Unmanned aerial vehicle wire threading erection control system based on image recognition and path planning

Abstract

The application discloses a UAV wire threading erection control system based on image recognition and path planning, relates to the technical field of UAV wire erection, and a visual perception module converts real-time images into three-dimensional space data, thereby obtaining three-dimensional perception capability. min A line machine state extraction module calculates a state vector F containing the shortest spatial distance d opt and the total hanging point tension T, so as to realize accurate quantification of a safety margin and a UAV load. A risk potential field establishment module establishes a composite risk potential field scalar Φ according to the state vector F, so as to unify the evaluation standard of collision and bearing capacity risks. A candidate trajectory deduction and optimal control decision module performs bimodal simulation based on the composite risk potential field scalar Φ, and selects an optimal path planning instruction a opt through a cost function, so as to realize forward-looking profit-seeking and hazard-avoiding decision and emergency treatment. A flight control instruction generation module converts the instruction into a bottom-layer PWM signal to drive the UAV, so as to improve the safety and efficiency of operation in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of UAV power line installation technology, specifically to a UAV power line installation control system based on image recognition and path planning. Background Technology

[0002] With the continuous advancement of urbanization, the construction and upgrading of urban power and communication network infrastructure are becoming increasingly frequent. Among these tasks, laying overhead power lines along both sides of urban roads is a common engineering project. A core aspect of this work is pulling and erecting new cables between poles or supports. To improve operational efficiency and reduce the impact on urban traffic, especially when laying lightweight cables such as fiber optic cables or low-voltage power lines, engineers aim to eliminate the traditional step of using guide ropes and instead use aerial equipment to directly pull the power lines. Unmanned aerial vehicles (UAVs), with their flexibility and high maneuverability, provide a feasible technical path for performing such tasks. To ensure the safe flight of UAVs in complex urban environments, especially among the common green trees along roadsides, intelligent planning of their flight paths is essential. The prerequisite for intelligent planning is that the system can perceive the environment accurately and in real time, which requires the application of image recognition technology. Therefore, a system that integrates image recognition and path planning to control UAVs for power line erection constitutes the research direction of this technical solution.

[0003] Currently, the method of using drones to directly pull power lines along urban roads for installation relies primarily on remote manual control by operators. This method has several inherent technical drawbacks: First, the trees lining urban roads vary in shape and have dense foliage, making it difficult for operators to accurately determine the three-dimensional spatial gap between the power line and branches using only two-dimensional video transmitted from the drone, thus lacking quantitative information on collision safety margins. Second, because the guide rope is omitted and the power line is pulled directly, even lightweight power lines have a much greater weight per unit length than the guide rope, placing higher demands on the drone's load-bearing capacity; however, operators cannot obtain real-time information on the specific tension value of the power line applied to the drone's attachment point during flight, leaving the core load status unknown. Finally, the entire control process is reactive; operators cannot predict how maneuvering to avoid a branch during climbing or turning commands will alter the catenary shape of the power line, potentially causing a sudden surge in tension.

[0004] The aforementioned technical deficiencies can lead to a series of abnormal effects in urban operating environments. Because the tension at the attachment points cannot be monitored in real time, when the traction cable accidentally snags on dense tree branches, the tension is instantly transmitted to the drone. If this tension, combined with the weight of the cable, exceeds the drone's maximum payload, the flight control system will be unable to provide sufficient lift, causing the drone to be unable to continue its flight commands, or even resulting in a sudden drop in altitude, loss of control, or a crash, posing a significant safety hazard to vehicles and pedestrians below. Furthermore, the lack of precise three-dimensional spatial perception capabilities makes it highly susceptible to misjudgment when operators attempt to navigate through gaps in the tree canopy, leading to direct collisions between the cable and branches. This can range from minor damage to the cable insulation and electrical hazards to serious issues like the cable or drone getting caught, causing the entire installation operation to fail and requiring post-installation cleanup in the complex urban environment, resulting in even greater time and economic costs. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a UAV power cable threading and erection control system based on image recognition and path planning, which solves the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a UAV power line threading and erection control system based on image recognition and path planning, including a visual perception module, a UAV state extraction module, a risk potential field establishment module, a candidate trajectory deduction module, an optimal control decision module, and a flight control command generation module;

[0007] The visual perception module acquires real-time two-dimensional images through the drone's camera and converts them into three-dimensional spatial data;

[0008] The wire state extraction module receives the converted three-dimensional spatial data, performs geometric and physical parameter calculations, and obtains a state vector containing the current state parameters of the wire and the UAV.

[0009] The risk potential field establishment module receives the state vector to assess the tension risk and collision risk of the wire, and obtains the risk potential field set.

[0010] The candidate trajectory deduction module receives the risk potential field set for risk assessment, performs simulation prediction based on the assessment results and the preset candidate action set, and re-substitutes them into the assessment calculation of the risk potential field scalar to obtain the candidate trajectory assessment dataset.

[0011] The optimal control decision module receives the candidate trajectory evaluation dataset, iterates through and calculates the combined cost of the risk term brought by the risk potential field scalar and the benefit term brought by the future state vector, and selects the action corresponding to the minimum value as the optimal path planning instruction.

[0012] The flight control command generation module converts the optimal path planning command into executable underlying physical commands and drives the UAV to execute the optimal path planning command.

[0013] Preferably, the visual perception module includes an environmental 3D mapping unit and a wire 3D morphology reconstruction unit;

[0014] The environmental 3D mapping unit uses the drone's overhead camera to collect the first video stream data below the drone before the power line installation. This data is synchronized with the flight speed data acquired by the onboard inertial measurement unit. The unit runs the ORB algorithm on consecutive image frames to automatically identify feature points. Feature point matching is performed between adjacent image frames to find the projection of the same spatial point in different images. Based on the matched feature points and the drone's flight speed, the Kalman filter algorithm is used to calculate the drone's position and attitude in 3D space in real time. Combining triangulation principles, the 2D feature points observed in images at different locations are back-projected to obtain their true coordinates in 3D space, constructing a 3D point cloud map of the environment. This results in an obstacle point cloud set P describing the surface shape of all static obstacles within the observation range of the drone. OBS .

[0015] Preferably, the three-dimensional shape reconstruction unit of the power line uses the rear-view camera of the wire threading drone and the airborne inertial measurement unit to collect the second video stream data behind the wire threading drone during the wire threading operation. By observing the front-view camera of the observation drone, an observation plane is formed at the position of the wire threading drone hanging point, the position of the starting tower hanging point of the power line, and the natural descent direction of the power line. The midpoint of the horizontal distance between the hanging point of the wire threading drone and the hanging point of the starting tower hanging point of the power line is used as a marker point to generate the normal vector of the observation plane, so that the observation drone is kept on the normal vector of the observation plane, and the third video stream data of the position of the wire threading drone and the descent trajectory of the power line are collected.

