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

Abstract

The application relates to the technical field of power equipment state detection, in particular to a pollution degree classification method for insulators based on a double-attention multi-scale enhancement network, which comprises the following steps: preparing ceramic insulator samples of different pollution levels, collecting hyperspectral image data of the samples, and constructing a double-attention multi-scale enhancement network framework model, wherein the model comprises a double-branch spectral enhancement module and a multi-scale spatial-spectral feature extraction module; a local region centered on a target pixel is extracted from the hyperspectral image data as network input, spectral correlation features are captured through the double-branch spectral enhancement module, multi-scale pollution region features and channel attention features are extracted through the multi-scale spatial-spectral feature extraction module and are fused; and a global average pooling and a full connection layer are used to output ceramic insulator pollution level classification results. The application realizes high-precision pixel-level classification of the pollution degree of ceramic insulators and is suitable for monitoring and evaluation of the pollution state of ceramic insulators in actual power systems.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology. Background Technology

[0002] Insulators are critical components of power distribution and transmission systems. Accumulation of contaminants on their surfaces can easily trigger flashover accidents, seriously threatening the stability of power supply. Traditional insulator contamination detection methods, such as salt density testing, suffer from low efficiency and slow response, making them unsuitable for meeting the real-time monitoring needs of insulators in large-scale transmission lines.

[0003] Hyperspectral images possess dozens to hundreds of spectral bands, revealing almost continuous spectral characteristic curves of materials. They contain both one-dimensional spectral information and two-dimensional spatial information, enabling sensitive capture of subtle differences in the composition of pollution layers and providing a new approach for classifying insulator pollution levels. However, traditional hyperspectral image classification generally relies on manual classification or machine learning to extract shallow features, resulting in problems such as "different spectra for the same object, and the same spectrum for different objects," leading to low classification accuracy. Current applications of deep learning in hyperspectral image classification are mostly concentrated in areas such as ground feature recognition and pest detection, with limited research on pixel-level classification of insulator pollution levels. Furthermore, existing deep learning methods suffer from insufficient spectral feature extraction and inadequate utilization of spatial multi-scale features, making it difficult to accurately distinguish ceramic insulators with different pollution levels.

[0004] Therefore, there is an urgent need for a method to classify the degree of contamination of ceramic insulators that can fully extract the spatial-spectral features of hyperspectral images and achieve high-precision pixel-level classification, so as to improve the efficiency and accuracy of insulator contamination detection. Summary of the Invention

[0005] To overcome the problems of insufficient spectral feature extraction, inadequate utilization of spatial multi-scale features, and difficulty in accurately distinguishing ceramic insulators with different pollution levels, this invention provides an insulator pollution level classification method based on a dual-attention multi-scale enhancement network.

[0006] The technical solution adopted by this invention to achieve the above objectives is: a method for classifying the pollution level of insulators based on a dual-attention multi-scale enhancement network, comprising the following steps:

[0007] 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.

[0008] 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.

[0009] 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. ;

[0010] 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. ;

[0011] 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. ;

[0012] 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.

[0013] Preferably, 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, thus obtaining ceramic insulator samples of different pollution levels.

[0014] Preferably, in step S2, a hyperspectral image acquisition platform is built, including a hyperspectral camera, a light source box and a correction whiteboard; the prepared ceramic insulator sample is placed on the acquisition platform, hyperspectral images of the sample are acquired, the pollution level corresponding to different pixels in the hyperspectral image is labeled, and the sample is established as a dataset.

[0015] Preferably, in step S3, local features of the insulator samples are captured through preliminary convolution operations, and the features are output. :

[0016] ;

[0017] in, This is a 2D convolution operation. Convolution weights For bias parameters, and These are batch normalization and ReLU activation function operations, respectively.

[0018] Preferably, in step S4, for the input features Perform max pooling and average pooling calculations:

[0019] ;

[0020] ;

[0021] in For position The value at channel c, It is a scaling factor;

[0022] Generate channel-level weights and filter key channels:

[0023] ;

[0024] in, It is the Sigmoid activation function. For matrix multiplication, The linear layer weight matrix, Shape , Shape ;

[0025] Generate cross-channel correlation features and fuse information between channels:

[0026] ;

[0027] Cascaded processing of channel feature attention information in the two branches using 1D convolution is used to capture channel relationships with different semantic meanings:

[0028] ;

[0029] in, This is a 1-D convolution operation. For splicing operations, This is the activation function for SigmoidLinearUnit (SiLU).

