A method and system for detecting defects in power equipment

Abstract

The present invention discloses a method and system for detecting defects in power equipment, belonging to the technical field of image detection. The method includes: collecting multi-angle images of the surface of a transformer through a camera device, and performing image preprocessing in combination with an adaptive light compensation algorithm; using the Faster R-CNN model to detect defect regions in the preprocessed image, and optimizing the positioning of defect regions through a non-maximum suppression algorithm; based on the optimized defect regions, using a multi-modal feature extraction network to extract defect features, generate feature vectors, and perform defect classification; performing correlation analysis on historical data and defect classification results through a graph neural network, determining the cause of the defect and generating a defect detection report, collecting on-site maintenance results and updating the defect sample library; realizing the automatic detection and classification of defects in power equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of image detection, and specifically relates to a method and system for detecting defects in power equipment. Background Art

[0002] During the long-term operation of power equipment, performance degradation or failures may occur due to natural aging or external factors such as overload and environmental corrosion. Traditional detection methods mainly rely on manual inspections or simple sensor monitoring, which have problems such as low efficiency and high misjudgment rates. In recent years, detection technologies based on image recognition have gradually been applied to the condition monitoring of power equipment. However, existing technologies are difficult to effectively distinguish normal aging from fault characteristics, resulting in a relatively high false alarm rate and affecting the reliability of detection results.

[0003] For example, the Chinese patent with the authorization announcement number CN115359054B discloses a method for detecting defects in power equipment based on the generation of a pseudo-defect space, including: artificially forging local irregularities with normal pictures as materials to generate a type of pseudo-defect image category. Based on the feature extraction and classification network for normal images, defect images, and pseudo-defect images, contrast learning is used to decouple the feature space region, and constraints are embedded in the features of pseudo-defect images, normal images, and defect images to enhance the anti-interference ability of complex backgrounds in difficult-to-distinguish samples and improve the recognition accuracy of the model for hidden defects in power equipment; using the power equipment defect model to predict the classification loss and obtaining edge pseudo-defect samples based on the classification loss. After several iterative trainings, the edge pseudo-defect samples are used to introduce a weakly supervised defect masking strategy for training to alleviate overfitting, help the model break through the training bottleneck, obtain the optimal model parameters and fix them, and use this power equipment defect detection model to detect the image to be detected.

[0004] The above existing technologies have the following problems: The quality of pseudo-defect images directly affects the performance of the model. If the generated pseudo-defects are significantly different from real defects, it may cause the model to learn incorrect features and reduce the detection accuracy; Although the recognition ability for hidden defects is enhanced through contrast learning and pseudo-defect generation, its effect highly depends on the design of pseudo-defects and the optimization of contrast learning; The consideration of the operating state of the equipment is lacking. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention proposes a method and system for detecting defects in power equipment. The method includes collecting multi-angle images of the surface of a transformer through a camera device, and performing image preprocessing in combination with an adaptive light compensation algorithm; using a Faster R-CNN model to detect defect regions in the preprocessed images, and optimizing the positioning of defect regions through a non-maximum suppression algorithm; based on the optimized defect regions, using a multi-modal feature extraction network to extract defect features, generate feature vectors, and perform defect classification; using a graph neural network to perform correlation analysis on historical data and defect classification results, determine the cause of the defect and generate a defect detection report, collect on-site maintenance results and update the defect sample library; realizing the automatic detection and classification of defects in power equipment.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for detecting defects in power equipment, comprising:

[0008] Step S1: Collect images of the surface of the transformer through a camera device, and in combination with an adaptive light compensation algorithm, obtain multi-angle images and perform preprocessing;

[0009] Step S2: Use a Faster R-CNN model to detect defect regions in the preprocessed multi-angle images, output the bounding box coordinates and confidence scores of the defect regions, and at the same time, introduce a non-maximum suppression optimization algorithm to optimize the positioning of the defect regions;

[0010] Step S3: According to the optimized defect regions, use a multi-modal feature extraction network to extract the features of the defect regions from the preprocessed multi-angle images, generate feature vectors, and based on the feature vectors, use an improved classification algorithm to classify the defects into defects caused by normal aging and defects caused by faults, and obtain the defect classification results;

[0011] Step S4: Combine the historical data of equipment operation to construct a multi-source data fusion model, and use a graph neural network model to perform correlation analysis on the historical data of equipment operation and the defect classification results, determine whether the defect is caused by normal aging or equipment failure, and output a defect detection report;

[0012] Step S5: Feed back the defect detection report to the equipment operation and maintenance team, collect on-site maintenance results, and construct a defect sample library.

