A distribution network insulator fault identification and risk assessment method based on semi-supervised learning

Abstract

This invention discloses a semi-supervised learning method for distribution network insulator fault identification and risk assessment, relating to the field of distribution network insulator identification technology. The method includes: a cloud server acquiring images of distribution network insulators and training a fault identification model; the cloud server performing grayscale processing on the real-time acquired images; constructing a grayscale value mapping function and a fusion evaluation function; optimizing the grayscale value mapping function parameters using an adversarial learning artificial lemming algorithm; obtaining the optimal parameters of the grayscale value mapping function to enhance the grayscale image; identifying the image through the fault identification model and determining whether the enhanced image contains an insulator fault; including images with a confidence level greater than a confidence threshold in the training set; and constructing a multi-dimensional risk assessment index system to quantify the distribution network risk of distribution network insulators in multiple dimensions and map the risk level. This invention effectively improves the accuracy and efficiency of insulator fault identification in complex environments, reduces manual annotation costs, and has good application prospects.

Description

Technical Field

[0001] This invention relates to the field of distribution network insulator identification technology, specifically a semi-supervised learning method for distribution network insulator fault identification and risk assessment. Background Technology

[0002] As a critical component of the power grid, distribution network insulators are exposed to complex natural environments for extended periods, making them susceptible to faults such as cracks, contamination, or spontaneous explosions. High-resolution aerial images of insulators obtained through drone inspections have become an effective means of ensuring the safe and stable operation of the power grid. However, the complexity of the real-world environment during inspections can significantly impact image quality: inclement weather such as rain, snow, and fog can lead to blurred images and reduced contrast; insufficient lighting can result in insufficient brightness, loss of texture, or color distortion; and strong sunlight can cause overexposure, masking surface defects. Furthermore, the construction of training image samples is highly dependent on manual labor, and the significant differences in insulator morphology in real-world scenarios lead to low annotation efficiency and a high risk of missed or incorrect annotations. Therefore, image enhancement preprocessing is necessary when identifying drone-captured images of distribution network insulators in complex real-world environments. This operation is of great significance for the maintenance of distribution network insulators.

[0003] Existing fault identification methods for distribution network insulators are mainly divided into traditional feature-based methods and deep learning-based methods. Traditional feature-based methods manually extract features such as color, texture, and shape, combined with threshold segmentation, contour analysis, and geometric operations. However, they are greatly affected by complex environments and have limited generalization ability. Deep learning-based methods, leveraging models such as the YOLO series, are optimized through single-stage detection, cascaded detection, and the integration of attention mechanisms and feature fusion. They better handle complex backgrounds and occlusion issues. However, they suffer from low fault detection accuracy due to poor image quality and high annotation costs due to scarce defect samples.

[0004] In addition, the existing risk assessment methods for distribution network insulators mainly rely on defect detection results combined with human experience, lacking a systematic risk assessment system and not yet forming a mature automated risk assessment method, making it difficult to achieve accurate quantification and dynamic early warning of insulator failure risks.

[0005] Based on this, a semi-supervised learning method for distribution network insulator fault identification and risk assessment is provided, which can eliminate the drawbacks of existing technical solutions. Summary of the Invention

[0006] The purpose of this invention is to provide a semi-supervised learning method for fault identification and risk assessment of distribution network insulators, in order to solve the problems in the background art, such as low fault detection accuracy of distribution network insulators due to poor image quality, strong reliance on manual data annotation, and lack of distribution network risk assessment methods for distribution network insulators.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A semi-supervised learning method for fault identification and risk assessment of distribution network insulators, specifically including the following steps:

[0009] Step S1: The cloud server collects images of distribution network insulators under different environments and weather conditions from the distribution network drone inspection platform, performs preliminary manual annotation to construct a training set of distribution network insulator images, trains the YOLOv12 model to obtain a fault identification model, and sets a confidence threshold and an effective threshold.

[0010] Step S2: The drone inspection equipment collects images of distribution network insulators in real time and uploads them to the cloud server. The cloud server obtains the real-time images of distribution network insulators and performs grayscale processing on the real-time images of distribution network insulators.

[0011] Step S3: Construct a grayscale value mapping function and a fusion evaluation function that integrates four indicators, namely, general quality index, edge pixel count, image entropy, and grayscale co-occurrence matrix contrast.

[0012] Step S4: Optimize the grayscale value mapping function parameters using the opposition learning artificial lemming algorithm. The parameters include nonlinear adjustment parameters, noise suppression parameters, contrast balance parameters, and gain adjustment parameters.

[0013] Step S5: Obtain the optimal parameters of the grayscale value mapping function and use the optimal parameters to enhance the grayscale image;

[0014] Step S6: Identify the enhanced image through the fault identification model, determine whether there is an insulator fault in the enhanced image, and process it according to the identification results. If the confidence level of the model output prediction is greater than the confidence threshold, then proceed to step S7; otherwise, identify whether the confidence level is less than the effective threshold.

