A method and device for defect detection in power distribution network equipment based on multi-model collaboration

Abstract

This invention discloses a method and apparatus for defect detection in power distribution network equipment based on multi-model collaboration, aiming to solve the problem of low defect detection accuracy caused by complex weather interference and noise labeling. The method includes: constructing a power distribution network defect detection dataset containing different weather interferences and noise labeling; building and training a multi-model collaborative detection model containing five modules: a weather interference quantification module calculates the interference coefficient ω; a multi-expert integration module constructs sub-models based on YOLOv8 and configures independent training conditions; a labeling consistency voting module marks suspicious samples; a suspicious sample adaptive relabeling module updates the labels; and a noise-aware loss function module dynamically adjusts the loss weights. A new image is input, inferred by the model, and the defect category, bounding box, and confidence score are output. This invention improves the robustness of detection under complex weather conditions, reduces the impact of noise labeling, and effectively improves the accuracy and efficiency of defect detection in power distribution network equipment. It is suitable for power distribution network inspection scenarios and has significant engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of power equipment testing technology, specifically relating to a method and device for detecting defects in power distribution network equipment based on multi-model collaboration. Background Technology

[0002] With the continuous advancement of smart grid construction, defect detection in distribution network equipment is rapidly transforming from manual inspection to automation and intelligence. In this process, computer vision-based target detection technology, due to its high efficiency, has become a core method for power grid defect detection. However, the performance of distribution network equipment defect detection is highly dependent on the quality of labeled data, and many key issues exist in practical applications.

[0003] When manually labeling defects in power distribution networks, noise can easily arise due to various factors. Blurred defect features lead to incorrect labeling categories; unclear defect boundaries cause bounding box offsets; and small defects are missed in massive datasets, resulting in unlabeled defects. This noise distorts the supervisory signals in the training data. Traditional detection models, such as YOLO and Faster R-CNN, passively fit incorrect labels, leading to decreased detection accuracy and instances where normal components are misclassified as defects or real defects are missed. Furthermore, outdoor inspections often face adverse weather conditions such as heavy rain, fog, sandstorms, and strong sunlight. These conditions reduce image contrast, blur details, increase noise, and obscure targets, further complicating model detection: heavy rain obscures images, masking defect features like insulator bursts and conductor corrosion; fog blurs the entire image, significantly reducing bounding box positioning accuracy; sandstorms cause abnormal grayscale values, easily misclassifying dusty areas as defects or missing real defects; and strong sunlight creates light spots, disrupting the grayscale difference between defects and the background.

[0004] Meanwhile, existing noise-resistant target detection solutions have significant limitations. Data cleaning methods, such as manual review, are labor-intensive and struggle to adapt to the massive growth of distribution network data. With the widespread use of drones and camera inspection equipment, the amount of data generated daily is increasing exponentially, making manual review unsustainable. Single-model robustness optimization, such as adding regularization, has poor adaptability to complex noise and cannot effectively distinguish between trustworthy and noisy labels. These methods often assume that noise is uniformly distributed, but distribution network data contains diverse noise types, including class misclassification, boundary bias, and missing labels, which simple regularization cannot handle. General noise-resistant algorithms are not designed for the defect characteristics of distribution network equipment, such as class imbalance and a high proportion of small samples. In distribution network scenarios, common defects such as insulator bursts have a high proportion, while rare defects such as high-voltage line corrosion have few samples, making the model more prone to overfitting the majority class and ignoring the minority class under noise interference, resulting in limited practical effectiveness.

[0005] Furthermore, distribution network scenarios have unique requirements; defect detection in distribution network equipment must meet both high reliability and real-time requirements. Labelable noise can cause fluctuations in model output, making it difficult to meet industrial-grade requirements. Therefore, there is an urgent need for a defect detection method for distribution network equipment that can adaptively identify labeling noise, dynamically adjust the learning strategy, and specifically overcome weather interference. This would improve the model's robustness under the dual influence of complex weather and noisy labeling, reduce manual labeling costs, and meet the practical needs of intelligent distribution network inspection. Summary of the Invention

[0006] This invention proposes a method and device for detecting defects in power distribution network equipment based on multi-model collaboration. By integrating multiple expert models, using a consistent voting module for annotation, an adaptive re-annotation mechanism for suspicious samples, and a noise perception loss function, the method achieves dynamic identification and robust learning of annotation noise.

[0007] A method for detecting defects in power distribution network equipment based on multi-model collaboration, characterized by comprising the following:

[0008] Prepare for the construction of the power distribution network defect detection dataset: acquire a large number of photos of power distribution network equipment defects under different weather interference conditions, which may include noise annotations;

[0009] Construction and training of a multi-model collaborative defect detection model for power distribution networks: The model includes a weather interference quantification module, a multi-expert integration module, a labeling consistency voting module, a suspicious sample adaptive relabeling module, and a noise perception loss function module. The specific process is as follows:

[0010] The weather interference quantification module operates as follows: Based on collected normal weather power distribution network photos, it calculates and stores the standard grayscale standard deviation σ0 and standard contrast C0 as benchmark values ​​for weather interference quantification, ensuring that the interference coefficients of different images can be compared horizontally. For each power distribution network image, it calculates its grayscale standard deviation σ and contrast C, and then substitutes them into the formula ω=1-σ·C / (σ0·C0) to calculate the weather interference intensity coefficient ω. The closer the ω value is to 0, the stronger the weather interference, and vice versa. The calculated ω value is then passed to the multi-expert integration module.

[0011] Multi-expert ensemble module operation: Based on YOLOv8, multiple sub-model frameworks are built. A Coordinate Attention layer is added to the feature extraction network to enhance the sensitivity to defect location information. Cross-scale skip connections are added to the PAN-FPN structure to fuse high-level and low-level features. Each sub-model output additionally includes a confidence score, consisting of classification confidence and bounding box. The product of the predicted values ​​is calculated; for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0012] The annotation consistency voting module operates as follows: Input the training set samples into the N sub-models that have been trained, collect the predicted class yij, predicted bounding box bij, and confidence score cij of each sample; calculate the class consistency probability P(yi) and the average bounding box IoUi of each sample, calculate the annotation confidence score Si, preset the balance constant α and the threshold τ for suspicious samples, and label the suspicious samples.

[0013] The adaptive relabeling module for suspicious samples operates as follows: when a suspicious sample shows high consistency in the prediction results of multiple sub-models, the prediction result is considered to have high credibility, and the original label is replaced by the model consensus result; if the prediction results of a suspicious sample differ greatly among the sub-models, it is sent to the manual review stage.

