Unmanned aerial vehicle inspection method and system for directional task of power transmission line and medium

Abstract

The invention relates to a power transmission line directional task unmanned aerial vehicle inspection method and system and a medium. The method comprises the steps of collecting point cloud data in a power transmission line unmanned aerial vehicle inspection range; carrying out pre-training on a point cloud active selection labeling model based on comparative learning; importing the collected unmanned aerial vehicle inspection point cloud data of the power transmission line into the model so as to mark a tower part; selecting a polling tower part to issue a polling task; and carrying out power transmission line unmanned aerial vehicle path planning for the selected tower part. According to the method, through the point cloud labeling technology combining comparative learning and active learning and the improved myxobacteria optimization path planning algorithm, the remarkable effects are achieved in the aspects of improving the inspection efficiency and reducing the risk of the power transmission line.

Description

Technical Field

[0001] This application relates to the field of power transmission line inspection by drones, and particularly to a method, system, and medium for inspecting power transmission lines by drones with directional tasks. Background Art

[0002] Drone inspection is a method of using drone technology to inspect and maintain power facilities such as power transmission lines and towers. It has the characteristics of high efficiency, safety, and precision, and is widely used in the operation and maintenance of power transmission lines in the power industry. Currently, the commonly used autonomous refined inspection is to inspect all parts of the tower based on the pre-planned flight path. However, when conducting a special inspection for a certain type of problem or performing a directional task, the autonomous refined inspection is no longer efficient. Currently, when performing a directional task, only the waypoint of the corresponding tower part is retained in the existing autonomous refined inspection flight path, and then the waypoints of each tower are connected step by step to generate a new flight path. Its disadvantages are large workload and complex operation, which are not suitable for the large-scale inspection work of drones. Summary of the Invention

[0003] The purpose of the embodiments of this application is to overcome the deficiencies of the existing technology, and provide a method, system, and medium for inspecting power transmission lines by drones with directional tasks, which can improve the inspection efficiency and reduce the risks of power transmission lines.

[0004] To achieve the above purpose, this application provides the following technical solutions:

[0005] In the first aspect, the embodiments of this application provide a method for inspecting power transmission lines by drones with directional tasks, including the following steps:

[0006] Collect point cloud data within the inspection range of the power transmission line by drones;

[0007] Pre-train a point cloud active selection annotation model based on contrastive learning;

[0008] Import the collected point cloud data of the power transmission line inspected by drones into the model for tower part annotation;

[0009] Select the inspected tower part and issue an inspection task;

[0010] Plan the path of the power transmission line drone for the selected tower part.

[0011] The specific method of collecting the point cloud data within the inspection range of the power transmission line by drones is to use the lidar point cloud on the drone to scan the power transmission line and collect the three-dimensional point cloud data of the power transmission line.

[0012] The specific method of pre-training the point cloud active selection annotation model based on contrastive learning is

[0013] S1. Unsupervised feature extraction pre-training based on contrastive learning. For unlabeled point cloud data, first extract effective features through a feature extraction module; then use the information of the data itself to constrain the extracted features to achieve effective pre-training of the model; after the training is completed, save the model parameters at this time;

[0014] S2. Active annotation of unlabeled data based on active learning. Use the trained model to take unlabeled point cloud data as input; after extracting the point cloud features, based on the obtained features and the calculated pseudo-labels, select some data from all unlabeled point clouds for manual annotation according to the designed selection metrics;

[0015] After performing the above steps, complete one round of model training and data annotation; then input the labeled data and the remaining unlabeled data into step S1 to repeat the above process. At this time, not only calculate the contrastive learning loss, but also calculate the cross-entropy loss of the labeled data to update the model, and continue to input the unlabeled data into step S2 to execute. This process alternates until the maximum number of annotations is reached, or the obtained model meets the accuracy requirements.