[0016] Based on the second video stream data, the U-Net image segmentation network is used to separate the wire portion from the background in the latest frame of the second video stream image, generating a binary mask image. A centerline fitting algorithm is then used to generate a centerline representing the wire on the two-dimensional image plane. The pixels on the centerline are converted into an ordered sequence of pixel coordinates. Based on the third video stream data, the U-Net image segmentation network is used to separate the wire portion from the background in the latest frame of the third video stream image, generating a catenary curve formed by gravity. The catenary curve is projected onto the two-dimensional image plane and compared with the pixel coordinate sequence converted from the centerline. Through optimization algorithms, a three-dimensional catenary curve that best matches the projection result with the actual observation result is generated, resulting in a point cloud set P containing obstacles. OBSUnder the same world coordinate system, an ordered sequence of three-dimensional spatial coordinates P of the power line describing the spatial curve from the power line attachment point of the UAV to the power line starting tower attachment point. ROP =[p r1 p r2 ... p rN ], where N represents the three-dimensional spatial coordinate sequence P of the wire. ROP The total number of coordinate points in p r1 p represents the coordinates of the attachment point for the drone. rN Indicates the coordinates of the starting tower attachment point for the power line;

[0017] Combined with obstacle point cloud set P OBS With the three-dimensional spatial coordinate sequence P of the wire ROP Together they constitute the 3D scene dataset D.

[0018] Preferably, the wire state extraction module includes a spatial set relationship parsing unit and a wire physical state parsing unit;

[0019] The spatial set relation parsing unit receives a 3D scene dataset D and, based on the obstacle point cloud set P in the 3D scene dataset D... OBS Using a kd-number construction algorithm, an obstacle kd-tree is generated, and the three-dimensional spatial coordinate sequence P of the wire is traversed. ROP For each 3D coordinate point, initiate a nearest neighbor query to the obstacle kd-tree to find the current 3D spatial coordinate sequence P of the wire. ROP The system iterates through the nearest obstacle point corresponding to each traversal point and calculates the Euclidean distance d between the two points. This process is repeated after traversing the three-dimensional spatial coordinate sequence P of the wire. ROP Then, by comparing all the Euclidean distances d, the shortest spatial distance d is found. min ;

[0020] Based on the obtained shortest spatial distance d min The corresponding points are the target points p of the wire. r and obstacle point p o Calculate the target point p of the wire r Pointing to obstacle point p o The normalized vector yields the danger direction vector d. vec The specific calculation method is to calculate the target point p of the wire. r With obstacle point p o The coordinate difference is divided by the target point p of the wire. r With obstacle point p o The distance between them.

[0021] Preferably, the wire physical state analysis unit receives the wire three-dimensional spatial coordinate sequence P from the three-dimensional scene dataset D. ROP First, set the coordinates p of the starting tower suspension point of the power line.rN Generate a simulated vertical line segment in the vertical direction to guide the coordinates p of the UAV mounting point. r1 Using the generated simulated vertical line segment as a reference point, draw a horizontal perpendicular line to obtain the spatial line segment. Finally, traverse the three-dimensional spatial coordinate sequence P of the power line. ROP Calculate the perpendicular distance S from each coordinate to the defined spatial line segment, and find the maximum perpendicular distance S by comparing all perpendicular distances S. MAX Marked as sag depth S d and the maximum vertical distance S MAX The corresponding coordinate point is marked as the maximum perpendicular point p. rM Next, calculate the coordinates p of the attachment point of the drone. r1 and the maximum perpendicular point p rM The projected distance on the horizontal plane is marked as the horizontal segment span L. d The specific calculation method is to calculate the square root of the square of the difference between the X-axis coordinates of the two points and the square of the difference between the Y-axis coordinates, according to the Pythagorean theorem.

[0022] Based on the horizontal segmented span Ld and sag depth S d Using the formula for calculating the total tension of a catenary in physics, the total tension T at the hanging point is calculated. First, the tension T is calculated from the point of maximum verticality p. rM The horizontal tension component generated by extending to both ends, wherein the horizontal tension component is related to the horizontal segment span L. d It is proportional to the square of the sag depth S d Inversely proportional, secondly, the calculation is based on the attachment point p of the pilot drone. r1 The vertical tension component that needs to be withstood due to the weight of the wire, wherein the vertical tension component is related to the distance from the attachment point p of the threaded drone. r1 to the maximum perpendicular point p rM Since the weight of some wires is equal, the square root of the sum of the squares of the horizontal and vertical tension components is then taken to obtain the result at the point p where the drone is attached. r1 Total tension T at the hanging point;

[0023] The formula for calculating the total tension T at the hanging point is as follows:

[0024]

[0025] In the formula, w represents the preset weight of the wire per unit length, which is obtained through pre-measurement and is expressed in Newtons per meter.

[0026] Combining the preset flight velocity vector v of the penetrating drone and the shortest spatial distance d min Danger direction vector d vec Given the total tension T at the hanging point, generate the state vector F of the current state parameters of the power line and the drone.

[0027] Preferably, the risk potential field establishment module includes a tension and spatial risk assessment unit;

[0028] The tension and space risk assessment unit constructs a composite risk model combining wire tension risk and space collision risk by receiving the state vector F of the current state parameters of the power line and the UAV. First, for the wire tension risk, the quadratic cost function concept from control theory is used to compare the total tension T at the attachment point with the preset ideal load force T of the UAV. idl The deviation is normalized and squared. Secondly, the dynamic concept of a preset prediction time step t is incorporated into the spatial collision risk. This is achieved by calculating the distance to be approached and the shortest spatial distance d within the prediction time step t. min The ratio of the two values ​​is used to obtain an index of the collision risk between the wire and the obstacle. Finally, by adding the wire tension risk and the spatial collision risk together, a composite risk potential field scalar Φ that comprehensively reflects the wire tension and spatial collision is obtained.

[0029] The formula for calculating the scalar Φ of the composite risk potential field is as follows:

[0030]

[0031] In the formula, T idl The preset ideal payload capacity for the drone is set by relevant personnel based on the wiring materials and industry safety regulations. (T) tol This represents the preset safety drone payload capacity fluctuation threshold, based on the project construction technical specifications and the preset ideal drone payload capacity T. idl Here, t represents the preset prediction time step, and max represents taking the maximum parameter within the function.

[0032] Preferably, the candidate trajectory deduction module includes a candidate trajectory deduction and evaluation unit;

[0033] The candidate trajectory extrapolation and evaluation unit receives the composite risk potential field scalar Φ and compares it with the preset risk potential field threshold Φ. th For comparison, if the composite risk potential scalar Φ < the preset risk potential threshold Φ th If the target value is not specified, the system will not enter risk adjustment mode and will maintain the preset UAV flight speed vector v. Otherwise, it will enter risk adjustment mode, whereby the observation UAV issues a stop-layout command to the ground-based cable laying equipment and performs simulation prediction based on a preset set of candidate actions. The candidate action set A includes a flight speed adjustment set A1 and a flight direction adjustment set A2. The digital mappings of the adjustment commands in the flight speed adjustment set A1 and the flight direction adjustment set A2 are traversed and combined to obtain the complete, computer-executable i-th path planning command a. i =(A 1i A 2i );

[0034] According to the path planning instruction a i By performing state simulation on the pilot-operated UAV, the predicted velocity vector New_v of the i-th pilot-operated UAV is obtained after a preset command execution time step t. i And the i-th pilot drone prediction attachment point New_p r1i Based on the i-th pilot drone's predicted attachment point New_p r1i , Coordinates of the starting tower suspension point p rN The length L of the wire during cable locking is used to obtain a new three-dimensional spatial coordinate sequence New_P for the i-th predicted wire, based on the catenary physical simulation model. ROPi Substituting this into Module 2, we obtain the i-th predicted future state vector, New_F. i The i-th predicted future state vector New_F i Substituting into Module 3, we obtain the i-th predicted composite risk potential scalar New_Φ. i and the corresponding i-th path planning instruction a i Data pairing is performed, and after traversing and simulating all candidate actions, the candidate trajectory evaluation dataset G = {(a1, New_Φ1), (a2, New_Φ2), ..., (a...} is obtained. i New_Φ i ), ..., (a 23 New_Φ 23 ), (a 24 New_Φ 24 )}.

