Insulator pollution degree classification method based on double-attention multi-scale enhancement network

Abstract

The invention relates to the technical field of power equipment state detection, in particular to an insulator pollution degree classification method based on a double-attention multi-scale enhancement network, which comprises the steps of preparing ceramic insulator samples of different pollution grades, collecting hyperspectral image data of the samples, constructing a double-attention multi-scale enhancement network framework model, and classifying the pollution degree of the ceramic insulator samples. The model comprises a double-branch spectrum enhancement module and a multi-scale spatial spectrum feature extraction module. A local area with a target pixel as the center is extracted from hyperspectral image data to serve as network input, spectrum related features are captured through a double-branch spectrum enhancement module, and multi-scale dirty area features and channel attention features are extracted and fused through a multi-scale spatial spectrum feature extraction module; and outputting a ceramic insulator pollution grade classification result by using global average pooling and a full connection layer. According to the method, high-precision pixel-level classification of the pollution degree of the ceramic insulator is realized, and the method is suitable for monitoring and evaluating the pollution state of the ceramic insulator of an actual power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state detection. BACKGROUND

[0002] The insulator is a key component of the power transmission system of the distribution network, and the surface contamination accumulation of the insulator is easy to cause a pollution flashover accident, which seriously threatens the stability of power supply. The traditional insulator contamination detection method, such as salt density testing, has defects such as low efficiency and response lag, and it is difficult to meet the real-time detection needs of large-scale power transmission lines.

[0003] The hyperspectral image has tens to hundreds of spectral bands, can show the almost continuous spectral characteristic curve of a substance, and contains one-dimensional spectral information and two-dimensional spatial information, which can sensitively capture the subtle component differences of the contamination layer and provide a new path for the contamination grade classification of the insulator. However, the traditional hyperspectral image classification generally relies on artificial classification or machine learning to extract shallow features, and has the problem of "same object different spectrum and different object same spectrum", and the classification accuracy is low. The existing deep learning is mostly applied to the field of ground object identification and pest detection, and there are few studies on pixel-level classification of insulator contamination grades, and there are problems of insufficient spectral feature extraction and insufficient use of spatial multi-scale features, which makes it difficult to accurately distinguish different contamination grades of ceramic insulators.

[0004] Therefore, there is an urgent need for a ceramic insulator contamination degree classification method which can fully extract the spatial and spectral features of the hyperspectral image and realize high-precision pixel-level classification, so as to improve the contamination detection efficiency and accuracy of the insulator. SUMMARY

[0005] In order to overcome the problems of insufficient spectral feature extraction, insufficient use of spatial multi-scale features and difficulty in accurately distinguishing different contamination grades of ceramic insulators, the present application provides an insulator contamination degree classification method based on a double-attention multi-scale enhancement network.

[0006] The technical scheme adopted by the present application to achieve the above-mentioned purpose is: an insulator contamination degree classification method based on a double-attention multi-scale enhancement network, comprising the following steps:

[0007] S1, setting the salt density range corresponding to the contamination grade of the insulator, mixing sodium chloride and kaolin in proportion and quantitatively coating on the surface of the insulator, and obtaining ceramic insulator samples of different contamination grades after drying;

[0008] S2, building a hyperspectral image acquisition platform, collecting hyperspectral images of the ceramic insulator samples, dividing the regions and labeling the contamination grades according to the image pixels, and establishing a data set;

[0009] S3, a dual-attention multi-scale enhancement network framework model is established, the model includes a dual-branch spectral enhancement module and a multi-scale spatial-spectral feature extraction module, a ceramic insulator sample hyperspectral image is selected from a data set as an input of the model, an operation is performed on the input data, and a feature is output .

[0010] S4, the feature is input into the dual-branch spectral enhancement module, spectral features of different contamination levels are captured, the spectral features of different contamination levels are processed through a channel attention mechanism, and a channel attention weighted feature is output .

[0011] S5, the feature is input into the multi-scale spatial-spectral feature extraction module, different scale contamination region features and local spatial relationships are captured, and a spatial fusion feature is obtained .

