Power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning

Abstract

The invention discloses a power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning, and relates to the technical field of intelligent inspection of power equipment, and the method comprises the following steps: collecting a visible light image of an insulator to construct a data set containing labels and unlabels, positioning and cutting an insulator string, normalizing the resolution, and obtaining a data set containing labels and unlabels; enhancing the annotated data, and retaining an original image of the unannotated data; constructing a dual-path feature extraction network based on ResNet-18, and generating antagonism enhancement features by matching a classifier and a U-Net structure adversarial sub-network; initializing model parameters and carrying out preheating training, carrying out dual-path evaluation on unlabeled samples, screening high-value samples and generating pseudo labels; generating high-value sample adversarial enhancement features, calculating multi-class loss, jointly optimizing and updating parameters, dynamically reducing a threshold value, and circularly training until convergence; and selecting an optimal model, solidifying and exporting the optimal model, integrating the optimal model to an inspection platform, marking suspected defects, pushing and rechecking the suspected defects, and accumulating newly added samples for incremental training.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent inspection of power equipment, and particularly relates to a power distribution network unmanned aerial vehicle (UAV) inspection defect identification method based on semi-supervised small sample learning. BACKGROUND

[0002] Insulator burst defect identification is a key task in UAV fine inspection, especially for transmission lines with voltage levels of 110 kV and above. Although such defects have a low occurrence rate, they are extremely harmful and are an important hidden danger that can cause line faults. The typical characteristics of such defects are cracks, defects, or even complete fragmentation on the surface or inside of porcelain insulators. These defects need to be distinguished from interference factors such as dirt, bird droppings, and normal textures. Accurate detection of such defects is of great significance for ensuring the safe and stable operation of the power grid.

[0003] In the field of UAV inspection image analysis based on deep learning, semi-supervised learning methods have become an important technical means for utilizing large amounts of unlabeled data and reducing dependence on expensive manual labeling. Existing technologies usually rely on the prediction confidence of the model for unlabeled samples, and set a threshold to filter high-confidence samples and assign pseudo-labels to expand the training set. However, these methods face a specific problem when dealing with extreme small sample and severe class imbalance scenarios such as insulator burst defects: in the initial defect labeling, the model's initial confidence judgment is unreliable, and the pseudo-labels generated based on a single confidence threshold contain a lot of noise. In particular, a large number of real defects are incorrectly labeled as "normal" because the model has not fully learned their characteristics. This type of false pseudo-labels misleads the model's learning direction, causing its decision boundary to deviate significantly from the true distribution. Not only does this fail to effectively utilize unlabeled data to improve the model's ability to recognize rare defects, but it also leads to a degradation or even complete failure of the model's performance. Therefore, to address the above problems, the following solutions are proposed. SUMMARY

[0004] To solve the above technical problems, the application is implemented by the following technical solutions:

[0005] The application is a power distribution network UAV inspection defect identification method based on semi-supervised small sample learning, comprising the following steps:

[0006] Step S1: Data preparation and preprocessing, collect visible light images of insulators to construct a dataset containing labeled and unlabeled data, locate and crop the insulator string and normalize the resolution, enhance the labeled data, and retain the original images of the unlabeled data;

[0007] Step S2: Model architecture construction and initialization, build a dual-path feature extraction network based on ResNet-18, and couple a classifier and a U-Net structure adversarial subnetwork to generate adversarial enhanced features;

[0008] Step S3: progressive pseudo-label screening and training cycle, initialize model parameters and preheat training, double-path evaluation on unlabeled samples, screen high-value samples and generate pseudo-labels;

[0009] Step S4: adversarial feature enhancement and joint optimization, generate high-value sample adversarial enhanced features, calculate multi-class loss and jointly optimize and update parameters, dynamically reduce threshold cycle training to convergence;

[0010] Step S5: model solidification and deployment application, select the optimal model and solidify and export, integrate into the inspection platform, mark suspected defects and push for review, accumulate new samples for incremental training.

[0011] Further, the step S1 includes the following steps:

[0012] Step S11: collect original data, obtain visible light images of insulators of 110kV lines in a specific area from a UAV inspection system, construct an original data set, which includes a labeled data set and an unlabeled data set; the labeled data set contains a small number of self-explosion defect samples and a small number of normal samples, the unlabeled data set only contains unlabeled image samples, and the number of samples in the labeled data set is much less than that in the unlabeled data set;

[0013] Step S12: defect region standardization cropping, use the insulator detection model to process all images in the labeled data set and the unlabeled data set, locate the insulator string in the image and crop it, ensure that each cropped input image mainly contains a single insulator string, and normalize the resolution of all cropped images to a uniform size;

[0014] Step S13: basic data enhancement, for each image in the labeled data set, apply random horizontal flipping, color jittering and slight rotation operations respectively, expand the number of labeled samples through these methods, and generate an enhanced labeled set; for all images in the unlabeled data set, do not perform any processing, and keep the original images for subsequent steps.