[0030] Through cross-channel attention features Input features Matrix multiplication is used to obtain channel attention-weighted features. :

[0031] .

[0032] Preferably, in step S5, the multi-scale spatial spectrum feature extraction module adopts a dual-branch structure, divided into an uplink branch and a downlink branch. In the uplink branch, the input features are processed... Perform channel reduction operation:

[0033] ;

[0034] in, and For 2D convolution parameters and biases, The number of input channels for 2D convolution. 2. Number of output channels after convolution;

[0035] 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:

[0036] ;

[0037] 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;

[0038] Integrate feature information at different scales, and simultaneously perform residual linking with the initial features:

[0039] ;

[0040] ;

[0041] 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:

[0042] ;

[0043] ;

[0044] Multiplication operations are used to fuse multi-scale spatial features with channel attention weights:

[0045] .

[0046] Preferably, in step S6, The pollution level classification results of the insulators were obtained.

[0047] Preferably, the method further includes step S7, which involves dividing the dataset into a training set, a validation set, and a test set; using the training set to train the dual-attention multi-scale augmentation network framework model; after training is complete, using the validation set to adjust the model parameters; using the test set to evaluate the model performance; and adjusting the parameters based on the validation and test results to obtain the trained model.

[0048] The beneficial effects of this invention are as follows:

[0049] This invention achieves high-precision pixel-level classification of the pollution level of ceramic insulators. By using a dual attention mechanism and multi-scale feature fusion, it effectively solves the problem of "different spectra for the same object and the same spectra for different objects", improves the classification robustness in complex scenarios, and improves the accuracy and kappa coefficient compared with traditional methods. It is suitable for monitoring and evaluating the pollution status of ceramic insulators in actual power systems. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the overall architecture of the dual-attention multi-scale enhancement network according to an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the dual-branch spectral enhancement module according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the multi-scale spatial spectrum feature extraction module according to an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the spectral curve of the light source box according to an embodiment of the present invention;

[0055] Figure 6 This is a sample area of ​​a polluted insulator and its corresponding pseudo-color map according to an embodiment of the present invention. Detailed Implementation

[0056] Example 1:

[0057] Embodiments of the present invention provide a method for classifying the pollution level of insulators based on a dual-attention multi-scale enhancement network, such as... Figure 1 As shown, it includes the following steps:

[0058] S1. Preparation of ceramic insulator samples:

[0059] Based on GB / T4585-2024 "Artificial Pollution Tests for High-Voltage Porcelain and Glass Insulators for AC Systems" and Q / GDW1152.1-2014 "Standards for Pollution Classification and External Insulation Selection in Power Systems Part 1: AC Systems", the salt density range corresponding to different pollution levels of ceramic insulators is determined. In this embodiment, the pollution levels of ceramic insulators are divided into the following five levels according to different salt densities (ESDD): Level A: 0.025 mg / cm², Level B: 0.025~0.05 mg / cm², Level C: 0.05~0.1 mg / cm², Level D: 0.1~0.25 mg / cm², and Level E: >0.25 mg / cm².

[0060] Sodium chloride and kaolin were weighed using an electronic balance with a ratio of 0.01:1 and mixed at a ratio of 1:20. The required mass of mixture for each pollution level was calculated based on the surface area of ​​the insulator. The required amount of sodium chloride and kaolin was weighed and added to an appropriate amount of pure water and stirred. The mixture was then applied evenly to the surface of the ceramic insulator disc using a quantitative brushing method. After the coating was completed, the insulator was placed in a 20°C room temperature environment to air dry naturally for 48 hours to obtain ceramic insulator samples of different pollution levels.

