Electric tower pole inspection method and device based on label identification

Abstract

The invention discloses an electric tower pole inspection method and device based on label identification, and the method comprises the steps: capturing an inspection image through an unmanned plane when the unmanned plane flies around an electric tower pole, carrying out the feature extraction and label identification of the inspection image through a label identification model, taking the image as a target inspection image when the inspection image is identified to have a radio frequency label, and carrying out the recognition of the radio frequency label in the target inspection image; and after the actual position of the radio frequency tag is obtained according to the current position of the unmanned aerial vehicle, the fixed position of the electric tower pole and the relative position information of the unmanned aerial vehicle relative to the electric tower pole, the unmanned aerial vehicle is controlled to fly to the actual position of the radio frequency tag, and the operation information in the radio frequency tag is automatically read to obtain the inspection data of the electric tower pole. According to the method, the unmanned aerial vehicle is adopted for routing inspection, and the routing inspection efficiency is improved through automatic tag identification and radio frequency tag reading, so that the operation information of the electric tower pole can be accurately obtained even under the conditions of severe weather or complex terrain, and the stability and the safety of a power system are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of inspection of electric tower poles, and particularly to an inspection method and device for electric tower poles based on tag recognition. Background Art

[0002] As an important part of the power system, the operating status of electric tower poles is directly related to the stability and safety of the power system. Inspecting electric tower poles can help understand the operating conditions of the electric tower poles and their equipment, including the environmental conditions along the line, providing a basis for equipment maintenance; it can also promptly detect and eliminate equipment defects, prevent accidents, and ensure the safe operation of power lines.

[0003] However, traditional inspection technologies mainly rely on manual inspection for electric tower poles. That is, inspection personnel need to go to the site in person, and often need to use high-altitude operations to inspect the radio frequency tags on the upper part of the tower poles. Therefore, traditional inspection technologies require inspection personnel to spend a lot of time and energy on site for inspection, resulting in low inspection efficiency. Moreover, during high-altitude operations in manual inspection, under harsh weather or complex terrain conditions, the observation angle of inspection personnel may be restricted, and the actual operating information of the electric tower poles cannot be accurately obtained, which may lead to the inability to promptly detect and eliminate equipment defects of the electric tower poles, resulting in low stability and safety of the power system. Summary of the Invention

[0004] The embodiments of the present invention provide an inspection method and device for electric tower poles based on tag recognition. Through automated tag recognition and operation data collection, the operating status of electric tower poles can be monitored in real time and accurately, effectively solving the problem in the prior art that the actual operating information of electric tower poles cannot be accurately obtained, resulting in the inability to promptly detect and eliminate equipment defects of electric tower poles.

[0005] An embodiment of the present invention provides an inspection method for electric tower poles based on tag recognition, including:

[0006] Repeatedly perform inspection image recognition operations until it is determined that there is a radio frequency tag in the inspection image, and use the inspection image with the radio frequency tag as the target inspection image; wherein, the radio frequency tag is used to display the operating information on the electric tower pole;

[0007] Generate the tag position of the radio frequency tag in the target inspection image according to the current position of the unmanned aerial vehicle, the fixed position of the electric tower pole, and the relative position of the unmanned aerial vehicle relative to the electric tower pole;

[0008] After obtaining the actual position of the radio frequency tag according to the tag position, control the unmanned aerial vehicle to fly to the actual position of the radio frequency tag, then control the unmanned aerial vehicle to read the operating information of the radio frequency tag, and generate the inspection data of the electric tower pole according to the read operating information;

[0009] Among them, the inspection image recognition operation includes:

[0010] Obtain the inspection images captured by the drone when flying around the electric tower pole;

[0011] Input the inspection images into a preset label recognition model, so that the label recognition model extracts the image features in the inspection images and outputs a radio frequency label recognition result according to the image features; among them, the radio frequency label recognition result includes: there is a radio frequency label in the inspection image, or there is no radio frequency label in the inspection image;

[0012] When it is determined that the radio frequency label recognition result is that there is no radio frequency label in the inspection image, perform the next inspection image recognition operation.

[0013] Preferably, before repeatedly performing the inspection image recognition operation, it further includes:

[0014] Repeatedly perform the regional image recognition operation until it is determined that the electric tower pole recognition result is that there is an electric tower pole in the regional image, and use the regional image with the electric tower pole as the target regional image;

[0015] Generate the position of the electric tower pole in the target regional image according to the current position of the drone and the relative position of the drone relative to the electric tower pole;

[0016] Obtain the fixed position of the electric tower pole according to the position in the target regional image;

[0017] Control the drone to fly to the fixed position of the electric tower pole, and then control the drone to fly around the electric tower pole and capture inspection images;

[0018] Among them, the regional image recognition operation includes:

[0019] Obtain the regional images collected by the drone in the inspection area of the electric tower pole;

[0020] Input the regional images into a preset electric tower pole recognition model, so that the electric tower pole recognition model extracts the global features and local features of the regional images, and outputs an electric tower pole recognition result according to the global features and the local features; among them, the electric tower pole recognition result includes: there is an electric tower pole in the regional image, or there is no electric tower pole in the regional image;

[0021] When it is determined that the electric tower pole recognition result is that there is no electric tower pole in the regional image, perform the next regional image recognition operation.

[0022] Preferably, before obtaining the regional images collected by the drone in the inspection area of the electric tower pole, it further includes:

[0023] Obtain the scanning width of the drone and the width of the inspection area of the electric tower pole;

[0024] According to the scanning width, vertically divide the inspection area of the electric tower pole to generate a number of sub-inspection areas;

[0025] Connect the vertical centerlines of all sub-inspection areas to generate a first inspection route;

[0026] Determine whether the starting position of the drone is within the inspection area of the electric tower pole;

[0027] If not, use the first inspection route as the target inspection route of the drone;

[0028] If so, use the sub-inspection area where the starting position of the drone is located as the target sub-inspection area, and use the end closest to the starting position of the drone on the vertical centerline of the target sub-inspection area as the inspection starting point; Obtain the number of sub-inspection areas on the left and right sides of the target sub-inspection area; Starting from the inspection starting point, connect the vertical centerlines of all sub-inspection areas on the side with the least number of sub-inspection areas to generate a second inspection route; Use the straight-line route between the end point of the second inspection route and the inspection starting point as the return route; Starting from the inspection starting point, and using the vertical centerline corresponding to the target sub-inspection area as the starting segment, connect the vertical centerlines of all sub-inspection areas on the side with the larger number of sub-inspection areas to obtain a third inspection route; Generate the target inspection route of the drone according to the second inspection route, the return route, and the third inspection route;

[0029] Control the drone to fly on the target inspection route, and at the same time control the drone to collect the area images corresponding to the inspection area of the electric tower pole.

[0030] Preferably, the generation of the label recognition model includes:

[0031] Repeat the model training operation until it is determined that the current iteration number is the same as the preset iteration number, and use the individual of the output hyperparameters with the maximum fitness as the target hyperparameter individual; Among them, the target hyperparameter individual includes a number of selected hyperparameters; The hyperparameters are used to adjust the model performance of the label recognition model;

[0032] Use each selected hyperparameter as the final hyperparameter of the label recognition model, and then generate a trained label recognition model;

[0033] Among them, the model training operation includes:

[0034] Obtain the current iteration number;

[0035] When it is determined that the current iteration number is less than the preset iteration number, several inspection image samples and the corresponding actual radio frequency tag recognition results of each inspection image sample are used as inputs, the predicted radio frequency tag recognition results of each inspection image sample are used as outputs, and the tag recognition model to be trained is iteratively trained according to several current hyperparameter individuals, and the loss value corresponding to each current hyperparameter individual during each iterative training is output; among them, when the model training operation is first executed, several current hyperparameter individuals are randomly generated;

[0036] The current hyperparameter individual with the smallest loss value is used as the first hyperparameter individual;

[0037] Randomly select each current hyperparameter individual except the first hyperparameter individual corresponding to a target iteration number during iterative training and mark them as to-be-processed hyperparameter individuals, and mark each hyperparameter in each to-be-processed hyperparameter individual as a to-be-processed hyperparameter;

[0038] According to the parameter values corresponding to the hyperparameters in the first hyperparameter individual, the parameter values of the to-be-processed hyperparameters of other hyperparameter individuals, the hyperparameter mean values corresponding to each to-be-processed hyperparameter individual, the parameter values of each to-be-processed hyperparameter, the preset upper limit value of each to-be-processed hyperparameter, and the preset lower limit value of each to-be-processed hyperparameter, each to-be-processed hyperparameter in each to-be-processed hyperparameter individual is searched once to generate several updated parameter values corresponding to each to-be-processed hyperparameter; among them, the other hyperparameter individuals are hyperparameter individuals randomly obtained in the target iteration number and except the to-be-processed hyperparameter individuals;

[0039] According to the several updated parameter values corresponding to each to-be-processed hyperparameter, several second hyperparameter individuals are generated;