[0015] Step S7: Include images with a confidence level greater than the confidence threshold into the training set, and obtain the insulator fault type, confidence level, and ambient humidity and temperature of the drone shooting environment;

[0016] Step S8: Construct a multi-dimensional risk assessment indicator system to quantify the distribution network risk of distribution network insulators in multiple dimensions, map the risk level, and feed back the identification results, confidence level, fault type, and risk level to the UAV inspection equipment.

[0017] Preferably, step S6 further includes: if the confidence level is less than the effective threshold, the corresponding recognition image is directly removed, a re-acquisition image command is sent to the UAV inspection device, and step S2 is executed again; otherwise, the recognition result and the corresponding recognition image are stored in the cache area. When the stored data in the cache area reaches the maximum time interval, the model retraining mechanism is triggered, and the process jumps to step S1 to retrain the model and uses the retrained model for secondary fault detection.

[0018] Preferably, the calculation formula for the grayscale value mapping function in step S3 is: ,in, For the enhanced grayscale value, For pixels in the image of distribution network insulators The original grayscale value at that location, This represents the global grayscale mean of the insulator's grayscale image. and They are in pixels The mean and standard deviation of gray levels within a local window centered on the insulator are used to capture local texture features. It is a non-linear adjustment parameter. These are noise suppression parameters. For contrast balance parameters, This is the gain adjustment parameter.

[0019] Preferably, in step S4, the specific process of optimizing the grayscale mapping function parameters using the oppositional learning artificial lemming algorithm includes:

[0020] Input the grayscale image of the distribution network insulator, the size of the artificial lemming population, and the maximum number of algorithm iterations. Use the four parameter values ​​of the grayscale mapping function as decision variables. Initialize the artificial lemming population using cubic chaotic mapping. Generate an opposing artificial lemming population based on the centroid of the artificial lemming population location. Then merge the artificial lemming population and the opposing artificial lemming population. Calculate the fitness function of the artificial lemming individuals. Select the artificial lemming individuals with the top 10% fitness values ​​to form the initial artificial lemming population.

[0021] Based on the energy factor of artificial lemmings, the exploration and development stages of artificial lemmings are determined, and the behavior selection and group location update operations of artificial lemmings are realized. In the exploration stage, artificial lemmings switch between exploration behavior and burrowing behavior according to the switching probability. In the development stage, artificial lemmings switch between foraging behavior and predator evasion behavior according to the switching probability.

[0022] Iterate through the artificial lemming population after the position update, calculate the mutation rate of each dimension of the artificial lemming, and perform mutation operation if the mutation rate is greater than the mutation threshold. Calculate the fitness function value of the artificial lemming population, select the artificial lemming individual with the smallest fitness function value as the optimal individual. If the maximum number of iterations is reached at this time, output the optimal individual and end the algorithm; otherwise, continue iterating.

[0023] Preferably, the fitness function value is calculated using a fusion evaluation function constructed based on a general quality index, the number of edge pixels, image entropy, and gray-level co-occurrence matrix contrast. The general quality index is calculated by comparing the local gray-level mean, variance, and covariance of the enhanced image and the original image. The number of edge pixels is obtained by extracting edge information using the Sobel operator and counting the total number of pixels with gradient magnitudes greater than a set edge pixel threshold. The image entropy is calculated based on the probability distribution of the gray-level histogram. The gray-level co-occurrence matrix contrast is calculated by the average contrast of the four directions of the gray-level co-occurrence matrix.

[0024] The fitness function is calculated using the following formula: ,in, To integrate the evaluation function, This is a general quality index. The number of pixels at the edge. Image entropy, The contrast of the gray-level co-occurrence matrix.

[0025] Preferably, the identification results in step S6 include reliable results, medium reliability results, and unreliable results. The reliable result is determined based on the confidence level of the model output prediction being greater than the confidence threshold. The medium reliability result is determined based on the confidence level of the model output prediction being between the confidence threshold and the effective threshold. The unreliable result is determined based on the confidence level of the model output prediction being less than the effective threshold.

[0026] Preferably, the risk assessment indicators in the multi-dimensional risk assessment indicator system in step S8 include fault type severity, fault identification confidence, and environmental impact coefficient.

[0027] Preferably, in step S8, the distribution network risk of distribution network insulators is quantified in multiple dimensions, and the risk level is mapped specifically including:

[0028] A distribution network risk quantification model is constructed based on risk assessment indicators. A comprehensive risk value is calculated through weighted fusion. The formula for calculating the comprehensive risk value is as follows: ,in, For the comprehensive risk value, For fault identification confidence, For the severity of the fault type, This is the environmental weighting coefficient, with a value range of [value range missing]. , This is the environmental impact coefficient;

[0029] Based on the comprehensive risk value and the impact of insulator faults on the distribution network, a risk level mapping mechanism is established, classifying the comprehensive risk value into low risk, medium risk, high risk, and extremely high risk. The specific mapping rules are as follows:

[0030] .

[0031] Preferably, the YOLOv12 model training process in step S1 specifically includes:

[0032] Manually labeled data is input into the YOLOv12 model for training to obtain the initial fault identification model;

[0033] Input the image data of the distribution network insulator to be detected, use the initial fault identification model for prediction, and perform hierarchical processing on the image data according to the identification results of the initial fault identification model. Image data identified as reliable results are stored in the training set, and the size of the training set is increased. Image data identified as medium reliability results are stored in the buffer, and image data identified as unreliable results are directly deleted. When the stored data in the buffer reaches the maximum time interval, the initial fault identification model is retrained using the enhanced training set to obtain an updated fault identification model. The updated fault identification model is then used to perform secondary detection on the data stored in the buffer.

