Thermal defect identification method and system for high-voltage switch equipment, and computer equipment

Abstract

The invention belongs to the technical field of fault diagnosis, and discloses a thermal defect identification method and system for a high-voltage switchgear, and computer equipment, and the method comprises the steps: firstly segmenting an infrared image through employing a transfer learning optimized Mask R-CNN model, reducing the dependence of annotated data through sharing pre-training parameters, and achieving the region extraction of pixel-level equipment; secondly, multi-dimensional temperature information is extracted in combination with a gray histogram and a gray co-occurrence matrix, and key features are screened through PCA to enhance noise immunity; and finally, the LSSVM is adopted for classification, so that the training efficiency is remarkably improved. According to the method, the equipment area is automatically segmented through deep learning, temperature distribution is quantified in combination with multi-dimensional features, man-made misjudgment is reduced, pre-training model parameter sharing is utilized, new tasks are rapidly adapted in a small sample scene, the generalization ability and efficiency are improved, feature dimensions are compressed through principal component analysis, the real-time monitoring requirement is met, and the monitoring efficiency is improved.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology and relates to a method and system for identifying thermal defects in high-voltage switchgear, as well as computer equipment. Background Technology

[0002] High-voltage switchgear exists in numerous front-line work sites. Due to prolonged exposure to outdoor environments and continuous operation, switchgear inevitably experiences faults of varying degrees, with thermal faults being a particularly common type. In the early stages of a thermal fault, high-voltage switchgear often exhibits a thermal defect state accompanied by abnormal localized temperature rises. In this state, the switchgear continues to operate without immediate damage or power outages. Therefore, if the thermal defect state of high-voltage switchgear can be identified and remedial measures implemented promptly before it deteriorates into a thermal fault, power system instability caused by equipment failure can be nipped in the bud, significantly improving the safety and reliability of the power grid.

[0003] Infrared thermal imaging technology, as a high-precision, high-sensitivity, non-contact temperature distribution measurement technology, has become an indispensable tool for condition monitoring and fault diagnosis of numerous devices, and is widely used in many fields such as power, agriculture, railway, construction, and petrochemicals. Infrared thermal imaging technology uses infrared detectors to collect the infrared radiation and thermal field distribution of target high-voltage switchgear, and processes the infrared thermal signals into electrical signals to obtain visualized infrared thermal images. Based on the infrared images containing switchgear temperature information and the heating characteristics of the equipment under thermal defect conditions, effective monitoring and identification of thermal defects in high-voltage switchgear can be achieved.

[0004] For example, Chinese patent application CN110619623A discloses an automatic identification method for overheating of transformer equipment joints. This method uses labeled images of transformer equipment to train a deep learning target segmentation model, Mask Region-based Convolutional Neural Network (Mask R-CNN), and a machine learning model, Support Vector Machine. Then, the trained model is used to locate the conductive joints in the images of the transformer equipment to be identified. Finally, the overheating defect in the conductive joint area is determined according to relevant defect grading standards, and the overheating identification result is output. However, faced with massive amounts of accumulated infrared image data of switchgear, algorithms based on threshold segmentation and temperature fitting are limited by environmental interference (such as illumination and radiation noise) and insufficient generalization ability, making it difficult to meet the accurate monitoring needs of complex working conditions in a big data environment. Furthermore, traditional temperature estimation algorithms are sensitive to background noise and are prone to misjudgment when the ambient temperature and humidity fluctuate. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the monitoring accuracy of high-voltage switchgear under complex operating conditions in big data scenarios.

[0006] This invention is achieved through the following technical solution: A method for identifying thermal defects in high-voltage switchgear includes the following steps: Step 1: Acquire infrared images of high-voltage disconnect switches and perform image segmentation on the infrared images of high-voltage disconnect switches based on a Mask R-CNN model optimized by transfer learning; Step 2: Based on the gray-level histogram and gray-level co-occurrence matrix, feature extraction is performed on the segmented infrared image of the high-voltage disconnector to obtain first-order and second-order features, respectively; Step 3: Perform feature filtering on the extracted first-order and second-order features based on principal component analysis; Step 4: Input the filtered first-order and second-order features into the classification model based on least squares support vector machine to identify thermal defects in high-voltage switchgear.

[0007] Preferably, in step 1, the Mask R-CNN model optimized based on transfer learning is specifically: the Mask R-CNN model is pre-trained using a benchmark dataset, and the weight parameters obtained from the pre-training are transferred and shared to the image segmentation model of high-voltage switchgear, thereby completing the optimization of the Mask R-CNN model.

