A Deep Learning-Based Method for Insulator Icing Detection and De-icing

Abstract

The application discloses to the technical field of insulators, and particularly relates to an insulator icing detection and deicing method based on deep learning, which comprises the following specific steps: S1: constructing an insulator image data set and pre-processing the images in the data set; S2: constructing an insulator image segmentation model for segmenting the insulators present in the images; S3: training the insulator image segmentation model; S4: applying the insulator image segmentation model to calculate the icing degree of the insulators; and S5: monitoring the icing degree of the insulators and performing deicing treatment on the insulators with serious icing. The application effectively improves the efficiency of insulator icing detection, greatly reduces the manpower required for detection, and reduces the influence of environmental factors on detection.

Description

Technical Field

[0001] This invention relates to the field of insulator technology, specifically to a method for detecting and melting ice accretion on insulators based on deep learning. Background Technology

[0002] my country's power grid has a vast coverage area, and a large-scale high-voltage transmission network has been built. At the same time, the country is vigorously developing and constructing ultra-high-voltage AC and DC transmission and transformation projects, which represent the highest technological level in the world to date. Therefore, transmission line icing, as a major cause of flashover accidents, has significant research and practical implications. Numerous statistics, including those from the State Grid Corporation of China, show that icing of transmission networks can lead to accidents such as tower collapse, conductor breakage, and insulator failure.

[0003] The main measures for preventing and controlling icing on power transmission lines are "avoidance, resistance, modification, prevention, and melting." "Avoidance" means designing transmission lines to avoid heavily icing areas; "resistance" means improving the icing resistance of transmission lines; "modification" means upgrading lines with weak icing resistance; "prevention" means adopting relevant measures to prevent icing on transmission lines; and "melting" means using alternating current or direct current to melt the ice. For newly constructed transmission lines, anti-icing and melting measures are mainly used; for existing old transmission lines, modification, anti-icing, and melting measures are mainly used; and for planned new lines, avoidance and anti-icing measures are mainly used. Although there are many methods for dealing with power grid icing disasters, they can ultimately be summarized into three technical means: de-icing, melting, and prevention. Among these, the most commonly used melting technology in power grids is direct current (DC) melting technology.

[0004] In high-voltage transmission lines, insulators play a crucial role. They primarily function as insulation and support, effectively isolating conductors from towers, preventing current leakage, ensuring the directional transmission of electrical energy along the conductors, and avoiding faults such as short circuits, thereby guaranteeing the safety and stability of power transmission. Insulators possess high insulation performance, high strength, and corrosion resistance, and are typically made of ceramics, glass, or composite materials, capable of withstanding significant mechanical stress and high-voltage surges. Due to their relatively complex structure, insulator monitoring is relatively difficult, and current monitoring methods are relatively limited. The main monitoring methods currently available for insulators fall into three categories: leakage current method, ultraviolet pulse method, and image monitoring method.

[0005] Leakage current can reflect operating voltage, insulator contamination and icing levels, and ambient humidity, and can be used to dynamically monitor environmental changes through waveform, enabling real-time monitoring. However, using leakage current to determine insulator reliability has problems: it has a weak correlation with insulator surface conductivity and icing status, requires extraction of easily disturbed high-order components for analysis, and the waveform is affected by multiple factors, making it difficult to establish authoritative standards and requiring manual judgment, thus reducing the reliability of the results. Furthermore, during icing, the transmission network is affected by severe conditions, making it difficult to guarantee a reliable power supply voltage for monitoring leakage current, hindering its development in online monitoring.

[0006] Currently, ultraviolet (UV) imaging analysis of iced insulators primarily assesses their safe operating status based on changes in UV photon quantity during icing and discharge. However, UV imaging discharge detection is affected by factors such as gain, detection distance, viewing angle, pollution, air pressure, and atmospheric humidity, and there is a lack of consensus and standards for quantitative analysis of UV characteristics under various conditions. Therefore, despite its advantages, UV imaging presents challenges in practical application.

[0007] In summary, the current method has the following main drawbacks:

[0008] 1. Determining the icing status of insulators based on leakage current requires manual assistance, reducing the reliability of the results. During icing, the transmission network is affected by severe conditions, making it difficult to guarantee a reliable power supply voltage for monitoring leakage current, which is detrimental to the development of online monitoring.

[0009] 2. The detection of icing status of insulators by ultraviolet imaging discharge is affected by factors such as gain, detection distance, viewing angle, pollution, air pressure and atmospheric humidity. Furthermore, there is a lack of consensus and standards for the quantitative analysis of ultraviolet characteristic quantities under various conditions, making practical application difficult. Summary of the Invention

[0010] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0011] A deep learning-based method for detecting and melting ice accretion in insulators includes the following specific steps:

[0012] S1: Construct an insulator image dataset and preprocess the images in the dataset;

[0013] S2: Construct an insulator image segmentation model to segment the insulators present in the image;

[0014] S3: Training the insulator image segmentation model;

[0015] S4: Apply the insulator image segmentation model to calculate the degree of icing on the insulator;

[0016] S5: Monitor the degree of icing on insulators and perform de-icing treatment on severely iced insulators.

[0017] As a preferred embodiment of the deep learning-based insulator icing detection and de-icing method described in this invention, the specific steps of S1 are as follows:

[0018] S11: Construct an insulator image dataset, which includes two types of images: images of insulators without ice and images of insulators with ice.