[0014] The noise perception loss function module operates as follows: The noise loss function consists of two parts: classification loss and regression loss. By introducing labeled confidence as a regulating factor, the classification loss and regression loss are dynamically adjusted to weaken the influence of low-confidence samples and are used in the sub-model training process.

[0015] Input a new power distribution network image, perform inference on the trained multi-model collaborative defect detection model, and output defect detection results, including defect category, bounding box, and confidence level.

[0016] Furthermore, the multi-expert integration module operation refers to: building multiple sub-model frameworks based on YOLOv8, optimizing the feature extraction network of each sub-model, embedding a Coordinate Attention layer after the C3 module of YOLOv8, and focusing on key defect areas such as insulator edges and conductor connection points by combining channel attention with spatial coordinate information; then, using cross-scale skip connections to fuse high and low layer features to solve the problem of defect scale differences; the output layer additionally generates prediction confidence, which is composed of classification confidence and bounding box. The product of the predicted values ​​is calculated; for the feature map after cross-scale fusion, feature enhancement is performed by linearly adjusting the amplitude; for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0017] Furthermore, the weather interference quantification module calculates the weather interference coefficient using the grayscale standard deviation and contrast of the power distribution network image. The specific calculation formula is as follows:

[0018]

[0019] in, The grayscale standard deviation of the input image. The contrast of the input image. This represents the standard grayscale standard deviation of a power distribution network image under normal weather conditions. Standard contrast for power distribution network images under normal weather conditions;

[0020] The formula for calculating the standard deviation of gray level is as follows:

[0021]

[0022] in, These are the height and width of the image, respectively. The grayscale value of a pixel. The average gray value of the image;

[0023] Contrast The calculation formula is as follows:

[0024]

[0025] in, , These represent the maximum and minimum grayscale values ​​of the image, respectively. Using the above formula, the degree of interference from weather events such as heavy rain, fog, sandstorms, and strong sunlight can be quickly quantified without the need for complex feature extraction or model training.

[0026] Furthermore, the multi-expert integration module comprises N structurally similar but functionally independent target detection sub-models, denoted as... All sub-models are based on the YOLOv8 framework and improved upon.

[0027] In the feature extraction network part of the sub-model, a CoordinateAttention layer is added after the C3 module of YOLOv8. Coordinate attention combines channel attention with spatial coordinate information, and its calculation formula is as follows:

[0028]

[0029] in, The feature map output by module C3. For global average pooling, It is a fully connected layer. It is the Sigmoid activation function. For convolution operations; after the YOLOv8 PAN-FPN result, a cross-scale skip connection is added to more tightly fuse high-level semantic features with high-level and low-level detail features. After unifying the number of channels through 1×1 convolution, element-wise summation is performed, as shown in the formula:

[0030]

[0031] in Characterized by high-level features These are low-level features. Finally, the output layer is adjusted; in addition to the usual class probabilities, the output layer of each sub-model is adjusted. and bounding box coordinates In addition, output confidence level Confidence level is used to measure how confident a model is in its own predictions. Classification confidence and bounding box The product of the predicted values ​​is calculated using the following formula:

[0032]

[0033] Subsequently, a unified lightweight feature enhancement strategy was adopted to address the common problems of all severe weather conditions:

[0034]

[0035] in, To enhance the coefficients, the feature map amplitude is linearly adjusted to increase the difference between the defect area and the background, thus offsetting the decrease in contrast caused by weather interference.

[0036] Furthermore, for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0037] Initialize parameters randomly, with all sub-models using different random seeds for weight initialization; let the sub-models... The initial weights are ,but Furthermore, the random seeds for each model are mutually exclusive;

[0038] The data augmentation strategy is differentiated by assigning an independent data augmentation pipeline to each sub-model. By changing the visual features of the image, such as brightness, scale, and angle, the sub-model is forced to learn the defect features from different perspectives.

[0039] The training hyperparameters are fine-tuned, with each sub-model using slightly different training hyperparameters, including the learning rate and weight decay coefficient.

[0040] Furthermore, the consistency voting module is executed, and the specific process is as follows:

[0041] First, we measure the consistency of category labeling. Let's assume that for each input sample... Each sub-model Output predicted category during forward inference. Predicting bounding boxes and the corresponding prediction confidence level ;

[0042] First, measure the consistency of category labeling. Let the labeled categories of the training set samples be... Then the probability of class consistency can be expressed as:

[0043]

[0044] in This is an indicator function; it takes the value 1 when the predicted category and the labeled category are the same, and 0 otherwise. A higher value indicates stronger consistency between the sub-models and the labeled category. Secondly, the consistency of bounding box annotations is measured using the average IoU as the metric.

[0045]

[0046] in, The labeled bounding boxes provided in the training data;

[0047] Finally, define the label confidence score. Its form is:

[0048]

[0049] in, This is a balancing constant used to control the proportion of category consistency and bounding box consistency in the overall confidence score. If the sample is not identified as a suspicious sample, it will be automatically marked as a suspicious sample and will proceed to the subsequent adaptive relabeling stage.

[0050] Furthermore, the adaptive relabeling module for suspicious samples operates as follows: when a suspicious sample shows high consistency in the prediction results of multiple sub-models, the prediction result is considered to have high credibility, and the original label is replaced by the model consensus result. The implementation method is as follows:

[0051]

[0052]

[0053] in For the first Sub-model prediction samples The probability, This represents the confidence level of the corresponding prediction. A weighted average approach yields more robust bounding box predictions. ;

[0054] Manual verification refers to situations where there is significant discrepancy in the prediction results of suspicious samples among the sub-models, specifically when the voting ratio of the largest category is below a threshold. It is then sent to the manual review stage.

[0055] Furthermore, the noise perception loss function consists of two parts: classification loss and regression loss. The classification loss adopts the form of cross-entropy, and its definition is:

[0056]

[0057] in The total number of categories, For the sample Category The probability, One-hot vectors labeled with categories; weight coefficients With label confidence Dynamic correlation, in the form of:

[0058]

[0059] in, To adjust the parameters, when When the value is low, the classification loss weights are automatically reduced to avoid excessive interference from noise labels on the model;

[0060] For the regression loss, confidence-weighted methods are used for bounding box prediction. loss:

[0061]

[0062] in, To predict the bounding box, To annotate the bounding box, weights are introduced. This can reduce the impact of the sample on the regression task when the confidence level is low;

[0063] Finally, the total loss function combines the classification loss and the regression loss, defining the overall training loss as:

[0064]

[0065] in, For batch size, This is a balancing coefficient used to control the relative importance of classification and regression tasks.