[0015] Further, the step S2 includes the following steps:

[0016] Step S21: construct a dual-path information perception feature extraction network, which uses ResNet-18 as the backbone network, removes the fully connected layer after the global average pooling layer, and determines the output of the last residual block of the network as the feature map;

[0017] Step S211: a classification uncertainty evaluation branch is constructed, and a lightweight convolution module is connected on the branch, which sequentially comprises a 1x1 convolution layer, a ReLU activation function and a 1x1 convolution layer, the channel number of the feature map is reduced to the class number through the module, a logits map of a corresponding dimension is output, and a spatial global average pooling operation is performed to obtain a classification prediction vector;

[0018] Step S212: a feature space consistency evaluation branch is constructed, and a non-parametric local feature aggregation module is connected on the branch, for each spatial position feature vector in the feature map, cosine similarity of the feature vector with all other position feature vectors in the feature map is calculated, a plurality of feature vectors with the highest similarity are selected for average calculation to obtain an aggregated feature of the position, and a 1x1 convolution layer is used to map the aggregated feature to a space with the same dimension as the original feature map to output a consistency enhanced feature map;

[0019] Step S22: a classifier is constructed, which is a full connection layer with a corresponding class number output unit, and the input of the classifier is the prediction vector finally output by the classification uncertainty evaluation branch, the prediction vector needs to be expanded first, and then a full connection layer is used for dimension reduction processing, and the processed feature vector is used as the input of the classifier;

[0020] Step S23: an adversarial feature enhancement subnetwork is constructed, which is a small U-Net structure generator, the input data includes two parts, which are the consistency enhanced feature map output by the feature space consistency evaluation branch and the shallow layer feature of the original input image after encoding corresponding to the consistency enhanced feature map, and the output is a feature layer disturbance field with the same spatial size as the consistency enhanced feature map.

[0021] Further, the step S3 comprises the following steps:

[0022] Step S31: initialize the model parameters, use the pre-trained weights on the ImageNet dataset to initialize the parameters of the backbone of the dual-path information perception feature extraction network; the parameters of the two parallel branches of the network, the classifier and the adversarial feature enhancement subnetwork are initialized in a random manner;

[0023] Step S32: first preheating stage training, only using the enhanced labeled dataset, training the dual-path information perception feature extraction network, the classification uncertainty evaluation branch and the classifier, freezing the parameters of the feature space consistency evaluation branch and the adversarial feature enhancement subnetwork during the training; using the standard cross-entropy loss as the supervision loss function, training for a specified number of rounds, so that the model has a preliminary classification ability;

[0024] Step S33: double-path information perception evaluation is performed on each unlabeled sample, specifically including the following sub-steps:

[0025] Step S331: forward propagation calculation, the unlabeled sample is input into the double-path information perception feature extraction network to obtain a classification prediction vector and a consistency enhanced feature map corresponding to the sample respectively;

[0026] Step S332: calculating the classification uncertainty score, the confidence degree of the model in the classification prediction of the unlabeled sample is quantified through information entropy related calculation, and the higher the uncertainty, the larger the score;

[0027] Step S333: calculating the feature space consistency score, the consistency of the features in the sample is evaluated by calculating the average cosine similarity between all non-repeated spatial position feature vectors in the consistency enhanced feature map, and the higher the similarity, the higher the consistency score;

[0028] Step S334: calculating the comprehensive information quantity score, two hyperparameters are introduced to generate a formula positively correlating low uncertainty with high consistency, and the classification uncertainty score and the feature space consistency score are integrated to obtain the comprehensive information quantity score of the sample;

[0029] Step S34: constructing a gradual trusted pseudo-label screening pool, setting a dynamic threshold, the threshold is linearly decayed from a higher value to a lower value with the training round; for the unlabeled samples in the current batch, the samples satisfying both the condition that the comprehensive information quantity score is greater than the dynamic threshold and the condition that the maximum probability of the classification prediction is greater than another fixed high threshold are classified into a high-value sample subset;

[0030] Step S35: generating pseudo-labels for the high-value sample subset, for each high-value sample, the class with the maximum probability in the classification prediction vector is obtained, and the class is taken as the pseudo-label of the sample.

[0031] Further, the step S4 includes the following steps:

[0032] Step S41: generating an adversarial disturbance, only for the high-value sample subset, the original image and the consistency enhanced feature map corresponding to each high-value sample are input into an adversarial feature enhancement subnetwork to generate a feature level disturbance field corresponding to the sample, and then the disturbance field is superimposed on the original feature map at a set proportion to obtain an adversarial enhanced feature of the sample;

[0033] Step S42: adversarial feature classification and loss calculation, the adversarial enhanced feature of each high-value sample is sent into the classification uncertainty evaluation branch and the classifier to obtain the adversarial prediction result of the sample; a cross-entropy loss function is used to calculate the pseudo-label consistency loss, and the loss is used to force the prediction made by the adversarial enhanced feature to be consistent with the pseudo-label of the original high-value sample;

[0034] Step S43: Calculate the supervised loss of the labeled data, input the labeled data in the current training batch into the dual-path information perception feature extraction network and the classifier to obtain the classification prediction result, and calculate the supervised loss of the labeled data through the cross-entropy loss function;

[0035] Step S44: Calculate the feature smoothing regularization loss, use the Frobenius norm to measure the overall difference between the original feature map and the adversarial enhanced feature map, and the difference value is the feature smoothing regularization loss, which is used to constrain the adversarial enhanced feature from deviating too far from the original feature;

[0036] Step S45: Joint optimization and parameter update, define the total loss function as the weighted sum of the supervised loss, the adversarial pseudo label consistency loss and the feature smoothing regularization loss, and control the contribution weights of the adversarial pseudo label consistency loss and the feature smoothing regularization loss through two hyperparameters respectively; update all parameters of the dual-path information perception feature extraction network, the classifier and the adversarial feature enhancement subnetwork simultaneously through the back propagation algorithm;

[0037] Step S46: Gradual threshold update and cycle, after completing the current round of training, reduce the dynamic threshold according to the preset plan, return to step S33 to start a new round of training, gradually include more information-rich and gradually increasing reliable unlabeled samples in the training process, until the model performance converges on the validation set or reaches the preset training round.