[0019] S12: Use the padding method to pad the images in the dataset with the specified value to unify the image size, and use Labelme annotation software to annotate the power transmission lines and line icing in the images respectively.

[0020] As a preferred embodiment of the deep learning-based insulator icing detection and de-icing method described in this invention, the specific steps of S2 are as follows:

[0021] S21: Construct a linear inverse residual network LIRNet and use LIRNet as the backbone network of the insulator image segmentation model;

[0022] S22: Design a multi-scale feature fusion module (MFF) to balance feature information in insulator feature maps of different scales;

[0023] S23: Design an adaptive spatial attention mechanism (ASA) to make the model pay more attention to important regions;

[0024] S24: Construct an insulator image segmentation model to segment the insulators present in the image.

[0025] As a preferred embodiment of the deep learning-based insulator icing detection and de-icing method described in this invention, the specific steps of S21 are as follows:

[0026] S211: Design the inverse residual dual-bottleneck module IRDB to effectively extract feature information while reducing computational load;

[0027] S212: Use the IRDB module to construct the linear inverse residual network LIRNet as the backbone network of the insulator image segmentation model.

[0028] As a preferred embodiment of the deep learning-based insulator icing detection and melting method described in this invention, the specific steps of S211 are as follows: The insulator feature map R2 is input into the inverse residual dual-bottleneck module IRDB. The insulator feature map R2 is then subjected to dimensionality upscaling by a convolutional layer Conv with a kernel size of 1×1 and a stride of 1 to obtain the inverse residual feature map B1. The inverse residual feature map B1 is then subjected to a depthwise separable convolutional layer DW with a kernel size of 3×3 and a stride of 2 to obtain the inverse residual feature map B2. The inverse residual feature map B2 is then subjected to a convolutional block attention layer CBAM to obtain the inverse residual feature map B3. Figure B3 is passed through a depth-separable convolutional layer (DW) with a kernel size of 3×3 and a stride of 1 to obtain the inverted residual feature map B4. The inverted residual feature map B4 is then passed through a convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1 for dimensionality reduction, and then through a convolutional block attention layer (CBAM) to obtain the inverted residual feature map B5. The inverted residual feature map B3 is then passed through a convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1 for dimensionality reduction to obtain the inverted residual feature map B6. Finally, the inverted residual feature maps B5 and B6 are added together to output the inverted feature map R3 as the result of the inverted residual dual bottleneck module (IRDB).

[0029] As a preferred embodiment of the deep learning-based insulator icing detection and de-icing method described in this invention, the specific steps of S212 are as follows: the insulator image R1 first passes through a convolutional layer Conv with a kernel size of 3×3 and a stride of 2 to obtain an insulator feature map R2. The insulator feature map R2 passes through a first IRDB module to obtain an insulator feature map R3. The insulator feature map R3 passes through a second IRDB module to obtain an insulator feature map R4. The insulator feature map R4 passes through a third IRDB module to obtain an insulator feature map R5. The insulator feature map R5 passes through a fourth IRDB module to obtain an insulator feature map R6. The insulator feature map R6 passes through a convolutional layer Conv with a kernel size of 3×3 and a stride of 2 to output an insulator feature map R7 as the result of the linear inverse residual network LIRNet.

[0030] As a preferred embodiment of the deep learning-based insulator icing detection and melting method described in this invention, the specific steps of S22 are as follows: Insulator feature maps R3, R5, and R7 are input into the multi-scale feature fusion module MFF. Then, insulator feature map R3 undergoes 4x downsampling using average pooling (AvgPool) and channel adjustment using a 1×1 convolutional layer (Conv) with a stride of 1 to obtain the insulator multi-scale feature map F1. The multi-scale feature map F1 and the insulator feature map R5 are weighted and added together using the WAdd operation to obtain the multi-scale feature map F2 of the insulator. Then, the multi-scale feature map F2 of the insulator is downsampled by 4 times using average pooling (AvgPool) and the number of channels is adjusted by a convolutional layer Conv with a kernel size of 1×1 and a stride of 1 to obtain the multi-scale feature map F3 of the insulator. Finally, the multi-scale feature map F3 of the insulator is weighted and added together with the insulator feature map R7 using the WAdd operation to obtain the insulator feature map R8. The feature fusion process can be expressed by the following formula:

[0031]

[0032] In the formula, Avg represents average pooling, and a, b, c, d are the calculated weight factors obtained through training.

[0033] As a preferred embodiment of the deep learning-based insulator icing detection and melting method described in this invention, the specific steps of S23 are as follows: The insulator feature map R8 is input into the adaptive spatial attention mechanism ASA. Global max pooling (GMP) and global average pooling (GAP) operations are performed on the insulator feature map R8 at the channel dimension. The purpose of these two operations is to compress the number of channels, obtaining insulator spatial attention feature maps S1 and S2 respectively. Then, the insulator spatial attention feature maps S1 and S2 are weighted and summed (WAdd) to obtain the insulator spatial attention feature map S3. The calculation formula is as follows:

[0034] S3=α·GMP(R8)+(1-α)GAP(R8)

[0035] In the formula, GMP is global max pooling, GAP is global average pooling, and α is the adaptive weight factor obtained through training. Then, S3 obtains the spatial attention weight γ through the Sigmoid activation function. Finally, γ is multiplied by the insulator feature map R8 to assign weights, and the output insulator feature map R9 is used as the result of the adaptive spatial attention mechanism ASA. The calculation formula is as follows:

[0036] R9=σ(S3)·R8

[0037] In the formula, σ is the Sigmoid activation function.

