hammer detachments, and bird's nests.

[0008] Step 2: Expand the inspection image obtained in step 1;

[0009] Step 3: Label the augmented dataset using the LabelImg annotation tool;

[0010] Step 4: Improve the convolution operation process for obtaining data features in the YOLOv5s base model;

[0011] Step 5: Adjust the target box localization loss function;

[0012] Step 6: Set the image size, momentum term, learning rate, batch size, number of training epochs, and weight decay regularization term, and train the improved YOLOv5s model;

[0013] Step 7: Implant the trained model into the edge device, perform fault detection on the inspection blurred images, and obtain the performance evaluation values.

[0014] Preferably, in Step 2, to prevent overfitting in subsequent model training and affect the fault detection effect, the following data augmentations are performed on the samples: video segmentation, mosaic processing, random mirroring, and random rotation to generate similar but different training samples, thereby expanding the scale of the training dataset and improving the generalization ability of the model after training.

[0015] Preferably, the improvement of the convolution operation process for obtaining data features in Step 4 uses the Ghost module technology. The specific steps for obtaining the required n feature maps are as follows:

[0016] 1) First, obtain m feature maps Y' through a limited number of conventional convolutions:

[0017] Y' = X * ω' + b (4)

[0018] where Y' represents the output of m feature maps with heights and widths of h' and w' respectively, the convolution filter is ω', * represents the convolution operation, b is the bias term, and m ≤ n. The parameters such as the convolution kernel size, stride, and spatial size are the same as those of the conventional convolution.

[0019] 2) To obtain a total of n feature maps required, use a series of simple linear calculations to generate s features based on the feature map Y':

[0020] y i,j = Φ i,j (y i )', i ∈ [1, m], j ∈ [1, s] (7)

[0021] where y i ' represents the i-th directly concatenated map in the feature map Y', with a total of m; Φ i,j represents the j-th feature map y i,j, so by using y i,j m×s = n feature maps can be generated;

[0022] 3) Finally, directly splice the feature map Y' and the feature map y i ' in the specified dimension to obtain the required n feature maps.

[0023] Preferably, in the fifth step, the target box localization loss function is set as:

[0024] Replace the cross-entropy loss with a distribution loss function based on KL divergence:

[0025]

[0026] where Q D (x) is the true label coordinate probability distribution, and P Φ (x) is the predicted coordinate probability distribution. The higher the coincidence degree of the two probability distributions, the better the model robustness; x is the coordinate sample value, x e is the predicted coordinate position, x g is the true value of the target box coordinate, Φ represents a set of parameters to be learned, σ is the standard deviation of the distribution, and D is the parameter set expressing x g .

[0027] Preferably, the specific parameter settings for training the model in the sixth step are as follows: uniformly adjust the input image size to 640×640 pixels to accelerate the model training speed; use the asynchronous stochastic gradient descent method with a momentum term of 0.9 during training. Set the learning rate in the first 150 rounds to 0.001, and as the number of iteration rounds increases, reduce the learning rate in the next 150 rounds to 0.0001; to prevent model overfitting, set the weight decay regularization term to 0.0005, and after 300 rounds of model iteration training, obtain the final model weights.

[0028] Preferably, in the seventh step, implant the improved YOLOv5s algorithm into the edge device to detect the transmission line inspection blurred image, and verify the accuracy, size, and speed deployment of the detection model on the edge computing device through the model parameter quantity, mean average precision, and frame rate.

[0029] The present invention provides a real-time detection method for power transmission line faults based on edge computing. The method can be used in edge computing devices with limited computing power and memory, and can be used in conjunction with drones to perform real-time detection of insulator self-explosion, anti-vibration hammer shedding, and bird's nests in power transmission lines. It can also achieve high detection accuracy for blurred input target images caused during flight. The Ghost module is used to reconstruct the operational process of the YOLOv5s algorithm to obtain feature data, reducing the amount of calculation and parameters required to obtain all feature data using Conv convolution, thereby improving the detection speed without affecting the accuracy. A loss function based on the KL divergence distribution is used as the target frame positioning loss function. During the model training process, the variance between the movement and position of the target frame is learned, thereby enhancing the robustness of the detection model after training and improving the detection accuracy of blurred fault images. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0031] Figure 1 This is a flow chart of the real-time detection method for power transmission line faults based on edge computing of the present invention;