[0013] Specifically, the specific steps for obtaining multi-angle images in step S1 include:

[0014] S1.1: Use a camera device to take multi-angle photos of the surface of the transformer and collect images of the surface of the transformer;

[0015] S1.2: Perform illumination analysis on the acquired surface images of the transformer to obtain the average brightness and contrast results of the surface images of the transformer. If the average brightness is less than the preset minimum brightness threshold or greater than the preset maximum brightness threshold, or the contrast is less than the preset minimum contrast threshold, then by calculating the product of the pixel value at the coordinate of the surface image of the transformer and the local illumination enhancement coefficient, and calculating the summation result of multiplying the pixel value at the coordinate of the surface image of the transformer by the Gaussian kernel at the point and adding the illumination smoothing term, by calculating the ratio of the product to the summation result, and then adding the global brightness offset term, obtain the illumination compensation result, and adjust the illumination conditions according to the illumination compensation result;

[0016] S1.3: Generate multi-angle images with uniform illumination according to the illumination compensation result.

[0017] Specifically, the specific steps of step S2 include:

[0018] S2.1: Input the preprocessed multi-angle images into the Faster R-CNN model, introduce the dynamic anchor generation mechanism, use the feature map to dynamically generate N anchors on the preprocessed multi-angle images, and output the classification score and bounding box offset information for each anchor; each anchor corresponds to a potential candidate region;

[0019] S2.2: Sort the classification scores in descending order, select the top M anchors corresponding to the classification scores, generate candidate regions in combination with the bounding box offset information, and extract features for each candidate region through a convolutional neural network to generate high-dimensional feature vectors;

[0020] S2.3: Use the high-dimensional feature vectors of each candidate region as input, use the fully connected layer to map the high-dimensional feature vectors to the category space, and apply the Softmax function to the output vector of the fully connected layer to calculate the probability of each category;

[0021] S2.4: According to the output of the Softmax function, obtain the probability that each candidate region belongs to each category;

[0022] S2.5: Map the high-dimensional feature vectors of each candidate region to the bounding box offset space through the regression layer to obtain the offset, and use the offset to adjust the bounding box of the candidate region to output the bounding box coordinates.

[0023] Specifically, the specific steps of step S2 further include:

[0024] S2.6: Calibrate the class probabilities of each candidate region using the temperature scaling method to obtain the calibrated class probabilities, and use the calibrated class probabilities as confidence scores. Combine the bounding box coordinates and output the bounding box coordinates and confidence scores of the calibrated defect regions.

[0025] S2.7: According to the bounding box coordinates and confidence scores of all detected defect regions, calculate the intersection area and union area of any two bounding boxes respectively, and obtain the intersection over union (IoU) of the two bounding boxes by calculating the ratio of the two. At the same time, sort all defect regions in descending order of confidence scores to obtain a sorted list of defect regions.

[0026] S2.8: Select the defect region with the highest confidence, and delete the non - highest - confidence defect regions whose IoU with the defect region with the highest confidence is greater than a preset threshold.

[0027] S2.9: Iteratively repeat steps S2.7 - S2.8 for the remaining defect regions, and use the filtered defect regions as the optimized defect regions. At the same time, output their bounding box coordinates and confidence scores.

[0028] Specifically, the specific steps of step S3 include:

[0029] S3.1: Obtain the bounding box coordinates of each optimized defect region, and extract the image patches of the defect regions from the pre - processed multi - angle images according to the bounding box coordinates.

[0030] S3.2: Use a convolutional neural network to extract the image features of the defect regions from the image patches of the defect regions.

[0031] S3.3: According to the bounding box coordinates of the optimized defect regions, extract their neighborhood image patches, and input the neighborhood image patches into a graph neural network to extract context features.

[0032] S3.4: Concatenate the image features and context features to generate a multi - modal feature vector.

[0033] Specifically, the specific steps of step S3 also include:

[0034] S3.5: Input the multi - modal feature vector into a multi - scale gradient - boosting tree, use n decision trees to step - by - step optimize the classification result, and output the defect class according to the classification results of the decision trees; the multi - scale gradient - boosting tree contains n decision trees.

[0035] The process of outputting the defect class is as follows:

[0036] Based on the multi - modal feature vector, calculate the prediction result of the t - th tree at the l - th scale for the multi - modal feature vector, and combine the weights of different scales to obtain the first dependent variable.

[0037] Calculate additional features at different scales through a scale feature function, and combine regularization parameters at different scales to obtain a second dependent variable; the additional features are context features.

[0038] Sum the first dependent variable and the second dependent variable to obtain a local prediction value at the current scale, and use it as the third dependent variable.

[0039] Based on the third dependent variable, sum up the prediction values of all decision trees at the current scale to obtain a comprehensive prediction value at the current scale. Combine weights at different scales and sum them up to obtain a final prediction value after multi-scale fusion, and use it as the fourth dependent variable.