[0066] A defect detection device for power distribution network equipment based on multi-model collaboration, comprising:

[0067] The system comprises a dataset preparation module, a weather interference quantification module, a multi-expert integration module, a label consistency voting module, a suspicious sample adaptive relabeling module, a noise perception loss function module, and an inference output module. These modules work in tandem to achieve accurate detection of defects in power distribution network equipment.

[0068] The dataset preparation module is used to acquire a large number of photos of power distribution network equipment defects under different weather interference conditions, and to construct a power distribution network defect detection dataset with noise annotation, so as to provide data support for subsequent model construction and training.

[0069] The weather interference quantification module is used to calculate and store the standard grayscale standard deviation σ0 and standard contrast C0 as the benchmark values ​​for weather interference quantification based on the normal weather power distribution network photos provided by the dataset preparation module, so as to ensure the horizontal comparability of interference coefficients of different images; it is also used to calculate the grayscale standard deviation σ and contrast C of each power distribution network image to be processed, and substitute them into the formula ω=1-σ·C / (σ0·C0) to calculate the weather interference intensity coefficient ω. The closer the ω value is to 0, the stronger the weather interference is, and the farther away from 0, the weaker the interference is. The calculated ω value is then passed to the multi-expert integration module.

[0070] The multi-expert integration module is used to build multiple sub-model frameworks based on YOLOv8. A Coordinate Attention layer is added to the feature extraction network of each sub-model to enhance sensitivity to defect location information. Cross-scale skip connections are added to the PAN-FPN structure to fuse high-level and low-level features. The output of each sub-model includes defect category, bounding box, and additional confidence score, which is calculated as the product of classification confidence score and bounding box IoU prediction value. It is also used to configure independent training conditions for each sub-model through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning. It receives the ω value transmitted by the weather interference quantification module and combines it with the sub-model training and inference process.

[0071] The annotation consistency voting module is used to input the training set samples constructed by the dataset preparation module into the N sub-models that have been trained, collect the predicted class yij, predicted bounding box bij and confidence score cij of each sub-model corresponding to each sample; it is also used to calculate the class consistency probability P(yi) and the average bounding box IoUi of each sample, and then calculate the annotation confidence score Si, preset the balance constant α and the suspicious sample threshold τ, mark the suspicious samples according to the comparison result of Si and τ, and pass the marking result to the suspicious sample adaptive re-labeling module;

[0072] The adaptive relabeling module for suspicious samples is used to receive suspicious samples marked by the labeling consistency voting module and the prediction results of the corresponding sub-models. If a suspicious sample shows high consistency in the prediction results of multiple sub-models, the original label is replaced by the model consensus result and the dataset is updated. If the prediction results of a suspicious sample differ greatly among the sub-models, the sample is sent to the manual review stage, and the labeling information is updated after the review is completed.

[0073] The noise-perceived loss function module is used to construct a noise loss function consisting of classification loss and regression loss. By introducing the label confidence Si output by the label consistency voting module as an adjustment factor, the weights of classification loss and regression loss are dynamically adjusted to weaken the impact of low-confidence samples on model training and provide loss calculation support for the training of sub-models in the multi-expert integration module.

[0074] The inference output module is used to receive the new power distribution network image to be detected, input it into the trained multi-model collaborative defect detection model, which is composed of a weather interference quantification module, a multi-expert integration module, a labeling consistency voting module, a suspicious sample adaptive relabeling module, and a noise perception loss function module, to perform inference and output the defect detection result, which includes the defect category, bounding box, and confidence level.

[0075] Furthermore, in the weather interference quantification module, the grayscale standard deviation σ is obtained by calculating the square root of the sum of the squares of the differences between the image pixel grayscale values ​​and the average grayscale values, and the contrast ratio C is obtained by calculating the difference between the maximum grayscale value and the minimum grayscale value in the image.

[0076] The beneficial effects of this invention are as follows:

[0077] This invention, by introducing multi-expert model ensemble, a label consistency voting module, an adaptive relabeling mechanism for suspicious samples, and a noise-aware loss function, effectively improves the robustness and detection accuracy of the model in noisy data environments for power distribution network equipment defect detection. Compared with existing noise-resistant methods that rely on manual data cleaning or single-model regularization, this invention has the following significant advantages: First, multi-expert model ensemble, while ensuring real-time detection, enhances the ability to identify erroneous labels by utilizing the differences between sub-models; Second, the label consistency voting module, by quantifying the consistency between categories and bounding boxes, can automatically screen potential noisy samples, achieving dynamic evaluation of data quality; Third, the adaptive relabeling mechanism for suspicious samples effectively reduces the cost of manual review, allowing the training dataset to be gradually purified during iteration, ensuring the reliability of the supervision signal; Fourth, the noise-aware loss function introduces a label confidence adjustment factor in the loss calculation, enabling reinforcement learning on high-quality samples and suppression training on low-quality samples, thereby avoiding model overfitting to noisy labels. In summary, the present invention can maintain a high accuracy and stability in defect identification even when there is large-scale annotation noise in the distribution network inspection images, thus meeting the actual needs of smart grids for high reliability and real-time performance, and has broad application value and promotion significance. Attached Figure Description

[0078] Figure 1 This is a flowchart of a method for detecting defects in power distribution network equipment based on multi-model collaboration.

[0079] Figure 2 This is an architecture diagram of a multi-model collaborative defect detection method for power distribution network equipment.

[0080] Figure 3 This is a flowchart for defect detection based on a multi-model collaborative defect detection model for power distribution network equipment. Detailed Implementation

[0081] The present invention will be further described below with reference to the accompanying drawings.

[0082] A method for detecting defects in power distribution network equipment based on multi-model collaboration includes the following:

[0083] Prepare for the construction of the power distribution network defect detection dataset: acquire a large number of photos of power distribution network equipment defects under different weather interference conditions, which may include noise annotations;

[0084] Construction and training of a multi-model collaborative defect detection model for power distribution networks: The model includes a weather interference quantification module, a multi-expert integration module, a labeling consistency voting module, a suspicious sample adaptive relabeling module, and a noise perception loss function module. The specific process is as follows:

[0085] The weather interference quantification module operates as follows: Based on collected normal weather power distribution network photos, it calculates and stores the standard grayscale standard deviation σ0 and standard contrast C0 as benchmark values ​​for weather interference quantification, ensuring that the interference coefficients of different images can be compared horizontally. For each power distribution network image, it calculates its grayscale standard deviation σ and contrast C, and then substitutes them into the formula ω=1-σ·C / (σ0·C0) to calculate the weather interference intensity coefficient ω. The closer the ω value is to 0, the stronger the weather interference, and vice versa. The calculated ω value is then passed to the multi-expert integration module.

