Power transmission line insulator state identification method based on IF-TransUNet

Abstract

The invention relates to a power transmission line insulator state identification method based on IF-TransUNet, and the method comprises the following steps: enabling an unmanned plane to fly above two sides of a wind-light power transmission line according to a preset route, employing a multi-height layered collection strategy, carrying out the wide-angle cruise shooting at a preset height from the wind-light power transmission line, and generating an image data set for insulator fault identification; preprocessing the image data set; and inputting the preprocessed image data set into a TransUNet network model to carry out insulator fault identification. According to the method, a full-process identification system of data acquisition, preprocessing, model reasoning and result output is constructed, the upgrade from manual sampling inspection to full-autonomous accurate detection is realized, the traditional manual inspection is replaced, and the accuracy of the detection result is improved through deep cooperation of image preprocessing and model improvement. The method comprehensively improves the efficiency and accuracy of the fault recognition of the insulator of the wind-solar power transmission line, and provides an efficient and reliable intelligent technical support for the safe operation and maintenance of a wind-solar power station.

Description

Technical Field

[0001] This invention relates to the field of transmission line insulator fault detection technology, specifically to a method for identifying the condition of transmission line insulators based on IF-TransUNet. Background Technology

[0002] With the rapid development of the new energy industry, the proportion of wind power and photovoltaic power generation in the power system continues to increase. As the core channel for transmitting new energy power, the operational stability of wind and solar transmission lines directly determines the reliability of new energy power supply. Insulators, as key insulating components of wind and solar transmission lines, bear the core functions of supporting conductors and isolating current. Once damage, flashover, or other abnormal operating conditions occur, it will directly lead to a decline in the insulation performance of the line, and in severe cases, it may even cause line tripping and large-scale power outages, resulting in huge economic losses. Therefore, insulator condition monitoring is a core link in the operation and maintenance of wind and solar transmission lines.

[0003] Currently, the condition monitoring of insulators in wind and solar transmission lines mainly relies on two methods: manual inspection, where maintenance personnel visually inspect, use binoculars, or climb towers – this is the traditional mainstream method; and machine-assisted inspection, which uses drone aerial images combined with machine vision or early deep learning technologies for fault identification. However, existing inspection technologies have many shortcomings and deficiencies, making it difficult to meet the needs of efficient and accurate maintenance of wind and solar transmission lines. Among these, manual inspection is particularly limited, with extremely low efficiency; inspecting a single line often requires multiple people working together for several days. The large investment of manpower and transportation resources leads to high inspection costs. Climbing towers carries the risk of falls, making safety difficult to guarantee. Furthermore, it is significantly affected by weather conditions; inspections cannot be effectively carried out in rainy, foggy, or nighttime conditions, and human error can easily lead to missed detection of minor damage, early flashovers, and other hidden faults.

[0004] Traditional machine vision technology also has obvious shortcomings. It relies on hand-designed features such as SIFT and HOG, which are extremely unreliable to complex interference in wind and light scenes. It is difficult to effectively distinguish insulators from the background, and the accuracy of fault identification is generally low. Moreover, it can only identify large-sized damage and cannot detect small cracks or blurred and weak flashover areas caused by humidity.

[0005] Existing deep learning methods also have significant shortcomings. Basic segmentation models such as Unet rely solely on CNNs to extract local features, failing to capture long-distance dependencies between insulators and fault areas, such as scattered small damages and flashover marks across the skirts, resulting in low accuracy in small target detection. While the original TransUnet combines the advantages of Transformer and Unet, it is not optimized for wind and solar scenarios, making it unable to withstand scene noise such as dynamic interference from wind turbine blades and photovoltaic reflections. Furthermore, it does not focus on the core features of insulator faults, resulting in fault classification accuracy and segmentation Dice coefficients below 80%, which is insufficient to meet the precision requirements of operation and maintenance. In addition, most deep learning models rely on cloud computing power and cannot be deployed on edge devices such as drones, resulting in severely insufficient real-time performance. Single-graph inference is time-consuming, failing to meet the actual needs of real-time feedback at the inspection site. Summary of the Invention

