Power distribution network fault identification method based on multi-spectral data of unmanned aerial vehicle

Abstract

The invention discloses a power distribution network fault identification method based on unmanned aerial vehicle multispectral data. The method comprises the following steps: S1, acquiring power distribution network original multispectral data acquired by a multispectral sensor array carried by an unmanned aerial vehicle; s2, performing primary processing such as data synchronization, image registration and format conversion on the original multispectral data of the power distribution network in different wavebands to obtain primarily processed multispectral data; s3, constructing an adaptive deep learning denoising model and performing pre-training, and performing denoising and fault feature enhancement processing on the preliminarily processed multispectral data by using the pre-trained adaptive deep learning denoising model to generate target denoised multispectral data; and S4, transmitting the target de-noised multispectral data to a fault identification and analysis module, and identifying the fault type of the power distribution network. According to the power distribution network fault identification method based on the unmanned aerial vehicle multispectral data provided by the invention, the problem that noise and weak fault features are difficult to distinguish in a processing method in the prior art is solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault identification technology for power distribution networks, and specifically relates to a method for identifying power distribution network faults based on UAV multispectral data. Background Technology

[0002] With the rapid development of power systems, the scale of distribution networks is expanding daily, and their safe and stable operation is crucial to social production and people's lives. Traditional distribution network inspection methods rely on manual inspections, which are inefficient and pose safety risks. In recent years, the combination of drone technology and multispectral imaging technology has provided a new solution for intelligent distribution network inspection. Drones equipped with multispectral sensors can acquire spectral information of distribution network equipment, such as lines, insulators, and towers, in multiple bands, including visible light, near-infrared, and thermal infrared, providing a rich data source for fault identification.

[0003] However, when drones collect multispectral data in the field, they are highly susceptible to interference from various environmental factors, such as uneven lighting, cloud cover, weather changes, obstruction by vegetation or obstacles, and sensor noise. This results in a large amount of noise, artifacts, and outliers in the raw multispectral data. This low-quality data severely affects the accuracy and robustness of subsequent AI-based fault identification algorithms, potentially leading to false alarms, missed alarms, or even masking subtle fault characteristics.

[0004] Existing data preprocessing methods, such as simple filtering and thresholding, often struggle to effectively distinguish noise from the subtle fault characteristics of distribution network equipment. They may lose useful fault information while removing noise, or lack the ability to handle complex noise patterns. Therefore, how to efficiently and adaptively denoise and enhance the features of UAV multispectral data to improve the signal-to-noise ratio and identification accuracy of fault features is a key technical problem that urgently needs to be solved in the field of intelligent distribution network inspection. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying power distribution network faults based on UAV multispectral data, which solves the problem that existing processing methods have difficulty distinguishing between noise and weak fault characteristics.

[0006] The technical solution adopted in this invention is a power distribution network fault identification method based on UAV multispectral data, as detailed below: S1. Acquire raw multispectral data of the power distribution network collected by the multispectral sensor array carried by the UAV; S2. Perform preliminary processing on the raw multispectral data of the distribution network in different bands, such as data synchronization, image registration and format conversion, to obtain the preliminary processed multispectral data; S3. Construct an adaptive deep learning denoising model and pre-train it. Use the pre-trained adaptive deep learning denoising model to denoise and enhance fault features on the pre-processed multispectral data to generate target denoised multispectral data. S4. Transmit the target denoised multispectral data to the fault identification and analysis module to identify the fault types of the power distribution network.

[0007] The invention is further characterized by: The multispectral sensor array in S1 includes a visible light camera, a near-infrared camera, or a thermal infrared camera; the raw multispectral data of the power distribution network includes image information of power distribution lines, towers, and equipment in different spectral bands, and the image information contains noise and potential fault characteristics; During the data acquisition process, the multispectral sensor array synchronously captures images at a frequency of 3 frames per second. At the same time, it records the GPS location information (longitude, latitude, altitude) and UAV attitude information (pitch angle, roll angle, and heading angle) corresponding to each frame of image. This information is then linked to the original image data through timestamps to form raw multispectral data with geospatial labels.

[0008] The methods for preliminary processing of raw multispectral data in S2 also include: Radiometric calibration is performed on the raw multispectral data of the power distribution network, and atmospheric correction is performed on the radiometrically calibrated data. In the initial processing stage, data synchronization adopts the timestamp alignment method. Based on the system time of the UAV, image data collected by different sensors are matched according to the standard that the timestamp difference is less than 10 milliseconds to ensure the correspondence of data in different bands at the same time. Image registration adopts an improved SIFT algorithm. First, feature points of each band image are extracted and matched by the FLANN matcher. Then, the random sampling consensus algorithm is used to eliminate mismatches. Finally, bilinear interpolation is used to perform geometric transformation on the image so that the average reprojection error of the registered image is controlled within 1 pixel. The format conversion converts the original RAW format image to 16-bit TIFF format to retain more spectral information.

