A drone-assisted precision spraying maintenance system for cable maintenance

Abstract

This invention relates to a drone-based collaborative precision spraying maintenance system for cable maintenance, specifically in the industrial application field of multi-drone collaborative operations. Through dynamic modeling, the system quantifies the physical collisions, airflow disturbances, and paint contamination risks of drone operations into a time-varying risk field, providing a quantitative benchmark for collaborative decision-making. Based on this, an initial trajectory without hard conflicts, balancing efficiency and risk, is planned in four-dimensional spacetime, and a safety corridor is generated. Online scheduling utilizes graph neural networks to perceive the global collaborative state in real time, driving multiple drones to dynamically fine-tune to proactively avoid interference. Simultaneously, forward-looking digital twin simulation continuously verifies and predicts; once a deviation is detected, compensation control is immediately triggered, forming a safety closed loop. By comparing predicted and measured data, the model and strategy are continuously optimized, thereby simultaneously achieving safety and reliability, uniform and clean coating, and improved overall operational efficiency in close multi-drone collaborative operations.

Description

Technical Field

[0001] This invention relates to the field of industrial applications of multi-drone collaborative operation, and more specifically, to a drone collaborative precision spraying maintenance system for cable maintenance. Background Technology

[0002] In the preventive maintenance and repair of large bridge cable systems, traditional manual methods using suspended scaffolds or mobile work vehicles suffer from numerous drawbacks, including low efficiency, high safety risks, and significant traffic disruption. With the maturity of drone technology, using drones equipped with spraying devices for high-altitude operations has become a promising alternative. However, when faced with complex spatial networks such as those in suspension bridges (composed of main cables, numerous suspenders, and densely packed cable clamps) or the confined spaces and intertwined cables around the towers of cable-stayed bridges, the capabilities of a single drone are insufficient. To improve the overall efficiency of maintenance operations, it is essential to employ a swarm of multiple drones working collaboratively. These drones need to traverse closely in three-dimensional space, each undertaking precise spraying tasks around specific cable sections or components. This operational scenario is not... Beyond complex spatial path planning, this scenario introduces dynamic and interconnected operational conflicts. There is a direct risk of physical collisions between drones and between drones and the rigid structure of the bridge. More complexly, the painting operation itself introduces new dimensions of interference: paint droplets sprayed by one drone may drift due to unstable airflow induced by its own rotor or those of adjacent drones, contaminating nearby precision components that should not be painted, such as sensors or metal mating surfaces at the cable anchoring ends. Simultaneously, the downdraft generated by the high-speed rotating rotors may disrupt the paint mist being formed by another drone, resulting in uneven deposition on the cable surface and severely impacting coating quality. Therefore, the requirements for multi-drone systems in this scenario far exceed simple spatial obstacle avoidance; they necessitate a deep collaboration encompassing flight trajectories, operational timing, and the impact of actions.

[0003] Existing technologies involving multi-mobile agent path planning, such as those applied in warehousing and logistics or agricultural plant protection, primarily focus on planning collision-free movement paths for multiple individuals in two-dimensional or three-dimensional space to achieve area coverage. These technologies typically treat the agent as a purely moving point mass or an object with a simple geometric envelope, with the core of their planning being resolving spatial location conflicts. However, when directly applied to collaborative spraying scenarios in cable networks, significant shortcomings exist. First, existing planning methods generally ignore the continuous impact of the specific task actions performed by the agent on the shared environment. In spraying operations, the drone is not a static geometric body; its "spraying action" continuously generates a dynamically changing "impact envelope." This envelope includes the spatial range of the paint mist cone and the spatial area potentially affected by rotor airflow disturbances. If only the path of the drone itself is planned without incorporating this operational impact envelope into the constraints of collaborative planning, then… This can easily lead to situations where, even without physical collision, the spraying operation of one drone can contaminate the clean surface that another drone is about to operate on, or its rotor airflow can severely interfere with the spray deposition process of neighboring drones, causing functional failure. Secondly, existing methods mostly focus on conflict resolution in the spatial dimension, lacking joint optimization with the task state in the temporal dimension. In dense cable networks, simple spatial path planning often leads to multiple drones getting stuck in narrow passages or critical nodes, or waiting for each other for a long time, because the algorithm does not consider "sharing" critical spatial resources by precisely scheduling the start and stop times of spraying actions, resulting in low overall operational efficiency. Therefore, current technology lacks a collaborative planning framework that can uniformly manage the spatial, temporal, and task domains, and cannot simultaneously optimize the quality of spraying operations and the overall efficiency of multi-drone systems while ensuring absolute safety. This has become the core bottleneck restricting the large-scale application of this technology in highly complex real-world scenarios. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a drone-based collaborative precision spraying maintenance system for cable maintenance. This system utilizes an influence envelope modeling module, a global spatiotemporal planning module, an online collaborative scheduling module, and a closed-loop verification execution module to solve the problems mentioned in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: an influence envelope modeling module, a global spatiotemporal planning module, an online collaborative scheduling module, and a closed-loop verification execution module connected in sequence; Impact envelope modeling module: Before collaborative operation, based on the rotor parameters, spraying parameters and environmental wind field data of the UAV, a corresponding dynamic operation impact envelope model is established for each UAV. This model is used to quantify the physical collision risk, rotor airflow disturbance risk and paint drift pollution risk generated by the UAV in space into a time-varying spatial risk field function. Global spatiotemporal planning module: Based on the preset task path and bridge structure model, in a four-dimensional spatiotemporal space including the time dimension, an improved conflict-based search algorithm is used to plan an initial spatiotemporal trajectory for each UAV and calculate the corresponding spatiotemporal corridor constraints. The planning process aims to minimize the total task time and the superposition integral of the operation risk field calculated based on the spatial risk field function. Online Cooperative Scheduling Module: Taking the dynamic operation influence envelope model and spatiotemporal corridor constraints as input, a dynamic heterogeneous graph is constructed, and a pre-trained graph neural network is used to extract the global cooperative state features of the dynamic heterogeneous graph. This drives the central scheduler in the online cooperative scheduling module to generate real-time control commands, which are used to fine-tune the initial spatiotemporal trajectory. Closed-loop verification execution module: While real-time control commands are issued, the real-time control commands and the measured data collected by the UAV's onboard sensors are input into the digital twin unit within the closed-loop verification execution module for forward-looking simulation. This predicts the collaborative operation status of multiple UAVs within a short time window and calculates the deviation from the spatiotemporal corridor. When the deviation exceeds a preset safety threshold, the online collaborative scheduling module is triggered to perform emergency compensation control. At the same time, the difference between the predicted data from the forward-looking simulation and the measured data is used to periodically optimize the dynamic operation influence envelope model and the online collaborative scheduling module.

[0006] In a preferred embodiment, the specific process of establishing a corresponding dynamic operational influence envelope model for each UAV based on the UAV's rotor parameters, coating parameters, and environmental wind field data is as follows: It receives and processes three types of inputs: rotor parameters, spraying parameters, and ambient wind field data. Rotor parameters include the equivalent collision radius of the UAV body, rotor radius, position and normal vector of the rotor plane in the body coordinate system, and rated downdraft velocity; Spraying parameters include spraying start / stop status, nozzle position and direction vector in the machine coordinate system, initial paint spraying speed, spray cone angle, and average paint particle size; The environmental wind field data consists of real-time environmental wind speed and direction vectors; Based on rotor parameters, coating parameters, and environmental wind field data, a time-varying spatial risk field function defined in three-dimensional space is constructed for each UAV at every moment. The value range of the spatial risk field function is defined between zero and one. The function value of the spatial risk field function at any point is used to quantify the risk intensity of the corresponding point in three-dimensional space that is adversely affected by UAV operation into a scalar. A scalar value of one represents the highest risk, and a scalar value of zero represents no risk. Risks include physical collision risk, airflow disturbance risk, and paint contamination risk.