[0008] Preferably, the Mask R-CNN model includes an image feature extraction module, a region proposal module, and a prediction regression module; The image feature extraction module consists of a backbone network composed of a ResNet101 network based on residual neural network and a feature pyramid FPN network, which is used to mine and extract feature information from infrared images of high-voltage switchgear. The region proposal module consists of a region proposal network (RPN) and a region of interest (ROI) alignment module (ROI Align), and is used to obtain proposal boxes for the region of interest. The prediction regression module consists of bounding box branches, category branches, and mask branches, and is used to complete the final category regression and segmentation tasks.

[0009] Preferably, the image feature extraction module is constructed as follows: First, two basic modules incorporating residual learning structures are set in the ResNet101 network, namely Conv Block and Identity Block; then, the backbone structure of the ResNet101 network is built using Conv Block and Identity Block; finally, the constructed ResNet101 network is combined with the Feature Pyramid Network (FPN) to complete the construction of the image feature extraction module of the Mask R-CNN model. The region proposal module consists of a Region Proposal Network (RPN) and a Region of Interest Alignment (ROI) module. It is used to obtain proposal boxes for the regions of interest. Specifically, firstly, the coordinates of each point in the feature map obtained by the image feature module are used as the center, and three prior boxes are preset with different heights and widths. Then, the RPN network calculates a correction bias by regression and uses the correction bias to optimize and correct the center coordinates, height, and width of each prior box to obtain a new batch of prior boxes. Finally, the proposal boxes for the regions of interest are obtained through ROI Align calibration. The prediction regression module consists of a bounding box branch, a category branch, and a mask branch, used to complete the final category regression and segmentation tasks. Specifically, the region proposal module inputs the acquired feature map into a fully connected layer network, outputs the category label of the high-voltage switchgear through the category regression branch, and outputs the bounding box of the target region through the bounding box regression branch. At the same time, the feature map is input into a fully convolutional network. After extracting deep semantic features through convolutional layers, the feature map is upsampled through deconvolutional layers to restore the size of the feature map to be consistent with the input image. A mask is generated pixel by pixel in the upsampled feature map to complete the image segmentation.

[0010] Preferably, the first-order statistical features used in the first-order feature extraction process based on gray-level histograms include average gray intensity, variance, skewness, kurtosis, entropy, and maximum gray intensity.

[0011] Preferably, the second-order statistical features used in the second-order feature extraction process, based on the gray-level co-occurrence matrix, include contrast, energy, correlation, and homogeneity.

[0012] Preferably, the feature selection based on principal component analysis for the extracted first-order and second-order features is as follows: First, the originally correlated high-dimensional data features are transformed into uncorrelated features of the same dimension through orthogonal transformation. Then, variance analysis is performed on the uncorrelated features of the same dimension, and each feature is sorted from high to low according to its variance contribution rate. The feature with the highest variance contribution is selected as the first principal component, the second highest as the second principal component, and so on. Finally, the high-ranking principal component features with a cumulative variance contribution rate greater than a threshold are retained, thus completing the feature selection of first-order and second-order features.

[0013] Preferably, in step S4, the regression function of the least squares support vector machine model is:

[0014] In the formula, The coefficients corresponding to each sample, For bias terms, Kernel function, For each training sample, Value to be predicted This is the final regression function.

[0015] A thermal defect identification system for high-voltage switchgear includes: An image segmentation module is used to acquire infrared images of high-voltage disconnect switches and perform image segmentation on the acquired infrared images of high-voltage disconnect switches based on a Mask R-CNN model optimized by transfer learning. The feature extraction module is used to extract first-order and second-order features from the segmented infrared image of the high-voltage disconnector based on the gray-level histogram and gray-level co-occurrence matrix, respectively. A feature filtering module is used to filter the extracted first-order and second-order features based on principal component analysis. The identification decision module is used to input the filtered first-order features and second-order features into a classification model based on least squares support vector machine to identify thermal defects in high-voltage switchgear.