[0032] Figure 2 This is a diagram of the edge-based transmission line fault detection framework of the present invention;

[0033] Figure 3 It is a Conv convolution operation diagram of the feature calculation application in the implementation of the present invention;

[0034] Figure 4 It is an operation diagram of the Ghost module of the feature calculation application in the implementation of the present invention;

[0035] Figure 5 It is a structural diagram of the improved YOLOv5s algorithm used in the implementation of the present invention;

[0036] Figure 6 It is a schematic diagram of a data set for image preprocessing of the present invention;

[0037] Figure 7 This is the loss function value (Loss) convergence curve of the improved model of the present invention and the original model

[0038] Figure 8 This is the mean average precision (mAP) curve of the improved model of the present invention and the original model

[0039] Figure 9 This is the F1 curve diagram of the improved model of the present invention and the original model

[0040] Figure 10 This is the result of the SSD algorithm of the present invention detecting fuzzy fault images

[0041] Figure 11 This is the result of the YOLOv3 algorithm detecting fuzzy fault images.

[0042] Figure 12 This is the result of the YOLOv4 algorithm detecting fuzzy fault images.

[0043] Figure 13 This is the result of the YOLOv5s algorithm detecting fuzzy fault images.

[0044] Figure 14 This is the result of detecting blurred fault images using the improved YOLOv5s algorithm of the present invention. DETAILED DESCRIPTION

[0045] like Figure 1-2 As shown, a real-time detection method for transmission line faults based on edge computing includes the following steps:

[0046] Step 1: Obtain an inspection image dataset, which includes images of insulator explosions, shock-absorbing hammer detachments, and bird's nests.

[0047] Step 2: Expand the inspection image obtained in step 1;

[0048] Step 3: Use the LabelImg annotation tool to label the expanded dataset;

[0049] Step 4: Improve the convolution operation process of obtaining data features in the YOLOv5s basic model;

[0050] Step 5: Adjust the target box positioning loss function;

[0051] Step 6: Set the image size, momentum term, learning rate, batch size, number of training times, and weight decay regularization term to train the improved YOLOv5s model;

[0052] Step 7: Implant the trained model into the edge device, perform fault detection on the inspection fuzzy image, and obtain performance evaluation values.

[0053] Specifically, we used fault image data from a State Grid transmission line inspection, including 361 images of insulator self-explosion, 160 images of shock-proof hammer shedding, and 270 images of bird's nests.

[0054] Preferably, in step 2, to prevent overfitting during subsequent model training, which could affect fault detection, the following data augmentation techniques were applied to the samples: video segmentation, mosaic processing, random mirroring, and random rotation were used to generate similar yet distinct training samples, thereby expanding the size of the training dataset and improving the generalization ability of the trained model. The resulting dataset contained 1,575 fault images, including 723 images of insulator explosions, 319 images of shock absorbers falling off, and 533 images of the Bird's Nest.

[0055] Preferably, in the process of generating the feature map by using the Conv convolution in the Bottleneck module of the original YOLOv5 in step four, as Figure 3 shown. Given the input data X ∈ R c×h×w , where c is the number of input data channels, and h and w are the height and width of the input data respectively. After Conv convolution calculation, n features are generated:

[0056] Y = X * ω + b (1)

[0057] Among them, Y represents the n feature maps with height h' and width w' generated, * is the convolution operation, and b is the bias term. ω represents the convolution filter of the convolution kernel k×k. The floating-point operations per second F1 and the number of parameters P1 in the Conv convolution can be expressed as:

[0058] F1 = n × h' × w' × c × k × k (2)

[0059] P1 = n × c × k × k (3)

[0060] As Figure 4 shown, during the operation of the Ghost module, assume that a part of m (m ≤ n) feature maps Y' are generated by the Conv convolution:

[0061] Y' = X * ω' + b (4)