[0034] Repeated data prediction and retraining operations are performed, with the model version iterated after each round of retraining, until the detection accuracy of the final fault identification model reaches the preset target.

[0035] Preferably, an edge-cloud collaborative system is applied, comprising a drone inspection device and a cloud server. The drone inspection device is equipped with a high-definition industrial camera and several sensors to capture real-time images of distribution network insulators and environmental data during the inspection process, and simultaneously uploads the collected images, humidity, and temperature data to the cloud server, awaiting instructions and recognition results from the cloud server. The cloud server is deployed with a YOLOv12 model and a distribution network risk quantification model. After receiving the images uploaded by the drone inspection device, it enhances the distribution network insulator images through an adversarial learning artificial lemming algorithm. The trained YOLOv12 model outputs the fault type and confidence level, and the distribution network risk quantification model quantifies the risk of the distribution network insulators.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention acquires insulator images from real, complex environments using drones, and optimizes the grayscale mapping function parameters using an adversarial learning artificial lemming algorithm. This optimizes the overall quality index, edge pixels, image entropy, and grayscale co-occurrence matrix contrast, effectively improving image quality. A semi-supervised lightweight target detection model is included, and iterative optimization is achieved through a tiered recognition result and retraining trigger mechanism, reducing reliance on manual annotation. Furthermore, this invention constructs a multi-dimensional risk assessment index system, calculating a weighted comprehensive risk value and mapping it to a risk level, enabling precise quantification and dynamic early warning of fault risks. This addresses the problems of low fault detection accuracy, strong reliance on manual data annotation, and lack of risk assessment methods for distribution network insulators caused by poor image quality acquired by drones in real, complex environments. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the steps of the distribution network insulator fault identification and risk assessment method of the present invention.

[0039] Figure 2 This is a flowchart illustrating step S4 of the present invention.

[0040] Figure 3 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0042] Example 1

[0043] In this embodiment, as Figures 1-3 As shown, a semi-supervised learning method for distribution network insulator fault identification and risk assessment includes the following steps:

[0044] Step S1: The cloud server collects images of distribution network insulators under different environments and weather conditions from the distribution network drone inspection platform. The distribution network drone inspection platform is a comprehensive system integrating drone control, mission planning, data communication, and image uploading functions. It can be used to schedule drone inspection equipment to autonomously inspect distribution network lines according to a predetermined route, and transmit insulator images collected by high-definition industrial cameras and environmental data collected by sensors in real time to the cloud server. Images of complex environments such as rain and fog account for 40%. A training set of distribution network insulator images is initially manually labeled to build a fault identification model. A YOLOv12 model is trained to obtain a fault identification model. A confidence threshold and an effective threshold are set. In this embodiment, the confidence threshold is set to... Set the effective threshold to The threshold can be set based on the accuracy of the validation set or historical data. The confidence threshold is a high confidence threshold. Recognition results above this value are considered highly reliable and can be directly used for output and training set expansion. The effective threshold is a low confidence threshold. Recognition results below this value are directly rejected due to insufficient reliability.

[0045] Step S2: The drone inspection equipment collects images of distribution network insulators in real time and uploads them to the cloud server. The cloud server obtains the real-time images of distribution network insulators and performs grayscale processing on the real-time images of distribution network insulators.

[0046] Step S3: Construct a grayscale value mapping function and a fusion evaluation function that integrates four indicators, including general quality index, edge pixel count, image entropy, and grayscale co-occurrence matrix contrast.

[0047] Step S4: Optimize the grayscale value mapping function parameters using the opposition learning artificial lemming algorithm. The parameters include nonlinear adjustment parameters, noise suppression parameters, contrast balance parameters, and gain adjustment parameters.

[0048] Step S5: Obtain the optimal parameters of the grayscale value mapping function and use the optimal parameters to enhance the grayscale image;

[0049] Step S6: Identify the enhanced image through the fault identification model, determine whether there is an insulator fault in the enhanced image, and process it according to the identification results. If the confidence level of the model output prediction is greater than the confidence threshold, then proceed to step S7; otherwise, identify whether the confidence level is less than the effective threshold.

[0050] Step S7: Include images with a confidence level greater than the confidence threshold into the training set, and obtain the insulator fault type, confidence level, and ambient humidity and temperature of the drone shooting environment;

[0051] Step S8: Construct a multi-dimensional risk assessment index system to quantify the distribution network risk of distribution network insulators in multiple dimensions, map the risk level, and feed back the identification results, confidence level, fault type, and risk level to the UAV inspection equipment.