[0009] The adaptive deep learning denoising model built in S3 is a generative adversarial network model, which includes a generator and a discriminator. The generator is used to learn the mapping relationship from the noisy pre-processed multispectral data to the noise-free target denoised multispectral data, and to generate the target denoised multispectral data. After the target denoised multispectral data is generated, a data index is established, and the target denoised multispectral data, along with the corresponding original multispectral data, processing time, and geographic location information, are stored in the database built by the data storage and management module for historical data tracing and model iteration. The discriminator is used to learn to distinguish between real noise-free multispectral data and target denoised multispectral data generated by the generator, and outputs the discrimination result to guide the generator's training process.

[0010] The generator adopts the U-Net architecture, which includes an encoder path and a decoder path. The encoder path contains 6 convolutional blocks, each consisting of two 3×3 convolutional layers and a 2×2 max pooling layer. The stride of the convolutional layers is 1, and a non-linear transformation is performed using the ReLU activation function. After each convolutional block, the size of the feature map is halved, and the number of channels is doubled. The decoder path also contains 6 convolutional blocks, each consisting of two 3×3 transposed convolutional layers. The stride of the transposed convolutions is 2, which upsamples the feature map. At the same time, the feature maps of the corresponding layers in the encoder path are fused through skip connections to preserve details. Finally, a 1×1 convolutional layer converts the feature map into 3-channel target denoised multispectral data with the same size as the input. The discriminator adopts the PatchGAN architecture, consisting of 5 convolutional blocks. Each convolutional block contains a 4×4 convolutional layer with a stride of 2. It is processed by the LeakyReLU activation function with a negative slope of 0.2. Finally, it outputs a 64×64 discrimination result image through a 1×1 convolutional layer. During the discrimination process, the input image is divided into 70×70 image blocks. Each image block is processed by the discriminator to obtain a discrimination value. The average of the discrimination values ​​of all image blocks is taken as the final output, which is used to determine the authenticity of the input image.

[0011] The specific method for pre-training the adaptive deep learning denoising model in S3 is as follows: S3.1. Pre-collect and construct a training dataset. The training dataset includes N sets of data pairs, where N is a positive integer greater than 1. Each set of data pairs includes a noisy multispectral image and a corresponding real noise-free multispectral image. S3.2. Divide the training dataset into a training set and a validation set in an 8:2 ratio; S3.3. Use the noisy multispectral images in the training set as input to the generator, compare the generator output with the corresponding real noise-free multispectral images in the training set, and calculate the generation loss. S3.4 Input the image output by the generator and the real noise-free images in the training set into the discriminator. The discriminator calculates the discrimination loss through the binary cross-entropy loss function to distinguish between real and fake images. Its feedback is used to jointly optimize the parameters of the generator and the discriminator. S3.5. An adversarial training mechanism is adopted to jointly optimize the network parameters of the generator and the discriminator by minimizing the generation loss and maximizing the discrimination loss. S3.6. Evaluate the model performance on the validation set until the performance index of the generative adversarial network model reaches the preset convergence threshold, then stop training and obtain the pre-trained adaptive deep learning denoising model.

[0012] In S3.3, the generation loss is calculated by comparing the pixel-level mean square error between the denoised multispectral data of the target generated by the generator and the real noise-free multispectral image. The specific calculation method is as follows: The generator takes a pre-processed noisy multispectral image as input and performs feature extraction and reconstruction through an encoder-decoder structure. The mapping relationship of the generator can be represented as follows: (1); in, The noisy multispectral data after preliminary processing. For the network parameters of the generator, This generates denoised multispectral data for the target. The discriminator then judges the authenticity of the input image patches, and its output is: (2); in, The input image patch can be a real, noise-free image patch or an image patch output by the generator. For the network parameters of the discriminator, This is the feature extraction and nonlinear transformation function of the discriminator. The output value ranges from (0,1), and the closer it is to 1, the more likely it is to be a real noise-free image patch. During model training, the generation loss uses mean squared error loss: (3) in, , , These are the image's height, width, and number of spectral channels, respectively. The generator outputs the image at the 1st line, number Column, No. Pixel values ​​of the channel, The pixel values ​​correspond to real noise-free images, and the units are all DN values ​​(digital quantization values).

[0013] The discriminant loss in S3.4 is calculated using the binary cross-entropy loss function: (4); in, The number of image patches in the training batch. Label 1 represents a real image patch, and 0 represents a generated image patch. For the first Image patches. By alternately optimizing the generator and discriminator, the generator can effectively remove noise while preserving fault features.

