Intelligent inspection method for transmission lines based on image recognition and deep learning technology

Abstract

This invention relates to an intelligent inspection method for power transmission lines based on image recognition and deep learning technologies, comprising the following steps: S1: Using a conventional inspection drone equipped with a high-definition camera, a preliminary inspection of the power transmission line is conducted based on a preset inspection path to acquire image data of the power transmission line; S2: Preprocessing the acquired image data; S3: Automatically identifying and locating power transmission line equipment using an anomaly detection model to detect anomalies; S4: Based on the anomalies, using a joint analysis method, predicting associated fault points and acquiring possible associated fault points; S5: Constructing a refined inspection path based on the anomalies and associated fault points; S6: Conducting inspection based on the refined inspection path; S7: Using an evaluation model to assess the severity of the faults and generating an inspection report for the power transmission line equipment. This invention achieves automated detection and analysis of power transmission line equipment, improving inspection efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of power transmission line inspection, and in particular to an intelligent inspection method for power transmission lines based on image recognition and deep learning technologies. Background Technology

[0002] With the rapid development of the domestic economy, the demand for electricity is gradually increasing. High-voltage and ultra-high-voltage overhead power lines are the main methods for long-distance power transmission and distribution. Therefore, the safe operation of ultra-high-voltage and high-capacity power lines is a strong power guarantee for my country's economic construction. Reducing power transmission faults and power outages are key issues that need to be addressed. Manual inspection is affected by terrain, weather and personnel factors, resulting in low efficiency, low data accuracy, and difficulty in ensuring personnel safety.

[0003] Currently, drones are used for inspections. In intelligent power line inspections, drones and other intelligent devices first collect data on transmission lines. Then, the collected images and other data are transmitted to a server, where computer vision technology is used for automatic processing and intelligent detection, significantly saving manpower and improving efficiency. However, due to the limitations of vision technology, it is prone to missed detections, and relying solely on image data cannot assess the severity of defects, severely impacting the efficiency and accuracy of inspections. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide an intelligent inspection method for power transmission lines based on image recognition and deep learning technologies, thereby enabling automated detection and analysis of power transmission line equipment and improving inspection efficiency and accuracy.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent inspection of power transmission lines based on image recognition and deep learning technologies includes the following steps:

[0007] S1: Using a conventional inspection drone equipped with a high-definition camera, a preliminary inspection of the power transmission line is carried out based on a preset inspection path to obtain image data of the power transmission line;

[0008] S2: Based on the preprocessing of the acquired image data, including image denoising, enhancement, cropping and annotation;

[0009] S3: An anomaly detection model is used to automatically identify and locate transmission line equipment and detect anomalies;

[0010] S4: Based on outliers, use joint analysis methods to predict associated fault points and obtain possible associated fault points;

[0011] S5: Construct refined inspection paths based on anomalies and associated fault points;

[0012] S6: The inspection drone equipped with a high-definition camera and infrared sensor conducts inspections based on a refined inspection path, and acquires multi-angle image data and infrared data of anomalies and related fault points;

[0013] S7: Based on the anomaly detection model, multi-angle image data of associated fault points are detected to obtain associated anomalies; and based on the multi-angle image data and infrared data of the anomalies and associated anomalies, an evaluation model is used to assess the severity and generate an inspection report for the transmission line equipment.

[0014] Further preprocessing is performed as follows:

[0015] Gaussian filtering is used to remove noise from the image and enhance the image, including contrast enhancement, brightness adjustment, and color saturation enhancement.

[0016] Based on the size and location of the power transmission line equipment, the image is cropped to remove irrelevant background information and focus on the key equipment.

[0017] Annotate the image, marking the locations and bounding boxes of transmission towers, insulators, and conductor equipment. Specifically:

[0018] Define the categories of transmission towers, insulators, and conductor equipment, and assign a unique label to each category.

[0019] The location of each device in the image is labeled using a rectangular bounding box to mark the location and size of the device.

[0020] The labeled data is saved in a standard format for subsequent training of the target detection model.