[0040] Map the fourth dependent variable through a sign function to obtain a defect classification result.

[0041] S3.6: According to the defect classification result, classify the defects into defects caused by normal aging and defects caused by faults.

[0042] Specifically, the specific steps of step S4 include:

[0043] S4.1: Obtain the device operation historical data and the defect classification result; the defect classification result includes the defect category and its confidence score.

[0044] S4.2: Use statistical methods to extract features from the device operation historical data, and fuse the features of the device operation historical data with the defect classification result to generate a multi-source feature vector.

[0045] S4.3: Construct a graph structure according to the multi-source feature vector, where nodes represent device states or defects, and edges represent the associations between device states and defects.

[0046] S4.4: Use a graph neural network to model the graph structure and learn the feature representations of nodes and edges.

[0047] S4.5: Determine whether the defect is caused by normal aging or device failure according to the learned feature representations of nodes and edges.

[0048] S4.6: Generate a defect detection report according to the determination result, including the defect location, category, and cause analysis.

[0049] A power equipment defect detection system includes: an image acquisition module, a defect detection and positioning module, a feature extraction and classification module, a defect analysis module, and a feedback and management module.

[0050] The image acquisition module is used to collect multi-angle images of the transformer surface and perform preprocessing.

[0051] The defect detection and location module uses the Faster R-CNN model to detect defect areas in images and locates the defect positions through an optimization algorithm;

[0052] The feature extraction and classification module is used to extract features from the detected defect areas and classify the defects based on the extracted features;

[0053] The defect analysis module is used to analyze the defect causes and generate a detection report by combining the device operation historical data and the defect classification results;

[0054] The feedback and management module is used to feedback the detection report to the operation and maintenance team, collect the on-site maintenance results, and update the defect sample library.

[0055] Specifically, the defect detection and location module includes: a defect detection unit and a location optimization unit;

[0056] The defect detection unit uses the Faster R-CNN model to detect defect areas in the preprocessed image and outputs the bounding box coordinates and confidence scores of the defect areas;

[0057] The location optimization unit is used to introduce the non-maximum suppression algorithm to optimize the location of the defect areas.

[0058] Specifically, the defect analysis module includes: a data fusion unit and a defect analysis unit;

[0059] The data fusion unit is used to build a multi-source data fusion model to integrate the device operation historical data and the defect classification results; the device operation historical data includes temperature, load, and operation time;

[0060] The defect analysis unit uses a graph neural network to perform correlation analysis on the fused data, determines whether the defect is caused by normal aging or device failure, and generates a defect detection report.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] 1. The present invention proposes a power equipment defect detection system and has optimized improvements in terms of architecture, operation steps, and processes. The system has the advantages of simple processes, low investment and operation costs, and low production work costs.

[0063] 2. The present invention proposes a method for detecting defects in power equipment. By combining technologies such as image recognition, multi-modal feature extraction, multi-source data fusion, and graph neural networks, it realizes the automatic detection and classification of defects on the surface of transformers. First, through adaptive light compensation and multi-angle image acquisition, the image quality is ensured; the Faster R-CNN model is used to accurately locate the defect area, and through multi-modal feature extraction and an improved classification algorithm, the defects are classified into those caused by normal aging or faults; secondly, by combining the historical operation data of the equipment, a multi-source data fusion model is constructed, and the graph neural network is used for correlation analysis to further determine the cause of the defect and generate a defect detection report; finally, through feedback on the on-site maintenance results and updating the defect sample library, a closed-loop optimization mechanism is formed to continuously improve the detection accuracy and system performance.

[0064] 3. The method for detecting defects in power equipment proposed by the present invention is characterized by high precision and high efficiency, and can effectively distinguish between defects caused by normal aging and faults, reducing misdetection and missed detection; at the same time, through the closed-loop feedback mechanism, the system can continuously optimize the model to adapt to complex and changing actual scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic diagram of a method for detecting defects in power equipment according to the present invention;

[0066] Figure 2 It is a principle flowchart of a method for detecting defects in power equipment according to the present invention;

[0067] Figure 3 It is an architecture diagram of a system for detecting defects in power equipment according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0068] Example 1

[0069] Please refer to Figure 1 and Figure 2 A method for detecting defects in power equipment provided by the present invention includes the following steps:

[0070] Step S1: Use a camera device to collect images of the surface of the transformer, combine with an adaptive light compensation algorithm to obtain multi-angle images, and perform preprocessing;

[0071] Step S2: Use the Faster R-CNN model to detect the defect area in the preprocessed multi-angle images, output the bounding box coordinates and confidence scores of the defect area, and at the same time, introduce a non-maximum suppression optimization algorithm to optimize the positioning of the defect area;