[0052] In this embodiment, the opposition learning artificial lemming optimization algorithm is an intelligent optimization method that integrates opposition learning strategy and artificial lemming optimization algorithm. It aims to improve the search efficiency and convergence speed of traditional optimization algorithms. The algorithm initializes the population through chaotic elite opposition learning, generates initial solutions using cubic chaotic mapping, constructs opposition solutions based on the population centroid, and selects high-quality individuals to form the initial group, thereby enhancing population diversity. The semi-supervised lightweight object detection model is an object detection framework that integrates semi-supervised learning paradigm and lightweight design concept. It aims to solve the problems of strong dependence on large-scale labeled data and low deployment efficiency in complex scenarios of traditional supervised detection models. This method selects the YOLOv12 lightweight object detection model as the backbone model, sets up a buffer to store data, reduces the problem of missed detection, and effectively reduces the manual dependence on data labeling.

[0053] Among them, such as Figure 3 As shown, step S6 also includes: if the confidence level is less than the effective threshold, the corresponding recognition image is directly removed, a re-acquisition image command is sent to the UAV inspection device, and step S2 is executed again; otherwise, the recognition result and the corresponding recognition image are stored in the cache area. When the stored data in the cache area reaches the maximum time interval, the model retraining mechanism is triggered, and the process jumps to step S1 to retrain the model and uses the retrained model for secondary fault detection.

[0054] Among them, such as Figure 3 As shown, the identification results in step S6 include reliable results, medium reliability results, and unreliable results. The reliable result is judged based on the confidence level of the model output prediction being greater than the confidence threshold. The medium reliability result is judged based on the confidence level of the model output prediction being between the confidence threshold and the effective threshold. The unreliable result is judged based on the confidence level of the model output prediction being less than the effective threshold.

[0055] In this embodiment, since prediction results with different confidence levels have different values ​​for model training, this method classifies the recognition results according to their confidence levels. Prediction results with confidence levels higher than the confidence threshold have had their accuracy fully verified and can be directly output to provide positive feedback for model parameter updates. Prediction results with confidence levels between the confidence threshold and the effective threshold, although having some uncertainty, still contain some effective information. If the recognition results are directly discarded, it will cause the target to be missed. Therefore, the data is stored in the buffer. Prediction results with confidence levels lower than the effective threshold cannot be guaranteed to be accurate and need to be directly removed. The buffer is a temporary storage area for results with moderate reliability and has the functions of data temporary storage, data extraction, clearing and resetting. When temporary storage, the original image information, the initial prediction confidence level and the prediction time of the data need to be recorded. In order to ensure the effectiveness of the data in the buffer, a data storage validity period is set. The storage validity period can be configured according to the inspection frequency. In this embodiment, it is set to 3 hours. Data stored beyond the validity period will be automatically deleted.

[0056] Among them, such as Figures 1-3 As shown, the calculation formula for the grayscale value mapping function in step S3 is:

[0057] ,in, For the enhanced grayscale value, For pixels in the image of distribution network insulators The original grayscale value at that location, This represents the global grayscale mean of the insulator's grayscale image. and They are in pixels The mean and standard deviation of gray levels within a local window centered on the insulator are used to capture local texture features. It is a non-linear adjustment parameter. These are noise suppression parameters. For contrast balance parameters, This is the gain adjustment parameter;

[0058] In this embodiment, pixels are used. A 3x3 square field window centered on the center, containing pixels. and its 8 adjacent pixels, This represents the average grayscale value of all pixels within the window, calculated using the following formula: ,in, For pixels grayscale value, Reflects the overall brightness level of the window area. The degree of dispersion of grayscale values ​​within the quantization window is calculated using the following formula:

[0059] ;

[0060] In this embodiment, the nonlinear adjustment parameter Adjusting the local grayscale mean through exponential calculation This achieves differential enhancement of bright and dark areas in an image, with a value range of [value range missing]. Noise suppression parameters By adjusting the denominator Adjust the dynamic range of the local standard deviation to avoid excessive amplification in noise regions; the value range is [value range missing]. Contrast balance parameters pass The term adjusts the weighting of the difference between the local and global means to control the aggressiveness of contrast enhancement; its value range is [value range missing]. Gain adjustment parameters The gain coefficient for overall enhancement intensity is controlled by scaling the ratio of the global mean to the local standard deviation, with a value range of [value range missing]. The above four parameters together determine the image enhancement effect of distribution network insulators. Therefore, better image enhancement effect can be achieved by optimizing the parameters. The image enhancement effect can be quantitatively evaluated by four indicators: general quality index, edge pixel count, image entropy, and gray-level co-occurrence matrix contrast.

[0061] Among them, such as Figure 2 As shown, the fitness function value is calculated using a fusion evaluation function constructed based on the general quality index, the number of edge pixels, image entropy, and gray-level co-occurrence matrix contrast. The general quality index is calculated by comparing the local gray-level mean, variance, and covariance of the enhanced image and the original image. The number of edge pixels is obtained by extracting edge information using the Sobel operator and counting the total number of pixels with gradient magnitudes greater than a set edge pixel threshold. The image entropy is calculated based on the probability distribution of the gray-level histogram. The gray-level co-occurrence matrix contrast is calculated by the average contrast of the four directions of the gray-level co-occurrence matrix.