[0061] S2. Acquire hyperspectral image data:

[0062] A hyperspectral image acquisition platform was constructed, including a hyperspectral camera (Figspec FS2X), a light source box, and a correction whiteboard. The hyperspectral camera has a spectral range of 400-1000 nm, a spatial resolution of 1920×1920, and a spectral resolution of 2.5 nm. The light source box was turned on, and data was acquired within the 410 nm-870 nm band range based on its spectral curve, ensuring the acquired spectral data could effectively distinguish different pollution components. The ceramic insulator samples prepared in step S1 were placed on the acquisition platform, and the distance between the camera and the samples was adjusted to ensure image clarity. Hyperspectral images of the samples were acquired; each pixel in these images contains spectral data in 387 bands (410-870 nm), reflecting the material composition of the pixel (e.g., polluted or clean insulators). ENVIclassic software was used to precisely divide the acquired hyperspectral images into regions, labeling the pollution levels corresponding to different pixels. After labeling, the samples were compiled into a dataset.

[0063] S3. Establish a dual-attention multi-scale augmentation network framework model, such as Figure 2 As shown, it includes 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, while the multi-scale spatial spectral feature extraction module is used to extract the rich spectral features extracted by the dual-branch spectral enhancement module in a more refined manner and to identify the features of pollution regions at different scales. While capturing the spectral features of different pollution levels, it also captures the local spatial relationships.

[0064] Extract a 7×7 local region centered on the target pixel from the dataset established in step S2 as network input. After preliminary convolution operations, capture the local features of the insulator samples and output the features. :

[0065] ;

[0066] in, This is a 2D convolution operation. Convolution weights For bias parameters, and These are batch normalization and ReLU activation function operations, respectively.

[0067] Features Input the dual-branch spectral enhancement module.

[0068] S4. Based on a dual-attention multi-scale augmentation network framework model, develop such as Figure 3 The dual-branch spectral enhancement module shown has clearly defined characteristics. :

[0069] The dual-branch spectral enhancement module captures spectral features from both global trend features and local salient features through a dual-branch mode, enabling refined modeling of spectral features at different pollution levels.

[0070] For input features Perform max pooling and average pooling calculations:

[0071] ;

[0072] ;

[0073] in For position The value at channel c. Feature compression and reconstruction are performed through a linear layer to generate spectral attention weights. The scaling factor was determined to be [value] after parameter tuning experiments and optimization analysis. The value is 4. Global channel feature attention information is generated through average pooling and max pooling to capture spectral data from different aspects. The average pooling channel feature attention information represents the global spectral response trend, while the max pooling channel feature attention information represents local salient features.

[0074] Generate channel-level weights and filter key channels:

[0075] ;

[0076] in, It is the Sigmoid activation function. For matrix multiplication, The linear layer weight matrix, Shape , Shape ;

[0077] Generate cross-channel correlation features and fuse information between channels:

[0078] ;

[0079] By using weighted filtering and cross-channel correlation, the model can accurately capture subtle spectral differences between different levels of contamination, providing core spectral discrimination criteria for subsequent calculations.

[0080] Cascaded processing of channel feature attention information in the two branches using 1D convolution is used to capture channel relationships with different semantic meanings:

[0081] ;

[0082] in, This is a 1-D convolution operation. For splicing operations, The activation function is SigmoidLinearUnit (SiLU). Compared to ReLU, SiLU allows negative values, thus preserving more information and being smoother. This method effectively captures cross-channel attention features with very few parameters.

[0083] Through cross-channel attention features Input features Matrix multiplication is used to obtain channel attention-weighted features. :

[0084] ;

[0085] By extracting multi-branch spectral features, the model's ability to capture spectral information is enhanced, allowing the model to focus on spectral bands that are useful for waste classification and weakening irrelevant noise.

[0086] S5. Based on a dual-attention multi-scale augmentation network framework model, develop such as Figure 4 The multi-scale spatial spectrum feature extraction module shown clearly defines the features. :

[0087] The multi-scale spatial spectral feature extraction module is used to extract more refined spectral features from the dual-branch spectral enhancement module, and to identify features of polluted areas at different scales. While capturing the spectral features of different levels of pollution, it also captures local spatial relationships.

[0088] The multi-scale spatial spectral feature extraction module adopts a dual-branch structure, consisting of an uplink branch and a downlink branch. In hyperspectral classification tasks, insulator aging can not only be determined by spectral information, but also usually depends on the information of the surrounding area. Therefore, the uplink branch extracts multi-scale local features.

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

[0090] ;

[0091] in, and For 2D convolution parameters and biases, The number of input channels for 2D convolution. 2. The number of output channels after convolution was determined experimentally. .