[0035] Preferably, the optimal control decision module includes a candidate action cost calculation unit and an optimal control command selection unit;

[0036] The candidate action cost calculation unit receives the candidate trajectory evaluation dataset G, and by traversing the candidate trajectory evaluation dataset G and combining optimal control theory, uses the predicted i-th composite risk potential field scalar New_Φ. i With the preset risk potential threshold Φ th Perform division to obtain the i-th path planning instruction a i The normalized security cost, combined with the predicted velocity vector New_v of the i-th penetrating UAV. i The preset drone flight speed vector v and the preset unit vector d of the destination pylon attachment point direction. B Calculate path planning instruction a i The normalized efficiency gain is used to obtain the execution path planning instruction a. i Total cost C i ;

[0037] Among them, the total cost C iThe calculation expression is as follows:

[0038]

[0039] In the formula, d B This represents a unit vector indicating the preset direction of the final hanging point on the tower, set according to the positions of the starting and ending hanging points. |d B | represents the unit vector d indicating the preset direction of the endpoint hanging point on the tower. B The model;

[0040] After traversing and simulating all composite risk potential scalars New_Φ, the total cost dataset H = {(a1, C1), (a2, C2), ..., (a...} is obtained. i C i ), ..., (a 23 C 23 ), (a 24 C 24 )}.

[0041] Preferably, the optimal control instruction selection unit receives the total cost dataset H, uses the minimum value search method to traverse all data pairs in the total cost dataset H, obtains the item with the minimum total cost C, and extracts the corresponding candidate action a as the optimal path planning instruction a. opt .

[0042] Preferably, the flight control command generation module includes a flight control command generation unit;

[0043] The flight control command generation unit receives the optimal path planning command a opt The drone is guided according to the optimal path planning instruction a opt A PWM signal with a corresponding duty cycle is generated for the electronic speed controller connected to each motor, driving the pilot flight to execute the optimal path planning command a. opt It achieves the target flight attitude.

[0044] This invention provides a UAV power cable threading and erection control system based on image recognition and path planning, which has the following advantages:

[0045] (1) During system operation, real-time 3D perception, multi-target risk assessment, forward-looking trajectory simulation, and optimal control decision-making are systematically integrated to replace the traditional remote control operation method that relies on the operator's personal experience with data-driven automated closed-loop control. This not only significantly improves the safety of UAVs in complex environments where environmental cleanup is not possible, such as power line erection operations in urban streets, avoiding flight loss of control and collision accidents caused by overload or spatial misjudgment, but also ensures the stability and reliability of flight actions. Ultimately, while ensuring construction safety and quality, it effectively improves the level of automation and operational efficiency.

[0046] (2) Through the visual perception module, the system can construct an obstacle point cloud set P containing static obstacles. OBS And the precise three-dimensional spatial coordinate sequence P of dynamic wires ROP This gives it a quantitative three-dimensional spatial perception capability that surpasses the human eye, fundamentally solving the problem that operators cannot accurately judge spatial margins due to the limitations of a two-dimensional perspective. Based on this three-dimensional scene dataset D, the line machine state extraction module can accurately analyze the shortest spatial distance d between the guide rope and the obstacle. min This includes the real-time changes in the total tension T at the drone's attachment point. This addresses the core pain point of operators being unable to accurately perceive safety margins and actual loads, providing objective and accurate data input for subsequent risk assessments. Furthermore, the risk potential field establishment module utilizes these precise state vectors F to quantify collision risks and load-bearing capacity risks of different physical dimensions into a dimensionless composite risk potential field scalar Φ through a composite risk model. This enables the system to, for the first time, possess a comprehensive and real-time judgment capability regarding the degree of danger of the current operational state, laying a solid foundation for achieving forward-looking intelligent decision-making.

[0047] (3) After acquiring accurate situational awareness capabilities, based on a precise assessment of the current risks, the candidate trajectory inference module can reach the risk potential field threshold Φ when the composite risk potential field scalar Φ is reached. th At that time, the system actively triggers a locking command and switches to a dedicated emergency simulation model to perform parallel simulations of the future consequences of various evasive actions in the candidate action set A. This forward-looking simulation capability enables the system to anticipate and select the optimal response strategy, avoiding the reactive, manual remedial approach of traditional methods. Based on a detailed candidate trajectory evaluation dataset G, the optimal control decision module rationally evaluates all possibilities using a cost function that balances safety costs and efficiency gains, ensuring the final selection of the optimal path planning command a. opt It represents the optimal solution that maximizes benefits and minimizes risks under the current conditions, achieving an intelligent dynamic balance between ensuring safety and completing the mission. Finally, the flight control command generation module generates this high-level optimal path planning command a optThis is precisely translated into physical commands that can be executed by the underlying hardware, forming a complete closed loop from perception, assessment, prediction, decision-making to execution. This ensures that the UAV can stably and reliably execute the optimal flight path and ultimately complete the task of erecting power lines in complex urban environments safely and efficiently. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the UAV power cable threading and erection control system based on image recognition and path planning according to the present invention.

[0049] Figure 2 This is a schematic diagram of the data processing flow of the UAV power line threading and erection control system based on image recognition and path planning according to the present invention;

[0050] Figure 3 This is a scalar data graph of the composite risk potential field Φ. Detailed Implementation

[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0052] Example 1

[0053] This invention provides a UAV power cable threading and erection control system based on image recognition and path planning. Please refer to [link / reference]. Figure 1 It includes a visual perception module, a flight state extraction module, a risk potential field establishment module, a candidate trajectory deduction module, an optimal control decision module, and a flight control command generation module;

[0054] The visual perception module acquires real-time two-dimensional images through the drone's camera and converts them into three-dimensional spatial data;

[0055] The wire state extraction module receives the converted three-dimensional spatial data, performs geometric and physical parameter calculations, and obtains a state vector containing the current state parameters of the wire and the UAV.

[0056] The risk potential field establishment module receives the state vector to assess the tension risk and collision risk of the wire, and obtains the risk potential field set.

[0057] The candidate trajectory deduction module receives the risk potential field set for risk assessment, performs simulation prediction based on the assessment results and the preset candidate action set, and re-substitutes them into the assessment calculation of the risk potential field scalar to obtain the candidate trajectory assessment dataset.

[0058] The optimal control decision module receives the candidate trajectory evaluation dataset, iterates through and calculates the combined cost of the risk term brought by the risk potential field scalar and the benefit term brought by the future state vector, and selects the action corresponding to the minimum value as the optimal path planning instruction.

[0059] The flight control command generation module converts the optimal path planning command into executable underlying physical commands and drives the UAV to execute the optimal path planning command.