[0007] In a preferred embodiment, the spatial risk field function is constructed by taking the maximum value of each component in a composite risk field; The composite risk field is composed of the superposition of the physical collision risk field, the rotor airflow disturbance risk field, and the paint drift pollution risk field. The function value of the physical collision risk field depends on the Euclidean distance between the spatial point in three-dimensional space and the real-time position of the UAV at the corresponding moment, and decays exponentially with the Euclidean distance as the variable. The function value of the rotor airflow disturbance risk field is determined by the rotor operating state coefficient, the ratio of the rated downdraft velocity to a reference wind speed, and the geometric relationship of a spatial point in three-dimensional space relative to the rotor, and decays with increasing distance. The function value of the paint drift pollution risk field is controlled by the spraying state and is obtained by time integration of the probability density of the sprayed droplets appearing at spatial points in three-dimensional space over an effective spraying time period from the current moment back. The dynamic operation influences the envelope model output spatial risk field function numerically described on a discretized three-dimensional spatial grid.

[0008] In a preferred embodiment, the specific process of planning within a four-dimensional spacetime containing the time dimension, based on a preset task path and bridge structure model, in the global spatiotemporal planning module is as follows: First, a numerical description of the spatial risk field function is received, and a four-dimensional spatiotemporal grid map is constructed by combining it with the bridge structure model. Define a spatiotemporal trajectory for each drone from the start time to the end time; plan with the objective of minimizing a multi-objective cost function, which is the sum of the costs of all drone spatiotemporal trajectories; The cost of each spatiotemporal trajectory is a weighted sum of the time cost corresponding to its duration and the risk cost corresponding to the integral of its operational risk field. The calculation process of the superimposed integral of the operation risk field is as follows: for a spatiotemporal trajectory, from the start time to the end time, it is gradually advanced at a fixed time step; At each discrete moment reached, the spatial position of the UAV at that discrete moment is first determined, and then the spatial risk field function value of each UAV at that spatial position and discrete moment is obtained. Next, all the obtained spatial risk field function values ​​are multiplied by preset risk coupling weights representing the importance of different risk sources and then summed to obtain an instantaneous superimposed risk value at that discrete moment. Finally, multiply this instantaneous superimposed risk value by the time step, and sum up the results of such multiplications at all discrete moments to obtain the superimposed integral of the operational risk field for this spatiotemporal trajectory.

[0009] In a preferred embodiment, the specific process of employing the improved conflict-based search algorithm is as follows: The conflict-based search algorithm maintains a search tree containing candidate spatiotemporal trajectories for each UAV and manages conflict constraints on a constraint tree; and defines conflict expansion as hard conflict and soft conflict. Hard collisions include geometric collisions and collisions that occur when the sum of the spatial risk field functions of two drones at any point in time exceeds a hard collision safety threshold. Soft conflict occurs when the sum of values ​​exceeds an optimization threshold that is below the hard conflict safety threshold but does not reach the hard conflict safety threshold. The iterative process of the conflict-based search algorithm is as follows: each UAV searches for its initial spatiotemporal trajectory while ignoring other UAVs; hard and soft conflicts are detected among all spatiotemporal trajectories. Resolve hard conflicts and replan the affected drones; for soft conflicts, guide the process at higher nodes of the constraint tree by comparing the total cost of containing the conflict with the cost of resolving it. Finally, a set of spatiotemporal trajectory solutions without hard conflicts are generated, and a corresponding spatiotemporal corridor constraint is generated for the trajectory of each UAV. The spatiotemporal corridor constraint is a four-dimensional spatiotemporal volume jointly determined by the spatiotemporal trajectory itself, the UAV body envelope, and the risk isosurface defined by the optimization threshold.

[0010] In a preferred embodiment, the specific process of constructing a dynamic heterogeneous graph using the dynamic operation influence envelope model and spatiotemporal corridor constraints as inputs is as follows: The system receives a dynamic job influence envelope model for real-time computation, an initial spatiotemporal trajectory generated for each UAV and its corresponding spatiotemporal corridor constraints, and real-time position, real-time speed, and remaining task information of the UAV from airborne sensors; based on the input, it constructs a dynamic heterogeneous graph that changes with job time. The node set of the dynamic heterogeneous graph includes drone nodes, cable segment nodes, and sensitive component nodes, each with its own attributes; The edges of the dynamic heterogeneous graph constitute an edge set, which includes at least spatial proximity edges, task-assigned edges, and risk-affected edges. The weight of the spatial proximity edges is determined by the distance between nodes. The task-assigned edges have fixed weights. The weight of the risk-affected edges is calculated in real time by calling the dynamic job influence envelope model, and the specific value is equal to the value of the spatial risk field function caused by the UAV corresponding to the source UAV node to the target node position.

[0011] In a preferred embodiment, the specific process of extracting the global cooperative state features of the dynamic heterogeneous graph using a pre-trained graph neural network and driving the central scheduler to generate real-time control commands is as follows: First, the dynamic heterogeneous graph is input into a pre-trained graph neural network, and a node embedding vector is generated for each drone node through multi-layer message passing, which serves as a global collaborative state feature. Next, the central scheduler generates real-time control commands based on the node embedding vectors of all UAV nodes through a pre-built policy network trained by multi-agent reinforcement learning. The policy network is trained by optimizing a multi-objective reward function, which is a weighted combination of task progress reward, corridor violation penalty and cooperation penalty. The collaborative penalty term is obtained by nonlinearly amplifying and integrating the interaction results of the spatial risk field functions of each pair of UAVs in the overlapping region, and is used to drive the policy network to generate fine-tuning actions to reduce the overlap of high-risk areas. The generated raw motion commands are processed by secure projection to ensure that the expected state of the UAV after application falls within its spatiotemporal corridor constraints, resulting in fine-tuning of flight speed and slight translation of the spraying start and stop sequence.

[0012] In a preferred embodiment, the specific process of inputting real-time control commands and measured data into the digital twin unit for forward-looking simulation and predicting the collaborative operation status within a short time window is as follows: While real-time control commands are being issued, the digital twin unit receives real-time control commands, measured position and speed data of the UAV from airborne sensors, as well as dynamic operation influence envelope model and spatiotemporal corridor constraints. The digital twin unit maintains a high-fidelity simulation environment, which embeds a real-time calculation engine for the dynamic operation influence envelope model and stores all UAV state copies, environmental wind field estimates, and bridge structure models. Based on the input, whenever a new set of real-time control commands is generated, the digital twin unit immediately starts a multi-agent cooperative dynamics simulation that runs faster than the actual time, using the latest measured UAV position and speed as the initial state and the real-time control commands as the control input, to deduce the state evolution of all UAVs and their interactions over a preset period of time in the future. During the multi-agent cooperative dynamics simulation, the dynamic operation influence envelope model is synchronously invoked to calculate the spatial risk field function value of each UAV and its distribution in three-dimensional space at each simulation step. This yields the predicted risk field distribution of the entire operation airspace over time within a short future time window. The simulation process generates a spatial position for each UAV at each moment, which is called the simulation prediction position.