[0006] To address the technical challenges in existing wind and solar power transmission line insulator inspections, including low efficiency and high risk of manual inspections, poor anti-interference capabilities of traditional machine vision, and the large parameter size of existing deep learning models making them difficult to deploy on UAV edge devices, as well as inaccurate extraction and segmentation accuracy of minor fault features under strong photovoltaic panel reflections and complex dynamic background interference, this invention proposes a transmission line insulator condition identification method based on IF-TransUNet, comprising the following steps: Step S1: The drone flies along the preset route on both sides of the wind and solar power transmission line, and adopts a multi-altitude layered acquisition strategy to carry out wide-angle cruise shooting at a predetermined altitude above the wind and solar power transmission line to generate an image dataset for insulator fault identification. Step S2: Preprocess the image dataset; Step S3: Input the preprocessed image dataset into the TransUNet network model for insulator fault identification; The TransUNet network model construction method includes: using depthwise separable convolution to reconstruct the network backbone and decoding module, embedding the IF-DFAM attention mechanism in the feature fusion nodes of the encoder and decoder, designing a cross-layer feature dynamic fusion module, and forming the IF-TransUNet model.

[0007] In step S2, the preprocessing of the image dataset includes adaptive median filtering for noise reduction, proportional scaling and black filling, and Retinex illumination correction.

[0008] Among them, proportional scaling and black fill include: For insulator images taken laterally, the height is reduced proportionally based on the target width. For insulator images taken vertically, the width is reduced proportionally based on the target height. Calculate the size difference of the scaled insulator image and fill the image with black pixels symmetrically on the left and right or top and bottom.

[0009] Retinex illumination correction includes: The grayscale information of the insulator image is decomposed into "illuminance component" and "reflection component". The illumination component is extracted by Gaussian filtering, and then the illumination component is subtracted from the original image to separate the reflection component containing fault features. The brightness of the reflected component is mapped to a predetermined range.

[0010] Step S3 includes: Step S31: Use depthwise separable convolution to reconstruct the network backbone and decoding module to process the input feature map and complete the channel mapping; Step S32: Perform directional convolution on the feature map through the IF-DFAM attention mechanism to capture edge features at different angles, and extract high-frequency components from the feature map through Fourier transform to enhance fault features; Step S33: The cross-layer feature dynamic fusion module dynamically weights and fuses the shallow high-resolution features, same-layer features, and deep high-semantic features of the feature map.

[0011] In step S31, the backbone and decoding module of the depthwise separable convolutional reconstruction network adopt a split design of "depth convolution + pointwise convolution". The depth convolution only processes the corresponding channel of the input feature map, focusing on the extraction of local features in the spatial dimension, while the pointwise convolution integrates the feature map output by the depth convolution in the channel dimension.

[0012] In step S32, the IF-DFAM attention mechanism includes a direction-aware branch and a frequency-attention branch. The direction-aware branch performs directional convolution on the feature map to capture edge features at different angles, while the frequency-attention branch extracts high-frequency components from the feature map through Fourier transform to enhance fault features.

[0013] The formula for directional convolution of the feature map by the orientation-aware branch is as follows: ; in, For the first Gabor filters in one direction, focusing on the tangential direction of the ring edge; As directional weights, they are automatically assigned through learning, specifically for the direction of the ring edge. Significantly higher than other directions; F′ is the feature map after direction enhancement, F′ is the feature map input to the direction-aware branch of the IF-DFAM attention mechanism, and k is the direction index, which takes values ​​from 1 to 8, representing 8 different directions.

[0014] The formula for the frequency attention branch to extract high-frequency components from the feature map through Fourier transform is as follows: ; Wherein, FFT and IFFT are the Fourier transform and inverse transform, respectively. For high-frequency mask, This is the feature map after high-frequency enhancement.

[0015] This invention presents an IF-TransUNet-based method for identifying the condition of insulators in transmission lines. Addressing industry pain points such as poor image quality, difficulty in model deployment, and frequent missed detections of minor faults in drone inspections of wind and solar transmission lines, and considering the specific needs of drone inspections of wind and solar transmission lines, this method constructs a comprehensive identification system encompassing "data acquisition - preprocessing - model inference - result output." This upgrades the process from manual sampling to fully autonomous and accurate detection, replacing traditional manual inspections. Through deep collaboration between image preprocessing and model improvement, it comprehensively enhances the efficiency and accuracy of insulator fault identification in wind and solar transmission lines. This represents a breakthrough in the industrial application of insulator fault detection, ensuring the safe and stable operation of new energy power transmission channels and providing efficient and reliable intelligent technical support for the safe operation and maintenance of wind and solar power plants. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the IF-TransUNet-based method for identifying the state of insulators in transmission lines according to the present invention.