[0040] According to the fitness of the first hyperparameter individual, the fitness corresponding to each second hyperparameter individual, and the Euclidean distance between each second hyperparameter individual and the first hyperparameter individual, each second hyperparameter individual is searched twice to generate several updated parameter values corresponding to each hyperparameter in each second hyperparameter individual; among them, the fitness is used to characterize the degree of reduction of the loss value of the hyperparameter individual to the model;

[0041] According to the several updated parameter values corresponding to each hyperparameter in each second hyperparameter individual, a third hyperparameter individual is generated;

[0042] According to the fitness of the first hyperparameter individual, the fitness corresponding to each third hyperparameter individual, and the Euclidean distance between each third hyperparameter individual and the first hyperparameter individual, each third hyperparameter individual is searched three times to generate several updated parameter values corresponding to each hyperparameter in each third hyperparameter individual;

[0043] Generate a fourth hyperparameter individual based on a number of updated parameter values corresponding to each hyperparameter in each third hyperparameter individual;

[0044] Perform four searches on the hyperparameters in each fourth hyperparameter individual according to the fitness of each fourth hyperparameter individual and a number of random numbers, and generate a number of updated parameter values corresponding to each hyperparameter in each fourth hyperparameter individual;

[0045] Generate a fifth hyperparameter individual based on a number of updated parameter values corresponding to each hyperparameter in each fourth hyperparameter individual;

[0046] Obtain a first fitness mean value according to the fitness of each fifth hyperparameter individual and the total number of fifth hyperparameter individuals; obtain a second fitness mean value according to the fitness of each fourth hyperparameter individual and the total number of fourth hyperparameter individuals;

[0047] When it is determined that the first fitness mean value is greater than the second fitness mean value, then use each fifth hyperparameter individual as the hyperparameter individual to be output; when it is determined that the first fitness mean value is not greater than the second fitness mean value, then use each fourth hyperparameter individual as the hyperparameter individual to be output;

[0048] Add the current iteration number value to the preset iteration number increment to obtain an updated iteration number, and use the updated iteration number as the iteration number for the next execution of the model training operation;

[0049] Use each hyperparameter individual to be output as each current hyperparameter individual for the next execution of the model training operation.

[0050] Preferably, the generation process of a number of updated parameter values corresponding to each hyperparameter to be processed includes:

[0051] For each hyperparameter to be processed in each hyperparameter individual to be processed, generate a search range control factor corresponding to each hyperparameter to be processed according to the parameter value corresponding to the hyperparameter in the first hyperparameter individual and the parameter values of the hyperparameters to be processed in other hyperparameter individuals;

[0052] For each hyperparameter individual to be processed, generate a hyperparameter mean value corresponding to the hyperparameter individual to be processed according to the parameter value of each hyperparameter to be processed in the hyperparameter individual to be processed and the total number of hyperparameters to be processed in the hyperparameter individual to be processed;

[0053] For each hyperparameter to be processed in each hyperparameter individual to be processed, generate a search step size for the hyperparameter to be processed according to the parameter value corresponding to the hyperparameter in the first hyperparameter individual, the parameter value of the hyperparameter to be processed, the hyperparameter mean value corresponding to the hyperparameter individual to be processed, the preset upper limit value of the hyperparameter to be processed, and the preset lower limit value of the hyperparameter to be processed;

[0054] According to the parameter values of each hyperparameter in the first hyperparameter individual, each search range control factor corresponding to each hyperparameter to be processed, and each search step of each hyperparameter to be processed, perform a search on each hyperparameter to be processed in the hyperparameter individual to be processed, and generate several updated parameter values corresponding to each hyperparameter to be processed.

[0055] Preferably, the generation process of several updated parameter values corresponding to each hyperparameter in each second hyperparameter individual includes:

[0056] According to the preset fitness function formula, calculate the fitness corresponding to the first hyperparameter individual and each second hyperparameter individual respectively;

[0057] For each second hyperparameter individual, generate the regional importance degree of the second hyperparameter individual according to the fitness corresponding to the second hyperparameter individual, the fitness corresponding to the first hyperparameter individual, and the minimum fitness among each second hyperparameter individual; wherein, the regional importance degree of the second hyperparameter individual is used to characterize the closeness between the second hyperparameter individual and the first hyperparameter individual;

[0058] Perform a neighborhood search on each second hyperparameter individual, and then obtain several first neighborhood hyperparameter individuals within the preset Euclidean distance range, and take the first neighborhood hyperparameter individual with the largest regional importance degree as the first target neighborhood hyperparameter individual;

[0059] For each second hyperparameter individual, calculate the influence factor corresponding to the second hyperparameter individual according to the fitness corresponding to the first hyperparameter individual and the Euclidean distance between the second hyperparameter individual and the first hyperparameter individual; wherein, the influence factor corresponding to the second hyperparameter individual is used to characterize the influence degree of the first hyperparameter individual on the second hyperparameter individual;

[0060] According to the regional importance degree of each second hyperparameter individual, the first target neighborhood hyperparameter individual, and the influence factors corresponding to each second hyperparameter individual, perform a secondary search on the hyperparameters in each second hyperparameter individual to generate several updated parameter values corresponding to each hyperparameter in each second hyperparameter individual.

[0061] Preferably, the generation process of several updated parameter values corresponding to each hyperparameter in each third hyperparameter individual includes:

[0062] According to the preset fitness function formula, calculate the fitness corresponding to each third hyperparameter individual;

[0063] For each third hyperparameter individual, generate the regional importance degree of the third hyperparameter individual according to the fitness corresponding to the third hyperparameter individual, the fitness corresponding to the first hyperparameter individual, and the minimum fitness among all third hyperparameter individuals; wherein, the regional importance degree of the third hyperparameter individual is used to characterize the proximity between the third hyperparameter individual and the first hyperparameter individual;

[0064] Perform neighborhood search on each third hyperparameter individual, and then obtain a number of second neighborhood hyperparameter individuals within a preset Euclidean distance range. Take the second neighborhood hyperparameter individual with the largest regional importance degree as the second target neighborhood hyperparameter individual; take the second neighborhood hyperparameter individual with the closest distance to the second target neighborhood hyperparameter individual as the third target neighborhood hyperparameter individual;

[0065] For each third hyperparameter individual, calculate the influence factor corresponding to the third hyperparameter individual according to the fitness corresponding to the first hyperparameter individual and the Euclidean distance between the third hyperparameter individual and the first hyperparameter individual; wherein, the influence factor corresponding to the third hyperparameter individual is used to characterize the influence degree of the first hyperparameter individual on the third hyperparameter individual;

[0066] According to the regional importance degree of each third hyperparameter individual, the second target neighborhood hyperparameter individual, the third target neighborhood hyperparameter individual, and the influence factor corresponding to each third hyperparameter individual, perform three searches on the hyperparameters in each third hyperparameter individual to generate a number of updated parameter values corresponding to each hyperparameter in each third hyperparameter individual.

[0067] Preferably, the generation process of a number of updated parameter values corresponding to each hyperparameter in each fourth hyperparameter individual includes:

[0068] Calculate the fitness corresponding to each fourth hyperparameter individual according to a preset fitness function formula;

[0069] According to the fourth hyperparameter individual with the maximum fitness, the fourth hyperparameter individual with the minimum fitness, and a number of random numbers, perform four searches on the hyperparameters in each fourth hyperparameter individual to generate a number of updated parameter values corresponding to each hyperparameter in each fourth hyperparameter individual.

[0070] Preferably, the generation of the electric tower pole recognition model includes:

[0071] Use a number of regional image samples and the actual electric tower pole recognition results corresponding to each regional image sample as inputs, and the predicted electric tower pole recognition results of each regional image sample as outputs, and perform iterative training on the electric tower pole recognition model to be trained until the model converges, and generate a preset electric tower pole recognition model.

[0072] Based on the above method embodiments, the present invention correspondingly provides device embodiments.

[0073] An embodiment of the present invention provides a tower pole inspection device based on tag recognition, including: a target inspection image generation module, a tag position determination module, and an inspection data generation module;

[0074] The target inspection image generation module is configured to repeatedly execute an inspection image recognition operation until, when it is determined that the radio frequency tag recognition result indicates that there is a radio frequency tag in the inspection image, the inspection image with the radio frequency tag is used as the target inspection image; wherein, the radio frequency tag is used to display the operation information on the tower pole;

[0075] The tag position determination module is configured to generate the tag position of the radio frequency tag in the target inspection image according to the current position of the unmanned aerial vehicle, the fixed position of the tower pole, and the relative position of the unmanned aerial vehicle relative to the tower pole;

[0076] The inspection data generation module is configured to, after obtaining the actual position of the radio frequency tag according to the tag position, control the unmanned aerial vehicle to fly to the actual position of the radio frequency tag, and then control the unmanned aerial vehicle to read the operation information of the radio frequency tag, and generate the inspection data of the tower pole according to the read operation information;

[0077] Among them, the inspection image recognition operation includes:

[0078] Obtain the inspection image captured by the unmanned aerial vehicle when flying around the tower pole;

[0079] Input the inspection image into a preset tag recognition model, so that the tag recognition model extracts the image features in the inspection image, and outputs a radio frequency tag recognition result according to the image features; wherein, the radio frequency tag recognition result includes: there is a radio frequency tag in the inspection image, or there is no radio frequency tag in the inspection image;

[0080] When it is determined that the radio frequency tag recognition result indicates that there is no radio frequency tag in the inspection image, the next inspection image recognition operation is executed.