[0038] Further, the step S5 comprises the following steps:

[0039] Step S51: Model selection, after training, select the model checkpoint with the highest F1 score in identifying insulator self-explosion defects based on the validation set as the evaluation basis, and determine it as the final model;

[0040] Step S52: Model solidification, save the trained dual-path information perception feature extraction network, classification uncertainty evaluation branch and classifier, and export them in a format that can be loaded by the inference engine; the adversarial feature enhancement subnetwork only plays a role in the training stage and does not need to be loaded in the deployment stage;

[0041] Step S53: Integrated deployment, integrate the solidified model into the unmanned aerial vehicle inspection data processing platform of the power grid company; when a new insulator image is input into the platform, the system first performs insulator string positioning and cutting according to the method of step S12, and then inputs the cut image into the solidified model, which outputs the probability that the image belongs to a self-explosion defect; set an application threshold, when the output probability exceeds the threshold, the system automatically marks the image as a suspected defect and pushes it to the manual review interface;

[0042] Step S54: manual review and model incremental data preparation, the suspected defect images marked by the system are finally confirmed by manual; the new defect samples and their annotation information confirmed by manual are stored in the incremental database; when the samples in the incremental database accumulate to a certain number, the newly added annotation data and the original annotation data are combined and used to start a new round of model incremental training to continuously improve the recognition performance of the model.

[0043] The present application has the following advantages:

[0044] 1. The present application effectively improves the utilization efficiency and security of massive unlabeled data under the condition of few labeled samples by designing a double-path information perception and progressive screening mechanism. This method avoids the extensive strategy of simply assigning pseudo-labels to all high-confidence samples, and instead evaluates the information value and reliability of samples from two dimensions of classification uncertainty and feature consistency, and dynamically adjusts the screening threshold. This process can include more high-quality unlabeled samples and their generated pseudo-labels in the training cycle, reducing the risk of model performance deterioration caused by the introduction of false pseudo-labels, thereby achieving stable improvement of model performance while reducing dependence on a large amount of manual annotation.

[0045] 2. The present application effectively enhances the model's generalization recognition ability for defect features, especially for defect patterns similar to normal samples or with fuzzy boundaries, by introducing a targeted adversarial feature enhancement mechanism. This mechanism does not perform general data enhancement, but focuses on high-value samples selected in the early stage to generate challenging adversarial samples at the feature level. This targeted consolidation forces the main classification network to learn more discriminative feature representations, especially strengthening the classification ability near the key decision boundary, so that the model performs better in real-world scenarios with diverse shapes and complex backgrounds.

[0046] 3. The method proposed in the present application can improve the learning effect and reliability of the final output of the model in the case of severe class imbalance; for the case of few defect samples, the present method tends to select those few-class (defect) potential samples with clear features and consistent model judgments by comprehensive information evaluation, and focuses on optimizing the feature space structure around these samples through adversarial enhancement strategy. This helps to balance the attention of the model to different classes and suppresses the tendency of the model to excessively favor the majority class due to the large number of samples, thereby improving the recall rate of rare defect classes and overall recognition accuracy, making the system output more reliable.

[0047] Of course, any product implementing the present application does not necessarily need to achieve all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to make the technical solutions of the embodiments of the present application clearer, the drawings needed for the description of the embodiments will be briefly introduced as follows. Obviously, the drawings described below are only some of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0049] Figure 1 The flowchart of the power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning of the present application is shown. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, and not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0051] Please refer to Figure 1 The power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning of the present application includes the following steps:

[0052] Step S1: data preparation and preprocessing, collecting original data to construct an original data set containing a labeled data set and an unlabeled data set, positioning and cropping the insulator string in all images and normalizing the resolution, performing basic data enhancement on the labeled data set, and retaining the original images of the unlabeled data set;

[0053] Step S2: model architecture construction and initialization, a dual-path information perception feature extraction network is constructed, which is based on ResNet-18 backbone and removes the global average pooling layer and the full connection layer after it, and the output of the last residual block is taken as the feature map, and then two parallel branches are set; wherein the classification uncertainty evaluation branch reduces the channel number to the class number through a lightweight convolution module (including two 1x1 convolution layers and ReLU activation functions), outputs the logits map of the corresponding dimension, and then obtains the classification prediction vector through spatial global average pooling; the feature space consistency evaluation branch passes through a non-parametric local feature aggregation module, calculates the cosine similarity between the feature vector at each spatial position in the feature map and all other position feature vectors, selects a number of most similar feature vectors, averages them to obtain aggregated features, and then maps the aggregated features to the same dimension as the original feature map through a 1x1 convolution layer to output a consistency-enhanced feature map; meanwhile, a classifier and an adversarial feature enhancement subnetwork are constructed, the classifier is a full connection layer with output units corresponding to the number of classes, and the input is the feature vector obtained by unfolding and dimensionality reduction of the prediction vector output by the classification branch; the adversarial feature enhancement subnetwork is a small U-Net structure generator, the input is the consistency-enhanced feature map and the corresponding original input image encoded shallow features, and the output is a feature level disturbance field with the same spatial size as the consistency-enhanced feature map, which is superimposed on the original feature according to a certain proportion to generate an adversarial enhanced feature;