[0086] Multi-expert ensemble module operation: Based on YOLOv8, multiple sub-model frameworks are built. A Coordinate Attention layer is added to the feature extraction network to enhance the sensitivity to defect location information. Cross-scale skip connections are added to the PAN-FPN structure to fuse high-level and low-level features. Each sub-model output additionally includes a confidence score, consisting of classification confidence and bounding box. The product of the predicted values ​​is calculated; for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0087] The annotation consistency voting module operates as follows: Input the training set samples into the N sub-models that have been trained, collect the predicted class yij, predicted bounding box bij, and confidence score cij of each sample; calculate the class consistency probability P(yi) and the average bounding box IoUi of each sample, calculate the annotation confidence score Si, preset the balance constant α and the threshold τ for suspicious samples, and label the suspicious samples.

[0088] The adaptive relabeling module for suspicious samples operates as follows: when a suspicious sample shows high consistency in the prediction results of multiple sub-models, the prediction result is considered to have high credibility, and the original label is replaced by the model consensus result; if the prediction results of a suspicious sample differ greatly among the sub-models, it is sent to the manual review stage.

[0089] The noise perception loss function module operates as follows: The noise loss function consists of two parts: classification loss and regression loss. By introducing labeled confidence as a regulating factor, the classification loss and regression loss are dynamically adjusted to weaken the influence of low-confidence samples and are used in the sub-model training process.

[0090] Input a new power distribution network image, perform inference on the trained multi-model collaborative defect detection model, and output defect detection results, including defect category, bounding box, and confidence level.

[0091] Furthermore, in another embodiment of the present invention, the operation of the multi-expert integration module refers to: building multiple sub-model frameworks based on YOLOv8, optimizing the feature extraction network of each sub-model, embedding a Coordinate Attention layer after the C3 module of YOLOv8, and focusing on key defect areas such as insulator edges and conductor connection points by combining channel attention with spatial coordinate information; then, using cross-scale skip connections to fuse high and low layer features to solve the problem of defect scale differences; the output layer additionally generates prediction confidence, which is composed of classification confidence and bounding box. The product of the predicted values ​​is calculated; for the feature map after cross-scale fusion, feature enhancement is performed by linearly adjusting the amplitude; for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0092] Furthermore, in another embodiment of the present invention, for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0093] Initialize parameters randomly, with all sub-models using different random seeds for weight initialization; let the sub-models... The initial weights are ,but Furthermore, the random seeds for each model are mutually exclusive;

[0094] The data augmentation strategy is differentiated by assigning an independent data augmentation pipeline to each sub-model. By changing the visual features of the image, such as brightness, scale, and angle, the sub-model is forced to learn the defect features from different perspectives.

[0095] The training hyperparameters are fine-tuned, with each sub-model using slightly different training hyperparameters, including the learning rate and weight decay coefficient.

[0096] Furthermore, in another embodiment of the present invention, the weather interference quantification module calculates the weather interference coefficient using the grayscale standard deviation and contrast of the power distribution network image. The specific calculation formula is as follows:

[0097]

[0098] in, The grayscale standard deviation of the input image. The contrast of the input image. This represents the standard grayscale standard deviation of a power distribution network image under normal weather conditions. This represents the standard contrast of a power distribution network image under normal weather conditions. The formula for calculating the standard deviation of grayscale is as follows:

[0099]

[0100] in, These are the height and width of the image, respectively. The grayscale value of a pixel. This represents the average grayscale value of the image.

[0101] Contrast The calculation formula is as follows:

[0102]

[0103] in, , These represent the maximum and minimum grayscale values ​​of the image, respectively. Using the above formula, the degree of interference from weather events such as heavy rain, fog, sandstorms, and strong sunlight can be quickly quantified without the need for complex feature extraction or model training.

[0104] The multi-expert integration module contains N structurally similar but functionally independent object detection sub-models, denoted as... All sub-models are based on the YOLOv8 framework and improved to balance detection accuracy and real-time requirements.

[0105] The improvements to the sub-model are mainly reflected in the feature extraction network. A Coordinate Attention layer is added after the C3 module of YOLOv8 to enhance the model's sensitivity to defect location information. Coordinate Attention, by combining channel attention with spatial coordinate information, can effectively focus on key defect areas such as insulator edges and conductor connection points, suppressing background interference. Its calculation formula is as follows:

[0106]

[0107] in, The feature map output by module C3. For global average pooling, It is a fully connected layer. It is the Sigmoid activation function. This is a convolution operation. To address the large scale differences in defects in distribution network equipment, a cross-scale skip connection is added after the YOLOv8 PAN-FPN results. This more tightly integrates high-level semantic features with high-level and low-level detail features. After unifying the number of channels through a 1×1 convolution, element-wise summation is performed, as shown in the formula:

[0108]

[0109] in Characterized by high-level features These are low-level features. Finally, the output layer is adjusted; in addition to the usual class probabilities, the output layer of each sub-model is adjusted. and bounding box coordinates In addition, output confidence level Confidence level is used to measure how confident a model is in its own predictions. Classification confidence and bounding box The product of the predicted values ​​is calculated using the following formula:

[0110]

[0111] Subsequently, a unified lightweight feature enhancement strategy was adopted to address the common problems of all severe weather conditions.

[0112]

[0113] in, To enhance the coefficients, the feature map amplitude is linearly adjusted to increase the difference between the defect area and the background, thus offsetting the decrease in contrast caused by weather interference.

[0114] The diversity of sub-models is a core prerequisite for the multi-expert ensemble module to identify annotation noise. If all sub-model outputs are highly consistent, it is impossible to distinguish noise from true annotations through voting. Therefore, this invention ensures model diversity through the following three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

[0115] In this process, the initialization parameters are randomized, with all sub-models using different random seeds for weight initialization. Let the sub-model... The initial weights are ,but Furthermore, the random seeds of each model are mutually exclusive. This initialization difference causes the sub-models to have different prediction biases for the same defective sample in the early stages of training, providing discriminative power for subsequent voting.

