Defect diagnosis method for overhead insulated cable

Abstract

The invention discloses an overhead insulated cable defect diagnosis method, and relates to the technical field of intelligent operation and maintenance and state monitoring of overhead power lines. Comprising the following steps: controlling a first unmanned aerial vehicle and a second unmanned aerial vehicle to cooperatively fly and accurately position, and scanning the overhead insulated cable by using an X-ray emitter and a detector to obtain an original X-ray image; deblurring the original X-ray image through an image deblurring module to obtain a clear image; inputting the obtained clear image into an improved YOLOX-Tiny target detection module to obtain a bounding box of the cable, a state classification label of the cable and a direction label of the cable, and realizing defect diagnosis of the cable; based on the direction label, generating a course guide instruction flying along the cable extension direction through the first unmanned aerial vehicle and the second unmanned aerial vehicle to control the unmanned aerial vehicles to fly along the cable extension direction; and continuously scanning and diagnosing the whole section of overhead insulated cable. According to the invention, internal defect diagnosis of the overhead insulated cable can be realized through an automatic means.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance and condition monitoring technology for overhead power lines, and in particular to a method for diagnosing defects in overhead insulated cables. Background Technology

[0002] Overhead insulated cables, as a key carrier of power distribution networks, are widely used in power supply lines in cities, rural areas, and industrial parks. They are exposed to harsh outdoor environments for extended periods, facing not only natural aging from sunlight, rain, and wind vibration, but also direct impacts from lightning strikes, overloads, and external collisions. The cable's insulation and sheath are barriers ensuring its electrical safety and mechanical strength. If the internal conductors or connections break, melt, or deteriorate due to lightning strikes, electrochemical corrosion, or mechanical fatigue, it will directly lead to serious faults such as discharge, short circuits, and even fires and line breaks, causing widespread power outages and posing a significant threat to power supply reliability. Therefore, regular, efficient, and accurate internal condition inspection and defect diagnosis of operating overhead insulated cables have crucial engineering value and socio-economic significance.

[0003] The use of drones equipped with visible light cameras and infrared thermal imagers for automated power line inspection has become common practice. Visible light images from drones, combined with machine learning algorithms, can automatically identify obvious damage or foreign objects on the cable surface; while infrared thermal imagers can detect abnormal temperature rises in the line and locate defects such as overheating at connection points. However, these mainstream technologies all have a fundamental limitation: they can only sense the surface of the line or indirectly infer its external state based on the temperature field. They cannot penetrate the cable's insulation sheath to directly detect physical defects in the internal conductors or early degradation of the insulation layer. These internal defects are highly concealed but are the core causes of failures. For overhead insulated cables, how to utilize available inspection data and innovate models and methods to indirectly and accurately diagnose their potential internal defects remains a challenge and a gap in the current technological field.

[0004] In summary, existing technologies still fall short of meeting the demand for highly reliable diagnosis of hidden internal defects in cables with complex backgrounds, small targets, and weak features due to insufficient accuracy, high false alarm rate, and weak generalization ability. Summary of the Invention

[0005] Therefore, it is necessary to provide a method for diagnosing defects in overhead insulated cables to address the aforementioned technical problems.

[0006] The following technical solution is adopted in this specification: This specification provides a method for diagnosing defects in overhead insulated cables, including: Obtain the raw X-ray image of the target overhead insulated cable; An improved YOLOX-Tiny object detection model is constructed using pre-training. The improved YOLOX-Tiny object detection model includes a backbone network, a feature pyramid network, and a decoupling head connected in sequence, as well as a directional classification branch network parallel to the decoupling head. Based on the original X-ray image, downsampling and feature extraction are performed through the backbone network to obtain multi-scale features. The multi-scale features are then fused through the feature pyramid network to obtain three enhanced feature maps at different scales. The decoupling head is used to convolve each enhanced feature map at each scale to obtain the bounding box information and state classification label of the target overhead insulated cable. The directional features of the enhanced feature maps are extracted through the directional classification branch network, and feature fusion and classification are performed based on the directional features to obtain the directional label of the cable. Based on the bounding box coordinates, state classification labels, and direction labels of the cable, the defects of the target overhead insulated cable are diagnosed.