[0081] By implementing the present invention, the following beneficial effects are achieved:

[0082] The embodiment of the present invention provides a method and device for inspecting electric tower poles based on tag recognition. The present invention uses a drone to capture inspection images while flying around the electric tower poles, and uses a preset tag recognition model to perform feature extraction and tag recognition on the inspection images. When a radio frequency tag is identified in the inspection image, the image is used as a target inspection image, and the position of the radio frequency tag in the target inspection image is generated according to the current position of the drone, the fixed position of the electric tower pole, and the relative position information of the drone relative to the electric tower pole. Therefore, after the drone is controlled to fly to the tag position, the operation information in the radio frequency tag is automatically read to obtain the inspection data of the electric tower pole. Through the inspection and image recognition of the drone, the automatic collection and transmission of the inspection information is realized. Compared with the prior art, the present invention adopts drone inspection, and through automatic tag recognition and reading of radio frequency tags, manual intervention is reduced and inspection efficiency is improved. Moreover, the present invention utilizes drones to continuously take pictures and continuously identify tags until the radio frequency tag can be identified. This ensures accurate acquisition of the operation information of the electric tower even in severe weather or complex terrain conditions. The present invention can monitor the operation status of the electric tower in real time and accurately through automated tag identification and operation data collection, and can promptly discover and eliminate equipment defects of the electric tower, thereby ensuring the stability and safety of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 The present invention is a flowchart of a method for inspecting electric towers based on tag recognition according to an embodiment of the present invention.

[0084] Figure 2 It is a structural schematic diagram of an electric tower inspection device based on tag recognition provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] like Figure 1 FIG. 1 is a flow chart of a method for inspecting electric towers based on tag identification according to an embodiment of the present invention. The method for inspecting electric towers based on tag identification comprises:

[0087] Step S1: Repeatedly perform the inspection image recognition operation until it is determined that there is a radio frequency tag in the inspection image, and then use the inspection image with the radio frequency tag as the target inspection image; wherein, the radio frequency tag is used to display the operation information on the electric tower pole.

[0088] Among them, the inspection image recognition operation includes:

[0089] Obtain the inspection images captured by the drone when flying around the electric tower pole.

[0090] Input the inspection image into a preset tag recognition model, so that the tag recognition model extracts the image features in the inspection image and outputs the radio frequency tag recognition result according to the image features; wherein, the radio frequency tag recognition result includes: there is a radio frequency tag in the inspection image, or there is no radio frequency tag in the inspection image.

[0091] When it is determined that the radio frequency tag recognition result is that there is no radio frequency tag in the inspection image, perform the next inspection image recognition operation.

[0092] Step S2: Generate the tag position of the radio frequency tag in the target inspection image according to the current position of the drone, the fixed position of the electric tower pole, and the relative position of the drone relative to the electric tower pole.

[0093] After obtaining the actual position of the radio frequency tag according to the tag position, control the drone to fly to the actual position of the radio frequency tag, and then control the drone to read the operation information of the radio frequency tag, and generate the inspection data of the electric tower pole according to the read operation information.

[0094] For step S1, in a preferred embodiment, before the present invention recognizes the inspection image, it can also recognize the electric tower pole from the area images collected in the inspection area of the electric tower pole, and then based on the recognized electric tower pole, control the drone to fly to the periphery of the electric tower pole to collect the inspection images about the radio frequency tag, then there is:

[0095] Repeatedly perform the area image recognition operation until it is determined that the electric tower pole recognition result is that there is an electric tower pole in the area image, and use the area image with the electric tower pole as the target area image.

[0096] Generate the position of the electric tower pole in the target area image according to the current position of the drone and the relative position of the drone relative to the electric tower pole.

[0097] Obtain the fixed position of the electric tower pole according to the position in the target area image.

[0098] Control the drone to fly to the fixed position of the electric tower pole, and then control the drone to fly around the electric tower pole and capture the inspection images.

[0099] Among them, the regional image recognition operation includes:

[0100] Obtain the regional image collected by the drone in the inspection area of the electric tower pole;

[0101] Input the regional image into a preset electric tower pole recognition model, so that the electric tower pole recognition model extracts the global features and local features of the regional image, and outputs an electric tower pole recognition result according to the global features and the local features; among them, the electric tower pole recognition result includes: there is an electric tower pole in the regional image, or there is no electric tower pole in the regional image;

[0102] When it is determined that the electric tower pole recognition result is that there is no electric tower pole in the regional image, the next regional image recognition operation is executed.

[0103] Specifically, when controlling the drone to fly to a fixed position of the electric tower pole, the drone can be controlled to collect inspection images in the nearby area of the electric tower pole. Moreover, by first identifying the position of the electric tower pole, the drone can fly more targeted to the surrounding of the electric tower pole to collect inspection images, which can avoid the drone flying blindly in the inspection area, reduce unnecessary flight time and energy consumption, and thus improve the inspection efficiency.

[0104] Furthermore, before collecting the regional image, the inspection route of the drone can also be automatically planned, so that the drone can quickly collect images of the electric tower pole in the inspection area where there is an electric tower pole within the planned route to assist subsequent image analysis. Then, before obtaining the regional image collected by the drone in the inspection area of the electric tower pole, it further includes:

[0105] Obtain the scanning width of the drone and the width of the inspection area of the electric tower pole;

[0106] According to the scanning width, divide the inspection area of the electric tower pole vertically to generate a number of sub-inspection areas;

[0107] Connect the vertical center lines of all sub-inspection areas to generate a first inspection route;

[0108] Judge whether the starting position of the drone is within the inspection area of the electric tower pole;

[0109] If not, use the first inspection route as the target inspection route of the drone;

[0110] If so, take the sub-inspection area where the starting position of the UAV is located as the target sub-inspection area, and take the end of the vertical center line of the target sub-inspection area that is closest to the starting position of the UAV as the starting point of the inspection; obtain the number of sub-inspection areas on the left and right sides of the target sub-inspection area; starting from the starting point of the inspection, connect the vertical center lines of all the sub-inspection areas on the side with the smallest number of sub-inspection areas to generate a second inspection route; take the straight-line route between the end point of the second inspection route and the starting point of the inspection as the return route; starting from the starting point of the inspection and taking the vertical center line corresponding to the target sub-inspection area as the starting segment, connect the vertical center lines of all the sub-inspection areas on the side with the larger number of sub-inspection areas to obtain a third inspection route; generate the target inspection route of the UAV according to the second inspection route, the return route and the third inspection route.

[0111] Control the UAV to fly on the target inspection route, and at the same time control the UAV to collect the area images corresponding to the inspection area of the electric tower pole.

[0112] It can be understood that first, the scanning area input by the staff through human-computer interaction can be obtained, and the UAV inspection route can be planned in the scanning area by using a route planning algorithm; a map can also be provided for the staff through a third-party map, and then multiple coordinate points determined by the staff through human-computer interaction are received, and when the multiple coordinate points enclose a closed area, the scanning area is obtained. Since the shooting range of the UAV is fixed and the scanning area cannot be inspected in one go, it is necessary to use a route planning algorithm to plan the UAV inspection route in the scanning area so that the UAV can perform inspections in the scanning area. Then there is:

[0113] Based on the scanning width of the UAV, the scanning area is vertically divided into multiple sub-scanning areas. When the width of the remaining area is less than the scanning width, the remaining area is taken as a separate sub-scanning area, and the vertical center line corresponding to each sub-scanning area is determined. The vertical center lines of all the scanning areas are connected in a serpentine shape to obtain a first inspection route.

[0114] Determine the starting position corresponding to the UAV to perform the inspection task, and judge whether the starting position is outside the scanning area. If so, take the initial inspection route as the UAV inspection route, otherwise perform secondary planning.

[0115] Determine the sub-scanning area where the starting position of the UAV is located as the target sub-scanning area, and take the end of the center line corresponding to the target sub-scanning area that is closer to the starting position of the UAV as the starting point of the inspection.

[0116] Determine the number of sub-scanning areas on the left and right sides of the target sub-scanning area, and based on the starting point of the inspection tour, connect the vertical centerlines of all sub-scanning areas on the side with the smaller number of sub-scanning areas in a serpentine manner to obtain the second inspection tour route;

[0117] Use the straight-line route between the end point of the first inspection tour route and the starting point of the inspection tour as the return route;

[0118] Based on the starting point of the inspection tour, and using the centerline corresponding to the target sub-scanning area as the starting segment, connect the vertical centerlines of all sub-scanning areas on the side with the larger number of sub-scanning areas in a serpentine manner to obtain the third inspection tour route;

[0119] Use the second inspection tour route, the return route, and the third inspection tour route together as the inspection tour route of the drone during secondary planning.