[0054] Step S3: information perception-based progressive pseudo-label screening and training cycle, initialize model parameters, preheat training using enhanced labeled data, then perform dual-path information perception evaluation on each unlabeled sample, calculate the classification uncertainty score, feature space consistency score and comprehensive information quantity score in turn, construct a progressive trusted pseudo-label screening pool, and generate pseudo-labels for the high-value sample subset screened out; wherein the classification uncertainty score quantifies the confidence of the model classification prediction through information entropy related calculation, the higher the uncertainty, the higher the score; the feature space consistency score evaluates the consistency of the internal features of the sample by calculating the average cosine similarity between all non-repeated spatial position feature vectors in the feature map; the comprehensive information quantity score integrates the first two scores to construct an expression that positively correlates low uncertainty (high confidence) and high consistency, and preferentially selects "high-value" samples with high model prediction confidence and strong feature internal consistency;

[0055] Step S4: Adversarial feature enhancement and joint optimization, generate adversarial perturbations and adversarial enhanced features for the high-value sample subset, input them into the classification branch and the classifier to obtain adversarial predictions and calculate the pseudo-label consistency loss, while calculating the supervision loss of the labeled data and the feature smoothing regularization loss, jointly optimize through the total loss function (the weighted sum of the supervision loss, the adversarial pseudo-label consistency loss and the feature smoothing regularization loss), update all related parameters of the feature extraction network, the classifier and the adversarial subnetwork through the back propagation algorithm; then reduce the dynamic threshold according to the predetermined plan, return to the dual-path information perception evaluation step to start a new round of training, gradually include more qualified unlabeled samples in the training, until the model performance converges on the validation set or reaches the preset number of rounds; wherein the pseudo-label consistency loss is realized by cross-entropy calculation, which forces the prediction results of the adversarial enhanced features to be consistent with the pseudo-labels of the original high-value samples, improves the anti-interference ability of the model to small perturbations in the feature space, and consolidates the decision boundary; the feature smoothing regularization loss is to measure the overall difference between the original feature map and the adversarial enhanced feature map, to constrain the adversarial enhanced features from deviating too far from the original features, and to avoid distortion of the feature space;

[0056] Step S5: Model solidification and deployment application, select the model checkpoint with the highest F1 score in identifying insulator self-explosion defects on the validation set as the final model, solidify the dual-path information perception feature extraction network, the classification uncertainty evaluation branch and the classifier, and export them in a format that can be loaded by the inference engine, the adversarial subnetwork is only used in the training stage and does not need to be loaded in the deployment stage; integrate the solidified model into the unmanned aerial vehicle inspection data processing platform of the power grid company, after inputting a new inspection insulator image, first perform insulator string positioning and cropping, then input the solidified model to obtain the probability that it belongs to a self-explosion defect; set an application threshold, when the probability exceeds the threshold, the system automatically marks it as a suspected defect and pushes it to the manual review interface; manually confirm the suspected defect, store the confirmed new defect sample and its label in the incremental database, and accumulate to a certain number, then combine with the original labeled data to start a new round of model incremental training and continuously improve the model performance;

[0057] Step S1 includes the following steps:

[0058] Step S11: Collect original data; obtain visible light images of insulators of a specific 110kV line from an unmanned aerial vehicle inspection system; build an original data set , wherein the labeled data set contains a small number of self-explosion defect samples and a small number of normal samples , the unlabeled data set , ;

[0059] Step S12: Defect region standardization cropping; locate and crop insulator string in each image in and using insulator detection model (e.g. pre-trained YOLOv5), ensuring that each input image mainly contains a single insulator string, and normalizing the resolution of all cropped images to (e.g. );

[0060] Step S13: Basic data augmentation; for each image in the labeled dataset , apply random horizontal flipping, color jittering and slight rotation (angle limited within ±10°) to expand the sample, generating the augmented labeled dataset ; for the unlabeled dataset , keep the original images for later processing.

[0061] Step S2 includes the following steps:

[0062] Step S21: Constructing dual-path information-aware feature extraction network ; the network is based on ResNet-18 backbone, removing the fully connected layer after the global average pooling layer; the output feature map of the last residual block is denoted as ; after that, two parallel branches are constructed:

[0063] Step S211: Classification uncertainty evaluation branch ; on this branch, a lightweight convolution module (containing one convolution layer, ReLU activation function and one convolution layer) is connected, reducing the channel number to the class number (here , i.e. defect and normal), outputting a logits map of , and then obtaining the classification prediction vector through spatial global average pooling;

[0064] Step S212: Feature space consistency evaluation branch ; on this branch, a non-parametric local feature aggregation module is connected; specifically, for the feature vector of each spatial position in the feature map , calculate the cosine similarity with all other position feature vectors in the feature map , select the most similar feature vectors (excluding itself) for averaging, obtaining the aggregated feature of the position ; this operation aims to enhance the intra-class feature consistency; then pass through a The convolutional layer maps the aggregated features to a space of the same dimension, outputting consistency-enhanced feature maps ; ;

[0065] Step S22: Constructing the classifier ; it is a fully connected layer with output units, taking as input the prediction vector output by the classification branch in step S211 after flattening and dimensionality reduction by a fully connected layer;

[0066] Step S23: Constructing the adversarial feature enhancer network ; it is a small U-Net structured generator, taking as input the consistency-enhanced feature maps from the branch and the corresponding original input image after encoding, outputting a feature-level perturbation field of the same spatial dimensions as ; the perturbation field is added to the original features , generating adversarial enhanced features , where is a small scalar controlling the perturbation strength.