[0060] In this embodiment, the system first uses a visual perception module and a drone camera to collect real-time images of the green trees and power lines on both sides of the city road, and then converts them into a point cloud set P containing obstacles. OBS and the three-dimensional spatial coordinate sequence P of the wire ROP The precise three-dimensional spatial data enables it to achieve quantitative three-dimensional spatial perception, solving the problem that operators cannot accurately judge spatial margins due to the limitations of a two-dimensional perspective. Based on this, the wire machine status extraction module receives this three-dimensional data and, through efficient spatial analysis and physical model calculation, accurately calculates the shortest spatial distance d between the wire and the tree branch. min The step of precisely quantifying the core safety and load status, which is imperceptible to the operator, along with the total tension T acting on the drone's attachment points, provides objective and accurate data input for subsequent intelligent decision-making. Furthermore, the risk potential field establishment module, based on the aforementioned quantified state parameters and combined with the preset ideal drone mounting force T, [further details on this step are needed]. idl Based on engineering standards, a composite risk potential field scalar Φ is established and calculated, unifying the two different types of risks—collision and load-bearing capacity—into a single assessment dimension. This endows the system with a comprehensive and real-time ability to assess the current level of danger. Subsequently, the candidate trajectory deduction module receives this risk value and compares it with the risk potential field threshold Φ. th A comparison was made. Since the risk value was below the threshold, the system determined the state was safe and did not enter the risk adjustment mode. However, under the assumption that the risk exceeded the threshold, this module could proactively trigger the locking of the line and switch to the emergency simulation model. This forward-looking simulation capability effectively avoided the problem of delayed response in manual operation when an accident occurred. Next, based on detailed simulation results, the optimal control decision module rationally evaluated the situation using a cost function that balanced safety costs and efficiency gains, ultimately selecting the optimal path planning instruction 'a' for the current state. optThis system achieves an intelligent dynamic balance between ensuring flight safety and completing the towing task, completely replacing the operator's subjective judgment. Finally, the flight control command generation module accurately translates this high-level optimal decision into low-level PWM signals to drive the UAV to execute. The entire process forms a complete closed loop from perception, evaluation, prediction, decision-making to execution, ensuring that the UAV can safely and stably execute the optimal flight path, ultimately efficiently completing the power line erection task in the complex urban environment and ensuring a high level of standardization in the operation.

[0061] Example 2

[0062] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: the visual perception module includes an environmental 3D mapping unit and a wire 3D morphology reconstruction unit;

[0063] The environmental 3D mapping unit uses the drone's overhead camera to collect the first video stream data below the drone before the power line installation. This data is synchronized with the flight speed data acquired by the onboard inertial measurement unit. The unit runs the ORB algorithm on consecutive image frames to automatically identify feature points. Feature point matching is performed between adjacent image frames to find the projection of the same spatial point in different images. Based on the matched feature points and the drone's flight speed, the Kalman filter algorithm is used to calculate the drone's position and attitude in 3D space in real time. Combining triangulation principles, the 2D feature points observed in images at different locations are back-projected to obtain their true coordinates in 3D space, constructing a 3D point cloud map of the environment. This results in an obstacle point cloud set P describing the surface shape of all static obstacles within the observation range of the drone. OBS ;

[0064] The three-dimensional morphology reconstruction unit of the power line uses the rear-view camera of the wire threading drone and the airborne inertial measurement unit to collect the second video stream data behind the wire threading drone during the wire threading operation. By observing the front-view camera of the observation drone, an observation plane is formed at the position of the wire threading drone hanging point, the position of the starting tower hanging point of the power line, and the natural descent direction of the power line. The midpoint of the horizontal distance between the hanging point of the wire threading drone and the hanging point of the starting tower of the power line is used as a marker point to generate the normal vector of the observation plane, so that the observation drone is kept on the normal vector of the observation plane, and the third video stream data of the position of the wire threading drone and the trajectory of the power line descent are collected.

[0065] Based on the second video stream data, the U-Net image segmentation network is used to separate the wire portion from the background in the latest frame of the second video stream image, generating a binary mask image. A centerline fitting algorithm is then used to generate a centerline representing the wire on the two-dimensional image plane. The pixels on the centerline are converted into an ordered sequence of pixel coordinates. Based on the third video stream data, the U-Net image segmentation network is used to separate the wire portion from the background in the latest frame of the third video stream image, generating a catenary curve formed by gravity. The catenary curve is projected onto the two-dimensional image plane and compared with the pixel coordinate sequence converted from the centerline. Through optimization algorithms, a three-dimensional catenary curve that best matches the projection result with the actual observation result is generated, resulting in a point cloud set P containing obstacles. OBS Under the same world coordinate system, an ordered sequence of three-dimensional spatial coordinates P of the power line describing the spatial curve from the power line attachment point of the UAV to the power line starting tower attachment point. ROP =[p r1 p r2 ... p rN ], where N represents the three-dimensional spatial coordinate sequence P of the wire. ROP The total number of coordinate points in p r1 p represents the coordinates of the attachment point for the drone. rN Indicates the coordinates of the starting tower attachment point for the power line;

[0066] The U-Net image segmentation network employs a classic encoder-decoder architecture. The encoder consists of five consecutive convolutional modules, each containing two 3x3 convolutional layers using ReLU as the activation function, followed by a 2x2 max-pooling layer for downsampling. The decoder, correspondingly, consists of five upsampling modules. Upsampling is achieved through transposed convolutions, and skip connections are used to concatenate the feature maps of the corresponding encoder layers with those of the current decoder layer, fusing deep semantic information and shallow spatial details. The final output layer is a 1x1 convolutional layer, using a sigmoid activation function to output a pixel-level binary segmentation mask image. During training, the network uses a weighted sum of the Dice loss function and the binary cross-entropy loss function as its optimization target loss function to address the issue of uneven pixel proportions in the image.

[0067] The optimization objective function of the optimization algorithm is defined as minimizing the root mean square error between the predicted projection and the actual observation. Specifically, for a given 3D catenary model parameter, the system projects it onto a 2D image plane to obtain a set of predicted pixel coordinate sequences. At the same time, it extracts the actual center line pixel coordinate sequence from the mask image. The objective function is the square root of the mean of the sum of squares of the Euclidean distances between all corresponding points in these two sets of sequences. The optimization process uses the Levenberg-Marquardt algorithm for iterative optimization. The iteration termination condition is that the number of iterations reaches the preset maximum iteration limit. The preset maximum iteration limit is set according to the performance of the processor carried by the UAV.

[0068] Combined with obstacle point cloud set P OBS With the three-dimensional spatial coordinate sequence P of the wire ROP Together they constitute the 3D scene dataset D.

[0069] In this embodiment, firstly, an observation drone is instructed to conduct a preliminary reconnaissance flight along the operation path via an environmental 3D mapping unit. The first video stream data collected by its overhead camera is processed by the ORB algorithm and the Kalman filter algorithm to construct a static obstacle point cloud set P containing the precise contours of all roadside trees. OBS Decoupling the two core tasks of environmental perception and cable laying in time offers a unique advantage: all subsequent dynamic analysis and path planning are based on a pre-built, interference-free, high-precision static 3D map, significantly improving the stability of environmental data. Subsequently, when the cable laying drone begins its towing operation, the cable 3D morphology reconstruction unit is activated. At this point, the observation drone dynamically moves to the preset observation plane normal direction, acquiring third video stream data of the overall cable morphology from the optimal side view, and coordinating it with the second video stream data acquired by the rear-view camera of the cable laying drone. The system fuses the visual information from these two different perspectives through U-Net image segmentation network and 3D catenary model fitting optimization, ultimately generating a high-precision 3D spatial coordinate sequence P of the cable. ROP This multi-drone, multi-view collaborative working mode offers the unique advantage of perfectly solving the technical challenges of narrow rear-view angles and inaccurate shape recognition due to complex background interference in single-drone systems. It significantly improves the accuracy of the three-dimensional spatial coordinate sequence P of power lines. ROP The accuracy and robustness of the reconstruction. Ultimately, the results generated by these two units together constitute a 3D scene dataset D, which is a high-fidelity digital twin scene, providing a basis for subsequent modules to accurately calculate the shortest spatial distance d. min The system provides a more reliable and unambiguous data foundation than a stand-alone system, along with core safety parameters such as the total tension T at the hanging point.