[0062] The formula for calculating the fitness function is: ,in, To integrate the evaluation function, This is a general quality index. The number of pixels at the edge. Image entropy, For gray-level co-occurrence matrix contrast, To enhance the image;

[0063] In this embodiment, the general quality index can measure the original image. With enhanced image The structural consistency is measured, with a value closer to 1 indicating better structural consistency. The calculation formula is as follows: ,in, Original image The average gray level within a local window To enhance the image The average gray level within a local window Original image The grayscale variance within a local window To enhance the image The grayscale variance within a local window Original image With enhanced image The covariance within a local window reflects the correlation between the pixel grayscale values ​​of the original image patch and the enhanced image patch. The calculation formula is as follows: , and These are the original images. and image enhancement The pixel grayscale value at the location;

[0064] In this embodiment, the number of edge pixels can be used to measure the enhanced image. The richness of the mid-edges indicates that the higher the value, the richer the edge details and the clearer the image visual effect. The calculation process includes: first calculating the enhanced image... medium pixel Gradients in the horizontal and vertical directions , The formula is Then, based on the gradients in the horizontal and vertical directions... , The gradient magnitude is calculated using the following formula: When the gradient magnitude Greater than the gradient threshold At that time, the pixel is considered an edge pixel, and the gradient threshold is applied. Set as Statistical gradient magnitude Greater than the gradient threshold The total number of pixels is the same as the number of edge pixels, calculated using the following formula: , where the matrix For the Sobel operator's convolution kernel in the horizontal direction, the matrix... The Sobel operator's convolution kernel in the vertical direction, For indicator functions, when hour, Otherwise, it is 0;

[0065] In this embodiment, a higher image entropy value indicates richer surface texture of the insulator. The calculation formula is as follows: , To enhance the grayscale values ​​in an image The probability of its occurrence;

[0066] In this embodiment, a higher gray-level co-occurrence matrix (GLCM) contrast value indicates a more significant gray-level difference between adjacent pixels in the image. The calculation process includes: constructing the GLCM matrix, then calculating the contrast, and enhancing the image. In the middle, select the distance pixels Furthermore, the neighborhood relationships in the directions of 0°, 45°, 90°, and 135° are used to count adjacent pixel pairs. probability of occurrence The calculation formula is: ,in, As an indicator function, when the image is enhanced median coordinate The pixel grayscale is And its neighboring pixels have a gray level of hour, Otherwise, it is 0. X and Y are the width and height of the image, respectively. Based on the GLCM matrix, the contrast calculation formula is: The numerator is the sum of the products of the squares of the gray level differences of all adjacent pixels and their corresponding probabilities, and the denominator is the normalization coefficient of the GLCM matrix.

[0067] Among them, such as Figure 1 and Figure 2 As shown, to achieve the optimal enhancement effect for distribution network insulator images, this method employs the opposition-learning artificial lemming algorithm to optimize the optimal parameters of the grayscale value mapping function. The specific process of optimizing the grayscale value mapping function parameters using the opposition-learning artificial lemming algorithm includes:

[0068] Input grayscale image of distribution network insulator, artificial lemming population size N, and maximum number of algorithm iterations. Map the grayscale values ​​to the four parameters of the function. As a decision variable, the artificial lemming population is initialized using a cubic chaotic mapping, and the formula for this operation is as follows: ,in, For the first The artificially brooded lemming in the first The initial position of the dimension. and For the first Upper and lower bounds of dimensional parameters It is the chaotic value of the cubic mapping function iteration. ,until Stop iterating, and generate an opposing artificial lemming population based on the centroid of the artificial lemming population location. The formula for this operation is: ,in, For the first A pair of opposing artificial lemmings in the first The position of the dimension The centroid of the artificially created lemming population is located in the 2nd dimension. Then, the artificially brooded lemming population and the opposing artificially brooded lemming population were merged, and the fitness function of each artificially brooded lemming individual was calculated. The artificially brooded lemming individuals with the highest fitness values ​​were selected to form the initial artificially brooded lemming population. It is an abbreviation for Maximum Iteration, meaning the maximum number of iterations.

[0069] Based on artificial lemming energy factors To determine the exploration and development stage of artificially broodstock, and to implement behavioral selection and population location updates, the energy factor calculation formula is as follows: ,when At this point, the artificially brooded lemming enters the exploration phase. During this phase, the artificially brooded lemming switches between exploration and digging behaviors based on the switching probability, as follows: ,in, It is the number of iterations of the algorithm. This is the current optimal position. For direction value, Set to ±1, For Brownian motion random vectors, It is a random vector, and its value range is... , and For random individual locations, The probability of behavior switching, with a value range of [value missing]. , To explore range parameters, ,when At this point, the artificially brooded lemmings enter the development phase. During this phase, the artificially brooded lemmings switch between foraging behavior and predator avoidance behavior based on the switching probability, as shown below: ,in, For the escape coefficient, , For the Lévy flight random number of the artificial lemming algorithm; This represents a mathematical model of Levy's flight in D-dimensional space. This represents a spiral function, used to simulate an individual moving towards the optimal solution along a spiral trajectory. The movement behavior reflects the spiral approach mechanism of local search in the algorithm, enhancing the fine exploration of the current optimal region. It is an abbreviation for random number, which usually refers to a random number that follows a uniform distribution within the interval [0,1]. It is used to introduce randomness, making the search process of the algorithm more diverse and avoiding getting trapped in local optima.

