Power transmission line insulator abnormity marking system and method based on edge intelligence

Abstract

The invention belongs to the technical field of mobile edge computing, and particularly relates to a power transmission line insulator abnormity marking system and method based on edge intelligence. The method comprises the following steps: S1, preprocessing a power transmission line video image data set based on a fuzzy enhancement and segmentation algorithm to obtain a training set; s2, constructing an insulator abnormity marking network comprising a student model and a teacher model; s3, the knowledge distillation module inputs the training set to the student model and the teacher model at the same time, and trains the insulator abnormity marking network to obtain an insulator abnormity marking model; and S4, inputting the to-be-processed image into the insulator abnormity marking model for marking, and completing abnormity marking of the insulator. Abnormal problems such as power transmission line insulator damage and fracture can be detected. According to the invention, the insulator can be quickly identified and positioned, the insulator abnormity can be timely detected, and the safe operation of the power transmission line is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the field of mobile edge computing technology, and in particular relates to a transmission line insulator abnormality marking system and method based on edge intelligence. Background Art

[0002] With the interconnection of national energy, UHV transmission projects play a key role. Transmission lines and their accessories, such as poles and insulators, are exposed to harsh environments for a long time and face damage from a variety of natural and biological factors. This requires the power sector to regularly monitor the status of insulators to ensure timely maintenance and replacement to prevent power outages. As a key accessory, the status monitoring of insulators is crucial to power transmission. At present, my country has turned to an automated inspection mode based on machine inspections, using artificial intelligence algorithms to perform intelligent diagnosis of drone aerial images, promoting the development of inspection methods towards intelligence and automation, and providing solid guarantees for the stable operation of the power system.

[0003] Mobile edge computing (MEC) is a new type of processing device. Unlike traditional cloud computing, MEC can effectively solve the delay in data transmission backhaul, help collect and process data on servers near the power grid, and reduce the transmission cost and processing time of data computing. At the same time, it greatly reduces the delay and power consumption of data processing. At the same time, edge intelligence technology embeds artificial intelligence (AI) algorithms on the edge side to achieve data preprocessing and local computing, reduce cloud communication requirements, and reduce latency and power consumption.

[0004] However, existing single-stage target marking algorithms based on deep learning at home and abroad, such as YOLO and SSD, have the advantages of simple structure and good real-time performance, but there are problems when they are directly applied to the marking of abnormal insulators on transmission lines. In particular, the marking effect is poor under fuzzy image data or complex background conditions, and the algorithm model is huge, the amount of calculation is large, and the marking speed is slow. When marking abnormalities of key components of transmission lines, deep learning algorithms face the following challenges: 1) Complex backgrounds lead to low recognition accuracy. Affected by shooting angles, lighting, and weather, it is crucial to accurately identify the boundaries and contours of components. 2) Inefficient resource scheduling in the cloud prolongs the feedback time and affects the recovery of the line. Mobile edge computing technology provides a solution for fast response and efficient resource allocation, but computing power limitations may affect marking speed and accuracy. Summary of the invention

[0005] In view of this, the present invention aims to provide a transmission line insulator abnormality marking system and method based on edge intelligence to solve the problems of low accuracy in identifying insulators in complex backgrounds, inefficient cloud server resource scheduling, slow feedback time of intelligent analysis and processing, and untimely marking of insulator abnormalities. The present invention provides guarantee for the safe operation of transmission lines.

[0006] To achieve the above object, the technical solution created by the present invention is implemented as follows: An abnormality marking system for insulators of power transmission lines based on edge intelligence includes a drone, an edge end, a cloud, and a knowledge distillation module. The edge end is mounted on the drone and performs cloud-edge collaboration with the cloud. A YOLOv8s model is deployed on the cloud, and an improved YOLOv8s network is deployed on the edge. The UAV collects the transmission line video image dataset, and pre-processes the transmission line video image dataset based on fuzzy enhancement and segmentation algorithm to obtain the training set; The improved YOLOv8s network on the edge is used as the student model, and the YOLOv8s model on the cloud is used as the teacher model to build an insulator anomaly marking network including the student model and the teacher model. The knowledge distillation module inputs the training set into the student model and the teacher model at the same time, and trains the insulator abnormality marking network to obtain the insulator abnormality marking model; The image to be processed is input into the insulator abnormality marking model for marking, thereby completing the abnormality marking of the insulator.