[0007] Furthermore, the acquisition of the original X-ray image specifically includes: Based on a real-time dynamic positioning system, the first UAV equipped with an X-ray emitter and the second UAV equipped with an X-ray detector are coordinated in flight and positioning control, so that the two UAVs remain synchronized. By aligning the X-ray emitter, the inspected portion of the target overhead insulated cable, and the X-ray detector in a straight line, an X-ray scan of the inspected portion of the target overhead insulated cable is performed to obtain an original X-ray image.

[0008] Furthermore, the original X-ray image is deblurred using a generative adversarial network model to obtain a clear image suitable for the improved YOLOX-Tiny target detection model.

[0009] Furthermore, the generative adversarial network model consists of a generator network G and a discriminator network D; The generator network G adopts a U-Net structure, consisting of an encoder and a decoder; the encoder extracts features through multi-layer convolution and downsampling, and the decoder fuses features and reconstructs the image through multi-layer upsampling and convolution. The discriminator network D adopts the PatchGAN structure and is used to distinguish the authenticity of the input clear image.

[0010] Furthermore, the training process of the generative adversarial network model specifically includes: Construct an image dataset with "blurred-sharp" pairings; Based on the image dataset, the generative adversarial network model is trained by configuring a composite loss function to obtain a trained adversarial network model. The composite loss function includes the adversarial loss L of the generator network G. adv Pixel-level L1 loss L1 and perceived loss L perc ; The resistance loss L adv The Wasserstein distance with gradient penalty is expressed as follows: ; Among them, I sharp For a clear reference image; I blurry The input is a blurred image; λ represents a random interpolation sample between the sharp image and the generated image. gp E[⋅] represents the preset weight coefficient of the gradient penalty term; E[⋅] represents the expected value of the variable in parentheses; G(⋅) represents the forward propagation operation of the generator network G; D(⋅) represents the forward propagation operation of the discriminator network D; Represents the function Request regarding input The gradient; ||⋅||2 represents the L2 norm of the vector inside the brackets, i.e., the Euclidean length; The pixel-level L1 loss L L1 The expression is: ; Where ||⋅||1 represents the sum of the absolute values ​​of all elements in the image within the parentheses, i.e., the L1 norm; The perceived loss L perc The expression is: ; in, ϕ (⋅) represents a feature extraction function of a pre-trained deep convolutional neural network, whose input is an image and whose output is a feature map extracted from a specific network layer; This means that after taking the square root of the sum of the squares of all elements of the difference between the two feature maps within the parentheses to obtain the L2 norm, the norm value is then squared. The composite loss function is composed of the adversarial loss L of the generator network G. adv Pixel-level L1 loss L1 and perceived loss L perc The expression, derived through weighted summation, is: ; Where, λ L1 and λ perc These are the preset weighting coefficients.

[0011] Furthermore, obtaining the direction label of the cable specifically includes: Based on the feature maps of three different scales output by the feature pyramid network, a 1×1 convolutional layer is used to perform dimensionality reduction convolution on the feature maps of each scale, reducing the number of channels to a uniform 256. Based on the reduced feature map, two 3×3 convolutional layers are passed sequentially to enhance the representation ability of the features; The enhanced features are mapped to a low-level representation for orientation classification through a 1×1 convolutional layer, resulting in feature maps of three scales for orientation classification. The feature maps of the three scales used for orientation classification are fused, and a result containing scores of 8 predefined orientation categories is obtained through classification. The results are converted into a probability distribution using the Softmax function, and the category with the highest probability is used as the predicted cable direction label.

[0012] Furthermore, the total loss function of the improved YOLOX-Tiny object detection model during training specifically includes: The original loss L of the main network yolox and the loss L of the directional classification branch direction The total loss function L is constructed by weighted summation. total The expression is: ; Where γ is the preset weighting coefficient for balancing the two losses; The directional classification loss L direction The cross-entropy loss function is used for calculation.

