Power transmission unmanned aerial vehicle self-adaptive inspection method and system based on AI auxiliary photographing

Abstract

The invention discloses a power transmission unmanned aerial vehicle self-adaptive inspection method and system based on AI auxiliary photographing, and relates to the field of power transmission inspection, and the method comprises the following steps: obtaining historical inspection photos and annotation information, and collecting original position information and actual photographing point information corresponding to the historical inspection photos; training an AI image recognition model by using the historical inspection photos and the annotation information, and training a position deviation correction model by using the historical inspection photos, the original position information and the actual shooting point position information; correcting a shooting point location of the unmanned aerial vehicle by using a position deviation correction model; adjusting the angle and position of the unmanned aerial vehicle during shooting; defects in the image are automatically detected and marked; the system comprises an information acquisition module, a model training module, a position correction module, a shooting adjustment module and a defect detection module. Through unmanned aerial vehicle position correction and multi-angle shooting, the unmanned aerial vehicle inspection image quality is improved.

Description

Technical Field

[0001] The present invention relates to the field of power transmission inspection, and in particular to a power transmission drone adaptive inspection method and system based on AI-assisted photography. Background Art

[0002] Transmission drones are able to conduct remote visual inspections of transmission lines, towers, insulators, and other power facilities at high altitudes. These drones can fly over difficult-to-reach or dangerous terrain, providing real-time images and data to help power companies detect potential faults or defects in a timely manner, thereby ensuring the continuity and safety of power supply. When performing inspection tasks, drones can automatically adjust their flight paths and parameters according to different environments and conditions. This adaptive capability enables drones to maintain efficient and accurate inspections in complex and changing environments.

[0003] In terms of refined inspection of power transmission based on routes, the following problems exist: the inspection process has a high rejection rate of inspection photos. During autonomous inspections according to routes, due to factors such as the deviation of waypoint shooting points caused by weak positioning signals, the inspection photos taken have problems such as the target equipment cannot be placed in the center of the picture, and defective parts cannot be photographed from multiple angles. This leads to a high rejection rate of inspection photos and the inspection quality cannot meet the requirements.

[0004] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0005] In response to the problems in the related technology, the present invention proposes an adaptive inspection method and system for power transmission drones based on AI-assisted photography to overcome the above-mentioned technical problems existing in the existing related technology.

[0006] To this end, the specific technical solution adopted by the present invention is as follows:

[0007] According to one aspect of the present invention, a power transmission drone adaptive inspection method based on AI-assisted photography is provided, and the power transmission drone adaptive inspection method based on AI-assisted photography includes the following steps:

[0008] S1. Obtain historical inspection photos and annotation information, and collect the original location information and actual shooting point information corresponding to the historical inspection photos.

[0009] S2. Use historical inspection photos and annotation information to train the AI ​​image recognition model, and use historical inspection photos, original location information and actual shooting point information to train the position deviation correction model.

[0010] S3. Based on the scheduled route inspection, the position deviation correction model is used to correct the shooting points of the drone.

[0011] S4. Based on the AI ​​image recognition model, adjust the angle and position of the drone during shooting to ensure that the target device is in the center of the picture and shoot from multiple angles.

[0012] S5. Obtain inspection images and location information of target equipment taken from multiple angles, and use the trained AI image recognition model to automatically detect defects in the images and mark them.

[0013] Furthermore, obtaining the actual shooting point information includes the following steps:

[0014] Pre-process historical inspection photos and use edge detection algorithms to detect the contours of target equipment in historical inspection photos;

[0015] Use feature extraction algorithm to extract key points and feature descriptors of target devices;

[0016] Match the feature descriptors with the features in the database, filter out feature pairs through the matching results, and calculate a preliminary estimate of the target device location;

[0017] According to the matching results and original position information, the actual shooting point is calculated based on the least squares method.

[0018] Furthermore, using historical inspection photos and annotation information to train the AI ​​image recognition model includes the following steps:

[0019] Build deep learning models;

[0020] Divide historical inspection photos and annotation information into training sets, validation sets, and test sets according to a pre-set ratio;

[0021] Use the training set data to train the deep learning model, use the performance of the validation set to adjust the hyperparameters, and use the test set to evaluate the performance of the deep learning model.