[0092] 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:

[0093] ;

[0094] in To divide the feature map into four sub-feature maps along the channel dimension, the convolutional kernel size in this embodiment is determined to be 1, 3, 5, and 7. and This represents the weights and biases of convolutional kernels with sizes of 1, 3, 5, and 7. For the first The convolution outputs a sub-feature map; capturing multi-scale spatial information helps the model understand local and global information in the image.

[0095] The output features are concatenated along the channel dimension to integrate feature information at different scales. Simultaneously, residual links are established with the initial features to maintain the richness of multi-scale features while supplementing information and preventing model degradation.

[0096] ;

[0097] ;

[0098] Finally, 2D convolution is used to further enhance the spatial feature representation capability.

[0099] In the downlink branch of the multi-scale spatial spectral feature extraction module, channel-dimensional attention extraction is implemented through a lightweight design. First, global average pooling is performed on the input features. Then, a one-dimensional convolution with a kernel size of 3 is used to capture the relationships between channels. Finally, the channel weights are generated through the Sigmoid function. The calculation process is as follows:

[0100] ;

[0101] ;

[0102] The entire downlink branch introduces only a small number of parameters to fully capture channel features and improve the model's feature representation capabilities.

[0103] Multiplication operations are used to fuse multi-scale spatial features with channel attention weights to obtain spatial fusion features. :

[0104] .

[0105] This method improves the model's ability to capture key features and enhances the utilization of spatial-spectral joint features, resulting in "spectral + spatial fusion features" for each pixel, which include both compositional and locational distribution information.

[0106] S6. Extract a 7×7 local region centered on the target pixel from the dataset established in step S2 as network input. After preliminary convolution operations, capture the local features of the insulator samples and output the features. Capture features Mid-spectral correlation characteristics The formula is:

[0107] ;

[0108] in, This refers to the calculations performed in the dual-branch spectral enhancement module in step S4.

[0109] Then the features Input is fed into the multi-scale spatial spectrum feature extraction module, and the output features are... :

[0110] ;

[0111] in, This is the operation in the multi-scale spatial spectral feature extraction module in step S5, used to extract multi-scale pollution region features and refine spectral features.

[0112] Using residual structures to reuse features yields:

[0113] ;

[0114] The residual structure is used to prevent information loss during the operation and ensure that features are not lost after the module is processed; 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 filth levels) are built to determine the classification results and output format.

[0115] S7. Divide the dataset into training, validation, and test sets. Use the training set to train the dual-attention multi-scale augmentation network framework model. After training, use the validation set to adjust the model parameters and the test set to evaluate the model performance. Adjust the parameters based on the validation and test results to obtain a trained model used for classifying the pollution level of ceramic insulators. The model assigns a pollution level to each pixel in the insulator image to be detected, determining the spectral information of a single pixel (to distinguish pollution components) and its surrounding spatial information (to distinguish pollution distribution). The algorithm combines these two information to accurately match the preset pollution level standard. In this embodiment, 5 pollution levels are set, so each pixel will get 5 probability values. The model selects the level with the highest probability as the final classification result for that pixel. For example, if the probability of a pixel being "Level C" is 98%, then the pixel is determined to belong to Level C. Combining the classification results of all pixels forms a "pixel-level pollution level distribution map" of the entire ceramic insulator image, such as... Figure 6 As shown.

[0116] Example 2:

[0117] According to the latest national standard GB / T4585-2024 "Artificial Pollution Test for High Voltage Porcelain and Glass Insulators for AC Systems" and Q / GDW1152.1-2014 "Standard for Pollution Classification and External Insulation Selection of Power Systems Part 1 AC Systems", the pollution levels of ceramic insulators are divided into the following five levels according to different salt density (ESDD): Level A: 0.025 mg / cm², Level B: 0.025~0.05 mg / cm², Level C: 0.05~0.1 mg / cm², Level D: 0.1~0.25 mg / cm², and Level E: >0.25 mg / cm². Sodium chloride and kaolin are mixed in a ratio of 1:20, and the required amount of sodium chloride and kaolin is weighed using an electronic balance with a precision of 0.01% to ensure accuracy.