[0021] Furthermore, the anomaly detection model is constructed as follows:

[0022] The VGG network is used as the base network to extract low-level features of images. An FPN structure is introduced, utilizing multi-scale feature maps to enhance the detection accuracy of small targets. The FPN results include top-down and bottom-up feature fusion, specifically:

[0023] Bottom-up: P l =Conv l (C l );

[0024] Top-down: P l =Upsample(P l+1 )+Conv l (C l );

[0025] Among them, P l C represents the feature map of the l-th layer. l This represents the convolutional features of the l-th layer, Conv() represents the convolution operation, and Upsample() represents the upsampling operation.

[0026] An SE module is added after each feature layer to adjust the importance of the feature maps using learned attention weights, thereby increasing attention to outliers and enhancing the model's representational ability. The calculation process is as follows:

[0027] z = Gap(F);

[0028] s=σ(W2δ(W1z));

[0029] F att =s·F;

[0030] Where Gap represents global average pooling, F is the input feature map, z is the feature after global average pooling, s is the attention weight after activation function, W1 and W2 are the learned weight parameters, and σ and δ are activation functions.

[0031] Use a multi-task loss function, including classification loss, localization loss, bounding box regression loss, and center offset loss:

[0032]

[0033] Where L(x,c,l,g) is the multi-task loss; L conf (x,c) represents the classification loss; L loc (x,l,g) represents the localization loss; L reg (x,l,g) represents the bounding box regression loss; L center (x,l,g) represents the center offset loss; N is the batch size; α, β, γ are the weighting coefficients; x represents the model output, c represents the target category, l represents the target's location information, and g represents the target's ground truth label information.

[0034] Based on labeled historical transmission line image data, a training dataset is constructed, and an anomaly detection model is obtained by training on the training dataset.

[0035] Further, the anomaly detection model training is as follows:

[0036] The parameters are updated by calculating the gradient of the loss function with respect to the model parameters:

[0037] θ=θ-η▽ θ L Z

[0038] Where θ represents the model parameters, and η is the learning rate; ▽θ L Z This represents the gradient of the loss function with respect to the model parameters;

[0039] Optimize model training using the Adam optimizer:

[0040]

[0041] in, and These are the first and second moment estimates of the gradient, respectively, where ε is a constant and t is the current iteration number;

[0042] Pre-defined steps during training reduce the learning rate:

[0043]

[0044] Where η0 is the initial learning rate, γ s It is the attenuation factor, s j It is the decay step size.

[0045] Furthermore, based on outliers, a joint analysis method is used to predict associated fault points and obtain possible associated fault points, as detailed below:

[0046] Joint analysis methods predict user interests or item characteristics based on the similarity between users or items, treating outliers as users and potential failure points as items, and predicting potential associated failure points by analyzing the interactions between them.

[0047] Construct an anomaly-potential failure point correlation matrix R: the elements r in the matrix ui This indicates the correlation strength between anomaly point u and potential fault point i;

[0048] Calculate the similarity between outliers:

[0049]

[0050] Where, r vi This represents the correlation strength between anomaly point v and possible fault point i, where I is the set of all possible fault points;

[0051] The predicted score is calculated based on similarity, and possible related fault points are screened based on the predicted score;

[0052]

[0053] Where U is the set of all users.

[0054] Furthermore, based on anomalies and associated fault points, a refined inspection path is constructed, specifically as follows:

[0055] Multiple optimization objectives are set, including minimizing the total path length and minimizing energy consumption. The following fitness function is used to evaluate the merits of each path, which takes into account the different optimization objectives:

[0056] F(X) = a1 × PL(X) + a2 × EC(X)

[0057] Where a is the weighting coefficient, PL(X) is the total path length function, and EC(X) is the energy consumption function;

[0058] Genetic Algorithm:

[0059] Represent the path as a chromosome;

[0060] Select high-quality chromosomes for reproduction based on the fitness function;

[0061] New chromosomes are generated through crossover operations;

[0062] Chromosome mutation

[0063] Repeat the selection, crossover, and mutation operations until the preset number of iterations or fitness threshold is reached to obtain a preliminary optimal solution;

[0064] Simulated annealing algorithm:

[0065] Starting with the initial optimal solution, new solution spaces are explored by randomly adjusting the path;

[0066] According to the Metropolis criterion, a worse solution is accepted with a probability P, thereby avoiding getting trapped in local optima;

[0067]

[0068] Where ΔE is the fitness difference between the new solution and the current solution, T is the current temperature, and exp is an exponential function;

[0069] By gradually lowering the temperature, the probability of accepting a worse solution is reduced, eventually stabilizing at the optimal solution.