[0067] Step S3 includes the following steps:

[0068] Step S31: Initializing model parameters; using pre-trained weights on ImageNet to initialize the backbone of the feature extraction network , its parallel branches , and the classifier , adversarial subnetwork are randomly initialized;

[0069] Step S32: First preheating stage training; only using enhanced labeled data to train the network (freezing and ); using standard cross-entropy loss for supervised training for a certain number of epochs (e.g., 50 rounds), so that the model has preliminary classification ability;

[0070] Step S33: For each unlabeled sample , perform dual-path information perception evaluation;

[0071] Step S331: Forward propagation calculation; input into the network , respectively, to obtain the classification prediction vector​ And consistency enhancement feature map ;

[0072] Step S332: Calculate the classification uncertainty score; this score quantifies the confidence level of the model in predicting the sample classification. The higher the uncertainty (the greater the entropy), the greater the potential noise risk of the false label. This calculation is achieved using the information entropy formula, as follows: In the formula, For the first Classification uncertainty score of an unlabeled sample; The total number of categories (2 in this case); for The The component represents the model's prediction that the sample belongs to the first element. The probability of a class;

[0073] Step S333: Calculate the feature space consistency score; this score aims to evaluate the internal consistency of samples from the perspective of feature representation; for a sample with clear and explicit feature responses, the feature vectors at different spatial locations in its deep feature map should exhibit high similarity; this score is obtained by calculating the average cosine similarity between feature vector pairs at all spatial locations within the feature map, as shown in the following formula: In the formula, For the first The feature space consistency score of each sample; and These are the height and width of the feature map, respectively; and For spatial location index on the feature map; For feature map In position The eigenvector at that location; This is the normalization coefficient, used to calculate the average value among all non-repeating position pairs;

[0074] Step S334: Calculate the comprehensive information content score; to integrate information from the two dimensions of classification uncertainty and feature consistency, a comprehensive scoring function is designed; this function aims to prioritize "high-value" samples with high model prediction confidence and strong internal feature consistency; a formula is used to construct a formula that positively correlates low uncertainty (high confidence) with high consistency, as follows: In the formula, For the first The overall information content score of each sample; This is the factor used to normalize the uncertainty score; It is a positive hyperparameter, called the consistency sensitivity coefficient, which is used to control the amplification intensity of the consistency score; is an hyper-parameter, called consistency reference value, used to adjust the offset of the exponential function;

[0075] Step S34: Constructing the progressive trusted pseudo-label filtering pool; set a dynamic threshold , which linearly decays from a higher value (e.g. 0.95) to a lower value (e.g. 0.75) with the training round ; for the unlabeled samples in the current batch, select those whose comprehensive information quantity scores are greater than the threshold , and whose classification prediction maximum probabilities are greater than another fixed high threshold (e.g. 0.9), and put them into the high-value sample subset ;

[0076] Step S35: generating pseudo labels for the high-value sample subset ; for , take the class corresponding to the maximum probability in its prediction vector as the pseudo label .

[0077] Step S4 includes the following steps:

[0078] Step S41: adversarial perturbation generation; only for the high-value sample subset filtered out in step S34 , input its corresponding original image and consistency enhanced feature map into the adversarial subnetwork , generate a feature perturbation field , and then get the adversarial enhanced feature ;

[0079] Step S42: adversarial feature classification and loss calculation; send the adversarial enhanced feature into the classification branch and the classifier , get its adversarial prediction ; calculate the pseudo label consistency loss; this loss forces the prediction made by the adversarial enhanced feature to be consistent with the pseudo label of the original high-value sample; its purpose is to make the model insensitive to small perturbations around the high-value sample in the feature space through such anti-interference training, thereby consolidating the decision boundary and improving the model robustness; this loss is realized through cross-entropy calculation, and the formula is as follows: wherein, is the adversarial pseudo label consistency loss; is the number of samples in the high-value sample subset ; is the summation of all samples in the subset; is the sample the adversarial enhanced predicted probability distribution; the pseudo label for the sample; the cross-entropy loss function;

[0080] Step S43: calculating the supervised loss of the labeled data; in the current training batch, both the labeled data and the unlabeled data are included ;

[0081] Step S44: calculating the feature smoothing regularization loss; in order to prevent the generated perturbation of the adversarial subnetwork from distorting the original features too much, leading to distortion of the feature space, a regularization constraint is introduced; this loss constraint requires that the adversarial enhanced features should not deviate too far from the original features , ensuring that the feature enhancement is within a reasonable and smooth range; this loss uses the Frobenius norm to measure the overall difference between two feature maps, and the formula is as follows: wherein, is the feature smoothing regularization loss; is the original feature map; is the adversarial enhanced feature map;

[0082] Step S45: joint optimization and parameter update; the total loss function is defined as the weighted sum of the supervised loss, the adversarial pseudo label consistency loss, and the feature smoothing regularization loss, and the formula is as follows: wherein, is the total loss function; is the supervised loss of the labeled data; is the pseudo label loss weight factor, which is a non-negative hyperparameter used to balance the contribution of the pseudo label loss; is a non-negative hyperparameter, called the smoothing regularization weight factor, used to balance the contribution of the feature smoothing regularization loss;

[0083] the parameters of the feature extraction network (including ), the classifier , and the adversarial subnetwork are updated simultaneously through the backpropagation algorithm ;

[0084] Step S46: gradual threshold update and cycle; after completing the current round of training, the dynamic threshold is lowered according to the predetermined plan ; return to step S33, start a new round of training, so gradually more information-rich, increasing the credibility of the unlabeled sample into the training process; this cycle continues until the model on the validation set (from a small amount of labeled data) performance convergence or reach the preset round.