[0022] Furthermore, building a deep learning model includes the following steps:

[0023] Select a basic model architecture for image recognition and design several convolutional layers to extract image features, with each convolutional layer followed by an activation function and pooling layers added between convolutional layers;

[0024] Generate target candidate regions using target detection algorithms and adjust candidate regions of different sizes to a uniform size;

[0025] Construct several fully connected layers, followed by activation functions, to classify candidate regions;

[0026] Construct a network structure for bounding box regression to predict the position and size of the bounding box, and define the classification loss function and regression loss function.

[0027] Furthermore, using historical inspection photos, original location information, and actual shooting point information to train a position deviation correction model includes the following steps:

[0028] Obtain image features from historical inspection photos;

[0029] Extracting position deviation information from original position information and actual shooting point information;

[0030] Match the image features with the position deviation information to form training samples, and each sample contains the image features and the corresponding position deviation label;

[0031] Initialize the support vector machine model and set the penalty factor and kernel function parameters;

[0032] Search for the optimal combination of penalty factor and kernel function parameters based on the optimization algorithm;

[0033] The support vector machine model is set using the optimized parameter combination, and is trained using training samples, and the support vector machine model is cross-validated.

[0034] Furthermore, searching for the optimal penalty factor and kernel function parameter combination based on the optimization algorithm includes the following steps:

[0035] Set the population size, number of iterations and step size of the optimization algorithm;

[0036] Define the fitness function, determine the search range of the penalty factor and kernel function parameters, and create the initial population;

[0037] For each individual, use the current penalty factor and kernel function parameters to train and verify the support vector machine model, and calculate the fitness function value;

[0038] Update the position of the individual according to the fitness value, so that the individual moves to a position where the error is less than the preset threshold;

[0039] Enable individuals to conduct global and local searches;

[0040] Update the position where the error is less than the preset threshold according to the fitness value;

[0041] Repeat the iteration until the stopping condition is met, and select the penalty factor and kernel function parameters corresponding to the individual with the best fitness value as the optimal solution.

[0042] Furthermore, the formula of the fitness function is:

[0043]

[0044] In the formula, F represents the fitness value;

[0045] N represents the total number of samples;

[0046] Y i Represents the actual position deviation of the i-th sample;

[0047] Represents the predicted position deviation of the i-th sample.

[0048] Furthermore, based on the scheduled route inspection, the position deviation correction model is used to correct the shooting position of the drone, including the following steps:

[0049] S31. Use geographic information system to plan routes and set the navigation path of drones to ensure that all inspection areas are covered;

[0050] S32, using enhanced global positioning system to improve positioning accuracy;

[0051] S33. Use the position deviation correction model to correct the shooting points of the drone in real time.

[0052] Furthermore, based on the AI ​​image recognition model, adjusting the angle and position of the drone during shooting, ensuring that the target device is in the center of the picture and shooting from multiple angles includes the following steps:

[0053] S41. Use AI image recognition model to identify target devices in the transmission line in real time, and use target tracking algorithm to track the location and movement trajectory of the target devices;

[0054] S42. Automatically adjust the shooting angle and position of the drone according to the position and movement trajectory of the target device.

[0055] According to another aspect of the present invention, an adaptive inspection system for power transmission drones based on AI-assisted photography is provided. The adaptive inspection system for power transmission drones based on AI-assisted photography includes an information acquisition module, a model training module, a position correction module, a shooting adjustment module and a defect detection module.

[0056] Among them, the information acquisition module is used to obtain historical inspection photos and annotation information, and collect the original location information and actual shooting point information corresponding to the historical inspection photos.

[0057] The model training module is used to train the AI ​​image recognition model using historical inspection photos and annotation information, and to train the position deviation correction model using historical inspection photos, original location information and actual shooting point information.

[0058] The position correction module is used to make the drone patrol according to the predetermined route and use the position deviation correction model to correct the shooting point of the drone.