[0012] S6, the data set is input into the dual-attention multi-scale enhancement network framework model, and a contamination level classification result of the insulator is obtained through the dual-branch spectral enhancement module and the multi-scale spatial-spectral feature extraction module.

[0013] Preferably, in step S1, sodium chloride and kaolin are weighed and mixed in a ratio of 1:20, the required mixture mass of each contamination level is calculated according to the surface area of the insulator, the required sodium chloride and kaolin are weighed and added to a proper amount of pure water, stirred and mixed, coated on the surface of the ceramic insulator disc, and after coating, the insulator is placed in a room temperature environment for natural air drying, thereby obtaining ceramic insulator samples of different contamination levels.

[0014] Preferably, in step S2, a hyperspectral image acquisition platform is built, including a hyperspectral camera, a light source box and a correction white board; the prepared ceramic insulator samples are placed on the acquisition platform, the sample hyperspectral images are acquired, the contamination levels corresponding to different pixels in the hyperspectral images are labeled, and the samples are established as a data set.

[0015] Preferably, in step S3, local features of the insulator sample are captured through preliminary convolution operation, and a feature is output

[0016] .

[0017] wherein, is a 2-D convolution operation, is a convolution weight is a bias parameter, and are batch normalization and ReLU activation function operations, respectively.

[0018] Preferably, in step S4, for the input feature max-pooling and average-pooling calculation is performed:

[0019] ;

[0020] ;

[0021] wherein is a value of the position at the channel c, is a scale factor;

[0022] Channel-level weights are generated to filter key channels:

[0023] ;

[0024] wherein is a Sigmoid activation function, is a matrix multiplication, is a linear layer weight matrix, has a shape of , has a shape of ;

[0025] Cross-channel correlation features are generated to fuse inter-channel information:

[0026] ;

[0027] A 1D convolution is used to cascade the channel feature attention information of the double branches to capture the channel relationship of different semantics:

[0028] ;

[0029] wherein is a 1-D convolution operation, is a concatenation operation, is a Sigmoid Linear Unit (SiLU) activation function;

[0030] The cross-channel attention feature is multiplied by the input feature to obtain the channel attention weighted feature :

[0031] .

[0032] Preferably, in step S5, the multi-scale spectral feature extraction module adopts a double-branch structure, which is divided into an ascending branch and a descending branch. In the ascending branch, a channel reduction operation is performed on the input feature :

[0033] ;

[0034] wherein, with is a 2-D convolution parameter and bias, is a 2-D convolution input channel number, 2 is a 2-D convolution output channel number;

[0035] According to the channel size, the feature map is divided into a plurality of sub-feature maps along the channel, and multi-scale feature extraction is performed by using convolution kernels of different sizes:

[0036] ;

[0037] wherein, is divided into four sub-feature maps along the channel dimension, is the first sub-feature map output after convolution;

[0038] The different scale feature information is integrated, and residual connection is performed with the initial feature:

[0039] ;

[0040] ;

[0041] In the descending branch of the multi-scale spatial-spectral feature extraction module, the input feature is subjected to global average pooling, the inter-channel relationship is captured, and a channel weight is generated by using a Sigmoid function:

[0042] ;

[0043] ;

[0044] The multi-scale spatial feature and the channel attention weight are fused by using a multiplication operation:

[0045] .

[0046] Preferably, in step S6, , the insulation sub-pollution level classification result is obtained.

[0047] Preferably, the method further comprises the following step S7: dividing the data set into a training set, a validation set and a test set; using the training set, training the dual-attention multi-scale enhancement network framework model, using the validation set to adjust the model parameters, using the test set to evaluate the model performance, adjusting the parameters according to the results of the validation and the test, and obtaining the trained model.