[0070] Furthermore, the construction of the evaluation model is as follows:

[0071] The historical dataset is obtained by acquiring image and infrared data labeled with the severity of the fault and performing preprocessing and feature extraction.

[0072] Using deep learning methods, a multi-layer neural network is constructed, including convolutional layers, pooling layers, fully connected layers, and an output layer. The output layer does not use an activation function and directly outputs continuous values.

[0073] Suppose the neural network has L layers, and the output of each layer l is h. (l) The weight is W (l)The bias is b (l) Then, a multilayer neural network can be represented as:

[0074] Z (l) =W (l) h (l-1) +b (l) ;

[0075] h (l) =activation(Z (l) );

[0076] Wherein, activation represents the activation function;

[0077] The mean squared error loss function is used to measure the difference between the model output and the true label;

[0078] The gradient descent algorithm RMSprop is used to minimize the loss function and update the parameters of the neural network.

[0079] The historical dataset is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to adjust the hyperparameters to prevent overfitting, thus obtaining an evaluation model.

[0080] The present invention has the following beneficial effects:

[0081] 1. This invention enables automated detection and analysis of power transmission line equipment, improving inspection efficiency and accuracy.

[0082] 2. This invention uses the SSD model to automatically identify and locate transmission line equipment and detect abnormal points. By introducing the FPN structure and SE module, the improved SSD model can improve accuracy and focus when detecting abnormal points in transmission line equipment, further enhancing detection performance and accuracy.

[0083] 3. Based on anomalies, this invention utilizes a joint analysis method to effectively mine the association rules between anomalies and related fault points, predict related fault points, obtain possible related fault points, reduce the missed detection rate of inspections, and effectively improve the quality of inspections. Attached Figure Description

[0084] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0085] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0086] refer to Figure 1 This embodiment provides an intelligent inspection method for power transmission lines based on image recognition and deep learning technologies, characterized by the following steps:

[0087] S1: Using a conventional inspection drone equipped with a high-definition camera, a preliminary inspection of the power transmission line is carried out based on a preset inspection path to obtain image data of the power transmission line, including photos and videos of equipment such as transmission towers, insulators, and conductors;

[0088] S2: Based on the preprocessing of the acquired image data, including image denoising, enhancement, cropping and annotation, to improve the accuracy of subsequent image recognition algorithms;

[0089] S3: An anomaly detection model is used to automatically identify and locate transmission line equipment and detect anomalies, including abnormal conditions such as transmission tower tilting, insulator damage, and conductor breakage.

[0090] S4: Based on outliers, use joint analysis methods to predict associated fault points and obtain possible associated fault points;

[0091] S5: Construct refined inspection paths based on anomalies and associated fault points;

[0092] S6: The inspection drone equipped with a high-definition camera and infrared sensor conducts inspections based on a refined inspection path, and acquires multi-angle image data and infrared data of anomalies and related fault points;

[0093] S7: Based on the anomaly detection model, multi-angle image data of associated fault points are detected to obtain associated anomalies; and based on the multi-angle image data and infrared data of the anomalies and associated anomalies, an assessment model is used to evaluate the severity, generating an inspection report for the transmission line equipment. This report includes information such as the location, type, and severity of the anomalies.

[0094] In this embodiment, the preprocessing is as follows:

[0095] Gaussian filtering is used to remove noise from the image and enhance the image, including contrast enhancement, brightness adjustment, and color saturation enhancement.

[0096] Based on the size and location of the power transmission line equipment, the image is cropped to remove irrelevant background information and focus on the key equipment.

[0097] Annotate the image, marking the locations and bounding boxes of transmission towers, insulators, and conductor equipment. Specifically:

[0098] Define the categories of transmission towers, insulators, and conductors, and assign a unique label to each category (e.g., 0 for transmission towers, 1 for insulators, and 2 for conductors).

[0099] The location of each device in the image is labeled using a rectangular bounding box to mark the location and size of the device.

[0100] The labeled data is saved in a standard format (such as PASCAL VOC, COCO) for subsequent training of the object detection model.