[0007] Furthermore, the improved YOLOv8s network improves the Bottleneck module of the YOLOv8s model by replacing the C3 module in the Bottleneck module of the YOLOv8s model with a Ghost convolution module and a C2F module, and using the output of the Ghost convolution module as the input of the C2F module.

[0008] Furthermore, in the process of training the insulator abnormality labeling network: the cross entropy of the teacher model and the student model themselves are calculated respectively, and the insulator abnormality labeling model is updated, the teacher model obtains the first label, and the student model obtains the second label; the KL-divergence between the first label and the second label is calculated, and the result of the proportional sum of the cross entropy and the KL-divergence is used as the loss function of the training.

[0009] Furthermore, the fuzzy enhancement and segmentation algorithm is composed of an ant colony intelligence algorithm and a weighted variable FCM algorithm.

[0010] Furthermore, preprocessing the transmission line video image data set includes: performing fuzzy image enhancement and image segmentation on each transmission line video image to obtain an insulator image and a background image.

[0011] A transmission line insulator abnormality marking method based on edge intelligence is implemented by using a transmission line insulator abnormality marking system based on edge intelligence, and specifically includes the following steps: S1: The UAV collects transmission line video images, and pre-processes the transmission line video image dataset based on fuzzy enhancement and segmentation algorithms to obtain a training set; S2: The improved YOLOv8s network on the edge is used as the student model, and the YOLOv8s model on the cloud is used as the teacher model to build an insulator anomaly marking network including the student model and the teacher model; S3: The knowledge distillation module inputs the training set into the student model and the teacher model at the same time, and trains the insulator abnormality marking network to obtain the insulator abnormality marking model; S4: Input the image to be processed into the insulator abnormality marking model for marking, thereby completing the abnormality marking of the insulator.

[0012] Furthermore, step S1 specifically includes: S11: construct a sliding window on the transmission line video image, and calculate the mean and standard deviation of the sliding window; S12: Calculate the adaptive threshold T' based on the mean and standard deviation of the current sliding window: T'=m+k×σ; Among them, m is the mean, σ is the standard deviation, and k is a constant; S13: using the adaptive threshold T' to perform binarization processing on the current sliding window to obtain a fuzzy set membership diagram of the current sliding window; S14: repeating steps S12-S13 until the entire transmission line video image is traversed, and the fuzzy set membership graphs obtained by each sliding window are spliced ​​to obtain a complete fuzzy set membership graph; S15: Fuse the original transmission line video image with the corresponding complete fuzzy set membership graph to obtain a fuzzy enhanced image: ; ; Among them, I(x,y) is the original transmission line video image, The pixel (x, y) of the original image to the blurred center A The distance is the calibration parameter, which is 2 here, and m is the fuzzy factor, which is 1.5 here. is the complete fuzzy set membership graph, Enhance images for blur; S16: Calculate local feature information of the fuzzy enhanced image, where the local feature information includes grayscale value, gradient amplitude and texture feature; S17: Initialize the parameters of the ant colony intelligence algorithm, including the number of ants and the concentration of pheromones; S18: using an ant colony intelligent algorithm to classify the fuzzy enhanced image, dividing the fuzzy enhanced image into a plurality of preliminary segmentation regions; and determining an initial clustering center along the distribution of local feature information of the fuzzy enhanced image; S19: Based on the initial cluster centers, the fuzzy enhancement and segmentation algorithms are used to optimize and refine each preliminary segmentation area to obtain the insulator image and the background image.

[0013] Furthermore, step S19 specifically includes the following steps: S191: Initialize the initial cluster center and membership matrix, the membership matrix is ​​obtained based on the complete fuzzy set membership graph; S192: Iteratively update the initial cluster center and membership matrix using the following formula: ; ; in, is the i-th sample point The degree of membership to the kth category, ∈[0, 1], is the i-th sample point With cluster center The distance is the i-th sample point With cluster center distance, c is the expected number of categories, m is the fuzzy factor, is the center of the kth cluster, is the center of the jth cluster, N is the number of samples involved in the cluster analysis, is the i-th sample point The weight of is the i-th sample point For the mth power of the membership degree of the kth category; S193: Obtain a final segmentation result according to the updated membership matrix, and post-process the final segmentation result to obtain an insulator image and a background image.