[0038] As a preferred embodiment of the deep learning-based insulator icing detection and melting method described in this invention, the specific steps of S24 are as follows: First, obtain an insulator image R1 from the insulator image dataset; input the insulator image R1 into the linear inverse residual network (LIRNet) in the insulator image segmentation model to obtain insulator feature maps R3, R5, and R7; input the insulator feature maps R3, R5, and R7 into the multi-scale feature fusion module (MFF) for feature fusion to obtain insulator feature map R8; pass the insulator feature map R3 through the adaptive spatial attention mechanism (ASA) to obtain insulator feature map R9; pass the insulator feature map R5 through the adaptive spatial attention mechanism (ASA) to obtain insulator feature map R10; pass the insulator feature map R8 through the adaptive spatial attention mechanism (ASA) to obtain insulator feature map R11; pass the insulator feature map R11 through the spatial attention mechanism (ASA)... After extracting features using the Spatial Convolutional Pooling Pyramid (ASPP), insulator feature map R12 is obtained. Insulator feature map R12 is then passed through a convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1, followed by a 4x upsampling operation (UpSample) to obtain insulator feature map R13. Insulator feature map R10 is concatenated with insulator feature map R13 (Concat) to obtain insulator feature map R14. Insulator feature map R14 is then passed through a convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1, followed by a 4x upsampling operation (UpSample) to obtain insulator feature map R15. Insulator feature map R9 is then concatenated with insulator feature map R15 (Concat) to obtain insulator feature map R16. Insulator feature map R16 is then passed through a convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1, followed by a 4x upsampling operation (UpSample) to output the segmented insulator image.

[0039] As a preferred embodiment of the deep learning-based insulator icing detection and de-icing method described in this invention, the specific steps of S3 are as follows:

[0040] S31: Input a batch of 32 training images into the insulator image segmentation model for training, and calculate the training loss value;

[0041] S32: After training the insulator image segmentation model using all batches of insulator images in the training set, input the validation set into the insulator image segmentation model according to batches to obtain the corresponding batch loss value batch_loss.

[0042] S33: During training and validation, the parameters of the insulator image segmentation model will be automatically learned and adjusted based on each loss and batch loss.

[0043] S34: When the loss and batch_loss values ​​tend to converge and the difference between them is less than a certain threshold, the training of the insulator image segmentation model ends; the threshold needs to be determined according to the actual training situation.

[0044] S35: After training, use the test set to test the model's performance. The trend of the loss value in the test set can determine whether the model has learned effectively and whether overfitting is possible.

[0045] The specific steps of S4 are as follows:

[0046] S41: Take a picture of an unice-covered insulator T0, input it into the insulator image segmentation model, obtain a local insulator image T1 segmented from the insulator image T0, and calculate the number of insulator pixels N0 in the insulator image T1.

[0047] S42: Take an insulator image T2 every 30 minutes and input it into the insulator image segmentation model. Segment the insulator image T2 to obtain a local insulator image T3, and calculate the number of insulator pixels N1 in the insulator image T3. Calculate the ratio of the number of insulator pixels in the insulator image T3 to the number of insulator pixels in the insulator image T1, using the following formula:

[0048]

[0049] In the formula, P is the ratio of the number of insulator pixels in T3 to the number of insulator pixels in T1, which represents the degree of icing on the insulator.

[0050] The specific steps of S5 are as follows:

[0051] S51: Real-time monitoring of insulator icing level P;

[0052] S52: When the degree of icing P of the insulator reaches the level that affects the normal operation of the insulator, an alarm is immediately issued, and the power grid staff will carry out de-icing treatment on the iced insulator. A short-time high current is applied to the insulator string to melt the ice using the Joule heating effect, or the ice is removed remotely using laser or ultrasonic technology.

[0053] Compared with existing technologies:

[0054] This invention constructs an insulator image segmentation model, which can accurately identify insulators in segmented images to monitor the degree of icing on insulators in real time. When the degree of icing on an insulator reaches a level that affects its normal operation, an alarm is immediately issued to the staff, and the iced insulator is melted. Based on this, this invention effectively improves the efficiency of insulator icing detection, significantly reduces the manpower required for detection, and reduces the impact of environmental factors on detection. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0056] Figure 2 This is a schematic diagram of the inverted residual dual bottleneck module IRDB structure of the present invention;

[0057] Figure 3 This is a schematic diagram of the linear inverse residual network (LIRNet) structure of the present invention;

[0058] Figure 4 This is a schematic diagram of the multi-scale feature fusion module (MFF) structure of the present invention;

[0059] Figure 5 This is a schematic diagram of the Adaptive Spatial Attention (ASA) mechanism structure of the present invention;

[0060] Figure 6 This is a schematic diagram of the insulator image segmentation model structure of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0062] This invention provides a deep learning-based method for detecting and melting ice accretion in insulators. Please refer to [link / reference]. Figures 1-6 The specific steps are as follows:

[0063] S1: Construct an insulator image dataset and preprocess the images in the dataset.

[0064] The specific steps of S1 are as follows:

[0065] S11: Construct an insulator image dataset, which includes two types of images: images of insulators without ice and images of insulators with ice.