[0070] Example 3

[0071] This embodiment is an explanation based on Embodiment 2. Please refer to it. Figure 1 Specifically: the wire machine status extraction module includes a spatial set relationship parsing unit and a wire physical state parsing unit;

[0072] The spatial set relation parsing unit receives a 3D scene dataset D and, based on the obstacle point cloud set P in the 3D scene dataset D... OBS Using a kd-number construction algorithm, an obstacle kd-tree is generated, and the three-dimensional spatial coordinate sequence P of the wire is traversed. ROP For each 3D coordinate point, initiate a nearest neighbor query to the obstacle kd-tree to find the current 3D spatial coordinate sequence P of the wire. ROP The system iterates through the nearest obstacle point corresponding to each traversal point and calculates the Euclidean distance d between the two points. This process is repeated after traversing the three-dimensional spatial coordinate sequence P of the wire. ROP Then, by comparing all the Euclidean distances d, the shortest spatial distance d is found. min ;

[0073] Based on the obtained shortest spatial distance d min The corresponding points are the target points p of the wire. r and obstacle point p o Calculate the target point p of the wire r Pointing to obstacle point p o The normalized vector yields the danger direction vector d. vec The specific calculation method is to calculate the target point p of the wire. r With obstacle point p o The coordinate difference is divided by the target point p of the wire. r With obstacle point p o The distance between them;

[0074] The wire physical state analysis unit receives the wire 3D spatial coordinate sequence P from the 3D scene dataset D. ROP First, set the coordinates p of the starting tower suspension point of the power line. rN Generate a simulated vertical line segment in the vertical direction to guide the coordinates p of the UAV mounting point. r1 Using the generated simulated vertical line segment as a reference point, draw a horizontal perpendicular line to obtain the spatial line segment. Finally, traverse the three-dimensional spatial coordinate sequence P of the power line. ROP Calculate the perpendicular distance S from each coordinate to the defined spatial line segment, and find the maximum perpendicular distance S by comparing all perpendicular distances S. MAX Marked as sag depth S d and the maximum vertical distance S MAX The corresponding coordinate point is marked as the maximum perpendicular point p. rM Next, calculate the coordinates p of the attachment point of the drone. r1and the maximum perpendicular point p rM The projected distance on the horizontal plane is marked as the horizontal segment span L. d The specific calculation method is to calculate the square root of the square of the difference between the X-axis coordinates of the two points and the square of the difference between the Y-axis coordinates, according to the Pythagorean theorem.

[0075] Based on the horizontal segmented span Ld and sag depth S d Using the formula for calculating the total tension of a catenary in physics, the total tension T at the hanging point is calculated. First, the tension T is calculated from the point of maximum verticality p. rM The horizontal tension component generated by extending to both ends, wherein the horizontal tension component is related to the horizontal segment span L. d It is proportional to the square of the sag depth S d Inversely proportional, secondly, the calculation is based on the attachment point p of the pilot drone. r1 The vertical tension component that needs to be withstood due to the weight of the wire, wherein the vertical tension component is related to the distance from the attachment point p of the threaded drone. r1 to the maximum perpendicular point p rM Since the weight of some wires is equal, the square root of the sum of the squares of the horizontal and vertical tension components is then taken to obtain the result at the point p where the drone is attached. r1 Total tension T at the hanging point;

[0076] The formula for calculating the total tension T at the hanging point is as follows:

[0077]

[0078] In the formula, w represents the preset weight of the wire per unit length, which is obtained through pre-measurement and is expressed in Newtons per meter.

[0079] Combining the preset flight velocity vector v of the penetrating drone and the shortest spatial distance d min Danger direction vector d vec Given the total tension T at the hanging point, generate the state vector F of the current state parameters of the power line and the drone.

[0080] In this embodiment, the obstacle point cloud set P is received. OBS and the three-dimensional spatial coordinate sequence P of the wire ROP The three-dimensional scene dataset D is jointly formed. Firstly, to improve computational efficiency, the spatial set relationship parsing unit within the module constructs an algorithm based on the kd-tree to analyze the obstacle point cloud set P. OBS Spatial indexing is performed, and then the three-dimensional spatial coordinate sequence P of the wire is traversed. ROP And perform a nearest neighbor query to calculate the shortest spatial distance d between the current power line and the obstacle. min and its corresponding danger direction vector d vecBy employing the efficient kd-tree data structure, the complex spatial safety search is transformed into a fast query, ensuring that the system can update the shortest spatial distance d frequently and in real time. min This provides the necessary time for immediate risk perception and avoidance. Simultaneously, the wire physical state analysis unit begins processing the wire's three-dimensional spatial coordinate sequence P. ROP To calculate the drone's payload. Considering the drone is in a climbing attitude, this unit traverses the three-dimensional spatial coordinate sequence P of the power line. ROP Directly locate the actual maximum perpendicular point p rM Based on this, the horizontal segmented span L on the UAV side was calculated. d and sag depth S d Subsequently, based on the physical principles of catenary cables and the weight w of the wire per unit length, the unit calculated the drone's attachment point p by vector synthesis of the horizontal and vertical tension components. r1 The total tension T borne by the hanging point. This is determined by directly locating the actual maximum vertical point p. rM Furthermore, asymmetric parameters are calculated to ensure highly accurate total tension T at the attachment point, even under complex flight conditions such as climb and descent. This precisely quantifies the real-time load of the UAV, providing a solid foundation for subsequent load prediction and risk assessment. Finally, this module integrates the calculated key parameters and outputs a result including the shortest spatial distance d. min Danger direction vector d vec The state vector F of the total tension T at the hanging point efficiently and accurately extracts a raw and complex three-dimensional visual scene into a low-dimensional structured state vector F, where each component has a clear physical meaning. This provides a high-quality, high-information-density digital foundation for all subsequent risk assessments and intelligent decisions.

[0081] Example 4

[0082] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 and Figure 3 Specifically: the risk potential field establishment module includes a tension and spatial risk assessment unit;

[0083] The tension and space risk assessment unit constructs a composite risk model combining wire tension risk and space collision risk by receiving the state vector F of the current state parameters of the power line and the UAV. First, for the wire tension risk, the quadratic cost function concept from control theory is used to compare the total tension T at the attachment point with the preset ideal load force T of the UAV. idl The deviation is normalized and squared. Secondly, the dynamic concept of a preset prediction time step t is incorporated into the spatial collision risk. This is achieved by calculating the distance to be approached and the shortest spatial distance d within the prediction time step t.min The ratio of the two values ​​is used to obtain an index of the collision risk between the wire and the obstacle. Finally, by adding the wire tension risk and the spatial collision risk together, a composite risk potential field scalar Φ that comprehensively reflects the wire tension and spatial collision is obtained.