[0120] In a preferred embodiment, in order to simplify the inspection tour route of the drone, the scanning area can also be rotated first so that the longest side of the scanning area is in a vertical state, thereby reducing the number of sub-scanning areas and achieving the purpose of simplifying the route.

[0121] Furthermore, control the drone to conduct inspections within the scanning area according to the inspection tour route of the drone, and during the inspection process, collect inspection video frame data in real time, and collect the video frame data during the inspection process at a preset data sampling frequency to obtain the inspection video frame data, thereby obtaining the area image corresponding to the inspection area of the electric tower pole.

[0122] In a preferred implementation, an electric tower pole recognition model can be obtained by constructing a deep learning model.

[0123] In a preferred implementation, the generation of the electric tower pole recognition model includes:

[0124] Using a number of area image samples and the corresponding actual electric tower pole recognition results of each area image sample as inputs, and the predicted electric tower pole recognition results of each area image sample as outputs, iteratively train the electric tower pole recognition model to be trained until the model converges, and generate a preset electric tower pole recognition model.

[0125] In a preferred implementation, a label recognition model can also be obtained by constructing another deep learning model.

[0126] In a preferred implementation, the generation of the label recognition model includes:

[0127] Repeat the model training operation until, when it is determined that the current iteration number is the same as the preset iteration number, the hyperparameter individual with the maximum fitness among the hyperparameters to be output is used as the target hyperparameter individual; wherein, the target hyperparameter individual includes several selected hyperparameters; the hyperparameters are used to adjust the model performance of the label recognition model;

[0128] Use each selected hyperparameter as the final hyperparameter of the label recognition model, and then generate a trained label recognition model;

[0129] Among them, the model training operation includes:

[0130] Obtain the current iteration number;

[0131] When it is determined that the current iteration number is less than the preset iteration number, use several inspection image samples and the corresponding actual radio frequency label recognition results of each inspection image sample as inputs, use the predicted radio frequency label recognition results of each inspection image sample as outputs, and iteratively train the label recognition model to be trained according to several current hyperparameter individuals, and output the loss value corresponding to each current hyperparameter individual during each iterative training; wherein, when the model training operation is initially executed, several current hyperparameter individuals are randomly generated;

[0132] Use the current hyperparameter individual with the minimum loss value as the first hyperparameter individual;

[0133] Randomly select each current hyperparameter individual corresponding to a target iteration number during iterative training except the first hyperparameter individual and mark it as a hyperparameter individual to be processed, and mark each hyperparameter in each hyperparameter individual to be processed as a hyperparameter to be processed;

[0134] According to the parameter values corresponding to the hyperparameters in the first hyperparameter individual, the parameter values of the hyperparameters to be processed of other hyperparameter individuals, the hyperparameter mean values corresponding to each hyperparameter individual to be processed, the parameter values of each hyperparameter to be processed, the preset upper limit value of each hyperparameter to be processed, and the preset lower limit value of each hyperparameter to be processed, perform a search on each hyperparameter to be processed in each hyperparameter individual to be processed to generate several updated parameter values corresponding to each hyperparameter to be processed; wherein, the other hyperparameter individuals are hyperparameter individuals randomly obtained in the target iteration number and except the hyperparameter individuals to be processed;

[0135] Generate several second hyperparameter individuals according to the several updated parameter values corresponding to each hyperparameter to be processed;

[0136] Perform a secondary search on each second hyperparameter individual based on the fitness of the first hyperparameter individual, the fitness corresponding to each second hyperparameter individual, and the Euclidean distance between each second hyperparameter individual and the first hyperparameter individual, to generate several updated parameter values corresponding to each hyperparameter in each second hyperparameter individual; wherein, the fitness is used to characterize the degree of reduction of the loss value of the hyperparameter individual to the model;

[0137] Generate a third hyperparameter individual based on several updated parameter values corresponding to each hyperparameter in each second hyperparameter individual;

[0138] Perform a tertiary search on each third hyperparameter individual based on the fitness of the first hyperparameter individual, the fitness corresponding to each third hyperparameter individual, and the Euclidean distance between each third hyperparameter individual and the first hyperparameter individual, to generate several updated parameter values corresponding to each hyperparameter in each third hyperparameter individual;

[0139] Generate a fourth hyperparameter individual based on several updated parameter values corresponding to each hyperparameter in each third hyperparameter individual;

[0140] Perform a quaternary search on the hyperparameters in each fourth hyperparameter individual based on the fitness of each fourth hyperparameter individual and several random numbers, to generate several updated parameter values corresponding to each hyperparameter in each fourth hyperparameter individual;

[0141] Generate a fifth hyperparameter individual based on several updated parameter values corresponding to each hyperparameter in each fourth hyperparameter individual;

[0142] Obtain a first fitness mean value according to the fitness corresponding to each fifth hyperparameter individual and the total number of fifth hyperparameter individuals; obtain a second fitness mean value according to the fitness corresponding to each fourth hyperparameter individual and the total number of fourth hyperparameter individuals;

[0143] When it is determined that the first fitness mean value is greater than the second fitness mean value, then use each fifth hyperparameter individual as the hyperparameter individual to be output; when it is determined that the first fitness mean value is not greater than the second fitness mean value, then use each fourth hyperparameter individual as the hyperparameter individual to be output;

[0144] Add the preset iteration number increment to the current iteration number value to obtain an updated iteration number, and use the updated iteration number as the iteration number for the next execution of the model training operation;

[0145] Use each hyperparameter individual to be output as each current hyperparameter individual for the next execution of the model training operation.

[0146] Schematically, in the process of generating the label recognition model of the present invention, the final label recognition model can be generated by means of iterative training and optimizing hyperparameters. In each iteration, the inspection image samples and their corresponding actual radio frequency label recognition results are used as inputs to train the label recognition model. In the initial iteration, a number of hyperparameter individuals (i.e., different combinations of hyperparameters) are randomly generated, and then through multiple searches and updates, new hyperparameter individuals are generated to find the optimal combination of hyperparameters. Then, this combination of hyperparameters (i.e., the target hyperparameter individual) is used as the final hyperparameters of the label recognition model, and then the trained label recognition model is generated.

[0147] It can be understood that hyperparameters are used to adjust the model performance of the label recognition model. For each hyperparameter individual, it may include: the size of the data volume loaded each time during model training, the learning rate, the number of hidden layers, the size of the hidden layers, the regularization parameter, and the dropout rate, etc.

[0148] The present invention first selects the optimal current hyperparameter individual, i.e., the first hyperparameter individual, according to the loss value. Then, multiple searches and updates are performed on other hyperparameter individuals to generate new hyperparameter individuals (such as the second, third, fourth, and fifth hyperparameter individuals), and each search is based on different strategies, such as based on parameter values, hyperparameter means, preset upper and lower limits, fitness, and Euclidean distance, etc. By continuously optimizing the hyperparameters, the optimal model configuration can be found, thereby improving the recognition performance and accuracy of the model. Moreover, multiple iterations and different types of search methods can enable the model to be trained under different combinations of hyperparameters, thereby enhancing the generalization ability of the model.

[0149] In a preferred embodiment, when obtaining a number of updated parameter values corresponding to each hyperparameter to be processed through one search, specifically:

[0150] For each hyperparameter to be processed in each hyperparameter individual to be processed, a search range control factor corresponding to each hyperparameter to be processed is generated according to the parameter value corresponding to the hyperparameter in the first hyperparameter individual and the parameter values of the hyperparameters to be processed in other hyperparameter individuals.

[0151] For each hyperparameter individual to be processed, a hyperparameter mean corresponding to the hyperparameter individual to be processed is generated according to the parameter value of each hyperparameter to be processed in the hyperparameter individual to be processed and the total number of hyperparameters to be processed in the hyperparameter individual to be processed.

[0152] For each to-be-processed hyperparameter in each to-be-processed hyperparameter individual, according to the parameter value corresponding to the hyperparameter in the first hyperparameter individual, the parameter value of the to-be-processed hyperparameter, the hyperparameter mean corresponding to the to-be-processed hyperparameter individual, the preset upper limit value of the to-be-processed hyperparameter, and the preset lower limit value of the to-be-processed hyperparameter, generate the search step size of the to-be-processed hyperparameter;

[0153] According to the parameter values of the various hyperparameters in the first hyperparameter individual, the respective search range control factors corresponding to each to-be-processed hyperparameter, and the respective search step sizes of each to-be-processed hyperparameter, perform a search on each to-be-processed hyperparameter in the to-be-processed hyperparameter individual once, and generate several updated parameter values corresponding to each to-be-processed hyperparameter.