[0048] The present application has the following advantages:

[0049] The application realizes high-precision pixel-level classification of the contamination degree of ceramic insulators, effectively solves the problem of "same object different spectrum and different object same spectrum" through a double attention mechanism and multi-scale feature fusion, improves the classification robustness in complex scenes, and improves the accuracy and kappa coefficient compared with traditional methods, and is suitable for monitoring and evaluation of the contamination state of ceramic insulators in actual power systems. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a whole process schematic diagram of the embodiment of the application;

[0051] Figure 2 is a whole architecture schematic diagram of the double attention multi-scale enhancement network of the embodiment of the application;

[0052] Figure 3 is a double-branch spectral enhancement module schematic diagram of the embodiment of the application;

[0053] Figure 4 is a multi-scale space spectrum feature extraction module schematic diagram of the embodiment of the application;

[0054] Figure 5 is a light source box spectrum curve schematic diagram of the embodiment of the application;

[0055] Figure 6 is a contaminated insulator sample area and corresponding pseudo color chart of the embodiment of the application. DETAILED DESCRIPTION

[0056] Embodiment one:

[0057] The embodiment of the application provides a kind of insulator contamination degree classification method based on double attention multi-scale enhancement network, as shown in Figure 1 The method comprises the following steps:

[0058] S1, ceramic insulator sample is prepared:

[0059] According to GB / T4585-2024 "Artificial Contamination Test for High Voltage Porcelain and Glass Insulators for AC Systems" and Q / GDW1152.1-2014 "Power System Contamination Area Classification and External Insulation Selection Standard Part 1 AC System", the salt density range corresponding to different ceramic insulator contamination levels is determined. In this embodiment, the ceramic insulator contamination level is divided into the following five levels according to different salt density (ESDD): A level: 0.025 mg / cm², B level: 0.025~0.05 mg / cm², C level: 0.05~0.1 mg / cm², D level: 0.1~0.25 mg / cm², E level: >0.25 mg / cm².

[0060] The electronic balance of 1 / 10,000 is used to weigh sodium chloride and kaolin, which are mixed according to a ratio of 1:20. The required mixture mass of each pollution level is calculated according to the surface area of the insulator. The required sodium chloride and kaolin are weighed and added to a proper amount of pure water to mix by stirring. The quantitative brushing method is used to uniformly apply the mixture on the surface of the ceramic insulator disc. After the application is completed, the insulator is placed in a 20°C room temperature environment for natural air drying for 48 hours to obtain ceramic insulator samples of different pollution levels.

[0061] S2, collect hyperspectral image data:

[0062] A hyperspectral image acquisition platform is built, including a hyperspectral camera (Figspec FS2X), a light source box and a correction whiteboard. The spectral range of the hyperspectral camera is 400-1000 nm, the spatial resolution is 1920x1920, and the spectral resolution is 2.5 nm. The light source box is turned on. According to the spectral curve of the light source box, the waveband data in the waveband range of 410-870 nm is selected for acquisition, so as to ensure that the acquired spectral data can effectively distinguish different pollution components. The ceramic insulator samples prepared in step S1 are placed on the acquisition platform. The distance between the camera and the sample is adjusted to ensure clear images. The hyperspectral images of the samples are collected. Each pixel of these images contains spectral data of 387 wavebands (410-870 nm), which can reflect the material composition of the pixel (such as pollution or clean insulator). The collected hyperspectral images are accurately regionally divided using ENVI classic software. The pollution levels corresponding to different pixels are labeled. After labeling is completed, the sample is established as a data set.

[0063] S3, a dual-attention multi-scale enhancement network framework model is established, as shown in Figure 2 , including a dual-branch spectral enhancement module and a multi-scale spatial-spectral feature extraction module. The dual-branch spectral enhancement module is used to fully capture the spectral features of different pollution levels. The multi-scale spatial-spectral feature extraction module is used to extract more detailed features from the rich spectral features extracted by the dual-branch spectral enhancement module, and to identify features of different scale pollution regions. While capturing the spectral features of different pollution levels, the local spatial relationship is captured.

[0064] A 7x7 local region centered on the target pixel is extracted from the data set established in step S2 as the network input. After preliminary convolution operation, the local features of the insulator sample are captured, and the output feature :

[0065] ;

[0066] wherein, is a 2-D convolution operation, is a convolution weight is a bias parameter, and Batch normalization and ReLU activation function operation respectively;

[0067] The features Input the double-branch spectral enhancement module.