[0016] The unsupervised feature extraction pre-training based on contrastive learning is specifically as follows. The input data is where N is the number of points input into the model, and each point is represented by its coordinates in space; construct a transformation set {T i} for transforming the coordinates of the input point cloud;

[0017] The upper and lower two branches h1 and h2 of the feature extraction part process P1 and P2 respectively, and the two branches adopt exactly the same structure; then input the extracted point-by-point features into the feature space projection modules M1 and M2 with the same structure, and further project them into the feature space to obtain normalized features. The whole process uses formula (1):

[0018]

[0019] Calculate PointInfo Loss in the normalized feature space and perform backpropagation to train the model. The parameters θ1 of the upper branch h1 and M1 are updated through backpropagation; according to the structure of MoCo, the parameters θ2 of the lower branch h2 and M2 are updated by momentum according to θ1, and the formula is:

[0020] θ2←mθ2+(1 - m)θ1(2)

[0021] where m represents the momentum update parameter. Fix the model parameters obtained after training for selecting data for annotation in the subsequent active learning process. The specific form of this loss function is:

[0022]

[0023] Among them, P o represents the set of positive example pairs composed of the origin point cloud and the transformed point cloud; N e represents the set of negative example pairs stored in the MoCo queue structure; the subscripts i and j represent different transformation features of the same point cloud; f i and f j constitute a positive example, and f i and f k constitute a negative example; τ represents the temperature parameter, which is used to control the proportion of positive and negative examples participating in the calculation.

[0024] If there are labeled tags in the input point cloud data, this part of the point cloud will be continuously input into the classification module, which is composed of a multi-layer fully connected network. The normalized features are regressed to obtain the classification probability of this point. Cross-entropy is used for loss calculation in this part, and the formula is:

[0025]

[0026] Among them, K represents the number of categories in the dataset; y ij represents the sign function. When the sample i belongs to the category j, y ij =1, otherwise y ij =0; p ij represents the predicted probability that the point i belongs to the category j obtained by calculation. After training, the classification model parameters obtained by fixing this part of the training are also used for subsequent use.

[0027] Specifically, the selected inspection tower part is given an inspection task by applying the trained model to the transmission line to identify the tower part of the transmission line and remotely issue an inspection instruction for a specific tower part.

[0028] The path planning of the UAV for the transmission line for the selected tower part is to use an improved slime mold optimization algorithm for path planning, which mainly includes the following processes:

[0029] S11. Set the population size, maximum number of iterations and related parameters;

[0030] S12. Initialize the population using the improved Logistic chaotic map, calculate the optimal and worst fitness of the population and sort them;

[0031] S13. Judge whether the random number r < c. If it is satisfied, the slime mold approaches the food, and a non-linear adaptive weight factor is added during the process. If it is not satisfied, go to step S15;

[0032] S14. Calculate the population fitness and update the optimal position;

[0033] S15. Sort according to the difference in each fitness value;

[0034] S16. Perform adaptive Cauchy compilation on the current position, calculate the fitness value, and press the optimal position;

[0035] S17. Whether the maximum number of iterations is reached. If not satisfied, repeat the above steps until the conditions are met.

[0036] In a second aspect, an embodiment of the present application provides an unmanned aerial vehicle inspection system for transmission line directivity tasks, including a memory and a processor. The memory includes a program for the unmanned aerial vehicle inspection method for transmission line directivity tasks. When the program for the unmanned aerial vehicle inspection method for transmission line directivity tasks is executed by the processor, the following steps are implemented: collecting point cloud data within the inspection range of the unmanned aerial vehicle for transmission lines; pre-training a point cloud active selection annotation model based on contrastive learning; importing the collected point cloud data of the unmanned aerial vehicle inspection for transmission lines into the model for tower part annotation; selecting the inspected tower part to issue an inspection task; and performing path planning for the unmanned aerial vehicle for transmission lines for the selected tower part.

[0037] In a third aspect, an embodiment of the present application provides a computer-readable storage medium. When the program code stored in the computer-readable storage medium is executed by a processor, the steps of the above-mentioned unmanned aerial vehicle inspection method for transmission line directivity tasks are implemented.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows: By combining the point cloud annotation technology of contrastive learning and active learning and the improved slime mold optimization path planning algorithm, remarkable results have been achieved in improving the inspection efficiency and reducing the risk of transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0040] Figure 1 It is a flowchart of the method of the present application;

[0041] Figure 2 It is a flowchart of path planning for the unmanned aerial vehicle for transmission lines for the selected tower part. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings in the embodiments of the present application. It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0043] The term "comprising", "including" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising a..." does not exclude the presence of additional identical elements in the process, method, article or device comprising the element.