[0014] The specific method for identifying fault types in the distribution network in S4 is as follows: S4.1 Receiving Target Denoising Multispectral Data: When transmitting target denoising multispectral data to the fault identification and analysis module, fiber optic communication is used, and the transmission rate is not less than 100Mbps. S4.2 Analyze the denoised multispectral data of the target using the target detection model and image classification model to identify the fault types of the power distribution network. The fault types include insulator damage, foreign object hanging on the line, tower corrosion, conductor strand breakage, and abnormal heating at line connection points.

[0015] The beneficial effects of this invention are: This invention presents a power distribution network fault identification method based on UAV multispectral data. It utilizes a UAV equipped with a multispectral sensor array to collect multi-band data from the power distribution network. Preliminary processing, including data synchronization, registration, radiometric calibration, and atmospheric correction, improves data consistency. A pre-trained generative adversarial network model is then used for adaptive denoising and enhancement of fault features. This effectively solves the problem of traditional preprocessing methods struggling to distinguish between noise and weak fault features. It removes environmental interference such as light and clouds while retaining key fault information such as insulator damage and conductor strand breakage, providing high-quality data for subsequent fault identification and improving accuracy and robustness. Furthermore, data storage management supports historical traceability and model iteration, making it more efficient and secure than manual inspection, providing reliable technical support for intelligent power distribution network inspection. Attached Figure Description

[0016] Figure 1 This is a flowchart of the power distribution network fault identification method based on UAV multispectral data of the present invention; Figure 2 This is a structural diagram of the power distribution network fault identification system based on UAV multispectral data according to the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0018] This invention relates to a power distribution network fault identification method based on UAV multispectral data, such as... Figure 1 As shown, the details are as follows: S1. Acquire raw multispectral data of the power distribution network collected by the multispectral sensor array carried by the UAV; The multispectral sensor array includes a visible light camera, a near-infrared camera, or a thermal infrared camera; the original multispectral data of the power distribution network contains image information of the power distribution network lines, towers, and equipment in different spectral bands, and the image information contains noise and potential fault characteristics; The specific method for acquiring raw multispectral data of the power distribution network collected by a multispectral sensor array carried by a drone is as follows: Plan the inspection routes for drones, covering the power distribution network areas to be inspected; Control the drone to fly along the inspection route and instruct the multispectral sensor array to synchronously or quasi-synchronously collect data on power distribution lines, towers and equipment during the flight; The raw image data collected in each band is correlated with the position and attitude information of the UAV to form raw multispectral data with geospatial labels.

[0019] During the data acquisition process, the multispectral sensor array synchronously captures images at a frequency of 3 frames per second. At the same time, it records the GPS location information (longitude, latitude, altitude) and UAV attitude information (pitch angle, roll angle, and heading angle) corresponding to each frame of image. This information is then linked to the original image data through timestamps to form raw multispectral data with geospatial labels.

[0020] S2. Perform preliminary processing on the raw multispectral data of the distribution network in different bands, such as data synchronization, image registration and format conversion, to obtain the preliminary processed multispectral data; Further methods for preliminary processing of raw multispectral data include: Radiometric calibration is performed on the raw multispectral data of the power distribution network to convert the digital quantization values ​​recorded by the sensors into physically meaningful radiance values. Atmospheric correction is performed on the radiometrically calibrated data to eliminate the influence of atmospheric scattering and absorption on spectral information. Specifically, radiometric calibration employs a combination of laboratory and field calibration. Laboratory calibration obtains the sensor's response function; for visible and near-infrared cameras, the formula for calculating radiance values ​​is: , in, This is the radiance value, measured in watts per square meter. sphericity micrometer, The quantization value is 0-65535. This is the gain factor, measured in watts per square meter. sphericity micrometer DN, This is the offset, in watts per square meter. sphericity Micrometers. For example, visible light cameras. The value is 0.001. The value is set to 0.1. The thermal infrared camera obtains the conversion relationship between radiance and DN value through linear fitting based on blackbody calibration data.

[0021] Atmospheric correction employs a method based on radiative transfer models, such as the 6S model. First, atmospheric parameters at the time of imaging are acquired, including atmospheric aerosol optical thickness, water vapor content, and ozone content. These can be measured in real-time using atmospheric sensors mounted on a drone or obtained from weather stations. Then, the 6S model is used to calculate atmospheric transmittance and atmospheric path radiatives, and atmospheric correction is performed using the following formula: , in, The value of radiance reflected by the Earth's surface, measured in watts per square meter. sphericity micrometer, The values ​​are headspace radiance values ​​after radiometric calibration, in the same units as above. This refers to the atmospheric path radiance value, in the same units as above. Atmospheric transmittance.