[0066] S12: Use the padding method to pad the images in the dataset with the specified value to unify the image size, and use Labelme annotation software to annotate the power transmission lines and line icing in the images respectively.

[0067] Padding is a common method for modifying image size during the dataset preprocessing of neural network models. This method does not change the original shape of the data and does not lose the original information of the image.

[0068] Labelme annotation software is a commonly used tool for annotating image information, and is not an innovation of this invention.

[0069] S2: Construct an insulator image segmentation model to segment the insulators present in the image;

[0070] This enables the input of an insulator image R1 from an insulator image dataset into an insulator image segmentation model to segment the insulators present in the image.

[0071] The specific steps of S2 are as follows:

[0072] S21: Construct a linear inverted residual network (LIRNet) and use LIRNet as the backbone network of the insulator image segmentation model;

[0073] R1 is first passed through a convolutional layer Conv with a kernel size of 3×3 and a stride of 2 to obtain the insulator feature map R2;

[0074] The specific steps of S21 are as follows:

[0075] S211: Design an Inverted Residual Double Bottleneck Module (IRDB) to effectively extract feature information while reducing computational load.

[0076] The specific steps of S211 are as follows: The insulator feature map R2 is input into the inverse residual dual-bottleneck module IRDB. The insulator feature map R2 is then subjected to dimensionality upscaling via a convolutional layer Conv with a kernel size of 1×1 and a stride of 1 to obtain the inverse residual feature map B1. The inverse residual feature map B1 is then subjected to a depthwise separable convolutional layer DW with a kernel size of 3×3 and a stride of 2 to obtain the inverse residual feature map B2. The inverse residual feature map B2 is then subjected to a convolutional block attention layer CBAM (Convolutional Block Attention). The insulator inverse residual feature map B3 is obtained from the module. B3 is then passed through a depth-separated convolutional layer (DW) with a kernel size of 3×3 and a stride of 1 to obtain insulator inverse residual feature map B4. B4 is then passed through a convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1 for dimensionality reduction, followed by a convolutional block attention layer (CBAM) to obtain insulator inverse residual feature map B5. B3 is then passed through another convolutional layer (Conv) with a kernel size of 1×1 and a stride of 1 for dimensionality reduction to obtain insulator inverse residual feature map B6. Finally, insulator inverse residual feature maps B5 and B6 are added together to output insulator feature map R3, which serves as the result of the inverse residual dual-bottleneck module IRDB. The IRDB structure is as follows: Figure 2 As shown.

[0077] The core idea of ​​the Convolutional Block Attention Module (CBAM) is to refine the input feature map in two stages: first, the channel attention module focuses on "which channels are important," and then the spatial attention module focuses on "where" is an informational part. This dual attention mechanism enables CBAM to comprehensively capture the key information in the features.

[0078] S212: Use the IRDB module to construct the linear inverse residual network LIRNet as the backbone network of the insulator image segmentation model.

[0079] The IRDB module designed in S211 is used to construct a linear inverse residual network (LIRNet) as the backbone network of the insulator image segmentation model. The specific steps of S212 are as follows: Insulator image R1 first passes through a convolutional layer Conv with a kernel size of 3×3 and a stride of 2 to obtain insulator feature map R2. Insulator feature map R2 passes through the first IRDB module to obtain insulator feature map R3. Insulator feature map R3 passes through the second IRDB module to obtain insulator feature map R4. Insulator feature map R4 passes through the third IRDB module to obtain insulator feature map R5. Insulator feature map R5 passes through the fourth IRDB module to obtain insulator feature map R6. Insulator feature map R6 passes through a convolutional layer Conv with a kernel size of 3×3 and a stride of 2 to output insulator feature map R7 as the result of the linear inverse residual network (LIRNet). The LIRNet structure is as follows: Figure 3 As shown. Among them, the first IRDB module, the second IRDB module, the third IRDB module, and the fourth IRDB module have the same structure.

[0080] Example: An insulator image R1 is obtained from the insulator image dataset. R1 has a size of 512×512×3 (corresponding to the length, width, and number of channels of the feature map). R1 ​​is first passed through a convolutional layer Conv with a kernel size of 3×3 and a stride of 2 to obtain an insulator feature map R2 with a size of 256×256×32. R2 is input into the first IRDB module. R2 is passed through a Conv layer with a kernel size of 1×1 and a stride of 1 for dimensionality upscaling to obtain an inverse residual feature map B1 with a size of 256×256×128. B1 is passed through a depthwise separable convolutional layer DW with a kernel size of 3×3 and a stride of 2 to obtain an inverse residual feature map B2 with a size of 128×128×128. B2 is then passed through CBAM to obtain... The insulator inverse residual feature map B3, with a size of 128×128×128, is obtained. After passing through a DW layer with a convolution kernel size of 3×3 and a stride of 1, B4, the insulator inverse residual feature map B4, with a size of 128×128×128, is obtained. After passing through a Conv layer with a convolution kernel size of 1×1 and a stride of 1 for dimensionality reduction, B4 is passed through a CBAM layer to obtain an insulator inverse residual feature map B5, with a size of 128×128×64. After passing through a Conv layer with a convolution kernel size of 1×1 and a stride of 1 for dimensionality reduction, B3 is obtained to obtain an insulator inverse residual feature map B6, with a size of 128×128×64. Finally, B5 and B6 are added together to output an insulator feature map R3, with a size of 128×128×64, which is the result of the first IRDB module. R3 passes through the second IRDB module to obtain an insulator feature map R4 with a size of 64×64×128. R4 passes through the third IRDB module to obtain an insulator feature map R5 with a size of 64×64×256. R5 passes through the fourth IRDB module to obtain an insulator feature map R6 with a size of 32×32×512. R6 passes through a Conv layer with a convolution kernel size of 3×3 and a stride of 2 to output an insulator feature map R7 with a size of 16×16×1024, which is the result of LIRNet.