[0062] Among them, the convolution filter is ω'. At this time, the floating-point operations per second F2 and the number of parameters P2 are:

[0063] F2 = m × h' × w' × c × k × k (5)

[0064] P2 = m × c × k × k (6)

[0065] The remaining n - m feature maps need to be linearly transformed t - 1 times for m features, where t = n / m, and the subtracted "1" is the Figure 5 "1" time of linearly calculating m identical features in i . By using equation (7), m × t = n feature maps can be generated. By performing the Φ i,j linear operation on each basic feature of the feature y i ', m × (t - 1) operations are still required, and then they are directly concatenated with the feature map Y' into n features (this process does not generate computational costs). The linear operation is:

[0066] y i,j = Φ i,j (y i '), i ∈ [1, m], j ∈ [1, t] (7)

[0067] Among them, y i'represents the i-th feature map in Y', Φ i,j represents the j-th linear operation on the i-th feature map, y i,j The j-th feature map generated for this process. Let the linear operation kernel size be d×d, the floating-point operations per second F3 and the number of parameters P3 generated by the above linear process are:

[0068] F3 = (t - 1)×h'×w'×m×d×d (8)

[0069] P3 = (t - 1)×m×d×d (9)

[0070] Then, according to equations (2), (3), (5), (6), (8), (9), the computational cost ratio T of the Ghost module to the Conv convolution can be calculated f and the parameter ratio T p :

[0071]

[0072]

[0073] Taking k×k to be the same size as d×d and t = c, equations (10) and (11) can be further simplified to:

[0074]

[0075] It can be seen that the computational cost and the number of parameters of using the Ghost module compared to the Conv convolution are approximately 1 / t. The Ghost module replaces some of the Conv convolution operations with lower-cost linear operations, achieving the goal of reducing the computational cost and the number of parameters. Therefore, as Figure 5 in the improvement part, this paper uses the Ghost module to replace the Conv convolution in the Bottleneck module, which can not only ensure that the feature redundancy does not affect the accuracy of the YOLOv5s detection model; but also reduce the overall convolutional computational cost and the number of parameters of the model, improving the model fault detection speed in edge devices.

[0076] Preferably, in step five, the object confidence loss L conf 、the object category loss L cla 、the object bounding box localization loss L loc together constitute the loss function L of the YOLOv5s algorithm, and its expression is:

[0077] L = L conf + L cla + L loc (13)

[0078] The object bounding box loss function L locThe mean squared error loss form makes the loss value very sensitive to the changes in the target bounding boxes, thus affecting the robustness of the detection model after training. Considering that the KL loss can learn the variance between the movement and position of the target bounding boxes to reduce the value of the loss function L and improve the robustness of the model. Therefore, in this paper, L loc adopts the KL loss, and its definition is as follows:

[0079]

[0080] where Q D (x) is the probability distribution of the true label coordinates, P Φ (x) is the probability distribution of the predicted coordinates. The higher the coincidence degree of the two probability distributions, the better the robustness of the model; x is the coordinate sample value, x e is the predicted coordinate position, x g is the true value of the target bounding box coordinates, Φ represents a set of parameters to be learned, σ is the standard deviation of the distribution, and D is the parameter set expressing x g .

[0081] Preferably, the specific parameter settings for training the model in step six are as follows:

[0082] The input image size is uniformly adjusted to 640×640 pixels to accelerate the model training speed. The asynchronous stochastic gradient descent method with a momentum term of 0.9 is adopted during training, and each batch of training Batch size contains 64 pictures. The learning rate during training is set to 0.001 in the first 150 rounds, and as the number of iteration rounds increases, the learning rate of training is reduced to 0.0001 in the next 150 rounds. To prevent the model from overfitting, the weight decay regularization term is set to 0.0005, and after 300 rounds of model iteration training, the final model weights are obtained.

[0083] The training performance of the model before and after improvement is mainly reflected by the change of the loss function, the mean average precision, and the F1 change curve.