[0022] By radiometric calibration and atmospheric correction, the original digital quantization values ​​are converted into physically meaningful surface radiance, eliminating spectral distortion caused by atmospheric scattering and absorption. This makes multispectral data collected at different times and locations comparable, improves the spectral consistency of the data, and provides a more reliable spectral basis for subsequent fault feature extraction.

[0023] In the initial processing stage, data synchronization adopts the timestamp alignment method. Based on the system time of the UAV, image data collected by different sensors are matched according to the standard that the timestamp difference is less than 10 milliseconds to ensure the correspondence of data in different bands at the same time. Image registration adopts an improved SIFT algorithm. First, feature points of each band image are extracted and matched by the FLANN matcher. Then, the random sampling consensus algorithm is used to eliminate mismatches. Finally, bilinear interpolation is used to perform geometric transformation on the image so that the average reprojection error of the registered image is controlled within 1 pixel. The format conversion converts the original RAW format image to 16-bit TIFF format to retain more spectral information.

[0024] S3. Construct and pre-train an adaptive deep learning denoising model. Use the pre-trained adaptive deep learning denoising model to denoise and enhance fault features on the initially processed multispectral data to generate target denoised multispectral data. The specific method is as follows: S3.1. Pre-collect and construct a training dataset. The training dataset includes N sets of data pairs, where N is a positive integer greater than 1. Each set of data pairs includes a noisy multispectral image and a corresponding real noise-free multispectral image. The adaptive deep learning denoising model constructed is a generative adversarial network model, which includes a generator and a discriminator. Furthermore, the generator is used to learn the mapping relationship from the noisy pre-processed multispectral data to the noise-free target denoised multispectral data, and to generate the target denoised multispectral data. After the target denoised multispectral data is generated, a data index is established, and the target denoised multispectral data, along with the corresponding original multispectral data, processing time, and geographic location information, are stored in the database built by the data storage and management module for historical data tracing and model iteration. The generator adopts a U-Net architecture, which includes an encoder path and a decoder path. The encoder path contains 6 convolutional blocks, each consisting of two 3×3 convolutional layers and a 2×2 max pooling layer. The stride of the convolutional layers is 1, and a non-linear transformation is performed using the ReLU activation function. After each convolutional block, the size of the feature map is halved, and the number of channels is doubled. The decoder path also contains 6 convolutional blocks, each consisting of two 3×3 transposed convolutional layers. The stride of the transposed convolutions is 2, which upsamples the feature map. At the same time, the feature map is fused with the feature map of the corresponding layer of the encoder path through skip connections to preserve details. Finally, a 1×1 convolutional layer converts the feature map into 3-channel target denoised multispectral data with the same size as the input.

[0025] The discriminator is used to learn to distinguish between real noise-free multispectral data and target denoised multispectral data generated by the generator, and outputs the discrimination result to guide the generator's training process.

[0026] The discriminator adopts the PatchGAN architecture, consisting of 5 convolutional blocks. Each convolutional block contains a 4×4 convolutional layer with a stride of 2. It is processed by the LeakyReLU activation function with a negative slope of 0.2. Finally, it outputs a 64×64 discrimination result image through a 1×1 convolutional layer. During the discrimination process, the input image is divided into 70×70 image blocks. Each image block is processed by the discriminator to obtain a discrimination value. The average of the discrimination values ​​of all image blocks is taken as the final output, which is used to determine the authenticity of the input image.

[0027] S3.2. Divide the training dataset into a training set and a validation set in an 8:2 ratio; S3.3. Use the noisy multispectral images in the training set as input to the generator, compare the generator output with the corresponding real noise-free multispectral images in the training set, and calculate the generation loss. Furthermore, the generation loss is calculated by comparing the pixel-level mean square error between the denoised multispectral data of the target generated by the generator and the real noise-free multispectral image. The specific calculation method is as follows: The generator takes a pre-processed noisy multispectral image as input and performs feature extraction and reconstruction through an encoder-decoder structure. The mapping relationship of the generator can be represented as follows: (1); in, The noisy multispectral data after preliminary processing. For the network parameters of the generator, This generates denoised multispectral data for the target. The discriminator then judges the authenticity of the input image patches, and its output is: (2); in, The input image patch can be a real, noise-free image patch or an image patch output by the generator. For the network parameters of the discriminator, This is the feature extraction and nonlinear transformation function of the discriminator. The output value ranges from (0,1), and the closer it is to 1, the more likely it is to be a real noise-free image patch. During model training, the generation loss uses mean squared error loss: (3) in, , , These are the image's height, width, and number of spectral channels, respectively. The generator outputs the image at the 1st line, number Column, No. Pixel values ​​of the channel, The pixel values ​​correspond to real noise-free images, and the units are all DN values ​​(digital quantization values).