[0116] The data augmentation strategy differentiates each sub-model by assigning an independent data augmentation pipeline. By altering the visual features of the image, such as brightness, scale, and angle, the sub-models are forced to learn defect features from different viewpoints.

[0117] In this process, the training hyperparameters are fine-tuned, with each sub-model using slightly different training hyperparameters, including learning rate and weight decay coefficient. These differences in hyperparameters result in varying convergence speeds and overfitting tendencies among the sub-models, further enhancing diversity.

[0118] The core logic of the annotation consistency voting module is to construct an annotation confidence evaluation system based on the prediction results of N sub-models output by the multi-expert model integration module. This is achieved by quantifying the matching degree between sub-model predictions and human annotations, as well as the consistency degree between predictions among sub-models, ultimately realizing accurate annotation of suspicious samples. This module simultaneously handles category annotation noise and bounding box annotation noise. The specific process is as follows:

[0119] First, we measure the consistency of category labeling. Let's assume that for each input sample... Each sub-model Output predicted category during forward inference. Predicting bounding boxes and the corresponding prediction confidence level

[0120] First, we measure the consistency of category labeling. Let the labeled categories of the training set samples be... Then the probability of class consistency can be expressed as:

[0121]

[0122] in This is an indicator function; it takes the value 1 when the predicted category and the labeled category are the same, and 0 otherwise. A higher value indicates stronger consistency between the sub-models and the labeled category. Secondly, the consistency of bounding box annotations is measured. Average IoU is used as the metric:

[0123]

[0124] in, This refers to the labeled bounding boxes provided in the training data. This metric measures the degree of match between the predicted bounding boxes and the labeled bounding boxes.

[0125] Finally, define the label confidence score. Its form is:

[0126]

[0127] in, This is a balancing constant used to control the proportion of category consistency and bounding box consistency in the overall confidence score. If If the sample is not identified as a suspicious sample, it will be automatically marked as a suspicious sample and will proceed to the subsequent adaptive relabeling stage.

[0128] The adaptive relabeling module for suspicious samples is designed to automatically correct noisy labels in the dataset while reducing manual intervention, thereby ensuring the overall reliability of the training data. This mechanism mainly includes two strategies: automatic relabeling based on model consensus and manual-assisted relabeling.

[0129] Automatic relabeling based on model consensus considers a suspicious sample to have high reliability when the predictions from multiple sub-models show high consistency. The original labeling is then replaced with the model consensus result. The implementation method is as follows:

[0130]

[0131]

[0132] in For the first Sub-model prediction samples The probability, This represents the confidence level of the corresponding prediction. A weighted average approach yields more robust bounding box predictions. .

[0133] Manual verification refers to situations where there is significant discrepancy in the prediction results of suspicious samples among the sub-models, specifically when the voting ratio of the largest category is below a threshold. If a sample is deemed unsuitable for automatic model correction, the system will filter it out and send it for manual review. By manually intervening only in a few high-uncertainty samples, the overall cost of manual annotation can be significantly reduced.

[0134] By organically combining the two mechanisms mentioned above, the training dataset is gradually purified during the iterative update process, thereby improving the overall quality of the labeled data and the stability of model training.

[0135] Noise-aware loss function module: To address the problem that traditional object detection loss functions are easily misled in noisy labeling scenarios, this loss function introduces label confidence as an adjustment factor to achieve reinforcement learning for high-confidence samples and weakening constraints for low-confidence samples, thereby improving the robustness of the model in noisy environments.

[0136] Specifically, the noise loss function consists of two parts: classification loss and regression loss. The classification loss takes the form of cross-entropy, and is defined as follows:

[0137]

[0138] in The total number of categories, For the sample Category The probability, This is a one-hot vector labeled with the category. Weight coefficients. With label confidence Dynamic correlation, in the form of:

[0139]

[0140] in, To adjust the parameters, when When the threshold is low, the classification loss weights are automatically reduced to avoid excessive interference from noisy labels on the model. For the regression loss, confidence-weighted methods are used for bounding box prediction. loss:

[0141]

[0142] in, To predict the bounding box, To annotate the bounding box, weights are introduced. This can reduce the impact of the sample on the regression task when the confidence level is low.

[0143] Finally, the total loss function combines the classification loss and the regression loss, defining the overall training loss as:

[0144]

[0145] in, For batch size, This is a balancing coefficient used to control the relative importance of classification and regression tasks.

[0146] Example 2

[0147] like Figure 1 As shown, the present invention discloses a method for detecting defects in distribution network equipment based on multi-model collaboration. First, it is necessary to collect images of distribution network defects under severe weather conditions, and then input them into a pre-trained model to obtain the defect detection results.

[0148] like Figure 2 As shown, a defect detection device for power distribution network equipment based on multi-model collaboration is characterized by comprising:

[0149] The system comprises a dataset preparation module, a weather interference quantification module, a multi-expert integration module, a label consistency voting module, a suspicious sample adaptive relabeling module, a noise perception loss function module, and an inference output module. These modules work in tandem to achieve accurate detection of defects in power distribution network equipment.

[0150] The dataset preparation module is used to acquire a large number of photos of power distribution network equipment defects under different weather interference conditions, and to construct a power distribution network defect detection dataset with noise annotation, so as to provide data support for subsequent model construction and training.

[0151] The weather interference quantification module is used to calculate and store the standard grayscale standard deviation σ0 and standard contrast C0 as the benchmark values ​​for weather interference quantification based on the normal weather power distribution network photos provided by the dataset preparation module, so as to ensure the horizontal comparability of interference coefficients of different images; it is also used to calculate the grayscale standard deviation σ and contrast C of each power distribution network image to be processed, and substitute them into the formula ω=1-σ·C / (σ0·C0) to calculate the weather interference intensity coefficient ω. The closer the ω value is to 0, the stronger the weather interference is, and the farther away from 0, the weaker the interference is. The calculated ω value is then passed to the multi-expert integration module.

[0152] The multi-expert integration module is used to build multiple sub-model frameworks based on YOLOv8. A Coordinate Attention layer is added to the feature extraction network of each sub-model to enhance sensitivity to defect location information. Cross-scale skip connections are added to the PAN-FPN structure to fuse high-level and low-level features. The output of each sub-model includes defect category, bounding box, and additional confidence score, which is calculated as the product of classification confidence score and bounding box IoU prediction value. It is also used to configure independent training conditions for each sub-model through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning. It receives the ω value transmitted by the weather interference quantification module and combines it with the sub-model training and inference process.