[0084] The formula for calculating the scalar Φ of the composite risk potential field is as follows:

[0085]

[0086] In the formula, T idl This indicates the preset ideal payload capacity for the drone, set by relevant personnel based on the drone's performance, wiring materials, and industry safety regulations. (T) tol This represents the preset safe drone payload capacity fluctuation threshold, based on project construction technical specifications, drone performance, and the preset ideal drone payload capacity T. idl The setting, specifically, is the preset ideal payload capacity T of the drone. idl The safety redundancy threshold T for the payload capacity fluctuation of the safe drone is 5% to 15%. tol t represents the preset prediction time step, and max represents taking the maximum parameter within the function;

[0087] The candidate trajectory deduction module includes a candidate trajectory deduction and evaluation unit;

[0088] The candidate trajectory extrapolation and evaluation unit receives the composite risk potential field scalar Φ and compares it with the preset risk potential field threshold Φ. th For comparison, if the composite risk potential scalar Φ < the preset risk potential threshold Φ th If the target value is not specified, the system will not enter risk adjustment mode and will maintain the preset UAV flight speed vector v. Otherwise, it will enter risk adjustment mode, whereby the observation UAV issues a stop-layout command to the ground-based cable laying equipment and performs simulation prediction based on a preset set of candidate actions. The candidate action set A includes a flight speed adjustment set A1 and a flight direction adjustment set A2. The digital mappings of the adjustment commands in the flight speed adjustment set A1 and the flight direction adjustment set A2 are traversed and combined to obtain the complete, computer-executable i-th path planning command a. i =(A 1i A 2i );

[0089] Among them, the flight speed adjustment set A1 includes three speed adjustment amounts: an increase of 1 m / s, an increase of 2 m / s, and an increase of 3 m / s, which are digitally mapped to 1 to 3 respectively. The flight direction adjustment set A2 includes clockwise rotation along the vertical plane of the preset flight direction, with each 45° increment being a standard, and a total of eight direction adjustment amounts, which are digitally mapped to 1 to 8 respectively.

[0090] According to the path planning instruction ai By performing state simulation on the pilot-operated UAV, the predicted velocity vector New_v of the i-th pilot-operated UAV is obtained after a preset command execution time step t. i And the i-th pilot drone prediction attachment point New_p r1i Based on the i-th pilot drone's predicted attachment point New_p r1i , Coordinates of the starting tower suspension point p rN The length L of the wire during cable locking is used to obtain a new three-dimensional spatial coordinate sequence New_P for the i-th predicted wire, based on the catenary physical simulation model. ROPi Substituting this into Module 2, we obtain the i-th predicted future state vector, New_F. i The i-th predicted future state vector New_F i Substituting into Module 3, we obtain the i-th predicted composite risk potential scalar New_Φ. i and the corresponding i-th path planning instruction a i Data pairing is performed, and after traversing and simulating all candidate actions, the candidate trajectory evaluation dataset G = {(a1, New_Φ1), (a2, New_Φ2), ..., (a...} is obtained. i New_Φ i ), ..., (a 23 New_Φ 23 ), (a 24 New_Φ 24 )};

[0091] The following is an example of drone path planning and control:

[0092] Task: To lay power lines between roads, with a total span of 100 meters;

[0093] Cable parameters: The weight w of the wire per unit length is 2.5 N / m;

[0094] Ideal payload capacity of UAV (T) idl =400N, Safety Drone Load Capacity Fluctuation Threshold T tol =50N, prediction time step t=3s;

[0095] Risk potential field threshold Φ th =1;

[0096] The current state vector F is calculated in real time using the visual perception module and the linear machine state extraction module.

[0097] Total tension at hanging point T = 430 N, shortest spatial distance d min =25m, danger direction vector d vec = (0.1, 0, -0.995);

[0098] The UAV's flight velocity vector is v = (3.0, 0, 0) m / s;

[0099] Danger direction vector d vec = (0, 1.0, 0);

[0100] The following is an example of calculating the scalar Φ of a complex risk potential field:

[0101]

[0102] The candidate trajectory inference module receives a composite risk potential scalar Φ = 0.36399 and compares it with the preset risk potential threshold Φ. th =1 for comparison, the composite risk potential scalar Φ < the preset risk potential threshold Φ th It will not enter risk adjustment mode.

[0103] In this embodiment, upon receiving the state vector F, the system quantifies and calculates the current comprehensive composite risk potential field scalar Φ, using information such as the total tension T at the hanging point and the shortest spatial distance dmin. This mathematical unification and quantification of two physically different hazard sources enables the system to comprehensively and systematically measure complex comprehensive risks. Subsequently, the candidate trajectory deduction module receives this risk value and compares it with the preset risk potential field threshold Φ. th A comparison is performed. If the current risk exceeds the limit, the system immediately enters risk adjustment mode and automatically executes the built-in emergency response protocol: it first issues a stop-layout command to the ground-based cable-laying equipment via the observation drone, and marks the current total cable length as the fixed cable length L when locking the cable, thereby proactively changing the physical constraints of the problem to prevent the risk from worsening. In this emergency mode, the system continues to process each path planning instruction a in the candidate action set A. i Simulations were performed to deduce the predicted attachment point New_p of the drone corresponding to a certain command. r1i Based on a fixed wire length L during cable locking, a novel three-dimensional spatial coordinate sequence New_P for the predicted wire is calculated using a catenary physical simulation model. ROPi The system recalculates this prediction result to obtain the predicted composite risk potential scalar New_Φ corresponding to this action. i This process elevates the decision-making basis from the current state to the future consequences by calculating the future risk consequences for each candidate action. Finally, after traversing all candidate actions, a candidate trajectory evaluation dataset G containing various avoidance actions and their corresponding quantified risks is generated, providing comprehensive data support for subsequent modules to select the optimal escape strategy.

[0104] Example 5

[0105] This embodiment is an explanation based on Embodiment 4. Please refer to it. Figure 1 and Figure 3 Specifically: the optimal control decision module includes a candidate action cost calculation unit and an optimal control command selection unit;

[0106] The candidate action cost calculation unit receives the candidate trajectory evaluation dataset G, and by traversing the candidate trajectory evaluation dataset G and combining optimal control theory, uses the predicted i-th composite risk potential field scalar New_Φ. i With the preset risk potential threshold Φ th Perform division to obtain the i-th path planning instruction a i The normalized security cost, combined with the predicted velocity vector New_v of the i-th penetrating UAV. i The preset drone flight speed vector v and the preset unit vector d of the destination pylon attachment point direction. B Calculate path planning instruction a i The normalized efficiency gain is used to obtain the execution path planning instruction a. i Total cost C i ;

[0107] Among them, the total cost C i The calculation expression is as follows:

[0108]

[0109] In the formula, d B This represents a unit vector indicating the preset direction of the final hanging point on the tower, set according to the positions of the starting and ending hanging points. |d B | represents the unit vector d indicating the preset direction of the endpoint hanging point on the tower. B The model;

[0110] After traversing and simulating all composite risk potential scalars New_Φ, the total cost dataset H = {(a1, C1), (a2, C2), ..., (a...} is obtained. i C i ), ..., (a 23 C 23 ), (a 24 C 24 )};

[0111] The optimal control instruction selection unit receives the total cost dataset H, and uses the minimum value search method to traverse all data pairs in the total cost dataset H, obtaining the item with the minimum total cost C and extracting the corresponding candidate action a as the optimal path planning instruction a. opt ;

[0112] The flight control command generation module includes a flight control command generation unit;