[0017] Figure 2 This is a diagram illustrating the effect of insulator damage identification according to the present invention.

[0018] Figure 3 This is a diagram illustrating the effect of insulator flashover identification according to the present invention. Detailed Implementation

[0019] To gain a better understanding of the technical solution and beneficial effects of the present invention, the technical solution of the present invention and its beneficial effects are described in detail below with reference to the accompanying drawings.

[0020] This invention provides a method for identifying the condition of insulators on transmission lines based on IF-TransUNet, aiming to solve problems such as low efficiency of traditional manual inspections, weak anti-interference capabilities of existing machine detection, large model parameters making deployment on UAV edge devices, and insufficient accuracy in identifying small faults. The core of this invention lies in combining deep learning, UAV technology, and computer vision technology with the operation and maintenance scenarios of wind and solar power transmission lines. More importantly, it focuses on constructing a lightweight target localization and improved TransUNet segmentation deep learning scheme. Through depthwise separable convolution parameter reduction, IF-DFAM attention anti-interference, and targeted data preprocessing optimization, it achieves automated and high-precision identification of "damage edges and flashover textures" of insulators in wind and solar power scenarios. It is also compatible with UAV edge computing devices, meeting the needs of real-time on-site inspections, and is suitable for daily inspections and safety assurance of wind and solar power transmission lines. The flowchart of the overall solution is as follows: Figure 1 As shown.

[0021] I. Image Acquisition - Obtaining Image Datasets of Wind and Solar Transmission Line Insulators under Complex Backgrounds Based on on-site investigations and literature reviews, no drone inspection technology has yet been found that combines lightweight deployment capabilities with high fault identification accuracy for wind and solar transmission line insulator scenarios. Traditional inspection methods require multiple people working together and taking several days to complete the inspection of a single line. This not only incurs high labor and transportation costs but also poses risks of falls while climbing towers, and is unsuitable for operation in rainy or foggy weather. Existing machine-aided inspection methods rely on manually designed features such as SIFT and HOG, which have extremely poor robustness to wind turbine blade afterimages and photovoltaic reflections. Therefore, this invention deeply integrates drone technology with improved deep learning algorithms. Through targeted scenario adaptation and model optimization, it achieves efficient, accurate, and real-time detection of insulator faults in wind and solar transmission lines, significantly reducing maintenance costs and safety risks.

[0022] Specifically, this invention uses a professional drone equipped with a high-definition RGB camera to fly along a preset route on both sides of the wind and solar power transmission line. It adopts a multi-altitude layered acquisition strategy, conducting wide-angle cruise photography at an altitude of 50-100 meters above the line to obtain global information about the line. In the area of ​​key equipment identified, it automatically descends to an altitude of 10-30 meters and adopts a multi-angle surround shooting mode, including orthogonal shooting, 45° oblique shooting, and 60° oblique shooting, to ensure that the surface details and potential defect features of the insulators are captured.

[0023] During flight, the camera avoids the rotation radius of wind turbine blades, maintains a safe distance, and avoids areas directly exposed to solar panel reflections. It ensures that the camera takes pictures at a small downward angle, while covering typical interference scenarios such as sunny days, cloudy days, rainy days, wind turbine blade shadows, solar panel reflections, and tree obstructions. Simultaneously, it records images and positioning data to generate an image dataset for insulator fault identification.

[0024] II. Image Preprocessing The environment along wind and solar power transmission lines combines the complexity of high-voltage transmission scenarios with the unique characteristics of new energy scenarios. On the one hand, drone aerial images are easily affected by meteorological interference such as raindrop noise in rainy weather and scattering blurring in foggy weather, and there are also uneven lighting issues such as local overbrightness caused by photovoltaic panel reflections and overall darkness caused by weak light on cloudy days. On the other hand, cluttered background factors such as tree obstruction can easily be confused with insulator targets. In addition, the insulator itself has a ring-shaped umbrella skirt structure, and directly stretching the image can easily lead to structural distortion. The combination of these problems means that if the images are directly input into target detection or segmentation models, serious false detections will occur due to issues such as noise masking faults and insufficient feature contrast.