[0044] Terms such as "first", "second", etc. are only used to distinguish one entity or operation from another entity or operation, and cannot be construed as indicating or implying relative importance, nor can they be construed as requiring or implying any actual relationship or order between these entities or operations.

[0045] Please refer to Figure 1 , a method for UAV inspection of transmission line directivity tasks combined with contrastive active learning, mainly including the following steps:

[0046] (1) Collect point cloud data within the UAV inspection range of the transmission line;

[0047] (2) Pre-train the point cloud active selection annotation model based on contrastive learning;

[0048] (3) Import the collected point cloud data of UAV inspection of the transmission line into the model for tower part annotation;

[0049] (4) Select the tower part to be inspected and issue an inspection task;

[0050] (5) Plan the UAV path of the transmission line for the selected tower part.

[0051] A method for UAV inspection of transmission line directivity tasks combined with contrastive active learning uses an airborne lidar of the UAV to scan the transmission line and collect three-dimensional point cloud data of the transmission line.

[0052] The pre-training of the point cloud active selection annotation model based on contrastive learning includes the following processes:

[0053] (1) Unsupervised feature extraction pre-training based on contrastive learning. For unlabeled point cloud data, firstly, effective features are extracted through the feature extraction module; then the extracted features are constrained by using the information of the data itself to achieve effective pre-training of the model; after the training is completed, the model parameters are saved;

[0054] (2) Active labeling of unlabeled data based on active learning. The model trained in step (1) is used as input for unlabeled point cloud data. After extracting point cloud features, based on the features obtained and the calculated pseudo-labels, some data is selected from all unlabeled point clouds for manual labeling according to the design selection index.

[0055] (3) After executing the above steps, a round of model training and data labeling is completed; then the labeled data and the remaining unlabeled data are input into step (1) again to repeat the above process. At this time, not only the contrastive learning loss is calculated, but also the cross entropy loss of the labeled data is calculated, the model is updated, and the unlabeled data is input into step (2) again. This process is repeated until the maximum number of labels is reached or the obtained model meets the accuracy requirements.

[0056] Unsupervised feature extraction pre-training based on contrastive learning adopts the MoCo design learning framework, which mainly includes the following processes:

[0057] (1) First, the input data is transformed and then input into the upper and lower two identical encoder branches for feature extraction;

[0058] (2) Then, feature projection is performed through a multi-layer perceptron, and contrastive learning is performed in this feature space. This method achieves the maintenance of a large number of negative examples by constructing a queue, thereby improving the performance of contrastive learning;

[0059] (3) Finally, the upper branch is updated by gradient back propagation, while the parameters of the lower branch are updated by momentum. Different transformations of the same input data constitute positive pairs, while other data constitute negative pairs. In the feature space, the feature distance between positive examples is shortened and the feature distance between negative examples is extended, thus realizing the training process of the model.

[0060] Contrasting learning pre-training, set the input data to Where N is the number of points in the input model, and each point is represented by its coordinates in space; construct a transformation set {T i} is used to transform the coordinates of the input point cloud.

[0061] The upper and lower two branches h1 and h2 of the feature extraction part process P1 and P2 respectively, and the two branches adopt exactly the same structure; then the point-by-point features extracted are input into the feature space projection modules M1 and M2 with the same structure, and further projected into the feature space to obtain normalized features. The whole process uses formula (1):

[0062]

[0063] Calculate PointInfo Loss in the normalized feature space and perform backpropagation to train the model. The parameters θ1 of the upper branch h1 and M1 are updated through backpropagation; according to the structure of MoCo, the parameters θ2 of the lower branch h2 and M2 are updated by momentum according to θ1, and the formula is:

[0064] θ2←mθ2+(1 - m)θ1 (2)

[0065] Among them, m represents the momentum update parameter. Fix the model parameters obtained after training for subsequent data selection and annotation in the active learning process. The specific form of this loss function is:

[0066]

[0067] Among them, P o represents the set of positive example pairs composed of the original point cloud and the transformed point cloud; N e represents the set of negative example pairs stored in the MoCo queue structure; the subscripts i and j represent different transformed features of the same point cloud; f i and f j constitute positive examples, f i and f k constitute negative examples; τ represents the temperature parameter, which is used to control the proportion of positive and negative examples participating in the calculation.