[0013] Furthermore, the status classification labels include: health labels, fracture labels, and melting labels.

[0014] Furthermore, the directional label is used to generate guidance instructions for the UAV's flight heading when the UAV scans the target overhead insulated cable.

[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: In the overhead insulated cable defect diagnosis method provided in this specification, an internal image of the cable is acquired by a collaborative UAV equipped with a low-dose X-ray device, and a modified YOLOX-Tiny model is used to realize defect identification and inspection path guidance. The modified YOLOX-Tiny model, by simultaneously outputting defect type and cable direction labels in a single inference, seamlessly combines intelligent defect identification with UAV autonomous inspection path guidance, realizing continuous operation of "diagnosis and inspection at the same time". The internal image of the cable acquired by the collaborative X-ray imaging of dual UAVs can directly and non-destructively see through the internal state of the high-altitude live overhead insulated cable, solving the fundamental problem that traditional surface inspection methods cannot detect hidden defects. Furthermore, by using a generative adversarial network model to deblur the original X-ray images, the problem of dynamic imaging blurring is effectively overcome. Combined with a closed-loop process from imaging and analysis to guidance, the efficiency, accuracy, and automation level of long-distance cable inspection are significantly improved, and safety risks are reduced. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 This document provides a flowchart illustrating a method for diagnosing defects in overhead insulated cables. Figure 2 This is a schematic diagram of the network structure of the generative adversarial network model for image deblurring provided in this specification. Figure 3 This is a schematic diagram of the network structure of the improved YOLOX-Tiny target detection model for defect diagnosis provided in this specification. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0019] X-ray imaging technology, with its non-destructive imaging capabilities, is maturely applied in industrial inspection. However, its application to the inspection of high-altitude, energized, and continuously laid overhead insulated cables still faces significant engineering challenges. Traditional fixed or handheld X-ray inspection methods are extremely inefficient and difficult to implement without power interruption or tower climbing. Existing technologies lack a complete solution that can deeply integrate X-ray imaging capabilities with the high-altitude mobility and precise positioning capabilities of UAVs, enabling real-time intelligent defect diagnosis and automatic inspection path guidance. Therefore, this invention solves the technical challenge of efficient, uninterrupted, and automated intelligent diagnosis of the internal condition of overhead insulated cables through an improved YOLOX-Tiny model-based defect diagnosis method.

[0020] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating a defect diagnosis method for overhead insulated cables as described in this specification, which specifically includes the following steps: S101: Control the first and second UAVs to fly together and accurately locate, and use an X-ray emitter and X-ray detector to scan the overhead insulated cable to obtain raw X-ray images.

[0022] The first UAV is equipped with an X-ray emitter, and the second UAV is equipped with an X-ray detector. The two UAVs are synchronized based on a real-time dynamic positioning system, so that the emitter, the cable under inspection and the detector are in the same straight line to complete the X-ray scanning imaging.

[0023] S102: The original X-ray image is deblurred using the image deblurring module to obtain a clear image.

[0024] The specific steps for deblurring the original X-ray image are as follows: 9. Construct a generative adversarial network model; like Figure 2 As shown, the image deblurring module consists of a generator network G and a discriminator network D. The generator network G adopts a U-Net structure, where the encoder extracts features through multi-layer convolution and downsampling, and the decoder fuses features and reconstructs the image through multi-layer upsampling and convolution. The discriminator network D adopts a PatchGAN structure and is used to distinguish between real and fake input images. This generative adversarial network needs to be trained adversarially using a pre-constructed "blurred-clear" paired image dataset. The trained generator network G serves as the core of the image deblurring module.