[0072] Step S3: According to the optimized defect area, use a multi-modal feature extraction network to extract the features of the defect area from the pre-processed multi-angle images, generate a feature vector. Based on the feature vector, use an improved classification algorithm to classify the defects into defects caused by normal aging and defects caused by faults, and obtain the defect classification result;

[0073] Step S4: Combine the device operation historical data to construct a multi-source data fusion model, and use a graph neural network model to perform correlation analysis on the device operation historical data and the defect classification result to determine whether the defect is caused by normal aging or device failure, and output a defect detection report;

[0074] Step S5: Feed back the defect detection report to the device operation and maintenance team, and collect the on-site maintenance results to construct a defect sample library.

[0075] The specific steps for obtaining multi-angle images in Step S1 include:

[0076] S1.1: Use a camera device to take multi-angle photos of the transformer surface and collect transformer surface images;

[0077] S1.2: Conduct a lighting analysis on the collected transformer surface images to obtain the average brightness and contrast results. If the average brightness is less than the preset minimum brightness threshold or greater than the preset maximum brightness threshold, or the contrast is less than the preset minimum contrast threshold, then by calculating the product of the pixel value of the transformer surface image at the coordinate and the local lighting enhancement coefficient , and calculating the product of the pixel value of the transformer surface image at the coordinate multiplied by the Gaussian kernel at that point and then adding the lighting smoothing term sum result, by calculating the ratio of the product to the sum result, and then adding the global brightness offset term , thus obtaining the lighting compensation result , and adjusting the lighting conditions according to the lighting compensation result, where represents the pixel value of the transformer surface image at the coordinate , W and H are the width and height of the transformer surface image respectively, x ≤ W, y ≤ H, E represents the total number of pixels of the transformer surface image, and represent the maximum and minimum pixel values of the transformer surface image respectively, represents the local lighting enhancement coefficient, and , represents the gradient value of the transformer surface image at the coordinate , represents the adjustment coefficient, Represents the average illumination intensity of the local area, Represents the smoothing coefficient, Represents the global average brightness, Represents the target brightness, Represents the offset coefficient;

[0078] S1.3: Generate multi-angle images with uniform illumination according to the illumination compensation result.

[0079] The specific steps of step S2 include:

[0080] S2.1: Input the preprocessed multi-angle images into the Faster R-CNN model, introduce the dynamic anchor generation mechanism, and use the feature map to dynamically generate N anchors on the preprocessed multi-angle images. The dynamic anchor generation formula is: , and , and output the classification score and the bounding box offset information, the formula is: ; Each anchor corresponds to a potential candidate region. The classification score indicates whether the anchor contains the target; The bounding box offset information indicates the offset between the anchor and the true bounding box, where, Represents the i-th anchor, Represents the center coordinates of the anchor, and Represent the width and height of the anchor respectively, Represents the dynamically adjusted scale, Represents the dynamically adjusted proportionality coefficient to adapt to defects of different sizes and shapes, Represents the activation function, and Represent the weights and biases of the classification layer respectively, Represents the feature map at The feature vector at the place, and Represent the weights and biases of the regression layer respectively;

[0081] S2.2: Sort the classification scores in descending order, select the anchors corresponding to the top M classification scores, and the screening formula is: , generate the candidate region by combining the bounding box offset information, and , , and extract features from each candidate region through the convolutional neural network to generate high-dimensional feature vectors, where, Anchor j Represents the j-th anchor, Represents the center coordinates of the j-th anchor, and represent the width and height of the j-th anchor point respectively, represent the candidate region corresponding to the generated j-th anchor point, represent the bounding box offset of the j-th anchor point, 、 、 and represent the horizontal offset of the center coordinate of the anchor point, the vertical offset of the center coordinate, the width offset, and the height offset respectively. Threshold represents the preset score threshold, represent the exponential function;

[0082] Furthermore, the specific steps of generating high-dimensional feature vectors by performing feature extraction on each candidate region through a convolutional neural network include:

[0083] (1) For each candidate region, use RoI Pooling to extract features of a fixed size from the feature map, where RoI Pooling represents Region of Interest Pooling. RoI Pooling is the prior art content in the field and not the creative solution of this application, so it will not be elaborated here;

[0084] (2) Divide the candidate region into a grid of size 7*7, and perform max pooling on the features within each grid;

[0085] (3) Output a feature map of size 7*7*A, where A represents the number of feature channels;

[0086] (4) Flatten the feature map of size 7*7*A into a one-dimensional vector as the feature representation of the candidate region.