[0014] Further, in step S193, post-processing includes small area merging and boundary smoothing.

[0015] Furthermore, in step S3, the Adam optimization algorithm is used to update the weight of the student model to minimize the loss function; at the same time, the model pruning technology or quantization compression technology is applied to reduce the student model so that the student model matches the computing power of the edge.

[0016] Compared with the prior art, the invention can achieve the following beneficial effects: (1) The present invention creates the transmission line insulator abnormality marking system and method based on edge intelligence, proposes a fuzzy enhancement and segmentation algorithm, combines the global exploration capability of the ant colony intelligence algorithm and the local optimization capability of C-means clustering (FCM), and can better handle the segmentation problem of complex images. The ant colony algorithm first finds the preliminary segmentation area to determine the initial clustering center, and the fuzzy enhancement and segmentation algorithm further optimizes and refines the segmentation result, thereby having a more accurate segmentation effect and achieving the separation of insulators and the background of complex transmission lines.

[0017] (2) The present invention creates the transmission line insulator abnormality marking system and method based on edge intelligence, designs a lightweight deep learning algorithm for real-time abnormality marking of intelligent transmission line insulators, which is suitable for running on the edge. It solves the problems of large computational complexity, abnormalities (insulator string drop, sheet drop, cracks and pores, etc.) and slow marking speed of the existing insulator abnormality marking model, and significantly improves the speed and efficiency of insulator abnormality marking.

[0018] (3) The present invention creates the transmission line insulator abnormality marking system and method based on edge intelligence, and proposes a method for autonomous online learning and updating models based on knowledge distillation. Based on the knowledge distillation module, the lightweight deep learning model can be optimized in real time and adapt to changing environmental conditions. The present invention allows the improved YOLOv8s network on the edge to remain lightweight while also obtaining the advanced feature representation and knowledge of the YOLOv8s model on the cloud. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings constituting part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments and descriptions of the present invention are used to explain the present invention and do not constitute an improper limitation on the present invention. In the drawings: Figure 1 A diagram showing the execution effect of the method for marking abnormality of insulators of a power transmission line based on edge intelligence according to an embodiment of the present invention; Figure 2 A schematic diagram of the structure of the autonomous online learning and updating model based on knowledge distillation according to an embodiment of the present invention; Figure 3 A flow chart of a method for marking abnormality of insulators of a transmission line based on edge intelligence as described in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and do not constitute a limitation of the invention.

[0021] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0022] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0023] In the description of the invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installation", "connection" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the invention can be understood according to specific circumstances.

[0024] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0025] On the one hand, the present invention provides a transmission line insulator abnormality marking system based on edge intelligence, including a drone, an edge end, a cloud and a knowledge distillation module, the edge end is mounted on the drone, and performs cloud-edge collaboration with the cloud; a YOLOv8s model is deployed on the cloud, and an improved YOLOv8s network is deployed on the edge; the drone collects a transmission line video image data set, and pre-processes the transmission line video image data set based on a fuzzy enhancement and segmentation algorithm to obtain a training set; the improved YOLOv8s network on the edge end is used as a student model, and the YOLOv8s model on the cloud end is used as a teacher model to construct an insulator abnormality marking network including a student model and a teacher model; the knowledge distillation module simultaneously inputs the training set into the student model and the teacher model, and trains the insulator abnormality marking network to obtain an insulator abnormality marking model; the image to be processed is input into the insulator abnormality marking model for marking, and the abnormal marking of the insulator is completed.

[0026] like Figure 1 As shown, the method proposed in this method can firstly segment the insulators from the complex transmission line background and mark them, and then based on the marking results, the lightweight anomaly representation deep model (i.e., the improved YOLOv8s network) built at the edge can be used to detect and mark the insulator abnormality information in real time.