[0025] 10. Configure a composite loss function to train the generative adversarial network model; The total loss function L of the generator network GG From the counter-loss L adv Pixel-level L1 loss L1 and perceived loss L perc The weighted summation is expressed as follows: ; Where, λ L1 and λ perc These are preset weighting coefficients; The resistance loss L adv The Wasserstein distance with gradient penalty is used, and its expression is: ; Among them, I sharp For a clear reference image, I blurry For the input blurred image, λ represents a random interpolation sample between the sharp image and the generated image. gp Here, E[⋅] represents the preset weight coefficients for the gradient penalty term, E[⋅] represents the expected value of the variables within the parentheses, G(⋅) represents the forward propagation operation of the generator network G, and D(⋅) represents the forward propagation operation of the discriminator network D. Represents the function Request regarding input The gradient of ||⋅||2 represents the L2 norm of the vector within the brackets, i.e., the Euclidean length.

[0026] The pixel-level L1 loss L L1 The expression is: ; Where ||⋅||1 represents the sum of the absolute values ​​of all elements in the image within the parentheses, i.e., the L1 norm; The perceived loss L perc The expression is: ; in, ϕ (⋅) represents a feature extraction function of a pre-trained deep convolutional neural network, whose input is an image and whose output is a feature map extracted from a specific network layer; This means that after taking the square root of the sum of the squares of all elements of the difference between the two feature maps within the parentheses to obtain the L2 norm, the norm value is then squared.

[0027] 11. Perform image deblurring; The raw X-ray image obtained in step S101 is input into the trained generator network G, and the generator network G directly outputs the processed clear image.

[0028] S103: Input the obtained clear image into the improved YOLOX-Tiny target detection module, which will output the cable's bounding box, cable status classification label, and cable orientation label, thereby realizing the defect diagnosis of the cable.

[0029] The specific steps of the overhead insulated cable defect diagnosis method based on the improved YOLOX-Tiny target detection module are as follows: 1. Construct the overall architecture of the improved YOLOX-Tiny object detection module: such as... Figure 3 As shown, the module is an improvement on the classic YOLOX-Tiny model. The classic YOLOX-Tiny model serves as the main network, which includes a backbone network for feature extraction, a feature pyramid network for multi-scale feature fusion, and a decoupling head for the main network to predict bounding boxes and categories. In this improvement, an independent directional classification branch network is added in parallel with the decoupling head of the main network at the output of the feature pyramid network. 2. Design of the main network: The backbone network of the main network adopts the Darknet53-tiny structure to downsample and extract features from the input clear image; the feature pyramid network receives the multi-scale features output by the backbone network, performs top-down and bottom-up feature fusion, and outputs three enhanced feature maps at different scales; the decoupling head performs convolution processing on the feature map at each scale, and outputs the bounding box coordinates, confidence score and state classification probability of the target in parallel. 3. Design the orientation classification branch network: The orientation classification branch network takes the feature maps at three different scales output by the feature pyramid network as input, and performs the following processing independently on the feature map at each scale to extract orientation features, then fuses and classifies them. The specific steps are as follows: a) Apply a 1×1 convolutional layer to the feature map at each scale to reduce the dimensionality, thereby reducing the number of channels to a uniform 256; b) The reduced features are processed by two 3×3 convolutional layers in sequence to enhance their representational power. c) Apply another 1×1 convolutional layer to map the processed features into a low-level representation for orientation classification; d) The features obtained after processing the above three scales are fused and then passed through a final classification output layer to produce a result containing scores for 8 predefined directional categories; e) The result is converted into a probability distribution using the Softmax function, where the category with the highest probability is the predicted cable direction label.

[0030] 4. Configure the total loss function of the improved YOLOX-Tiny object detection module: the total loss function L during training of the module. total The original loss L of the main network yolox The loss L of the directional classification branch direction The weighted summation is expressed as follows: ; Where γ is a preset weighting coefficient balancing the two losses, and the directional classification loss L... direction The cross-entropy loss function is used for calculation.

[0031] 5. Perform target detection and output: Input the clear image obtained in step S102 into the improved YOLOX-Tiny target detection module that has been trained; the main network of the module outputs the bounding box coordinates of the cable target and its status classification label in parallel, the status classification label including "healthy", "broken" and "melted"; at the same time, the direction classification branch network of the module outputs the extension direction label of the cable, the direction label is used to calculate and generate the guidance command for the UAV flight heading in step S104.