[0016] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described above.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects: This invention discloses a method for identifying thermal defects in high-voltage switchgear. The method first segments the acquired infrared images of high-voltage disconnecting switches using a Mask R-CNN model optimized through transfer learning, reducing the model's dependence on labeled data and improving segmentation robustness. Then, it captures the overall characteristics of temperature distribution using gray-level histograms (i.e., first-order statistics), analyzes texture details using gray-level co-occurrence matrices (i.e., second-order statistics), and further filters key features using principal component analysis (PCA) to remove redundant information. Multi-dimensional feature fusion not only quantifies the spatial heterogeneity of temperature distribution but also enhances the model's robustness to noise and reduces overfitting through PCA filtering. Least squares support vector machines significantly improve training efficiency through least squares optimization. Simultaneously, its high generalization ability can adapt to complex classification requirements for different defect types (such as contact oxidation and ablation). Therefore, this method automatically segments equipment regions through deep learning, quantifies temperature distribution using multi-dimensional features to reduce human error, and utilizes pre-trained model parameter sharing to quickly adapt to new tasks in small-sample scenarios, improving generalization ability and efficiency. Furthermore, principal component analysis compresses feature dimensions to meet real-time monitoring requirements and improve monitoring efficiency. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for identifying thermal defects in high-voltage switchgear according to the present invention. Figure 2 This is the residual neural network structure in Embodiment 2 of the present invention; Figure 3 This is the Feature Pyramid (FPN) network structure in Embodiment 2 of the present invention; Figure 4 This is a schematic diagram illustrating the basic principle of principal component analysis in Embodiment 2 of the present invention; Figure 5 This is the thermal defect identification process for high-voltage switchgear according to Embodiment 2 of the present invention; Figure 6 This is a schematic diagram of the structure of a thermal defect identification system for high-voltage switchgear according to the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 like Figure 1 As shown, this invention discloses a method for identifying thermal defects in high-voltage switchgear, comprising the following steps: Step 1: Acquire infrared images of high-voltage disconnect switches and perform image segmentation on the acquired infrared images of high-voltage disconnect switches based on a Mask Region-based Convolutional Neural Network (Mask R-CNN) deep learning model optimized by transfer learning; Specifically, the Mask R-CNN model is pre-trained using a benchmark dataset, and the pre-trained weight parameters are transferred and shared to the Mask R-CNN model of high-voltage switchgear to complete the optimization of the Mask R-CNN model.

[0024] The Mask R-CNN model includes an image feature extraction module, a region proposal module, and a prediction and regression module; The image feature extraction module consists of a backbone network composed of a fusion of a Residual Network 101 (ResNet101) and a Feature Pyramid Network (FPN), used to mine and extract feature information from infrared images of high-voltage switchgear. The region proposal module consists of a Region Proposal Network (RPN) and a Region of Interest Align (ROI Align) module, and is used to obtain proposal boxes for the region of interest. The prediction regression module consists of bounding box branch, class branch, and mask branch, and is used to complete the final class regression and segmentation tasks.

[0025] Taking the image feature extraction module as an example, which is composed of a backbone network based on the fusion of ResNet101 and FPN networks, the construction of the image feature extraction module is as follows: First, two basic modules incorporating residual learning structures are set in the ResNet101 network, namely the Convolutional Block and the Identity Block; then, the backbone structure of the ResNet101 network is built using the Convolutional Block and the Identity Block; finally, the built ResNet101 network is combined with the constructed FPN network to obtain the image feature extraction module of the Mask R-CNN model. The region proposal module consists of a Region Proposal Network (RPN) and a Region of Interest (ROI) Align module. It is used to obtain proposal boxes for regions of interest. Specifically, firstly, the coordinates of each point in the feature map obtained by the image feature module are used as the center, and three prior boxes are preset with different heights and widths. Then, the RPN calculates a correction bias through regression and uses the correction bias to optimize and correct the center coordinates, height, and width of each prior box to obtain a new batch of prior boxes. Finally, the proposal boxes for the regions of interest are obtained through calibration by the ROI Align module.

[0026] The prediction regression module consists of a bounding box branch, a category branch, and a mask branch, used to complete the final category regression and segmentation tasks. Specifically, the region proposal module inputs the acquired feature map into a fully connected layer network, outputs the category label of the high-voltage switchgear through the category regression branch, and outputs the bounding box of the target region through the bounding box regression branch. At the same time, the feature map is input into a fully convolutional network. After extracting deep semantic features through layer-by-layer convolution, the feature map is upsampled through a deconvolution layer to restore the size of the feature map to be consistent with the input image. A mask is generated pixel by pixel in the upsampled feature map to cover the precise shape of the high-voltage switchgear in the image, thus completing the image segmentation.

[0027] In this embodiment, the Mask R-CNN model was optimized. Due to the complex structure and large number of parameters of the Mask R-CNN model, and the fact that the initial default values ​​of all the parameters are too far from the features of the infrared image of the high-voltage switchgear, training from the initial state of the model often results in slow convergence, excessive time spent on iterative training, and even a surge in the loss function value, which remains high for several generations, causing the model to fail to converge smoothly.