[0013] In a preferred embodiment, the specific process of calculating the deviation from the spatiotemporal corridor and triggering emergency compensation control is as follows: Based on the predicted risk field distribution obtained from forward-looking simulation, a deviation called the spatiotemporal consistency index is calculated. The calculation process of the deviation is as follows: within the short future time window of the simulation, a continuous time interval is selected, and for each discrete time node in this time interval, the risk intensity compliance judgment and statistics of each spatial point in the three-dimensional space are performed in the entire range of the preset three-dimensional work space. After the statistics are completed, the proportion of the total volume of the three-dimensional space points that are judged to be compliant within the entire work space at that time point is calculated to obtain the space compliance rate at that time point. Next, the spatial compliance rates of all discrete time nodes within the selected time interval are averaged to obtain a spatiotemporal compliance rate; finally, the spatiotemporal compliance rate is subtracted from one to obtain the deviation; the calculated deviation is compared with a preset safety threshold between zero and one. If the calculated deviation is less than or equal to the safety threshold, no intervention will be taken; If the calculated deviation exceeds the safety threshold, emergency compensation control is triggered; an interrupt signal is sent to the online collaborative scheduling module. Upon receiving an interruption signal, the online collaborative scheduling module recalculates and generates a new set of compensatory real-time control instructions based on the latest emergency status within its policy network framework.

[0014] In a preferred embodiment, the specific process of using the difference data between the predicted data from the prospective simulation and the measured data for periodic optimization is as follows: We continuously collect and buffer two types of discrepancy data: the first type is called trajectory tracking discrepancy; the second type is called risk perception discrepancy. The two types of differential data together constitute the risk perception differential data used for optimization. Risk perception discrepancy data is buffered and used to perform the following two periodic optimization processes: The first optimization is to optimize the dynamic operation influence envelope model. Specifically, the collected risk perception difference data is used as training samples, and an online learning algorithm is used to periodically adjust the key physical parameters in the dynamic operation influence envelope model so that the prediction output of the dynamic operation influence envelope model for the future risk field gradually approaches the real physical effect. The second optimization is to optimize the online collaborative scheduling module. Specifically, the experience tuples generated during the actual operation, consisting of the UAV status, real-time control commands issued, subsequent status, and actual operation cost, are stored as new experience samples in the experience replay buffer maintained by the policy network within the online collaborative scheduling module. This buffer is used to store historical experience data for training purposes. During a preset period when computing resources are idle, offline fine-tuning training of the policy network is triggered. New experience samples from the real world are used to update the internal parameters of the policy network, so that the cooperative policies generated by the policy network can better adapt to the actual dynamic environment.

[0015] The beneficial effects of this invention are as follows: By dynamically modeling, the physical collisions, airflow disturbances, and coating contamination risks of UAV operations are uniformly quantified into a time-varying risk field, providing a quantitative benchmark for collaborative decision-making; based on this, an initial trajectory without hard conflicts and balancing efficiency and risk is planned in four-dimensional spacetime, and a safe corridor is generated; online scheduling utilizes graph neural networks to perceive the global collaborative state in real time, driving multiple UAVs to perform dynamic fine-tuning to actively avoid interference; at the same time, forward-looking digital twin simulation continuously verifies and predicts, and once a deviation is detected, compensation control is immediately triggered to form a safety closed loop, and by comparing predicted and measured data, the model and strategy are continuously optimized, thereby simultaneously achieving safety and reliability, uniform and clean coating, and improved overall operational efficiency in close collaborative operation of multiple UAVs. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application. Example

[0020] This embodiment provides, for example Figure 1-2 The UAV collaborative precision spraying maintenance system for cable maintenance is shown, specifically including: an influence envelope modeling module, a global spatiotemporal planning module, an online collaborative scheduling module, and a closed-loop verification execution module connected in sequence; Impact envelope modeling module: Before collaborative operation, based on the rotor parameters, spraying parameters and environmental wind field data of the UAV, a corresponding dynamic operation impact envelope model is established for each UAV. This model is used to quantify the physical collision risk, rotor airflow disturbance risk and paint drift pollution risk generated by the UAV in space into a time-varying spatial risk field function. Global spatiotemporal planning module: Based on the preset task path and bridge structure model, in a four-dimensional spatiotemporal space including the time dimension, an improved conflict-based search algorithm is used to plan an initial spatiotemporal trajectory for each UAV and calculate the corresponding spatiotemporal corridor constraints. The planning process aims to minimize the total task time and the superposition integral of the operation risk field calculated based on the spatial risk field function. Online collaborative scheduling module: Taking the dynamic operation influence envelope model and spatiotemporal corridor constraints as input, a dynamic heterogeneous graph is constructed. The nodes of the dynamic heterogeneous graph include entities representing each UAV, each section of cable to be maintained, and each sensitive component. The global collaborative state features of the dynamic heterogeneous graph are extracted using a pre-trained graph neural network, which drives the central scheduler in the online collaborative scheduling module to generate real-time control commands. The real-time control commands are used to fine-tune the initial spatiotemporal trajectory. Closed-loop verification execution module: While real-time control commands are issued, the real-time control commands and the measured data collected by the UAV's onboard sensors are input into the digital twin unit within the closed-loop verification execution module for forward-looking simulation. This predicts the collaborative operation status of multiple UAVs within a short time window and calculates the deviation from the spatiotemporal corridor. When the deviation exceeds a preset safety threshold, the online collaborative scheduling module is triggered to perform emergency compensation control. At the same time, the difference between the predicted data from the forward-looking simulation and the measured data is used to periodically optimize the dynamic operation influence envelope model and the online collaborative scheduling module.