[0027] like Figure 2 As shown in the figure, in the cloud, the input image is first processed using cloud resources to run the complete YOLOv8s model. The YOLOv8s model includes a backbone network, a neck network, and a head network. The backbone network and the neck network are connected using the C3 module, and the neck network and the head network are connected using the C3 module. Among them, the C3 module uses an 8x8 convolution with a step length of 4 and a SiLU activation function, followed by a 3x3 convolution with a step length of 1 and a SiLU activation function. The neck network is used for feature fusion and enhancement. The head network uses the Softmax function to output the detection results, and the loss distribution of the teacher model is migrated to the edge through the knowledge distillation technology. The specific cloud processing flow is as follows: the input image is subjected to preliminary feature extraction through the backbone network, the extracted features are further processed through the C3 module, and the processed features are passed to the neck network for feature fusion. Finally, the features are classified through the head network, and the detection results are output using the Softmax function.

[0028] At the edge, an improved lightweight YOLOv8s network is run to adapt to the resource constraints of edge devices. The original C3 module is replaced with Ghost convolution and C2f modules to reduce the amount of computation. The backbone network is connected to the neck network using Ghost+C2f, and the neck network and the head network are connected with the C2f module. The neck network also contains a C2f module for feature processing. The head network uses the Softmax function for output and calculates the loss function through KL divergence to optimize the model. The edge processing flow is as follows: the input image is extracted through a lightweight backbone network, specifically using Ghost convolution and C2f modules, the extracted features are processed by the C2f module of the neck network, the processed features are passed to the head network for classification, and the detection results are output using the Softmax function.

[0029] In the process of knowledge distillation, the output of the cloud teacher model is transferred to the edge student model through knowledge distillation. The edge student model improves performance by imitating the behavior of the cloud model. The specific process is as follows: The output of the cloud teacher model is converted into a probability distribution through the Softmax function. The output of the edge student model is also converted into a probability distribution through the Softmax function. The KL divergence between the two probability distributions is calculated as the loss function. By optimizing this loss function, the edge student model learns the behavior of the cloud teacher model and realizes knowledge distillation.

[0030] At the edge, the model is trained by minimizing the loss function, which includes KL divergence and model loss distribution. In this way, the edge student model can achieve performance close to that of the cloud teacher model under resource constraints, significantly reduce the number of parameters, and increase the training speed to achieve real-time detection.

[0031] It should be noted that model evaluation indicators usually include accuracy, precision, recall, F1 score, parameter size, etc. The present invention mainly considers three aspects: accuracy, parameter size and operand size.

[0032] In some embodiments, the improved YOLOv8s network improves the Bottleneck module of the YOLOv8s model, replaces the C3 module in the Bottleneck module of the YOLOv8s model with a Ghost convolution module and a C2F module, and uses the output of the Ghost convolution module as the input of the C2F module.

[0033] In some embodiments, during the training of the insulator abnormality marking network: the cross entropy of the teacher model and the student model are calculated respectively, and the insulator abnormality marking model is updated, the teacher model obtains the first label, and the student model obtains the second label; the KL-divergence between the first label and the second label is calculated, and the result of the proportional summation of the cross entropy and the KL-divergence is used as the loss function of the training.

[0034] It should be noted that the C3 module is replaced by the C2f module and multi-feature linear transformation is performed, and the convolution layer in the YOLOv8s model is replaced by the Ghost convolution module, which greatly reduces the number of parameters in the original model; a lightweight improved YOLOv8s model is used for feature extraction, and the Ghost convolution module is used for image segmentation, and then the SiLU activation function is used for nonlinear transformation, making it more suitable for insulator images with complex backgrounds, which will reduce the number of model parameters and calculations and improve the network convergence speed.

[0035] In some embodiments, preprocessing the power transmission line video image data set includes: performing fuzzy image enhancement and image segmentation on each power transmission line video image to obtain an insulator image and a background image.

[0036] It should be noted that the preprocessing of the transmission line video image improves the generalization ability of the lightweight insulator abnormality marking model running on the edge drone; the model is trained using the labeled training set, and the network weights are updated through the back propagation algorithm. In addition, in the process of fuzzy image enhancement, the relative grayscale level of the pixel is selected as the fuzzy feature, and the transmission line video image is fuzzy enhanced by combining the adaptive selection method of the adaptive threshold. The transmission line video image speeds up the selection of the optimal parameters to ensure the quality of the enhanced image, and improves the feasibility and efficiency of the fuzzy enhancement and segmentation algorithm.