[0028] When calculating the generation loss, the mean square error loss formula is used. The difference between the target denoised multispectral image output by the generator and the real noise-free multispectral image at each pixel position and each spectral channel is included in the calculation to ensure that the loss value can accurately reflect the fidelity of the generated image.

[0029] S3.4 Input the image output by the generator and the real noise-free images in the training set into the discriminator. The discriminator calculates the discrimination loss through the binary cross-entropy loss function to distinguish between real and fake images. Its feedback is used to jointly optimize the parameters of the generator and the discriminator. Furthermore, the discriminant loss is calculated using the binary cross-entropy loss function: (4); in, The number of image patches in the training batch. Label 1 represents a real image patch, and 0 represents a generated image patch. For the first Image patches. By alternately optimizing the generator and discriminator, the generator can effectively remove noise while preserving fault features.

[0030] The binary cross-entropy loss formula is adopted. The images output by the generator and the real noise-free images are batched and input into the discriminator in a 1:1 ratio. The loss is calculated by the difference between the output of the discriminator and the real label, so that the discriminator can effectively distinguish between the two.

[0031] S3.5. An adversarial training mechanism is adopted to jointly optimize the network parameters of the generator and the discriminator by minimizing the generation loss and maximizing the discrimination loss. S3.6. Evaluate the model performance on the validation set until the performance index of the generative adversarial network model reaches the preset convergence threshold, then stop training and obtain the pre-trained adaptive deep learning denoising model.

[0032] S4. Transmit the target denoised multispectral data to the fault identification and analysis module to identify the fault types in the distribution network, as follows: S4.1 Receiving Target Denoising Multispectral Data: When transmitting target denoising multispectral data to the fault identification and analysis module, fiber optic communication is used, and the transmission rate is not less than 100Mbps. S4.2 Analyze the denoised multispectral data of the target using the target detection model and image classification model to identify the fault types of the power distribution network. The fault types include insulator damage, foreign object hanging on the line, tower corrosion, conductor strand breakage, and abnormal heating at line connection points.

[0033] A power grid fault data preprocessing system based on UAV multispectral data, such as Figure 2 As shown, it includes: Data acquisition module: used to acquire raw multispectral data of the power distribution network collected by the multispectral sensor array carried by the UAV. The multispectral sensor array includes a visible light camera, a near-infrared camera and a thermal infrared camera. The raw multispectral data contains image information of power distribution lines, towers and equipment in different spectral bands, and the image information contains noise and potential fault characteristics. Preliminary processing module: Used to perform preliminary processing on the raw multispectral data to obtain preliminarily processed multispectral data. Preliminary processing includes data synchronization, image registration and format conversion of raw multispectral data of different bands; Adaptive deep learning denoising module: It is used to receive the pre-processed multispectral data and use a pre-trained adaptive deep learning denoising model to denoise and enhance the fault features of the pre-processed multispectral data to generate target denoised multispectral data. Data interface module: Used to transmit the target denoised multispectral data to the external fault identification and analysis module, so that the fault identification and analysis module can identify the fault type of the distribution network.

[0034] Example 1 This embodiment presents a power grid fault identification method based on UAV multispectral data, as detailed below: S1. Acquire raw multispectral data of the power distribution network collected by the multispectral sensor array carried by the UAV; S2. Perform preliminary processing on the raw multispectral data of the distribution network in different bands, such as data synchronization, image registration and format conversion, to obtain the preliminary processed multispectral data; S3. Construct an adaptive deep learning denoising model and pre-train it. Use the pre-trained adaptive deep learning denoising model to denoise and enhance fault features on the pre-processed multispectral data to generate target denoised multispectral data. S4. Transmit the target denoised multispectral data to the fault identification and analysis module to identify the fault types of the power distribution network.

[0035] Example 2 This embodiment is a power distribution network fault identification method based on UAV multispectral data. Based on embodiment 1, the multispectral sensor array in S1 includes a visible light camera, a near-infrared camera, or a thermal infrared camera; the original multispectral data of the power distribution network includes image information of power distribution network lines, towers, and equipment in different spectral bands, and the image information contains noise and potential fault features. During the data acquisition process, the multispectral sensor array synchronously captures images at a frequency of 3 frames per second. At the same time, it records the GPS location information (longitude, latitude, altitude) and UAV attitude information (pitch angle, roll angle, and heading angle) corresponding to each frame of image. This information is then linked to the original image data through timestamps to form raw multispectral data with geospatial labels.