[0025] Traditional general image preprocessing methods are not designed for the special imaging conditions of insulators of wind and solar power transmission lines. Therefore, this invention proposes a multi-level adaptive preprocessing process of "spatial denoising-structure conformity preservation-detail enhancement" for UAV inspection of insulators of wind and solar power transmission lines. This process provides high-quality input image data for the model and lays the foundation for accurate recognition. Specifically, it includes the following three steps.

[0026] 1. Adaptive median filtering for noise reduction Landscape inspections often encounter extreme weather: raindrops hitting the lens in rainy weather produce "salt and pepper noise," and foggy weather causes "Gaussian noise" due to air scattering. Both types of noise can make small damage difficult to detect. Traditional fixed-window filtering can remove noise, but it smooths out the grayscale gradient at the edges of the damage. Adaptive median filtering, by dynamically adjusting the window, can filter out noise while preserving the original image details, making it the optimal choice for landscape scenes. This processing aims to solve the interference of weather noise and preserve the edge details of the insulator damage.

[0027] First, set the initial window size to 3×3 and the maximum size to 7×7. A smaller window can reduce edge blurring. In each median filtering operation, first set the window size to the initial value, calculate the minimum, median, and maximum pixel values ​​in the current window. For each pixel in the image: if the gray value of the pixel is between the minimum and maximum values, it means there is no noise and it is directly retained; if there are pixels whose pixel values ​​are out of range, increase the window size and recalculate until the pixel value of each pixel in the image is between the minimum and maximum values ​​of the window pixel values. If, after the window is increased to the maximum, there are still pixels that are not between the minimum and maximum values ​​of the window pixel values, they are identified as noise pixels and replaced with the window median. The core of the whole process is to use dynamic judgment and precise replacement to avoid blindly removing filtering damage. The core logic formula for pixel replacement is shown in Equation (1).

[0028] (1) in, These are the original pixel values. These are the pixel values ​​after noise reduction. This is the current filtering window. , , These are the minimum, median, and maximum pixel values ​​within the window, respectively.

[0029] In the image preprocessing stage of UAV inspection, adaptive median filtering addresses the motion blur caused by the high-speed flight of the UAV. While preserving the details of minor damage edges, it can improve the removal rate of blade afterimages and dust noise to 87%, preventing fault features from being obscured by dynamic interference.

[0030] 2. Proportional scaling and black fill like Figure 1 As shown, the insulator has a "ring-shaped skirt" structure, while drone aerial images are mostly rectangular. However, the subsequent improvement of the TransUNet model requires a 512×512 square input. If the rectangular image is directly stretched, the ring-shaped skirt will become elliptical. Small damage that is originally at the edge of the skirt may be misjudged as being inside the skirt, seriously affecting the segmentation accuracy. Therefore, it is necessary to follow the steps of scaling proportionally first and then symmetrically filling to preserve the ring-shaped structure of the insulator.

[0031] The first step is to determine the aspect ratio of the image: If the wind power line is photographed horizontally, the width of the image is greater than its height, so the target width is 512, and the height is reduced proportionally; if the photovoltaic surrounding line is photographed vertically, the height is greater than its width, so the target height is 512, and the width is reduced proportionally. The core of scaling is "proportional" to avoid structural deformation. The proportional calculation logic can be simplified as shown in formula (2): (2) in, It's the scaling ratio. This is the target size required by the model; here, 512×512 is used. It is the baseline dimension (width or height) of the original image.

[0032] The second step involves black filling: Calculate the size difference between the scaled image and the 512×512 image, and symmetrically fill the image with black pixels (grayscale value 0) on the left and right or top and bottom to ensure that the insulator remains centered in the image and that subsequent positioning errors are not caused by fill position shifts. This process avoids shape distortion and false detection of edge breakage of the insulator skirts during image stretching, perfectly adapts to the skirt deformation problem caused by drone tilt shooting, maintains the integrity of the ring insulator structure, ensures that the fault position does not shift, and provides a data foundation for subsequent accurate segmentation of small fault areas.