[0081] S22: Design a multi-scale feature fusion module (MFF) to balance feature information in insulator feature maps of different scales.

[0082] The specific steps of S22 are as follows: Insulator feature maps R3, R5, and R7 are input into the multi-scale feature fusion module MFF. Then, insulator feature map R3 is downsampled by 4x using average pooling (AvgPool) and its channel count is adjusted using a 1×1 convolutional layer (Conv) with a stride of 1 to obtain insulator multi-scale feature map F1. Insulator multi-scale feature map F1 and insulator feature map R5 are weighted and added together using WAdd (Weighted addition) to obtain insulator multi-scale feature map F2. Insulator multi-scale feature map F2 is then downsampled by 4x using average pooling (AvgPool) and its channel count is adjusted using a 1×1 convolutional layer (Conv) with a stride of 1 to obtain insulator multi-scale feature map F3. Finally, insulator multi-scale feature map F3 and insulator feature map R7 are weighted and added together using WAdd to obtain insulator feature map R8. The feature fusion process can be expressed by the following formula:

[0083]

[0084] In the formula, Avg represents average pooling, and a, b, c, d are the calculated weight factors obtained through training; the MFF structure is as follows: Figure 4 As shown.

[0085] Example: Input R3, R5, and R7 into MFF. Then, R3 is downsampled by 4x using AvgPool and adjusted for the number of channels using a Conv layer with a kernel size of 1×1 and a stride of 1 to obtain an insulator multi-scale feature map F1 with a size of 64×64×256. F1 and R5 are then combined with WAdd to obtain an insulator multi-scale feature map F2 with a size of 64×64×256. F2 is then downsampled by 4x using AvgPool and adjusted for the number of channels using a Conv layer with a kernel size of 1×1 and a stride of 1 to obtain an insulator multi-scale feature map F3 with a size of 16×16×1024. Finally, F3 and R7 are combined with WAdd to obtain an insulator feature map R8 with a size of 16×16×1024, which is the result of MFF.

[0086] S23: Design an adaptive spatial attention module (ASA) to make the model pay more attention to important regions.

[0087] The specific steps of S23 are as follows: The insulator feature map R8 is input into the adaptive spatial attention mechanism ASA. Global Max Pooling (GMP) and Global Average Pooling (GAP) operations are performed on the insulator feature map R8 at the channel dimension. The purpose of these two operations is to compress the number of channels, resulting in insulator spatial attention feature maps S1 and S2, respectively. Then, insulator spatial attention feature maps S1 and S2 are weighted and summed (WAdd) to obtain insulator spatial attention feature map S3. The calculation formula is as follows:

[0088] S3=α·GMP(R8)+(1-α)GAP(R8)

[0089] In the formula, GMP is global max pooling, GAP is global average pooling, and α is the adaptive weight factor obtained through training. Then, S3 obtains the spatial attention weight γ through the Sigmoid activation function. Finally, γ is multiplied by the insulator feature map R8 to assign weights, and the output insulator feature map R9 is used as the result of the adaptive spatial attention mechanism ASA. The calculation formula is as follows:

[0090] R9=σ(S3)·R8

[0091] In the formula, σ is the Sigmoid activation function; a schematic diagram of the ASA structure is shown below. Figure 5 As shown.

[0092] Example: Input the insulator feature map R8 into ASA, perform GMP and GAP on R8 in the channel dimension to compress the number of channels, and obtain insulator spatial attention feature maps S1 and S2 respectively. The size of S1 and S2 is 16×16×1. Then, perform WAdd on S1 and S2 to obtain insulator spatial attention feature map S3 with a size of 16×16×1. Then, S3 is activated by the Sigmoid activation function to obtain the spatial attention weight γ. Finally, multiply γ with R8 to assign weights and output insulator feature map R9 with a size of 16×16×1024 as the result of ASA.

[0093] S24: Construct an insulator image segmentation model to segment the insulators present in the image.

[0094] An insulator image segmentation model is constructed using the LIRNet backbone network built in S21, the MFF multi-scale feature fusion module designed in S22, and the ASA adaptive spatial attention mechanism designed in S23. The model structure is as follows: Figure 6 As shown.

[0095] The specific steps of S24 are as follows: First, obtain an insulator image R1 from the insulator image dataset. Input the insulator image R1 into the linear inverse residual network (LIRNet) in the insulator image segmentation model to obtain insulator feature maps R3, R5, and R7. Input the insulator feature maps R3, R5, and R7 into the multi-scale feature fusion module (MFF) for feature fusion to obtain insulator feature map R8. Pass the insulator feature map R3 through the adaptive spatial attention mechanism (ASA) to obtain insulator feature map R9. Pass the insulator feature map R5 through the adaptive spatial attention mechanism (ASA) to obtain insulator feature map R10. Pass the insulator feature map R8 through the adaptive spatial attention mechanism (ASA) to obtain insulator feature map R11. Pass the insulator feature map R11 through the atrous spatial convolutional pooling pyramid (ASPP) process. After feature extraction using pooling, insulator feature map R12 is obtained. R12 is then passed through a convolutional layer Conv with a kernel size of 1×1 and a stride of 1, followed by a 4x upsampling operation UpSample to obtain insulator feature map R13. Insulator feature map R10 and R13 are concatenated (Concat) to obtain insulator feature map R14. R14 is then passed through a convolutional layer Conv with a kernel size of 1×1 and a stride of 1, followed by a 4x upsampling operation UpSample to obtain insulator feature map R15. Insulator feature map R9 and R15 are then concatenated (Concat) to obtain insulator feature map R16. Finally, R16 is passed through a convolutional layer Conv with a kernel size of 1×1 and a stride of 1, followed by a 4x upsampling operation UpSample to output the segmented insulator image.