[0036] Example 3 This embodiment of the power distribution network fault identification method based on UAV multispectral data, based on embodiment 1, further includes the following method in S2 for preliminary processing of the raw multispectral data: Radiometric calibration is performed on the raw multispectral data of the power distribution network, and atmospheric correction is performed on the radiometrically calibrated data. In the initial processing stage, data synchronization adopts the timestamp alignment method. Based on the system time of the UAV, image data collected by different sensors are matched according to the standard that the timestamp difference is less than 10 milliseconds to ensure the correspondence of data in different bands at the same time. Image registration adopts an improved SIFT algorithm. First, feature points of each band image are extracted and matched by the FLANN matcher. Then, the random sampling consensus algorithm is used to eliminate mismatches. Finally, bilinear interpolation is used to perform geometric transformation on the image so that the average reprojection error of the registered image is controlled within 1 pixel. The format conversion converts the original RAW format image to 16-bit TIFF format to retain more spectral information.

[0037] Example 4 This embodiment is a power grid fault identification method based on UAV multispectral data. Based on embodiment 1, the adaptive deep learning denoising model constructed in S3 is a generative adversarial network model, which includes a generator and a discriminator. The generator is used to learn the mapping relationship from the noisy pre-processed multispectral data to the noise-free target denoised multispectral data, and to generate the target denoised multispectral data. After the target denoised multispectral data is generated, a data index is established, and the target denoised multispectral data, along with the corresponding original multispectral data, processing time, and geographic location information, are stored in the database built by the data storage and management module for historical data tracing and model iteration. The discriminator is used to learn to distinguish between real noise-free multispectral data and target denoised multispectral data generated by the generator, and outputs the discrimination result to guide the generator's training process.

[0038] The generator employs a U-Net architecture, which includes an encoder path and a decoder path. The encoder path contains six convolutional blocks, each consisting of two 3×3 convolutional layers and a 2×2 max-pooling layer. The stride of the convolutional layers is 1, and a non-linear transformation is performed using the ReLU activation function. After each convolutional block, the size of the feature map is halved, while the number of channels is doubled. The decoder path also contains six convolutional blocks, each consisting of two 3×3 transposed convolutional layers. The stride of the transposed convolutions is 2, achieving upsampling of the feature maps. Simultaneously, the feature maps from the corresponding layers of the encoder path are fused through skip connections to preserve detailed information. Finally, a 1×1 convolutional layer transforms the feature maps into 3-channel target denoised multispectral data of the same size as the input. The discriminator adopts the PatchGAN architecture, consisting of 5 convolutional blocks. Each convolutional block contains a 4×4 convolutional layer with a stride of 2. It is processed by the LeakyReLU activation function with a negative slope of 0.2. Finally, it outputs a 64×64 discrimination result image through a 1×1 convolutional layer. During the discrimination process, the input image is divided into 70×70 image blocks. Each image block is processed by the discriminator to obtain a discrimination value. The average of the discrimination values ​​of all image blocks is taken as the final output, which is used to determine the authenticity of the input image.

[0039] Example 5 This embodiment of the power distribution network fault identification method based on UAV multispectral data, builds upon embodiment 4. The specific method for pre-training the adaptive deep learning denoising model in S3 is as follows: S3.1. Pre-collect and construct a training dataset. The training dataset includes N sets of data pairs, where N is a positive integer greater than 1. Each set of data pairs includes a noisy multispectral image and a corresponding real noise-free multispectral image. S3.2. Divide the training dataset into a training set and a validation set in an 8:2 ratio; S3.3. Use the noisy multispectral images in the training set as input to the generator, compare the generator output with the corresponding real noise-free multispectral images in the training set, and calculate the generation loss. S3.4 Input the image output by the generator and the real noise-free images in the training set into the discriminator. The discriminator calculates the discrimination loss through the binary cross-entropy loss function to distinguish between real and fake images. Its feedback is used to jointly optimize the parameters of the generator and the discriminator. S3.5. An adversarial training mechanism is adopted to jointly optimize the network parameters of the generator and the discriminator by minimizing the generation loss and maximizing the discrimination loss. S3.6. Evaluate the model performance on the validation set until the performance index of the generative adversarial network model reaches the preset convergence threshold, then stop training and obtain the pre-trained adaptive deep learning denoising model.

[0040] Example 6 This embodiment of the power distribution network fault identification method based on UAV multispectral data, builds upon embodiment 5. In step S3.3, the generation loss is calculated by comparing the pixel-level mean square error between the denoised multispectral data of the target generated by the generator and the real noise-free multispectral image. The specific calculation method is as follows: The generator takes a pre-processed noisy multispectral image as input and performs feature extraction and reconstruction through an encoder-decoder structure. The mapping relationship of the generator can be represented as follows: (1); in, The noisy multispectral data after preliminary processing. For the network parameters of the generator, This generates denoised multispectral data for the target. The discriminator then judges the authenticity of the input image patches, and its output is: (2); in, The input image patch can be a real, noise-free image patch or an image patch output by the generator. For the network parameters of the discriminator, This is the feature extraction and nonlinear transformation function of the discriminator. The output value ranges from (0,1), and the closer it is to 1, the more likely it is to be a real noise-free image patch. During model training, the generation loss uses mean squared error loss: (3) in, , , These are the image's height, width, and number of spectral channels, respectively. The generator outputs the image at the 1st line, number Column, No. Pixel values ​​of the channel, The pixel values ​​correspond to real noise-free images, and the units are all DN values ​​(digital quantization values).

