UAV autonomous inspection processing method and system based on distribution network overhead line equipment

Abstract

The present invention discloses a method and system for autonomous inspection and processing of drones based on distribution network overhead line equipment, belonging to the technical field of drone inspection. The method first obtains the laser point cloud inspection information set and inspection route record of the target drone, generates the target inspection point cloud element knowledge through the point cloud element recognition branch in the inspection scheduling processing network, and considers the line crossing information between drones, improving the inspection safety. The line optimization decision branch combines the target inspection point cloud element knowledge and the inspection route record to generate the target route planning suggestion features, comprehensively considering multiple factors to ensure flight safety, inspection efficiency, and equipment coverage. The route clustering processing branch clusters the target route planning suggestion features, and finally determines the drone inspection planning cluster through the inspection energy efficiency discrimination weight of the clustering result. Combining the ball query safety verification technology, the inspection efficiency is improved and safety accidents are reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of UAV inspection, and particularly relates to a method and system for autonomous inspection and processing of UAVs based on distribution network overhead line equipment. Background Art

[0002] UAVs have extensive and important applications in the field of inspection of distribution network overhead lines. With its own mobility and a variety of detection devices carried, such as high-definition cameras, lidar, thermal imagers, etc., it conducts a full-range inspection on equipment such as pole bodies, tower heads, tower bodies, tower bases, cross arms, transformers, switches, etc. in the distribution network overhead lines, as well as the line orientation, corridors and surrounding environments. Through the high-definition camera, the appearance condition of the equipment can be directly obtained to check for problems such as damage, corrosion, and loose connectors; the laser point cloud data collected by lidar can accurately construct a three-dimensional model of the equipment and the environment to analyze structural deformation and other situations; the thermal imager is used to detect the temperature distribution of equipment such as transformers to determine whether there is overheating abnormality. At the same time, when the UAV inspects the corridor, it can monitor the distance between the surrounding trees and the line, whether the buildings are illegal construction and their interference with the electromagnetic environment. However, in the actual application process of inspection of distribution network overhead lines, there are certain shortcomings in the traditional UAV route planning method, which is difficult to ensure inspection efficiency and safety. Summary of the Invention

[0003] The present invention provides a method and system for autonomous inspection and processing of UAVs based on distribution network overhead line equipment, which can solve or partially solve the technical problems involved in the above background art.

[0004] The present invention provides a method for autonomous inspection and processing of drones based on distribution network overhead line equipment, which is applied to an autonomous inspection and processing system of drones. The method includes: obtaining X laser point cloud inspection information sets and X inspection route records of X target drones, where each laser point cloud inspection information set in the X laser point cloud inspection information sets includes key point detection information and flight state monitoring information. The key point detection information is used to represent the information obtained after the laser point cloud scanning and collection of the distribution network overhead line equipment by the target drone within a specified inspection task cycle, and the flight state monitoring information is used to represent the monitoring result corresponding to the attitude change of the target drone within a specified inspection task cycle. X is greater than or equal to 1; using the X laser point cloud inspection information sets as the incoming information of the point cloud element recognition branch in the inspection scheduling processing network, and generating X target inspection point cloud element knowledge through the point cloud element recognition branch, where the inspection scheduling processing network is debugged based on the initial inspection route record of the drone sample, the first inspection route record sample, and the second inspection route record sample. The initial inspection route record is used to represent the line crossing information between the drone sample and Y collaborative drones corresponding to the drone sample, the first inspection route record sample is used to represent the line crossing information between the drone sample and Z collaborative drones determined from the Y collaborative drones, and the second inspection route record sample is used to represent the line crossing information corresponding to the conflict drone. Z is greater than or equal to 1 and less than or equal to Y; using the X target inspection point cloud element knowledge and the X inspection route records as the incoming information of the line optimization decision branch in the inspection scheduling processing network, and generating X target route planning suggestion features through the line optimization decision branch; clustering the X target route planning suggestion features based on the route clustering processing branch to generate Q clustering results, where Q is greater than or equal to 1; identifying the inspection energy efficiency discrimination weights of the Q clustering results, and determining a drone inspection planning cluster based on the inspection energy efficiency discrimination weights of the Q clustering results, where the drone inspection planning cluster includes at least one of the target drones.

[0005] The present invention provides an autonomous inspection and processing system for drones, including at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the above method.

[0006] The present invention provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0007] Traditional technologies are difficult to comprehensively obtain detailed information on the equipment of distribution network overhead lines and the flight status of drones, which leads to inaccurate judgment of equipment status and the inability to detect potential fault risks in a timely manner. During the collaborative inspection of drones, due to the failure to fully consider the line crossing information between drones, problems such as flight path conflicts are likely to occur, affecting the inspection efficiency and safety. The flight path planning lacks comprehensive consideration of multiple factors, such as flight safety, inspection efficiency, and equipment coverage, and cannot be balanced, resulting in poor inspection results. Moreover, there is no effective classification processing for flight path planning, and no effective basis can be provided for determining the final inspection cluster. The traditional method also lacks a method for determining the cluster in combination with safety verification, and cannot improve the inspection efficiency and avoid safety accidents.

[0008] Based on this, the present invention realizes the effective planning of the drone inspection flight path through a series of steps. First, the laser point cloud inspection information set and inspection flight path record of the target drone are obtained, including the key point detection information and flight status monitoring information, which provide the basis for the equipment status and drone flight status for subsequent processing. Through the point cloud element recognition branch in the inspection scheduling processing network, the target inspection point cloud element knowledge is generated. Thanks to the debugging of the network based on multiple inspection flight path records, the line crossing information between drones can be fully considered, improving the accuracy of the evaluation of the equipment of the distribution network overhead lines. The line optimization decision branch combines the target inspection point cloud element knowledge with the inspection flight path record to generate the target flight path planning suggestion features, comprehensively considering multiple factors to ensure flight safety, inspection efficiency, and equipment coverage. The flight path clustering processing branch clusters the target flight path planning suggestion features to provide a basis for determining the drone inspection planning cluster. Finally, the drone inspection planning cluster is determined by identifying the inspection energy efficiency discrimination weight of the clustering result, and combined with the sphere query safety verification technology, the inspection efficiency is improved and safety accidents are reduced. Brief Description of the Drawings

[0009] Figure 1 It is a flowchart of a method for autonomous inspection of drones based on the equipment of distribution network overhead lines provided by the present invention.

[0010] Figure 2 It is a schematic structural diagram of a system for autonomous inspection of drones provided by the present invention. Detailed Embodiments

[0011] Next, the technical solutions in the present invention will be clearly and completely described in conjunction with the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without making creative efforts based on the embodiments in the present invention belong to the protection scope of the present invention.

[0012] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, rather than to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually of the same type, and the number of objects is not limited. For example, the first object can be one or multiple. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates an "or" relationship between the associated objects before and after.

[0013] Figure 1 Disclosed is a method for autonomous inspection and processing of an unmanned aerial vehicle (UAV) based on distribution network overhead line equipment, which is applied to an autonomous inspection and processing system of a UAV. The method includes the following steps 110 to 150.

[0014] Step 110: Obtain X laser point cloud inspection information sets and X inspection route records of X target UAVs.

[0015] In the present invention, each laser point cloud inspection information set in the X laser point cloud inspection information sets includes key point detection information and flight state monitoring information. The key point detection information is used to represent the information obtained after laser point cloud scanning and acquisition of distribution network overhead line equipment by the target UAV within a specified inspection task cycle, and the flight state monitoring information is used to represent the monitoring result corresponding to the attitude change of the target UAV within a specified inspection task cycle, where X is greater than or equal to 1.

[0016] Furthermore, the objects of laser point cloud scanning and acquisition processing for distribution network overhead line equipment include, but are not limited to: pole tower body, tower head, tower body, tower base, alignment, corridor, cross arm, transformer, switch, and surrounding environment.

[0017] In the present invention, for each target UAV, there is a corresponding laser point cloud inspection information set. Here, X represents the number of target UAVs, and X is greater than or equal to 1. Each laser point cloud inspection information set contains two important pieces of information, namely key point detection information and flight state monitoring information.

[0018] The key point detection information is the information obtained after the target UAV performs laser point cloud scanning and acquisition on the distribution network overhead line equipment during the specified inspection task cycle. The objects of laser point cloud scanning and acquisition processing for distribution network overhead line equipment are very extensive, such as the pole tower body, tower head, tower body, tower base, orientation, corridor, cross arm, transformer, switch, surrounding environment, etc. For example, when performing laser point cloud scanning and acquisition on the pole tower body, coordinate information, shape information, etc. of each part of the pole tower body may be obtained. For example, if the pole tower body is in the shape of a cylinder, through laser point cloud scanning and acquisition, the coordinates of a large number of points on the surface of the cylinder can be obtained, so as to determine key dimension information such as its height and diameter. These information are very important for judging the structural integrity of the pole tower body.

[0019] The flight state monitoring information is the monitoring result based on the attitude changes of the target UAV during the specified inspection task cycle. For example, the attitude of the UAV includes its pitch angle, roll angle, and yaw angle. During the inspection process, if the UAV encounters airflow interference, its attitude will change. For example, at a certain moment, the pitch angle of the UAV suddenly changes from 0° to 5°, and this change will be recorded by the flight state monitoring system. By monitoring these attitude changes, the flight stability of the UAV can be analyzed to determine whether there are abnormal flight conditions, such as whether it is affected by strong wind interference or whether its own flight control system has failed, etc.

[0020] At the same time, it is also necessary to obtain the inspection route records of each target UAV. These inspection route records are of great significance for subsequent route planning and optimization decisions. For example, the inspection route records may include information such as the take-off point of the UAV, the coordinates of each inspection point passed by, flight altitude, flight speed, and the final landing point.

[0021] Step 120: Use the X laser point cloud inspection information sets as the incoming information of the point cloud element recognition branch in the inspection scheduling processing network, and generate X target inspection point cloud element knowledge through the point cloud element recognition branch.

[0022] In the present invention, the inspection scheduling processing network is debugged based on the initial inspection route record of the UAV sample, the first inspection route record sample, and the second inspection route record sample. The initial inspection route record is used to characterize the line crossing information between the UAV sample and the Y cooperative UAVs corresponding to the UAV sample. The first inspection route record sample is used to characterize the line crossing information between the UAV sample and the Z cooperative UAVs determined from the Y cooperative UAVs. The second inspection route record sample is used to characterize the line crossing information corresponding to the conflict UAV, where Z is greater than or equal to 1 and less than or equal to Y.

[0023] In the present invention, the initial inspection route record is used to characterize the line crossing information between the UAV sample and the Y cooperative UAVs corresponding to the UAV sample. For example, there is a UAV sample, and the number of cooperative UAVs Y = 3. The initial inspection route record will record the line crossing situation between this UAV sample and these 3 cooperative UAVs during flight. If the route of the UAV sample is line A and the route of cooperative UAV 1 is line B, when line A and line B intersect, the initial inspection route record will record information such as the coordinates of the intersection point and the flight altitude at the time of crossing.

[0024] The first inspection route record sample is used to characterize the line crossing information between the UAV sample and the Z cooperative UAVs determined from the Y cooperative UAVs, where Z is greater than or equal to 1 and less than or equal to Y. For example, Y = 5 and Z = 3. This means that 3 cooperative UAVs are selected from 5 cooperative UAVs to determine the line crossing information with the UAV sample. For example, if the selected cooperative UAVs are 1, 3, and 5, the first inspection route record sample will detail the line crossing situation between the UAV sample and these 3 cooperative UAVs, such as the crossing frequency, relative speed at the time of crossing, etc.

[0025] The second inspection route record sample is used to characterize the line crossing information corresponding to the UAV sample and the conflict UAV. The conflict UAV may be a UAV that has a potential conflict with the UAV sample due to unreasonable route planning or unexpected situations. For example, when a UAV sample is performing inspection according to a normal route, suddenly an unplanned UAV enters its inspection area, and this UAV is regarded as a conflict UAV. The second inspection route record sample will record the line crossing situation between them, including information such as the crossing angle and distance.

[0026] Furthermore, the working principle of the point cloud feature recognition branch is as follows: When X laser point cloud inspection information sets are input into the point cloud feature recognition branch of the inspection scheduling processing network, the point cloud feature recognition branch will process this information. The point cloud feature recognition branch may adopt a convolutional neural network (CNN) model in deep learning algorithms. For example, in the CNN model, the input laser point cloud inspection information sets are processed through convolutional layers, pooling layers, etc. For example, the convolutional layer uses a 3*3 convolutional kernel with a stride of 1, and through convolutional operations, the features in the laser point cloud inspection information sets can be extracted. The pooling layer uses a 2*2 max pooling operation, which can reduce the data dimension and computational amount. After multiple layers of convolution and pooling operations, finally, X target inspection point cloud feature knowledges are generated through the fully connected layer. These target inspection point cloud feature knowledges can include more accurate feature descriptions of the distribution network overhead line equipment, such as the damage degree assessment of the pole body and the working state analysis of the transformer, etc., knowledge based on point cloud data.

[0027] Step 130: Use the X target inspection point cloud element knowledge and the X inspection route records as the incoming information for the route optimization decision branch in the inspection scheduling processing network, and generate X target route planning suggestion features through the route optimization decision branch.