[0153] The annotation consistency voting module is used to input the training set samples constructed by the dataset preparation module into the N sub-models that have been trained, collect the predicted class yij, predicted bounding box bij and confidence score cij of each sub-model corresponding to each sample; it is also used to calculate the class consistency probability P(yi) and the average bounding box IoUi of each sample, and then calculate the annotation confidence score Si, preset the balance constant α and the suspicious sample threshold τ, mark the suspicious samples according to the comparison result of Si and τ, and pass the marking result to the suspicious sample adaptive re-labeling module;

[0154] The adaptive relabeling module for suspicious samples is used to receive suspicious samples marked by the labeling consistency voting module and the prediction results of the corresponding sub-models. If a suspicious sample shows high consistency in the prediction results of multiple sub-models, the original label is replaced by the model consensus result and the dataset is updated. If the prediction results of a suspicious sample differ greatly among the sub-models, the sample is sent to the manual review stage, and the labeling information is updated after the review is completed.

[0155] The noise-perceived loss function module is used to construct a noise loss function consisting of classification loss and regression loss. By introducing the label confidence Si output by the label consistency voting module as an adjustment factor, the weights of classification loss and regression loss are dynamically adjusted to weaken the impact of low-confidence samples on model training and provide loss calculation support for the training of sub-models in the multi-expert integration module.

[0156] The inference output module is used to receive the new power distribution network image to be detected, input it into the trained multi-model collaborative defect detection model, which is composed of a weather interference quantification module, a multi-expert integration module, a labeling consistency voting module, a suspicious sample adaptive relabeling module, and a noise perception loss function module, to perform inference and output the defect detection result, which includes the defect category, bounding box, and confidence level.

[0157] In the weather interference quantification module, the grayscale standard deviation σ is obtained by calculating the square root of the sum of the squares of the differences between the image pixel grayscale values ​​and the average grayscale values, and the contrast ratio C is obtained by calculating the difference between the maximum grayscale value and the minimum grayscale value in the image.

[0158] Furthermore, in another embodiment of the present invention, the multi-expert integration module trains multiple target detection sub-models with similar structures but different initializations. By using methods such as random seed initialization, differential data augmentation, and training hyperparameter fine-tuning, the diversity among the sub-models is ensured, thereby achieving independent learning and collaborative prediction on noisy datasets, enhancing the algorithm's ability to identify labeled noise and improving detection accuracy.

[0159] Furthermore, in another embodiment of the present invention, the annotation consistency voting module, based on the prediction results of N sub-models of the multi-expert model integration module, quantifies the matching degree between the predicted category and the manually labeled category, as well as the average of the predicted bounding box and the labeled bounding box. Value, combining the two to form a balance constant Constructing labeled confidence scores ,like Below the preset threshold If the sample is automatically marked as a suspicious sample, it can achieve accurate screening of category and bounding box noise and trigger subsequent relabeling.

[0160] Furthermore, in another embodiment of the present invention, the adaptive relabeling module for suspicious samples automatically corrects noise labels through two strategies. When the prediction results of the sub-models are highly consistent, new labels and bounding box coordinates are generated using a model consensus weighted average for automatic relabeling; if the prediction discrepancies are large, samples are selected and sent to manual review for expert correction, thereby gradually purifying the training dataset while reducing manual intervention and improving the overall labeling quality and model stability.

[0161] Furthermore, in another embodiment of the present invention, the noise-aware loss function module dynamically adjusts the classification loss and regression loss by introducing labeled confidence as an adjustment factor, strengthens learning on high-confidence samples while weakening constraints on low-confidence samples, avoids the misleading influence of noise labels on model training, and thus significantly improves the robustness and detection performance of the model in noisy environments.

[0162] Application example:

[0163] like Figure 3The diagram shows the flowchart of the method of this invention. This method is applied to the inspection scenario of a power distribution network in a prefecture-level city. It targets core equipment such as 10kV line insulators and circuit breakers, solving the problems of missed and false detections of defects under complex weather conditions such as rainy days, heavy fog, and strong sunlight, thereby improving inspection efficiency and accuracy. This application covered 200km of power distribution network lines in three jurisdictions, deployed 50 high-definition inspection drones, and processed more than 8,000 images.

[0164] During the dataset construction phase, drones were used to collect images of equipment under different weather conditions, including 4,000 images on sunny days, 2,000 images on rainy days, 1,200 images on foggy days, and 800 images on bright sunlight days. These images cover six types of defects, such as insulator damage, circuit breaker overheating, and conductor corrosion. Some manual annotations were noisy due to rain and fog obscuring the images. The final dataset consists of 7,200 training images and 800 test images.

[0165] In the model building and training phase, the weather interference quantification module first calculates the standard deviation of standard grayscale σ0=80 and the standard contrast C0=0.4 for clear day images as baseline values. After calculating σ and C for each image, the ω value is obtained. The mean ω value for foggy day images is 0.21 (strong interference), and the mean ω value for clear day images is 0.83 (weak interference). These values ​​are then simultaneously transmitted to the multi-expert integration module.

[0166] The multi-expert ensemble module is built on YOLOv8 with 5 sub-models, each with a Coordinate Attention layer and cross-scale jump connection. Independent training is achieved by randomly initializing parameters, differential data augmentation (such as adding Gaussian noise to sub-model 1 and adjusting brightness in sub-model 2) and fine-tuning the learning rate. The confidence of the sub-model is calculated by multiplying the classification confidence and the IoU prediction value.

[0167] The consistency voting module votes on the training set samples, setting α=0.6 and τ=0.5, identifying 412 suspicious samples. The adaptive relabeling module for suspicious samples automatically updates the labels on 308 samples whose sub-model predictions are consistent (over 60% of sub-models share the same class, IoU ≥ 0.6). The remaining 104 divergent samples are manually reviewed and corrected to improve dataset quality. The noise-aware loss function module introduces Si to dynamically adjust the loss weights, mitigating the impact of low-confidence samples. Subsequently, in the noise-aware loss function module, the classification loss is defined as weighted cross-entropy, where the weight coefficients are positively correlated with the label confidence and adjusted via an exponential function. The loss weights for low-confidence samples are dynamically adjusted to avoid misleading the system. The CIOU loss, which is a confidence-weighted regression loss, also incorporates a weighting factor. The total loss function is a weighted sum of the classification and regression losses, using a balance coefficient. =1.0 controls the weighting of both and optimizes them in batch processing.