[0037] In some embodiments, the drone is equipped with a high-definition camera that can shoot transmission lines from different angles and heights to capture high-definition images of insulators. These images contain a lot of background information, including the sky, vegetation, towers, etc., which makes the identification of insulators complicated.

[0038] In some embodiments, the fuzzy enhancement and segmentation algorithm is composed of an ant colony intelligence algorithm and a weighted variable FCM algorithm.

[0039] It should be noted that the combination of the ant colony intelligence algorithm and the weighted variable FCM algorithm can better solve the problem of random initialization of cluster centers and improve the accuracy of segmenting insulators from complex backgrounds. The calculation formula of the weighted variable FCM algorithm is: ; in, is the center of the kth cluster, N is the number of samples involved in the cluster analysis, is the i-th sample point The weight of is the i-th sample point The mth power of the membership degree of the kth category. Weights are usually used to adjust the influence of each sample in the calculation of the cluster center. The weights can be set according to the importance, confidence or other characteristics of the sample.

[0040] In order to better understand the relationship between the cloud, edge, and knowledge distillation modules, the following is a detailed introduction to the connection between the three. The cloud is the cloud service end, and the edge is the edge computing device end. The cloud server and the edge computing device end perform cloud-edge collaboration to train and update the lightweight deep learning model (i.e., the improved YOLOv8s network) on the edge end. The cloud and edge end will jointly participate in the training and updating process of the model. First, the cloud server processes large-scale data sets such as aerial images and deep learning parameters to train and optimize complex neural network models. The cloud quickly iterates the model through powerful computing power to improve the accuracy and robustness of the model. After training, the model is compressed and optimized to adapt the model to the computing power of the edge device; then, the edge end is responsible for real-time processing of image data collected by the drone, and uses the iterated model for rapid anomaly detection. Through cloud-edge collaboration, the edge end can send detection results and feedback information to the cloud end, and the cloud end further optimizes the model based on this information and updates the edge end model regularly or as needed.

[0041] In order to improve the updating efficiency of the model, ensure that the model can adapt to the changing environment and conditions, and improve the accuracy and real-time performance of transmission line insulator anomaly detection, the knowledge distillation module is used for online learning and model updating. First, the lightweight model on the edge is used as the student model, and the complete model on the cloud is used as the teacher model. Through knowledge distillation, the student model can learn the decision boundary and feature representation of the teacher model, and maintain high detection performance even under resource constraints; then, the online learning mechanism of the knowledge distillation module allows the model to be continuously updated and optimized based on the data collected in real time. Whenever the drone captures new image data, the lightweight model on the edge performs real-time detection and feeds the results back to the cloud. Based on this feedback information, the cloud uses knowledge distillation technology to update the teacher model, and then passes the updated knowledge to the student model to achieve continuous learning and evolution of the model.

[0042] like Figure 3 As shown, on the other hand, the present invention proposes a transmission line insulator abnormality marking method based on edge intelligence, which is implemented using a transmission line insulator abnormality marking system based on edge intelligence, and specifically includes the following steps: S1: The UAV collects transmission line video images, and preprocesses the transmission line video image data set based on the fuzzy enhancement and segmentation algorithm to obtain a training set; S2: The improved YOLOv8s network on the edge is used as the student model, and the YOLOv8s model on the cloud is used as the teacher model to construct an insulator abnormality marking network including the student model and the teacher model; S3: The knowledge distillation module inputs the training set into the student model and the teacher model at the same time, and trains the insulator abnormality marking network to obtain an insulator abnormality marking model; S4: The image to be processed is input into the insulator abnormality marking model for marking, and the abnormal marking of the insulator is completed.