[0068] S4, based on the double-attention multi-scale enhancement network framework model, develops a double-branch spectral enhancement module as shown in Figure 3 The features :

[0069] The double-branch spectral enhancement module captures spectral features from two dimensions of global trend features and local significant features through a double-branch mode, achieving fine modeling of spectral features of different contamination levels.

[0070] For the input features , maximum pooling and average pooling calculations are performed:

[0071] ;

[0072] ;

[0073] Wherein is the value of the position in the channel c. Feature compression and feature reconstruction are performed through a linear layer to generate spectral attention weights, is a scale factor, which is finally determined as 4 through parameter tuning experiment optimization analysis. Global channel feature attention information is generated through average pooling and maximum pooling to capture different aspects of spectral data, wherein the average pooling channel feature attention information represents the global spectral response trend, and the maximum pooling channel feature attention information represents the local significant feature.

[0074] Channel-level weights are generated to filter key channels:

[0075] ;

[0076] Wherein, is a Sigmoid activation function, is matrix multiplication, is a linear layer weight matrix, has a shape of , has a shape of ;

[0077] Cross-channel correlation features are generated to fuse inter-channel information:

[0078] ;

[0079] Through weight screening and cross-channel association, the model can accurately capture the subtle spectral differences of different contamination levels, providing the core spectral discrimination basis for subsequent calculation.

[0080] The 1D convolution is used to cascade the channel feature attention information of the double-branch, capturing the channel relationship of different semantics:

[0081] ;

[0082] wherein, is a 1-D convolution operation, is a splicing operation, is a Sigmoid Linear Unit (SiLU) activation function. Compared with ReLU, SiLU allows negative values to exist, so it can retain more information and is smoother. This method effectively captures the cross-channel attention features with few parameters.

[0083] Through the matrix multiplication of the cross-channel attention features and the input features , the channel attention weighted features are obtained:

[0084] ;

[0085] Through multi-branch spectral feature extraction, the model's capture of spectral information is enhanced, focusing on spectral bands useful for contamination classification and weakening irrelevant noise.

[0086] Based on the double-attention multi-scale enhancement network framework model, a multi-scale spatial-spectral feature extraction module is developed as shown in Figure 4 , and the features are explicitly defined as :

[0087] The multi-scale spatial-spectral feature extraction module is used to extract more refined features from the rich spectral features extracted by the double-branch spectral enhancement module, and to identify features of different scale contamination regions. It captures the spectral features of different levels of contamination while capturing the local spatial relationship.

[0088] The multi-scale spatial-spectral feature extraction module adopts a double-branch structure, divided into an ascending branch and a descending branch. In the hyperspectral classification task, insulator aging can not only be distinguished by spectral information, but also usually depends on the information of the surrounding area. Therefore, the multi-scale local features are extracted in the ascending branch.

[0089] The input of the multi-scale spatial-spectral feature extraction module is the output of the double-branch spectral enhancement module . In order to reduce the overhead of calculation and storage, a 1x1 convolution operation is used to perform a channel reduction operation on the input features, and the calculation formula is as follows:

[0090] ;

[0091] wherein, and are 2-D convolution parameters and bias, is the number of 2-D convolution input channels, 2 is the number of convolution output channels, and .

[0092] According to the channel size, the feature map is divided into multiple sub-feature maps along the channel, and different size convolution kernels are used for multi-scale feature extraction:

[0093] ;

[0094] wherein is divided into four sub-feature maps along the channel dimension, and the convolution kernel size in the embodiment is determined as 1, 3, 5, and 7, and represent the weights and bias of the convolution kernel with a size of 1, 3, 5, and 7, is the i-th sub-feature map output by convolution, Capture multi-scale spatial information, which helps the model to understand the local and global information in the image.

[0095] The output features are spliced along the channel dimension, integrating different scale feature information, while the initial features are linked with residual, maintaining the richness of multi-scale features while supplementing information to prevent model degradation:

[0096] ;

[0097] ;

[0098] Finally, the spatial feature expression ability is further enhanced through 2-D convolution.