[0032] S104: Based on the cable direction tag output in step S3, the first UAV and the second UAV automatically calculate and generate a heading guidance command for flying along the cable extension direction, and control the UAV to fly along the cable extension direction accordingly. S105: Repeat steps S1 to S4 until the continuous scanning and diagnosis of the entire overhead insulated cable is completed.

[0033] Following the above design steps, an example of defect diagnosis for overhead insulated cables based on the improved YOLOX-Tiny model is given, and the implementation effect of the method of the present invention is explained in detail and analyzed.

[0034] The specific design and training of the image deblurring module based on generative adversarial networks used in step S102 of this invention are as follows: A generative adversarial network (GAN) model is constructed, consisting of a generator network G and a discriminator network D. It requires adversarial training using a pre-built "blurred-sharp" paired X-ray image dataset. The generator network G adopts a U-Net architecture; its encoder part consists of four downsampling blocks, progressively downsampling the input 1×1024×1024 image to a 1×256×256 feature map. The decoder part symmetrically consists of four upsampling blocks, each upsampling through transposed convolution and concatenating with the feature map of the corresponding layer of the encoder. Finally, a 1×1024×1024 deblurred image is output through a convolutional layer with a tanh activation function. The discriminator network D adopts a PatchGAN architecture, consisting of five stacked 4×4 convolutional layers with a stride of 2. In its final output feature map, each pixel corresponds to the discrimination result of a local region of the input image. After training, the generator network G serves as the core of the image deblurring module.

[0035] The generative adversarial network model is trained using a composite loss function: the total loss function L of the generator network G. G From the counter-loss L adv Pixel-level L1 loss L1 and perception-based feature loss L perc It consists of three weighted parts, namely L G =L adv +λ L1 L L1 +λ perc L perc Among them, the countermeasure loss L adv The Wasserstein distance form with gradient penalty is used to improve training stability; pixel loss L L1 Constraining pixel-level consistency; perceptual loss L perc Features are extracted using a pre-trained VGG network to ensure the reconstructed image is semantically realistic. In this embodiment, the weight coefficients are set to λ. L1 =100, λ perc =0.1, gradient penalty coefficient λ gp =10; The training process uses a specially constructed X-ray image dataset containing 1968 precisely labeled images. The optimizer is Adam, and the learning rate for both the generator and discriminator is set to 1×10. -4 The batch size is 4, and the number of iterations is set to 300.

[0036] The specific design and training of the improved YOLOX-Tiny target detection module used in step S103 of this invention are as follows: An improved YOLOX-Tiny target detection network model is constructed: the main network is based on the YOLOX-Tiny framework, the backbone network adopts Darknet53-tiny, the feature pyramid network is used for multi-scale feature fusion, and the decoupling head is responsible for outputting bounding boxes and state classification; the orientation classification branch network takes the three-scale feature maps output by the feature pyramid network as input, and performs the following processing on the feature map of each scale independently to extract orientation features, which are then fused and classified. The specific steps are as follows: a) Apply a 1×1 convolutional layer to the feature map of each scale to reduce the dimensionality, and uniformly reduce the number of channels to 256; b) Process the dimensionality-reduced features with two 3×3 convolutional layers in sequence to enhance the representation ability of the features; c) Apply another 1×1 convolutional layer to map the processed features to the low-level representation for orientation classification; d) Fuse the features obtained from the above three scales and generate a result containing 8 predefined orientation category scores through a final classification output layer; e) Convert the result into a probability distribution through the Softmax function, where the category with the highest probability is the predicted cable orientation label.