[0043] In some embodiments, step S1 specifically includes: S11: construct a sliding window on the transmission line video image, and calculate the mean and standard deviation of the sliding window; S12: Calculate the adaptive threshold T' based on the mean and standard deviation of the current sliding window: T'=m+k×σ; Among them, m is the mean, σ is the standard deviation, and k is a constant; S13: using the adaptive threshold T' to perform binarization processing on the current sliding window to obtain a fuzzy set membership diagram of the current sliding window; S14: repeating steps S12-S13 until the entire transmission line video image is traversed, and the fuzzy set membership graphs obtained by each sliding window are spliced ​​to obtain a complete fuzzy set membership graph; S15: Fuse the original transmission line video image with the corresponding complete fuzzy set membership graph to obtain a fuzzy enhanced image: ; ; Among them, I(x,y) is the original transmission line video image, The pixel (x, y) of the original image to the blurred center A The distance is the calibration parameter, which is 2 here, and m is the fuzzy factor, which is 1.5 here. is the complete fuzzy set membership graph, Enhance images for blur; S16: Calculate local feature information of the fuzzy enhanced image, where the local feature information includes grayscale value, gradient amplitude and texture feature; S17: Initialize the parameters of the ant colony intelligence algorithm, including the number of ants and the concentration of pheromones; S18: using an ant colony intelligent algorithm to classify the fuzzy enhanced image, dividing the fuzzy enhanced image into a plurality of preliminary segmentation regions; and determining an initial clustering center along the distribution of local feature information of the fuzzy enhanced image; S19: Based on the initial cluster centers, the fuzzy enhancement and segmentation algorithms are used to optimize and refine each preliminary segmentation area to obtain the insulator image and the background image.

[0044] It should be noted that in step S15, the weights can be adjusted during fusion. to control the blur enhancement effect.

[0045] It should be noted that in step S18, the ant agent explores along the local feature distribution of the image, gradually finds the optimal classification path, and thus determines the initial clustering center.

[0046] In some embodiments, step S19 specifically includes the following steps: S191: Initialize the initial cluster center and membership matrix, the membership matrix is ​​obtained based on the complete fuzzy set membership graph; S192: Iteratively update the initial cluster center and membership matrix using the following formula: ; ; in, is the i-th sample point The degree of membership to the kth category, ∈[0, 1], is the i-th sample point With cluster center The distance is the i-th sample point With cluster center distance, c is the expected number of categories, m is the fuzzy factor, is the center of the kth cluster, is the center of the jth cluster, N is the number of samples involved in the cluster analysis, is the i-th sample point The weight of is the i-th sample point For the mth power of the membership of the kth category; generally the value is between (1, 2). When m approaches 1, the degree of fuzziness decreases and the membership function is more like a hard segmentation; when m approaches 2, the degree of fuzziness increases and the membership function is smoother. is the cluster center of the kth category, N is the number of samples involved in the cluster analysis, For sample points The weight is usually used to adjust the influence of each sample point in the cluster center calculation. The weight can be set according to the importance, confidence or other characteristics of the sample. For sample points The mth power of the membership degree of the kth category indicates the influence of the sample point in the calculation of the cluster center.

[0047] S193: Obtain a final segmentation result according to the updated membership matrix, and post-process the final segmentation result to obtain an insulator image and a background image.

[0048] It should be noted that the combination of the global exploration capability of the ant colony intelligence algorithm and the local optimization capability of the weighted variable FCM algorithm can better handle complex image segmentation problems. The ant colony intelligence algorithm first finds the preliminary segmentation area to determine the initial cluster center, and the fuzzy enhancement and segmentation algorithm further optimizes and refines the segmentation results, thereby obtaining a more accurate segmentation effect and achieving the effect of separating insulators from complex transmission line backgrounds.

[0049] Further, in step S193, after multiple update iterations of step S192, when the membership matrix converges, the final category of each sample point can be determined. The specific steps include: 1) Selecting a category: For each sample point , select the cluster center with the highest membership as its category, which will assign a category label to each pixel (e.g., insulator or background). 2) Generate a mask based on the final category label, and draw the insulator area and background area into different colors or binary images respectively. Extract the insulator image and background image according to the mask to obtain the final segmentation result.

[0050] In some embodiments, in step S193, post-processing includes small area merging and boundary smoothing.

[0051] In some embodiments, in step S3, the Adam optimization algorithm is used to update the weight of the student model to minimize the loss function; at the same time, the model pruning technology or quantization compression technology is applied to reduce the student model so that the student model matches the computing power of the edge.

[0052] In some embodiments, the improved YOLOv8s network applies batch normalization to accelerate the training of a lightweight insulator anomaly marking model running on an edge drone and improve its stability; a cross entropy loss function is defined for error calculation during training, the cross entropy loss function updates the YOLOv8s model and the improved YOLOv8s network respectively, the Adam optimizer is selected, parameters are adjusted during training, and then the performance is evaluated using a validation set to monitor the accuracy and loss value during training.