[0085] Step S5 includes the following steps:

[0086] Step S51: model selection; after training, select the model checkpoint with the highest F1 score for identifying insulator self-explosion defects on the validation set as the final model;

[0087] Step S52: model solidification; save the trained double-path information perception feature extraction network , classification uncertainty evaluation branch and classifier Export to a format that can be loaded by the inference engine (such as ONNX); the adversarial subnetwork Only used in the training stage, no need to load in the deployment stage;

[0088] Step S53: integrated deployment; integrate the solidified model into the power grid company's unmanned aerial vehicle inspection data processing platform; when a new insulator image is input, the system first performs insulator string positioning and cutting (step S12), then inputs the cut image into the solidified model, and the model outputs the probability that it belongs to a self-explosion defect; Set an application threshold, when the probability exceeds the threshold, the system automatically marks the image as a suspected defect and pushes it to the manual review interface;

[0089] Step S54: manual review and model incremental data preparation; confirm the suspected defects marked by the system through manual review; confirmed new defect samples (and their labels) will be stored in an incremental database; when a certain amount is accumulated, these newly labeled data can be combined with the original Merge and use to start a new round of model incremental training, so as to continuously improve the performance of the model.

[0090] Working principle:

[0091] The core of the scheme adopts an information perception-based progressive pseudo-label screening and adversarial enhancement mechanism. The mechanism constructs a parallel evaluation classification uncertainty branch and a feature space consistency branch, dynamically screens high-value unlabeled samples, and generates targeted feature disturbance to consolidate the model's decision boundary in the fuzzy area. The method uses the visual feature response of the insulator sample itself in the unmanned aerial vehicle inspection image to quantify the model's prediction reliability and internal feature consistency for the unlabeled sample, and builds a progressive pseudo-label credibility pool, thereby improving the recognition generalization ability of insulator self-explosion defects in the initial defect sample extreme scarcity scenario. In the above scheme, the output of the feature extraction network is directed to two parallel paths of classification prediction and local feature aggregation, and an adversarial subnetwork is introduced to generate feature-level adversarial disturbance for the screened high-value samples, so that the model can use the prediction confidence of the sample and the structural information of the feature space at the same time, avoiding the propagation of pseudo-label noise while strengthening the feature learning of difficult example samples.

[0092] In the method, the classification uncertainty score and the feature space consistency score are calculated based on the soft prediction probability distribution of the model and the internal similarity of the feature map, respectively. The uncertainty score quantifies the ambiguity of the model's sample class judgment by prediction entropy, and the consistency score evaluates the coordination degree of the model's internal feature response by calculating the average similarity between all spatial position feature vectors in the feature map. The two scores are fused by a designed comprehensive information amount formula to quantify the overall training value of the unlabeled sample. The core of the progressive screening is to combine the dynamic decay threshold with the innovative double-path evaluation results, and through the dual filtering of the comprehensive information amount threshold and the prediction confidence threshold, a gradually expanding high-value sample subset is constructed. Specifically, the screening mechanism forces the model to preferentially learn samples that are highly determined by the model itself and have consistent internal feature responses: if an unlabeled sample is classified into a certain class by the model with high probability, and its activated features show high consistency in space, then the sample is more likely to be assigned a correct pseudo-label and added to the training. This mechanism can effectively utilize the high-confidence information in the massive unlabeled data, without the need for a large amount of manual labeling, but only relies on a small number of initial labels;

[0093] Through the above screening and enhancement mechanism, the method can be used for identifying the self-explosion defect of the 110kV power transmission line ceramic insulator string. In the initial stage, the self-explosion defect sample can only have 3-5 pictures, and it is difficult to distinguish from the normal insulator or the insulator containing stains and bird droppings under specific illumination and angle. The dual-path information perception module can identify the reliable sample in the unlabeled data which is adjacent to the labeled defect sample in the feature space and is cognitively unified in the model by parallel evaluation of the predicted certainty and the consistency of the features. The adversarial enhancer network focuses on these screened high-value samples, generates a small perturbation in the feature space, and aims to "challenge" the decision robustness of the main classifier for these key samples, so as to enlarge the interval between the defect class and the easily confused normal class in the feature space in the subsequent joint optimization, and make the decision boundary in the fuzzy area more clear.

[0094] Specifically, for an unlabeled unmanned aerial vehicle image containing a suspected self-explosion insulator, the system first extracts the features of the image and calculates the comprehensive information score thereof. If the score is high and the probability of the model predicting the image as "defect" is extremely high, the image is selected into the high-value pool and is given a "defect" pseudo-label. Then, the adversarial sub-network generates a small perturbation field for the feature map of the image. The perturbation is added to the original features to form an adversarial feature version. The training target requires the main classifier not only to make a correct pseudo-label prediction for the original features, but also to maintain the same prediction for the adversarial features subjected to the targeted perturbation. This process is equivalent to locally "strengthening" the high-value sample points in the feature space, so that the defect class feature representation learned by the model is not only effective for the original sample, but also robust to the small changes of the sample in the feature space neighborhood, thereby significantly improving the recognition ability of the model for the defect samples with slightly changed morphology and illumination in the real scene.