[0021] In this embodiment, the specific process of establishing a corresponding dynamic operational influence envelope model for each UAV based on its rotor parameters, coating parameters, and environmental wind field data is as follows: It receives and processes three types of inputs: rotor parameters, spraying parameters, and ambient wind field data. Rotor parameters include the equivalent collision radius of the UAV body, rotor radius, position and normal vector of the rotor plane in the body coordinate system, and rated downdraft velocity; the equivalent collision radius of the UAV body is determined by the maximum outer dimensions of the UAV, and its value range is usually 0.5 meters to 1.5 meters; the rated downdraft velocity is obtained through ground static testing, and its typical value is 3 meters per second to 8 meters per second; Spraying parameters include spraying start / stop status, nozzle position and direction vector in the machine coordinate system, initial paint spraying speed, spray cone angle, and average paint particle size. The initial paint spraying speed is set according to the spraying pump pressure and nozzle diameter, ranging from 0.5 meters per second to 5 meters per second. The spray cone angle is usually between 30 degrees and 80 degrees. The average paint particle size is between 50 micrometers and 200 micrometers. The environmental wind field data consists of real-time environmental wind speed and direction vectors. The real-time environmental wind speed and direction vectors are collected in real time by wind speed and direction sensors on airborne or ground meteorological stations, with an update frequency of no less than once per second. Based on rotor parameters, coating parameters, and environmental wind field data, a time-varying spatial risk field function defined in three-dimensional space is constructed for each UAV at every moment. The value range of the spatial risk field function is defined between zero and one. The function value at any point is used to quantify the risk intensity of the adverse effects of UAV operation on the corresponding point in three-dimensional space into a scalar. A scalar value of one represents the highest risk, and a scalar value of zero represents no risk. Risks include physical collision risk, airflow disturbance risk, and paint contamination risk. The quantification of "risk intensity" enables the subsequent planning and scheduling modules to perform unified and comparable mathematical processing on operational interference of different natures and sources. This is the foundation for achieving accurate quantification of multi-UAV collaborative conflicts. The spatial risk field function is constructed by taking the maximum value of each component in a composite risk field; The composite risk field is composed of the superposition of the physical collision risk field, the rotor airflow disturbance risk field, and the paint drift pollution risk field. The physical collision risk field characterizes the risk of collision with the UAV body. The function value of the physical collision risk field depends on the Euclidean distance between a spatial point in three-dimensional space and the real-time position of the UAV at the corresponding moment. The risk intensity distribution is centered on the instantaneous centroid position of the UAV. The risk intensity value decreases exponentially with the increase of the Euclidean distance, according to the attenuation rate determined by the product of the equivalent collision radius of the UAV body and a preset safety factor greater than one. The preset safety factor is used to add a buffer zone outside the geometric collision boundary. Its value is usually set to 1.2 to 1.5 to cope with positioning errors and dynamic response delays. The exponential decay form ensures that the risk value drops sharply at the safety boundary, thereby ensuring safety without overly conservatively restricting the UAV's passable space. The rotor airflow disturbance risk field characterizes the risk of interference from the rotor downwash airflow to the surrounding environment. The function value of the rotor airflow disturbance risk field is determined by the rotor operating state coefficient, the ratio of the rated downwash airflow velocity to a reference wind speed, a radial attenuation factor based on the vertical distance from a spatial point in three-dimensional space to the rotor central axis, an axial attenuation factor based on the absolute value of the directional distance of a spatial point in three-dimensional space relative to the rotor plane in the rotor plane normal direction, and a directional cosine factor based on the angle between the position vector of the spatial point in three-dimensional space relative to the rotor center and the rotor plane normal vector. Among them, the radial attenuation factor and the axial attenuation factor attenuate the distance in the form of a negative exponential function based on the preset radial attenuation rate parameters and axial attenuation rate parameters, respectively; the reference wind speed is used to normalize the downdraft velocity, usually taken as 1 meter per second; the radial attenuation rate parameters and the axial attenuation rate parameters are obtained by fitting computational fluid dynamics simulation or wind tunnel experimental data, respectively describing the attenuation characteristics of airflow disturbance in the horizontal and vertical directions, and their values ​​determine the spatial range of the risk field; the directional cosine factor makes the risk field mainly concentrated in the mainstream direction of the rotor downwash, which is consistent with physical reality; The paint drift pollution risk field characterizes the pollution risk caused by paint droplet diffusion. The function value of the paint drift pollution risk field is controlled by the spraying start and stop status and is obtained by time integration of the probability density of all sprayed droplets appearing at spatial points in three-dimensional space within an effective spraying time period from the current moment. The effective spraying time period is set according to the residence time of paint droplets in the air, usually a time window of 1 to 5 seconds before the current moment. The time integration operation transforms the instantaneous spraying action into a spatial risk distribution with a continuous impact, thereby enabling the assessment of the cumulative pollution effect of spraying on a spatiotemporal region. The expected spatial position of a single droplet at any time after ejection is determined by a uniformly accelerated trajectory jointly determined by the nozzle position, the product of the initial spray velocity of the coating and the nozzle direction vector, the real-time ambient wind speed vector, and the gravitational acceleration vector. This trajectory model treats the droplet as a point mass and comprehensively considers the initial kinetic energy, ambient wind load, and gravitational effect, which is the basic physical model for predicting the spatial drift of the coating. The probability density distribution of a single droplet appearing at a expected spatial location is described by a time-varying covariance matrix, which is determined by parameters related to the spray cone angle, turbulence intensity, and average particle size of the coating. The diffusion range of the time-varying covariance matrix expands with time. The principal axis of the time-varying covariance matrix is ​​related to the spray cone angle, and its increasing characteristic over time reflects the diffusion process of the droplet swarm under turbulence. The turbulence intensity can be estimated from environmental wind field data or set as an empirical value. For each drone and each moment requiring assessment, the dynamic operation impact envelope model outputs a numerical description of the spatial risk field function on a discretized three-dimensional spatial grid. The three-dimensional spatial grid covers the operational airspace, and the grid resolution is set according to the balance between accuracy requirements and computational resources, for example, 0.1 meters to 0.5 meters. The numerical description is a three-dimensional array, where each element corresponds to a scalar value of risk intensity at the center point of the grid. This data structure is easy for subsequent algorithms to store, transmit, and compute. The output spatial risk field function, numerically described on a discretized grid, encapsulates all external influences determined by the UAV at a given moment based on rotor parameters, spraying parameters, and environmental wind field data. This function serves as direct input data for the global spatiotemporal planning module. The module uses the numerical description to calculate the superposition integral of the operational risk field associated with any planned spatiotemporal trajectory, and to determine whether the spatial risk field functions corresponding to different UAVs exhibit illegal overlap in the spatiotemporal dimension. The "superposition integral of the operational risk field" refers to the integral calculated along a spatiotemporal trajectory for each node traversed by the trajectory. The risk field function values ​​corresponding to the spatiotemporal points are accumulated or integrated to obtain the overall risk cost of the trajectory. The criterion for "illegal overlap" is: if the sum of the spatial risk field function values ​​corresponding to two or more UAVs at the same spatiotemporal point exceeds a preset overlap risk threshold, it is determined to be illegal overlap. The setting of this threshold needs to take into account safety redundancy and operational efficiency, for example, it can be set to 0.7. By calling a unified numerical description, the global spatiotemporal planning module can efficiently evaluate the path safety and cleanliness with a standardized data interface, realizing the decoupling and collaboration between risk assessment and path planning.