[0028] In this step, the route optimization decision branch receives the X target inspection point cloud element knowledge and the X inspection route records as input information. These input information contain the analysis results of the overhead distribution line equipment in the previous steps and the flight history routes of the drones. The route optimization decision branch may use a genetic algorithm to generate the route planning suggestion features. In the genetic algorithm, first, the target inspection point cloud element knowledge and the inspection route records are encoded. For example, information such as the flight altitude, speed, and passing point coordinates of the drone are encoded into binary strings. For example, the target inspection point cloud element knowledge includes the damage condition of the pole body (represented by a value between 0 and 1, 0 means no damage, 1 means severe damage) and the working state of the transformer (represented by a value between -1 and 1, -1 means failure, 1 means normal operation), as well as information such as the flight altitude of 100 meters and speed of 5 m / s in the inspection route record. After encoding this information, an initial population is formed.

[0029] Then, according to the set fitness function, evaluate the quality of each individual (the encoded route plan). The fitness function may comprehensively consider factors such as the flight safety of the drone, the inspection efficiency, and the coverage of the overhead distribution line equipment. For example, the fitness function can be expressed as: F = a * S + b * E + c * C, where F is the fitness value, S is the flight safety index (value range 0 - 1, 1 means the safest), E is the inspection efficiency index (such as the number of equipment inspected per unit time), C is the coverage of the overhead distribution line equipment (value range 0 - 1, 1 means full coverage), and a, b, c are weight coefficients set according to actual needs.

[0030] Through selection, crossover, and mutation operations, continuously evolve the population, and finally obtain X target route planning suggestion features. For example, after multiple generations of evolution, the obtained target route planning suggestion features may include new flight altitude, speed adjustment suggestions, and optimization of passing points. For example, adjust the flight altitude from 100 meters to 80 meters, the speed from 5 m / s to 6 m / s, and add a new passing point to more comprehensively inspect a certain key equipment, etc.

[0031] Step 140: Cluster the X target route planning suggestion features based on the route clustering processing branch to generate Q clustering results.

[0032] In the present invention, Q is greater than or equal to 1. Among them, the principle of grouping in the route clustering processing branch is as follows: The route clustering processing branch performs a grouping operation on X target route planning suggestion features. Here, clustering algorithms such as the K-Means clustering algorithm may be used. For example, X = 10, that is, there are 10 target route planning suggestion features.

[0033] In the K-Means clustering algorithm, first, the number of clusters Q needs to be determined. If Q = 3, the algorithm will randomly select 3 initial cluster centers. For example, 3 are randomly selected from 10 target route planning suggestion features as the initial cluster centers. Then, the distance from each target route planning suggestion feature to these 3 cluster centers is calculated. The distance can be calculated using the Euclidean distance formula. For each dimension in the target route planning suggestion feature (such as flight altitude, speed, waypoint coordinates, etc.), calculate the sum of the squares of the differences between it and the corresponding dimension of the cluster center, and then take the square root to obtain the distance.

[0034] According to the distance, each target route planning suggestion feature is assigned to the class to which the nearest cluster center belongs. Then, the cluster center of each class is recalculated, that is, the average value of all target route planning suggestion features in each class in each dimension is calculated as the new cluster center. Repeat this process until the cluster center no longer changes significantly or reaches the preset number of iterations. Finally, Q = 3 grouping results are obtained, and each grouping result contains several target route planning suggestion features with similar characteristics.

[0035] Step 150: Identify the inspection energy efficiency discrimination weights of the Q grouping results, and determine the UAV inspection planning cluster based on the inspection energy efficiency discrimination weights of the Q grouping results.

[0036] In the present invention, at least one of the target UAVs is included in the UAV inspection planning cluster. Further, in the process of determining the UAV inspection planning cluster, the spherical query safety verification technology is also combined, so that the route planning situation of each target UAV in the UAV inspection planning cluster that needs to perform autonomous route planning can be accurately and safely determined, improving the inspection efficiency and reducing safety accidents during the inspection process.

[0037] Specifically, step 150 involves two key steps: identifying the inspection energy efficiency discrimination weights and determining the UAV inspection planning cluster.

[0038] Identification of the inspection energy efficiency discrimination weight: For Q clustering results, it is necessary to identify their inspection energy efficiency discrimination weights. The determination of the inspection energy efficiency discrimination weight may involve multiple factors. For example, if the target flight route planning suggestion feature in a clustering result can cover more key overhead distribution network line equipment, its inspection energy efficiency discrimination weight may be higher. For example, the target flight route planning suggestion feature in clustering result 1 can cover 80% of the key equipment, while the target flight route planning suggestion feature in clustering result 2 can only cover 60% of the key equipment. Then, the inspection energy efficiency discrimination weight of clustering result 1 may be higher than that of clustering result 2.

[0039] In addition, the flight safety and efficiency of the drone when executing the flight route planning suggestion feature in this clustering result may also be considered. For example, although the number of key equipment covered by the flight route planning suggestion feature in clustering result 3 is not the largest, its flight safety is the highest and the flight efficiency is also relatively high (such as the shortest flight path and the most stable flight speed, etc.). Then, its inspection energy efficiency discrimination weight will also be relatively high.

[0040] Determination of the drone inspection planning cluster: The ball query safety verification technology is also combined in the process of determining the drone inspection planning cluster. The ball query safety verification technology can be understood as constructing a virtual spherical space centered on the drone. For example, if the safety radius of the drone is R = 10 meters, within this radius range, if there are other obstacles or drones, there may be safety risks.

[0041] According to the inspection energy efficiency discrimination weights of the Q clustering results, the target drones in the clustering results with higher weights are included in the drone inspection planning cluster. For example, if the inspection energy efficiency discrimination weight of clustering result 1 is the highest and it includes target drones 1, 2, and 3, then target drones 1, 2, and 3 will be determined as members of the drone inspection planning cluster. In this way, the line planning situation of each target drone in the drone inspection planning cluster that needs to perform autonomous flight route planning can be accurately and safely determined, improving the inspection efficiency and reducing safety accidents during the inspection process.

[0042] It can be understood that the further introduction of the above-mentioned target inspection point cloud element knowledge and target flight route planning suggestion features is as follows.

[0043] I. Quantitative features of the target inspection point cloud element knowledge

[0044] Quantitative features related to the structure of overhead distribution network line equipment

[0045] 1) Quantification of the tower structure feature

[0046] Height deviation quantification: For the pole tower body, the actual height of the pole tower can be calculated from the point cloud data collected by laser point cloud scanning. For example, in an ideal state, the designed height of the pole tower is (H_d = 50) meters, and the actual height calculated through point cloud analysis is (H_a). Then the height deviation (DeltaH = |H_d - H_a|). For example, if (H_a = 49.5) meters, then (DeltaH = |50 - 49.5| = 0.5) meters. This quantified value can be used as part of the knowledge of the target inspection point cloud elements to determine whether there is settlement or foundation damage in the pole tower, etc.

[0047] Tower body inclination quantification: A coordinate system is established with the center of the pole tower bottom as the origin, and the offset of the top of the pole tower relative to the bottom center in the horizontal direction is measured. For example, the radius of the pole tower bottom is (r = 2) meters, the offset of the top of the pole tower in the (x) direction is measured as (x_o = 0.2) meters, and the offset in the (y) direction is (y_o = 0.1) meters. Then the tower body inclination (theta = arctan(sqrt{x_o^2 + y_o^2} / h)), where (h) is the height of the pole tower. If (h = 50) meters, (theta = arctan(sqrt{0.2^2 + 0.1^2} / 50) ≈ 0.0025) radians. This quantified inclination value is an important indicator of the structural stability of the pole tower in the knowledge of the target inspection point cloud elements.

[0048] 2) Cross-arm deformation quantification

[0049] Cross-arm length change quantification: The designed length of the cross-arm is (L_d = 5) meters, and the actual length of the cross-arm obtained through laser point cloud analysis is (L_a). The cross-arm length change amount (DeltaL = |L_d - L_a|). For example, when (L_a = 4.9) meters, (DeltaL = |5 - 4.9| = 0.1) meters. This quantified value helps to determine whether the cross-arm has undergone tensile or compressive deformation.

[0050] Cross-arm curvature quantification: The cross-arm is regarded as a straight line segment (ideal state), and the actual shape curve of the cross-arm is fitted through point cloud data. Calculate the maximum vertical distance (d) between this curve and the ideal straight line, and based on the cross-arm length (L), the curvature (k = d / L). If (d = 0.1) meters and (L = 5) meters, then (k = 0.1 / 5 = 0.02). This quantified curvature is a key quantified feature of the cross-arm structural integrity in the knowledge of the target inspection point cloud elements.

[0051] Quantified features related to the equipment operation status

[0052] 1) Transformer oil temperature quantification

[0053] When the transformer is operating normally, the oil temperature has a reasonable designed range, for example, from (T_{min} = 50^{\circ}C) to (T_{max} = 80^{\circ}C). The actual oil temperature (T_a) of the transformer is obtained through laser point cloud scanning in cooperation with temperature sensors (for example, the point cloud data can assist in positioning the data obtained by the temperature sensors). If (T_a = 70^{\circ}C), this oil temperature value is a quantitative feature in the target inspection point cloud element knowledge and can be used to judge whether the operation state of the transformer is normal.

[0054] 2) Quantification of switch opening and closing states

[0055] Define the opening and closing state of the switch as a binary variable. (S = 0) indicates that the switch is closed, and (S = 1) indicates that the switch is open. The actual opening and closing state of the switch is determined through laser point cloud data combined with image recognition technology (for example, the point cloud data can assist in positioning the switch and performing image recognition). This (S) value, as part of the target inspection point cloud element knowledge, directly reflects the operation state of the switch.

[0056] Quantitative features related to the surrounding environment

[0057] 1) Quantification of the safety distance between surrounding vegetation and the line

[0058] For the overhead distribution network line, it is stipulated that the safety distance between the surrounding vegetation and the line is (D_s = 3) meters. The actual distance (D_a) between the vegetation and the line is obtained through laser point cloud scanning. The distance deviation (\(\Delta D=|D_s - D_a|\)). For example, when (D_a = 2.5) meters, (\(\Delta D=|3 - 2.5| = 0.5\) meters). This quantified distance deviation is an important indicator in the target inspection point cloud element knowledge regarding the impact of the surrounding environment on the line safety.

[0059] 2) Quantification of the distance between surrounding buildings and the pole tower

[0060] For example, it is stipulated that the safety distance between the surrounding buildings and the pole tower is (D_{b - t}=10) meters. The actual distance (D_{a - b - t}) is calculated through laser point cloud data. The distance deviation (\(\Delta D_{b - t}=|D_{b - t}-D_{a - b - t}|\)). For example, when (D_{a - b - t}=8) meters, (\(\Delta D_{b - t}=|10 - 8| = 2\) meters). This quantified value can reflect the potential impact of the surrounding buildings on the stability of the pole tower and is part of the target inspection point cloud element knowledge.

[0061] II. Quantitative features of the target flight path planning suggestions

[0062] Quantitative features related to the flight path

[0063] 1) Quantification of flight distance

[0064] The flight distance in the target route planning suggestion feature is an important quantification index. For example, for a drone flying from the takeoff point (P_1(x_1, y_1, z_1)) to the landing point (P_2(x_2, y_2, z_2)), the flight distance (L = sqrt{(x_2 - x_1)^2+(y_2 - y_1)^2+(z_2 - z_1)^2}). For example, when (x_1 = 0, y_1 = 0, z_1 = 0), (x_2 = 100, y_2 = 100, z_2 = 50), then: (L = sqrt{(100 - 0)^2+(100 - 0)^2+(50 - 0)^2}=sqrt{10000 + 10000 + 2500}=sqrt{22500}=150) meters. This quantified value of the flight distance directly affects the inspection efficiency and energy consumption.

[0065] 2) Quantification of flight trajectory smoothness

[0066] Regard the flight trajectory as a curve composed of a series of discrete points (P_i(x_i, y_i, z_i)) ((i = 1, 2, cdots, n)). Calculate the vector between adjacent two points (vec{v}_i = overrightarrow{P_{i}P_{i+1}}=(x_{i+1}-x_i, y_{i+1}-y_i, z_{i+1}-z_i)).

[0067] Then calculate the angle between adjacent vectors: (theta_i = arccos(frac{vec{v}_icdotvec{v}_{i+1}}{|vec{v}_i|*|vec{v}_{i+1}|})). The flight trajectory smoothness (S = frac{1}{n - 1}sum_{i = 1}^{n - 1}theta_i). A smaller (S) value indicates a smoother flight trajectory. For example, (S = 0.1) radian means the trajectory is relatively smooth, which helps to reduce the energy consumption during the drone flight and improve the flight stability.

[0068] Quantification features related to flight parameters

[0069] 1) Quantification of flight speed

[0070] The target route planning suggestion features include the flight speed (v) of the UAV, with the unit of meters per second. For example, (v = 5) m / s means the UAV flies at this speed. Different flight tasks and environments may require different flight speeds. In the inspection task, it is necessary to comprehensively consider the inspection accuracy and efficiency to determine the appropriate flight speed. If a more detailed inspection is needed (such as a close-up inspection of key equipment), the flight speed may need to be reduced; while in the long-distance line inspection, the flight speed can be appropriately increased.

[0071] 2) Quantification of flight altitude

[0072] The flight altitude (h) is also a key quantified feature, with the unit of meters. For example, (h = 100) m. The selection of the flight altitude needs to consider multiple factors, such as avoiding obstacles and obtaining a better inspection perspective. For the inspection of distribution network overhead lines, it may be necessary to determine the appropriate flight altitude according to the height of the line, the surrounding environment (such as the height of buildings and vegetation), etc.