[0028] This invention introduces a transfer learning method, which pre-trains the model on an existing mature and high-quality dataset, and then transfers and shares the pre-trained weight parameters to the image segmentation model of high-voltage switchgear, thereby completing the initialization of the Mask R-CNN model, narrowing the feature distance with the infrared images of switchgear, accelerating the model convergence process, and to some extent making up for the deficiencies of the infrared image dataset of switchgear.

[0029] Step 2: Based on the gray-level histogram and gray-level co-occurrence matrix, perform first-order feature extraction and second-order feature extraction on the segmented infrared image of the high-voltage disconnector switch, respectively.

[0030] 2.1 The first-order feature extraction process described above uses first-order statistical features based on gray-level histograms, including average gray-level intensity, variance, skewness, kurtosis, entropy, and maximum gray-level intensity. The specific extraction process is as follows.

[0031] First-order statistical features based on gray-level histograms are a common method for describing the asymmetry of infrared images. A gray-level histogram is defined as a graphical representation of the number of pixels in an image at different gray-level intensity levels, and it is a simple and commonly used way to describe the gray-level distribution of an image.

[0032] Gray-level histograms can only provide a visual description of switching devices based on the most basic gray-level value distribution. To extract gray-level features that better represent image information, a series of first-order statistical values ​​need to be calculated for feature representation. The first-order statistical features based on gray-level histograms used in this invention are average gray-level intensity, variance, skewness, kurtosis, entropy, and maximum gray-level intensity.

[0033] Average gray intensity, characterized by the average pixel gray value, describes the overall brightness of the segmented infrared image and can be expressed as:

[0034] In the formula, The average gray intensity, , These represent the width and height of the image, respectively. Representing different grayscale values, Indicates the possible grayscale intensities in an image. Indicates grayscale value The number of pixels.

[0035] The standard deviation measures the dispersion of grayscale pixels in the infrared image of a disconnector switch, and can be expressed as:

[0036] In the formula, Standard deviation, , These represent the width and height of the image, respectively. Representing different grayscale values, Indicates the possible grayscale intensities in an image. Indicates grayscale value The number of pixels, This is the average of all grayscale values.

[0037] Skewness S This determines the asymmetric nature of the gray-level histogram's deviation from the average gray-level intensity. If the skewness... S If the skewness is negative, then the gray intensity of most pixels in the entire image is less than the average gray intensity; if the skewness is negative... S If it is positive, then the opposite is true. Its mathematical expression is as follows:

[0038] In the formula, For skewness, , These represent the width and height of the image, respectively. Representing different grayscale values, Indicates the possible grayscale intensities in an image. Indicates grayscale value The number of pixels, Standard deviation of grayscale pixels This is the average of all grayscale values.

[0039] Kudo K This measure assesses the change in the peak value or flatness of the grayscale intensity distribution relative to a normal distribution, summarizing the distribution of image grayscale values. It can be calculated using the following formula:

[0040] In the formula, , These represent the width and height of the image, respectively. Representing different grayscale values, Indicates the possible grayscale intensities in an image. Indicates grayscale value The number of pixels, The standard deviation of grayscale pixels. This is the average of all grayscale values.

[0041] entropy E Used to measure the disorder of infrared images of disconnecting switches. If entropy... E A larger entropy indicates that the image grayscale values ​​are distributed in a relatively dispersed manner; if the entropy is large... E A smaller value indicates that the image grayscale values ​​are centrally distributed. This statistical characteristic can be expressed by the following formula:

[0042] In the formula, For entropy, , These represent the width and height of the image, respectively. Representing different grayscale values, Indicates the possible grayscale intensities in an image. Indicates grayscale value The number of pixels.

[0043] Maximum grayscale intensity defines the maximum pixel intensity of a grayscale image, and can be expressed by the following formula:

[0044] In the formula, Representing different grayscale values, Indicates the possible grayscale intensities in an image. Indicates grayscale value The number of pixels.

[0045] In the second-order feature extraction process described in 2.2, the second-order statistical features used based on the gray-level co-occurrence matrix include contrast, energy, correlation, and homogeneity. The specific extraction process is as follows.