[0084] Preferably, in step seven, the improved YOLOv5s algorithm is first implanted into the Huawei Atlas 200DK edge device to detect the blurred images of power transmission line inspection. The accuracy, size, and speed of the detection model are verified to be suitable for deployment on edge computing devices through the number of model parameters, the mean average precision, and the frame rate.

[0085] In Example 1, a total of 2 models are trained, namely the original YOLOv5s model and the improved YOLOv5s model in this paper.

[0086] Using the dataset shown in Figure 6 to train the model, the convergence curves of the loss function values (Loss) of the network models before and after improvement are as shown in Figure 7As shown in the figure. Compared with the original model, the initial Loss value of the improved model is reduced to 11, which is more conducive to the convergence of the model. This is because the improved loss function reduces the interference during the backpropagation process, resulting in a smaller initial Loss. As the number of iterations increases, the Loss value of the improved model drops to around 2 after 100 rounds of iteration and finally converges to around 1, while the final Loss value of the original model converges to around 2.5. The data shows that the optimization of the improved model in the selection of the target box localization loss function makes the model have better convergence and training effect.

[0087] The mean average precision (mAP) curves of the improved model and the original model are as Figure 8 shown. It can be seen that after the original model and the improved model are iterated to 50 and 30 rounds respectively, the mAP value rises to 0.6, and finally stabilizes at 0.79 and 0.84 respectively. The results show that compared with the original model, the improved model has a faster detection convergence speed and higher detection accuracy.

[0088] The harmonic mean F1 curve of precision and recall is as Figure 9 shown. The F1 value is an indicator that can represent that both precision and recall reach the highest and achieve a balance, and can better reflect the detection performance of the model. It can be seen from the figure that the F1 value of the improved model gradually becomes higher than that of the original model as the number of iterations increases, and finally stabilizes at 0.81, which is higher than 0.78 of the original model. Therefore, the improved model has an obvious advantage in the detection effect considering both precision and recall.

[0089] Furthermore, to verify the detection effects of different improvement strategies on the YOLOv5s model, the improvement strategy selections shown in Table 1 are set and tested respectively. Number 1 is the original YOLOv5s model. The Ghost module improvement is to improve the detection speed of the model, while the KL divergence loss is to improve the detection accuracy of the model. Therefore, the mAP values of models 3 and 4 are significantly improved compared with those of models 1 and 2, and their detection accuracy is improved by more than 5% compared with the original YOLOv5s model.

[0090] Table 1 Comparison of the effects of different improvement strategies of the YOLOv5s model

[0091]

[0092] In Example 2, this specific implementation designed five different sets of comparative fault detection experiments to verify the superiority of the improved YOLOv5s model based on the Ghost convolution module in fault detection speed in edge computing devices. The mainstream real-time detection models shown in Table 2 were deployed on the Huawei Atlas 200DK edge device for testing. All models were trained using the same data samples and parameters. The results of the model parameters, mean average precision (mAP), and frame rate (FPS) are shown in Table 2.

[0093] Table 2 Performance comparison of mainstream models

[0094]

[0095] Analysis of the data in Table 2 shows that the YOLOv4 model achieves the highest mAP value of 85.32%. However, its model parameter size is the largest at 246MB, and its FPS is the slowest, resulting in slower detection speed. The YOLOv4 model, along with its similar SSD and YOLOv3 models, is not suitable for real-time detection on edge computing devices. While YOLOv5s and our proposed method achieve slightly lower detection accuracy than YOLOv4, their detection speed is 3.1 times faster, meeting both high accuracy and real-time detection requirements. Compared to YOLOv5s, our proposed model, improved with the Ghost module, reduces the model parameter size from 7.46MB to 4.39MB and increases the FPS from 23 to 34, making it more suitable for edge computing devices with limited computing power and memory. When used in edge computing modules on drones, our proposed model ensures high detection accuracy and good real-time performance for power transmission line faults.