[0053] In some embodiments, the trained improved YOLOv8s network is deployed to an edge server to ensure that the improved YOLOv8s network can operate in a resource-constrained environment. On the edge computing device side, the deployed model is used to perform fault detection on real-time insulator image data, and manual review or automatic feedback adjustment is performed based on the output fault detection results.

[0054] It should be noted that the improved YOLOv8s network is a lightweight model suitable for running on edge computing devices. The lightweight model has fewer parameters and computing requirements.

[0055] In some embodiments, a teacher model is used to process input data and extract high-level feature representations; a student model is trained to mimic the behavior of the teacher model and minimize the difference between the student model output and the teacher model output.

[0056] It should be noted that the trained and compressed student model is deployed on the edge computing device, where the student model acts as an online intelligent agent to process new input data in real time and make predictions.

[0057] In some embodiments, a lightweight insulator anomaly indication model running on an edge drone is trained to optimize the model's online learning process; computing resources of edge computing devices are managed, and the model is updated according to the guidance of the policy network; the performance of the improved YOLOv8s network at the edge is regularly evaluated to ensure its adaptability and meet real-time requirements; and the improved YOLOv8s network is continuously updated and iterated based on the results of online learning and performance evaluation.

[0058] As shown in Table 1, the model proposed in the present invention (a lightweight insulator anomaly marking model running on an edge drone) is compared with the YOLOV8s, MobileNetV3, and Ghost models in terms of performance, including the number of operations (in GFLOPS), accuracy, and parameter quantity (in megaparameter quantity / M). It can be seen that since the YOLOV8s model is a non-lightweight model, although its detection accuracy is the highest, the model operation and parameter quantity are both large, and it is not suitable for running on the edge. The MobileNetV3 model and the Ghost model are lightweight models suitable for running on the edge, but their operation and parameter quantities are higher than the model proposed in this article, and since the present invention proposes an autonomous online learning and updating method based on knowledge distillation, the accuracy is greatly improved with a small number of operations and parameters.

[0059] Table 1

[0060] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0061] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A transmission line insulator abnormality marking system based on edge intelligence, characterized by: It includes a drone, an edge end, a cloud end, and a knowledge distillation module. The edge end is mounted on the drone and performs cloud-edge collaboration with the cloud end. The cloud end is deployed with a YOLOv8s model, and the edge end is deployed with an improved YOLOv8s network. The drone collects a data set of transmission line video images, and pre-processes the data set of transmission line video images based on a fuzzy enhancement and segmentation algorithm to obtain a training set; The improved YOLOv8s network on the edge is used as the student model, and the YOLOv8s model on the cloud is used as the teacher model to build an insulator anomaly marking network including the student model and the teacher model. The knowledge distillation module inputs the training set into the student model and the teacher model at the same time, and trains the insulator abnormality marking network to obtain the insulator abnormality marking model; The image to be processed is input into the insulator abnormality marking model for marking, thereby completing the abnormality marking of the insulator.

2. The power transmission line insulator abnormality marking system based on edge intelligence according to claim 1 is characterized by: The improved YOLOv8s network improves the Bottleneck module of the YOLOv8s model. The C3 module in the Bottleneck module of the YOLOv8s model is replaced with a Ghost convolution module and a C2F module, and the output of the Ghost convolution module is used as the input of the C2F module.

3. The power transmission line insulator abnormality marking system based on edge intelligence according to claim 1 is characterized in that: During the training of the insulator abnormality marking network: the cross entropy of the teacher model and the student model are calculated respectively, and the insulator abnormality marking model is updated, the teacher model obtains the first label, and the student model obtains the second label; the KL-divergence between the first label and the second label is calculated, and the result of proportional summation of the cross entropy and the KL-divergence is used as the loss function of the training.

4. The power transmission line insulator abnormality marking system based on edge intelligence according to claim 1 is characterized in that: The fuzzy enhancement and segmentation algorithm is composed of ant colony intelligence algorithm and weighted variable FCM algorithm.