[0033] 3. Retinex illumination correction The lighting conditions in wind and solar power transmission lines are highly unique. Reflections from solar panels can cause localized overexposure, resulting in pixels with grayscale values ​​exceeding 200, obscuring the blurred textures of weak flashover. During inspections in cloudy weather, the overall image is dark, reducing the contrast at the edges of small damage areas, thus impairing the model's recognition capabilities. Traditional brightness adjustments are prone to overexposure, while the Retinex algorithm, specifically designed for the strong lighting variations in wind and solar power transmission lines, can separate lighting interference from target features, specifically enhancing fault areas. This is crucial for resolving uneven lighting on insulators in wind and solar power transmission lines.

[0034] First, the grayscale information of the image is decomposed into "illuminance component" and "reflection component", which represent interference such as photovoltaic reflection and weak light on cloudy days, and the damage and flashover characteristics of the insulator itself, respectively. The relationship between the two is "original image = illumination component × reflection component". After extracting the illumination component by Gaussian filtering (smoothing), the illumination component is subtracted from the original image to separate the reflection component containing the fault characteristics, as shown in formula (3).

[0035] (3) in, For the original image in coordinates The pixel grayscale value at that location, It is the light component. This is the reflection component, which represents the characteristics of an insulator fault. The logarithm is used to facilitate separation and calculation.

[0036] Finally, the brightness range of the reflected component is mapped to 0-255, which expands the grayscale difference between weak flashover and normal umbrella skirt, and significantly improves the contrast of small damaged edges.

[0037] Figure 3 The following is an illustration of the insulator flashover identification effect of the present invention: Through the above-mentioned targeted preprocessing process, the weather noise removal rate is significantly improved, the image clarity in foggy weather is improved by 15-20dB, and the gray-scale gradient retention rate of the damaged edge reaches more than 90%. The deformation rate of the insulator annular skirt is reduced by 10%, and the gray-scale contrast of the weak flashover area is improved by 40%. This ensures that the UAV can collect identifiable fault features at inspection heights of 60-150 meters, laying a structural foundation for accurate model segmentation.

[0038] III. TransUNet Network Model Design and Insulator Fault Identification With the deepening of smart grid construction, drone power line inspection has become a core means of identifying the condition of insulators on transmission lines due to its high efficiency and flexibility. However, aerial images taken in wind and solar power scenarios present three major technical challenges: first, the scale differences of fault targets are significant; second, the background interference caused by photovoltaic panel reflections and wind turbine blade shadows is strong; and third, the computing power of drone edge devices is limited. These factors place stringent requirements on the detection models for both "lightweight deployment" and "high-precision segmentation." Existing mainstream methods are mostly based on optimizations of the Unet series models, but they are insufficient for fine tasks such as weak flashover texture recognition of insulators and pixel-level localization of small damage, making it difficult to adapt to the special inspection needs of wind and solar power transmission lines.

[0039] Based on the above challenges, this invention proposes a multi-scale fault fine segmentation model for UAV inspection scenarios—IF-TransUNet, which achieves a balance between lightweight and high accuracy through three core improvements.

[0040] 1. TransUNet Network Model Design and Training (1) Model building and training preparation A depthwise separable convolutional network is used to reconstruct the backbone and decoding module. An IF-DFAM attention mechanism is embedded in the feature fusion nodes of the encoder and decoder, and a cross-layer dynamic feature fusion module is designed to form the IF-TransUNet model. After completing the model structure configuration, the corresponding training scripts and configuration files are prepared.

[0041] (2) Dataset preparation and annotation Aerial images of wind and solar power transmission lines are stored in a designated directory. The labelme tool is used to annotate the bounding boxes of insulators in the images, generating corresponding mask images, which are then divided into training and testing sets.

[0042] (3) Model training After configuring the device category information file, load the pre-trained weights, use the training script to train the IF-TransUNet model, and obtain the optimized weight file.

[0043] 2. TransUNet Network Model Detection and Image Extraction The preprocessed insulator images are input into the TransUNet network model. The trained model is used to perform inference detection on the insulator fault dataset, segment the insulator contours, filter out insulator images containing faults, and label the fault type and confidence level to prepare for subsequent fault quantification assessment. The processing steps of the model include the following (1)-(6): (1) Input layer receives preprocessed feature map: The model input layer receives a 512×512 pixel three-channel RGB image after preprocessing.

[0044] (2) The depth-separable convolutional reconstruction network backbone and decoding module process the input feature map and complete the channel mapping.