[0073] Quantified features related to inspection coverage

[0074] 1) Quantification of equipment coverage ratio

[0075] Define the distribution network overhead line equipment set as (E = {e_1, e_2, cdots, e_m}), and the equipment set covered by the target route planning suggestion features as (E_c = {e_{c1}, e_{c2}, cdots, e_{ck}}). The equipment coverage ratio (C = frac{k}{m}). For example, if there are a total of (m = 10) devices, and the target route planning suggestion features can cover (k = 8) devices, then (C = frac{8}{10} = 0.8). This quantified value reflects the comprehensiveness of the route planning for equipment inspection.

[0076] 2) Quantification of the coverage of key equipment

[0077] For the key equipment set (E_{key} = {e_{k1}, e_{k2}, cdots, e_{kn}}) ((nleqm)), define the key equipment coverage ratio (C_{key} = frac{n_{covered}}{n}), where (n_{covered}) is the number of key equipment covered by the target route planning suggestion features. For example, if there are (n = 5) key devices, and the target route planning suggestion features can cover (n_{covered} = 4) key devices, then (C_{key} = frac{4}{5} = 0.8). This quantified feature focuses on the inspection coverage of key equipment because the normal operation of key equipment is crucial to the entire distribution network overhead line system.

[0078] Applying the above technical solution, the laser point cloud inspection information set obtained in step 110 contains rich information. Among them, the key point detection information covers detailed data of many components of the distribution network overhead line equipment, such as the coordinates and shapes of the pole tower body, etc. This helps to accurately judge the structural integrity of the equipment. For example, it can timely detect minor deformations or damages of the pole tower and avoid potential line fault risks. The flight state monitoring information can monitor the attitude changes of the UAV in real time. For example, by recording the changes in the pitch angle, roll angle and yaw angle, it can timely detect abnormal situations during the UAV flight, whether it is external airflow interference or its own control system failure, to ensure the smooth progress of the inspection task. The inspection route record provides basic data for subsequent planning and decision-making, clarifying key parameters such as the flight path and speed of the UAV.

[0079] In step 120, the point cloud element recognition branch in the inspection scheduling processing network generates target inspection point cloud element knowledge. This branch is debugged based on multiple inspection route records and can fully consider the line crossing information between UAVs, whether it is the situation of cooperative UAVs or conflicting UAVs. A convolutional neural network (CNN) model is used for processing. For example, by using specific convolutional kernels and pooling operations, features can be effectively extracted. The finally generated target inspection point cloud element knowledge can more accurately evaluate the distribution network overhead line equipment, such as accurately evaluating the damage degree of the pole tower and the working state of the transformer, improving the accuracy of the inspection.

[0080] In step 130, the line optimization decision branch combines the target inspection point cloud element knowledge and the inspection route record, and uses the genetic algorithm to generate target route planning suggestion features. This method can comprehensively consider various factors such as flight safety, inspection efficiency and equipment coverage. Through fitness function evaluation and population evolution operations, an optimized route planning suggestion is obtained, such as adjusting the flight height, speed and passing points, which can not only improve the inspection efficiency but also ensure the comprehensiveness of the inspection.

[0081] The route clustering processing branch in step 140 clusters the target route planning suggestion features. Using a clustering algorithm (such as the K-Means clustering algorithm) can group similar route planning features into one category, making subsequent processing more efficient. This helps to find the commonalities and differences in different types of route planning and provides a basis for determining the final UAV inspection planning cluster.

[0082] Finally, step 150 determines the UAV inspection planning cluster by identifying the inspection energy efficiency discrimination weights of the clustering results and combining the sphere query security verification technology. The inspection energy efficiency discrimination weights determined by considering multiple factors can reasonably screen out the efficient route planning clustering results. Then, combined with the sphere query security verification technology, a safety space (such as a sphere space with a radius of 10 meters) is constructed with the UAV as the center to avoid interference from obstacles and other UAVs, accurately determine the UAV inspection planning cluster, improve the inspection efficiency and reduce safety accidents at the same time.

[0083] In some alternative embodiments, the X laser point cloud inspection information sets include X key point detection information and X flight state monitoring information. The point cloud element recognition branch includes a key point quantization recognition module, an attitude point cloud recognition module, and a fully connected module. Then, the step 120 of using the X laser point cloud inspection information sets as the incoming information of the point cloud element recognition branch in the inspection scheduling processing network and generating X target inspection point cloud element knowledge through the point cloud element recognition branch includes: using the X key point detection information as the incoming information of the key point quantization recognition module and generating X key point quantization knowledge codes through the key point quantization recognition module; using the X flight state monitoring information as the incoming information of the attitude point cloud recognition module and generating X attitude point cloud knowledge codes through the point cloud element recognition branch; and using the X key point quantization knowledge codes and the X attitude point cloud knowledge codes as the incoming information of the fully connected module and generating X target inspection point cloud element knowledge through the incoming information of the fully connected module.

[0084] In addition, in the following Preferred Embodiment 1, the key point quantization and recognition module includes a vector space conversion unit and a knowledge collision unit; based on this, taking the X key point detection information as the incoming information of the key point quantization and recognition module, X key point quantization knowledge encodings are generated by the key point quantization and recognition module, including: generating X local point cloud description semantic maps based on the key point image semantics of the X key point detection information and the X key point detection information; taking the X local point cloud description semantic maps as the incoming information of the vector space conversion unit, and generating X local point cloud space migration encodings based on the vector space conversion unit, where the vector space conversion unit is used to sum the corresponding feature members in each local point cloud description semantic map of the X local point cloud description semantic maps; taking the X local point cloud description semantic maps as the incoming information of the knowledge collision unit, and generating X local point cloud advanced space migration encodings through the knowledge collision unit, where the knowledge collision unit is used to multiply the corresponding feature members in each local point cloud description semantic map of the X local point cloud description semantic maps and sum the results of multiplying the corresponding feature members; integrating the X local point cloud space migration encodings based on the X local point cloud advanced space migration encodings to obtain X key point quantization knowledge encodings.

[0085] In addition, in the following Preferred Embodiment 2, the pose point cloud recognition module includes a residual connection unit and a point cloud reconstruction unit; taking the X flight state monitoring information as the incoming information of the pose point cloud recognition module, X pose point cloud knowledge encodings are generated through the point cloud element recognition branch, including: taking the X flight state monitoring information as the incoming information of the residual connection unit, and generating X*R pose monitoring data groups through the residual connection unit, where the pose monitoring data group is composed of the sensing monitoring data in at least one of the X flight state monitoring information, and R is an integer greater than 1; fusing the X*R pose monitoring data groups through a pose fusion algorithm to obtain X pose monitoring event sets; taking the X pose monitoring event sets as the incoming information of the point cloud reconstruction unit, and generating X pose point cloud knowledge encodings through the point cloud reconstruction unit.

[0086] In addition, in the following preferred Embodiment 3, the fully connected module includes a data weighting unit and an activation unit; the quantization knowledge encoding of the X key points and the pose point cloud knowledge encoding of the X pose points are used as the incoming information of the fully connected module, and X target inspection point cloud element knowledge is generated through the incoming information of the fully connected module, including: integrating each key point quantization knowledge encoding in the quantization knowledge encoding of the X key points with each pose point cloud knowledge encoding in the pose point cloud knowledge encoding of the X pose points to obtain X integrated knowledge encodings; using the X integrated knowledge encodings as the incoming information of the data weighting unit, and generating X multi-modal inspection knowledge encodings through the data weighting unit; using the X multi-modal inspection knowledge encodings as the incoming information of the activation unit, and generating X target inspection point cloud element knowledge through the activation unit.

[0087] It can be understood that the above replaceable embodiment and the corresponding three preferred embodiments focus on the processing of the laser point cloud inspection information set in the inspection of the overhead distribution network by the unmanned aerial vehicle. It mainly involves the processing process of the X laser point cloud inspection information sets in the point cloud element recognition branch of the inspection scheduling processing network to generate X target inspection point cloud element knowledge. Among them, the laser point cloud inspection information set includes X key point detection information and X flight state monitoring information, and the point cloud element recognition branch includes a key point quantization recognition module, a pose point cloud recognition module and a fully connected module.

[0088] 1. Key point quantization recognition module

[0089] 1.1. Generate a local point cloud description semantic map based on the key point detection information

[0090] In this module, first, X local point cloud description semantic maps are generated based on the key point image semantics of the X key point detection information and the X key point detection information itself. For example, when X = 3, for the first key point detection information, it may be the detection information about a certain part of the pole tower body. The image semantics of this part (such as the shape features, texture features, etc. represented in the image) are combined with the detection information (such as coordinates, relative positions with other components, etc.) to generate a local point cloud description semantic map. This semantic map can describe in detail the various feature relationships of this local point cloud.

[0091] 1.2. Working process of the vector space conversion unit

[0092] Next, take the X local point cloud description semantic maps as the incoming information of the vector space conversion unit. The vector space conversion unit is used to sum the corresponding feature members in each local point cloud description semantic map among the X local point cloud description semantic maps. For example, for the coordinate feature in the local point cloud description semantic map, if the coordinate feature vector in one local point cloud description semantic map is ((x_1, y_1, z_1)) and another is ((x_2, y_2, z_2)), after summation, we get ((x_1 + x_2, y_1 + y_2, z_1 + z_2)). Through such operations, each local point cloud description semantic map is processed, and finally X local point cloud based spatial migration encodings are generated. Taking X = 3 as an example, 3 local point cloud based spatial migration encodings will be obtained.

[0093] 1.3. Working process of the knowledge collision unit

[0094] Meanwhile, take the X local point cloud description semantic maps as the incoming information of the knowledge collision unit. The operation of the knowledge collision unit is to multiply the corresponding feature members in each local point cloud description semantic map among the X local point cloud description semantic maps, and then sum the results of multiplying the corresponding feature members. For example, for the coordinate feature vectors ((x_1, y_1, z_1)) and ((x_2, y_2, z_2)) mentioned above, the product is ((x_1x_2, y_1y_2, z_1z_2)), and then sum all the results after multiplication. In this way, X local point cloud advanced spatial migration encodings are generated.

[0095] 1.4. Integrate to obtain the key point quantization knowledge encoding

[0096] Finally, integrate the X local point cloud based spatial migration encodings and the X local point cloud advanced spatial migration encodings to obtain X key point quantization knowledge encodings. The integration method can be to combine the corresponding encodings according to certain rules, such as adding the corresponding elements in the spatial migration encoding and the advanced spatial migration encoding or adding them according to certain weights, etc.

[0097] 2. Pose point cloud recognition module

[0098] 2.1. The residual connection unit generates the pose monitoring data group

[0099] The residual connection unit in the attitude point cloud recognition module takes X flight state monitoring information as the incoming information. The flight state monitoring information includes information such as the pitch angle, roll angle, and yaw angle of the drone. For example, the flight state monitoring information is collected in a certain time series, and there is a set of data at each time point. The residual connection unit processes these data to generate X*R attitude monitoring data groups, where R is an integer greater than 1. For example, if X = 2 and R = 3, then 2*3 = 6 attitude monitoring data groups will be generated. Here, the attitude monitoring data group is composed of sensing monitoring data in at least one flight state monitoring information. The residual connection unit may adopt the residual connection idea in the Residual Network (ResNet), by adding the input information to the output information after some operations such as convolutional layers and activation functions, so as to retain more original information and improve the learning ability of the network.

[0100] 2.2. The attitude fusion algorithm fuses the attitude monitoring data groups

[0101] Then, through the attitude fusion algorithm, the X*R attitude monitoring data groups are fused to obtain X attitude monitoring event sets. The attitude fusion algorithm can adopt the Extended Kalman Filter (EKF) algorithm. For example, for the pitch angle data in the attitude monitoring data group, there may be certain errors and uncertainties in different attitude monitoring data groups. The EKF algorithm fuses the pitch angle data in multiple attitude monitoring data groups through prediction and update steps, using the state equation and measurement equation of the system. Taking the pitch angle data (theta_1), (theta_2), (theta_3) in three attitude monitoring data groups as an example, the EKF algorithm will calculate the fused pitch angle (theta) according to information such as their measurement covariance matrices, and finally obtain X attitude monitoring event sets.

[0102] 2.3. The point cloud reconstruction unit generates the attitude point cloud knowledge encoding

[0103] Finally, taking the X attitude monitoring event sets as the incoming information of the point cloud reconstruction unit, the point cloud reconstruction unit generates X attitude point cloud knowledge encodings. The point cloud reconstruction unit may adopt a deep learning-based point cloud generation algorithm. For example, taking the attitude monitoring event set as the input, after being processed by a multi-layer neural network, each layer of the neural network extracts and transforms the input data. For example, the first layer may be a convolutional layer, using a 3×3 convolutional kernel with a stride of 1 to perform a convolutional operation on the input attitude monitoring event set to extract features. After multiple such operations, finally X attitude point cloud knowledge encodings are generated.

[0104] 3. Fully connected module

[0105] 3.1. Integrate the knowledge encoding

[0106] The fully connected module first integrates each keypoint quantization knowledge encoding in the X keypoint quantization knowledge encodings with each pose point cloud knowledge encoding in the X pose point cloud knowledge encodings to obtain X integrated knowledge encodings. The integration method can be a concatenation operation. For example, if the keypoint quantization knowledge encoding is (K=(k_1, k_2, \cdots, k_n)) and the pose point cloud knowledge encoding is (A=(a_1, a_2, \cdots, a_m)), the integrated knowledge encoding can be (I=(k_1, k_2, \cdots, k_n, a_1, a_2, \cdots, a_m)).