[0154] Schematically, by considering the hyperparameter values in the first hyperparameter individual and the to-be-processed hyperparameter values of other hyperparameter individuals, the search range of each to-be-processed hyperparameter can be dynamically adjusted, so that a more detailed search can be realized in the region more likely to contain the optimal solution, thereby improving the search efficiency. According to the current value, mean, preset upper and lower limits of the to-be-processed hyperparameter, and the value of the first hyperparameter individual, a suitable search step size can be calculated, ensuring that the search process is neither too rough to miss the optimal solution nor too fine to waste computing resources.

[0155] Calculate the mean of the to-be-processed hyperparameter individual, then the overall performance level of the current hyperparameter configuration can be evaluated. When using it as the basis for adjusting the search strategy, for example, when the mean is low, the search range can be increased or the search direction can be changed.

[0156] In the above-mentioned one-time search process, the entire search process is dynamically adjusted based on the current hyperparameter configuration and search results. Then, as the search progresses, the search strategy will gradually converge to the direction more likely to find the optimal solution.

[0157] Schematically, the above-mentioned one-time search adopts a fast search algorithm, and then the search range control factor can be obtained according to the following formula:

[0158]

[0159] where, represents the search range control factor corresponding to the d-th hyperparameter of the i-th to-be-processed hyperparameter individual in the t-th training process, d = 1, 2,..., D, D represents the total number of hyperparameters in the to-be-processed hyperparameter individual, i = 1, 2,..., N, N represents the total number of to-be-processed hyperparameter individuals; schematically, the t-th time is a certain target iteration number randomly selected during the above-mentioned iterative training;

[0160] represents the d-th hyperparameter of the to-be-processed hyperparameter individual in the t-th training process, represents the d-th hyperparameter corresponding to the randomly selected hyperparameter individual to be processed during the t-th training process, excluding the first hyperparameter individual. ε represents the third constant term, and ε < 0.0001;

[0161] For the i-th hyperparameter individual to be processed during the t-th training process, the hyperparameter mean value obtained can be:

[0162]

[0163] where represents the d-th dimensional hyperparameter of the i-th hyperparameter individual to be processed during the t-th training process, and χ i represents the hyperparameter mean value corresponding to the i-th hyperparameter individual to be processed during the t-th training process;

[0164] Based on the first hyperparameter individual and the hyperparameter mean values of the hyperparameter individuals to be processed, the search step size is obtained as:

[0165]

[0166] where represents the search step size corresponding to the d-th hyperparameter of the i-th hyperparameter individual to be processed during the t-th training process. c represents the information interaction influence factor for controlling the training process, and ub d represents the upper limit value of the d-th hyperparameter to be processed, and lb d represents the lower limit value of the d-th hyperparameter to be processed;

[0167] Then, based on the first hyperparameter individual, the search range control factor, and the search step size, the hyperparameter individuals to be processed can be quickly searched, and the second hyperparameter individual after the quick search can be:

[0168]

[0169] where represents the d-th dimensional hyperparameter of the second hyperparameter individual after the i-th quick search. γ represents the search accuracy, and r1 represents the first random number between (0, 1).

[0170] In a preferred embodiment, when obtaining several updated parameter values corresponding to each hyperparameter among the second hyperparameter individuals through secondary search, specifically:

[0171] According to the preset fitness function formula, the fitness values corresponding to the first hyperparameter individual and each of the second hyperparameter individuals are calculated respectively;

[0172] For each second hyperparameter individual, based on the fitness corresponding to the second hyperparameter individual, the fitness corresponding to the first hyperparameter individual, and the minimum fitness among the second hyperparameter individuals, the regional importance degree of the second hyperparameter individual is generated; wherein, the regional importance degree of the second hyperparameter individual is used to characterize the proximity of the second hyperparameter individual to the first hyperparameter individual;

[0173] Perform neighborhood search on each second hyperparameter individual, and then obtain a number of first neighborhood hyperparameter individuals within a preset Euclidean distance range. The first neighborhood hyperparameter individual with the largest regional importance degree is used as the first target neighborhood hyperparameter individual;

[0174] For each second hyperparameter individual, based on the fitness corresponding to the first hyperparameter individual and the Euclidean distance between the second hyperparameter individual and the first hyperparameter individual, the influence factor corresponding to the second hyperparameter individual is calculated; wherein, the influence factor corresponding to the second hyperparameter individual is used to characterize the influence degree of the first hyperparameter individual on the second hyperparameter individual;

[0175] Based on the regional importance degrees of the second hyperparameter individuals, the first target neighborhood hyperparameter individual, and the influence factors corresponding to the second hyperparameter individuals, perform a secondary search on the hyperparameters in each second hyperparameter individual to generate a number of updated parameter values corresponding to each hyperparameter in each second hyperparameter individual.

[0176] Illustratively, in the above process, by calculating the regional importance degree of the second hyperparameter individual, it is possible to identify which hyperparameter individuals are closer to the region where the optimal solution is located, enabling more targeted exploration of these important regions in subsequent searches, thereby improving the search efficiency. The influence factor is used to characterize the influence degree of the first hyperparameter individual on the second hyperparameter individual. By calculating the influence factor, the search direction and step size of the second hyperparameter individual can be intelligently adjusted, and decisions can be made based on the current situation, thus more effectively finding the optimal solution.

[0177] Not only is neighborhood search performed to obtain the first neighborhood hyperparameter individuals within a preset Euclidean distance range, but also the idea of global search is combined to evaluate the fitness of each second hyperparameter individual. This combination can maintain the local search accuracy while also having the ability of global search, preventing getting stuck in local optimal solutions during the search process and enabling exploration of the possible solution space in a wider range.

[0178] If the above secondary search adopts a slow search strategy, the slow search strategy provided by the embodiments of the present invention enables the second hyperparameter individuals to learn the information of the first hyperparameter individuals, and then combines the positions of the second hyperparameter individuals themselves, enabling the second hyperparameter individuals with better positions to perform more refined searches and the second hyperparameter individuals with worse positions to perform faster searches, thereby achieving neighborhood search in the optimal direction and being able to effectively search the local area.

[0179] In a preferred embodiment, the regional importance of each second hyperparameter individual can be determined according to the following formula:

[0180]

[0181] Where SV m represents the regional importance corresponding to the m-th second hyperparameter individual after fast search, m = 1, 2,..., N, N represents the total number of second hyperparameter individuals, f best represents the fitness corresponding to the first hyperparameter individual, f m represents the fitness corresponding to the second hyperparameter individual after fast search, f worst represents the fitness corresponding to the worst individual among the first hyperparameter individuals after fast search, and the fitness is obtained by adding the loss function value and the third constant term and taking the reciprocal;

[0182] Obtain M neighborhood hyperparameter individuals with the smallest Euclidean distance from the first hyperparameter individual within the neighborhood range, and determine the optimal neighborhood hyperparameter individual with the greatest regional importance according to the M neighborhood hyperparameter individuals;

[0183] Based on the regional importance corresponding to the first hyperparameter individual, the influence factor of the first hyperparameter individual on the second hyperparameter individual after fast search is determined as:

[0184]

[0185] Where ξ mbest represents the influence factor of the first hyperparameter individual on the m-th second hyperparameter individual, SV best represents the regional importance corresponding to the first hyperparameter individual, e represents the natural constant, dist mbest represents the Euclidean distance between the first hyperparameter individual and the m-th second hyperparameter individual;

[0186] According to the influence factor, regional importance, and optimal neighborhood hyperparameter individual, perform a slow search on the second hyperparameter individual, and the third hyperparameter individual after slow search is obtained as:

[0187]

[0188] Where represents the d-th hyperparameter of the m-th second hyperparameter individual during the t-th training process, represents the d-th hyperparameter of the m-th third hyperparameter individual, represents the d-th hyperparameter of the first target neighborhood hyperparameter individual corresponding to the m-th second hyperparameter individual, and r2 represents a second random number between (0, 1), represents the d-th hyperparameter of the first hyperparameter individual during the t-th training process, and r3 represents a third random number between (-1, 1).

[0189] In a preferred embodiment, when obtaining several updated parameter values corresponding to each hyperparameter in each third hyperparameter individual through three searches, specifically:

[0190] Calculate the fitness corresponding to each third hyperparameter individual according to the preset fitness function formula;

[0191] For each third hyperparameter individual, generate the regional importance degree of the third hyperparameter individual according to the fitness corresponding to the third hyperparameter individual, the fitness corresponding to the first hyperparameter individual, and the minimum fitness among all third hyperparameter individuals; wherein, the regional importance degree of the third hyperparameter individual is used to characterize the closeness between the third hyperparameter individual and the first hyperparameter individual;

[0192] Perform neighborhood search on each third hyperparameter individual, and then obtain several second neighborhood hyperparameter individuals within the preset Euclidean distance range. Take the second neighborhood hyperparameter individual with the largest regional importance degree as the second target neighborhood hyperparameter individual; take the second neighborhood hyperparameter individual closest to the second target neighborhood hyperparameter individual as the third target neighborhood hyperparameter individual;

[0193] For each third hyperparameter individual, calculate the influence factor corresponding to the third hyperparameter individual according to the fitness corresponding to the first hyperparameter individual and the Euclidean distance between the third hyperparameter individual and the first hyperparameter individual; wherein, the influence factor corresponding to the third hyperparameter individual is used to characterize the influence degree of the first hyperparameter individual on the third hyperparameter individual;

[0194] Perform three searches on the hyperparameters in each third hyperparameter individual according to the regional importance degree of each third hyperparameter individual, the second target neighborhood hyperparameter individual, the third target neighborhood hyperparameter individual, and the influence factor corresponding to each third hyperparameter individual, and generate several updated parameter values corresponding to each hyperparameter in each third hyperparameter individual.