[0045] The first core improvement of the TransUNet network model in this invention is the reconstruction of the network backbone and decoding module by the depthwise separable convolution: the ordinary convolution of the original TransUNet needs to process "spatial dimension feature extraction" and "channel dimension information integration" at the same time, resulting in a huge number of parameters and making it impossible to deploy on edge devices; while the depthwise separable convolution of this invention, through the split design of "depth convolution + pointwise convolution", significantly reduces the number of model parameters and computational cost while retaining key features such as the edge of the insulator skirt and small damaged texture.

[0046] Depthwise convolution uses a single-channel convolution kernel to process only the corresponding channel of the input feature map. The insulator damage edge channel uses only a single 3×3 convolution kernel for spatial convolution, focusing on local feature extraction in the spatial dimension and extracting independent planar spatial features for each channel without involving redundant calculations between channels. In pointwise convolution, a 1×1 convolution kernel is used to integrate the channel dimensions of the output feature map of depthwise convolution, such as fusing features from the damage edge channel and the umbrella skirt contour channel, and channel mapping can be completed with only a small number of parameters.

[0047] The number of parameters in a regular convolution is determined by the number of input channels, the number of output channels, and the size of the convolution kernel, as shown in formula (4); the number of parameters in a depthwise separable convolution is shown in formula (5).

[0048] (4) (5) in, These are the parameters for a normal convolution; The total number of parameters for depthwise separable convolution; The number of parameters for depthwise convolution; The number of parameters for pointwise convolution; The number of channels in the input insulator feature map; The number of channels in the output feature map; and These represent the height and width of the depthwise convolution kernel, respectively, while the size of the pointwise convolution kernel is fixed at 1×1.

[0049] like =64, =128, =3, then the number of parameters in ordinary convolution is 3. =73728, total parameters of depthwise separable convolution. =8768. Calculations using the above formula show that depthwise separable convolution reduces the number of parameters by approximately 80% compared to ordinary convolution, thus reducing the number of model parameters and making it more suitable for the computing power of edge devices like drones. Depthwise convolution focuses only on local features in a single channel, improving the gradient preservation rate of small damage edges on insulators and providing accurate detail support for subsequent segmentation.

[0050] (3) The feature map is convolved in the direction of the IF-DFAM attention mechanism to capture edge features at different angles, and high-frequency components are extracted by Fourier transform to enhance fault features.

[0051] The IF-DFAM attention mechanism is the second core improvement of the TransUNet network model in this invention: Addressing the shortcomings of strong photovoltaic panel reflections and difficulty in locating small faults in the ring structure of insulator skirts in wind and solar transmission lines, this invention specifically improves the original CBAM attention mechanism by adding a direction-aware branch and a frequency-attention branch—embedding the IF-DFAM attention mechanism, suitable for insulator fault detection, into the feature fusion node of the encoder and decoder. This mechanism dynamically strengthens the channel containing the fault feature through "channel attention" while suppressing irrelevant background channels. It also enhances the pixel weights in areas of small damage and weak flashover through "spatial attention," reducing interference from photovoltaic reflections and blade afterimages.

[0052] (3)-1: Direction-aware branch The insulator skirts are arranged in a periodic ring, and small faults are mostly distributed in the tangential direction of the ring edge. Therefore, the direction-aware branch introduces a Gabor filter group to perform directional convolution on the feature map output by the channel attention, capturing edge features at different angles. In view of the irregular linear characteristics of insulator damage and cracks, multi-scale directional convolution kernels are used to scan the feature map in the horizontal, vertical and diagonal directions respectively to capture edge texture features at different angles, solving the problem of insufficient sensitivity of traditional square convolution kernels to slender cracks, as shown in formula (6).

[0053] (6) F′ is the feature map input to the direction-aware branch of the IF-DFAM attention mechanism, and k is the direction index, which takes values ​​from 1 to 8, representing 8 different directions.

[0054] in, For the first Gabor filters in one direction, focusing on the tangential direction of the ring edge; As directional weights, they are automatically assigned through learning, specifically for the direction of the ring edge. Significantly higher than other directions; This is the feature map after directional enhancement.