[0107] 3.2. The data weighting unit generates multimodal inspection knowledge encodings

[0108] Next, the X integrated knowledge encodings are used as the input information of the data weighting unit, and the data weighting unit generates X multimodal inspection knowledge encodings. The data weighting unit may assign different weights according to different feature importances. For example, higher weights are assigned to some key features in the integrated knowledge encoding (such as the representation of the structural features of key parts of the tower in the keypoint quantization knowledge encoding), while lower weights are assigned to some relatively unimportant features. For example, if the integrated knowledge encoding is (I=(i_1, i_2, \cdots, i_p)) and the weight vector is (W=(w_1, w_2, \cdots, w_p)), then the multimodal inspection knowledge encoding is (M=(i_1w_1, i_2w_2, \cdots, i_pw_p)).

[0109] 3.3. The excitation unit generates target inspection point cloud element knowledge

[0110] Finally, the X multimodal inspection knowledge encodings are used as the input information of the excitation unit, and the excitation unit generates X target inspection point cloud element knowledge. The excitation unit can adopt an activation function, such as the ReLU (Rectified Linear Unit) function. For example, for each element (m_i) in the multimodal inspection knowledge encoding, after being processed by the ReLU function, (y_i = max(0, m_i)) is obtained, and the final X (y_i) form X target inspection point cloud element knowledge.

[0111] With such a design, first of all, in the key point quantization and recognition module, through the unique operations of the vector space conversion unit and the knowledge collision unit, the features in the key point detection information can be more comprehensively and deeply mined to generate accurate key point quantization knowledge codes. For example, through operations such as summation and multiplication on the local point cloud description semantic map, the structural relationships of various components in the overhead distribution line equipment can be accurately captured. In the attitude point cloud recognition module, the application of the residual connection unit and the attitude fusion algorithm enables the effective fusion of flight state monitoring information to generate accurate attitude point cloud knowledge codes. For example, the residual connection unit retains more original information using the idea of the residual network, and the attitude fusion algorithm uses the EKF algorithm to improve the accuracy of attitude data. The collaborative work of the data weighting unit and the excitation unit in the fully connected module enables the integrated knowledge to be adjusted according to the feature importance, and the finally generated target inspection point cloud element knowledge can more accurately reflect the state of the overhead distribution line equipment and the flight state of the UAV, providing a reliable data basis for subsequent inspection decisions and improving the accuracy and reliability of the entire inspection system.

[0112] Based on the above, the method further includes the following training process: Obtain the laser point cloud inspection information set and the initial inspection route record of the UAV sample, the Y cooperative UAVs of the UAV sample, and the conflicting UAVs of the UAV sample. Among them, the laser point cloud inspection information set includes key point detection information and flight state monitoring information. The key point detection information is used to represent the information obtained after laser point cloud scanning and collection of the UAV sample for the distribution network overhead line equipment within a specified inspection task cycle. The flight state monitoring information is used to represent the monitoring results corresponding to the attitude changes of the UAV sample within a specified inspection task cycle. The initial inspection route record is used to represent the line crossing information between the UAV sample and the cooperative UAVs; Determine Z cooperative UAVs from the Y cooperative UAVs, and generate the first inspection route record sample of the UAV sample based on the initial inspection route record of the Z cooperative UAVs and the UAV sample; Simulate the line crossing information between the UAV sample and the conflicting UAVs to generate the second inspection route record sample of the UAV sample; Use the laser point cloud inspection information set as the input information of the point cloud element recognition branch, and generate an inspection point cloud element knowledge sample through the point cloud element recognition branch; Use the inspection point cloud element knowledge sample, the initial inspection route record, the first inspection route record sample, and the second inspection route record sample as the input information of the line optimization decision branch, and generate the first route planning recommendation feature sample, the initial route planning recommendation feature sample, and the second route planning recommendation feature sample through the line optimization decision branch; Generate a training evaluation function based on the first route planning recommendation feature sample, the initial route planning recommendation feature sample, and the second route planning recommendation feature sample; Update the neuron weights of the point cloud element recognition branch and the line optimization decision branch through the training evaluation function.

[0113] The training process in the above technical solution aims to optimize the point cloud element recognition branch and the line optimization decision branch to improve the efficiency and accuracy of UAVs in the inspection of distribution network overhead lines. The entire training process involves the acquisition, processing of multiple data, and the generation of samples, and finally updates the neuron weights of the relevant branches through the training evaluation function.

[0114] 1. Acquisition of training data

[0115] First, obtain the laser point cloud inspection information set and the initial inspection route record of the UAV sample. The laser point cloud inspection information set includes key point detection information and flight state monitoring information. The key point detection information is the information obtained after the laser point cloud scanning of the UAV sample for the distribution network overhead line equipment during the specified inspection task cycle. For example, for the laser point cloud scanning of the tower body, data such as the coordinates and shape of the tower may be obtained as the key point detection information. The flight state monitoring information represents the monitoring results corresponding to the attitude changes of the UAV sample during the specified inspection task cycle, such as the changes in the pitch angle, roll angle, and yaw angle of the UAV. At the same time, obtain the Y cooperative UAVs of the UAV sample and the conflicting UAVs of the UAV sample. The initial inspection route record is used to represent the line crossing information between the UAV sample and the cooperative UAVs.

[0116] 2. Generation of Inspection Route Record Samples

[0117] 2.1 Generation of the First Inspection Route Record Sample

[0118] Determine Z cooperative UAVs (Z is greater than or equal to 1 and less than or equal to Y) from the Y cooperative UAVs. Based on the initial inspection route record of the Z cooperative UAVs and the UAV sample, generate the first inspection route record sample of the UAV sample. For example, if Y = 5 and Z = 3, select specific 3 cooperative UAVs. For example, the initial inspection route record records the line crossing situations of the UAV sample and 5 cooperative UAVs in coordinate form, such as the UAV sample crossing the route of cooperative UAV 1 at the coordinate ((x_1, y_1, z_1)), crossing the route of cooperative UAV 2 at ((x_2, y_2, z_2)), etc. For the selected 3 cooperative UAVs, extract the corresponding crossing coordinate information and organize this information in a certain format to generate the first inspection route record sample. This sample details the line crossing situations between the UAV sample and specific cooperative UAVs, such as information like the crossing frequency and relative speed at the time of crossing.

[0119] 2.2 Generation of the Second Inspection Route Record Sample

[0120] Generate a second inspection route record sample of the UAV example by simulating the route crossing information between the UAV example and the conflicting UAV. For example, the conflicting UAV may conflict due to accidentally entering the inspection area of the UAV example. For example, the route of the conflicting UAV is a straight line (L_1), and the route of the UAV example is a straight line (L_2). Simulate and calculate the intersection coordinates ((x_3, y_3, z_3)) and the intersection angle (theta) (which can be obtained by vector calculation, such as (cos theta = frac{vec{v_1}cdotvec{v_2}}{|vec{v_1}|*|vec{v_2}|}), where (vec{v_1}) and (vec{v_2}) are the direction vectors of the two routes) and other information, and integrate this information to form a second inspection route record sample.

[0121] 3. Generation of knowledge examples of inspection point cloud elements

[0122] Use the acquired laser point cloud inspection information set as the input information for the point cloud element recognition branch. The point cloud element recognition branch includes the key point quantization recognition module, the pose point cloud recognition module, the fully connected module, etc. as described before. According to the previous processing flow, pass the key point detection information through the vector space conversion unit (such as the sum operation of the corresponding feature members in the local point cloud description semantic map) and the knowledge collision unit (such as the product and sum operation of the corresponding feature members) of the key point quantization recognition module to generate the key point quantization knowledge encoding; pass the flight state monitoring information through the residual connection unit (such as using the residual network idea to process the flight state monitoring information to generate the pose monitoring data set) and the point cloud reconstruction unit (such as using the deep learning-based point cloud generation algorithm to process the pose monitoring event set) of the pose point cloud recognition module to generate the pose point cloud knowledge encoding; and finally generate the inspection point cloud element knowledge example through operations such as integration, data weighting, and excitation of the fully connected module.

[0123] 4. Generation of feature samples of route planning suggestions

[0124] Take the inspection point cloud element knowledge sample, the initial inspection route record, the first inspection route record sample, and the second inspection route record sample as the incoming information for the line optimization decision branch. The line optimization decision branch may be processed by means such as a genetic algorithm. For example, encode these input information, such as the damage degree of the pole tower in the inspection point cloud element knowledge sample (represented by a value between 0 and 1, 0 means no damage, 1 means severe damage), the working state of the transformer (represented by a value between -1 and 1, -1 means failure, 1 means normal operation), and the flight altitude (such as 100 meters), speed (such as 5 m / s), etc. in the inspection route record into binary strings. Then, according to the set fitness function (such as (F = a*S + b*E + c*C), where (F) is the fitness value, (S) is the flight safety index (value range 0 - 1, 1 means the safest), (E) is the inspection efficiency index (for example, the number of devices inspected per unit time), (C) is the coverage of the overhead distribution network line equipment (value range 0 - 1, 1 means full coverage), (a), (b), (c) are weight coefficients) to evaluate the pros and cons of each individual (the encoded route plan). Through selection, crossover, and mutation operations, continuously evolve the population, and finally generate the first route planning recommendation feature sample (corresponding to the route planning related to the first inspection route record sample), the initial route planning recommendation feature sample (corresponding to the route planning related to the initial inspection route record), and the second route planning recommendation feature sample (corresponding to the route planning related to the second inspection route record sample).

[0125] 5. Generation and update of neuron weights for the training evaluation function

[0126] 5.1 Generation of the training evaluation function

[0127] Generate a training evaluation function based on the first route planning recommendation feature sample, the initial route planning recommendation feature sample, and the second route planning recommendation feature sample. This training evaluation function can comprehensively consider various factors. For example, calculate the differences in indicators such as the accuracy and efficiency of the route planning for each sample. For example, the accuracy index (A) is calculated by comparing the deviation between the planned route and the actual optimal route (such as the average deviation distance), and the efficiency index (E) is calculated based on the number of devices inspected per unit time or the flight distance, etc. The training evaluation function can be (T = d*(A_1 - A_0) + e*(E_1 - E_0)), where (T) is the training evaluation value, (d) and (e) are weight coefficients, (A_1) and (A_0) are the accuracy indices of the newly generated sample and the initial sample respectively, and (E_1) and (E_0) are the efficiency indices of the newly generated sample and the initial sample respectively.

[0128] 5.2 Neuron weight update

[0129] Update the neuron weights of the point cloud feature recognition branch and the route optimization decision branch by training the evaluation function. For example, in a neural network, according to the value of the training evaluation function, use the backpropagation algorithm to calculate the gradient of each neuron. For the neurons in the key point quantization recognition module, the pose point cloud recognition module, and the fully connected module in the point cloud feature recognition branch, as well as the relevant neurons in the route optimization decision branch, update the neuron weights according to the calculated gradient at a certain learning rate (such as (alpha = 0.01)):

[0130] (w_{new}=w_{old}-alpha*frac{partialT}{partialw}), where (w_{new}) is the updated weight and (w_{old}) is the original weight.

[0131] With such a design, through a detailed training process, using a variety of inspection route record samples and inspection point cloud feature knowledge examples, the point cloud feature recognition branch and the route optimization decision branch can be continuously optimized. When generating inspection route record samples, the situations of cooperative drones and conflicting drones are accurately considered, such as the calculation of specific route intersection information, providing a comprehensive reference for subsequent route planning. In the process of generating inspection point cloud feature knowledge examples, the fine processing of each module can effectively extract equipment and flight state information. The route optimization decision branch generates route planning suggestion feature samples based on various samples and evaluates them through a comprehensive training evaluation function, which can accurately adjust the neuron weights, improving the accuracy, efficiency of the entire UAV inspection system for the inspection of distribution network overhead line equipment, and the adaptability to various complex situations (such as the presence of cooperative and conflicting drones).

[0132] In an exemplary technical solution, taking the inspection point cloud feature knowledge example, the initial inspection route record, the first inspection route record sample, and the second inspection route record sample as the incoming information of the route optimization decision branch, and generating the first route planning suggestion feature sample, the initial route planning suggestion feature sample, and the second route planning suggestion feature sample through the route optimization decision branch includes: taking the inspection point cloud feature knowledge example and the initial inspection route record as the incoming information of the route optimization decision branch, and generating the initial route planning suggestion feature sample through the route optimization decision branch; taking the inspection point cloud feature knowledge example and the first inspection route record sample as the incoming information of the route optimization decision branch, and generating the first route planning suggestion feature sample through the route optimization decision branch; taking the inspection point cloud feature knowledge example and the second inspection route record sample as the incoming information of the route optimization decision branch, and generating the second route planning suggestion feature sample through the route optimization decision branch.

[0133] In the present invention, this technical solution focuses on how the line optimization decision branch generates specific route planning recommendation feature samples based on different input information, including initial route planning recommendation feature samples, first route planning recommendation feature samples, and second route planning recommendation feature samples. This process is a key link in optimizing the route planning in the entire UAV inspection system. By reasonably processing different combinations of input information, it provides an accurate basis for the route planning of UAVs in the inspection of distribution network overhead lines.