[0087] S2.3: Use the high-dimensional feature vector of each candidate region, such as 7*7*A, as the input, and use a fully connected layer to map the high-dimensional feature vector to the category space. The formula of the fully connected layer is: , and apply the Softmax function to the output vector of the fully connected layer to calculate the probability of each category , where Z represents the output vector, which is the score of each category, represents the weight matrix of the fully connected layer, X represents the input high-dimensional feature vector, B represents the bias vector of the fully connected layer, represents the probability that the candidate region belongs to category d, D represents the total number of categories, represents the score of category d, represents the score of category k, and ;

[0088] S2.4: According to the output of the Softmax function, obtain the probability that each candidate region belongs to each category;

[0089] S2.5: Map the high-dimensional feature vectors of each candidate region to the bounding box offset space through the regression layer to obtain the offsets , and use the offsets to adjust the bounding box coordinates of the candidate regions , and the output is the precise bounding box coordinates ;

[0090] S2.6: Use to calibrate the class probabilities of each candidate region to obtain the calibrated class probabilities , and use the calibrated class probabilities as confidence scores. Combine the precise bounding box coordinates to output the bounding box coordinates and confidence scores of the calibrated defect regions, where represents the adaptive weight of class d, represents the characteristic function of class d, represents the adaptive temperature parameter;

[0091] S2.7: According to the bounding box coordinates and confidence scores of all detected defect regions, calculate the intersection area and union area of any two bounding boxes respectively, and obtain the intersection over union (IoU) of the two bounding boxes by calculating the ratio of the two. At the same time, sort all defect regions in descending order of confidence scores to obtain the sorted list of defect regions, where represents the top-left coordinates of bounding box U, represents the bottom-right coordinates of bounding box U, represents the top-left coordinates of bounding box V, represents the bottom-right coordinates of bounding box V, represents the area of bounding box U, represents the area of bounding box V, represents the maximum value and minimum value functions respectively; and represent the maximum and minimum functions respectively;

[0092] S2.8: Select the defect region with the highest confidence, and delete the non-highest-confidence defect regions whose intersection over union with the defect region with the highest confidence is greater than the preset threshold;

[0093] S2.9: Iteratively repeat the process for the remaining defect regions according to S2.7 - S2.8, and use the filtered defect regions as the optimized defect regions. At the same time, output their bounding box coordinates and confidence scores.

[0094] The specific steps of step S3 include:

[0095] S3.1: Obtain the bounding box coordinates of each optimized defect region. According to the bounding box coordinates, extract the image patches of the defect regions from the preprocessed multi - angle images. ;

[0096] S3.2: Use a convolutional neural network to extract the image features of the defect regions from the image patches of the defect regions. , where represents the convolutional neural network function, and the convolutional neural network is the prior art content in this field and not the creative solution of this application, so it will not be elaborated here;

[0097] S3.3: According to the bounding box coordinates of the optimized defect regions, extract their neighborhood image patches. And input the neighborhood image patches into the graph neural network to extract context features. , where represents the graph neural network function, and the graph neural network is the prior art content in this field and not the creative solution of this application, so it will not be elaborated here;

[0098] S3.4: Concatenate the image features and context features to generate a multi - modal feature vector;

[0099] S3.5: Input the multi - modal feature vector into a multi - scale gradient boosting tree, use n decision trees to optimize the classification result step by step, and output the defect category according to the classification result of the decision trees; the multi - scale gradient boosting tree contains n decision trees;

[0100] The process of outputting the defect category is as follows:

[0101] Based on the multi - modal feature vector, calculate the prediction result of the t - th tree at the l - th scale for the multi - modal feature vector, and combine the weights of different scales to obtain the first dependent variable;

[0102] Calculate the additional features of different scales through the scale feature function, and combine the regularization parameters of different scales to obtain the second dependent variable; the additional features are context features;

[0103] Sum up the first dependent variable and the second dependent variable to obtain the local prediction value at the current scale, and use it as the third dependent variable;

[0104] Based on the third dependent variable, sum up the prediction values of all decision trees at the current scale to obtain the comprehensive prediction value at the current scale, combine the weights of different scales, and sum up to obtain the final prediction value after multi - scale fusion, and use it as the fourth dependent variable;

[0105] Map the fourth dependent variable through the sign function to obtain the defect classification result;

[0106] Among them, the formula is: , where Y represents the defect classification result, represents the multi-modal feature vector, represents the weight of the l-th scale, represents the weight of the t-th decision tree at the l-th scale, represents the prediction result of the t-th decision tree at the l-th scale, represents the regularization coefficient of the l-th scale, represents the additional feature function of the l-th scale, L represents the number of scales, and n represents the number of decision trees, represents the sign function;

[0107] S3.6: According to the defect classification result, classify the defects into defects caused by normal aging and defects caused by faults. Among them, the defect classification result is -1 or 1. Among them, 1 represents the defect caused by normal aging, and -1 represents the defect caused by a fault.