[0046] The Gray Level Co-occurrence Matrix (GLCM) is a method for analyzing the relative spatial information of an image. It characterizes image information by mining the spatial distribution features of the correlation between pixels in a grayscale image. Its basic principle is as follows: Suppose a two-dimensional grayscale image has the following properties in the horizontal and vertical directions: N x , N y Each pixel is used to equalize 256 grayscale values. k If each pixel's grayscale value can be assigned to one of these grayscale levels, then GLCM can be defined as follows: In the direction and at a distance of d Pixel pairs of specified gray levels ( x i ,y i ), ( x j , y j The number of times a number appears can be expressed by the following formula:

[0047] In the formula, To specify the number of times a pixel pair at a given gray level occurs. This indicates the number of elements in the set that satisfy the condition. , Representing pixel pairs ( x i , y i ), ( x j , y j The angle and distance between them , They represent , The corresponding grayscale level. and These represent the horizontal and vertical coordinates of the pixel relative to i, respectively. and Let j represent the horizontal and vertical coordinates of pixel pair j. The GLCM can be generated by counting the number of pixel pairs. Its matrix form is shown in the following formula. It can be seen that the GLCM is a symmetric matrix.

[0048]

[0049] In the process of generating a GLCM from a grayscale image, since there are 256 grayscale values ​​and each element in the matrix needs to be counted by traversing the entire image, this often requires a large amount of computation. Therefore, when quantizing grayscale levels, the number of grayscale levels should be set without affecting the extraction of image spatial information. k To save computational costs, this invention uses grayscale levels that are much smaller than 256. k =16.

[0050] To extract spatial information from infrared images of high-voltage disconnectors, this invention further statistically calculates second-order grayscale features based on GLCM. This invention selects four second-order features—contrast, energy, correlation, and homogeneity—to describe the relative spatial relationships between image pixels.

[0051] Contrast ratio measures the overall sharpness of a grayscale image. The higher the contrast ratio, the better and clearer the visual effect of the image. Its mathematical expression can be represented by the following formula:

[0052] In the formula, Indicates contrast. Grayscale levels , They represent , The corresponding gray levels, For two adjacent pixels (defined by a specific angle and distance) to simultaneously exhibit grayscale levels and The number of times.

[0053] energy The stability of texture changes in a grayscale image can be represented by the following formula: .

[0054] Correlation determines the degree of correlation between local pixels in the grayscale image of the isolating switch. When the correlation is high, the similarity between elements in the GLCM is higher in the row or column direction. This statistical characteristic can be expressed by the following formula:

[0055] In the formula, Indicates correlation Gray levels

[0056] in, , , , The representation is as follows:

[0057]

[0058]

[0059] .

[0060] Homogeneity reflects the density of GLCM elements relative to the diagonal. In grayscale images, greater homogeneity indicates a more uniform local distribution, which can be expressed by the following formula:

[0061] In the formula, Indicates homogeneity

[0062] Step 3: Perform feature filtering on the extracted first-order and second-order features based on Principal Component Analysis (PCA); Specifically, the process involves: first, transforming the originally correlated high-dimensional data features into uncorrelated features of the same dimension through orthogonal transformation; then, performing variance analysis on the uncorrelated features of the same dimension and sorting them from highest to lowest according to their variance contribution rate; selecting the feature with the highest variance contribution as the first principal component, the second highest as the second principal component, and so on; finally, retaining the high-ranking second principal component features with a cumulative variance contribution rate greater than a threshold (e.g., 90%), thus completing the feature selection for first-order and second-order features.

[0063] Step 4: Input the filtered first-order and second-order features into the classification model based on Least Squares Support Vector Machine (LSSVM) to complete the thermal defect identification of high-voltage switchgear.

[0064] The regression function of the classification model based on least squares support vector machine is:

[0065] In the formula, The coefficients corresponding to each sample, For bias terms, Kernel function, For each training sample, Value to be predicted The number of samples.

[0066] When using Support Vector Machines (SVMs) to solve multi-class classification tasks, it is necessary to construct an LSSVM multi-classifier. This invention employs the binary tree LSSVM algorithm, the basic principle of which is: first, the feature data of all samples are binary classified; then, the two resulting subclass nodes are further binary classified until all nodes have only one class. Therefore, for tasks containing... k This algorithm essentially uses a multi-class classification task with multiple categories. k -1 classifier transformed it into k -1 binary classification task.

[0067] This invention replaces the traditional approach of estimating temperature values ​​from infrared images by extracting multi-dimensional grayscale features that characterize the temperature information of infrared images of high-voltage disconnecting switches. At the same time, it has excellent accuracy and completeness performance for different types of thermal defects, and can adapt to the complex thermal defect identification tasks of high-voltage switchgear in actual engineering.