[0059] The shooting adjustment module is used to enable the drone to automatically adjust the angle and position when shooting based on the AI ​​image recognition model, ensuring that the target device is in the center of the picture and performing multi-angle shooting.

[0060] The defect detection module is used to obtain inspection images and location information of target equipment taken from multiple angles, and use the trained AI image recognition model to automatically detect defects in the image and mark them.

[0061] The beneficial effects of the present invention are:

[0062] (1) The power transmission drone adaptive inspection method and system based on AI-assisted photography provided by the present invention improves the drone inspection image quality through drone position correction and multi-angle shooting. Through the trained position deviation correction model, the drone can predict and correct the shooting point deviation caused by weak positioning signals. Even in an environment with poor signals, the drone can accurately aim at the target device for shooting, reducing waste film caused by positioning errors. Using the AI ​​image recognition model, the drone can automatically adjust the angle and position during shooting to ensure that the target device is always in the center of the picture. This automatic focus and alignment function greatly improves the success rate of shooting and reduces waste film caused by the device not being in the center of the picture.

[0063] (2) AI-assisted drones can shoot from multiple angles to ensure that defective parts are fully documented. This approach not only increases the probability of discovering defects, but also reduces the possibility of missing key information due to single-angle shooting. Through the trained AI image recognition model, drones can automatically detect defects in images and annotate them. This automated processing greatly improves inspection efficiency, reduces the need for manual inspections, and also reduces human errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0065] Figure 1 is a flow chart of a power transmission drone adaptive inspection method based on AI-assisted photography according to an embodiment of the present invention;

[0066] Figure 2 It is a principle block diagram of a power transmission drone adaptive inspection system based on AI-assisted photography according to an embodiment of the present invention.

[0067] In the figure:

[0068] 1. Information acquisition module; 2. Model training module; 3. Position correction module; 4. Shooting adjustment module; 5. Defect detection module. DETAILED DESCRIPTION

[0069] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0070] According to an embodiment of the present invention, a method and system for adaptive inspection of power transmission drones based on AI-assisted photography are provided.

[0071] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to one embodiment of the present invention, a power transmission drone adaptive inspection method based on AI-assisted photography is provided, and the power transmission drone adaptive inspection method based on AI-assisted photography includes the following steps:

[0072] S1. Obtain historical inspection photos and annotation information (category and location of equipment and defects), and collect the original location information (GPS data) and actual shooting point information (corrected location information) corresponding to the historical inspection photos.

[0073] In a further embodiment, obtaining the actual shooting point information includes the following steps:

[0074] Preprocess historical inspection photos and use edge detection algorithms to detect the contours of target devices in historical inspection photos.

[0075] Feature extraction algorithms (such as SIFT, SURF, ORB) are used to extract key points and feature descriptors of the target device.

[0076] The feature descriptors are matched with the features in the database, and reliable feature pairs are screened out through the matching results, and a preliminary estimate of the target device location is calculated.

[0077] The process of matching the feature descriptor with the features in the database and selecting reliable feature pairs through the matching results is as follows:

[0078] For each historical inspection photo, a feature extraction algorithm (such as SIFT, SURF, ORB) is used to extract the feature descriptor set {d i}, where d irepresents the i-th feature descriptor in the feature descriptor set, and the matching feature set in the database is {D j}, where D j Represents the jth matching feature descriptor in the matching feature set in the database.

[0079] All feature descriptors are preliminarily classified by the locality sensitive hashing (LSH) method. The LSH function h(·) maps the feature descriptors to a low-dimensional space: h(d i )=hash(d i ), h(D j )=hash(D j ), construct feature descriptors with the same hash value into feature descriptor pairs (d i , D j ).

[0080] For feature descriptor pairs with the same hash value, calculate their cosine similarity:

[0081]

[0082] In the formula, ts(d i , D j ) is the cosine similarity of the feature descriptor pair.

[0083] Select the feature pair with the largest cosine similarity as the preliminary matching result. For the preliminary matching results, further calculate their angular distance:

[0084]

[0085] In the formula, θ(d i , D j ) is the angular distance.