[0108] The specific steps of step S4 include:

[0109] S4.1: Obtain the device operation historical data and the defect classification result; the defect classification result includes the defect category and its confidence score;

[0110] S4.2: Use statistical methods to extract features from the device operation historical data, and fuse the features of the device operation historical data with the defect classification result to generate a multi-source feature vector. Among them, the statistical method is the prior art content in the field and is not the creative solution of this application, so it will not be elaborated here;

[0111] S4.3: Construct a graph structure according to the multi-source feature vector. Among them, the nodes represent the device state or defects, and the edges represent the association between the device state and the defects;

[0112] Further, the specific steps of constructing the graph structure include:

[0113] (1) Define nodes, where each node represents a device state or a defect;

[0114] For the device state node, the feature is the feature of the device operation historical data;

[0115] For the defect node, the feature is the feature of the defect classification result;

[0116] (2) Define edges, where the edges represent the association between the device state and the defects, and the weight of the edges is obtained based on the similarity between the device state and the defects;

[0117] (3) Use the device state and the defects as nodes to construct a node set;

[0118] (4) Construct an edge set based on the associations between nodes;

[0119] (5) Obtain the weights of the edges by calculating the cosine similarity between the eigenvectors of different nodes;

[0120] (6) Represent the graph structure using an adjacency matrix and represent the node features using a node feature matrix;

[0121] The adjacency matrix includes: if the edge connecting any two nodes is in the edge set, the adjacency matrix is equal to the weight of the edge between these two nodes; otherwise, the adjacency matrix is equal to 0;

[0122] The node feature matrix is composed of the eigenvectors of different nodes.

[0123] S4.4: Use a graph neural network to model the graph structure and learn the feature representations of nodes and edges;

[0124] Further, the specific steps of S4.4 include:

[0125] (1) Input the adjacency matrix and the node feature matrix into the graph neural network model;

[0126] (2) Use the graph neural network model to learn the feature representations of nodes and edges, where the graph neural network model is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here.

[0127] S4.5: Determine whether the defect is caused by normal aging or equipment failure based on the learned feature representations of nodes and edges;

[0128] S4.6: Generate a defect detection report according to the determination result, including the defect location, category, and cause analysis.

[0129] Embodiment 2

[0130] Please refer to Figure 3 , another embodiment provided by the present invention: A power equipment defect detection system includes:

[0131] An image acquisition module, a defect detection and positioning module, a feature extraction and classification module, a defect analysis module, and a feedback and management module;

[0132] The image acquisition module is used to acquire multi-angle images of the transformer surface and improve the image quality through preprocessing to provide high-quality input data for subsequent defect detection;

[0133] The defect detection and positioning module uses the Faster R-CNN model to detect the defect areas in the image and accurately locate the defect positions through an optimization algorithm;

[0134] A feature extraction and classification module, which is used to extract features from the detected defect areas and classify the defects based on the features;

[0135] A defect analysis module, which is used to analyze the defect causes and generate a detection report by combining the device operation historical data and the defect classification results;

[0136] A feedback and management module, which is used to feedback the detection report to the operation and maintenance team, collect the on-site maintenance results, and update the defect sample library.

[0137] The image acquisition module includes: an image acquisition unit, a light compensation unit, and a preprocessing unit;

[0138] The image acquisition unit performs multi-angle image acquisition on the surface of the transformer through a camera device;

[0139] The light compensation unit is used to adjust the image lighting conditions in combination with an adaptive light compensation algorithm to ensure the consistency of image quality;

[0140] The preprocessing unit is used to perform denoising, enhancement, and normalization processing on the acquired images to generate preprocessed images.

[0141] The defect detection and localization module includes: a defect detection unit and a localization optimization unit;

[0142] The defect detection unit uses the Faster R-CNN model to detect the defect areas in the preprocessed images and outputs the bounding box coordinates and confidence scores of the defect areas;

[0143] The localization optimization unit is used to introduce a non-maximum suppression algorithm to optimize the localization of the defect areas, remove redundant bounding boxes, and improve the detection accuracy.

[0144] The feature extraction and classification module includes: a feature extraction unit and a defect classification unit;

[0145] The feature extraction unit uses a multi-modal feature extraction network to extract the features of the defect areas from the preprocessed images and generates high-dimensional feature vectors;

[0146] The defect classification unit, based on the feature vectors, uses an improved classification algorithm to classify the defects into defects caused by normal aging and defects caused by faults.