[0022] In this embodiment, it is specifically necessary to explain the planning process in the global spatiotemporal planning module, which is based on a preset task path and bridge structure model, within a four-dimensional spatiotemporal space including the time dimension: First, the numerical description of the spatial risk field function pre-calculated for each UAV at discrete time points on a discretized three-dimensional spatial grid is received, and combined with the triangular mesh model of the bridge structure model to jointly construct a four-dimensional spatiotemporal grid map. The spatial dimension grid resolution of the four-dimensional spatiotemporal grid map is usually consistent with the resolution of the discretized three-dimensional spatial grid, for example, 0.1 meters to 0.5 meters; its time dimension step is set according to the UAV motion control accuracy and planning real-time requirements, usually between 0.1 seconds and 0.5 seconds. The process of constructing this map is essentially to encode the dynamic and continuous risk field and static obstacle information into a discrete and efficiently searchable data structure. The four-dimensional spatiotemporal grid map consists of a series of four-dimensional cells. Each cell records its own spatial location and temporal index, and stores two types of information: the first type is the static obstacle occupancy status, and the second type is the dynamic risk field superposition value obtained by querying and interpolating the numerical description of the spatial risk field function of each UAV at the corresponding spatial location and time in the four-dimensional cell on a discretized three-dimensional spatial grid. The "query and interpolation" operation, for a given spatial location and time, if it does not fall exactly at the pre-calculated grid point and time, estimates the spatial risk field function value of each UAV at that location through three-dimensional spatial linear interpolation and temporal linear interpolation, and then superimposes the estimated values ​​of all UAVs. This process ensures the continuity of spatial risk field assessment. The planned spatiotemporal trajectory for each drone is a four-dimensional path from the start time to the expected end time. The spatiotemporal trajectory explicitly specifies the spatial position of the drone at each discrete time step. This spatiotemporal trajectory is the direct object of subsequent optimization. It not only defines "when to arrive at where" but also implies the drone's movement sequence. The planning process aims to minimize the multi-objective cost function, which is the sum of the costs of all UAV spatiotemporal trajectories. The positive weighting coefficients, namely the time cost weighting coefficient and the risk cost weighting coefficient, need to be calibrated before simulation or actual operation to reflect different emphases on operational efficiency and safety / cleanliness. For example, the risk cost weighting coefficient can be increased in areas with strong winds or densely packed sensitive components. The cost of each spatiotemporal trajectory is composed of the time cost corresponding to the duration of the spatiotemporal trajectory and the risk cost corresponding to the superposition integral of the operational risk field of the spatiotemporal trajectory, multiplied by a positive weighting coefficient and then added together. The calculation process of the superimposed integral of the operation risk field is as follows: for a spatiotemporal trajectory, from the start time to the end time, it is gradually advanced at a fixed time step; At each discrete moment reached, the spatial position of the UAV at that discrete moment is first determined, and then the spatial risk field function value of each UAV at that spatial position and discrete moment is obtained. The preset principle of risk coupling weight is: the weight of "self-risk item" is usually set to 1, indicating that all risks generated by its own operation are fully included; for "other machine interference risk item", the weight of "airflow interference" from other machines can be set to 0.6 to 0.9, while the weight of "paint drift interference" can be set to 0.8 to 1.0, because paint pollution is usually irreversible and needs to be avoided more strictly; this differentiated weight setting allows the algorithm to make more refined trade-offs in the conflict resolution process. Next, all the obtained spatial risk field function values ​​are multiplied by preset risk coupling weights representing the importance of different risk sources and then summed to obtain an instantaneous superimposed risk value at that discrete moment. Finally, multiply this instantaneous superimposed risk value by the time step, and sum up the results of such multiplications at all discrete moments to obtain the superimposed integral of the operational risk field for this spatiotemporal trajectory. The risk coupling weights include the weights for its own risk items and the weights for the risk items of interference from other aircraft. The weights for the risk items of interference from other aircraft can be set to different values ​​according to different types of airflow interference or paint drift interference. The specific process of using the improved conflict-based search algorithm is as follows: The conflict-based search algorithm maintains a search tree containing candidate spatiotemporal trajectories for each UAV and manages conflict constraints on a constraint tree. The conflict-based search algorithm handles two types of conflicts. The first is hard conflict, which includes geometric collisions between drones or between a drone and a static obstacle, and at any given time and space point, the sum of the spatial risk field functions of the two drones at the corresponding point exceeds a hard conflict safety threshold between 0.8 and 1.0. The recommended hard conflict safety threshold is 0.9. This threshold is set close to the maximum value of 1, which means that only when the risk fields of the two drones highly overlap in time and space, causing the superimposed risk to approach the theoretical maximum risk, is it considered a hard safety problem that must be eliminated. This retains a certain degree of flexibility in planning and avoids the lack of a solution due to overly conservative hard constraints. The second type is soft conflict, which refers to the situation where, at any given point in time, the sum of the spatial risk field function values ​​of two drones at the corresponding point exceeds an optimization threshold between 0.2 and 0.5, which is less than the hard conflict safety threshold, but does not reach the hard conflict safety threshold. The optimization threshold is recommended to be set to 0.3. This threshold defines the risk level that needs to be optimized but is not absolutely prohibited. The existence of soft conflict increases the trajectory cost, guiding the algorithm to find solutions with less mutual interference, thereby further improving the overall cleanliness and smoothness of the operation while ensuring hard safety. The hard conflict safety threshold is a more stringent safety criterion than the preset overlap risk threshold used to judge illegal overlap. This relationship (hard conflict safety threshold > illegal overlap risk threshold) ensures that a more stringent standard is adopted in the planning stage than simply judging "whether there is illegal overlap", thus leaving a safety margin for the subsequent online scheduling module to cope with real-time disturbances. The iterative process of the conflict-based search algorithm is as follows: First, each UAV searches for an initial spatiotemporal trajectory that minimizes its own spatiotemporal trajectory cost in the four-dimensional spatiotemporal grid map, ignoring other UAVs. At this time, the integral of the operation risk field only includes its own risk and the risk of static obstacles. This step finds a locally optimal starting point for each UAV. Next, in the set of spatiotemporal trajectories consisting of the current spatiotemporal trajectories of all UAVs, it is necessary to detect whether there is a hard or soft conflict. During the detection, the sum of the spatial risk field function values ​​of each pair of UAVs at all overlapping spatiotemporal points needs to be calculated. Conflict detection is the core computational step of the algorithm, and its output is a list containing conflict type, conflicting UAV pairs, conflict spatiotemporal points, and conflict severity (overlapping value). If a hard conflict is detected, a constraint prohibiting the drones involved in the conflict from existing at the conflict point in time and space is added, and the affected drones are instructed to replan their time and space trajectories. This is the process of solving the absolute safety problem. By adding the "prohibition" constraint, the final solution is ensured to meet the hard conflict safety conditions. For soft conflicts, the conflict-based search algorithm compares the calculated result of the multi-objective cost function of the corresponding spatiotemporal trajectory set containing such conflicts with the additional time cost that resolving the conflict may introduce. This cost comparison is performed at higher-level nodes of the constraint tree to guide the search towards a smaller calculated result for the multi-objective cost function; this step is key to the algorithm's "improvement." Traditional CBS typically treats any conflict as a hard constraint that must be resolved; however, this method allows soft conflicts to exist but calculates the incremental risk cost they bring. At higher-level nodes of the constraint tree, the algorithm compares the total cost (time cost + risk cost) of nodes that "keep the current soft conflict but have a shorter total time" with those that "resolve the soft conflict but have a longer total time," selecting the branch with the smaller total cost to continue the search. This allows the algorithm to perform global, cross-agent cost trade-offs, thereby finding a solution with better overall efficiency and safety / cleanliness. When the search algorithm finds a set of spatiotemporal trajectory solutions without any hard conflicts, a corresponding spatiotemporal corridor constraint is generated for each UAV's spatiotemporal trajectory. The spatiotemporal corridor constraint is a four-dimensional spatiotemporal volume, the boundary of which is jointly determined by the spatiotemporal trajectory itself, the UAV's body envelope, and a risk isosurface defined by an optimization threshold. The risk isosurface refers to the surface formed by three-dimensional spatial points where the spatial risk field function value is equal to the optimization threshold at each discrete moment. The spatial range allowed by the spatiotemporal corridor constraint at each moment is the region located within this risk isosurface and connected to the trajectory points. The generated spatiotemporal corridor constraint not only includes the geometric channel of the trajectory but also clarifies the size of the "risk-safe zone" around the UAV's trajectory at each moment without triggering new soft conflicts. This defines a clear boundary for the fine-tuning behavior of the online collaborative scheduling module and is a key data interface connecting global planning and real-time scheduling.