[0195] Schematically, in the embodiments of the present invention, by selecting the individual closest to the second target neighborhood hyperparameter individual (i.e., the individual with better performance in the current search area) as the third target neighborhood hyperparameter individual, the algorithm can perform a more detailed search within a smaller range, so as to find a better hyperparameter combination within the local area, thereby improving the accuracy and efficiency of the search. Since the third target neighborhood hyperparameter individual is close to the second target neighborhood hyperparameter individual, their positions in the hyperparameter space are relatively close. Therefore, in the subsequent search process, the algorithm can more efficiently explore these adjacent areas, thereby reducing unnecessary search overhead and consumption of computing resources.

[0196] Schematically, the above three searches adopt a coordinated search strategy, which can effectively realize information interaction within the neighborhood range, so as to realize the search of more unfamiliar areas, and thus, without increasing a large number of search times, try to realize the traversability of the solution space, thereby increasing the possibility of finding the global optimal solution.

[0197] In a preferred embodiment, for the third hyperparameter individual after slow search, according to the fitness corresponding to each third hyperparameter individual, the regional importance degree of each third hyperparameter individual is determined as:

[0198]

[0199] where SV k represents the regional importance degree corresponding to the kth third hyperparameter individual after slow search, k = 1, 2,..., N, N represents the total number of third hyperparameter individuals, f best represents the fitness corresponding to the first hyperparameter individual, f k represents the fitness corresponding to the kth third hyperparameter individual after slow search, f worst represents the fitness corresponding to the worst individual among the third hyperparameter individuals after slow search; the fitness is obtained by adding the loss function value and the third constant term and taking the reciprocal;

[0200] Within the neighborhood range (a neighborhood radius can be set, and a neighborhood range can be determined through this neighborhood radius), M neighborhood hyperparameter individuals with the smallest Euclidean distance from the third hyperparameter individual are obtained, and based on the M neighborhood hyperparameter individuals, the optimal neighborhood hyperparameter individual with the largest regional importance degree is determined; at the same time, the relatively better neighborhood hyperparameter individual with the closest Euclidean distance to the first hyperparameter individual and a larger regional importance degree (i.e., the above-mentioned third target neighborhood hyperparameter individual) is determined;

[0201] Based on the regional importance degree corresponding to the optimal neighborhood hyperparameter individual, the influence factor of the optimal neighborhood hyperparameter individual on the third hyperparameter individual after slow search is determined as:

[0202]

[0203] Among them, ξ kG represents the influence factor of the optimal neighborhood hyperparameter individual on the third hyperparameter individual after the k-th slow search, SV G represents the importance degree of the region corresponding to the optimal neighborhood hyperparameter individual, e represents the natural constant, dist mG represents the Euclidean distance between the optimal neighborhood hyperparameter individual and the third hyperparameter individual after the m-th slow search;

[0204] According to the influence factor, the importance degree of the region, the optimal neighborhood hyperparameter individual, and the relatively optimal neighborhood hyperparameter individual, a coordinated search is performed on the first hyperparameter individual, and the fourth hyperparameter individual after the coordinated search is obtained as:

[0205]

[0206] Among them, represents the d-th hyperparameter of the third hyperparameter individual after the k-th slow search in the t-th training process, represents the d-th hyperparameter of the fourth hyperparameter individual after the k-th coordinated search, represents the d-th hyperparameter of the optimal neighborhood hyperparameter individual corresponding to the k-th third hyperparameter individual, r4 represents the fourth random number between (0, 1), r5 represents the fifth random number between (0, 1), represents the d-th hyperparameter of the relatively optimal neighborhood hyperparameter individual corresponding to the third hyperparameter individual.

[0207] In a preferred embodiment, when obtaining several updated parameter values corresponding to each hyperparameter in each of the fourth hyperparameter individuals through four searches, specifically:

[0208] According to the preset fitness function formula, the fitness corresponding to each of the fourth hyperparameter individuals is calculated;

[0209] According to the fourth hyperparameter individual with the maximum fitness, the fourth hyperparameter individual with the minimum fitness, and several random numbers, four searches are performed on the hyperparameters in each of the fourth hyperparameter individuals to generate several updated parameter values corresponding to each hyperparameter in each of the fourth hyperparameter individuals.

[0210] It can be understood that the above four-search process is based on the fitness of each of the fourth hyperparameter individuals. The individual with the maximum fitness (the best performance) and the individual with the minimum fitness (the worst performance) are used as the reference points for the search, so that the model can maintain excellent characteristics while also trying to improve the poor parts, thereby being able to more effectively guide the search direction and avoid blind search.

[0211] During the search process, random numbers are also used to jump out of local optima, explore a wider parameter space, and prevent premature convergence.

[0212] Schematically, embodiments of the present invention can obtain the fourth hyperparameter individual with the minimum fitness and the fourth hyperparameter individual with the maximum fitness for the fourth hyperparameter individual after coordinated search;

[0213] According to the fourth hyperparameter individual with the minimum fitness and the fourth hyperparameter individual with the maximum fitness, perform exploitation search on the first hyperparameter individual to obtain an exploitation search individual (i.e., the fifth hyperparameter individual), which can be:

[0214]

[0215] wherein, represents the p-th fourth hyperparameter individual after coordinated search in the t-th training process, is the fifth hyperparameter individual, represents the fourth hyperparameter individual with the maximum fitness, represents the fourth hyperparameter individual with the minimum fitness, and r6 represents the sixth random number between (0, 1).

[0216] Continue to judge whether the average fitness of the exploitation search individual (which can also be the loss value of the model) decreases. If so, accept this exploitation search and use each fifth hyperparameter individual after exploitation search as the hyperparameter individual to be output. Otherwise, directly use the original fourth hyperparameter individual after coordinated search as the hyperparameter individual to be output.

[0217] The above four searches adopt a greedy exploitation search strategy. Then, embodiments of the present invention can make the hyperparameter individual randomly move under the attraction of the individual with the optimal fitness and the repulsion of the individual with the worst fitness, thereby realizing position update, improving the diversity of the algorithm, and enabling the algorithm to perform global search.

[0218] It can be understood that the present invention can repeatedly execute the fast search strategy, slow search strategy, coordinated search strategy, and greedy exploitation search strategy until the iteration end condition is met. Then, according to the hyperparameter individual to be output after the last exploitation search, re-obtain the hyperparameter individual with the maximum fitness to be output, use it as the final target hyperparameter individual, and use each hyperparameter included in the target hyperparameter individual as the final hyperparameters of the label recognition model to obtain the label recognition model, and deploy the label recognition model on the unmanned aerial vehicle.

[0219] In a preferred embodiment, the electric tower pole recognition model of the present invention also adopts a training method and a hyperparameter search method similar to those of the tag recognition model. Then, the fast search strategy, the slow search strategy, the coordinated search strategy, and the greedy exploitation search strategy can be repeatedly executed until the iteration end condition is met. After that, among the hyperparameter individuals after the last exploitation search, the hyperparameter individual with the maximum fitness is re-obtained, which is used as the final hyperparameter individual. Each hyperparameter included in the final hyperparameter individual is used as the final hyperparameter of the electric tower pole recognition model to obtain the electric tower pole recognition model, and the electric tower pole recognition model is deployed on the unmanned aerial vehicle.

[0220] For steps S2 and S3, in a preferred embodiment, according to the relative position relationship between the unmanned aerial vehicle and the electric tower pole, and the position of the electric tower pole in the target inspection image, the coordinates of the radio frequency tag relative to the image center or the offset relative to the image boundary can be calculated.

[0221] Then, further based on the coordinates of the radio frequency tag relative to the image center or the offset relative to the image boundary, and information such as the current position of the unmanned aerial vehicle and the fixed position of the electric tower pole, the actual position of the radio frequency tag in the three-dimensional space is calculated.

[0222] After the unmanned aerial vehicle receives the actual position information of the radio frequency tag, it plans a flight path through its navigation system and controls the unmanned aerial vehicle to fly along this path to the position where the radio frequency tag is located. When the unmanned aerial vehicle arrives near the radio frequency tag, the RFID reader carried by it will activate and read the operation information stored in the radio frequency tag. Schematically, the operation information stored in the radio frequency tag includes the status of the electric tower pole, maintenance records, fault history, etc.

[0223] The unmanned aerial vehicle transmits the operation information of the radio frequency tag read to the ground station or the cloud server for further processing and analysis. Then, based on this information, an inspection data report of the electric tower pole is generated, and the inspection data report can include the health status, potential risks, maintenance suggestions, etc. of the electric tower pole.