[0147] The defect analysis module includes: a data fusion unit and a defect analysis unit;

[0148] The data fusion unit is used to construct a multi-source data fusion model to integrate the device operation historical data and the defect classification results; the device operation historical data includes temperature, load, and operation time;

[0149] The defect analysis unit uses a graph neural network to perform correlation analysis on the fused data, determines whether the defect is caused by normal aging or equipment failure, and generates a defect detection report.

[0150] The feedback and management module includes: a report feedback unit, a sample collection unit, and a sample library update unit;

[0151] The report feedback unit is used to send the defect detection report to the equipment operation and maintenance team to guide on-site maintenance work;

[0152] The sample collection unit is used to collect on-site maintenance results, including maintenance records and images after repair, and construct a defect sample library;

[0153] The sample library update unit is used to add newly discovered defect samples, including images, feature vectors, classification results, and maintenance results, to the defect sample library.

[0154] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make changes, modifications, substitutions, and variations to the above embodiments without departing from the purpose and scope of the present invention. These all fall within the protection scope of the present invention.

Claims

1. A method for detecting defects in electric power equipment, characterized in that: Including: Step S1: Use a camera device to collect images of the transformer surface, combine with an adaptive light compensation algorithm to obtain multi-angle images, and perform preprocessing. Step S2: Use the Faster R-CNN model to detect the defect areas in the preprocessed multi-angle images, output the bounding box coordinates and confidence scores of the defect areas. At the same time, introduce the non-maximum suppression optimization algorithm to optimize the positioning of the defect areas. Step S3: According to the optimized defect areas, use a multi-modal feature extraction network to extract the features of the defect areas from the preprocessed multi-angle images to generate feature vectors. Based on the feature vectors, use an improved classification algorithm to classify the defects into defects caused by normal aging and defects caused by faults to obtain the defect classification results. Step S4: Combine the device operation historical data to construct a multi-source data fusion model, and use a graph neural network model to perform correlation analysis on the device operation historical data and the defect classification results to determine whether the defect is caused by normal aging or device failure, and output a defect detection report. Step S5: Feed the defect detection report back to the device operation and maintenance team, collect the on-site maintenance results, and construct a defect sample library. The specific steps of the said Step S3 include: S3.1: Obtain the bounding box coordinates of each optimized defect area. According to the bounding box coordinates, extract the image patches of the defect areas from the preprocessed multi-angle images. S3.2: Use a convolutional neural network to extract the image features of the defect areas from the image patches of the defect areas. S3.3: According to the bounding box coordinates of the optimized defect areas, extract their neighborhood image patches, and input the neighborhood image patches into a graph neural network to extract context features. S3.4: Concatenate the image features and context features to generate a multi-modal feature vector. The specific steps of the said Step S3 also include: S3.5: Input the multimodal feature vector into the multi-scale gradient boosting tree, using n The decision tree optimizes the classification results step by step, and outputs the defect category according to the classification results of the decision tree; the multi-scale gradient boosting tree includes n decision trees; The process of outputting the defect category is: Based on the multi-modal feature vector, calculate the prediction result of the l th tree of the t th scale for the multi-modal feature vector, and combine the weights of different scales to obtain the first dependent variable; Calculate the additional features of different scales through a scale feature function, combine with different scale regularization parameters to obtain the second dependent variable; the additional features are context features. Sum the first dependent variable and the second dependent variable to obtain the local prediction value of the current scale, and use it as the third dependent variable. Based on the third dependent variable, accumulate and sum the prediction values of all decision trees at the current scale to obtain the comprehensive prediction value at the current scale. Combine the weights of different scales, accumulate and sum to obtain the final prediction value after multi-scale fusion, and use it as the fourth dependent variable. Map the fourth dependent variable through a sign function to obtain the defect classification result. S3.6: According to the defect classification result, classify the defects into defects caused by normal aging and defects caused by faults.

2. The method for detecting defects of a power device according to claim 1, characterized in that, The specific steps of obtaining the multi-angle images in the said Step S1 include: S1.1: Use a camera device to take multi-angle photos of the transformer surface and collect the transformer surface images. S1.2: Conduct a lighting analysis on the acquired transformer surface image to obtain the average brightness and contrast results of the transformer surface image. If the average brightness is less than the preset minimum brightness threshold or greater than the preset maximum brightness threshold, or the contrast is less than the preset minimum contrast threshold, then by calculating the product of the pixel value at the coordinate of the transformer surface image and the local lighting enhancement coefficient, and calculating the sum result of multiplying the pixel value at the coordinate of the transformer surface image by the Gaussian kernel at the point plus the lighting smoothing term, by calculating the ratio of the product to the sum result, and then adding the global brightness offset term, obtain the lighting compensation result, and adjust the lighting conditions according to the lighting compensation result; S1.3: Generate multi-angle images with uniform illumination according to the light compensation results.