[0041] The discriminant loss in S3.4 is calculated using the binary cross-entropy loss function: (4); in, The number of image patches in the training batch. Label 1 represents a real image patch, and 0 represents a generated image patch. For the first Image patches. By alternately optimizing the generator and discriminator, the generator can effectively remove noise while preserving fault features.

[0042] Example 7 This embodiment of the power distribution network fault identification method based on UAV multispectral data, builds upon Embodiment 1. The specific method for identifying the fault type of the power distribution network in S4 is as follows: S4.1 Receiving Target Denoising Multispectral Data: When transmitting target denoising multispectral data to the fault identification and analysis module, fiber optic communication is used, and the transmission rate is not less than 100Mbps. S4.2 Analyze the denoised multispectral data of the target using the target detection model and image classification model to identify the fault types of the power distribution network. The fault types include insulator damage, foreign object hanging on the line, tower corrosion, conductor strand breakage, and abnormal heating at line connection points.

Claims

1. A method for power distribution network fault identification based on UAV multispectral data, characterized in that, Specifically as follows: S1. Acquire raw multispectral data of the power distribution network collected by the multispectral sensor array carried by the UAV; S2. Perform preliminary processing on the raw multispectral data of the distribution network in different bands, such as data synchronization, image registration and format conversion, to obtain the preliminary processed multispectral data; S3. Construct an adaptive deep learning denoising model and pre-train it. Use the pre-trained adaptive deep learning denoising model to denoise and enhance fault features on the pre-processed multispectral data to generate target denoised multispectral data. S4. The target denoised multispectral data is transmitted to the fault identification and analysis module to identify the fault type of the power distribution network.

2. The power distribution network fault identification method based on UAV multispectral data according to claim 1, characterized in that, The multispectral sensor array in S1 includes a visible light camera, a near-infrared camera, or a thermal infrared camera; the original multispectral data of the power distribution network includes image information of power distribution network lines, towers, and equipment in different spectral bands, and the image information contains noise and potential fault characteristics; During the data acquisition process, the multispectral sensor array synchronously captures images at a frequency of 3 frames per second. At the same time, it records the GPS location information (longitude, latitude, altitude) and UAV attitude information (pitch angle, roll angle, and heading angle) corresponding to each frame of image. This information is then linked to the original image data through timestamps to form raw multispectral data with geospatial labels.

3. The power distribution network fault identification method based on UAV multispectral data according to claim 2, characterized in that, The method for preliminary processing of the raw multispectral data in S2 further includes: Radiometric calibration is performed on the raw multispectral data of the power distribution network, and atmospheric correction is performed on the radiometrically calibrated data. In the initial processing stage, data synchronization adopts the timestamp alignment method. Based on the system time of the UAV, image data collected by different sensors are matched according to the standard that the timestamp difference is less than 10 milliseconds to ensure the correspondence of data in different bands at the same time. Image registration adopts an improved SIFT algorithm. First, feature points of each band image are extracted and matched by the FLANN matcher. Then, the random sampling consensus algorithm is used to eliminate mismatches. Finally, bilinear interpolation is used to perform geometric transformation on the image so that the average reprojection error of the registered image is controlled within 1 pixel. The format conversion converts the original RAW format image to 16-bit TIFF format to retain more spectral information.

4. The method for power distribution network fault identification based on UAV multispectral data according to claim 1, characterized in that, The adaptive deep learning denoising model constructed in S3 is a generative adversarial network model, which includes a generator and a discriminator. The generator is used to learn the mapping relationship from noisy pre-processed multispectral data to noise-free target denoised multispectral data, and generate target denoised multispectral data. After the target denoised multispectral data is generated, a data index is established, and the target denoised multispectral data, along with the corresponding original multispectral data, processing time, and geographic location information, are stored in the database constructed by the data storage and management module for historical data tracing and model iteration. The discriminator is used to learn to distinguish between real noise-free multispectral data and target denoised multispectral data generated by the generator, and outputs the discrimination result to guide the training process of the generator.