[0037] The improved YOLOX-Tiny object detection network model is trained using a configured total loss function: L is the total loss function of the improved YOLOX-Tiny object detection module. total The main detection network loss L yolox With directional classification branch loss L direction The weighted sum, i.e. L total =L yolox +γL direction In this embodiment, the weight coefficient γ used to balance the two losses is set to 0.5. The training process uses the clear image dataset output by the aforementioned image deblurring module as training samples. This dataset is divided into training, validation, and test sets according to a specified ratio. During training, the input image size is uniformly adjusted to 320×320 pixels, the optimizer is Adam, and the initial learning rate is set to 1×10⁻⁶. -3 The number of iterations is set to 300.

[0038] based on Figure 1 The beneficial effects of the method for diagnosing defects in overhead insulated cables shown are as follows: 1. Breakthrough in the bottleneck of non-destructive testing of internal defects: Through the collaborative X-ray imaging of dual UAVs, direct and non-destructive imaging of the internal condition of high-altitude live overhead insulated cables is achieved, solving the fundamental problem that traditional surface inspection methods cannot detect hidden defects.

[0039] 2. Integration of diagnosis and inspection guidance: Through the improved YOLOX-Tiny model, the defect type and cable direction label are output simultaneously in a single inference, so that intelligent defect identification and UAV autonomous inspection heading guidance are seamlessly combined, realizing continuous operation of "diagnosing and inspecting at the same time".

[0040] 3. Improved automation and reliability of the entire process: The use of generative adversarial networks effectively overcomes the problem of dynamic imaging blurring. Combined with a closed-loop process from imaging and analysis to guidance, it significantly improves the efficiency, accuracy and automation level of long-distance cable inspection and reduces safety risks.

[0041] Furthermore, the model trained according to the specific design parameters described above was comprehensively evaluated on a standard test set independently partitioned from the total dataset. This standard test set contains 151 images, and the validation results are as follows: The defect diagnosis model achieved excellent defect diagnosis performance on the test set. It can diagnose defects in healthy, fractured, and molten states, with average precision of 99.19%, 95.48%, and 99.98%, respectively, and an overall average precision of 98.21%. The recall rates were 98.03%, 93.98%, and 98.74%, respectively, with an overall recall rate of 96.91%. These quantitative indicators fully demonstrate the model's high accuracy in identifying typical defects inside cables.

[0042] The directional classification branch of the defect diagnosis model performed stably on the test set, with an average recognition accuracy of over 99% for the cable extension direction. This result meets the technical requirements for providing reliable and continuous heading guidance for UAVs and is key to achieving a closed loop of autonomous inspection.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for diagnosing defects in overhead insulated cables, characterized in that, include: Obtain the raw X-ray image of the target overhead insulated cable; An improved YOLOX-Tiny object detection model is constructed using pre-training. The improved YOLOX-Tiny object detection model includes a backbone network, a feature pyramid network, and a decoupling head connected in sequence, as well as a directional classification branch network parallel to the decoupling head. Based on the original X-ray image, downsampling and feature extraction are performed through the backbone network to obtain multi-scale features; the multi-scale features are fused through the feature pyramid network to obtain three enhanced feature maps at different scales; the enhanced feature maps at each scale are convolved through the decoupling head to obtain the bounding box information and state classification label of the target overhead insulated cable. The directional features of the enhanced feature map are extracted through the directional classification branch network, and the features are fused and classified according to the directional features to obtain the directional label of the cable. Based on the bounding box coordinates, state classification labels, and direction labels of the cable, the defects of the target overhead insulated cable are diagnosed.

2. The method for diagnosing defects in overhead insulated cables as described in claim 1, characterized in that, The acquisition of the original X-ray image specifically includes: Based on a real-time dynamic positioning system, the first UAV equipped with an X-ray emitter and the second UAV equipped with an X-ray detector are coordinated in flight and positioning control, so that the two UAVs remain synchronized. By aligning the X-ray emitter, the inspected portion of the target overhead insulated cable, and the X-ray detector in a straight line, an X-ray scan of the inspected portion of the target overhead insulated cable is performed to obtain an original X-ray image.