3. A method for detecting defects in electric power equipment according to claim 2, characterized in that: The specific steps of the said Step S2 include: S2.1: Input the pre - processed multi - angle images into the Faster R - CNN model. Introduce a dynamic anchor generation mechanism, use the feature map to dynamically generate N anchors on the pre - processed multi - angle images, and output the classification score and bounding box offset information for each anchor; each anchor corresponds to a potential candidate region. S2.2: Sort the classification scores in descending order, select the anchors corresponding to the top M classification scores, generate candidate regions by combining the bounding box offset information, and extract features for each candidate region through a convolutional neural network to generate high - dimensional feature vectors. S2.3: Take the high - dimensional feature vector of each candidate region as input, use a fully - connected layer to map the high - dimensional feature vector to the class space, and apply the Softmax function to the output vector of the fully - connected layer to calculate the probability of each class. S2.4: According to the output of the Softmax function, obtain the probability that each candidate region belongs to each class. S2.5: Map the high - dimensional feature vector of each candidate region to the bounding box offset space through a regression layer to obtain the offset, and use the offset to adjust the bounding box of the candidate region to output the bounding box coordinates.

4. The method for detecting defects of a power device according to claim 3, characterized in that, The specific steps of step S2 also include: S2.6: Use the temperature scaling method to calibrate the class probabilities of each candidate region to obtain the calibrated class probabilities, and use the calibrated class probabilities as confidence scores. Combine with the bounding box coordinates to output the bounding box coordinates and confidence scores of the calibrated defect regions. S2.7: According to the bounding box coordinates and confidence scores of all detected defect regions, calculate the intersection area and union area of any two bounding boxes respectively, and obtain the intersection - over - union ratio of the two bounding boxes by calculating the ratio of the two. At the same time, sort all defect regions in descending order of confidence scores to obtain a sorted list of defect regions. S2.8: Select the defect region with the highest confidence, and delete the non - highest - confidence defect regions whose intersection - over - union ratio with the defect region with the highest confidence is greater than a preset threshold. S2.9: Iteratively repeat steps S2.7 - S2.8 for the remaining defect regions, and use the filtered defect regions as the optimized defect regions. At the same time, output their bounding box coordinates and confidence scores.

5. A method for detecting defects in electric power equipment according to claim 4, characterized in that: The specific steps of step S4 include: S4.1: Obtain the device operation historical data and defect classification results; the defect classification results include defect categories and their confidence scores. S4.2: Use statistical methods to extract features from the device operation historical data, and fuse the features of the device operation historical data with the defect classification results to generate multi - source feature vectors. S4.3: Construct a graph structure according to the multi - source feature vectors, where nodes represent device states or defects, and edges represent the associations between device states and defects. S4.4: Use a graph neural network to model the graph structure to learn the feature representations of nodes and edges. S4.5: According to the learned feature representations of nodes and edges, determine whether the defect is caused by normal aging or device failure. S4.6: According to the determination result, generate a defect detection report, including defect location, category, and cause analysis.

6. A power equipment defect detection system, which is used to implement the power equipment defect detection method described in any one of claims 1-5, and is characterized in that, Including: Image acquisition module, defect detection and localization module, feature extraction and classification module, defect analysis module, feedback and management module; The image acquisition module is used to acquire multi-angle images of the transformer surface and perform preprocessing; The defect detection and localization module uses the Faster R-CNN model to detect defect regions in the image and locates the defect positions through an optimization algorithm; The feature extraction and classification module is used to extract features from the detected defect regions and classify the defects based on the extracted features; The defect analysis module is used to analyze the defect causes and generate a detection report by combining the device operation historical data and the defect classification results; The feedback and management module is used to feedback the detection report to the operation and maintenance team, collect on-site maintenance results, and update the defect sample library.

7. The power equipment defect detection system according to claim 6, wherein The defect detection and localization module includes: a defect detection unit and a localization optimization unit; The defect detection unit uses the Faster R-CNN model to detect defect regions in the preprocessed image and outputs the bounding box coordinates and confidence scores of the defect regions; The localization optimization unit is used to introduce a non-maximum suppression algorithm to optimize the localization of defect regions.

8. The power equipment defect detection system according to claim 7, characterized in that, The defect analysis module includes: a data fusion unit and a defect analysis unit; The data fusion unit is used to construct a multi-source data fusion model to integrate the device operation historical data and the defect classification results; the device operation historical data includes temperature, load, and operation time; The defect analysis unit uses a graph neural network to perform correlation analysis on the fused data, determines whether the defect is caused by normal aging or device failure, and generates a defect detection report.

Images