[0224] Based on the fact that the unmanned aerial vehicle can quickly and accurately fly to the position of the radio frequency tag and read the operation information therein, the inspection time is shortened. Compared with the traditional manual inspection method, the unmanned aerial vehicle inspection does not require dispatching a large number of personnel to the site, reducing the labor cost and safety risks. Moreover, the present invention can accurately read the information in the radio frequency tag through the RFID reader carried by the unmanned aerial vehicle, avoiding the errors that may occur in manual reading.

[0225] Then, through the combination of the unmanned aerial vehicle and the radio frequency tag technology, the present invention can realize the real-time monitoring and early warning of the electric tower pole, and timely discover and handle potential safety hazards.

[0226] Such as Figure 2As shown in the figure, based on the above embodiments of various tower pole inspection methods based on tag recognition, the present invention correspondingly provides an apparatus embodiment;

[0227] An embodiment of the present invention provides a tower pole inspection apparatus based on tag recognition, including: a target inspection image generation module, a tag position determination module, and an inspection data generation module;

[0228] The target inspection image generation module is configured to repeatedly execute an inspection image recognition operation until, when it is determined that the radio frequency tag recognition result indicates that there is a radio frequency tag in the inspection image, the inspection image with the radio frequency tag is used as the target inspection image; wherein, the radio frequency tag is used to display the operation information on the tower pole;

[0229] The tag position determination module is configured to generate the tag position of the radio frequency tag in the target inspection image according to the current position of the unmanned aerial vehicle, the fixed position of the tower pole, and the relative position of the unmanned aerial vehicle relative to the tower pole;

[0230] The inspection data generation module is configured to, after obtaining the actual position of the radio frequency tag according to the tag position, control the unmanned aerial vehicle to fly to the actual position of the radio frequency tag, and then control the unmanned aerial vehicle to read the operation information of the radio frequency tag, and generate the inspection data of the tower pole according to the read operation information;

[0231] Wherein, the inspection image recognition operation includes:

[0232] Obtaining an inspection image captured by the unmanned aerial vehicle when flying around the tower pole;

[0233] Inputting the inspection image into a preset tag recognition model, so that the tag recognition model extracts the image features in the inspection image, and outputs a radio frequency tag recognition result according to the image features; wherein, the radio frequency tag recognition result includes: there is a radio frequency tag in the inspection image, or there is no radio frequency tag in the inspection image;

[0234] When it is determined that the radio frequency tag recognition result indicates that there is no radio frequency tag in the inspection image, the next inspection image recognition operation is executed.

[0235] It should be noted that the device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement this without creative efforts.

[0236] Those skilled in the art can clearly understand that for the convenience and brevity, the specific working process of the device described above can refer to the corresponding process in the foregoing method embodiments, and will not be elaborated herein.

[0237] The above is the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A method for inspecting electric tower poles based on tag recognition, characterized in that: include: Repeat the inspection image recognition operation until it is determined that the inspection image has an RFID tag in the RFID tag recognition result, and use the inspection image with the RFID tag as the target inspection image; wherein the RFID tag is used to display the operation information on the power tower pole; Generate a tag position of the radio frequency tag in the target inspection image according to the current position of the drone, the fixed position of the power tower, and the relative position of the drone relative to the power tower; After the actual position of the radio frequency tag is obtained according to the tag position, the drone is controlled to fly to the actual position of the radio frequency tag, and then the drone is controlled to read the operation information of the radio frequency tag, and the inspection data of the electric tower is generated according to the read operation information; The inspection image recognition operation includes: Get inspection images captured by a drone as it flies around a power tower; The inspection image is input into a preset tag recognition model so that the tag recognition model extracts image features in the inspection image and outputs a radio frequency tag recognition result according to the image features; wherein the radio frequency tag recognition result includes: the existence of a radio frequency tag in the inspection image, or the absence of a radio frequency tag in the inspection image; When it is determined that the radio frequency tag recognition result is that there is no radio frequency tag in the inspection image, the next inspection image recognition operation is performed.

2. The method for inspecting electric tower poles based on tag recognition according to claim 1, characterized in that: Before repeatedly executing the inspection image recognition operation, it also includes: Repeat the regional image recognition operation until it is determined that the power tower pole recognition result indicates that the regional image contains power tower poles, and use the regional image containing the power tower poles as the target regional image; Generate the position of the electric tower in the target area image according to the current position of the UAV and the relative position of the UAV to the electric tower; According to the position in the target area image, the fixed position of the electric tower pole is obtained; Control the drone to fly to a fixed position of the power tower, and then control the drone to fly around the power tower and capture inspection images; The regional image recognition operation includes: Obtain regional images collected by drones in the tower inspection area; The regional image is input into a preset electric tower pole recognition model, so that the electric tower pole recognition model extracts global features and local features of the regional image, and outputs an electric tower pole recognition result according to the global features and the local features; wherein the electric tower pole recognition result includes: there is an electric tower pole in the regional image, or there is no electric tower pole in the regional image; When it is determined that the power tower recognition result is that there is no power tower in the regional image, the next regional image recognition operation is performed.

3. The method for inspecting electric tower poles based on tag recognition according to claim 2, characterized in that: Before obtaining the regional images collected by the drone in the tower inspection area, it also includes: Get the scanning width of the drone and the width of the inspection area of ​​the power tower; According to the scanning width, the electric tower inspection area is vertically divided to generate a plurality of sub-inspection areas; Connect the vertical center lines of all sub-inspection areas to generate a first inspection route; Determine whether the starting position of the drone is within the inspection area of ​​the electric tower; If not, the first inspection route is used as the target inspection route of the drone; If so, the sub-inspection area where the starting position of the drone is located is taken as the target sub-inspection area, and the end of the vertical center line of the target sub-inspection area that is closest to the starting position of the drone is taken as the inspection starting point; the number of sub-inspection areas on the left and right sides of the target sub-inspection area is obtained; taking the inspection starting point as the starting point, the vertical center lines of all sub-inspection areas on the side with the least number of sub-inspection areas are connected to generate a second inspection route; the straight line between the end point of the second inspection route and the inspection starting point is taken as the return route; taking the inspection starting point as the starting point, and taking the vertical center line corresponding to the target sub-inspection area as the starting segment, the vertical center lines of all sub-inspection areas on the side with a larger number of sub-inspection areas are connected to obtain a third inspection route; based on the second inspection route, the return route and the third inspection route, the target inspection route of the drone is generated; The drone is controlled to fly on the target inspection route, and at the same time, the drone is controlled to collect regional images corresponding to the tower inspection area.

4. The method for inspecting electric tower poles based on tag recognition according to claim 3, characterized in that: The generation of the tag recognition model includes: Repeat the model training operation until it is determined that the current number of iterations is the same as the preset number of iterations, and use the hyperparameter individual to be output with the largest fitness as the target hyperparameter individual; wherein the target hyperparameter individual includes a plurality of selected hyperparameters; the hyperparameters are used to adjust the model performance of the label recognition model; Each selected hyperparameter is used as the final hyperparameter of the label recognition model, and then a trained label recognition model is generated; The model training operation includes: Get the current iteration number; When it is determined that the current number of iterations is less than the preset number of iterations, a number of inspection image samples and the actual radio frequency tag recognition results corresponding to each inspection image sample are used as input, and the predicted radio frequency tag recognition results of each inspection image sample are used as output, and the tag recognition model to be trained is iteratively trained according to a number of current hyperparameter individuals, and the loss value corresponding to each current hyperparameter individual in each iterative training is output; wherein, when the model training operation is first performed, a number of current hyperparameter individuals are randomly generated; The current hyperparameter individual with the smallest loss value is taken as the first hyperparameter individual; Randomly select each current hyperparameter individual except the first hyperparameter individual corresponding to a target number of iterations during iterative training and mark them as hyperparameter individuals to be processed, and mark each hyperparameter in each hyperparameter individual to be processed as a hyperparameter to be processed; According to the parameter value corresponding to the hyperparameter in the first hyperparameter individual, the parameter values ​​of the hyperparameters to be processed in other hyperparameter individuals, the hyperparameter mean corresponding to each hyperparameter individual to be processed, the parameter value of each hyperparameter to be processed, the preset upper limit value of each hyperparameter to be processed and the preset lower limit value of each hyperparameter to be processed, each hyperparameter to be processed in each hyperparameter individual to be processed is searched once to generate a number of updated parameter values ​​corresponding to each hyperparameter to be processed; wherein the other hyperparameter individuals are hyperparameter individuals randomly obtained in the target number of iterations and other than the hyperparameter individuals to be processed; Generate a plurality of second hyperparameter individuals according to a plurality of updated parameter values ​​corresponding to each of the hyperparameters to be processed; According to the fitness of the first hyperparameter individual, the fitness corresponding to each second hyperparameter individual, and the Euclidean distance between each second hyperparameter individual and the first hyperparameter individual, a secondary search is performed on each second hyperparameter individual to generate a number of updated parameter values ​​corresponding to each hyperparameter in each second hyperparameter individual; wherein the fitness is used to characterize the degree to which the hyperparameter individual reduces the loss value of the model; Generate a third hyperparameter individual according to a number of updated parameter values ​​corresponding to each hyperparameter in each second hyperparameter individual; According to the fitness of the first hyperparameter individual, the fitness corresponding to each third hyperparameter individual, and the Euclidean distance between each third hyperparameter individual and the first hyperparameter individual, each third hyperparameter individual is searched three times to generate a number of updated parameter values ​​corresponding to each hyperparameter in each third hyperparameter individual; Generate a fourth hyperparameter individual according to a number of updated parameter values ​​corresponding to each hyperparameter in each third hyperparameter individual; According to the fitness of each fourth hyperparameter individual and a number of random numbers, four searches are performed on the hyperparameters in each fourth hyperparameter individual to generate a number of updated parameter values ​​corresponding to each hyperparameter in each fourth hyperparameter individual; Generate a fifth hyperparameter individual according to a number of updated parameter values ​​corresponding to each hyperparameter in each fourth hyperparameter individual; According to the fitness corresponding to each fifth hyperparameter individual and the total number of the fifth hyperparameter individuals, the first fitness mean is obtained; according to the fitness corresponding to each fourth hyperparameter individual and the total number of the fourth hyperparameter individuals, the second fitness mean is obtained; When it is determined that the first fitness mean is greater than the second fitness mean, each fifth hyperparameter individual is used as the hyperparameter individual to be output; when it is determined that the first fitness mean is not greater than the second fitness mean, each fourth hyperparameter individual is used as the hyperparameter individual to be output; Adding the current number of iterations to the preset number of iterations increment obtains the updated number of iterations, and uses the updated number of iterations as the number of iterations for the next model training operation; Each hyperparameter individual to be output is used as each current hyperparameter individual when the model training operation is performed next time.