[0070] Iterate through the artificial lemming population after the position update and calculate the mutation rate of the artificial lemming in each dimension. The calculation formula is: , As a decision variable dimension, The number of dimensions to be mutated is randomly selected. If the artificial lemming's first dimension is... Variance of decision variables If the value is greater than the mutation threshold of 0.5, then mutation is performed. The mutation formula is: ,in, artificial lemmings The The value after mutation in each dimension This is a variable-asynchronous long factor, with a value of 0.2. Let be a real number that follows a normal distribution, and take values ​​in the range of . , Indicates the first The maximum value of a dimensional decision variable is the upper bound of the range of values ​​for that dimensional variable. Indicates the first The minimum value of a dimensional decision variable is the lower bound of the range of values ​​for that dimensional variable. and Used to define the range of values ​​for decision variables in the corresponding dimension, combined with variable-asynchronous long factor. and normal distribution real numbers The variation rate of artificial lemming individuals in this dimension is controlled to achieve effective exploration of the solution space. If the variation rate exceeds the variation threshold, a mutation operation is performed. The fitness function value of the artificial lemming population is calculated, and the artificial lemming individual with the smallest fitness function value is selected as the optimal individual. If the maximum number of iterations is reached at this point, that is... Then output the optimal individual. Output optimal parameters If the optimal parameter is obtained, the algorithm ends; otherwise, it continues to iterate until the optimal parameter is obtained, and then the image enhancement operation can continue.

[0071] Among them, such as Figure 1 As shown, this paper presents the risk differences between cracked, damaged, and dirty states and normal states. By adjusting the weights of environmental factors on different failure types, this method proposes a multi-dimensional risk assessment index system. The risk assessment indexes of the multi-dimensional risk assessment index system include failure type severity, failure identification confidence, and environmental impact coefficient.

[0072] Among them, such as Figure 1 As shown, in step S8, the distribution network risk of distribution network insulators is quantified in multiple dimensions, and the risk level is mapped specifically including:

[0073] A distribution network risk quantification model is constructed based on risk assessment indicators. The comprehensive risk value is calculated through weighted fusion. The formula for calculating the comprehensive risk value is as follows: ,in, For the comprehensive risk value, The fault identification confidence score is the model prediction output value, and its value range is [value range missing]. , For the severity of the fault type, This refers to the environmental weighting coefficient, which can be derived from historical fault data statistics, and its value range is [range missing]. , This is the environmental impact coefficient;

[0074] Based on the comprehensive risk value and the impact of insulator faults on the distribution network, a risk level mapping mechanism is established, classifying the comprehensive risk value into low risk, medium risk, high risk, and extremely high risk. The specific mapping rules are as follows:

[0075] The mapping rule values ​​are set based on historical data;

[0076] In this embodiment, since the direct harm caused by different types of insulator faults is significant, and the risks posed by different fault types to the distribution network also vary, this method directly quantifies the inherent risks of different insulator states. For example, an insulator in a damaged state no longer has the function of insulation and isolation, which may lead to a short circuit between the conductor and the tower, causing a tripping accident. This poses the most serious risk and is assigned a severity rating. Even if an insulator is cracked, it still retains its insulating function, but it poses a risk of leakage and may eventually break down. The potential for harm is thus assessed with a severity rating of [not specified]. If an insulator is in a polluted state, the surface resistance of the insulator will decrease due to the coverage of the pollution layer, affecting its insulation performance. The risk of damage is progressive, and a severity level of danger will be assigned. If the insulation performance and physical strength of the insulators under normal conditions meet the requirements of the distribution network and there is no hazard risk, then a hazard severity value is assigned. ;

[0077] In this embodiment, environmental factors have a significant impact on the condition of insulators. High humidity can easily cause a water film to form on the insulator surface, reducing insulation performance. Dust and salt adhering to the insulator surface can form conductive channels, triggering flashover. High temperatures accelerate material aging, while low temperatures may cause freezing and cracking. Strong winds can exacerbate pollutant accumulation and mechanical stress, shortening insulator life and threatening distribution network safety. Therefore, this method considers the impact of environmental factors on the condition of insulators. Based on the correlation characteristics between distribution network operating environment parameters and fault types, an environmental impact coefficient is calculated. The method is as follows: ,in, For real-time precipitation, For air humidity, environmental data can be obtained in real time through meteorological monitoring stations along the power distribution network;

[0078] Among them, such as Figure 3 As shown, to reduce the reliance on manual data annotation and improve model recognition accuracy, this method uses a semi-supervised lightweight model for distribution network insulator defect detection. The model needs to be trained and updated in a timely manner. The model training process specifically includes:

[0079] Manually labeled data is input into the YOLOv12 model for training to obtain the initial fault identification model;

[0080] Input the image data of the distribution network insulator to be detected, use the initial fault identification model for prediction, and perform hierarchical processing on the image data according to the identification results of the initial fault identification model. Image data identified as reliable results are stored in the training set, and the size of the training set is increased. Image data identified as medium reliability results are stored in the buffer, and image data identified as unreliable results are directly deleted. When the stored data in the buffer reaches the maximum time interval, the initial fault identification model is retrained using the enhanced training set to obtain an updated fault identification model. The updated fault identification model is then used to perform secondary detection on the data stored in the buffer.