[0101] In this embodiment, the anomaly detection model is constructed as follows:

[0102] The VGG network is used as the base network to extract low-level features of images. An FPN structure is introduced, utilizing multi-scale feature maps to enhance the detection accuracy of small targets. The FPN results include top-down and bottom-up feature fusion, specifically:

[0103] Bottom-up: P l =Conv l (C l )

[0104] Top-down: P l =Upsample(P l+1 )+Conv l (C l );

[0105] Among them, P l C represents the feature map of the l-th layer. l This represents the convolutional features of the l-th layer, Conv() represents the convolution operation, and Upsample() represents the upsampling operation.

[0106] An SE module is added after each feature layer to adjust the importance of the feature maps using learned attention weights, thereby increasing attention to outliers and enhancing the model's representational ability. The calculation process is as follows:

[0107] z = Gap(F);

[0108] s=σ(W2δ(W1z));

[0109] F att =s·F;

[0110] Where Gap represents global average pooling, F is the input feature map, z is the feature after global average pooling, s is the attention weight after activation function, W1 and W2 are the learned weight parameters, and σ and δ are activation functions.

[0111] Use a multi-task loss function, including classification loss, localization loss, bounding box regression loss, and center offset loss:

[0112]

[0113] Where L(x,c,l,g) is the multi-task loss; L conf (x,c) represents the classification loss; L loc(x,l,g) represents the localization loss; L reg (x,l,g) represents the bounding box regression loss; L center (x,l,g) represents the center offset loss; (the classification loss uses cross-entropy loss to calculate the difference between the predicted class and the true class; the localization loss uses SmoothL1 loss to calculate the difference between the predicted bounding box and the true bounding box; the bounding box regression loss and the center offset loss use L1 loss); N is the batch size; α, β, γ are weight coefficients; x represents the model output, c represents the target class, l represents the target's location information, and g represents the target's true annotation information;

[0114] Based on labeled historical transmission line image data, a training dataset is constructed, and an anomaly detection model is obtained by training on the training dataset.

[0115] In this embodiment, the anomaly detection model is trained as follows:

[0116] The parameters are updated by calculating the gradient of the loss function with respect to the model parameters:

[0117] θ=θ-η▽ θ L Z

[0118] Where θ represents the model parameters, and η is the learning rate; ▽ θ L Z This represents the gradient of the loss function with respect to the model parameters;

[0119] Optimize model training using the Adam optimizer:

[0120]

[0121] in, and These are the first and second moment estimates of the gradient, respectively, where ε is a constant and t is the current iteration number;

[0122] Pre-defined steps during training reduce the learning rate:

[0123]

[0124] Where η0 is the initial learning rate, γ s It is the attenuation factor, s j It is the decay step size.

[0125] In this embodiment, based on outliers, a joint analysis method is used to predict associated fault points and obtain possible associated fault points, as detailed below:

[0126] Joint analysis methods predict user interests or item characteristics based on the similarity between users or items, treating outliers as users and potential failure points as items, and predicting potential associated failure points by analyzing the interactions between them.

[0127] Construct an anomaly-potential failure point correlation matrix R: the elements r in the matrix ui This indicates the correlation strength between anomaly point u and potential fault point i;

[0128] Calculate the similarity between outliers:

[0129]

[0130] Where, r vi This represents the correlation strength between anomaly point v and possible fault point i, where I is the set of all possible fault points;

[0131] The predicted score is calculated based on similarity, and possible related fault points are screened based on the predicted score;

[0132]

[0133] Where U is the set of all users.

[0134] In this embodiment, a refined inspection path is constructed based on anomalies and associated fault points, specifically as follows:

[0135] Multiple optimization objectives are set, including minimizing the total path length and minimizing energy consumption. The following fitness function is used to evaluate the merits of each path, which takes into account the different optimization objectives:

[0136] F(X) = a1 × PL(X) + a2 × EC(X)

[0137] Where a is the weighting coefficient, PL(X) is the total path length function, and EC(X) is the energy consumption function;

[0138] Genetic Algorithm:

[0139] Represent the path as a chromosome (e.g., a sequence containing the order in which nodes are visited);

[0140] Select high-quality chromosomes for reproduction based on the fitness function;

[0141] New chromosomes are generated through crossover operations; such as partial matching crossover (PMX).

[0142] Chromosome mutations can be performed, such as swapping the positions of two nodes, to increase population diversity.