[0099] In the descending branch of the multi-scale spectral feature extraction module, the channel dimension attention is extracted through lightweight design. First, the input features are globally averaged pooled, then a one-dimensional convolution with a kernel size of 3 is used to capture the inter-channel relationship, and finally the channel weight is generated through the Sigmoid function, and the calculation process is as follows:

[0100] ;

[0101] ;

[0102] The entire descending branch only introduces a small number of parameters, fully captures the channel features, and improves the model feature representation ability.

[0103] The spatial fusion features are obtained by fusing the multi-scale spatial features and the channel attention weights through multiplication operation :

[0104] .

[0105] The method improves the capturing ability of the model for key features, enhances the utilization of spatial-spectral joint features, and obtains "spectrum + spatial fusion features" for each pixel, which contains component information and position distribution information.

[0106] S6, a 7x7 local region centered on the target pixel is extracted from the data set established in step S2 as network input, and the local features of the insulator sample are captured through preliminary convolution operation, and the output feature captures the features Spectral correlation features , the formula is:

[0107] ;

[0108] wherein, is the operation in the double-branch spectral enhancement module in step S4.

[0109] Then, the feature is input into the multi-scale spatial-spectral feature extraction module, and the output feature :

[0110] ;

[0111] wherein, is the operation in the multi-scale spatial-spectral feature extraction module in step S5, which is used to extract multi-scale dirty area features and refine spectral features.

[0112] The residual structure is used to reuse the features, and the following is obtained:

[0113] ;

[0114] The residual structure is used to prevent information loss during operation and ensure that the features after module processing are not lost; at the same time, a global average pooling layer (compressing the feature map into a vector) and a fully connected layer (outputting the probability of different dirty levels) are built to determine the classification result and output format.

[0115] S7, divide the data set into training set, validation set and test set; using the training set, the double attention multi-scale enhanced network framework model is trained, the training is completed, the model parameters are adjusted using the validation set, and the model performance is evaluated using the test set, the parameters are adjusted according to the results of validation and test, and the trained model is obtained, which is used for ceramic insulator contamination degree classification. The model will determine a contamination level for each pixel point in the detected insulator image, determine the spectral information of a single pixel (distinguish the contamination component) and the spatial information around it (distinguish the contamination distribution), combine the two through an algorithm, and accurately match the preset contamination level standard. In this embodiment, five contamination levels are set, so each pixel will get five probability values, and the model selects the highest probability level as the final classification result of the pixel. For example, the probability of a certain pixel "level C" is 98%, so the pixel belongs to level C. Combining the classification results of all pixels, a "pixel-level contamination level distribution map" of the entire ceramic insulator image is formed, as shown in Figure 6

[0116] Example two

[0117] According to the latest national standard GB / T4585-2024 "Artificial Contamination Test of High Voltage Porcelain and Glass Insulators for AC System" and Q / GDW1152.1-2014 "Power System Contamination Classification and External Insulation Selection Standard Part 1 AC System", the contamination level of ceramic insulator is divided into the following five levels according to different salt density (ESDD): A level: 0.025 mg / cm², B level: 0.025~0.05 mg / cm², C level: 0.05~0.1 mg / cm², D level: 0.1~0.25 mg / cm², E level: >0.25 mg / cm², sodium chloride and kaolin are mixed in a ratio of 1:20, and a one-hundredth electronic balance is used to weigh the required sodium chloride and kaolin to ensure accuracy.