[0096] Among them, ASPP (Atrous Spatial Pyramid Pooling) is a deep learning technique for extracting multi-scale features, commonly used in semantic segmentation tasks. It includes components such as 1×1 convolutions, dilated convolutions, and pooling pyramids. It obtains features from different receptive fields through dilated convolutions with different dilation rates, then obtains global information through global pooling and 1×1 convolutions, and finally enhances the model's expressive power by stacking and fusing these features.

[0097] Example: First, obtain an insulator image R1 from the insulator image dataset. R1 has a size of 512×512×3 (corresponding to the length, width, and number of channels of the feature map). Input R1 into the LIRNet module in the insulator image segmentation model to obtain an insulator feature map R3 with a size of 128×128×64, an insulator feature map R5 with a size of 64×64×256, and an insulator feature map R7 with a size of 16×16×1024. Then, combine R3, R5, and... R7 is input to MFF for feature fusion to obtain an insulator feature map R8 with a size of 16×16×1024. R3 is processed by ASA to obtain an insulator feature map R9 with a size of 128×128×64. R5 is processed by ASA to obtain an insulator feature map R10 with a size of 64×64×256. R8 is processed by ASA to obtain an insulator feature map R11 with a size of 16×16×1024. R11 is then processed by Atrous Spatial Pyramid Convolution Pooling Pyramid (ASPP). After feature extraction using pooling, an insulator feature map R12 with a size of 16×16×256 is obtained. R12 is then passed through a convolution kernel of size 1×1 with a stride of 1, and then upsampled by 4 times to obtain an insulator feature map R13 with a size of 64×64×256. R10 and R13 are concatenated to obtain an insulator feature map R14 with a size of 64×64×512. R14 is then passed through a convolution kernel of size 1×1 with a stride of 1, and then upsampled by 4 times to obtain an insulator feature map R15 with a size of 128×128×64. R9 and R15 are then concatenated to obtain an insulator feature map R16 with a size of 128×128×128. R16 is then passed through a convolution kernel of size 1×1 with a stride of 1, and then upsampled by 4 times to output the segmented insulator image.

[0098] S3: Training the insulator image segmentation model.

[0099] During the training and validation of the insulator image segmentation model, the insulator image dataset was divided into training, validation, and test sets in a 6:3:1 ratio. During training, in each epoch, the training images in the training set were randomly shuffled and divided into 300 complete batches with a batch size of 32. Each batch contained 32 consecutive training images and their annotations. The validation set maintained a fixed order and was divided into 94 batches in the same manner.

[0100] The specific steps of S3 are as follows:

[0101] S31: Input a batch of 32 training images into the insulator image segmentation model for training, and calculate the training loss value;

[0102] S32: After training the insulator image segmentation model using all batches of insulator images in the training set, input the validation set into the insulator image segmentation model according to batches to obtain the corresponding batch loss value batch_loss.

[0103] S33: During training and validation, the parameters of the insulator image segmentation model will be automatically learned and adjusted based on each loss and batch loss.

[0104] S34: When the loss and batch_loss values ​​tend to converge and the difference between them is less than a certain threshold, the training of the insulator image segmentation model ends; the threshold needs to be determined according to the actual training situation.

[0105] S35: After training, use the test set to test the model's performance. The trend of the loss value in the test set can determine whether the model has learned effectively and whether overfitting is possible.

[0106] S4: Apply the insulator image segmentation model to calculate the degree of icing on the insulator.

[0107] When installing fixed monitoring equipment such as cameras at the locations of insulators that require monitoring, the specific steps of S4 are as follows:

[0108] S41: Take a picture of an unice-covered insulator T0, input it into the insulator image segmentation model, obtain a local insulator image T1 segmented from the insulator image T0, and calculate the number of insulator pixels N0 in the insulator image T1.

[0109] S42: Take an insulator image T2 every 30 minutes and input it into the insulator image segmentation model. Segment the insulator image T2 to obtain a local insulator image T3, and calculate the number of insulator pixels N1 in the insulator image T3. Calculate the ratio of the number of insulator pixels in the insulator image T3 to the number of insulator pixels in the insulator image T1, using the following formula:

[0110]

[0111] In the formula, P is the ratio of the number of insulator pixels in T3 to the number of insulator pixels in T1, which represents the degree of icing on the insulator.

[0112] S5: Monitor the degree of icing on insulators and perform de-icing treatment on severely iced insulators.

[0113] The specific steps of S5 are as follows:

[0114] S51: Real-time monitoring of insulator icing level P;

[0115] S52: When the degree of icing P of the insulator reaches the level that affects the normal operation of the insulator, an alarm is immediately issued, and the power grid staff will carry out de-icing treatment on the iced insulator. A short-time high current is applied to the insulator string to melt the ice using the Joule heating effect, or the ice is removed remotely using laser or ultrasonic technology.