[0143] Repeat the selection, crossover, and mutation operations until the preset number of iterations or fitness threshold is reached to obtain a preliminary optimal solution;

[0144] Simulated annealing algorithm:

[0145] Starting with an initial optimal solution, new solution spaces are explored by randomly adjusting the path (e.g., swapping the positions of two nodes);

[0146] According to the Metropolis criterion, a worse solution is accepted with a probability P, thereby avoiding getting trapped in local optima;

[0147]

[0148] Where ΔE is the fitness difference between the new solution and the current solution, T is the current temperature, and exp is an exponential function;

[0149] By gradually lowering the temperature, the probability of accepting a worse solution is reduced, eventually stabilizing at the optimal solution.

[0150] In this embodiment, the construction of the evaluation model is specifically as follows:

[0151] The historical dataset is obtained by acquiring image and infrared data labeled with the severity of the fault and performing preprocessing and feature extraction.

[0152] Using deep learning methods, a multi-layer neural network is constructed, including convolutional layers, pooling layers, fully connected layers, and an output layer. The output layer does not use an activation function and directly outputs continuous values.

[0153] Suppose the neural network has L layers, and the output of each layer l is h. (l) The weight is W (l) The bias is b (l) Then, a multilayer neural network can be represented as:

[0154] Z (l) =W (l) h (l-1) +b (l) ;

[0155] h (l) =activation(Z (l) );

[0156] Wherein, activation represents the activation function;

[0157] The mean squared error loss function is used to measure the difference between the model output and the true label;

[0158] The gradient descent algorithm RMSprop is used to minimize the loss function and update the parameters of the neural network.

[0159] The historical dataset is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to adjust the hyperparameters to prevent overfitting, thus obtaining an evaluation model.

[0160] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0162] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent inspection of transmission lines based on image recognition and deep learning technologies, characterized in that, Includes the following steps: S1: Using a conventional inspection drone equipped with a high-definition camera, a preliminary inspection of the power transmission line is carried out based on a preset inspection path to obtain image data of the power transmission line; S2: Based on the preprocessing of the acquired image data, including image denoising, enhancement, cropping and annotation; S3: An anomaly detection model is used to automatically identify and locate transmission line equipment and detect anomalies; S4: Based on outliers, use joint analysis methods to predict associated fault points and obtain potential associated fault points; S5: Construct refined inspection paths based on anomalies and associated fault points; S6: The inspection drone equipped with a high-definition camera and infrared sensor conducts inspections based on a refined inspection path, and acquires multi-angle image data and infrared data of anomalies and related fault points; S7: Based on the anomaly detection model, multi-angle image data of associated fault points are detected to obtain associated anomalies; and based on the anomalies and associated anomalies, multi-angle image data and infrared data are used to evaluate the severity of the anomalies and generate an inspection report of the transmission line equipment. The method of predicting associated fault points and obtaining potential associated fault points based on outliers using joint analysis is as follows: Joint analysis methods predict user interests or item characteristics based on the similarity between users or items, treating outliers as users and potential failure points as items, and predicting potential associated failure points by analyzing the interactions between them. Construct an anomaly-potential failure point correlation matrix R: the elements r in the matrix ui This indicates the correlation strength between anomaly point u and potential fault point i; Calculate the similarity between outliers: Where, r vi This represents the correlation strength between anomaly point v and potential fault point i, where I is the set of all potential fault points; A predicted score is calculated based on similarity, and potential related fault points are screened based on the predicted score. Where U is the set of all users.

2. The intelligent inspection method for transmission lines based on image recognition and deep learning technology according to claim 1, characterized in that, The preprocessing is as follows: Gaussian filtering is used to remove noise from the image and enhance the image, including contrast enhancement, brightness adjustment, and color saturation enhancement. Based on the size and location of the power transmission line equipment, the image is cropped to remove irrelevant background information and focus on the key equipment. Annotate the image, marking the locations and bounding boxes of transmission towers, insulators, and conductor equipment. Specifically: Define the categories of transmission towers, insulators, and conductor equipment, and assign a unique label to each category; The location of each device in the image is labeled using a rectangular bounding box to mark the device's position and size; The labeled data is saved in a standard format for subsequent training of the target detection model.