5. The power distribution network fault identification method based on UAV multispectral data according to claim 4, characterized in that, The generator employs a U-Net architecture, which includes an encoder path and a decoder path. The encoder path contains six convolutional blocks, each consisting of two 3×3 convolutional layers and a 2×2 max-pooling layer. The stride of the convolutional layers is 1, and a non-linear transformation is performed using the ReLU activation function. After each convolutional block, the size of the feature map is halved, while the number of channels is doubled. The decoder path also contains six convolutional blocks, each consisting of two 3×3 transposed convolutional layers. The stride of the transposed convolutions is 2, achieving upsampling of the feature maps. Simultaneously, the feature maps from the corresponding layers of the encoder path are fused through skip connections to preserve detailed information. Finally, a 1×1 convolutional layer converts the feature maps into 3-channel target denoised multispectral data of the same size as the input. The discriminator adopts the PatchGAN architecture, which consists of 5 convolutional blocks. Each convolutional block contains a 4×4 convolutional layer with a stride of 2. It is processed by the LeakyReLU activation function with a negative slope of 0.

2. Finally, it outputs a 64×64 discrimination result image through a 1×1 convolutional layer. During the discrimination process, the input image is divided into 70×70 image blocks. Each image block is processed by the discriminator to obtain a discrimination value. The average of the discrimination values ​​of all image blocks is taken as the final output, which is used to determine the authenticity of the input image.

6. The power distribution network fault identification method based on UAV multispectral data according to claim 5, characterized in that, The specific method for pre-training the adaptive deep learning denoising model described in S3 is as follows: S3.

1. A training dataset is pre-collected and constructed. The training dataset includes N sets of data pairs, where N is a positive integer greater than 1. Each set of data pairs includes a noisy multispectral image and a corresponding real noise-free multispectral image. S3.

2. Divide the training dataset into a training set and a validation set in an 8:2 ratio; S3.

3. Use the noisy multispectral images in the training set as input to the generator, compare the output of the generator with the corresponding real noise-free multispectral images in the training set, and calculate the generation loss. S3.4 Input the image output by the generator and the real noise-free images in the training set into the discriminator. The discriminator calculates the discrimination loss through the binary cross-entropy loss function to distinguish between real and fake images. Its feedback is used to jointly optimize the parameters of the generator and the discriminator. S3.

5. An adversarial training mechanism is adopted to jointly optimize the network parameters of the generator and the discriminator by minimizing the generation loss and maximizing the discrimination loss. S3.

6. Evaluate the model performance on the validation set until the performance index of the generative adversarial network model reaches the preset convergence threshold, then stop training to obtain the pre-trained adaptive deep learning denoising model.

7. The method for power distribution network fault identification based on UAV multispectral data according to claim 6, characterized in that, In S3.3, the generation loss is calculated by comparing the pixel-level mean square error between the denoised multispectral data of the target generated by the generator and the real noise-free multispectral image. The specific calculation method is as follows: The generator takes a pre-processed noisy multispectral image as input and performs feature extraction and reconstruction through an encoder-decoder structure. The mapping relationship of the generator can be represented as follows: （1）； in, The noisy multispectral data after preliminary processing. For the network parameters of the generator, To generate the target denoised multispectral data, the discriminator judges the authenticity of the input image patches, and its output is: （2）； in, The input image patch can be a real, noise-free image patch or an image patch output by the generator. For the network parameters of the discriminator, This is the feature extraction and nonlinear transformation function of the discriminator. The output value ranges from (0,1), and the closer it is to 1, the more likely it is to be a real noise-free image patch. During model training, the generation loss uses mean squared error loss: （3） in, , , These are the image's height, width, and number of spectral channels, respectively. The generator outputs an image at the 1st... line, number Column, No. Pixel values ​​of the channel, The pixel values ​​correspond to real noise-free images, and the units are all DN values ​​(digital quantization values).

8. The method for power distribution network fault identification based on UAV multispectral data according to claim 6, characterized in that, The discriminant loss in S3.4 is calculated using the binary cross-entropy loss function: （4）； in, The number of image patches in the training batch. Label 1 represents a real image patch, and 0 represents a generated image patch. For the first Each image patch is optimized by alternating between the generator and the discriminator, enabling the generator to effectively remove noise while preserving fault features.

9. The method for power distribution network fault identification based on UAV multispectral data according to claim 1, characterized in that, The method for identifying the fault type of the distribution network in S4 is as follows: S4.1 Receiving Target Denoising Multispectral Data: When transmitting target denoising multispectral data to the fault identification and analysis module, fiber optic communication is used, and the transmission rate is not less than 100Mbps. S4.2 Analyze the denoised multispectral data of the target using the target detection model and the image classification model to identify the fault types of the power distribution network; the fault types include insulator damage, foreign object hanging on the line, tower corrosion, conductor strand breakage, and abnormal heating at line connection points.

Images