[0095] The preprocessed insulator image data and the optimized dual-path perception model are finally used for defect recognition of new inspection images, outputting a defect probability to evaluate the self-explosion risk, thereby providing accurate early warning for power grid operation and maintenance. The entire process is realized through an automatic and iterative pseudo-label screening and adversarial enhancement cycle, ensuring that the recognition model can fully utilize the internal structural information of the unlabeled data under the condition of extremely small sample, and obtain high precision and strong generalization performance.

[0096] In the description of the present specification, the description referring to the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0097] The preferred embodiments of the application disclosed above are only to facilitate the elucidation of the application. The preferred embodiments do not describe all the details of the application and limit the application to the specific embodiments described. Obviously, many modifications and variations can be made in light of the teachings above. The description is chosen and described in order to best explain the principles of the application and its practical application to thereby enable others skilled in the art to best utilize the application and get the best results from the application. The application is only limited by the claims and their full scope and equivalents.

Claims

1. A power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning, characterized in that: The method comprises the following steps: Step S1: data preparation and preprocessing, collecting insulator visible light images to construct a data set containing labeled and unlabeled data, locating and cropping insulator strings and normalizing resolution, enhancing labeled data, and retaining original images of unlabeled data; Step S2: model architecture construction and initialization, constructing a double-path feature extraction network based on ResNet-18, matching a classifier and a U-Net structure of an adversarial subnetwork to generate adversarial enhanced features; Step S3: progressive pseudo-label screening and training cycle, initializing model parameters and preheating training, double-path evaluation of unlabeled samples, screening high-value samples and generating pseudo-labels; Step S4: adversarial feature enhancement and joint optimization, generating adversarial enhanced features of high-value samples, calculating multi-class loss and updating parameters through joint optimization, and dynamically reducing threshold to train until convergence; Step S5: model solidification and deployment application, selecting the optimal model and solidifying and exporting it, integrating it into the inspection platform, marking suspected defects for review, and accumulating new samples for incremental training.

2. The power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning according to claim 1, characterized in that, The step S1 comprises the following steps: Step S11: collecting original data, obtaining insulator visible light images of 110kV lines in a specific area from a UAV inspection system, and constructing an original data set containing labeled and unlabeled data sets; the labeled data set contains a small number of self-explosion defect samples and a small number of normal samples, and the unlabeled data set contains only unlabeled image samples, and the number of samples in the labeled data set is much smaller than that in the unlabeled data set; Step S12: standardizing and cropping the defect area, using an insulator detection model to process all images in the labeled and unlabeled data sets, locating and cropping the insulator strings in the images to ensure that each cropped input image mainly contains a single insulator string, and normalizing the resolution of all cropped images to a uniform size; Step S13: basic data enhancement, for each image in the labeled data set, respectively applying random horizontal flipping, color jittering and slight rotation operations to expand the number of labeled samples and generate an enhanced labeled set; for all images in the unlabeled data set, no processing is performed, and the original images are retained for subsequent steps.

3. The power distribution grid drone inspection defect identification method based on semi-supervised small sample learning according to claim 1, characterized in that, The step S2 comprises the following steps: Step S21: constructing a double-path information perception feature extraction network, which uses ResNet-18 as the backbone network, removes the fully connected layer after the global average pooling layer, and determines the output of the last residual block of the network as the feature map; Step S211: constructing a classification uncertainty evaluation branch, connecting a lightweight convolution module to the branch, which contains a 1x1 convolution layer, a ReLU activation function and a 1x1 convolution layer in sequence, reducing the channel number of the feature map to the number of classes through the module, outputting logits corresponding to the dimension, and obtaining a classification prediction vector through spatial global average pooling operation; Step S212: a feature space consistency evaluation branch is constructed, a non-parametric local feature aggregation module is connected on the branch, for each spatial position feature vector in the feature map, the cosine similarity between the feature vector and all other position feature vectors in the feature map is calculated, a number of feature vectors with the highest similarity are selected for average calculation to obtain the aggregated feature of the position, and the aggregated feature is mapped to the same dimension space as the original feature map through a 1*1 convolution layer to output a consistency enhanced feature map; Step S22: a classifier is constructed, which is a full connection layer with an output unit corresponding to the number of categories, and the input of the classifier is the prediction vector finally output by the classification uncertainty evaluation branch. The prediction vector needs to be expanded first and then processed by a full connection layer for dimension reduction. The processed feature vector is used as the input of the classifier; Step S23: an adversarial feature enhancement subnetwork is constructed, which is a generator with a small U-Net structure. The input data includes two parts, which are the consistency enhanced feature map output by the feature space consistency evaluation branch and the shallow layer feature of the original input image after encoding corresponding to the consistency enhanced feature map. The output is a feature layer disturbance field with the same spatial size as the consistency enhanced feature map.