3. The method for diagnosing defects in overhead insulated cables as described in claim 1, characterized in that, The original X-ray image is deblurred using a generative adversarial network model to obtain a clear image suitable for the improved YOLOX-Tiny target detection model.

4. The method for diagnosing defects in overhead insulated cables as described in claim 3, characterized in that, The generative adversarial network model consists of a generator network G and a discriminator network D. The generator network G adopts a U-Net structure, consisting of an encoder and a decoder; the encoder extracts features through multi-layer convolution and downsampling, and the decoder fuses features and reconstructs the image through multi-layer upsampling and convolution. The discriminator network D adopts the PatchGAN structure and is used to distinguish the authenticity of the input clear image.

5. The method for diagnosing defects in overhead insulated cables as described in claim 3, characterized in that, The training process of the generative adversarial network model specifically includes: Construct an image dataset with "blurred-sharp" pairings; Based on the image dataset, the generative adversarial network model is trained by configuring a composite loss function to obtain a trained adversarial network model. The composite loss function includes the adversarial loss L of the generator network G. adv Pixel-level L1 loss L1 and perceived loss L perc ; The resistance loss L adv The Wasserstein distance with gradient penalty is expressed as follows: ； Among them, I sharp For a clear reference image; I blurry The input is a blurred image; λ represents a random interpolation sample between the sharp image and the generated image. gp E[⋅] represents the preset weight coefficient of the gradient penalty term; E[⋅] represents the expected value of the variable in parentheses; G(⋅) represents the forward propagation operation of the generator network G; D(⋅) represents the forward propagation operation of the discriminator network D; Represents the function Request regarding input The gradient; ||⋅||2 represents the L2 norm of the vector inside the brackets, i.e., the Euclidean length; The pixel-level L1 loss L L1 The expression is: ； Where ||⋅||1 represents the sum of the absolute values ​​of all elements in the image within the parentheses, i.e., the L1 norm; The perceived loss L perc The expression is: ； in, ϕ (⋅) represents a feature extraction function of a pre-trained deep convolutional neural network, whose input is an image and whose output is a feature map extracted from a specific network layer; This means that after taking the square root of the sum of the squares of all elements of the difference between the two feature maps within the parentheses to obtain the L2 norm, the norm value is then squared. The composite loss function is composed of the adversarial loss L of the generator network G. adv Pixel-level L1 loss L1 and perceived loss L perc The expression, derived through weighted summation, is: ； Where, λ L1 and λ perc These are the preset weighting coefficients.

6. The method for diagnosing defects in overhead insulated cables as described in claim 1, characterized in that, Obtaining the direction label of the cable specifically includes: Based on the feature maps of three different scales output by the feature pyramid network, a 1×1 convolutional layer is used to perform dimensionality reduction convolution on the feature maps of each scale, reducing the number of channels to a uniform 256. Based on the reduced feature map, two 3×3 convolutional layers are passed sequentially to enhance the representation ability of the features; The enhanced features are mapped to a low-level representation for orientation classification through a 1×1 convolutional layer, resulting in feature maps of three scales for orientation classification. The feature maps of the three scales used for orientation classification are fused, and a result containing scores of 8 predefined orientation categories is obtained through classification. The results are converted into a probability distribution using the Softmax function, and the category with the highest probability is used as the predicted cable direction label.

7. The method for diagnosing defects in overhead insulated cables as described in claim 1, characterized in that, The total loss function of the improved YOLOX-Tiny object detection model during training includes: The original loss L of the main network yolox and the loss L of the directional classification branch direction The total loss function L is constructed by weighted summation. total The expression is: ； Where γ is the preset weighting coefficient for balancing the two losses; The directional classification loss L direction The cross-entropy loss function is used for calculation.

8. The method for diagnosing defects in overhead insulated cables as described in claim 1, characterized in that, The status classification labels include: health label, fracture label, and melt label.

9. The method for diagnosing defects in overhead insulated cables as described in claim 1, characterized in that, The directional label is used to generate guidance instructions for the UAV's flight heading when the UAV scans the target overhead insulated cable.

Images