[0118] Polluted insulator samples were prepared using a quantitative coating method. First, based on the surface area of ​​the insulator samples and international standards, the required amounts of kaolin and sodium chloride for different pollution levels were calculated. The required sodium chloride and kaolin were weighed and added to an appropriate amount of purified water, then stirred and mixed. Finally, the prepared pollutant was evenly coated onto the surface of the insulator disc. Subsequently, the prepared polluted insulators were naturally air-dried at 20°C for 48 hours to obtain the final self-made polluted insulators. The specific salt density values ​​for the prepared insulators of different pollution levels are shown in the table below:

[0119]

[0120] Subsequently, hyperspectral image data of the self-made, contaminated ceramic insulators were acquired. The acquisition equipment mainly included a hyperspectral camera (Figspec FS2X), a light source box, and a correction whiteboard. The hyperspectral camera has a spectral range of 400-1000nm, including ultraviolet, visible, and infrared bands, with a spatial resolution of 1920*1920. The light source box was used to provide uniform illumination. According to... Figure 5 The spectral curves of the light source box shown were obtained by selecting data from 387 bands within the 410nm-870nm wavelength range for the experiment, with a spectral resolution of 2.5nm.

[0121] ENVIclassic was used to precisely divide different contaminated areas and create a dataset. For example... Figure 6 The images shown are pseudo-color images and real images of insulator samples at different pollution levels, each containing 524×481 pixels.

[0122] Overall accuracy (OA), average accuracy (AA), and Kappa coefficient were used as quantitative evaluation metrics for classification performance, and qualitative analysis was conducted in conjunction with visualization results. 20% of the samples from each category were randomly selected to form the training set, and the remaining samples were used as the test set. To minimize the influence of random factors, multiple experiments were conducted, and the optimal results were selected.

[0123] The dual-attention multi-scale augmentation network algorithm proposed in this embodiment is implemented in Python 3.7 and the PyTorch 1.13.1 deep learning framework, and validated based on this implementation. The configuration information is as follows: CPU is an AMD Ryzen™ CPU R7-7745HX 3.60GHz, GPU is an NVIDIA GeForce RTX 4060 laptop with 8GB of RAM. The Adam optimizer is used, with a batch size of 32, an initial learning rate of 3e-4, a feature cube size of 7×7×387, and a maximum training epoch of 100. To stabilize convergence in later training stages, exponential learning rate decay is adopted, with a decay factor for each epoch. The learning rate was initially set to 0.9, and then gradually decreased by 10% at each step. The experimental results are shown in the table below:

[0124]

[0125] The model captures complex spectral relationships through a bi-branch spectral enhancement module and models the relationships between different contamination locations. It achieves an overall classification accuracy of 95.81% and an average accuracy of 94.01% for samples of different contamination levels, with a kappa coefficient of 0.9381. The classification accuracy for contamination level 0 and contamination level BE both exceed 98%, demonstrating the model's accurate classification capability for different levels of contamination.

[0126] To further verify the effectiveness of the bi-branch spectral enhancement module and the multi-scale spatial spectral feature extraction module, an ablation experiment was designed, gradually introducing module components to analyze their impact on classification performance. In the ablation experiment, when the bi-branch spectral enhancement module was not used, a linear layer of the same depth and a 1d convolution were selected instead; when the multi-scale spatial spectral feature extraction module was not used, a 2d convolution of the same depth was used instead. The ablation experiment results are shown in the table below:

[0127]

[0128] The results show that introducing the dual-branch spectral enhancement module alone improved the overall accuracy by 19.86%, introducing the multi-scale spatial spectral feature extraction module alone improved the overall accuracy by 14.92%, and when both modules were introduced, the overall accuracy of the model improved by 24.01%, which confirms the effectiveness of the present invention in classifying the pollution level of insulators.

[0129] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

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. Specifically: 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: ； S6. Input the dataset into the dual-attention multi-scale enhancement network framework model. The dual-branch spectral enhancement module captures and processes the spectral features of different pollution levels, and the multi-scale spatial spectral feature extraction module extracts and fuses the multi-scale pollution region features and channel attention features. The residual structure is used to reuse the above features. ； Will After global average pooling and sequential processing of the fully connected layer, the pollution level classification results of the insulator are obtained.

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 built, including a hyperspectral camera, a light source box, and a correction whiteboard; the prepared ceramic insulator sample is placed on the acquisition platform, hyperspectral images of the sample are acquired, the pollution level corresponding to different pixels in the hyperspectral image is labeled, and the sample is established as 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. 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, 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