[0134] 1. Generate initial route planning recommendation feature samples

[0135] 1.1. Input and processing basis of the line optimization decision branch

[0136] First, the inspection point cloud element knowledge sample and the initial inspection route record are used as the incoming information of the line optimization decision branch. The inspection point cloud element knowledge sample contains the comprehensive knowledge about the distribution network overhead line equipment and the UAV flight state obtained through the previous point cloud element recognition branch. For example, the inspection point cloud element knowledge sample may include the quantified value of the damage degree of the tower (for example, the value range is 0 - 1, 0 means no damage, 1 means severe damage), and the quantified value of the attitude stability during the UAV flight (for example, the value range is -1 to 1, -1 means very unstable, 1 means very stable), etc. The initial inspection route record details the line intersection information between the UAV sample and the cooperative UAV, such as the coordinates of the intersection point (for example, ((x_0, y_0, z_0))), the relative height difference at the intersection (for example, (h_0) meters), etc.

[0137] 1.2. Internal processing mechanism of the line optimization decision branch

[0138] The line optimization decision branch may use optimization algorithms such as genetic algorithms for processing. During the processing, first, the input inspection point cloud element knowledge sample and the initial inspection route record are encoded. For example, for the tower damage degree 0.3, the UAV attitude stability 0.5 in the inspection point cloud element knowledge sample, and the intersection point coordinates ((10, 20, 30)), the relative height difference of 5 meters, etc. in the initial inspection route record, they are encoded into binary strings. For example, the encoded inspection point cloud element knowledge sample is (P = [010101]) (only for example), and the initial inspection route record is (R = [110010]).

[0139] Then, evaluate the quality of each individual (the encoded flight route plan) according to the set fitness function. The fitness function may comprehensively consider multiple factors, such as (F = a*S + b*E + c*C). Among them, (S) is the flight safety index, which can be calculated according to the situation of the intersection point and the attitude stability of the UAV. For example, (S = 0.8) (the value range is 0-1, and 1 represents the safest); (E) is the inspection efficiency index, such as calculated according to the distance from the current position of the UAV to the next inspection point and the expected flight speed. For example, (E = 0.6) (affected by factors such as the number of devices inspected per unit time); (C) is the coverage degree of the overhead line equipment of the distribution network. For example, judge the coverage of the current flight route on the equipment according to information such as the damage degree of the pole tower in the inspection point cloud element knowledge sample, (C = 0.7) (the value range is 0-1, and 1 represents full coverage); (a = 0.3), (b = 0.4), (c = 0.3) are weight coefficients.

[0140] Continuously evolve the population through selection, crossover, and mutation operations. For example, in the selection operation, select individuals with higher fitness according to the fitness function value to enter the next generation population. For example, the population size is (N = 100), and select the top (50) individuals in terms of fitness. In the crossover operation, use single-point crossover, randomly select a crossover point, and exchange the parts of the two individuals after the crossover point. For example, individual (A = [010101]) and individual (B = [110010]). For example, the crossover point is the 3rd bit, and after crossover, (A' = [010010]) and (B' = [110101]) are obtained. The mutation operation changes a certain bit in the individual with a certain probability (for example, (p = 0.01)). After evolution for multiple generations (for example, (G = 100) generations), finally obtain the initial flight route planning suggestion feature sample. This sample may contain new flight altitude suggestions (for example, adjusted from the original 100 meters to 90 meters), new flight speed suggestions (from 5 m / s to 6 m / s), and new inspection point order and other information.

[0141] 2. Generate the first flight route planning suggestion feature sample

[0142] 2.1. Input information and initial processing

[0143] Next, use the inspection point cloud element knowledge sample and the first inspection flight route record sample as the incoming information for the line optimization decision branch. The first inspection flight route record sample is generated based on the initial inspection flight route records of Z cooperative UAVs determined from Y cooperative UAVs and the UAV sample, and it contains more detailed line intersection information between the UAV sample and specific cooperative UAVs. For example, the first inspection flight route record sample may contain information such as the intersection frequency with specific cooperative UAVs (for example, intersecting once every 10 minutes) and the relative speed at the time of intersection (for example, 2 m / s).

[0144] 2.2 Processing Process of Route Optimization Decision Branch

[0145] Similarly, first encode the inspection point cloud element knowledge example and the first inspection route record sample. For example, the inspection point cloud element knowledge example is encoded as (P_1 = [101010]), and the first inspection route record sample is encoded as (R_1 = [011011]). Then evaluate the pros and cons of each individual according to the fitness function. The fitness function here is similar to that when generating the initial route planning suggestion feature sample, but may adjust the weights or add new considerations according to specific information in the first inspection route record sample. For example, due to the cross frequency and relative speed information in the first inspection route record sample, the fitness function may be adjusted to (F_1 = a_1*S_1 + b_1*E_1 + c_1*C_1 + d_1*F_q + e_1*V_r), where (F_q) is the quantization value of the cross frequency (calculated according to the deviation of the actual cross frequency from the ideal cross frequency), (V_r) is the quantization value of the relative speed (calculated according to the deviation of the relative speed from the safe relative speed), and (a_1 = 0.2), (b_1 = 0.3), (c_1 = 0.2), (d_1 = 0.2), (e_1 = 0.1) are weight coefficients.

[0146] Evolve the population according to the selection, crossover, and mutation operations of the genetic algorithm. For example, select the top (60) individuals in terms of fitness (the population size is still (N = 100)), adopt two-point crossover (randomly select two crossover points and exchange the parts between the two points), the mutation probability is (p = 0.02), and after (G = 150) generations of evolution, obtain the first route planning suggestion feature sample. This sample may contain route adjustments for specific cooperative drones, such as adjusting the flight path to avoid overly frequent crossings (adjusting the original straight path that may cause frequent crossings to a slightly curved path), or adjusting the flight speed according to the relative speed (if the relative speed is too fast, reducing its own flight speed to ensure safety), etc.

[0147] 3. Generating the Second Route Planning Suggestion Feature Sample

[0148] 3.1 Input Information and Its Characteristics

[0149] Finally, use the inspection point cloud element knowledge example and the second inspection route record sample as the incoming information for the route optimization decision branch. The second inspection route record sample is obtained by simulating the route crossing information between the drone example and the conflicting drone, and it contains key information such as the crossing angle (for example, (theta = 30^{circ})) and the distance at the time of crossing (for example, (d = 5) meters).

[0150] 3.2 Processing Logic of Route Optimization Decision Branch

[0151] Encode the inspection point cloud element knowledge example and the second inspection route record sample. Let the inspection point cloud element knowledge example encoding be (P_2 = [010111]), and the second inspection route record sample encoding be (R_2 = [100101]). The fitness function will focus on factors related to conflicts, such as (F_2 = a_2*S_2 + b_2*E_2 + c_2*C_2 + f_2*theta_d + g_2*d_d), where (theta_d) is the quantization value of the crossing angle (e.g., calculated based on the deviation of the crossing angle from the safe crossing angle. The safe crossing angle is, for example, (45^{circ}), then:

[0152] (theta_d = frac{|theta - 45^{circ}|}{45^{circ}})), (d_d) is the quantization value of the crossing distance (e.g., calculated based on the deviation of the crossing distance from the safe distance. The safe distance is, for example, 10 meters, then (d_d = frac{|d - 10|}{10})), and (a_2 = 0.3), (b_2 = 0.2), (c_2 = 0.2), (f_2 = 0.2), (g_2 = 0.1) are weight coefficients.

[0153] Evolve the population through the selection, crossover, and mutation operations of the genetic algorithm. For example, select the top (40) individuals in terms of fitness (population size (N = 100)), use uniform crossover (exchange each bit in the individual with a certain probability), and the mutation probability is (p = 0.03). After (G = 200) generations of evolution, obtain the second route planning suggestion feature sample. This sample may include route adjustments for emergency obstacle avoidance, such as when the crossing angle is small and the crossing distance is close, significantly changing the flight direction (e.g., changing the flight direction by (90^{circ})), or urgently reducing the flight altitude to avoid collision with conflicting drones, etc.

[0154] In this way, by generating different route planning suggestion feature samples based on different combinations of input information respectively, it is possible to more precisely handle different inspection situations. When generating the initial route planning suggestion feature sample, considering the equipment status and basic collaborative drone crossing information comprehensively, the route planning in general cases is optimized, improving the inspection efficiency and equipment coverage. When generating the first route planning suggestion feature sample, based on the detailed crossing information with specific collaborative drones, the route related to collaborative drones is further optimized, reducing the crossing risk and improving flight safety. When generating the second route planning suggestion feature sample, for the situation of conflicting drones, factors such as crossing angle and distance are considered, effectively avoiding potential collision risks and ensuring the safe progress of the drone inspection task. Overall, this method of handling cases separately improves the autonomy, safety, and efficiency of drones in the inspection of distribution network overhead lines.

[0155] In an alternative design idea A, it involves the relevant processing of the original sample: The initial inspection route record includes the initial UAV inspection route record and the initial spatial distribution of the distribution network overhead structure. The initial UAV inspection route record includes the initial directional UAV, and the initial directional UAV and the UAV sample correspond to the same inspection task event; The initial spatial distribution of the distribution network overhead structure includes the key point structure spatial distribution, and the key point structure spatial distribution and the UAV sample correspond to the same distribution network overhead structure area; The line optimization decision branch includes a UAV line intersection mining module, a structure spatial distribution intersection mining module, an inspection route record combination module, and a route planning output module; Taking the inspection point cloud element knowledge sample and the initial inspection route record as the input information of the line optimization decision branch, and generating an initial route planning recommendation feature sample through the line optimization decision branch, including: Taking the initial UAV inspection route record and the inspection point cloud element knowledge sample as the input information of the UAV line intersection mining module, and generating an initial UAV line intersection feature through the UAV line intersection mining module; Taking the initial spatial distribution of the distribution network overhead structure and the inspection point cloud element knowledge sample as the input information of the structure spatial distribution intersection mining module, and generating an initial structure spatial distribution intersection feature through the structure spatial distribution intersection mining module; Taking the initial UAV line intersection feature and the initial structure spatial distribution intersection feature as the input information of the inspection route record combination module, and generating an associated structure spatial distribution intersection heat map of the initial line intersection combination information and an associated UAV line intersection heat map of the initial structure spatial distribution intersection combination information, where the associated structure spatial distribution intersection heat map of the initial line intersection combination information is obtained based on the initial structure spatial distribution intersection feature and the structure spatial collaborative distribution relationship network, and the associated UAV line intersection heat map of the initial structure spatial distribution intersection combination information is obtained based on the initial UAV line intersection feature and the UAV collaborative line relationship network; Taking the associated structure spatial distribution intersection heat map of the initial line intersection combination information and the associated UAV line intersection heat map of the initial structure spatial distribution intersection combination information as the input information of the route planning output module, and generating an initial route planning recommendation feature sample through the route planning output module.

[0156] In an alternative design idea B, it involves the related processing of positive samples: The first inspection route record sample includes the first UAV inspection route example and the first spatial distribution example of the overhead distribution network structure. The first UAV inspection route example includes the first directional cruise line example, and the first directional cruise line example corresponds to the same inspection task event as the UAV example. The first spatial distribution example of the overhead distribution network structure includes the first key point structure spatial distribution example, and the first key point structure spatial distribution example corresponds to the same overhead distribution network structure area as the UAV example. The line optimization decision branch includes a UAV line intersection mining module, a structure spatial distribution intersection mining module, an inspection route record combination module, and a route planning output module. Taking the inspection point cloud element knowledge example and the first inspection route record sample as the input information of the line optimization decision branch, and generating the first route planning suggestion feature sample through the line optimization decision branch, including: Taking the first UAV inspection route example and the inspection point cloud element knowledge example as the input information of the UAV line intersection mining module, and generating the first UAV line intersection feature sample through the UAV line intersection mining module; Taking the first spatial distribution example of the overhead distribution network structure and the inspection point cloud element knowledge example as the input information of the structure spatial distribution intersection mining module, and generating the first structure spatial distribution intersection feature sample through the structure spatial distribution intersection mining module; Taking the first UAV line intersection feature sample and the first structure spatial distribution intersection feature sample as the input information of the inspection route record combination module, and generating the first associated distribution intersection heat map example of the first line intersection combination information example and the first associated line intersection heat map example of the first structure distribution involvement feature example through the inspection route record combination module. Among them, the first associated distribution intersection heat map example of the first line intersection combination information example is obtained based on the first structure spatial distribution intersection feature sample and the structure spatial collaborative distribution relationship network, and the first associated line intersection heat map example of the first structure distribution involvement feature example is obtained based on the first UAV line intersection feature sample and the UAV collaborative line relationship network; Taking the first associated distribution intersection heat map example of the first line intersection combination information example and the first associated line intersection heat map example of the first structure distribution involvement feature example as the input information of the route planning output module, and generating the first route planning suggestion feature sample through the route planning output module.