[0086] The feature pair with the smallest angular distance is selected as the final matching result.

[0087] For each final matched feature pair, the matching reliability metric R(d i , D j ):

[0088]

[0089] Where α is the weight factor of the cosine similarity stored in the database, and β is the weight factor of the angular distance stored in the database.

[0090] For the weight factors α and β stored in the database, cross-validation is performed through grid search: first prepare a training data set containing labeled feature descriptor pairs, define the search range of α and β (such as [0, 1]), then use k-fold cross-validation to split the data, train the model in turn and evaluate the performance of each combination of α and β, and finally select the weight coefficient combination that makes the model perform best on the validation set.

[0091] Using the LSH method for preliminary classification helps to quickly screen out possible matching feature descriptor pairs and reduce the amount of calculation. Combining cosine similarity and angular distance can more comprehensively measure the matching degree of feature descriptors and improve matching accuracy. By calculating the reliability measure of the matching, reliable feature pairs can be effectively screened out, further improving the credibility of the matching results. Through a multi-step matching process, the most suitable feature pairs can be gradually screened out, reducing the probability of mismatching.

[0092] According to the matching results and original position information, the actual shooting point is calculated based on the least squares method.

[0093] In addition, visualization tools (such as maps and 3D models) are used to show the changes in location information before and after correction. The least squares method is used to solve a function to minimize the sum of squared deviations of the data points. Specifically, it is used to correct the original GPS location information and estimate a more accurate shooting point based on the image feature matching results.

[0094] S2. Use historical inspection photos and annotation information to train the AI ​​image recognition model, and use historical inspection photos, original location information and actual shooting point information to train the position deviation correction model. The AI ​​image recognition model identifies target equipment and defects such as poles, insulators, and conductors. The position deviation correction model can predict and correct shooting point deviations caused by weak positioning signals.

[0095] In a further embodiment, training an AI image recognition model using historical inspection photos and annotation information includes the following steps:

[0096] Build a deep learning model; divide historical inspection photos and annotation information into training set, validation set and test set according to pre-set ratios.

[0097] Use the training set data to train the deep learning model, use the performance of the validation set to adjust hyperparameters such as learning rate and batch size, and use the test set to evaluate the performance of the deep learning model and make necessary adjustments.

[0098] In a further embodiment, building a deep learning model comprises the following steps:

[0099] Select a basic model architecture for image recognition, such as ResNet, Inception, or VGG, and design several convolutional layers to extract image features. Each convolutional layer is followed by an activation function. At the same time, pooling layers are added between convolutional layers to reduce the spatial dimension of the features and increase the invariance to image displacement.

[0100] Use the target detection algorithm (such as RPN, SSD or YOLO) to generate target candidate regions and adjust the candidate regions of different sizes to a uniform size.

[0101] Construct several fully connected layers, followed by activation functions, to classify the candidate regions.

[0102] Construct a network structure for bounding box regression, which usually includes several fully connected layers to predict the position and size of the bounding box, and define the classification loss function and regression loss function.

[0103] In a further embodiment, training the position deviation correction model using historical inspection photos, original position information, and actual shooting point information includes the following steps:

[0104] Get image features from historical inspection photos.

[0105] Extract position deviation information from the original position information and the actual shooting point information.

[0106] The image features are matched with the position deviation information to form training samples, and each sample contains the image features and the corresponding position deviation label.

[0107] Initialize the support vector machine model and set the penalty factor and kernel function parameters.

[0108] Search for the optimal combination of penalty factor and kernel function parameters based on the optimization algorithm.

[0109] The support vector machine model is set using the optimized parameter combination and trained using training samples so that the model can learn the relationship between image features and position deviations, and the support vector machine model is cross-validated.

[0110] Specifically, a cross-validation method was used to evaluate the stability and accuracy of the model. The model parameters were adjusted until satisfactory performance was obtained on an independent validation set.

[0111] In a further embodiment, searching for an optimal combination of penalty factor and kernel function parameters based on an optimization algorithm comprises the following steps:

[0112] Set the optimization algorithm parameters such as population size, number of iterations, and step size.