[0096] To further verify the superiority of the improved KL divergence loss function in detecting blurred target images in the actual application of the detection model, this paper uses blurred image samples of insulator self-explosion, shock-absorbing hammer shedding, and bird's nest to test and compare the real-time detection model of this paper with YOLOv3, YOLOv4, YOLOv5s, and SSD real-time detection models. The effects of different models on the detection of blurred or deformed inspection fault images are as follows. Figures 10-14 As shown, in comparison, Figures 10-14 The detection results of the four models shown in the figure all showed missed detection. Figure 14 The improved YOLOv5s algorithm in this paper can accurately detect three types of fault targets in blurred images: insulator explosion, shock-absorbing hammer detachment, and bird's nests, demonstrating stronger detection performance. Therefore, the improved model based on the KL divergence loss function has better detection results for blurred or deformed images.

[0097] The above embodiments are only the preferred technical solutions of the present invention and should not be regarded as limitations on the present invention. The protection scope of the present invention should be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A real-time fault detection method for transmission lines based on edge computing, characterized in that, The following steps are involved: Step 1: Obtain an inspection image dataset, which includes images of insulator explosions, shock-absorbing hammer detachments, and bird's nests. Step 2: Expand the inspection image obtained in step 1; Step 3: Use the LabelImg annotation tool to label the expanded dataset; Step 4: Improve the convolution operation process of obtaining data features in the YOLOv5s basic model; Step 5: Adjust the target box positioning loss function; Step 6: Set the image size, momentum term, learning rate, batch size, number of training times, and weight decay regularization term to train the improved YOLOv5s model; Step 7: Embed the trained model into the edge device, perform fault detection on the inspection fuzzy image, and obtain performance evaluation values; The convolution operation process for obtaining data features in step 4 is improved using the Ghost module technology to obtain the required specific steps for the number of feature maps are as follows: 1) First, obtain feature maps by means of a limited number of conventional convolutions: ；（4） Among them indicates that the output is feature maps with heights and widths of and respectively, and the convolutional filter is , represents the convolutional operation, is the bias term, ; the convolutional kernel size, stride, and spatial size parameters are the same as those of conventional convolution; 2) To obtain a total of feature maps, a series of simple linear calculations are used to generate on the basis of the feature maps features: ；(7) Among them represents the th directly concatenated graph in the feature map, with a total of ; represents the th feature map generated by the th linear operation. Therefore, is used to generate feature maps; ​ 3) Finally, the feature map and the feature map are directly concatenated in the specified dimension to obtain the required feature maps.

2. The real-time transmission line fault detection method based on edge computing according to claim 1, characterized in that In step 2, in order to prevent overfitting in subsequent model training and affect the fault detection effect, the following data enhancements were performed on the samples: video segmentation, mosaic processing, random mirroring, and random rotation to generate similar but different training samples, thereby expanding the scale of the training data set and improving the generalization ability of the model after training.

3. The real-time transmission line fault detection method based on edge computing according to claim 1, characterized in that, In step 5, the target frame positioning loss function is set to: Replace the cross entropy loss with a distribution loss function based on KL divergence: ；(14) Among them, is the probability distribution of the true label coordinates, is the probability distribution of the predicted coordinates. The higher the coincidence degree of the two probability distributions, the better the robustness of the model; is the value of the coordinate sample, is the predicted coordinate position, is the ground truth of the target box coordinates, represents a set of parameters to be learned, is the standard deviation of the distribution, and D is the parameter set representing .

4. The real-time transmission line fault detection method based on edge computing according to claim 1, wherein, The specific parameters of the training model in step six are set as follows: the input image size is uniformly adjusted to 640×640 pixels to speed up the model training speed; during training, an asynchronous stochastic gradient descent method with a momentum term of 0.9 is used, and the learning rate in training is set to 0.001 in the first 150 rounds. As the number of iterations increases, the learning rate in training is reduced to 0.0001 in the next 150 rounds; in order to prevent the model from overfitting, the weight decay regularization term is set to 0.0005, and the final model weight is obtained after 300 rounds of model iterative training.

5. The real-time transmission line fault detection method based on edge computing according to claim 1, characterized in that, In step seven, the improved YOLOv5s algorithm is implanted into the edge device to detect the fuzzy image of the power transmission inspection. The accuracy, size, and speed of the detection model are verified by the model parameter quantity, average accuracy mean, and frame rate. The model is deployed on the edge computing device.

Images