4. The power distribution network unmanned aerial vehicle inspection defect identification method based on semi-supervised small sample learning according to claim 1, characterized in that, The step S3 includes the following steps: Step S31: initialize the model parameters, use the pre-trained weights on the ImageNet dataset to initialize the parameters of the backbone of the double-path information perception feature extraction network; the parameters of the two parallel branches of the network, the classifier and the adversarial feature enhancement subnetwork are initialized in a random manner; Step S32: first preheating stage training, only using the enhanced labeled dataset, training the double-path information perception feature extraction network, the classification uncertainty evaluation branch and the classifier, freezing the parameters of the feature space consistency evaluation branch and the adversarial feature enhancement subnetwork during the training; using the standard cross-entropy loss as the supervision loss function, training for a specified number of rounds to make the model have preliminary classification ability; Step S33: double-path information perception evaluation is performed on each unlabeled sample, which includes the following sub-steps: Step S331: forward propagation calculation, inputting the unlabeled sample into the double-path information perception feature extraction network to obtain the classification prediction vector and the consistency enhanced feature map corresponding to the sample respectively; Step S332: calculate the classification uncertainty score, which quantifies the confidence degree of the model in the classification prediction of the unlabeled sample through information entropy related calculation method. The higher the uncertainty, the larger the score; Step S333: calculate the feature space consistency score, which evaluates the consistency of the internal features of the sample by calculating the average cosine similarity between all non-repeated spatial position feature vectors in the consistency enhanced feature map. The higher the similarity, the higher the consistency score; Step S334: calculate the comprehensive information quantity score, introduce two hyperparameters to generate a formula that positively correlates low uncertainty with high consistency, integrate the classification uncertainty score and the feature space consistency score to obtain the comprehensive information quantity score of the sample; Step S34: Constructing a progressive trusted pseudo-label screening pool, setting a dynamic threshold that linearly decays from a higher value to a lower value with training rounds; for the unlabeled samples in the current batch, the samples that simultaneously satisfy two conditions of the comprehensive information quantity score being greater than the dynamic threshold and the classification prediction maximum probability being greater than another fixed high threshold are classified into a high-value sample subset; Step S35: Generating pseudo-labels for the high-value sample subset, for each high-value sample, obtaining the class with the maximum probability in the classification prediction vector, and taking the class as the pseudo-label of the sample.

5. The power distribution grid drone inspection defect identification method based on semi-supervised few-shot learning according to claim 1, characterized in that, The step S4 includes the following steps: Step S41: Adversarial disturbance generation, only for the high-value sample subset, inputting the original image and the consistent enhanced feature map corresponding to each high-value sample into the adversarial feature enhancement subnetwork to generate a feature layer disturbance field corresponding to the sample, and then superimposing the disturbance field on the original feature map at a set proportion to obtain the adversarial enhanced feature of the sample; Step S42: Adversarial feature classification and loss calculation, inputting the adversarial enhanced feature of each high-value sample into the classification uncertainty evaluation branch and the classifier to obtain the adversarial prediction result of the sample; using a cross-entropy loss function to calculate the pseudo-label consistency loss, which is used to force the prediction made by the adversarial enhanced feature to be consistent with the pseudo-label of the original high-value sample; Step S43: Calculating the supervision loss of the labeled data, inputting the labeled data in the current training batch into the dual-path information perception feature extraction network and the classifier to obtain the classification prediction result, and calculating the supervision loss of the labeled data through a cross-entropy loss function; Step S44: Calculating the feature smoothing regularization loss, using the Frobenius norm to measure the overall difference between the original feature map and the adversarial enhanced feature map, and the difference value is the feature smoothing regularization loss, which is used to constrain the adversarial enhanced feature from deviating too far from the original feature; Step S45: Joint optimization and parameter update, defining the total loss function as the weighted sum of the supervision loss, the adversarial pseudo-label consistency loss, and the feature smoothing regularization loss, where two hyperparameters are used to control the contribution weights of the adversarial pseudo-label consistency loss and the feature smoothing regularization loss, respectively; all parameters of the dual-path information perception feature extraction network, the classifier, and the adversarial feature enhancement subnetwork are updated simultaneously through the back propagation algorithm; Step S46: Progressive threshold update and cycle, after completing the current round of training, the dynamic threshold is reduced according to the preset plan, and the step S33 is returned to start a new round of training, gradually including more unlabeled samples with rich information quantity and increasing trustworthiness in the training process, until the model performance converges on the validation set or reaches the preset training rounds.

6. The power distribution grid drone inspection defect identification method based on semi-supervised few-shot learning according to claim 1, characterized in that, The step S5 includes the following steps: Step S51: Model selection, after training, using the validation set as the evaluation basis to select the model checkpoint with the highest F1 score in identifying insulator self-explosion defects, and determining it as the final model; Step S52: model solidification, save the trained double-path information perception feature extraction network, classification uncertainty evaluation branch and classifier, and export them into a format loadable by the inference engine; the adversarial feature enhancement sub-network only plays a role in the training stage and does not need to be loaded in the deployment stage; Step S53: integrated deployment, integrate the solidified model into the unmanned aerial vehicle inspection data processing platform of the power grid company; when a new inspection insulator image is input into the platform, the system first performs insulator string positioning and cutting according to the manner of step S12, and then inputs the cut image into the solidified model, and the model outputs the probability that the image belongs to a self-explosion defect; set an application threshold, when the output probability exceeds the threshold, the system automatically marks the image as a suspected defect, and pushes it to the manual review interface; Step S54: manual review and model incremental data preparation, manually confirm the suspected defect images marked by the system; store the manually confirmed new defect samples and their annotation information in the incremental database; when the samples in the incremental database accumulate to a certain number, combine these newly added annotation data with the original annotation data to start a new round of model incremental training, and continuously improve the recognition performance of the model.

Images