[0113] Define the fitness function, determine the search range of the penalty factor and kernel function parameters, and create the initial population.

[0114] For each individual, the current penalty factor and kernel function parameters are used to train and validate the support vector machine model, and the fitness function value is calculated.

[0115] The position of the individual is updated according to the fitness value, so that the individual moves to a position where the error is less than a preset threshold or away from a position where the error is greater than a preset threshold.

[0116] Enable individuals to perform global search and local search.

[0117] The position where the error is less than the preset threshold or the position where the error is greater than the preset threshold is updated according to the fitness value.

[0118] Repeat the iteration until the stopping condition is met, such as reaching the maximum number of iterations or the fitness is no longer significantly improved, and select the penalty factor and kernel function parameters corresponding to the individual with the best fitness value as the optimal solution.

[0119] In a further embodiment, the formula of the fitness function is:

[0120]

[0121] In the formula, F represents the fitness value; N represents the total number of samples; Y i Represents the actual position deviation of the i-th sample; Represents the predicted position deviation of the i-th sample.

[0122] S3. Based on the scheduled route inspection, the position deviation correction model is used to correct the shooting points of the drone.

[0123] In a further embodiment, based on the scheduled route inspection, correcting the shooting position of the drone using the position deviation correction model includes the following steps:

[0124] S31. Use geographic information system (GIS) to plan routes and set the navigation path of the drone to ensure coverage of all inspection areas.

[0125] S32. Use enhanced global positioning systems (such as RTK-GPS) to improve positioning accuracy.

[0126] S33. Using the position deviation correction model, the shooting position of the drone is corrected in real time, and the shooting position is accurately aligned with the target device.

[0127] Among them, the GIS is a system used to capture, store, analyze and manage geospatial data. It combines computer hardware, software and spatial data to solve complex planning and management problems. The Enhanced Global Positioning System is a system that improves the positioning accuracy of the Global Positioning System through external information sources (such as ground reference stations and communication networks). The positioning accuracy is improved by using differential technology.

[0128] S4. Based on the AI ​​image recognition model, adjust the angle and position of the drone during shooting to ensure that the target device is in the center of the picture and shoot from multiple angles.

[0129] In a further embodiment, based on the AI ​​image recognition model, adjusting the angle and position of the drone during shooting to ensure that the target device is in the center of the picture and performing multi-angle shooting includes the following steps:

[0130] S41. Use AI image recognition models to identify target equipment in transmission lines, such as towers and insulators, in real time, and use target tracking algorithms to track the location and movement trajectory of target equipment.

[0131] S42. Automatically adjust the shooting angle and position of the drone according to the position and movement trajectory of the target device, ensure that the target device is located in the center of the picture, and perform multi-angle shooting.

[0132] S5. Obtain inspection images and location information of target equipment taken from multiple angles, and use the trained AI image recognition model to automatically detect defects in the image and mark them. The specific steps include:

[0133] Use the trained AI image recognition model to analyze the features in the inspection images of the target equipment to detect defects in the images.

[0134] If a defect is detected, the location of the defect is determined and marked on the image, usually including the defect type, size, location and other relevant information.

[0135] The process of determining the defect type is as follows:

[0136] The defect feature vector set extracted from the image is F = [f 1 , f 2 , ..., f a , ..., f m ], f a is the ath defect feature vector in the defect feature vector set, a=1,2,3,...,m, m is the total number of defect feature vectors, and each defect feature vector has a different weight w a .

[0137] Get the feature vector set of the b-th defect stored in the database as Gb =[g b1 , g b2 , ..., g bc , ..., g bN ],g bc is the cth defect feature vector in the feature vector set of the bth defect stored in the database, c=1,2,3,...N, N is the total number of defect feature vectors, where c corresponds to a, and m=N.

[0138]

[0139] In the formula, σ a is the scale parameter for normalizing the distance, which can be determined using the standard deviation of each feature component in the training data set, and e is a natural constant.