[0157] In an alternative design idea C, it involves the relevant processing of negative samples: The second inspection route record sample includes the second UAV inspection route example and the second spatial distribution example of the distribution network overhead structure. The second UAV inspection route example includes the second directional cruise line example, and the second directional cruise line example corresponds to the same inspection task event as the UAV example. The second spatial distribution example of the distribution network overhead structure includes the second key point structure spatial distribution example, and the second key point structure spatial distribution example corresponds to the same distribution network overhead structure area as the UAV example. The line optimization decision branch includes a UAV line intersection mining module, a structure spatial distribution intersection mining module, an inspection route record combination module, and a route planning output module. Taking the inspection point cloud element knowledge example and the second inspection route record sample as the incoming information of the line optimization decision branch, and generating a second route planning recommendation feature sample through the line optimization decision branch, including: Taking the second UAV inspection route example and the inspection point cloud element knowledge example as the incoming information of the UAV line intersection mining module, and generating a second UAV line intersection feature sample through the UAV line intersection mining module; Taking the second spatial distribution example of the distribution network overhead structure and the inspection point cloud element knowledge example as the incoming information of the structure spatial distribution intersection mining module, and generating a second structure spatial distribution intersection feature sample through the structure spatial distribution intersection mining module; Taking the second UAV line intersection feature sample and the second structure spatial distribution intersection feature sample as the incoming information of the inspection route record combination module, and generating a second associated distribution intersection heat map example of the second line intersection combination information example and a second associated line intersection heat map example of the second structure distribution involvement feature example through the inspection route record combination module. Among them, the second associated distribution intersection heat map example of the second line intersection combination information example is obtained based on the second structure spatial distribution intersection feature sample and the structure spatial collaborative distribution relationship network, and the second associated line intersection heat map example of the second structure distribution involvement feature example is obtained based on the second UAV line intersection feature sample and the UAV collaborative line relationship network; Taking the second associated distribution intersection heat map example of the second line intersection combination information example and the second associated line intersection heat map example of the second structure distribution involvement feature example as the incoming information of the route planning output module, and generating a second route planning recommendation feature sample through the route planning output module.

[0158] In the present invention, the above three design ideas revolve around the operation process of the line optimization decision branch when processing different types of samples (original samples, positive samples, negative samples) to generate different route planning recommendation feature samples (initial route planning recommendation feature samples, first route planning recommendation feature samples, second route planning recommendation feature samples). The line optimization decision branch includes a UAV line intersection mining module, a structural spatial distribution intersection mining module, an inspection route record combination module, and a route planning output module. Different types of samples include different inspection route records and content related to the spatial distribution of overhead power distribution structures. The generation of route planning recommendation feature samples is achieved through the collaborative processing of each module.

[0159] Design idea A - Processing of original samples to generate initial route planning recommendation feature samples

[0160] 1. Analysis of the content of the initial inspection route record

[0161] 1.1 Initial UAV inspection route record

[0162] The initial UAV inspection route record in the initial inspection route record includes an initial directional UAV, and the initial directional UAV and the UAV sample correspond to the same inspection task event. This means that in the same inspection task, there is a certain association between the route of the initial directional UAV and the route of the UAV sample, and this association is of great significance for subsequent operations such as line intersection mining.

[0163] 1.2 Initial spatial distribution of overhead power distribution structures

[0164] The initial spatial distribution of overhead power distribution structures includes the spatial distribution of key points, and the spatial distribution of key points and the UAV sample correspond to the same overhead power distribution structure area. For example, in a specific overhead power distribution line area, the spatial distribution of key points may include information such as the position coordinates of the poles and the spatial distribution of the cross arms. These information are the spatial descriptions of the overhead power distribution structures in this area and provide basic data for subsequent structural spatial distribution intersection mining.

[0165] 2. Processing flow of each module

[0166] 2.1 The UAV line intersection mining module generates initial UAV line intersection features

[0167] Use the initial UAV inspection route record and the knowledge sample of inspection point cloud elements as the input information for the UAV line intersection mining module. The knowledge sample of inspection point cloud elements contains comprehensive knowledge of the overhead distribution line equipment and the UAV flight state. For example, the knowledge sample of inspection point cloud elements includes the current position coordinates ((x_1, y_1, z_1)) of the UAV, the flight direction vector (vec{v} = (v_x, v_y, v_z)), and the position coordinates ((x_t, y_t, z_t)) of the overhead distribution line equipment (such as a pole tower). The initial UAV inspection route record may be represented in the form of a series of waypoint coordinates ((x_{p1}, y_{p1}, z_{p1})), ((x_{p2}, y_{p2}, z_{p2})), etc.

[0168] The UAV line intersection mining module may determine whether there is an intersection point by calculating the distance formula (d = frac{|vec{AB} * vec{AC}|}{|vec{AB}|}) between two routes (the UAV sample route and the initial oriented UAV route) (where (vec{AB}) and (vec{AC}) are vectors constructed based on the route coordinates). If the calculated distance (d) is less than a certain threshold (for example, (d < 5) meters, and here 5 meters is the safety distance threshold set according to actual requirements and experience), it is considered that there is a line intersection situation. Through the calculation and analysis of the entire route, an initial UAV line intersection feature is generated, which may be a vector containing information such as the intersection point coordinates and the possibility of intersection (expressed in probability form, such as 0.8 indicating an 80% possibility of intersection).

[0169] 2.2. The structural spatial distribution intersection mining module generates the initial structural spatial distribution intersection feature

[0170] Use the initial overhead distribution network structural spatial distribution and the knowledge sample of inspection point cloud elements as the input information for the structural spatial distribution intersection mining module. The key point structural spatial distribution information (such as the coordinates and dimensions of the pole tower) in the initial overhead distribution network structural spatial distribution is combined with the UAV position information in the knowledge sample of inspection point cloud elements. For example, a spatial geometry algorithm is used to calculate the spatial relationship between the UAV and the pole tower. If a coordinate system is established with the pole tower as the center, calculate the relative position coordinates ((x_r, y_r, z_r)) of the UAV in this coordinate system.

[0171] Generate the initial structural spatial distribution intersection feature according to the relative position relationship between the UAV and each key point of the overhead distribution network structure. This feature may include information such as the closest distance between the UAV and the key structure (such as the distance from the UAV to the nearest pole tower is (d_{min} = 10) meters) and the azimuth angle of the UAV relative to the key structure (such as the azimuth angle relative to a certain pole tower is (theta = 45^{circ})).

[0172] 2.3. The inspection route record combination module generates a heat map

[0173] Take the initial UAV line intersection feature and the initial structural space distribution intersection feature as the input information of the inspection route record combination module. For example, the structural space collaborative distribution relationship network is a graph structure with key points of the distribution network overhead structure as nodes and their spatial relationships (such as distance, azimuth, etc.) as edge weights. Based on the initial structural space distribution intersection feature and the structural space collaborative distribution relationship network, calculate the association degree of each node (key point) with the line intersection through a certain algorithm (such as an algorithm based on graph convolutional neural network), and generate an associated structural space distribution intersection heat map of the initial line intersection combination information. This heat map is represented in matrix form, and each element in the matrix represents the association strength between a key point and the line intersection, and the numerical range can be 0 - 1. For example, the element value corresponding to a certain tower is 0.6, indicating a relatively high association strength between the tower and the line intersection.

[0174] At the same time, based on the initial UAV line intersection feature and the UAV collaborative line relationship network (which may be a graph structure with UAV routes as nodes and the intersection relationships between routes as edge weights), use a similar algorithm to calculate the association degree of each UAV route node with the structural space distribution intersection, and generate an associated UAV line intersection heat map of the initial structural space distribution intersection combination information. For example, the element value corresponding to a certain UAV route is 0.7, indicating a relatively strong association between this route and the structural space distribution intersection.

[0175] 2.4. The route planning output module generates an initial route planning suggestion feature sample

[0176] Take the associated structural space distribution intersection heat map of the initial line intersection combination information and the associated UAV line intersection heat map of the initial structural space distribution intersection combination information as the input information of the route planning output module. The route planning output module may perform route planning using a path planning algorithm (such as the A* algorithm) based on the information in the heat map. For example, the A* algorithm calculates an optimal path based on the starting point (the current position of the UAV), the target point (the next inspection point), and the obstacle information in the heat map (such as the area with a high association strength of line intersection is regarded as an obstacle). This optimal path may include information such as new waypoint coordinates, flight altitude adjustment suggestions (such as adjusting from the current 100 meters to 90 meters), flight speed adjustment suggestions (such as adjusting from 5 m / s to 6 m / s), etc., so as to generate an initial route planning suggestion feature sample.

[0177] Design idea B - Positive sample processing to generate the first route planning suggestion feature sample

[0178] 1. Analysis of the content of the first inspection route record sample

[0179] 1.1. First UAV inspection route example

[0180] The first UAV inspection route example in the first inspection route record sample includes the first directional cruise route example, and the first directional cruise route example corresponds to the same inspection task event as the UAV example. This is similar to the initial UAV inspection route record, but for the positive sample situation, and its route information is of great significance for analyzing line intersections and other situations in specific cases.

[0181] 1.2. First spatial distribution example of the overhead structure of the distribution network

[0182] The first spatial distribution example of the overhead structure of the distribution network includes the first spatial distribution example of key points structure, which corresponds to the same overhead structure area of the distribution network as the UAV example. This part of the content is similar to the initial spatial distribution of the overhead structure of the distribution network, which is a description of the spatial information of the overhead structure of the distribution network, but for the positive sample situation, there may be some differences in its data, such as different tower layouts or equipment states, etc.

[0183] 2. Processing flow of each module

[0184] 2.1. The UAV line intersection mining module generates the first UAV line intersection feature sample

[0185] Take the first UAV inspection route example and the knowledge example of inspection point cloud elements as the input information of the UAV line intersection mining module. For example, the knowledge example of inspection point cloud elements contains information such as the attitude angles of the UAV (pitch angle (alpha = 10^{circ}), roll angle (beta = 5^{circ})), etc., and the first UAV inspection route example is given by a specific route trajectory equation (such as a quadratic curve equation (y = ax^2 + bx + c) representing the route trajectory).

[0186] The UAV line intersection mining module calculates the intersection situation with other relevant routes (such as the cooperative UAV route) by analyzing information such as the route trajectory equation and the attitude angles of the UAV. For example, by simultaneously solving the trajectory equations of two routes, the possible intersection point coordinates ((x_c, y_c)) can be obtained. At the same time, consider the influence of the attitude angles of the UAV on the intersection situation, such as the attitude angles may affect the flight radius of the UAV, etc. Through these calculations and analyses, the first UAV line intersection feature sample is generated, and this sample may contain information such as the coordinates of the intersection point, the attitude angle information at the intersection, the predicted time of intersection (such as it is expected to intersect in 10 minutes), etc.

[0187] 2.2. The structure spatial distribution intersection mining module generates the first structure spatial distribution intersection feature sample

[0188] Use the first distribution example of the overhead structure of the primary distribution network and the knowledge example of the inspection point cloud elements as the input information for the cross-mining module of the structural space distribution. The first key point structure space distribution example in the first distribution example of the overhead structure of the primary distribution network may contain more detailed equipment status information, such as the inclination of the pole tower (for example, (theta_t = 3^{circ})), the deformation of the cross arm (for example, (Delta l = 0.1) meters), etc.

[0189] Combine the UAV flight status information (such as the flight altitude is 100 meters) in the knowledge example of the inspection point cloud elements, and calculate the intersection situation between the UAV and the overhead structure of the distribution network through a spatial analysis algorithm (such as a three-dimensional space coordinate transformation algorithm). For example, calculate the collision risk between the UAV and the inclined pole tower in different flight postures. According to these calculation results, generate the first cross-characteristic sample of the structural space distribution, and this sample may include information such as the collision risk probability between the UAV and different equipment (such as the collision risk probability with a certain pole tower is 0.1), and the safety distance margin from the key structure (such as the safety distance margin from a certain cross arm is 2 meters).

[0190] 2.3. The inspection route record combination module generates a heat map

[0191] Use the first cross-characteristic sample of the UAV line and the first cross-characteristic sample of the structural space distribution as the input information for the inspection route record combination module. The structural space collaborative distribution relationship network and the UAV collaborative line relationship network may have different weights or connection relationships in the case of positive samples, reflecting the unique situation of positive samples.

[0192] Through an algorithm based on a graph neural network, according to the first cross-characteristic sample of the structural space distribution and the structural space collaborative distribution relationship network, calculate and generate the first associated distribution cross heat map example of the first line cross combination information example. The element values in this heat map represent the degree of association between different key points and the line cross, which may be different from the original sample situation. For example, the element value corresponding to a certain device is 0.4, which is lower than the value in the original sample, indicating that the degree of association between this device and the line cross is lower in the case of positive samples.

[0193] At the same time, according to the first cross-characteristic sample of the UAV line and the UAV collaborative line relationship network, generate the first associated line cross heat map example of the first structural distribution involvement characteristic example. The element values in this heat map represent the degree of association between different flight routes and the structural space distribution cross. For example, the element value corresponding to a certain flight route is 0.5, reflecting the association situation between this flight route and the structural space distribution cross in the case of positive samples.

[0194] 2.4. The route planning output module generates the first route planning suggestion characteristic sample

[0195] Take the first associated distribution cross-thermal pattern example of the first line crossing combination information example and the first associated line crossing thermal pattern example of the first structural distribution involved feature example as the incoming information of the route planning output module. The route planning output module uses an improved path planning algorithm (such as an improved version of the Dijkstra algorithm), considering the information in the thermal map and some specific constraints of positive samples (such as the key inspection requirements for certain equipment) to perform route planning. For example, on the premise of meeting the key inspection requirements, an optimized route is calculated according to the obstacle information and path cost in the thermal map. This route may include information such as a new inspection sequence (inspecting a certain key equipment first), adjusted flight speed (such as adjusting the speed according to the distance from the equipment), etc., so as to generate the first route planning suggestion feature sample.