[0068] If there are labeled tags in the input point cloud data, then this part of the point cloud is continued to be input into the classification module, which is composed of a multi-layer fully connected network, and the normalized features are regressed to obtain the classification probability of this point. This part uses cross-entropy for loss calculation, and the formula is:

[0069]

[0070] Among them, K represents the number of categories in the dataset; y ij represents the sign function, when sample i belongs to category j, y ij =1, otherwise y ij =0; p ij represents the predicted probability that this point i belongs to category j. After training, the classification model parameters obtained from this part of the training are also fixed for subsequent use.

[0071] Select the inspection tower part to issue the inspection task, apply the trained model to the transmission line, identify the tower part of the transmission line, and remotely issue the inspection instruction for the specific tower part.

[0072] Perform path planning for the UAV of the transmission line for the selected tower part. It is characterized in that an improved slime mold optimization algorithm is used for path planning, which mainly includes the following processes:

[0073] (1) Set the population size, maximum number of iterations and related parameters;

[0074] (2) Initialize the population using the improved Logistic chaotic map, calculate the worst fitness of the population and sort it;

[0075] (3) Judge whether the random number r < c. If it is satisfied, the slime mold approaches the food, and a non-linear adaptive weight factor is added during the process. If it is not satisfied, go to step 5;

[0076] (4) Calculate the population fitness and update the optimal position;

[0077] (5) Sort according to the different fitness values each time;

[0078] (6) Perform adaptive Cauchy compilation on the current position, calculate the fitness, and press the optimal position;

[0079] (7) Whether the maximum number of iterations is reached. If not, repeat the above steps until the condition is satisfied.

[0080] The present invention has achieved remarkable results in improving the inspection efficiency and reducing the risk of transmission lines by combining the point cloud annotation technology of contrastive learning and active learning and the improved slime mold optimization path planning algorithm. The specific application examples and effects are as follows:

[0081] Taking a local power supply company (with about 200 kilometers of transmission lines under its jurisdiction, including 800 towers) as an example, the efficiency differences between the present invention and manual directed task inspection and full-scale UAV autonomous inspection are compared from a technical level:

[0082] 1. Comparison of technical solutions

[0083]

[0084] 2. Analysis of technical advantages

[0085] (1) Efficient annotation technology for point cloud data

[0086] Contrastive Learning Pretraining: Through unsupervised feature extraction (MoCo framework), the model automatically learns the tower structure features using unlabeled point cloud data, reducing the dependence on manual labeling. Technical Effect: After pre-training, the accuracy of feature extraction for key components such as insulators and fittings is increased to 92%, and the amount of manual labeling is reduced by 70%. Active Learning Screening Mechanism: Based on pseudo-label confidence (Formula 3) and uncertainty sampling, the point cloud region with the largest amount of information is preferentially labeled. Technical Effect: The labeling efficiency is compressed from 6 hours / km (full-scale labeling) to 1.5 hours / km, and the labeling error rate is reduced from 8% to 3%. (2) Improved Slime Mould Optimization Path Planning Algorithm

[0087] Chaotic Initialization and Adaptive Weights: An improved Logistic chaotic map is used to generate the initial population (the formula involves non-linear factors) to avoid the traditional algorithm falling into local optima; an adaptive Cauchy mutation is added to enhance the global search ability.

[0088] Technical Effect: The path planning time is shortened from 10 minutes / tower to 2 minutes / tower, and the average length of the planned path is reduced by 18%.

[0089] Dynamic Obstacle Avoidance Mechanism: Combining point cloud data to generate a three-dimensional map of obstacles in real time, the algorithm dynamically adjusts the weight factor (Formula 2) to ensure that the safe distance between the drone and the wire > 1.5 meters.

[0090] 3. Examples of Efficiency Improvement (Data of Prefecture-level Power Companies)

[0091] In a certain prefecture-level power company for the task of "lightning strike tower grounding device defect inspection", the comparison results are as follows:

[0092] Index Manual directional inspection Full-scale autonomous inspection The method of the present invention Total time consumption for a single task 120 hours 45 hours 18 hours Defect recognition rate 78％ 88％ 96％ Frequency of manual intervention 100％ 30％ 5％ Energy consumption for path planning - High (full route coverage) Low (directional optimization

[0093] Note:

[0094] Comparison of Time Consumption: Through targeted labeling and path optimization, the time consumption of this invention is only 15% of manual inspection and 40% of full-scale inspection.