[0055] (3)-2: Frequency attention branch The reflective area has low-frequency smooth features, while the fault area has high-frequency edge features such as dotted or linear textures. High-frequency components are extracted by Fourier transform to enhance the fault features and transform them from the spatial domain to the frequency domain. High-frequency components representing edge abrupt changes and texture details are extracted using a high-pass filter, while low-frequency components representing smooth backgrounds, such as illuminated areas, are suppressed. Then, the inverse Fourier transform is used to restore the spatial domain, thereby significantly enhancing the feature response of weak flashover traces, as shown in formula (7).

[0056] (7) Wherein, FFT and IFFT are the Fourier transform and inverse transform, respectively. For high-frequency mask, This is the feature map after high-frequency enhancement.

[0057] Finally, the feature maps extracted from the two branches are used to generate spatial and channel weight maps through the Sigmoid activation function, and then multiplied with the original feature maps. This increases the weights of pixels in the fault region while suppressing interference from bright or complex background noise such as photovoltaic panel reflections and leaf shadows.

[0058] In summary, in the IF-DFAM attention mechanism of this invention, the direction perception branch captures the tangential direction features of the insulator skirt through Gabor filtering, filtering out the linear afterimages of wind turbine blades, thereby improving the edge response intensity of small insulator damage and reducing positioning errors. The frequency attention branch filters low-frequency reflective components and suppresses blocky reflections from photovoltaic panels, reducing the weight of reflective areas and maintaining a higher weight for fault areas. This enhances the model's ability to capture key fault features of insulators, while suppressing background interference such as wind turbine blade rotation afterimages, photovoltaic panel reflections, and tree shading, reducing the risk of missing small faults. This ensures that the fault identification accuracy remains stable at 90% in complex scenarios, with a 20% increase in recall, significantly reducing the cost of manual review.

[0059] IF-DFAM is not a general attention mechanism, but a scenario-specific design that is deeply bound to the annular shed structure characteristics of insulators, edge point fault features, and photovoltaic environment reflection or light interference. Through multi-faceted adjustment of "direction-frequency", it solves the core problems of the original model in wind and solar transmission line scenarios, such as difficulty in distinguishing faults from interference and poor dynamic environment adaptability.

[0060] (4) Cross-layer feature dynamic fusion module: In response to the multi-scale characteristics of insulator faults, this module breaks through the limitation of traditional skip connections in the same layer splicing and dynamically weights and fuses the shallow high-resolution features, same-layer features, and deep high-semantic features of the encoder. By adaptively allocating weights through feature information entropy and combining residual connections to retain historical features, the information entropy of features at different levels is calculated. The shallow high-resolution features, same-layer features, and deep high-semantic features of the encoder are adaptively weighted and fused, which effectively solves the semantic gap problem in traditional skip connections and ensures that more small target fault details are preserved when restoring the image size, so that the model can adapt to the segmentation requirements of faults at different scales at the same time.

[0061] (5) Decoder feature reconstruction and upsampling: The fused feature map enters the decoder path, where bilinear interpolation is used to upsample and restore the feature map size. Then, the depthwise separable convolution module is used again to decode and refine the features, gradually restoring the abstract semantic features to pixel-level feature maps with the same resolution as the input image.

[0062] (6) Output layer generates segmentation mask: Finally, the decoded feature map is channel-mapped through a 1×1 convolutional layer, and the final prediction result is output using an activation function to generate a binary mask, thus completing the pixel-level localization and segmentation of the insulator body and the fault area.

[0063] IV. Data Display and Storage The fault insulator image information filtered by the IF-TransUNet model is integrated with the GPS information in the original image, and compiled into an Excel file using Python for easy querying. It can also be combined with a MySQL database for data management.

[0064] In summary, the IF-TransUNet-based insulator condition identification method for transmission lines of this invention addresses the industry pain points of "poor image quality, difficult model deployment, and many missed minor faults" in UAV inspections of wind and solar transmission lines. It also addresses the specific needs of UAV inspections of wind and solar transmission lines by constructing a full-process identification system encompassing "data acquisition - preprocessing - model inference - result output." This upgrades the method from manual sampling to fully autonomous and accurate detection, replacing traditional manual inspections. Through deep collaboration between image preprocessing and model improvement, it comprehensively enhances the efficiency and accuracy of insulator fault identification in wind and solar transmission lines, achieving a breakthrough in the industrial application of insulator fault detection. This ensures the safe and stable operation of new energy power transmission channels and provides efficient and reliable intelligent technical support for the safe operation and maintenance of wind and solar power plants.