5. The power transmission line insulator abnormality marking system based on edge intelligence according to claim 1 is characterized in that: The preprocessing of the transmission line video image data set includes: performing fuzzy image enhancement and image segmentation on each transmission line video image to obtain insulator images and background images.

6. A method for marking abnormality of insulators of a power transmission line based on edge intelligence, implemented by using the abnormality marking system for insulators of a power transmission line based on edge intelligence according to claims 1 to 5, characterized in that: The specific steps include: S1: The UAV collects transmission line video images, and pre-processes the transmission line video image dataset based on fuzzy enhancement and segmentation algorithms to obtain a training set; S2: The improved YOLOv8s network on the edge is used as the student model, and the YOLOv8s model on the cloud is used as the teacher model to build an insulator anomaly marking network including the student model and the teacher model; S3: The knowledge distillation module inputs the training set into the student model and the teacher model at the same time, and trains the insulator abnormality marking network to obtain an insulator abnormality marking model; S4: Input the image to be processed into the insulator abnormality marking model for marking, thereby completing the abnormality marking of the insulator.

7. The method for marking abnormality of insulators of power transmission lines based on edge intelligence according to claim 6 is characterized in that: Step S1 specifically includes: S11: construct a sliding window on the transmission line video image, and calculate the mean and standard deviation of the sliding window; S12: Calculate the adaptive threshold T' based on the mean and standard deviation of the current sliding window: T'=m+k×σ; Among them, m is the mean, σ is the standard deviation, and k is a constant; S13: using the adaptive threshold T' to perform binarization processing on the current sliding window to obtain a fuzzy set membership diagram of the current sliding window; S14: repeating steps S12-S13 until the entire transmission line video image is traversed, and the fuzzy set membership graphs obtained by each sliding window are spliced ​​to obtain a complete fuzzy set membership graph; S15: Fuse the original transmission line video image with the corresponding complete fuzzy set membership graph to obtain a fuzzy enhanced image: ； ； Among them, I(x,y) is the original transmission line video image, The pixel (x, y) of the original image to the blurred center A The distance is the calibration parameter, which is 2 here, and m is the fuzzy factor, which is 1.5 here. is the complete fuzzy set membership graph, Enhance images for blur; S16: Calculate local feature information of the fuzzy enhanced image, where the local feature information includes grayscale value, gradient amplitude and texture feature; S17: Initializing parameters of the ant colony intelligence algorithm, the parameters including the number of ants and the concentration of pheromones; S18: using an ant colony intelligent algorithm to classify the fuzzy enhanced image, dividing the fuzzy enhanced image into a plurality of preliminary segmentation regions; and determining an initial clustering center along the distribution of local feature information of the fuzzy enhanced image; S19: Based on the initial cluster center, each preliminary segmentation area is optimized and refined using a fuzzy enhancement and segmentation algorithm to obtain an insulator image and a background image.

8. The method for marking abnormality of insulators of power transmission lines based on edge intelligence according to claim 7 is characterized in that: Step S19 specifically includes the following steps: S191: Initializing initial cluster centers and a membership matrix, wherein the membership matrix is ​​obtained based on a complete fuzzy set membership graph; S192: Iteratively update the initial cluster center and membership matrix using the following formula: ； ； in, is the i-th sample point The degree of membership to the kth category, ∈[0, 1], is the i-th sample point With cluster center The distance is the i-th sample point With cluster center distance, c is the expected number of categories, m is the fuzzy factor, is the center of the kth cluster, is the center of the jth cluster, N is the number of samples involved in the cluster analysis, is the i-th sample point The weight of is the i-th sample point For the mth power of the membership degree of the kth category; S193: Obtain a final segmentation result according to the updated membership matrix, and post-process the final segmentation result to obtain an insulator image and a background image.

9. The method for marking abnormality of insulators of power transmission lines based on edge intelligence according to claim 8, characterized in that: In step S193, post-processing includes small area merging and boundary smoothing.

10. The method for marking abnormality of insulators of power transmission lines based on edge intelligence according to claim 8, characterized in that: In step S3, the Adam optimization algorithm is used to update the weights of the student model to minimize the loss function; at the same time, model pruning technology or quantization compression technology is applied to reduce the student model so that the student model matches the computing power of the edge.

Images