[0113] The flight control command generation unit receives the optimal path planning command a opt The drone is guided according to the optimal path planning instruction a opt A PWM signal with a corresponding duty cycle is generated for the electronic speed controller connected to each motor, driving the pilot flight to execute the optimal path planning command a. opt To achieve the target flight attitude;

[0114] Example comparing traditional drone remote control and traditional manual operation:

[0115]

[0116] In this embodiment, a candidate trajectory evaluation dataset G is received. The unit then iterates through each pair of data in the dataset and applies the i-th total cost C. i The calculation expression is to convert the predicted composite risk potential field scalar New_Φ corresponding to the i-th action. i With risk potential threshold Φ th The normalized safety cost is obtained by division, and then combined with the predicted velocity vector New_v of the penetrating UAV. i The unit vector d in the direction of the final hanging point of the tower B Normalized efficiency gains are calculated using parameters such as safety and efficiency. This provides a rational and quantifiable decision-making framework by normalizing both safety and efficiency costs. The total cost C of the i-th action is then calculated. i This transforms the abstract goals of safety and efficiency into mathematical scalars that can be compared on the same dimension, thus replacing the fuzzy process of operators relying on subjective intuition for trade-offs. After calculating the total cost dataset H containing all candidate actions and their corresponding costs, the optimal control instruction selection unit within the module receives this dataset and performs a minimum value search method on it. By comparing the total cost C of all items in the dataset one by one, the item with the lowest cost is finally determined, and its corresponding candidate action is extracted as the optimal path planning instruction a. opt This ensures that the system selects the theoretically optimal solution from all calculated possibilities, achieving the best response to the current predicament, rather than simply providing a feasible, ordinary solution. Finally, the flight control command generation module receives this unique optimal path planning command a. opt It converts the signals into high-fidelity underlying PWM signals required to drive each motor of the drone, ensuring the accuracy of decision execution and avoiding the disconnect between decision and execution caused by delays or errors in manual operation. This ultimately forms a complete closed loop from perception, evaluation, prediction, decision-making to execution, ensuring that the drone can stably and reliably execute the optimal flight path.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A UAV power cable threading and erection control system based on image recognition and path planning, characterized in that: It includes a visual perception module, a flight state extraction module, a risk potential field establishment module, a candidate trajectory deduction module, an optimal control decision module, and a flight control command generation module; The visual perception module acquires real-time two-dimensional images through the drone's camera and converts them into three-dimensional spatial data; The wire state extraction module receives the converted three-dimensional spatial data, performs geometric and physical parameter calculations, and obtains a state vector containing the current state parameters of the wire and the UAV. The risk potential field establishment module receives the state vector to assess the tension risk and collision risk of the wire, and obtains the risk potential field set. The candidate trajectory deduction module receives the risk potential field set for risk assessment, performs simulation prediction based on the assessment results and the preset candidate action set, and re-substitutes them into the assessment calculation of the risk potential field scalar to obtain the candidate trajectory assessment dataset. The optimal control decision module receives the candidate trajectory evaluation dataset, iterates through and calculates the combined cost of the risk term brought by the risk potential field scalar and the benefit term brought by the future state vector, and selects the action corresponding to the minimum value as the optimal path planning instruction. The flight control command generation module converts the optimal path planning command into executable underlying physical commands and drives the UAV to execute the optimal path planning command. The wire machine status extraction module includes a spatial set relationship parsing unit and a wire physical state parsing unit; Combined with obstacle point cloud set P OBS With the three-dimensional spatial coordinate sequence P of the wire ROP Together they constitute a 3D scene dataset D. The spatial set relation parsing unit receives the 3D scene dataset D and, based on the obstacle point cloud set P in the 3D scene dataset D... OBS Using a kd-number construction algorithm, an obstacle kd-tree is generated, and the three-dimensional spatial coordinate sequence P of the wire is traversed. ROP For each 3D coordinate point, initiate a nearest neighbor query to the obstacle kd-tree to find the current 3D spatial coordinate sequence P of the wire. ROP The system iterates through the nearest obstacle point corresponding to each traversal point and calculates the Euclidean distance d between the two points. This process is repeated after traversing the three-dimensional spatial coordinate sequence P of the wire. ROP Then, by comparing all the Euclidean distances d, the shortest spatial distance d is found. min ; Based on the obtained shortest spatial distance d min The corresponding points are the target points p of the wire. r and obstacle point p o Calculate the target point p of the wire r Pointing to obstacle point p o The normalized vector yields the danger direction vector d. vec The specific calculation method is to calculate the target point p of the wire. r With obstacle point p o The coordinate difference is divided by the target point p of the wire. r With obstacle point p o The distance between them; The wire physical state analysis unit receives the wire 3D spatial coordinate sequence P from the 3D scene dataset D. ROP First, set the coordinates p of the starting tower suspension point of the power line. rN Generate a simulated vertical line segment in the vertical direction to guide the coordinates p of the UAV mounting point. r1 Using the generated simulated vertical line segment as a reference point, draw a horizontal perpendicular line to obtain the spatial line segment. Finally, traverse the three-dimensional spatial coordinate sequence P of the power line. ROP Calculate the perpendicular distance S from each coordinate to the defined spatial line segment, and find the maximum perpendicular distance S by comparing all perpendicular distances S. MAX Marked as sag depth S d and the maximum vertical distance S MAX The corresponding coordinate point is marked as the maximum perpendicular point p. rM Next, calculate the coordinates p of the attachment point of the drone. r1 and the maximum perpendicular point p rM The projected distance on the horizontal plane is marked as the horizontal segment span L. d The specific calculation method is to calculate the square root of the square of the difference between the X-axis coordinates of the two points and the square of the difference between the Y-axis coordinates, according to the Pythagorean theorem. According to the horizontal segmented span L d and sag depth S d Using the formula for calculating the total tension of a catenary in physics, the total tension T at the hanging point is calculated. First, the tension T is calculated from the point of maximum verticality p. rM The horizontal tension component generated by extending to both ends, wherein the horizontal tension component is related to the horizontal segment span L. d It is proportional to the square of the sag depth S d Inversely proportional, secondly, the calculation is based on the attachment point p of the pilot drone. r1 The vertical tension component that needs to be withstood due to the weight of the wire, wherein the vertical tension component is related to the distance from the attachment point p of the threaded drone. r1 to the maximum perpendicular point p rM Since the weight of some wires is equal, the square root of the sum of the squares of the horizontal and vertical tension components is then taken to obtain the result at the attachment point p of the drone. r1 Total tension T at the hanging point; The formula for calculating the total tension T at the hanging point is as follows: ； In the formula, w represents the preset weight of the wire per unit length, which is obtained through pre-measurement and is expressed in Newtons per meter. Combining the preset flight velocity vector v of the penetrating drone and the shortest spatial distance d min Danger direction vector d vec Given the total tension T at the hanging point, generate the state vector F of the current state parameters of the power line and the drone.

2. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 1, characterized in that: The visual perception module includes an environmental 3D mapping unit and a wire 3D morphology reconstruction unit; The environmental 3D mapping unit uses the drone's overhead camera to collect the first video stream data below the drone before the power line installation. This data is synchronized with the flight speed data acquired by the onboard inertial measurement unit. The unit runs the ORB algorithm on consecutive image frames to automatically identify feature points. Feature point matching is performed between adjacent image frames to find the projection of the same spatial point in different images. Based on the matched feature points and the drone's flight speed, the Kalman filter algorithm is used to calculate the drone's position and attitude in 3D space in real time. Combining triangulation principles, the 2D feature points observed in images at different locations are back-projected to obtain their true coordinates in 3D space, constructing a 3D point cloud map of the environment. This results in an obstacle point cloud set P describing the surface shape of all static obstacles within the observation range of the drone. OBS .

3. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 2, characterized in that: The three-dimensional morphology reconstruction unit of the power line uses the rear-view camera of the wire threading drone and the airborne inertial measurement unit to collect the second video stream data behind the wire threading drone during the wire threading operation. By observing the front-view camera of the observation drone, an observation plane is formed at the position of the wire threading drone hanging point, the position of the starting tower hanging point of the power line, and the natural descent direction of the power line. The midpoint of the horizontal distance between the hanging point of the wire threading drone and the hanging point of the starting tower of the power line is used as a marker point to generate the normal vector of the observation plane, so that the observation drone is kept on the normal vector of the observation plane, and the third video stream data of the position of the wire threading drone and the trajectory of the power line descent are collected. Based on the second video stream data, the U-Net image segmentation network is used to separate the wire portion from the background in the latest frame of the second video stream image, generating a binary mask image. A centerline fitting algorithm is then used to generate a centerline representing the wire on the two-dimensional image plane. The pixels on the centerline are converted into an ordered sequence of pixel coordinates. Based on the third video stream data, the U-Net image segmentation network is used to separate the wire portion from the background in the latest frame of the third video stream image, generating a catenary curve formed by gravity. The catenary curve is projected onto the two-dimensional image plane and compared with the pixel coordinate sequence converted from the centerline. Through optimization algorithms, a three-dimensional catenary curve that best matches the projection result with the actual observation result is generated, resulting in a point cloud set P containing obstacles. OBS Under the same world coordinate system, an ordered sequence of three-dimensional spatial coordinates P of the power line describing the spatial curve from the attachment point of the unmanned aerial vehicle (UAV) to the attachment point of the power line's starting tower. ROP =[p r1 p r2 ... p rN ], where N represents the three-dimensional spatial coordinate sequence P of the wire. ROP The total number of coordinate points in p r1 p represents the coordinates of the attachment point for the drone. rN This indicates the coordinates of the starting tower attachment point for the power line.

4. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 1, characterized in that: The risk potential field establishment module includes a tension and spatial risk assessment unit; The tension and space risk assessment unit constructs a composite risk model combining wire tension risk and space collision risk by receiving the state vector F of the current state parameters of the power line and the UAV. First, for the wire tension risk, the quadratic cost function concept from control theory is used to compare the total tension T at the attachment point with the preset ideal load force T of the UAV. idl The deviation is normalized and squared. Secondly, the dynamic concept of a preset prediction time step t is incorporated into the spatial collision risk. This is achieved by calculating the distance to be approached and the shortest spatial distance d within the prediction time step t. min The ratio of the two values ​​is used to obtain an index of the collision risk between the wire and the obstacle. Finally, by adding the wire tension risk and the spatial collision risk together, a composite risk potential field scalar Φ that comprehensively reflects the wire tension and spatial collision is obtained. The formula for calculating the scalar Φ of the composite risk potential field is as follows: ； In the formula, T idl The preset ideal payload capacity for the drone is set by relevant personnel based on the wiring materials and industry safety regulations. (T) tol This represents the preset safety drone payload capacity fluctuation threshold, based on the project construction technical specifications and the preset ideal drone payload capacity T. idl Here, t represents the preset prediction time step, and max represents taking the maximum parameter within the function.

5. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 4, characterized in that: The candidate trajectory deduction module includes a candidate trajectory deduction and evaluation unit; The candidate trajectory extrapolation and evaluation unit receives the composite risk potential field scalar Φ and compares it with the preset risk potential field threshold Φ. th For comparison, if the composite risk potential scalar Φ < the preset risk potential threshold Φ th If the target value is not specified, the system will not enter risk adjustment mode and will maintain the preset UAV flight speed vector v. Otherwise, it will enter risk adjustment mode, whereby the observation UAV issues a stop-layout command to the ground-based cable laying equipment and performs simulation prediction based on a preset set of candidate actions. The candidate action set A includes a flight speed adjustment set A1 and a flight direction adjustment set A2. The digital mappings of the adjustment commands in the flight speed adjustment set A1 and the flight direction adjustment set A2 are traversed and combined to obtain the complete, computer-executable i-th path planning command a. i = (A 1i A 2i ); According to the path planning instruction a i By performing state simulation on the pilot-operated UAV, the predicted velocity vector New_v of the i-th pilot-operated UAV is obtained after a preset command execution time step t. i And the i-th pilot drone prediction attachment point New_p r1i Based on the i-th pilot drone's predicted attachment point New_p r1i , Coordinates of the starting tower suspension point p rN The length L of the wire during cable locking is used to obtain a new three-dimensional spatial coordinate sequence New_P for the i-th predicted wire, based on the catenary physical simulation model. ROPi Then, by substituting the vectors into the line machine state extraction module, the i-th predicted future state vector New_F is obtained. i The i-th predicted future state vector New_F i Substituting into the risk potential field establishment module, we obtain the i-th predicted composite risk potential field scalar New_Φ. i and the corresponding i-th path planning instruction a i Data pairing is performed, and after traversing and simulating all candidate actions, the candidate trajectory evaluation dataset G={(a1, New_Φ1), (a2, New_Φ2), ..., (a... i New_Φ i ), ..., (a 23 New_Φ 23 (a 24 New_Φ 24 )}.

6. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 5, characterized in that: The optimal control decision module includes a candidate action cost calculation unit and an optimal control command selection unit; The candidate action cost calculation unit receives the candidate trajectory evaluation dataset G, and by traversing the candidate trajectory evaluation dataset G and combining optimal control theory, uses the predicted i-th composite risk potential field scalar New_Φ. i With the preset risk potential threshold Φ th Perform division to obtain the i-th path planning instruction a i The normalized security cost, combined with the predicted velocity vector New_v of the i-th penetrating UAV. i The preset drone flight speed vector v and the preset unit vector d of the destination pylon attachment point direction. B Calculate path planning instruction a i The normalized efficiency gain is used to obtain the execution path planning instruction a. i Total cost C i ; Among them, the total cost C i The calculation expression is as follows: ； In the formula, d B This represents a unit vector indicating the preset direction of the final hanging point on the tower, set according to the positions of the starting and ending hanging points. |d B | represents the unit vector d indicating the preset direction of the endpoint hanging point on the tower. B The model; After traversing and simulating all composite risk potential scalars New_Φ, the total cost dataset H = {(a1, C1), (a2, C2), ..., (a...} is obtained. i C i ), ..., (a 23 C 23 (a 24 C 24 )}.

7. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 6, characterized in that: The optimal control instruction selection unit receives the total cost dataset H, and uses the minimum value search method to traverse all data pairs in the total cost dataset H, obtaining the item with the minimum total cost C and extracting the corresponding candidate action a as the optimal path planning instruction a. opt .

8. The UAV power cable threading and erection control system based on image recognition and path planning according to claim 7, characterized in that: The flight control command generation module includes a flight control command generation unit; The flight control command generation unit receives the optimal path planning command a opt The drone is guided according to the optimal path planning instruction a opt A PWM signal with a corresponding duty cycle is generated for the electronic speed controller connected to each motor, driving the drone to execute the optimal path planning command a. opt It achieves the target flight attitude.

Images