3. The intelligent inspection method for transmission lines based on image recognition and deep learning technology according to claim 1, characterized in that, The anomaly detection model is constructed as follows: The VGG network is used as the base network to extract low-level features of images. An FPN structure is introduced to enhance the detection accuracy of small targets by utilizing multi-scale feature maps. The FPN structure includes top-down and bottom-up feature fusion, specifically: Bottom-up: P l =Conv l (C l ); Top-down: P l =Upsample(P l+1 )+Conv l (C l ); Among them, P l C represents the feature map of the l-th layer. l This represents the convolutional features of the l-th layer, Conv() represents the convolution operation, and Upsample() represents the upsampling operation. An SE module is added after each feature layer to adjust the importance of the feature maps using learned attention weights, thereby increasing attention to outliers and enhancing the model's representational ability. The calculation process is as follows: z = Gap(F); s=σ(W2δ(W1z)); F att =s·F; Where Gap represents global average pooling, F is the input feature map, z is the feature after global average pooling, s is the attention weight after activation function, W1 and W2 are the learned weight parameters, and σ and δ are activation functions. Use a multi-task loss function, including classification loss, localization loss, bounding box regression loss, and center offset loss: Among them, L conf (x,c) represents the classification loss; L loc (x,l,g) represents the localization loss; L reg (x,l,g) represents the bounding box regression loss; L center (x,l,g) represents the center offset loss; N is the batch size; α,β,γ are the weight coefficients; x represents the model output, c represents the target category, l represents the target's location information, and g represents the target's ground truth label information. Based on labeled historical transmission line image data, a training dataset is constructed, and an anomaly detection model is obtained by training on the training dataset.

4. The intelligent inspection method for transmission lines based on image recognition and deep learning technology according to claim 3, characterized in that, The training of the anomaly detection model is as follows: The parameters are updated by calculating the gradient of the loss function with respect to the model parameters: Where θ represents the model parameters and η is the learning rate; This represents the gradient of the loss function with respect to the model parameters; Optimize model training using the Adam optimizer: in, and These are the first and second moment estimates of the gradient, respectively, where ε is a constant and t is the current iteration number; Pre-defined steps during training reduce the learning rate: Where η0 is the initial learning rate, γ s It is the attenuation factor, s j It is the decay step size.

5. The intelligent inspection method for transmission lines based on image recognition and deep learning technology according to claim 1, characterized in that, The detailed inspection path is constructed based on anomalies and associated fault points, specifically as follows: Multiple optimization objectives are set, including minimizing the total path length and minimizing energy consumption. The following fitness function is used to evaluate the merits of each path, which takes into account the different optimization objectives: F(X) = a1 × PL(X) + a2 × EC(X) Where a is the weighting coefficient, PL(X) is the total path length function, and EC(X) is the energy consumption function; Genetic Algorithm: Represent the path as a chromosome; Select high-quality chromosomes for reproduction based on the fitness function; New chromosomes are generated through crossover operations; To perform mutation operations on chromosomes; Repeat the selection, crossover, and mutation operations until the preset number of iterations or fitness threshold is reached to obtain a preliminary optimal solution; Simulated annealing algorithm: Starting with the initial optimal solution, new solution spaces are explored by randomly adjusting the path; According to the Metropolis criterion, a worse solution is accepted with a probability P, thereby avoiding getting trapped in local optima; Where ΔE is the fitness difference between the new solution and the current solution, T is the current temperature, and exp is an exponential function; By gradually lowering the temperature, the probability of accepting a worse solution is reduced, eventually stabilizing at the optimal solution.

6. The intelligent inspection method for transmission lines based on image recognition and deep learning technology according to claim 1, characterized in that, The construction of the evaluation model is as follows: Acquire image and infrared data labeled with the severity of the fault, and perform preprocessing and feature extraction to obtain a historical dataset; Using deep learning methods, a multi-layer neural network is constructed, including convolutional layers, pooling layers, fully connected layers, and an output layer. The output layer does not use an activation function and directly outputs continuous values. Suppose the neural network has L layers, and the output of each layer l is h. (l) The weight is W (l) The bias is b (l) Then, a multilayer neural network can be represented as: Z (l) =W (l) h (l-1) +b (l) ; h (l) =activation(Z (l) ); Wherein, activation represents the activation function; The mean squared error loss function is used to measure the difference between the model output and the true label; The gradient descent algorithm RMSprop is used to minimize the loss function and update the parameters of the neural network. The historical dataset is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to adjust the hyperparameters to prevent overfitting, thus obtaining an evaluation model.

Images