[0118] A quantitative brushing method is used to prepare the contaminated insulator sample. First, according to the surface area size of the sample insulator to be prepared and the international standard, the required kaolin and sodium chloride for different contamination levels are calculated, and the required sodium chloride and kaolin are weighed and added to a suitable amount of pure water and stirred to mix. Finally, the prepared contamination is evenly coated on the surface of the insulator disc. Then, the prepared contaminated insulator is naturally air-dried at 20°C for 48 hours, and the final self-prepared contaminated insulator is obtained. The specific salt density values of the prepared insulators of different contamination levels are shown in the following table:

[0119]

[0120] ​Subsequently, the high-spectral image data of the self-made dirty ceramic insulator was collected, and the collection equipment mainly included a hyperspectral camera (Figspec FS2X), a light source box, and a correction whiteboard. The hyperspectral camera had a spectral range of 400-1000 nm, including ultraviolet, visible, and infrared bands, and a spatial resolution of 1920*1920. The light source box was used to provide uniform illumination. According to the spectral curve of the light source box as shown in Figure 5 , the data of 387 bands in the range of 410-870 nm were selected for the experiment, and the spectral resolution was 2.5 nm.

[0121] The ENVI classic was used to accurately divide different dirty areas and make data sets. As shown in the pseudo-color image and real image of the insulator sample with different dirty levels, the image contained 524*481 pixels. Figure 6

[0122] The overall accuracy (OA), average accuracy (AA), and Kappa coefficient were used as quantitative evaluation indicators of the classification performance, and the qualitative analysis was combined with the visual results. 20% of the samples in each category were randomly selected to form a training set, and the remaining samples were used for the test set. In order to minimize the influence of random factors, multiple experiments were conducted, and the best effect was selected.

[0123] The dual-attention multi-scale enhancement network algorithm proposed in this embodiment was implemented in Python 3.7 and PyTorch 1.13.1 deep learning framework, and was verified on this basis. The configuration information is as follows: the CPU is AMD Ryzen (TM) CPU R7-7745HX 3.60 GHz, and the GPU is 8 GB NVIDIA GeForce RTX 4060 laptop. The Adam optimizer was used, the batch size was set to 32, the learning rate was initialized to 3e-4, the feature cube size obtained by the model was 7*7*387, and the maximum training period was set to 100. In order to stabilize the convergence in the later training stage, the exponential learning rate decay was adopted, the decay factor of each period was 0.9, and the learning rate was reduced by 10% at each step. The experimental results are shown in the following table:

[0124]

[0125] The model captures complex spectral relationships through the dual-branch spectral enhancement module, and classifies samples of different dirty levels by modeling the relationship between different dirty locations. The overall accuracy of the model reached 95.81%, the average accuracy reached 94.01%, and the kappa coefficient was 0.9381. The classification accuracy of dirty levels 0 and B-E all exceeded 98%, which showed the accurate classification ability of the model for different dirty levels.

[0126] ​In order to further verify the effectiveness of the double-branch spectral enhancement module and the multi-scale spatial-spectral feature extraction module, an ablation experiment is designed to gradually introduce the module components to analyze their influence on the classification effect. In the ablation experiment, when the double-branch spectral enhancement module is not used, linear layers and 1d convolution of the same depth are selected instead; when the multi-scale spatial-spectral feature extraction module is not used, 2d convolution of the same depth is used instead. The ablation experiment results are shown in the following table:

[0127]

[0128] The results show that the overall accuracy is improved by 19.86% when the double-branch spectral enhancement module is introduced alone, the overall accuracy is improved by 14.92% when the multi-scale spatial-spectral feature extraction module is introduced alone, and the overall accuracy of the model is improved by 24.01% when both modules are introduced, which verifies the effectiveness of the application in insulator contamination degree classification.

[0129] The application is described through embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to these features and embodiments without departing from the spirit and scope of the application. In addition, under the guidance of the application, these features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the application. Therefore, the application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application belong to the protection scope of the application.