[0196] Design idea C - Negative sample processing to generate the second route planning suggestion feature sample

[0197] 1. Analysis of the content of the second inspection route record sample

[0198] 1.1 Second UAV inspection route example

[0199] The second UAV inspection route example in the second inspection route record sample includes the second directional cruise line example, corresponding to the same inspection task event as the UAV example. In the case of negative samples, there may be potential conflicts with other factors (such as conflicting UAVs) in this route, and its route information is an important basis for analyzing such conflict situations.

[0200] 1.2 Second distribution network overhead structure spatial distribution example

[0201] The second distribution network overhead structure spatial distribution example includes the second key point structure spatial distribution example, corresponding to the same distribution network overhead structure area as the UAV example. Similar to positive samples and original samples, but in the case of negative samples, due to factors such as conflicts, different considerations may be given to the spatial relationship of the distribution network overhead structure when analyzing route planning.

[0202] 2. Processing flow of each module

[0203] 2.1 The UAV line crossing mining module generates the second UAV line crossing feature sample

[0204] Use the second UAV inspection route example and the inspection point cloud element knowledge example as the input information for the UAV route intersection mining module. For example, the inspection point cloud element knowledge example contains the sensor status information of the UAV (such as a certain sensor failure, and the failure code is F001). The second UAV inspection route example is given in the form of discrete waypoints and flight direction vectors, such as waypoints ((x_{n1}, y_{n1}, z_{n1})), ((x_{n2}, y_{n2}, z_{n2})) and the flight direction vector (vec{v_n} = (v_{nx}, v_{ny}, v_{nz})).

[0205] Considering the sensor failure, the UAV route intersection mining module analyzes the intersection situation between the second UAV inspection route example and other relevant routes (such as conflicting UAV routes). For example, due to sensor failure, there may be a certain positioning error for the UAV, such as the positioning error range is (±5) meters. When calculating the intersection with the conflicting UAV route, this error range needs to be considered. Through complex calculations (such as the Monte Carlo simulation method, simulating various possible situations within the positioning error range), generate the second UAV route intersection feature sample. This sample may include the intersection probability considering the positioning error (such as 0.3, indicating a 30% probability of intersection), the possible intersection area (represented by a coordinate range, such as ((x_{r1}, y_{r1}, z_{r1})) to ((x_{r2}, y_{r2}, z_{r2}))), etc.

[0206] 2.2. The structure spatial distribution intersection mining module generates the second structure spatial distribution intersection feature sample

[0207] Use the second distribution network overhead structure spatial distribution example and the inspection point cloud element knowledge example as the input information for the structure spatial distribution intersection mining module. The second key point structure spatial distribution example in the second distribution network overhead structure spatial distribution example may contain some special situations, such as due to external interference (such as strong wind causing the pole tower to shake), the dynamic position change of the pole tower (for example, the pole tower shakes within a certain range, and the shaking center coordinates are ((x_{s}, y_{s}, z_{s})), and the shaking radius is (r = 1) meter).

[0208] Combined with the UAV flight status information in the inspection point cloud element knowledge example (such as the flight speed is unstable, and the speed fluctuation range is (±1) m / s), calculate the intersection situation between the UAV and the overhead distribution network structure through a dynamic spatial analysis algorithm (such as a spatial coordinate transformation algorithm considering time series). For example, calculate the collision risk between the UAV and the swaying pole tower at different times. According to these calculation results, generate a second structural spatial distribution intersection feature sample, which may include information such as the collision risk probability of the UAV with different devices in a dynamic situation (such as the collision risk probability with a certain swaying pole tower is 0.2), and the dynamic safety distance margin from the key structure (such as the dynamic safety distance margin from a certain cross-arm at a certain moment is 1.5 m).

[0209] 2.3. The inspection route record combination module generates a heat map

[0210] Take the second UAV line intersection feature sample and the second structural spatial distribution intersection feature sample as the input information of the inspection route record combination module. The structural spatial collaborative distribution relationship network and the UAV collaborative line relationship network will be adjusted according to the special situation of the negative sample in the case of negative samples. For example, higher weights are given to the routes and devices with conflict risks.

[0211] Through an algorithm based on a dynamic graph neural network (which can handle dynamically changing node and edge relationships), according to the second structural spatial distribution intersection feature sample and the structural spatial collaborative distribution relationship network, calculate and generate a second associated distribution intersection heat map example of the second line intersection combination information example. The element value in this heat map represents the degree of association between different key points and the line intersection in the case of negative samples. For example, the element value corresponding to a pole tower affected by strong wind is 0.7, which is higher than the value under normal circumstances, indicating that its degree of association with the line intersection is higher in the case of negative samples.

[0212] At the same time, according to the second UAV line intersection feature sample and the UAV collaborative line relationship network, generate a second associated line intersection heat map example of the second structural distribution involvement feature sample. The element value in this heat map represents the degree of association between different routes and the structural spatial distribution intersection in the case of negative samples. For example, the element value corresponding to a certain route with a conflict risk is 0.8, reflecting that this route has a strong association with the structural spatial distribution intersection in the case of negative samples.

[0213] 2.4. The route planning output module generates a second route planning recommendation feature sample

[0214] Take the second associated distribution cross heat map example of the second line crossing combination information example and the second associated line distribution cross heat map example of the second structure distribution involved feature example as the incoming information of the route planning output module. The route planning output module adopts an emergency route planning algorithm (such as a priority-based path planning algorithm), giving priority to avoiding conflicts and ensuring safety. For example, according to the information in the heat map, mark the areas with high conflict risks as high-priority avoidance areas, and calculate a route that can quickly avoid these areas. This route may include emergency steering instructions (such as turning left (90°)) and emergency reduction of flight altitude (such as reducing from 100 meters to 80 meters emergently), thus generating a second route planning suggestion feature sample.

[0215] With such a design, through different processing methods for the original sample, positive sample, and negative sample, the inspection routes of drones can be comprehensively and accurately planned. For the original sample, in the initial route planning, comprehensively consider the relationship between the drone line crossing and the structure space distribution crossing. The generated initial route planning suggestion feature sample can optimize the route under normal inspection conditions, improving the inspection efficiency and safety. When processing the positive sample, according to its specific route and structure space distribution, the generated first route planning suggestion feature sample can adapt to special inspection requirements, such as key inspections of specific equipment, while optimizing the relationship with other relevant factors. When processing the negative sample, the generated second route planning suggestion feature can take into account various abnormal situations (such as sensor failures, external interferences, etc.), thus improving the safety of the cruise.

[0216] In some other embodiments, it involves the specific determination method of the training evaluation function, that is, generating the training evaluation function based on the first route planning suggestion feature sample, the initial route planning suggestion feature sample, and the second route planning suggestion feature sample, including: generating a first training error index based on the commonality score between the first route planning suggestion feature sample and the initial route planning suggestion feature sample; generating a second training error index based on the commonality score between the second route planning suggestion feature sample and the initial route planning suggestion feature sample; generating the training evaluation function based on the first training error index and the second training error index.

[0217] It can be understood that this embodiment focuses on how to generate the training evaluation function based on the first route planning suggestion feature sample, the initial route planning suggestion feature sample, and the second route planning suggestion feature sample. This process mainly calculates the commonality scores between different samples, then obtains the training error index, and finally constructs the training evaluation function, which is of crucial significance for optimizing the entire drone inspection route planning system.

[0218] First, calculate the commonality score between the first route planning recommendation feature sample and the initial route planning recommendation feature sample. The first route planning recommendation feature sample and the initial route planning recommendation feature sample contain various information about the UAV inspection route planning. For example, the first route planning recommendation feature sample may include a new flight altitude recommendation (e.g., (h_1 = 90) meters), a new flight speed recommendation (e.g., (v_1 = 6) m / s), a new inspection point order (e.g., (P_1 = (p_{11}, p_{12}, p_{13}))), etc.; the initial route planning recommendation feature sample may include a flight altitude of (h_0 = 100) meters, a flight speed of (v_0 = 5) m / s, an inspection point order of (P_0 = (p_{01}, p_{02}, p_{03})), etc.

[0219] When calculating the commonality score, various methods can be adopted. One possible method is the weighted sum based on features. For example, for the flight altitude, define the altitude difference weight (w_h = 0.3), then the contribution of the altitude difference to the commonality score is (w_h * |h_1 - h_0| = 0.3 * |90 - 100| = 3). For the flight speed, the speed difference weight is (w_v = 0.3), and the contribution of the speed difference to the commonality score is (w_v * |v_1 - v_0| = 0.3 * |6 - 5| = 0.3). For the inspection point order, define an inspection point order difference function (d(P_1, P_0)), for example, the value calculated by (d(P_1, P_0)) is 0.4, and the inspection point order difference weight is (w_p = 0.4), then the contribution of the inspection point order difference to the commonality score is (w_p * d(P_1, P_0) = 0.4 * 0.4 = 0.16). Add these contributions to get the commonality score (S_{10} = 3 + 0.3 + 0.16 = 3.46) (the calculation here is only an example, and the actual calculation may be more complex).

[0220] After obtaining the commonality score (S_{10}), generate the first training error metric based on this. A common method is to take the reciprocal after normalizing the commonality score. For example, using linear normalization, map (S_{10}) to the interval of (0 - 1). For example, the mapped result is (S'_{10} = 0.6) (the specific normalization function is determined according to actual needs), then the first training error metric is (E_1 = \frac{1}{1 - S'_{10}} = \frac{1}{1 - 0.6} = 2.5). This first training error metric reflects the degree of difference between the first route planning recommendation feature sample and the initial route planning recommendation feature sample. The larger the error metric, the greater the difference.

[0221] Similarly, calculate the commonality score between the second route planning recommendation feature sample and the initial route planning recommendation feature sample. For example, the second route planning recommendation feature sample includes flight altitude (h_2 = 80) meters, flight speed (v_2 = 4) meters per second, inspection point sequence (P_2 = (p_{21}, p_{22}, p_{23})), etc.

[0222] According to a similar calculation method as before, for the flight altitude, with the altitude difference weight (w_h = 0.3), the contribution of the altitude difference to the commonality score is (w_h * |h_2 - h_0| = 0.3 * |80 - 100| = 6). For the flight speed, with the speed difference weight (w_v = 0.3), the contribution of the speed difference to the commonality score is (w_v * |v_2 - v_0| = 0.3 * |4 - 5| = 0.3). For the inspection point sequence, for example, the value calculated by (d(P_2, P_0)) is (0.5), and the inspection point sequence difference weight (w_p = 0.4), then the contribution of the inspection point sequence difference to the commonality score is (w_p * d(P_2, P_0) = 0.4 * 0.5 = 0.2). Adding these contributions together gives the commonality score (S_{20} = 6 + 0.3 + 0.2 = 6.5).

[0223] Then generate the second training error metric based on this commonality score. After the same normalization process and taking the reciprocal, for example, if the normalized result is (S'_{20} = 0.8), then the second training error metric (E_2 = \frac{1}{1 - S'_{20}} = \frac{1}{1 - 0.8} = 5). This second training error metric reflects the degree of difference between the second route planning recommendation feature sample and the initial route planning recommendation feature sample.

[0224] Generate a training evaluation function based on the first training error metric (E_1) and the second training error metric (E_2). One construction method is a linear combination. For example, the training evaluation function (F = a * E_1 + b * E_2), where (a) and (b) are weight coefficients. For example, (a = 0.4), (b = 0.6), then (F = 0.4 * 2.5 + 0.6 * 5 = 1 + 3 = 4). The value of this training evaluation function reflects the overall training error situation. Based on the value of this function, relevant algorithms or models can be adjusted and optimized. For example, if the value of (F) is greater than a certain threshold (for example, the threshold is (3)), it indicates that the difference between the current route planning recommendation feature samples is relatively large, and some parameters or algorithms in the route optimization decision branch may need to be adjusted, such as adjusting the crossover probability and mutation probability in the genetic algorithm, or adjusting the weight coefficients when calculating the commonality score, etc.

[0225] In this way, by calculating the commonality scores between the characteristic samples of different route planning suggestions to generate a training error index, and then constructing a training evaluation function, this method can quantitatively evaluate the differences between different samples. Specific numerical examples show how to accurately calculate the commonality scores, training error index, and training evaluation function, making the evaluation of route planning more precise. For example, when calculating the commonality scores, the comprehensive consideration of factors such as flight altitude, speed, and inspection point order can comprehensively reflect the similarity between samples. The training evaluation function can guide the adjustment of relevant algorithms or models based on the overall error situation, which helps to improve the accuracy and stability of the UAV inspection route planning, and reduce problems such as low inspection efficiency or safety risks caused by unreasonable route planning.

[0226] Furthermore, Figure 2 is a schematic structural diagram of an unmanned aerial vehicle (UAV) autonomous inspection processing system 200 provided by the present invention. As Figure 2 shown, the UAV autonomous inspection processing system 200 includes a processor 210. The processor 210 can call and run a computer program from a memory to implement the method in the present invention.

[0227] Optionally, as Figure 2 shown, the UAV autonomous inspection processing system 200 may further include a memory 230. Among them, the processor 210 can call and run a computer program from the memory 230 to implement the method in the present invention.

[0228] Among them, the memory 230 can be an independent device from the processor 210, or can be integrated in the processor 210.

[0229] Optionally, as Figure 2 shown, the UAV autonomous inspection processing system 200 may further include a transceiver 220. The processor 210 can control the transceiver 220 to interact with other devices. Specifically, it can send information or data to other devices, or receive information or data sent by other devices.