[0168] During the actual inference phase, the drone transmits new images in real time, which are then processed by the model to output the results. The accuracy rate for insulator damage detection in foggy weather reaches 89.3%, a 15.7% improvement compared to traditional single-model approaches; the missed detection rate for circuit breaker overheating defects in rainy weather drops to 3.2%. Detection results are simultaneously pushed to the operation and maintenance platform, annotating the defect location, type, and confidence level to guide maintenance personnel in precise repairs, shortening the inspection cycle from 15 days to 7 days, improving fault handling efficiency by 60%, and significantly reducing the risk of power distribution network outages.

[0169] The above description, in conjunction with specific / preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. Those skilled in the art can make various substitutions or modifications to these described embodiments without departing from the inventive concept, and all such substitutions or modifications should be considered within the scope of protection of the present invention.

[0170] The parts of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A method for defect detection in power distribution network equipment based on multi-model collaboration, characterized in that, Includes the following: Prepare for the construction of the power distribution network defect detection dataset: acquire a large number of photos of power distribution network equipment defects under different weather interference conditions, which may include noise annotations; Construction and training of a multi-model collaborative defect detection model for power distribution networks: The model includes a weather interference quantification module, a multi-expert integration module, a labeling consistency voting module, a suspicious sample adaptive relabeling module, and a noise perception loss function module. The specific process is as follows: The weather interference quantification module operates as follows: Based on collected normal weather power distribution network photos, it calculates and stores the standard grayscale standard deviation σ0 and standard contrast C0 as benchmark values ​​for weather interference quantification, ensuring that the interference coefficients of different images can be compared horizontally. For each power distribution network image, it calculates its grayscale standard deviation σ and contrast C, and then substitutes them into the formula ω=1-σ·C / (σ0·C0) to calculate the weather interference intensity coefficient ω. The closer the ω value is to 0, the stronger the weather interference, and vice versa. The calculated ω value is then passed to the multi-expert integration module. Multi-expert ensemble module operation: Based on YOLOv8, multiple sub-model frameworks are built. A Coordinate Attention layer is added to the feature extraction network to enhance the sensitivity to defect location information. Cross-scale skip connections are added to the PAN-FPN structure to fuse high-level and low-level features. Each sub-model output additionally includes a confidence score, consisting of classification confidence and bounding box. The product of the predicted values ​​is calculated. For each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning. The annotation consistency voting module operates as follows: Input the training set samples into the N sub-models that have been trained, collect the predicted class yij, predicted bounding box bij, and confidence score cij of each sample; calculate the class consistency probability P(yi) and the average bounding box IoUi of each sample, calculate the annotation confidence score Si, preset the balance constant α and the threshold τ for suspicious samples, and label the suspicious samples. The adaptive relabeling module for suspicious samples operates as follows: when a suspicious sample shows high consistency in the prediction results of multiple sub-models, the prediction result is considered to have high credibility, and the original label is replaced by the model consensus result; if the prediction results of a suspicious sample differ greatly among the sub-models, it is sent to the manual review stage. The noise perception loss function module operates as follows: The noise loss function consists of two parts: classification loss and regression loss. By introducing labeled confidence as a regulating factor, the classification loss and regression loss are dynamically adjusted to weaken the influence of low-confidence samples and are used in the sub-model training process. Input a new power distribution network image, perform inference on the trained multi-model collaborative defect detection model, and output defect detection results, including defect category, bounding box, and confidence level.

2. The method according to claim 1, characterized in that, The multi-expert ensemble module operation refers to: building multiple sub-model frameworks based on YOLOv8, optimizing the feature extraction network of each sub-model, embedding a Coordinate Attention layer after the C3 module of YOLOv8, and focusing on key defect areas such as insulator edges and conductor connection points by combining channel attention with spatial coordinate information; then, using cross-scale skip connections to fuse high and low layer features to solve the problem of defect scale differences; the output layer additionally generates prediction confidence, which is composed of classification confidence and bounding box. The product of the predicted values ​​is calculated. For the feature maps that have undergone cross-scale fusion, feature enhancement is performed by linearly adjusting the amplitude; for each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning.

3. The method according to claim 1, characterized in that, The aforementioned weather interference quantification module calculates the weather interference coefficient using the grayscale standard deviation and contrast of the power distribution network image. The specific calculation formula is as follows: in, The grayscale standard deviation of the input image. The contrast of the input image. This represents the standard grayscale standard deviation of a power distribution network image under normal weather conditions. Standard contrast for power distribution network images under normal weather conditions; The formula for calculating the standard deviation of gray level is as follows: in, These are the height and width of the image, respectively. The grayscale value of a pixel. The average gray value of the image; Contrast The calculation formula is as follows: in, , These are the maximum and minimum grayscale values ​​of the image, respectively. Using the above formula, the degree of interference from weather events such as heavy rain, fog, sandstorms, and strong light can be quickly quantified without the need for complex feature extraction or model training.

4. The method according to claim 1, characterized in that, The multi-expert integration module contains N structurally similar but functionally independent object detection sub-models, denoted as... All sub-models are based on the YOLOv8 framework and improved upon. In the feature extraction network part of the sub-model, a CoordinateAttention layer is added after the C3 module of YOLOv8. Coordinate attention combines channel attention with spatial coordinate information, and its calculation formula is as follows: in, The feature map output by module C3. For global average pooling, It is a fully connected layer. It is the Sigmoid activation function. For convolution operations; after the YOLOv8 PAN-FPN result, a cross-scale skip connection is added to more tightly fuse high-level semantic features with high-level and low-level detail features. After unifying the number of channels through 1×1 convolution, element-wise summation is performed, as shown in the formula: in Characterized by high-level features For low-level features, the output layer is adjusted last. In addition to the usual class probabilities, the output layer of each sub-model... and bounding box coordinates In addition, output confidence level Confidence level is used to measure the confidence level of a model in its own predictions. Classification confidence and bounding box The product of the predicted values ​​is calculated using the following formula: Subsequently, a unified lightweight feature enhancement strategy was adopted to address the common problems of all severe weather conditions: in, To enhance the coefficients, the feature map amplitude is linearly adjusted to increase the difference between the defect area and the background, thus offsetting the decrease in contrast caused by weather interference.

5. The method according to claim 4, characterized in that, For each sub-model, independent training conditions are configured through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning. Initialize parameters randomly, with all sub-models using different random seeds for weight initialization; let the sub-models... The initial weights are ,but Furthermore, the random seeds for each model are mutually exclusive; The data augmentation strategy is differentiated by assigning an independent data augmentation pipeline to each sub-model. By changing the visual features of the image, such as brightness, scale, and angle, the sub-model is forced to learn the defect features from different perspectives. The training hyperparameters are fine-tuned, with each sub-model using slightly different training hyperparameters, including the learning rate and weight decay coefficient.