[0023] In this embodiment, the specific process of constructing a dynamic heterogeneous graph using the dynamic operation influence envelope model and spatiotemporal corridor constraints as input is as follows: The system receives a dynamic job influence envelope model for real-time computation, an initial spatiotemporal trajectory generated for each UAV and its corresponding spatiotemporal corridor constraints, and real-time position, real-time speed, and remaining task information of the UAV from airborne sensors; based on the input, it constructs a dynamic heterogeneous graph that changes with job time. The node set of the dynamic heterogeneous graph contains three types of nodes: drone nodes, cable segment nodes to be maintained, and sensitive component nodes to be avoided. Each drone node has its real-time position, speed, remaining task length, and current spraying status as node attributes. Each cable segment node has its geometric centerline parameters, length, and current spraying coverage as node attributes. Each sensitive component node has its position and type as node attributes. "Remaining task length" is the remaining unsprayed physical length of the current cable segment to be sprayed. "Current spraying coverage" is the ratio of the sprayed area to the total target spraying area of ​​the cable segment. The edges of the dynamic heterogeneous graph are used to characterize the dynamic interaction relationships between the drone nodes, the cable segment nodes to be maintained, and the sensitive component nodes to be avoided in the graph. The edges of the dynamic heterogeneous graph constitute an edge set, which contains at least three types of edges: spatial proximity edges, task affiliation edges, and risk impact edges. Spatial proximity edges connect any two node entities whose spatial Euclidean distance is less than a preset proximity distance threshold. The weight of the edge is determined by a negative exponential function, which is the square of the Euclidean distance between the two node entities divided by the square of twice the distance decay scale parameter. The preset proximity distance threshold is set according to the safe interval and communication range of UAV operations, and is usually between 5 meters and 15 meters. The distance decay scale parameter controls the rate at which the weight decays with distance, and its value is usually set to about half of the proximity distance threshold, such as 2 meters to 5 meters, to ensure that the weight has decayed to near zero at the threshold. This negative exponential weight form can smoothly represent the continuous change of spatial proximity. The task assignment edge connects the drone node performing the spraying operation with the cable segment node it is currently responsible for maintaining, and has a fixed weight of one. This edge establishes a clear task assignment relationship, which is the basis for graph neural networks to understand the task structure. Risk impact edges connect drone nodes to any other node entity in the dynamic heterogeneous graph. The weight of the risk impact edge is calculated in real time by calling the dynamic operation impact envelope model. The specific value is equal to the spatial risk field function value of the drone corresponding to the drone node at the time of calculation, which causes the target node entity to be in a spatial position. This quantifies the continuous risk field into the dynamic interaction intensity between nodes in the dynamic heterogeneous graph. By seamlessly converting it into a graph data structure, the graph neural network can directly perceive and infer the dynamic impact between drones and the drones on the environment and mission objectives. When each node is computed in a graph neural network, it is represented as a high-dimensional numerical vector called the node embedding vector, which encodes the state of the node and its relationship with other nodes in the dynamic heterogeneous graph. The specific process of using a pre-trained graph neural network to extract global cooperative state features of dynamic heterogeneous graphs and drive the central scheduler to generate real-time control commands is as follows: First, the constructed dynamic heterogeneous graph is input into a pre-trained graph neural network; the graph neural network contains multiple sequentially connected message passing layers; In each messaging layer, for each node in the dynamic heterogeneous graph, the following operations are performed to compute the node embedding vector of that node in the current messaging layer: First, an aggregation operation is performed to collect the node embedding vectors of all neighboring nodes in the previous messaging layer, forming a set of neighbor embedding vectors. Then, an aggregation function is used to process the set of neighbor embedding vectors. The aggregation function summarizes the information of all neighbors. The summarization method includes, but is not limited to, attention-weighted summation of the neighbor embedding vectors. "Attention-weighted summation" means that different weights are assigned to different neighbor node embedding vectors during aggregation. These weights are usually calculated by combining the features of the neighboring nodes and the current node, allowing the graph neural network to focus on more important interaction relationships. Next, an update operation is performed. The node's embedding vector from the previous message passing layer and the aggregated result of the aggregation function are input into an update function. This update function generates a new node embedding vector for the node in the current message passing layer. The update function is usually a small feedforward neural network that is responsible for fusing the node's own historical information with the aggregated neighbor information. Through multi-layer message passing, the node embedding vector of each node can integrate the information of its multi-hop neighbor nodes. Finally, the node embedding vector generated for each drone node and updated after the last message passing layer is the extracted global collaborative state feature, which encodes the state of the drone itself, the state and intention of surrounding entities, and the real-time risk distribution from other drones perceived through risk impact edges. After, for example, 3 to 5 layers of message passing, the embedding vector of each drone node can perceive indirect, higher-order interaction effects, such as "how my actions affect another drone, and thus indirectly affect a third drone or a sensitive component", thereby forming a global collaborative perspective. Next, a central scheduler receives the node embedding vectors of all drone nodes and generates real-time control commands for each drone based on a pre-built policy network trained by multi-agent reinforcement learning. During the training phase, the policy network learns cooperative policies by optimizing a specific multi-objective reward function. The multi-objective reward function is the result of subtracting the weighted sum of task progress rewards, the weighted sum of corridor violation penalties, and the weighted sum of a cooperative penalty. The weighting coefficients are set before training; for example, the weight of task progress rewards can be 1.0, the weight of corridor violation penalties can be -5.0, and the weight of cooperative penalties can be -2.0. This structure forces the policy network to prioritize safety (avoiding corridor violations) and actively reduce mutual interference while pursuing efficiency. The task progress reward item is used to encourage drones to efficiently complete the painting operation of their respective cable segments; the specific calculation can be the increase in the painting length per unit time. The corridor violation penalty is activated and a negative reward is applied when the action suggested by the policy network causes the expected state of the drone to exceed the constraints of the corresponding spatiotemporal corridor. This penalty is usually set to a large fixed negative value (such as -10) to ensure that the policy network learns to strictly avoid behaviors that would lead to out-of-bounds behavior during training. The cooperative penalty term is used to drive multiple drones to actively cooperate in avoidance. The specific calculation process is as follows: For each pair of different drones, at the calculation time, the overlapping region of the spatial risk field function defined by the dynamic operation influence envelope model of the two drones in three-dimensional space is determined. Within this overlapping region, a specified mathematical interaction operation is performed on the spatial risk field function values ​​of the two drones to obtain an interaction result. This interaction result is then multiplied by a specified power greater than one. The result after the power operation is then integrated over the entire overlapping region. Finally, the integration results of all drone pairs are summed to obtain the cooperative penalty term. The specified mathematical interaction operation usually adopts multiplication, that is, the interaction result is the product of the two risk field values, which can sensitively capture the overlap of high-risk areas. The specified power greater than one is usually 2 or 3. Through this nonlinear amplification, an extremely severe penalty is imposed on the overlap of the core region (value close to 1) of the risk field of the two drones, forcing the policy network to learn fine cooperative actions that can significantly avoid high-risk spatiotemporal regions, rather than just reducing the overall overlap volume. Powers greater than one are used to nonlinearly amplify the high-value parts of the interaction results, thereby defining regions with high values ​​of the spatial risk field function as high-risk regions and imposing a more severe penalty on the spatiotemporal overlap of high-risk regions. This incentivizes the strategy network to generate fine-tuning actions that can significantly reduce the spatiotemporal overlap of high-risk areas; Finally, the original action commands output by the policy network need to undergo a secure projection process. This process ensures that the original action commands are mapped to new action commands, so that after applying these new commands, the expected state of each UAV, including its expected position and expected velocity, strictly falls within the four-dimensional spacetime bounded by the spatiotemporal corridor constraints corresponding to the UAV. Secure projection processing is usually achieved by solving a constrained optimization problem, that is, finding a feasible command that is closest to the original output command of the policy network within the boundaries of the spatiotemporal corridor constraints. Real-time control commands include fine-tuning of flight speed and small translations of the spraying start and stop timing. The speed fine-tuning range is usually ±10% of the nominal speed, and the spraying timing translation range is usually ±0.5 seconds. These fine-tunings achieve compensation for dynamic disturbances and real-time optimization of coordination without violating the global safety framework.