[0068] Example 2 The features of the present invention will be further described in detail with reference to the following embodiments.

[0069] This invention discloses a method for identifying thermal defects in high-voltage switchgear, specifically including the following steps: S1. Image segmentation of infrared images of high-voltage switchgear based on Mask R-CNN model optimized by transfer learning; S2. Based on the gray-level histogram and gray-level co-occurrence matrix, extract the first-order and second-order gray-level features of the infrared image of the high-voltage disconnector that has been detected and segmented in step 1. S3. Screening of a series of extracted grayscale features based on principal component analysis; S4. Using the filtered grayscale features as input, the thermal defect identification task of high-voltage switchgear is completed based on the least squares support vector machine. Specifically: In S1, the Mask R-CNN model based on convolutional neural networks mainly consists of three basic modules: image feature extraction module, region proposal module, and prediction / regression module. The specific process of S1 is as follows: S1.1 Based on the ResNet101 residual neural network, an image feature extraction module for the Mask R-CNN model is constructed. The basic idea of ​​the residual neural network is as follows: Figure 2 As shown, x and y are the input and output of the module unit, respectively, and F(x) is the output after processing by the convolutional layer. The ResNet101 network introduces a residual neural network into its structure, using residual units to construct skip-layer connections. This superimposes the unit's input with the output of the five convolutional blocks in the ResNet101 network, forming a new output stream for the next layer. The process of building the image feature extraction module of the Mask R-CNN model specifically includes: S1.11. Design two basic modules integrated into the residual learning structure: the Conv Block module and the Identity Block module. The Conv Block module achieves different input and output dimensions by changing the convolution stride of the residual units, thus altering the network's dimensionality. The Identity Block module has the same input and output dimensions, used to deepen the network structure. S1.12. Construct the backbone structure of the ResNet101 network using Conv Block and Identity Block. The ResNet101 network mainly consists of five convolutional blocks, from Convolutional Block 1 to Convolutional Block 5, connected in sequence. Each convolutional block contains several convolutional layers. Both the Conv Block and Identity Block modules are convolutional layers, and these two modules are based on the basic principles of residual neural networks. Convolutional block 1 consists of one convolution operation and one max pooling operation. After processing by convolutional block 1, the resolution of the feature map is reduced to 1 / 4 of the input image. Convolutional block 2 consists of one Conv Block and two Identity Blocks. By setting the convolution stride of the Conv Block residual units to 1, the output resolution of the Conv Block is kept consistent with the input. After processing by convolutional block 2, the resolution of the feature map remains unchanged. Convolutional block 3 contains one Conv Block and three Identity Blocks. The convolution stride of the Conv Block residual units is set to 2. After processing by convolutional block 3, the resolution of the feature map is reduced to 1 / 8 of the input image. Convolutional blocks 4 and 5 are basically similar to convolutional block 3, except that convolutional block 4 has 22 Identity Blocks, while convolutional block 5 has only 2. After processing by convolutional blocks 4 and 5, the resolution of the feature map is reduced to 1 / 16 and 1 / 32 of the input image, respectively.

[0070] S1.13. Construct a Feature Pyramid Network (FPN). The FPN network consists of three parts: a bottom-up channel, a top-down channel, and lateral connections. Its structure is as follows: Figure 3 As shown. In the Feature Pyramid Network (FPN), the bottom-up channel is formed by stacking feature maps of different scales extracted by the preceding deep neural network; in the top-down channel, the highest layer feature map is calculated by convolution operation of the highest layer of the bottom-up channel, and the other layers are obtained by fusion of upsampled high-level feature maps with feature maps of the same size in the lower layer through lateral connections; lateral connections represent the superposition and fusion calculation operation of two feature layers.

[0071] S1.14. Based on the feature maps extracted from the ResNet101 network at different scales, a Feature Pyramid Network (FPN) is superimposed to obtain the image feature extraction module of the Mask R-CNN model. After the input image is processed by the ResNet101 network to extract features, the output resolution is compressed to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image, respectively. These four different scale feature maps are then convolved once and stacked to form the bottom-up channels of the Feature Pyramid Network (FPN). Then, the top-level feature map in the bottom-up channel is convolved again. The convolution result and the max pooling result after convolution are used as the top two layers P5 and P6 of the top-down channel. The high-level feature maps are then upsampled by a factor of 2 and fused with the corresponding feature layers of the same scale through lateral connections to obtain the outputs of the remaining layers P4, P3, and P2. Finally, the P2, P3, P4, P5, and P6 feature layers together form a five-layer top-down feature pyramid, completing the feature extraction of the original image.