[0095] Improvement in Recognition Rate: By directionally analyzing the point cloud data of the arrester connection part, the recognition rate is increased by 8 percentage points compared with full-scale inspection.

[0096] 4. Comprehensive Technical Value

[0097] Economy: The annual inspection cost of prefecture-level power companies is reduced by 42%, mainly due to the savings in manual labeling and path planning resources.

[0098] Accuracy: Through the tower component features extracted by contrastive learning (such as the normalized projection in Formula 1), the cross-entropy loss of the classification model is reduced by 30% (Formula 4).

[0099] Scalability: It supports multi-machine collaborative inspection. The slime mold algorithm can perform parallel computing (Steps S12 - S17), and the task allocation efficiency is increased by 50%.

[0100] Through the "intelligent annotation - targeted planning" technology closed-loop, the present invention solves the small-scale and high-precision inspection requirements of local power companies, and is significantly superior to traditional methods in terms of efficiency, accuracy, and cost.

[0101] An embodiment of the present application provides a UAV inspection system for transmission line directivity tasks, including a memory and a processor. The memory includes a program for the UAV inspection method for transmission line directivity tasks. When the program for the UAV inspection method for transmission line directivity tasks is executed by the processor, the following steps are implemented: collecting point cloud data within the UAV inspection range of the transmission line; pre-training a point cloud active selection annotation model based on contrastive learning; importing the collected point cloud data of the UAV inspection of the transmission line into the model for tower part annotation; selecting the inspected tower part to issue an inspection task; and performing UAV path planning for the selected tower part of the transmission line.

[0102] An embodiment of the present application provides a computer-readable storage medium. The computer-readable storage medium stores program code. When the program code is executed by a processor, the steps of the UAV inspection method for transmission line directivity tasks as described above are implemented.

[0103] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0105] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction means that implements the function specified in one or more of the processes and / or blocks Figure 1 in one or more of the processes and / or blocks Figure 1 specified in one or more of the blocks or blocks.

[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, such that a series of operational steps are performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in one or more of the processes and / or blocks Figure 1 in one or more of the processes and / or blocks Figure 1 specified in one or more of the blocks or blocks.

[0107] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0108] The memory may include non-permanent memory in the computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of a computer-readable medium.

[0109] Computer-readable media includes permanent and non-permanent, removable and non-removable media implemented by any method or technology for information storage. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0110] The above are only embodiments of the present application and are not intended to limit the protection scope of the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A method for inspecting transmission lines by a directional task unmanned aerial vehicle, characterized in that, It includes the following steps: Collect the point cloud data within the range of UAV inspection for transmission lines; Pre-train the point cloud active selection annotation model based on contrastive learning; Import the collected point cloud data of UAV inspection for transmission lines into the model for tower part annotation; Select the inspection tower part and issue the inspection task; Plan the UAV path for the selected tower part of the transmission line.

2. The inspection method of the transmission line directivity task unmanned aerial vehicle according to claim 1, characterized in that The specific method for collecting the point cloud data within the range of UAV inspection for transmission lines is to use the on-board lidar of the UAV to scan the transmission line and collect the three-dimensional point cloud data of the transmission line.

3. The method for inspecting a transmission line by a directional task unmanned aerial vehicle according to claim 1, wherein The specific method for pre-training the point cloud active selection annotation model based on contrastive learning is as follows: S1. Unsupervised feature extraction pre-training based on contrastive learning. For the unlabeled point cloud data, first extract the effective features through the feature extraction module; then use the information of the data itself to constrain the extracted features to achieve effective pre-training of the model; Save the model parameters at this time after the training ends; S2. Active annotation of unlabeled data based on active learning. Use the trained model to take the unlabeled point cloud data as input; After extracting the point cloud features, based on the obtained features and the calculated pseudo-labels, select some data from all the unlabeled point clouds for manual annotation according to the designed selection index; After performing the above steps, complete one round of model training and data annotation; then input the labeled data and the remaining unlabeled data into step S1 to repeat the above process. At this time, not only calculate the contrastive learning loss, but also calculate the cross-entropy loss of the labeled data to update the model, and continue to input the unlabeled data into step S2 to execute. This process alternates until the maximum number of annotations is reached or the obtained model meets the accuracy requirements.