[0024] In this embodiment, the specific process of inputting real-time control commands and measured data into the digital twin unit for forward-looking simulation and predicting the collaborative operation status within a short time window is as follows: While real-time control commands are being issued, the digital twin unit receives real-time control commands, measured position and speed data of the UAV from airborne sensors, as well as dynamic operation influence envelope model and spatiotemporal corridor constraints. The digital twin unit maintains a high-fidelity simulation environment, which embeds a real-time calculation engine for the dynamic operation influence envelope model and stores state copies of all UAVs, environmental wind field estimates, and bridge structure models. The "high-fidelity simulation environment" is built by integrating a multibody dynamics engine, aerodynamic models, and environmental interaction models to ensure that its simulation results highly reflect real physical laws. The "state copy" refers to a digital mirror image that corresponds one-to-one with the real UAV and includes its position, velocity, attitude, and mission state. Based on the input, whenever a new set of real-time control commands is generated, the digital twin unit immediately starts a pre-built multi-agent cooperative dynamics simulation with the latest measured UAV position and velocity as the initial state and the real-time control commands as the control input. This simulation, which runs faster than the actual time, predicts the state evolution and interactions of all UAVs over a preset period of time. "Running faster than the actual time" means that the simulation calculation speed is several times faster than the actual physical time elapses, such as five to ten times, to ensure that the assessment of the situation several seconds in the future is completed before the real-time control commands actually act on the physical UAVs. The preset duration is usually two to five seconds to balance the advance warning and computational complexity. During the multi-agent collaborative dynamics simulation, the dynamic operation influence envelope model is synchronously invoked to calculate the spatial risk field function value of each UAV and its distribution in three-dimensional space at each simulation step, thereby obtaining the predicted risk field distribution of the entire operation airspace over time within the future short time window. The simulation process will generate a spatial position for each UAV at each moment, called the simulation prediction position. The "simulation step" is set according to the simulation accuracy and real-time calculation requirements, usually from 0.01 seconds to 0.1 seconds. The specific process for calculating the deviation from the spatiotemporal corridor and triggering emergency compensation control is as follows: Based on the predicted risk field distribution obtained from forward-looking simulation, a deviation called the spatiotemporal consistency index is calculated. The "spatiotemporal consistency index" is used to quantify the overall deviation between the predicted future operational situation and the safety benchmark set by the global plan. The deviation calculation process is as follows: within the short time window of the simulation, a continuous time interval is selected, and for each discrete time node in this time interval, the risk intensity compliance judgment and statistics of spatial points in the three-dimensional space are performed within the entire range of the preset three-dimensional operational space. The compliance assessment method is as follows: For any three-dimensional spatial point within the operational space, if at that time node, the sum of the predicted spatial risk field function values ​​of all UAVs at that point is not greater than the maximum allowable risk intensity at that point, which is jointly defined by the spatiotemporal corridor constraints of all UAVs and is determined by an optimization threshold, then the three-dimensional spatial point is deemed compliant at that time node. This assessment logic ensures that the evaluation is based on a direct comparison of risk field intensity, rather than simple geometric location, thereby capturing potential violations caused by the abnormal accumulation of risks. After the statistics are completed, the proportion of the total volume occupied by the three-dimensional space points that are judged to be compliant within the entire work space at that time point is calculated to obtain the space compliance rate at that time point; the "total volume proportion" is approximately calculated by the ratio of the number of compliant three-dimensional space points to the total number of grid points after discretization of the work space. Next, the spatial compliance rates of all discrete time nodes within the selected time interval are averaged to obtain a spatiotemporal compliance rate. Finally, the deviation is obtained by subtracting the spatiotemporal compliance rate from one. The deviation ranges from zero to one, where zero indicates that the predicted future fully conforms to the safety corridor, and one indicates complete deviation. The calculated deviation is compared with a preset safety threshold between zero and one. The safety threshold is a small value between zero and one, and it is recommended to set it between 0.01 and 0.1, more preferably 0.05. This safety threshold defines the upper limit of the prediction deviation that the system can tolerate. Its setting needs to strike a balance between safety redundancy and avoiding false alarms. 0.05 means that the system allows up to 5% of the spatiotemporal volume in the prediction to have the risk of exceeding the limit. If the calculated deviation is less than or equal to the safety threshold, the predicted operational status is determined to be safe, the closed-loop verification execution module does not intervene, and the original real-time control command continues to be executed. If the calculated deviation is greater than the safety threshold, emergency compensation control is triggered. When triggered, the closed-loop verification execution module immediately sends an interrupt signal containing the current measured status and high-risk warning information to the online collaborative scheduling module. The "interruption signal" forces the online collaborative scheduling module to suspend the current regular scheduling cycle and switch to emergency processing mode. Upon receiving an interruption signal, the online collaborative scheduling module recalculates and generates a new set of compensatory real-time control instructions based on the latest emergency status within its policy network framework. The primary goal of compensatory real-time control commands is to reduce the deviation to below the safety threshold in the next control cycle. To this end, compensatory commands tend to adopt a more conservative control strategy than regular commands. "More conservative control strategies" usually include, but are not limited to: instructing the UAV to hover temporarily, significantly reduce its flight speed, immediately stop the painting operation, or order it to retreat a safe distance along the center line of the spatiotemporal corridor. The specific process of using the difference between the predicted data from forward simulation and the measured data for periodic optimization is as follows: Two types of difference data are continuously collected and buffered. The first type is called trajectory tracking difference, which is defined as the coordinate difference between the actual position of each UAV and its simulated predicted position in the digital twin at the same time. The coordinate difference is a three-dimensional vector, which is calculated by measuring the difference between the actual position coordinates and the simulated predicted position coordinates on each coordinate axis. The second category is called risk perception difference, which is defined as the difference between actual observation data of the coating deposition effect on the cable surface obtained by airborne visual sensors and the spatial risk field function value of the coating drift pollution risk field obtained by simulation prediction of the dynamic operation influence envelope model by digital twin units at the corresponding spatial location and time. "Actual observation data" is obtained by processing cable images taken by airborne cameras, for example by analyzing the changes in color, gloss or texture features of wet coating in the images, and using a pre-calibrated model to convert it into an estimate of coating thickness or coverage that is comparable to the risk field function value. "Difference" is the difference between this estimate and the simulation prediction value. The two types of differential data together constitute the risk perception differential data used for optimization. Risk perception discrepancy data is buffered and used to perform the following two periodic optimization processes: The first optimization involves optimizing the dynamic operation impact envelope model. Specifically, the collected risk perception difference data is used as training samples. An online learning algorithm, such as recursive least squares or Bayesian update, is employed to periodically adjust the key physical parameters in the dynamic operation impact envelope model. These key physical parameters include at least the rotor airflow field intensity coefficient, which describes the intensity of the rotor downwash airflow, and the spray diffusion coefficient, which describes the diffusion rate of paint droplets. By adjusting these parameters, the predicted output of the dynamic operation impact envelope model for the future risk field gradually approximates the actual physical effect. This optimization process is essentially an online calibration of the model parameters. For example, if the measured paint dispersion range is consistently larger than the predicted range, the online learning algorithm will correspondingly increase the spray diffusion coefficient. The optimization cycle can be set according to the data accumulation rate and model stability requirements, such as once after each section of cable is sprayed or every few minutes. The second optimization involves optimizing the online collaborative scheduling module. Specifically, it involves taking the experience tuples generated during actual operation, which consist of the UAV status, real-time control commands issued, subsequent status, and actual operation costs, as new experience samples and storing them in the experience replay buffer maintained by the policy network within the online collaborative scheduling module. This buffer is used to store historical experience data for training purposes. The actual operation costs include the actual task completion time and the actual operation interference events. The "actual operation interference events" can be determined by analyzing sensor data (such as visual analysis of coating uniformity) or by recording the number of emergency compensation controls triggered by the closed-loop verification module. During preset periods when computing resources are idle or the load is low, offline fine-tuning training of the policy network is triggered. This training process samples batches of new and old experience samples from the experience replay buffer and uses new experience samples from the real world to update the internal parameters of the policy network, so that the cooperative policies generated by the policy network can better adapt to the actual dynamic environment and unmodeled perturbations. The "offline fine-tuning training" usually uses the same reinforcement learning algorithm as the original training of the policy network, but the learning rate is set lower to prevent catastrophic forgetting. This process enables the system to continuously learn from actual operating experience and continuously improve its cooperative avoidance and dynamic adjustment capabilities.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A drone-assisted precision spraying maintenance system for cable maintenance, characterized in that, Specifically, it includes: The system consists of an influence envelope modeling module, a global spatiotemporal planning module, an online collaborative scheduling module, and a closed-loop verification execution module, which are connected sequentially. Impact envelope modeling module: Before collaborative operation, based on the rotor parameters, spraying parameters and environmental wind field data of the UAV, a corresponding dynamic operation impact envelope model is established for each UAV. This model is used to quantify the physical collision risk, rotor airflow disturbance risk and paint drift pollution risk generated by the UAV in space into a time-varying spatial risk field function. The specific process of establishing a corresponding dynamic operational influence envelope model for each UAV based on its rotor parameters, coating parameters, and environmental wind field data is as follows: It receives and processes three types of inputs: rotor parameters, spraying parameters, and ambient wind field data. Rotor parameters include the equivalent collision radius of the UAV body, rotor radius, position and normal vector of the rotor plane in the body coordinate system, and rated downdraft velocity; Spraying parameters include spraying start / stop status, nozzle position and direction vector in the machine coordinate system, initial paint spraying speed, spray cone angle, and average paint particle size; The environmental wind field data consists of real-time environmental wind speed and direction vectors; Based on rotor parameters, coating parameters, and environmental wind field data, a time-varying spatial risk field function defined in three-dimensional space is constructed for each UAV at every moment. The value range of the spatial risk field function is defined between zero and one. The function value of the spatial risk field function at any point is used to quantify the risk intensity of the corresponding point in three-dimensional space that is adversely affected by UAV operation into a scalar. A scalar value of one represents the highest risk, and a scalar value of zero represents no risk. Risks include physical collision risk, airflow disturbance risk, and paint contamination risk. The spatial risk field function is constructed by taking the maximum value of each component in a composite risk field; The composite risk field is composed of the superposition of the physical collision risk field, the rotor airflow disturbance risk field, and the paint drift pollution risk field. The function value of the physical collision risk field depends on the Euclidean distance between the spatial point in three-dimensional space and the real-time position of the UAV at the corresponding moment, and decays exponentially with the Euclidean distance as the variable. The function value of the rotor airflow disturbance risk field is determined by the rotor operating state coefficient, the ratio of the rated downdraft velocity to a reference wind speed, and the geometric relationship of a spatial point in three-dimensional space relative to the rotor, and decays with increasing distance. The function value of the paint drift pollution risk field is controlled by the spraying state and is obtained by time integration of the probability density of the sprayed droplets appearing at spatial points in three-dimensional space over an effective spraying time period from the current moment back. Numerical description of the output spatial risk field function of the dynamic operation influence envelope model on a discretized three-dimensional spatial grid; Global spatiotemporal planning module: Based on the preset task path and bridge structure model, in a four-dimensional spatiotemporal space including the time dimension, an improved conflict-based search algorithm is used to plan an initial spatiotemporal trajectory for each UAV and calculate the corresponding spatiotemporal corridor constraints. The planning process aims to minimize the total task time and the superposition integral of the operation risk field calculated based on the spatial risk field function. Online Cooperative Scheduling Module: Taking the dynamic operation influence envelope model and spatiotemporal corridor constraints as input, a dynamic heterogeneous graph is constructed, and a pre-trained graph neural network is used to extract the global cooperative state features of the dynamic heterogeneous graph. This drives the central scheduler in the online cooperative scheduling module to generate real-time control commands, which are used to fine-tune the initial spatiotemporal trajectory. Closed-loop verification execution module: While real-time control commands are issued, the real-time control commands and the measured data collected by the UAV's onboard sensors are input into the digital twin unit within the closed-loop verification execution module for forward-looking simulation. This predicts the collaborative operation status of multiple UAVs within a short time window and calculates the deviation from the spatiotemporal corridor. When the deviation exceeds a preset safety threshold, the online collaborative scheduling module is triggered to perform emergency compensation control. At the same time, the difference between the predicted data from the forward-looking simulation and the measured data is used to periodically optimize the dynamic operation influence envelope model and the online collaborative scheduling module.