[0072] S1.2 After image feature extraction is completed in the image feature extraction module, the five feature layers P2, P3, P4, P5, and P6 are input into the RPN network of the region proposal module for the next step of obtaining proposal boxes.

[0073] S1.3 Input the feature map of the suggestion box region into the prediction regression module for category detection and image segmentation.

[0074] S2. Based on the gray-level histogram, first-order feature extraction is performed on the segmented infrared image of the high-voltage switchgear in S1. The first-order statistical features used based on the gray-level histogram are average gray intensity, variance, skewness, kurtosis, entropy, and maximum gray intensity. Based on the gray-level co-occurrence matrix, second-order feature extraction is performed on the segmented infrared image of the high-voltage switchgear in S1. The second-order statistical features used based on the gray-level co-occurrence matrix are contrast, energy, correlation, and homogeneity.

[0075] S3. The basic principles of principal component analysis are as follows: Figure 4 As shown, the process first transforms the originally correlated high-dimensional data features into a new set of uncorrelated features of the same dimension through orthogonal transformation. Then, analysis of variance is performed on the new features, and they are sorted from highest to lowest according to their variance contribution rate. The feature with the highest variance contribution is selected as the first principal component, the second highest as the second principal component, and so on. Finally, the high-ranking second principal component features with a cumulative variance contribution rate of over 90% are retained, completing the data dimensionality reduction and reconstruction process. The specific steps are as follows: S3.1. Perform PCA data dimensionality reduction and reconstruction on the 10 first-order and second-order features extracted based on the gray-level histogram and gray-level co-occurrence matrix, and select appropriate low-dimensional statistical features based on the results of principal component analysis. S3.2. Reconstruct the features of the infrared image of the high-voltage switchgear using the feature vector of the selected low-dimensional statistical features; S4. Perform binary classification on the feature data of all samples, and then continue binary classification on the two sub-class nodes until all nodes have only one unique class. Therefore, for a multi-class classification task with k classes, this algorithm essentially uses k-1 classifiers to transform it into a k-1 binary classification task. The classification process of the high-voltage switchgear thermal defect identification task of this invention is as follows: Figure 5 As shown, there are a total of 4 classification tasks, so 3 classifiers are used. LSSVM1, LSSVM2 and LSSVM3 are the 3 classifiers used. Categories 1, 2, 3 and 4 represent normal state, poor contact of terminal block, incomplete closing of knife switch and local contamination of insulator, respectively.

[0076] Example 3 In addition, such as Figure 6 As shown, the present invention also discloses a high-voltage switchgear thermal defect identification system corresponding to the high-voltage switchgear thermal defect identification method in Embodiments 1 and 2, comprising: An image segmentation module is used to acquire infrared images of high-voltage disconnect switches and perform image segmentation on the acquired infrared images of high-voltage disconnect switches based on a Mask R-CNN model optimized by transfer learning. The feature extraction module is used to extract first-order and second-order features from the segmented infrared image of the high-voltage disconnector based on the gray-level histogram and gray-level co-occurrence matrix, respectively. A feature filtering module is used to filter the extracted first-order and second-order features based on principal component analysis. The identification decision module is used to identify thermal defects in high-voltage switchgear by taking the filtered first-order features and second-order features as inputs and using a least-squares support vector machine.

[0077] The specific operations performed by each module of the high-voltage switchgear thermal defect identification system are the same as the corresponding steps in the high-voltage switchgear thermal defect identification methods in Embodiments 1 and 2.

[0078] Additionally, a schematic diagram of a terminal device according to an embodiment of the present invention is provided. This terminal device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.

[0079] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0080] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0081] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0082] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0083] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0084] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying thermal defects in high-voltage switchgear, characterized in that, Includes the following steps: Step 1: Acquire infrared images of high-voltage disconnect switches and perform image segmentation on the infrared images of high-voltage disconnect switches based on a Mask R-CNN model optimized by transfer learning; Step 2: Based on the gray-level histogram and gray-level co-occurrence matrix, feature extraction is performed on the segmented infrared image of the high-voltage disconnector to obtain first-order and second-order features, respectively; Step 3: Perform feature filtering on the extracted first-order and second-order features based on principal component analysis; Step 4: Input the filtered first-order and second-order features into the classification model based on least squares support vector machine to identify thermal defects in high-voltage switchgear.