5. The method for inspecting electric tower poles based on tag recognition according to claim 4, characterized in that: The process of generating several updated parameter values ​​corresponding to each hyperparameter to be processed includes: For each to-be-processed hyperparameter in each to-be-processed hyperparameter individual, a search range control factor corresponding to each to-be-processed hyperparameter is generated according to the parameter value corresponding to the hyperparameter in the first hyperparameter individual and the parameter values ​​of the to-be-processed hyperparameters in other hyperparameter individuals; For each hyperparameter individual to be processed, according to the parameter value of each hyperparameter to be processed in the hyperparameter individual to be processed and the total number of hyperparameters to be processed in the hyperparameter individual to be processed, generate the hyperparameter mean corresponding to the hyperparameter individual to be processed; For each unprocessed hyperparameter in each unprocessed hyperparameter individual, a search step length of the unprocessed hyperparameter is generated according to the parameter value corresponding to the hyperparameter in the first hyperparameter individual, the parameter value of the unprocessed hyperparameter, the hyperparameter mean corresponding to the unprocessed hyperparameter individual, the preset upper limit value of the unprocessed hyperparameter, and the preset lower limit value of the unprocessed hyperparameter; According to the parameter values ​​of each hyperparameter in the first hyperparameter individual, the search range control factors corresponding to each hyperparameter to be processed, and the search steps of each hyperparameter to be processed, a search is performed on each hyperparameter to be processed in the hyperparameter individual to be processed to generate several updated parameter values ​​corresponding to each hyperparameter to be processed.

6. The method for inspecting electric tower poles based on tag recognition according to claim 5, characterized in that: The process of generating a plurality of updated parameter values ​​corresponding to each hyperparameter in each second hyperparameter individual includes: According to the preset fitness function, the fitness corresponding to the first hyperparameter individual and each second hyperparameter individual is calculated respectively; For each second hyperparameter individual, the regional importance of the second hyperparameter individual is generated according to the fitness corresponding to the second hyperparameter individual, the fitness corresponding to the first hyperparameter individual, and the minimum fitness among the second hyperparameter individuals; wherein the regional importance of the second hyperparameter individual is used to characterize the closeness between the second hyperparameter individual and the first hyperparameter individual; Perform a neighborhood search on each second hyperparameter individual, and then obtain a number of first neighborhood hyperparameter individuals within a preset Euclidean distance range, and use the first neighborhood hyperparameter individual with the greatest regional importance as the first target neighborhood hyperparameter individual; For each second hyperparameter individual, the influence factor corresponding to the second hyperparameter individual is calculated according to the fitness corresponding to the first hyperparameter individual and the Euclidean distance between the second hyperparameter individual and the first hyperparameter individual; wherein the influence factor corresponding to the second hyperparameter individual is used to characterize the influence degree of the first hyperparameter individual on the second hyperparameter individual; According to the regional importance of each second hyperparameter individual, the first target neighborhood hyperparameter individual and the influencing factor corresponding to each second hyperparameter individual, a secondary search is performed on the hyperparameters in each second hyperparameter individual to generate several updated parameter values ​​corresponding to each hyperparameter in each second hyperparameter individual.

7. The method for inspecting electric tower poles based on tag recognition according to claim 6, characterized in that: The process of generating a plurality of updated parameter values ​​corresponding to each hyperparameter in each third hyperparameter individual includes: According to the preset fitness function, the fitness corresponding to each third hyperparameter individual is calculated; For each third hyperparameter individual, the regional importance of the third hyperparameter individual is generated according to the fitness corresponding to the third hyperparameter individual, the fitness corresponding to the first hyperparameter individual, and the minimum fitness among the third hyperparameter individuals; wherein the regional importance of the third hyperparameter individual is used to characterize the closeness between the third hyperparameter individual and the first hyperparameter individual; Perform a neighborhood search on each third hyperparameter individual, and then obtain several second neighborhood hyperparameter individuals within a preset Euclidean distance range, and use the second neighborhood hyperparameter individual with the greatest regional importance as the second target neighborhood hyperparameter individual; use the second neighborhood hyperparameter individual that is closest to the second target neighborhood hyperparameter individual as the third target neighborhood hyperparameter individual; For each third hyperparameter individual, the influence factor corresponding to the third hyperparameter individual is calculated according to the fitness corresponding to the first hyperparameter individual and the Euclidean distance between the third hyperparameter individual and the first hyperparameter individual; wherein the influence factor corresponding to the third hyperparameter individual is used to characterize the influence degree of the first hyperparameter individual on the third hyperparameter individual; According to the regional importance of each third hyperparameter individual, the second target neighborhood hyperparameter individual, the third target neighborhood hyperparameter individual and the influencing factors corresponding to each third hyperparameter individual, three searches are performed on the hyperparameters in each third hyperparameter individual to generate several updated parameter values ​​corresponding to each hyperparameter in each third hyperparameter individual.

8. The method for inspecting electric tower poles based on tag recognition according to claim 7, characterized in that: The process of generating a plurality of updated parameter values ​​corresponding to each hyperparameter in each fourth hyperparameter individual includes: According to the preset fitness function, the fitness corresponding to each fourth hyperparameter individual is calculated; According to the fourth hyperparameter individual with the largest fitness, the fourth hyperparameter individual with the smallest fitness and a number of random numbers, four searches are performed on the hyperparameters in each fourth hyperparameter individual to generate a number of updated parameter values ​​corresponding to each hyperparameter in each fourth hyperparameter individual.

9. The method for inspecting electric tower poles based on tag recognition according to claim 2, characterized in that: The generation of the electric tower identification model includes: Taking several regional image samples and the actual electric tower recognition results corresponding to each regional image sample as input and the predicted electric tower recognition results of each regional image sample as output, the electric tower recognition model to be trained is iteratively trained until the model converges to generate a preset electric tower recognition model.

10. An electric tower inspection device based on tag recognition, characterized in that: include: Target inspection image generation module, label position determination module and inspection data generation module; The target inspection image generation module is used to repeatedly perform the inspection image recognition operation until the inspection image with the RFID tag is determined to have a RFID tag in the inspection image, and the inspection image with the RFID tag is used as the target inspection image; wherein the RFID tag is used to display the operation information on the electric tower pole; The tag position determination module is used to generate the tag position of the radio frequency tag in the target inspection image according to the current position of the drone, the fixed position of the power tower and the relative position of the drone relative to the power tower; The inspection data generating module is used to control the drone to fly to the actual position of the radio frequency tag after obtaining the actual position of the radio frequency tag according to the tag position, and then control the drone to read the operation information of the radio frequency tag, and generate the inspection data of the power tower according to the read operation information; The inspection image recognition operation includes: Get inspection images captured by a drone as it flies around a power tower; The inspection image is input into a preset tag recognition model so that the tag recognition model extracts image features in the inspection image and outputs a radio frequency tag recognition result according to the image features; wherein the radio frequency tag recognition result includes: the existence of a radio frequency tag in the inspection image, or the absence of a radio frequency tag in the inspection image; When it is determined that the radio frequency tag recognition result is that there is no radio frequency tag in the inspection image, the next inspection image recognition operation is performed.

Images