[0116] In summary, this invention first constructs an insulator image dataset, containing both un-iced and iced insulator images, for subsequent model construction and training. Then, an insulator image segmentation model is designed to segment insulators present in the images. Within this model, a Linear Inverted Residual Network (LIRNet) is designed to extract semantic information from insulator feature maps at different scales. An Inverted Residual Double Bottleneck Module (IRDB) is designed within LIRNet to effectively extract feature information while reducing computational cost. A Multi-Scale Feature Fusion Module (MFF) is designed to balance feature information in insulator feature maps at different scales. Finally, an Adaptive Spatial Attention Module (ASA) is designed to make the model pay more attention to important regions. This invention employs an insulator image segmentation model to segment insulators within an insulator image. By calculating the ratio of the number of insulator pixels in the original insulator image to the number of insulator pixels in the image of un-iced insulators, the degree of icing is determined. When the icing level reaches a level that affects the normal operation of the insulator, an alarm is immediately issued to personnel, and the iced insulator is melted. This invention effectively improves the efficiency of insulator icing detection, significantly reduces the manpower required for detection, and minimizes the impact of environmental factors on the detection process.

[0117] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A deep learning-based method for detecting icing and de-icing of an insulator, the method comprising: The specific steps include the following: S1: Constructing an insulator image dataset, pre-processing the images in the dataset; S2: Constructing an insulator image segmentation model for segmenting the insulators present in the images, and the specific steps are as follows: S21: Constructing a linear inverse residual network LIRNet, taking the LIRNet as the backbone network of the insulator image segmentation model, and the specific steps are as follows: S211: Designing an inverse residual double-bottleneck module IRDB to effectively extract feature information while reducing the amount of calculation, and the specific steps are as follows: inputting an insulator feature map R2 into the inverse residual double-bottleneck module IRDB, performing dimensionality increasing on the insulator feature map R2 through a convolution layer Conv with a convolution kernel size of 1*1 and a step size of 1 to obtain an insulator inverse residual feature map B1, performing dimensionality increasing on the insulator inverse residual feature map B1 through a depth separable convolution layer DW with a convolution kernel size of 3*3 and a step size of 2 to obtain an insulator inverse residual feature map B2, obtaining an insulator inverse residual feature map B3 through a convolution block attention layer CBAM from the insulator inverse residual feature map B2, performing dimensionality decreasing on the insulator inverse residual feature map B4 through a depth separable convolution layer DW with a convolution kernel size of 3*3 and a step size of 1, and obtaining an insulator inverse residual feature map B5 through a convolution block attention layer CBAM layer, performing dimensionality decreasing on the insulator inverse residual feature map B3 through a convolution layer Conv with a convolution kernel size of 1*1 and a step size of 1 to obtain an insulator inverse residual feature map B6, and finally adding the insulator inverse residual feature map B5 and the insulator inverse residual feature map B6 to output an insulator feature map R3 as the result of the inverse residual double-bottleneck module IRDB; S212: Using the IRDB module to construct a linear inverse residual network LIRNet as the backbone network of the insulator image segmentation model; S22: Designing a multi-scale feature fusion module MFF for balancing the feature information in insulator feature maps of different scales; S23: Designing an adaptive spatial attention mechanism ASA to make the model pay more attention to important areas; S24: Constructing an insulator image segmentation model for segmenting the insulators present in the images; S3: Training the insulator image segmentation model; S4: Applying the insulator image segmentation model to calculate the icing degree of the insulator; S5: Monitoring the icing degree of the insulator and performing ice melting treatment on the seriously iced insulator.

2. The insulator icing detection and de-icing method based on deep learning according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Constructing an insulator image dataset, and the insulator image dataset includes no-icing insulator images and icing insulator images; S12: Using the pad padding method to fill the specified value around the images in the dataset, unifying the image size, and using the Labelme labeling software to mark the power transmission lines and line icing in the images.

3. The method of claim 1, wherein the method further comprises: The specific steps of S212 are as follows: the insulator image R1 is first subjected to a convolution layer Conv with a convolution kernel size of 3*3 and a step length of 2 to obtain an insulator feature map R2, the insulator feature map R2 is subjected to a first IRDB module to obtain an insulator feature map R3, the insulator feature map R3 is subjected to a second IRDB module to obtain an insulator feature map R4, the insulator feature map R4 is subjected to a third IRDB module to obtain an insulator feature map R5, the insulator feature map R5 is subjected to a fourth IRDB module to obtain an insulator feature map R6, and the insulator feature map R6 is subjected to a convolution layer Conv with a convolution kernel size of 3*3 and a step length of 2 to output an insulator feature map R7 as a result of the linear inverse residual network LIRNet.