6. The method according to claim 1, characterized in that, The consistency voting module runs as follows: First, we measure the consistency of category labeling. Let's assume that for each input sample... Each sub-model Output predicted category during forward inference. Predicting bounding boxes and the corresponding prediction confidence level ; First, measure the consistency of category labeling. Let the labeled categories of the training set samples be... Then the probability of class consistency can be expressed as: in This is an indicator function; it takes the value 1 when the predicted category and the labeled category match, and 0 otherwise. A higher value indicates stronger consistency between the sub-models and the labeled category. Secondly, the consistency of bounding box annotations is measured using the average IoU as the metric. in, The labeled bounding boxes provided in the training data; Finally, define the label confidence score. Its form is: in, This is a balancing constant used to control the proportion of category consistency and bounding box consistency in the overall confidence score. If the sample is not identified as a suspicious sample, it will be automatically marked as a suspicious sample and will proceed to the subsequent adaptive relabeling stage.

7. The method according to claim 1, characterized in that, The adaptive relabeling module for suspicious samples operates as follows: When a suspicious sample shows high consistency in the prediction results of multiple sub-models, the prediction result is considered to have high credibility, and the original label is replaced by the model consensus result. The implementation method is as follows: in For the first Sub-model prediction samples The probability, By using a weighted average to calculate the corresponding confidence level of the prediction, a more robust bounding box prediction result can be obtained. ; Manual verification refers to situations where there is significant discrepancy in the prediction results of suspicious samples among the sub-models, specifically when the voting ratio of the largest category is below a threshold. It is then sent to the manual review stage.

8. The method according to claim 1, characterized in that, The noise perception loss function consists of two parts: classification loss and regression loss. The classification loss adopts the form of cross-entropy and is defined as follows: in The total number of categories, For the sample Category The probability, One-hot vectors labeled with categories; weight coefficients With label confidence Dynamic correlation, in the form of: in, To adjust the parameters, when When the value is low, the classification loss weights are automatically reduced to avoid excessive interference from noise labels on the model; For the regression loss, confidence-weighted methods are used for bounding box prediction. loss: in, To predict the bounding box, To label the bounding box, weights are introduced. This can reduce the impact of the sample on the regression task when the confidence level is low; Finally, the total loss function combines the classification loss and the regression loss, defining the overall training loss as: in, For batch size, This is a balancing coefficient used to control the relative importance of classification and regression tasks.

9. A defect detection device for power distribution network equipment based on multi-model collaboration, characterized in that, include: The system includes modules for dataset preparation, weather interference quantification, multi-expert integration, annotation consistency voting, adaptive re-annotation of suspicious samples, noise-aware loss function, and inference output. Each module works in sequence and in concert to achieve accurate detection of defects in power distribution network equipment, including: The dataset preparation module is used to acquire a large number of photos of power distribution network equipment defects under different weather interference conditions, and to construct a power distribution network defect detection dataset with noise annotation, so as to provide data support for subsequent model construction and training. The weather interference quantification module is used to calculate and store the standard grayscale standard deviation σ0 and standard contrast C0 as the benchmark values ​​for weather interference quantification based on the normal weather power distribution network photos provided by the dataset preparation module, so as to ensure the horizontal comparability of interference coefficients of different images; it is also used to calculate the grayscale standard deviation σ and contrast C of each power distribution network image to be processed, and substitute them into the formula ω=1-σ·C / (σ0·C0) to calculate the weather interference intensity coefficient ω. The closer the ω value is to 0, the stronger the weather interference is, and the farther away from 0, the weaker the interference is. The calculated ω value is then passed to the multi-expert integration module. The multi-expert integration module is used to build multiple sub-model frameworks based on YOLOv8. A Coordinate Attention layer is added to the feature extraction network of each sub-model to enhance sensitivity to defect location information. Cross-scale skip connections are added to the PAN-FPN structure to fuse high-level and low-level features. The output of each sub-model includes defect category, bounding box, and additional confidence score, which is calculated as the product of classification confidence score and bounding box IoU prediction value. It is also used to configure independent training conditions for each sub-model through three mechanisms: initialization parameter randomization, data augmentation strategy differentiation, and training hyperparameter fine-tuning. It receives the ω value transmitted by the weather interference quantification module and combines it with the sub-model training and inference process. The annotation consistency voting module is used to input the training set samples constructed by the dataset preparation module into the N sub-models that have been trained, collect the predicted class yij, predicted bounding box bij and confidence score cij of each sub-model corresponding to each sample; it is also used to calculate the class consistency probability P(yi) and the average bounding box IoUi of each sample, and then calculate the annotation confidence score Si, preset the balance constant α and the suspicious sample threshold τ, mark the suspicious samples according to the comparison result of Si and τ, and pass the marking result to the suspicious sample adaptive re-labeling module; The adaptive relabeling module for suspicious samples is used to receive suspicious samples marked by the labeling consistency voting module and the prediction results of the corresponding sub-models. If a suspicious sample shows high consistency in the prediction results of multiple sub-models, the original label is replaced by the model consensus result and the dataset is updated. If the prediction results of a suspicious sample differ greatly among the sub-models, the sample is sent to the manual review stage, and the labeling information is updated after the review is completed. The noise-perceived loss function module is used to construct a noise loss function consisting of classification loss and regression loss. By introducing the label confidence Si output by the label consistency voting module as an adjustment factor, the weights of classification loss and regression loss are dynamically adjusted to weaken the impact of low-confidence samples on model training and provide loss calculation support for the training of sub-models in the multi-expert integration module. The inference output module is used to receive the new power distribution network image to be detected, input it into the trained multi-model collaborative defect detection model, which is composed of a weather interference quantification module, a multi-expert integration module, a labeling consistency voting module, a suspicious sample adaptive relabeling module, and a noise perception loss function module, to perform inference and output the defect detection result, which includes the defect category, bounding box, and confidence level.

10. The apparatus according to claim 9, characterized in that, In the weather interference quantification module, the grayscale standard deviation σ is obtained by calculating the square root of the sum of the squares of the differences between the image pixel grayscale values ​​and the average grayscale values, and the contrast ratio C is obtained by calculating the difference between the maximum grayscale value and the minimum grayscale value in the image.

Images