2. The method for identifying thermal defects in high-voltage switchgear according to claim 1, characterized in that, In step 1, the transfer learning optimization process of the Mask R-CNN model is as follows: the Mask R-CNN model is pre-trained using a benchmark dataset, and the weight parameters obtained from the pre-training are transferred and shared to the image segmentation model of high-voltage switchgear, thereby completing the optimization of the Mask R-CNN model.

3. A method for identifying thermal defects in high-voltage switchgear according to claim 2, characterized in that, The Mask R-CNN model includes an image feature extraction module, a region proposal module, and a prediction and regression module; The image feature extraction module consists of a backbone network based on the fusion of residual neural network and feature pyramid network, which is used to mine and extract feature information from infrared images of high-voltage switchgear. The region suggestion module consists of a region suggestion network and a region of interest calibration module, and is used to obtain suggestion boxes for the region of interest. The prediction regression module consists of bounding box branches, category branches, and mask branches, and is used to complete the final category regression and segmentation tasks.

4. The method for identifying thermal defects in high-voltage switchgear according to claim 3, characterized in that, The image feature extraction module is constructed as follows: First, two basic modules integrated into the residual learning structure are set in the residual neural network, namely a convolutional block and an identity block; then, the convolutional block and the identity block are used to build the backbone structure of the residual neural network; finally, the constructed residual neural network is combined with the feature pyramid network to complete the construction of the image feature extraction module of the Mask R-CNN model. The region proposal module consists of a region proposal network and a region of interest calibration module. It is used to obtain proposal boxes for regions of interest. Specifically, it first uses the coordinates of each point in the feature map obtained by the image feature module as the center and presets three prior boxes with different heights and widths. Then, the region proposal network obtains a correction bias by calculating regression, and uses the correction bias to optimize and correct the center coordinates, height, and width of each prior box to obtain a new batch of prior boxes. Finally, the region of interest proposal boxes are obtained through calibration by the region of interest calibration module. The prediction regression module consists of a bounding box branch, a category branch, and a mask branch, and is used to complete the final category regression and segmentation tasks. Specifically, the region proposal module inputs the acquired feature map into a fully connected layer network, outputs the category label of the high-voltage switchgear through the category regression branch, and outputs the bounding box of the target region through the bounding box regression branch. At the same time, the feature map is input into a fully convolutional network. After extracting deep semantic features through convolution layer by layer, the feature map is upsampled through a deconvolution layer to restore the size of the feature map to be consistent with the input image. A mask is generated pixel by pixel in the upsampled feature map to complete the image segmentation.

5. The method for identifying thermal defects in high-voltage switchgear according to claim 1, characterized in that, The first-order feature extraction process uses first-order statistical features based on gray-level histograms, including average gray intensity, variance, skewness, kurtosis, entropy, and maximum gray intensity.

6. The method for identifying thermal defects in high-voltage switchgear according to claim 1, characterized in that, The second-order feature extraction process uses second-order statistical features based on the gray-level co-occurrence matrix, including contrast, energy, correlation, and homogeneity.

7. The method for identifying thermal defects in high-voltage switchgear according to claim 1, characterized in that, The feature selection based on principal component analysis for the extracted first-order and second-order features is as follows: First, the originally correlated high-dimensional data features are transformed into uncorrelated features of the same dimension through orthogonal transformation. Then, variance analysis is performed on the uncorrelated features of the same dimension, and each feature is sorted from high to low according to its variance contribution rate. The feature with the highest variance contribution is selected as the first principal component, the second highest as the second principal component, and so on. Finally, the high-ranking principal component features with a cumulative variance contribution rate greater than a threshold are retained, thus completing the feature selection of first-order and second-order features.

8. The method for identifying thermal defects in high-voltage switchgear according to claim 1, characterized in that, In step S4, the regression function of the classification model based on least squares support vector machine is: In the formula, The coefficients corresponding to each sample, For bias terms, Kernel function, For each training sample, Value to be predicted The number of samples.

9. A thermal defect identification system for high-voltage switchgear, characterized in that, include: An image segmentation module is used to acquire infrared images of high-voltage disconnect switches and perform image segmentation on the acquired infrared images of high-voltage disconnect switches based on a Mask R-CNN model optimized by transfer learning. The feature extraction module is used to extract first-order and second-order features from the segmented infrared image of the high-voltage disconnector based on the gray-level histogram and gray-level co-occurrence matrix, respectively. A feature filtering module is used to filter the extracted first-order and second-order features based on principal component analysis. The identification decision module is used to input the filtered first-order features and second-order features into a classification model based on least squares support vector machine to identify thermal defects in high-voltage switchgear.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 8.

Images