[0140] The defect type with the highest similarity is taken as the final judgment result:

[0141]

[0142] Where haty is the defect type with the highest similarity score.

[0143] like Figure 2 As shown, according to another embodiment of the present invention, an adaptive inspection system for power transmission drones based on AI-assisted photography is also provided. The adaptive inspection system for power transmission drones based on AI-assisted photography includes an information acquisition module 1, a model training module 2, a position correction module 3, a shooting adjustment module 4 and a defect detection module 5.

[0144] Among them, the information acquisition module 1 is used to obtain historical inspection photos and annotation information, and collect the original location information and actual shooting point information corresponding to the historical inspection photos.

[0145] Model training module 2 is used to train the AI ​​image recognition model using historical inspection photos and annotation information, and to train the position deviation correction model using historical inspection photos, original location information and actual shooting point information.

[0146] The position correction module 3 is used to make the UAV patrol according to the predetermined route and use the position deviation correction model to correct the shooting point of the UAV.

[0147] The shooting adjustment module 4 is used to enable the drone to automatically adjust the angle and position when shooting based on the AI ​​image recognition model, ensure that the target device is in the center of the picture and perform multi-angle shooting.

[0148] The defect detection module 5 is used to obtain the inspection images and location information of the target equipment obtained by multi-angle shooting, and use the trained AI image recognition model to automatically detect defects in the image and mark them.

[0149] In summary, the adaptive inspection method and system of power transmission drone based on AI-assisted photography provided by the present invention improves the quality of drone inspection images through drone position correction and multi-angle shooting. Through the trained position deviation correction model, the drone can predict and correct the shooting point deviation caused by weak positioning signal. Even in an environment with poor signal, the drone can accurately aim at the target device for shooting, reducing the waste film caused by positioning error. Using the AI ​​image recognition model, the drone can automatically adjust the angle and position during shooting to ensure that the target device is always in the center of the picture. This automatic focus and alignment function greatly improves the success rate of shooting and reduces the waste film caused by the device not being in the center of the picture. AI-assisted drones can perform multi-angle shooting to ensure a full range of defective parts. This method not only increases the probability of discovering defects, but also reduces the possibility of missing key information due to single-angle shooting. Through the trained AI image recognition model, the drone can automatically detect defects in the image and mark them. This automated processing greatly improves the inspection efficiency, reduces the need for manual inspection, and also reduces human errors.

[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The power transmission drone adaptive inspection method based on AI-assisted photography is characterized by: The power transmission drone adaptive inspection method based on AI-assisted photography includes the following steps: S1. Obtain historical inspection photos and annotation information, and collect the original location information and actual shooting point information corresponding to the historical inspection photos; S2. Use historical inspection photos and annotation information to train the AI ​​image recognition model, and use historical inspection photos, original location information, and actual shooting point information to train the position deviation correction model; S3, based on the scheduled route inspection, and using the position deviation correction model to correct the shooting point of the drone; S4. Based on the AI ​​image recognition model, adjust the angle and position of the drone during shooting to ensure that the target device is in the center of the picture and shoot from multiple angles; S5. Obtain inspection images and location information of target equipment taken from multiple angles, and use the trained AI image recognition model to automatically detect defects in the images and mark them.

2. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 1 is characterized in that: The acquisition of actual shooting point information includes the following steps: Pre-process historical inspection photos and use edge detection algorithms to detect the contours of target equipment in historical inspection photos; Use feature extraction algorithm to extract key points and feature descriptors of target devices; Match the feature descriptors with the features in the database, filter out feature pairs based on the matching results, and calculate a preliminary estimate of the target device location; According to the matching results and original position information, the actual shooting point is calculated based on the least squares method.

3. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 2 is characterized in that: The method of training the AI ​​image recognition model using historical inspection photos and annotation information includes the following steps: Build deep learning models; Divide historical inspection photos and annotation information into training sets, validation sets, and test sets according to a pre-set ratio; Use the training set data to train the deep learning model, use the performance of the validation set to adjust the hyperparameters, and use the test set to evaluate the performance of the deep learning model.

4. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 3 is characterized in that: The construction of the deep learning model includes the following steps: Select a basic model architecture for image recognition and design several convolutional layers to extract image features, with each convolutional layer followed by an activation function and pooling layers added between convolutional layers; Generate target candidate regions using target detection algorithms and adjust candidate regions of different sizes to a uniform size; Construct several fully connected layers, followed by activation functions, to classify candidate regions; Construct a network structure for bounding box regression to predict the position and size of the bounding box, and define the classification loss function and regression loss function.

5. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 4 is characterized in that: The training of the position deviation correction model using historical inspection photos, original location information, and actual shooting point information includes the following steps: Obtain image features from historical inspection photos; Extracting position deviation information from original position information and actual shooting point information; Match the image features with the position deviation information to form training samples, and each sample contains the image features and the corresponding position deviation label; Initialize the support vector machine model and set the penalty factor and kernel function parameters; Search for the optimal combination of penalty factor and kernel function parameters based on the optimization algorithm; The support vector machine model is set using the optimized parameter combination, and is trained using training samples, and the support vector machine model is cross-validated.

6. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 5 is characterized in that: The method of searching for the optimal penalty factor and kernel function parameter combination based on the optimization algorithm comprises the following steps: Set the population size, number of iterations and step size of the optimization algorithm; Define the fitness function, determine the search range of the penalty factor and kernel function parameters, and create the initial population; For each individual, use the current penalty factor and kernel function parameters to train and verify the support vector machine model, and calculate the fitness function value; Update the position of the individual according to the fitness value, so that the individual moves to a position where the error is less than the preset threshold; Enable individuals to conduct global and local searches; Update the position where the error is less than the preset threshold according to the fitness value; Repeat the iteration until the stopping condition is met, and select the penalty factor and kernel function parameters corresponding to the individual with the best fitness value as the optimal solution.

7. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 6 is characterized in that: The formula of the fitness function is: In the formula, F represents the fitness value; N represents the total number of samples; Y i Represents the actual position deviation of the i-th sample; Represents the predicted position deviation of the i-th sample.

8. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 7 is characterized in that: The method of inspecting the scheduled route and correcting the shooting position of the drone using the position deviation correction model includes the following steps: S31. Use geographic information system to plan routes and set the navigation path of drones to ensure that all inspection areas are covered; S32, using enhanced global positioning system to improve positioning accuracy; S33. Use the position deviation correction model to correct the shooting points of the drone in real time.

9. The power transmission drone adaptive inspection method based on AI-assisted photography according to claim 8 is characterized in that: The method of adjusting the angle and position of the drone during shooting based on the AI ​​image recognition model to ensure that the target device is in the center of the picture and to perform multi-angle shooting includes the following steps: S41. Use AI image recognition model to identify target devices in the transmission line in real time, and use target tracking algorithm to track the location and movement trajectory of the target devices; S42. Automatically adjust the shooting angle and position of the drone according to the position and movement trajectory of the target device.

10. A power transmission drone adaptive inspection system based on AI-assisted photography, used to implement the power transmission drone adaptive inspection method based on AI-assisted photography as described in any one of claims 1 to 9, characterized in that: The power transmission drone adaptive inspection system based on AI-assisted photography includes an information acquisition module, a model training module, a position correction module, a shooting adjustment module and a defect detection module; The information acquisition module is used to obtain historical inspection photos and annotation information, and collect the original location information and actual shooting point information corresponding to the historical inspection photos; The model training module is used to train the AI ​​image recognition model using historical inspection photos and annotation information, and to train the position deviation correction model using historical inspection photos, original position information, and actual shooting point information; The position correction module is used to make the drone inspect according to the predetermined route and correct the shooting point of the drone using the position deviation correction model; The shooting adjustment module is used to automatically adjust the angle and position of the drone during shooting based on the AI ​​image recognition model to ensure that the target device is in the center of the picture and to perform multi-angle shooting; The defect detection module is used to obtain inspection images and location information of the target equipment obtained by multi-angle shooting, and use the trained AI image recognition model to automatically detect defects in the image and mark them.

Images