4. The insulator icing detection and de-icing method based on deep learning according to claim 1, characterized in that, The specific steps of S22 are as follows: the insulator feature maps R3, R5 and R7 are input into a multi-scale feature fusion module MFF, then the insulator feature map R3 is subjected to 4 times down-sampling by an average pooling AvgPool and adjustment of the number of channels by a convolution layer Conv with a convolution kernel size of 1*1 and a step length of 1 to obtain an insulator multi-scale feature map F1, the insulator multi-scale feature map F1 is subjected to a weighted addition WAdd operation with the insulator feature map R5 to obtain an insulator multi-scale feature map F2, the insulator multi-scale feature map F2 is subjected to 4 times down-sampling by an average pooling AvgPool and adjustment of the number of channels by a convolution layer Conv with a convolution kernel size of 1*1 and a step length of 1 to obtain an insulator multi-scale feature map F3, and finally the insulator multi-scale feature map F3 is subjected to a weighted addition WAdd operation with the insulator feature map R7 to obtain an insulator feature map R8, and the feature fusion process can be expressed by the following formula: wherein, is an average pooling, is a calculated weight factor obtained by training.

5. The method of claim 1, wherein the method is characterized by: The specific steps of S23 are as follows: the insulator feature map R8 is input into an adaptive spatial attention mechanism ASA, global maximum pooling GMP operation and global average pooling GAP operation are performed on the channel dimension of the insulator feature map R8, the purpose of the two operations is to compress the number of channels, and an insulator spatial attention feature map S1 and an insulator spatial attention feature map S2 are obtained, respectively, then the insulator spatial attention feature map S1 and the insulator spatial attention feature map S2 are subjected to a weighted addition WAdd operation to obtain an insulator spatial attention feature map S3, and the calculation formula is as follows: wherein, is a global max pooling, is a global average pooling, is an adaptive weight factor obtained by training; then S3 obtains the spatial attention weight through the Sigmoid activation function , and finally is multiplied by the insulator feature map R8 to weight and output the insulator feature map R9 as the result of the adaptive spatial attention mechanism ASA, and the calculation formula is as follows: wherein, is a Sigmoid activation function.

6. The method of claim 1, wherein the method is characterized by: The specific steps of S24 are as follows: first, an insulator image R1 is obtained from the insulator image dataset, the insulator image R1 is input into the linear inverse residual network LIRNet in the insulator image segmentation model to obtain the insulator feature map R3, the insulator feature map R5 and the insulator feature map R7; the insulator feature map R3, the insulator feature map R5 and the insulator feature map R7 are input into the multi-scale feature fusion module MFF for feature fusion to obtain the insulator feature map R8, the insulator feature map R3 is subjected to the adaptive spatial attention mechanism ASA to obtain the insulator feature map R9, the insulator feature map R5 is subjected to the adaptive spatial attention mechanism ASA to obtain the insulator feature map R10, the insulator feature map R8 is subjected to the adaptive spatial attention mechanism ASA to obtain the insulator feature map R11, the insulator feature map R11 is subjected to the feature extraction of the atrium spatial convolution pyramid ASPP to obtain the insulator feature map R12, the insulator feature map R12 is subjected to a convolution layer Conv with a convolution kernel size of 1*1 and a step of 1, and then subjected to a 4 times up sampling operation UpSample to obtain the insulator feature map R13, the insulator feature map R10 and the insulator feature map R13 are spliced Concat to obtain the insulator feature map R14, the insulator feature map R14 is subjected to a convolution layer Conv with a convolution kernel size of 1*1 and a step of 1, and then subjected to a 4 times up sampling operation UpSample to obtain the insulator feature map R15, the insulator feature map R9 and the insulator feature map R15 are spliced Concat to obtain the insulator feature map R16, the insulator feature map R16 is subjected to a convolution layer Conv with a convolution kernel size of 1*1 and a step of 1, and then subjected to a 4 times up sampling operation UpSample to output the segmented insulator image.

7. The method of claim 1, wherein the method further comprises: The specific steps of S3 are as follows: S31: input a batch of 32 training images into the insulator image segmentation model for training, and calculate the training loss value loss; S32: when the training of the insulator image segmentation model using all batches of insulator images in the training set is completed, input the validation set into the insulator image segmentation model according to batches to obtain the corresponding batch loss value batch_loss; S33: during training and verification, the parameters of the insulator image segmentation model are automatically learned and adjusted according to the loss and batch_loss each time; S34: when the loss and batch_loss values tend to converge and the difference between them is less than a certain threshold, the training of the insulator image segmentation model is ended; the threshold needs to be determined according to the actual training situation; S35: after the training is completed, the model effect is tested using the test set, and the loss value change trend of the test set can judge whether the model is effectively learned and whether overfitting may occur; The specific steps of S4 are as follows: S41: Take an image T0 of an insulator without ice cover, input an insulator image segmentation model, obtain a local insulator image T1 segmented from the insulator image T0, and calculate the number of insulator pixels in the insulator image T1 ; S42: take an insulator image T2 every 30 minutes, input the insulator image into the insulator image segmentation model, segment the local insulator image T3 from the insulator image T2, and calculate the number of insulator pixels in the insulator image T3 ; calculate the ratio of the number of insulator pixels in the insulator image T3 to the number of insulator pixels in the insulator image T1, and the calculation formula is as follows: wherein, is the ratio of the number of insulator pixels in T3 to the number of insulator pixels in T1, i.e. the icing degree of the insulator. The specific steps of S5 are as follows: S51: Real-time monitoring of icing degree of insulator ; S52: When the insulator is covered with ice When the level of ice accumulation reaches a level that affects the normal operation of the insulator, an alarm is immediately issued, and power grid personnel carry out de-icing treatment on the iced insulators. This involves applying a short-term high current to the insulator string to melt the ice using the Joule heating effect, or remotely removing the ice using laser or ultrasonic technology.

Images