Claims

1. A method for classifying the pollution level of insulators based on a dual-attention multi-scale reinforcement network, characterized in that, Includes the following steps: S1. Set the salt density range corresponding to the pollution level of the insulator, mix sodium chloride and kaolin in proportion and apply them quantitatively to the surface of the insulator, and obtain ceramic insulator samples with different pollution levels after air drying. S2. Build a hyperspectral image acquisition platform, acquire hyperspectral images of ceramic insulator samples, divide the regions and label the pollution level according to the image pixels, and establish a dataset. S3. Establish a dual-attention multi-scale enhancement network framework model. This model includes a dual-branch spectral enhancement module and a multi-scale spatial-spectral feature extraction module. Hyperspectral images of ceramic insulator samples are selected from the dataset as model input. Operations are performed on the input data, and the output features are determined. ; S4, Features The input is a dual-branch spectral enhancement module that captures spectral features at different levels of contamination. These features are then processed using a channel attention mechanism to output channel attention-weighted features. ; S5, Features The multi-scale spatial spectrum feature extraction module is used to capture the features of polluted areas at different scales and their local spatial relationships, resulting in spatial fusion features. ; S6. Input the dataset into the dual-attention multi-scale enhancement network framework model, and process it through the dual-branch spectral enhancement module and the multi-scale spatial spectral feature extraction module to obtain the insulator pollution level classification results.

2. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, In step S1, sodium chloride and kaolin are weighed and mixed in a ratio of 1:

20. Based on the surface area of ​​the insulator, the required mass of the mixture for each pollution level is calculated. The required amount of sodium chloride and kaolin is weighed and added to an appropriate amount of pure water, stirred and mixed, and then applied to the surface of the ceramic insulator disc. After the coating is completed, the insulator is placed in a room temperature environment to air dry naturally, and ceramic insulator samples of different pollution levels are obtained.

3. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, In step S2, a hyperspectral image acquisition platform is set up, including a hyperspectral camera, a light source box and a correction whiteboard; the prepared ceramic insulator samples are placed on the acquisition platform, hyperspectral images of the samples are acquired, the pollution levels corresponding to different pixels in the hyperspectral images are labeled, and the samples are built into a dataset.

4. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, In step S3, local features of the insulator samples are captured through preliminary convolution operations, and the features are output. : ； in, This is a 2D convolution operation. Convolution weights For bias parameters, and These are batch normalization and ReLU activation function operations, respectively.

5. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, In step S4, for the input features Perform max pooling and average pooling calculations: ； ； in For position The value at channel c, It is a scaling factor; Generate channel-level weights and filter key channels: ； in, It is the Sigmoid activation function. For matrix multiplication, The linear layer weight matrix, Shape , Shape ; Generate cross-channel correlation features and fuse information between channels: ； Cascaded processing of channel feature attention information in the two branches using 1D convolution is used to capture channel relationships with different semantic meanings: ； in, This is a 1-D convolution operation. For splicing operations, This is the activation function for SigmoidLinearUnit (SiLU). Through cross-channel attention features Input features Matrix multiplication is used to obtain channel attention-weighted features. : 。 6. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, In step S5, the multi-scale spatial spectrum feature extraction module adopts a dual-branch structure, consisting of an uplink branch and a downlink branch. In the uplink branch, the input features are processed... Perform channel reduction operation: ； in, and For 2D convolution parameters and biases, The number of input channels for 2D convolution.

2. Number of output channels after convolution; Based on the channel size, the feature map is segmented into multiple sub-feature maps along the channel, and multi-scale feature extraction is performed using convolutional kernels of different sizes: ； in, The feature map is divided into four sub-feature maps along the channel dimension. For the first A sub-feature map output by convolution; Integrate feature information at different scales, and simultaneously perform residual linking with the initial features: ； ； In the downlink branch of the multi-scale spatial spectrum feature extraction module, the input features are processed... Perform global average pooling, capture the relationships between channels, and generate channel weights using the Sigmoid function: ； ； Multiplication operations are used to fuse multi-scale spatial features with channel attention weights: 。 7. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, In step S6, The pollution level classification results of the insulators were obtained.

8. The insulator pollution degree classification method based on dual-attention multi-scale enhancement network according to claim 1, characterized in that, It also includes step S7, which divides the dataset into a training set, a validation set, and a test set; using the training set, the dual-attention multi-scale augmentation network framework model is trained; after training is completed, the model parameters are adjusted using the validation set, and the model performance is evaluated using the test set; the parameters are adjusted according to the results of validation and testing to obtain the trained model.

Images