[0081] Repeated data prediction and retraining operations are performed, with each retraining iteration updating the model version until the detection accuracy of the final fault identification model reaches a preset target. In this embodiment, the preset target is... ;

[0082] In this embodiment, the retraining mechanism is triggered under two conditions: first, when the amount of data in the buffer reaches a preset window threshold, the operation indicates that there is enough moderately reliable data available for model optimization; second, when the amount of data stored in the buffer reaches the maximum time interval, both the preset window threshold and the maximum time interval can be adjusted according to actual needs.

[0083] Example 2

[0084] Unlike Example 1, this method employs an edge-cloud collaborative system, which includes drone inspection equipment and a cloud server. The edge-cloud collaborative system refers to a distributed processing architecture composed of drone inspection equipment deployed at the inspection site and a cloud server. The two work collaboratively through a communication network. The drone inspection equipment is equipped with a high-definition industrial camera and several sensors to capture real-time images of distribution network insulators and environmental data during the inspection process. Simultaneously, it uploads the collected images, humidity, and temperature data to the cloud server, awaiting instructions and recognition results. The cloud server deploys a YOLOv12 model and a distribution network risk quantification model, as well as an image enhancement module and a risk assessment module. This allows the processed data to be fed back to the drone inspection equipment. After receiving the images uploaded by the drone inspection equipment, the distribution network insulator images are enhanced using an adversarial learning artificial lemming algorithm. The trained YOLOv12 model outputs the fault type and confidence level, while the distribution network risk quantification model quantifies the risk of the distribution network insulators.

[0085] In summary, this invention establishes an image enhancement method based on the adversarial learning artificial lemming algorithm, optimizes key parameters of the grayscale mapping function, and evaluates image quality based on a fusion evaluation function of general quality index, edge pixel count, image entropy, and grayscale co-occurrence matrix contrast, thereby achieving adaptive enhancement of insulator images in complex environments. This invention also uses recognition results based on the YOLOv12 model, designs a buffer and retraining trigger mechanism to achieve iterative optimization of the model, reduce reliance on manual annotation, and thus improve the efficiency of insulator fault detection.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A semi-supervised learning method for fault identification and risk assessment of distribution network insulators, characterized in that, Specifically, the following steps are included: Step S1: The cloud server collects images of distribution network insulators under different environments and weather conditions from the distribution network drone inspection platform, performs preliminary manual annotation to construct a training set of distribution network insulator images, trains the YOLOv12 model to obtain a fault identification model, and sets a confidence threshold and an effective threshold. Step S2: The drone inspection equipment collects images of distribution network insulators in real time and uploads them to the cloud server. The cloud server obtains the real-time images of distribution network insulators and performs grayscale processing on the real-time images of distribution network insulators. Step S3: Construct a grayscale value mapping function and a fusion evaluation function that integrates four indicators, namely, general quality index, edge pixel count, image entropy, and grayscale co-occurrence matrix contrast. Step S4: Optimize the grayscale value mapping function parameters using the opposition learning artificial lemming algorithm. The parameters include nonlinear adjustment parameters, noise suppression parameters, contrast balance parameters, and gain adjustment parameters. Step S5: Obtain the optimal parameters of the grayscale value mapping function and use the optimal parameters to enhance the grayscale image; Step S6: Identify the enhanced image through the fault identification model, determine whether there is an insulator fault in the enhanced image, and process it according to the identification results. If the confidence level of the model output prediction is greater than the confidence threshold, then proceed to step S7; otherwise, identify whether the confidence level is less than the effective threshold. Step S7: Include images with a confidence level greater than the confidence threshold into the training set, and obtain the insulator fault type, confidence level, and ambient humidity and temperature of the drone shooting environment; Step S8: Construct a multi-dimensional risk assessment index system to quantify the distribution network risk of distribution network insulators in multiple dimensions, map the risk level, and feed back the identification results, confidence level, fault type, and risk level to the UAV inspection equipment. In step S4, the specific process of optimizing the grayscale mapping function parameters using the oppositional learning artificial lemming algorithm includes: Input the grayscale image of the distribution network insulator, the size of the artificial lemming population, and the maximum number of algorithm iterations. Use the four parameter values ​​of the grayscale mapping function as decision variables. Initialize the artificial lemming population using cubic chaotic mapping. Generate an opposing artificial lemming population based on the centroid of the artificial lemming population location. Then merge the artificial lemming population and the opposing artificial lemming population. Calculate the fitness function of the artificial lemming individuals. Select the artificial lemming individuals with the top 10% fitness values ​​to form the initial artificial lemming population. Based on the energy factor of artificial lemmings, the exploration and development stages of artificial lemmings are determined, and the behavior selection and group location update operations of artificial lemmings are realized. In the exploration stage, artificial lemmings switch between exploration behavior and burrowing behavior according to the switching probability. In the development stage, artificial lemmings switch between foraging behavior and predator evasion behavior according to the switching probability. Iterate through the artificial lemming population after the position update, calculate the mutation rate of artificial lemmings in each dimension. If the mutation rate is greater than the mutation threshold, perform a mutation operation. Calculate the fitness function value of the artificial lemming population, select the artificial lemming individual with the smallest fitness function value as the optimal individual. If the maximum number of iterations is reached at this time, output the optimal individual and end the algorithm; otherwise, continue iterating. The fitness function value is calculated using a fusion evaluation function constructed based on a general quality index, edge pixel count, image entropy, and gray-level co-occurrence matrix contrast. The general quality index is calculated by comparing the local gray-level mean, variance, and covariance of the enhanced image and the original image. The edge pixel count is obtained by extracting edge information using the Sobel operator and counting the total number of pixels with gradient magnitudes greater than a set edge pixel threshold. The image entropy is calculated based on the probability distribution of the gray-level histogram. The gray-level co-occurrence matrix contrast is calculated by averaging the contrast in the four directions of the gray-level co-occurrence matrix. The fitness function is calculated using the following formula: ,in, To integrate the evaluation function, This is a general quality index. The number of pixels at the edge. Image entropy, The contrast of the gray-level co-occurrence matrix.

2. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 1, characterized in that, Step S6 further includes: if the confidence level is less than the effective threshold, the corresponding recognition image is directly removed, a re-acquisition image command is sent to the UAV inspection device, and step S2 is executed again; otherwise, the recognition result and the corresponding recognition image are stored in the cache area. When the stored data in the cache area reaches the maximum time interval, the model retraining mechanism is triggered, and the process jumps to step S1 to retrain the model and uses the retrained model for secondary fault detection.

3. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 1, characterized in that, The formula for calculating the grayscale value mapping function in step S3 is as follows: ,in, For the enhanced grayscale value, For pixels in the image of distribution network insulators The original grayscale value at that location, This represents the global grayscale mean of the insulator's grayscale image. and They are in pixels The mean and standard deviation of gray levels within a local window centered on the insulator are used to capture local texture features. It is a non-linear adjustment parameter. These are noise suppression parameters. For contrast balance parameters, This is the gain adjustment parameter.

4. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 1, characterized in that, The identification results in step S6 include reliable results, medium reliability results, and unreliable results. The reliable result is determined by the confidence level of the model output prediction being greater than the confidence threshold. The medium reliability result is determined by the confidence level of the model output prediction being between the confidence threshold and the effective threshold. The unreliable result is determined by the confidence level of the model output prediction being less than the effective threshold.

5. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 1, characterized in that, The risk assessment indicators in the multi-dimensional risk assessment indicator system in step S8 include fault type severity, fault identification confidence, and environmental impact coefficient.

6. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 5, characterized in that, In step S8, the distribution network risk of distribution network insulators is quantified from multiple dimensions, and the risk level is mapped specifically including: A distribution network risk quantification model is constructed based on risk assessment indicators. A comprehensive risk value is calculated through weighted fusion. The formula for calculating the comprehensive risk value is as follows: ,in, For the comprehensive risk value, For fault identification confidence, For the severity of the fault type, This is the environmental weighting coefficient, with a value range of [value range missing]. , This is the environmental impact coefficient; Based on the comprehensive risk value and the impact of insulator faults on the distribution network, a risk level mapping mechanism is established, classifying the comprehensive risk value into low risk, medium risk, high risk, and extremely high risk. The specific mapping rules are as follows: 。 7. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 4, characterized in that, The YOLOv12 model training process in step S1 specifically includes: Manually labeled data is input into the YOLOv12 model for training to obtain the initial fault identification model; Input the image data of the distribution network insulator to be detected, use the initial fault identification model for prediction, and perform hierarchical processing on the image data according to the identification results of the initial fault identification model. Image data identified as reliable results are stored in the training set, and the size of the training set is increased. Image data identified as medium reliability results are stored in the buffer, and image data identified as unreliable results are directly deleted. When the stored data in the buffer reaches the maximum time interval, the initial fault identification model is retrained using the enhanced training set to obtain an updated fault identification model. The updated fault identification model is then used to perform secondary detection on the data stored in the buffer. Repeated data prediction and retraining operations are performed, with the model version iterated after each round of retraining, until the detection accuracy of the final fault identification model reaches the preset target.

8. The method for semi-supervised learning-based fault identification and risk assessment of distribution network insulators according to claim 1, characterized in that, An edge-cloud collaborative system is applied, comprising a drone inspection device and a cloud server. The drone inspection device is equipped with a high-definition industrial camera and several sensors to capture real-time images of distribution network insulators and environmental data during the inspection process, and simultaneously uploads the collected images, humidity, and temperature data to the cloud server, awaiting instructions and recognition results from the cloud server. The cloud server is deployed with a YOLOv12 model and a distribution network risk quantification model. After receiving the images uploaded by the drone inspection device, it enhances the distribution network insulator images through an adversarial learning artificial lemming algorithm. The trained YOLOv12 model outputs the fault type and confidence level, and the distribution network risk quantification model quantifies the risk of the distribution network insulators.

Images