4. The inspection method of the directional task unmanned aerial vehicle for the transmission line according to claim 3, characterized in that The unsupervised feature extraction pre-training based on contrastive learning is specifically as follows. The input data is where N is the number of points in the input model, and each point is represented by its coordinates in space; construct a set of transformations {T i} for transforming the coordinates of the input point cloud; The upper and lower two branches h1 and h2 of the feature extraction part process P1 and P2 respectively, and the two branches adopt exactly the same structure; then input the extracted point-by-point features into the feature space projection modules M1 and M2 with the same structure, and further project them into the feature space to obtain the normalized features. The whole process uses formula (1): Calculate the PointInfo Loss in the normalized feature space and perform backpropagation to train the model. The parameters θ1 of the upper branch h1 and M1 are updated through backpropagation; according to the structure of MoCo, the parameters θ2 of the lower branch h2 and M2 are updated by momentum according to θ1, and the formula is: θ2←mθ2+(1-m)θ1(2) where m represents the momentum update parameter. Fix the model parameters obtained after training for selecting data for annotation in the subsequent active learning process. The specific form of this loss function is: Among them, P o represents the set of positive example pairs composed of the origin point cloud and the transformed point cloud; N e represents the set of negative example pairs stored in the MoCo queue structure; the subscripts i and j represent different transformation features of the same point cloud; f i and f j constitute a positive example, and f i and f k constitute a negative example; τ represents the temperature parameter, which is used to control the proportion of positive and negative examples participating in the calculation. If there are labeled labels in the input point cloud data, continue to input this part of the point cloud into the classification module, which is composed of a multi-layer fully connected network, and regress the normalized features to obtain the classification probability of this point. The cross-entropy is used for loss calculation in this part, and the formula is: Among them, K represents the number of categories in the dataset; y ij represents the sign function, where y ij = 1 when sample i belongs to category j, otherwise y ij = 0; p ij represents the predicted probability that point i belongs to category j obtained through calculation. After training, the classification model parameters obtained by fixing this part are also used for subsequent use.

5. The inspection method of a transmission line directional task unmanned aerial vehicle according to claim 1, characterized in that Select the inspection tower pole part and issue the inspection task. Specifically, apply the trained model to the transmission line to identify the tower pole part of the transmission line, and remotely issue the inspection instruction for the specific tower pole part.

6. The inspection method of a power transmission line directivity task unmanned aerial vehicle according to claim 1, wherein The path planning of the UAV for the selected tower pole part of the transmission line is to use the improved slime mold optimization algorithm for path planning, which mainly includes the following processes: S11. Set the population size, the maximum number of iterations and related parameters; S12. Initialize the population using the improved Logistic chaotic map, calculate and sort the optimal and worst fitness of the population; S13. Judge whether the random number r < c. If it is satisfied, the slime mold approaches the food, and a non-linear adaptive weight factor is added during the process. If it is not satisfied, execute step S15; S14. Calculate the population fitness and update the optimal position; S15. Sort according to the different fitness each time; S16. Perform adaptive Cauchy compilation on the current position, calculate the fitness, and press the optimal position; S17. Whether the maximum number of iterations is reached. If not, repeat the above steps until the conditions are met.

7. A directional task unmanned aerial vehicle inspection system for transmission lines, characterized in that, It includes a memory and a processor. The memory includes a program for the UAV inspection method for the directional task of the transmission line. When the program for the UAV inspection method for the directional task of the transmission line is executed by the processor, the following steps are implemented: Collect the point cloud data within the inspection range of the UAV for the transmission line; Pre-train the point cloud active selection and annotation model based on contrastive learning; Import the collected point cloud data of the UAV inspection for the transmission line into the model for tower pole part annotation; Select the inspection tower pole part and issue the inspection task; Perform path planning of the UAV for the selected tower pole part of the transmission line.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program codes. When the program codes are executed by the processor, the steps of the UAV inspection method for the directional task of the transmission line as described in any one of claims 1 to 6 are implemented.

Images