[0065] Although the present invention has been described using the above preferred embodiments, it is not intended to limit the scope of protection of the present invention. Any changes and modifications made by those skilled in the art to the above embodiments without departing from the spirit and scope of the present invention shall still fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be defined by the claims.

Claims

1. A power line insulator state recognition method based on IF-TransUNet, characterized in that, The method comprises the following steps: Step S1: flying along the preset flight route above the both sides of the wind and light transmission line by the unmanned aerial vehicle, adopting a multi-height layered collection strategy, and performing wide-angle cruising shooting at a predetermined height from the wind and light transmission line to generate an image dataset for insulator fault identification; Step S2: preprocessing the image dataset; Step S3: inputting the preprocessed image dataset into a TransUNet network model for insulator fault identification; The building method of the TransUNet network model comprises the following steps: reconstructing a network backbone and a decoding module by using a deep separable convolution, embedding an IF-DFAM attention mechanism at a feature fusion node of an encoder and a decoder, and designing a cross-layer feature dynamic fusion module to form an IF-TransUNet model.

2. The IF-TransUNet-based power transmission line insulator state recognition method of claim 1, wherein: In the step S2, the preprocessing of the image dataset comprises adaptive median filter denoising, equal proportion scaling and black filling, and Retinex illumination correction.

3. The IF-TransUNet-based power transmission line insulator state recognition method of claim 2, wherein: The equal proportion scaling and black filling comprise the following steps: For a horizontally shot insulator image, reducing the height in proportion to the target width; For a vertically shot insulator image, reducing the width in proportion to the target height; Calculating the size difference of the scaled insulator image, and filling black pixels symmetrically on the left and right or above and below of the image.

4. The IF-TransUNet-based power transmission line insulator state recognition method of claim 2, wherein: The Retinex illumination correction comprises the following steps: Decomposing the gray information of the insulator image into a "light component" and a "reflection component", extracting the light component by Gaussian filtering, and then subtracting the light component from the original image to separate out the reflection component containing the fault features; Mapping the brightness of the reflection component to a predetermined range.

5. The IF-TransUNet-based power transmission line insulator state recognition method of claim 1, wherein: The step S3 comprises the following steps: Step S31: processing the input feature map by using the deep separable convolution to reconstruct the network backbone and the decoding module, and completing channel mapping; Step S32: performing directional convolution on the feature map by using the IF-DFAM attention mechanism to capture edge features at different angles, and extracting high-frequency components from the feature map by using Fourier transform to strengthen the fault features; Step S33: dynamically weighting and fusing the shallow high-resolution features, the same layer features and the deep high-semantic features of the feature map by using the cross-layer feature dynamic fusion module.

6. The IF-TransUNet-based power transmission line insulator state recognition method of claim 5, wherein: In the step S31, the deep separable convolution reconstructs the network backbone and the decoding module by using a "depth convolution + point-by-point convolution" split design; wherein the depth convolution only processes the corresponding channels of the input feature map, focuses on local feature extraction in the spatial dimension, and the point-by-point convolution integrates the feature map output by the depth convolution in the channel dimension.

7. The IF-TransUNet-based power transmission line insulator state recognition method of claim 5, wherein: In the step S32, the IF-DFAM attention mechanism comprises a direction perception branch and a frequency attention branch; the direction perception branch performs directional convolution on the feature map to capture edge features at different angles, and the frequency attention branch extracts high-frequency components from the feature map by using Fourier transform to strengthen the fault features.

8. The IF-TransUNet-based power transmission line insulator state recognition method of claim 7, wherein: The formula for the direction perception branch to perform directional convolution on the feature map is: ； wherein, is the Gabor filter for the th direction, focusing on the tangential direction of the annular edge; is the direction weight, automatically assigned by learning, the annular edge direction is significantly higher than other directions; is the direction enhanced feature map, F' is the feature map input to the IF-DFAM attention mechanism direction perception branch, k is the index of the direction, taking values from 1 to 8, representing 8 different directions.

9. The IF-TransUNet-based power transmission line insulator state recognition method of claim 7, wherein: The formula for the frequency attention branch to extract high-frequency components from the feature map by using Fourier transform is: ； wherein FFT and IFFT are Fourier transform and inverse transform, respectively, is a high-frequency mask, is a high-frequency enhanced feature map.

Images