[0230] Optionally, the UAV autonomous inspection processing system 200 can implement the corresponding processes of the storage engine or components (such as processing modules) in the storage engine or the device deployed with the storage engine in each method of the present invention. For the sake of brevity, it will not be elaborated here.

[0231] It should be understood that the processor of the present invention may be an integrated circuit chip with signal processing capabilities.

[0232] It can be understood that the memory in the present invention can be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. It should be noted that the memory of the system and method described herein is intended to include but not be limited to suitable types of memories.

[0233] On the basis described above, a readable storage medium is provided, and a program or instruction is stored on the readable storage medium. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0234] Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc), and includes several instructions for causing a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

[0235] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the purpose of the present invention and the scope protected by the present invention, and all of them belong to the protection scope of the present invention.

Claims

1. A method for autonomous inspection of overhead line equipment using a UAV, characterized in that: The method is applied to a UAV autonomous inspection processing system, and the method comprises: Obtain X laser point cloud inspection information sets and X inspection route records of X target UAVs, wherein each of the X laser point cloud inspection information sets includes key point detection information and flight status monitoring information, the key point detection information is used to characterize information obtained after laser point cloud scanning and collection of distribution network overhead line equipment for the target UAV within a specified inspection task cycle, and the flight status monitoring information is used to characterize monitoring results corresponding to posture changes of the target UAV within the specified inspection task cycle, and X is greater than or equal to 1; The X laser point cloud inspection information sets are used as the input information of the point cloud element recognition branch in the inspection scheduling processing network, and X target inspection point cloud element knowledge is generated through the point cloud element recognition branch, wherein the inspection scheduling processing network is debugged based on the initial inspection route record, the first inspection route record sample and the second inspection route record sample of the drone sample, the initial inspection route record is used to characterize the route intersection information of the drone sample and the Y coordinated drones corresponding to the drone sample, the first inspection route record sample is used to characterize the route intersection information of the drone sample and the Z coordinated drones determined from the Y coordinated drones, and the second inspection route record sample is used to characterize the route intersection information corresponding to the drone sample and the conflicting drones, and Z is greater than or equal to 1 and less than or equal to Y; The X target inspection point cloud element knowledge and the X inspection route records are used as input information of a route optimization decision branch in the inspection scheduling processing network, and X target route planning suggestion features are generated through the route optimization decision branch; Based on the route clustering processing branch, the X target route planning suggestion features are grouped to generate Q grouping results, where Q is greater than or equal to 1; Identify the inspection energy efficiency judgment weights of the Q sub-cluster results, and determine a drone inspection planning cluster based on the inspection energy efficiency judgment weights of the Q sub-cluster results, wherein the drone inspection planning cluster includes at least one of the target drones.

2. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 1, characterized in that: The method further comprises: Obtain a laser point cloud inspection information set and an initial inspection route record of a drone sample, Y cooperative drones of the drone sample, and a conflicting drone of the drone sample, wherein the laser point cloud inspection information set includes key point detection information and flight status monitoring information, the key point detection information is used to characterize information obtained after laser point cloud scanning and collection of distribution network overhead line equipment for the drone sample within a specified inspection task cycle, the flight status monitoring information is used to characterize monitoring results corresponding to posture changes of the drone sample within a specified inspection task cycle, and the initial inspection route record is used to characterize line intersection information between the drone sample and the cooperative drone; Determine Z cooperative drones from the Y cooperative drones, and generate a first inspection route record sample of the drone sample based on the Z cooperative drones and the initial inspection route records of the drone sample; Simulate the route intersection information of the drone sample and the conflicting drone to generate a second inspection route record sample of the drone sample; Using the laser point cloud inspection information set as the input information of the point cloud element recognition branch, and generating inspection point cloud element knowledge samples through the point cloud element recognition branch; The inspection point cloud element knowledge sample, the initial inspection route record, the first inspection route record sample and the second inspection route record sample are used as input information of the route optimization decision branch, and the first route planning suggestion feature sample, the initial route planning suggestion feature sample and the second route planning suggestion feature sample are generated through the route optimization decision branch; Generate a training evaluation function based on the first route planning suggestion feature sample, the initial route planning suggestion feature sample and the second route planning suggestion feature sample; The neuron weights of the point cloud element recognition branch and the route optimization decision branch are updated through the training evaluation function.

3. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 2, characterized in that: The method uses the inspection point cloud element knowledge sample, the initial inspection route record, the first inspection route record sample and the second inspection route record sample as input information of a route optimization decision branch, and generates a first route planning suggestion feature sample, an initial route planning suggestion feature sample and a second route planning suggestion feature sample through the route optimization decision branch, including: The inspection point cloud element knowledge sample and the initial inspection route record are used as input information of the route optimization decision branch, and an initial route planning suggestion feature sample is generated through the route optimization decision branch; The inspection point cloud element knowledge sample and the first inspection route record sample are used as input information of the route optimization decision branch, and a first route planning suggestion feature sample is generated through the route optimization decision branch; The inspection point cloud element knowledge sample and the second inspection route record sample are used as input information of the route optimization decision branch, and the second route planning suggestion feature sample is generated through the route optimization decision branch.

4. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 3 is characterized in that: The initial inspection route record includes an initial drone inspection route record and an initial distribution network overhead structure spatial distribution, wherein the initial drone inspection route record includes an initial directional drone, and the initial directional drone and the drone sample correspond to the same inspection task event; The initial distribution network overhead structure spatial distribution includes key point structure spatial distribution, and the key point structure spatial distribution and the drone sample correspond to the same distribution network overhead structure area; the line optimization decision branch includes a drone line intersection mining module, a structure space distribution intersection mining module, an inspection route record combination module and a route planning output module; The method of using the inspection point cloud element knowledge sample and the initial inspection route record as input information of a route optimization decision branch, and generating an initial route planning suggestion feature sample through the route optimization decision branch, includes: The initial drone inspection route record and the inspection point cloud element knowledge sample are used as input information of the drone route intersection mining module, and the initial drone route intersection feature is generated by the drone route intersection mining module; The initial distribution network overhead structure spatial distribution and the inspection point cloud element knowledge sample are used as input information of the structure spatial distribution cross mining module, and the initial structure spatial distribution cross features are generated by the structure spatial distribution cross mining module; The initial UAV route intersection feature and the initial structural spatial distribution intersection feature are used as input information of the inspection route record combination module, and the associated structural spatial distribution intersection heat map of the initial route intersection combination information and the associated UAV route intersection heat map of the initial structural spatial distribution intersection combination information are generated by the inspection route record combination module, wherein the associated structural spatial distribution intersection heat map of the initial route intersection combination information is obtained based on the initial structural spatial distribution intersection feature and the structural spatial collaborative distribution relationship network, and the associated UAV route intersection heat map of the initial structural spatial distribution intersection combination information is obtained based on the initial UAV route intersection feature and the UAV collaborative route relationship network; The associated structural spatial distribution cross heat map of the initial route intersection combination information and the associated drone route cross heat map of the initial structural spatial distribution cross combination information are used as input information of the route planning output module, and the initial route planning suggestion feature sample is generated through the route planning output module.

5. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 3 is characterized in that: The first inspection route record sample includes a first drone inspection route sample and a first distribution network overhead structure spatial distribution sample, the first drone inspection route sample includes a first directional cruise route sample, and the first directional cruise route sample and the drone sample correspond to the same inspection task event; The first distribution network overhead structure spatial distribution sample includes a first key point structure spatial distribution sample, and the first key point structure spatial distribution sample and the drone sample correspond to the same distribution network overhead structure area; the line optimization decision branch includes a drone line intersection mining module, a structure space distribution intersection mining module, an inspection route record combination module and a route planning output module; The method of using the inspection point cloud element knowledge sample and the first inspection route record sample as input information of a route optimization decision branch, and generating a first route planning suggestion feature sample through the route optimization decision branch, includes: Using the first drone inspection route sample and the inspection point cloud element knowledge sample as input information of the drone route intersection mining module, and generating a first drone route intersection feature sample through the drone route intersection mining module; The first distribution network overhead structure spatial distribution sample and the inspection point cloud element knowledge sample are used as input information of the structure spatial distribution cross mining module, and the first structure spatial distribution cross feature sample is generated by the structure spatial distribution cross mining module; The first UAV route intersection feature sample and the first structural space distribution intersection feature sample are used as input information of the inspection route record combination module, and the first associated distribution cross heat map sample of the first route intersection combination information sample and the first associated route cross heat map sample of the first structural distribution involved feature sample are generated by the inspection route record combination module, wherein the first associated distribution cross heat map sample of the first route intersection combination information sample is obtained based on the first structural space distribution cross feature sample and the structural space collaborative distribution relationship network, and the first associated route cross heat map sample of the first structural distribution involved feature sample is obtained based on the first UAV route intersection feature sample and the UAV collaborative route relationship network; The first associated distribution cross heat map sample of the first route intersection combination information sample and the first associated route cross heat map sample of the first structural distribution involved feature sample are used as input information of the route planning output module, and the first route planning suggestion feature sample is generated through the route planning output module.

6. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 3, characterized in that: The second inspection route record sample includes a second UAV inspection route sample and a second distribution network overhead structure spatial distribution sample, the second UAV inspection route sample includes a second directional cruise route sample, and the second directional cruise route sample and the UAV sample correspond to the same inspection task event; The second distribution network overhead structure spatial distribution sample includes a second key point structure spatial distribution sample, and the second key point structure spatial distribution sample and the drone sample correspond to the same distribution network overhead structure area; the line optimization decision branch includes a drone line intersection mining module, a structure space distribution intersection mining module, an inspection route record combination module and a route planning output module; The method of using the inspection point cloud element knowledge sample and the second inspection route record sample as input information of a route optimization decision branch, and generating a second route planning suggestion feature sample through the route optimization decision branch, includes: Using the second drone inspection route sample and the inspection point cloud element knowledge sample as input information of the drone route intersection mining module, and generating a second drone route intersection feature sample through the drone route intersection mining module; The second distribution network overhead structure spatial distribution sample and the inspection point cloud element knowledge sample are used as input information of the structure spatial distribution cross mining module, and a second structure spatial distribution cross feature sample is generated by the structure spatial distribution cross mining module; The second UAV route intersection feature sample and the second structural space distribution intersection feature sample are used as input information of the inspection route record combination module, and the second associated distribution intersection heat map sample of the second route intersection combination information sample and the second associated route intersection heat map sample of the second structural distribution involved feature sample are generated by the inspection route record combination module, wherein the second associated distribution cross heat map sample of the second route intersection combination information sample is obtained based on the second structural space distribution intersection feature sample and the structural space collaborative distribution relationship network, and the second associated route intersection heat map sample of the second structural distribution involved feature sample is obtained based on the second UAV route intersection feature sample and the UAV collaborative route relationship network; The second associated distribution cross heat map sample of the second route intersection combination information sample and the second associated route cross heat map sample of the second structural distribution involved feature sample are used as input information of the route planning output module, and the second route planning suggestion feature sample is generated through the route planning output module.

7. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 2, characterized in that: The generating of a training evaluation function based on the first route planning suggestion feature sample, the initial route planning suggestion feature sample and the second route planning suggestion feature sample comprises: generating a first training error indicator based on a commonality score between the first route planning suggestion feature sample and the initial route planning suggestion feature sample; generating a second training error indicator based on a commonality score between the second route planning suggestion feature sample and the initial route planning suggestion feature sample; A training evaluation function is generated based on the first training error indicator and the second training error indicator.

8. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 1, characterized in that: The X laser point cloud inspection information sets include X key point detection information and X flight status monitoring information, and the point cloud element recognition branch includes a key point quantization recognition module, a posture point cloud recognition module and a full connection module; The method of using the X laser point cloud inspection information sets as the incoming information of the point cloud element recognition branch in the inspection scheduling processing network, and generating X target inspection point cloud element knowledge through the point cloud element recognition branch, includes: Using the X key point detection information as input information of the key point quantization identification module, and generating X key point quantization knowledge codes through the key point quantization identification module; Using the X flight status monitoring information as input information of the attitude point cloud recognition module, and generating X attitude point cloud knowledge codes through the point cloud element recognition branch; The X key point quantization knowledge codes and the X posture point cloud knowledge codes are used as input information of the fully connected module, and X target inspection point cloud element knowledge is generated through the input information of the fully connected module.

9. The autonomous inspection and processing method of UAV based on distribution network overhead line equipment according to claim 8, characterized in that: The key point quantization identification module includes a vector space conversion unit and a knowledge collision unit; The step of using the X key point detection information as input information of the key point quantization identification module and generating X key point quantization knowledge codes through the key point quantization identification module includes: Based on the key point image semantics of the X key point detection information and the X key point detection information, generate X local point cloud description semantic graphs; The X local point cloud description semantic graphs are used as input information of the vector space conversion unit, and X local point clouds are generated based on space migration encoding by the vector space conversion unit, wherein the vector space conversion unit is used to sum the corresponding feature members in each local point cloud description semantic graph in the X local point cloud description semantic graphs; The X local point cloud description semantic graphs are used as input information of the knowledge collision unit, and X local point cloud advanced space migration codes are generated by the knowledge collision unit, wherein the knowledge collision unit is used to calculate the product of corresponding feature members in each local point cloud description semantic graph in the X local point cloud description semantic graphs, and sum the results of calculating the product of corresponding feature members; The X local point clouds are integrated based on the space migration coding and the X local point clouds are integrated based on the space migration coding to obtain X key point quantization knowledge coding.

10. An autonomous inspection and processing system for unmanned aerial vehicles, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images