2. The UAV-assisted precision spraying maintenance system for cable maintenance according to claim 1, characterized in that: In the global spatiotemporal planning module, the specific process of planning within a four-dimensional spatiotemporal space including the time dimension, based on a preset task path and bridge structure model, is as follows: First, a numerical description of the spatial risk field function is received, and a four-dimensional spatiotemporal grid map is constructed by combining it with the bridge structure model. Define a spatiotemporal trajectory for each drone from the start time to the end time; plan with the objective of minimizing a multi-objective cost function, which is the sum of the costs of all drone spatiotemporal trajectories; The cost of each spatiotemporal trajectory is a weighted sum of the time cost corresponding to its duration and the risk cost corresponding to the integral of its operational risk field. The calculation process of the superimposed integral of the operation risk field is as follows: for a spatiotemporal trajectory, from the start time to the end time, it is gradually advanced at a fixed time step; At each discrete moment reached, the spatial position of the UAV at that discrete moment is first determined, and then the spatial risk field function value of each UAV at that spatial position and discrete moment is obtained. Next, all the obtained spatial risk field function values ​​are multiplied by preset risk coupling weights representing the importance of different risk sources and then summed to obtain an instantaneous superimposed risk value at that discrete moment. Finally, multiply this instantaneous superimposed risk value by the time step, and sum up the results of such multiplications at all discrete moments to obtain the superimposed integral of the operational risk field for this spatiotemporal trajectory.

3. The UAV-assisted precision spraying maintenance system for cable maintenance according to claim 2, characterized in that: The specific process of employing the improved conflict-based search algorithm is as follows: The conflict-based search algorithm maintains a search tree containing candidate spatiotemporal trajectories for each UAV and manages conflict constraints on a constraint tree; and defines conflict expansion as hard conflict and soft conflict. Hard collisions include geometric collisions and collisions that occur when the sum of the spatial risk field functions of two drones at any point in time exceeds a hard collision safety threshold. Soft conflict occurs when the sum of values ​​exceeds an optimization threshold that is below the hard conflict safety threshold but does not reach the hard conflict safety threshold. The iterative process of the conflict-based search algorithm is as follows: each UAV searches for its initial spatiotemporal trajectory while ignoring other UAVs; Detect hard and soft conflicts between all spatiotemporal trajectories; Resolve hard conflicts and replan for affected drones; For soft conflicts, guidance is provided at higher-level nodes of the constraint tree by comparing the total cost of containing the conflict with the cost of resolving the conflict; Finally, a set of spatiotemporal trajectory solutions without hard conflicts are generated, and a corresponding spatiotemporal corridor constraint is generated for the trajectory of each UAV. The spatiotemporal corridor constraint is a four-dimensional spatiotemporal volume jointly determined by the spatiotemporal trajectory itself, the UAV body envelope, and the risk isosurface defined by the optimization threshold.

4. The UAV-assisted precision spraying maintenance system for cable maintenance according to claim 3, characterized in that: The specific process of constructing a dynamic heterogeneous graph using the dynamic operation influence envelope model and spatiotemporal corridor constraints as input is as follows: The system receives a dynamic job influence envelope model for real-time computation, an initial spatiotemporal trajectory generated for each UAV and its corresponding spatiotemporal corridor constraints, and real-time position, real-time speed, and remaining task information of the UAV from airborne sensors; based on the input, it constructs a dynamic heterogeneous graph that changes with job time. The node set of the dynamic heterogeneous graph includes drone nodes, cable segment nodes, and sensitive component nodes, each with its own attributes; The edges of the dynamic heterogeneous graph constitute an edge set, which includes at least spatial proximity edges, task-assigned edges, and risk-affected edges. The weight of the spatial proximity edges is determined by the distance between nodes. The task-assigned edges have fixed weights. The weight of the risk-affected edges is calculated in real time by calling the dynamic job influence envelope model, and the specific value is equal to the value of the spatial risk field function caused by the UAV corresponding to the source UAV node to the target node position.

5. A drone-assisted precision spraying maintenance system for cable maintenance according to claim 4, characterized in that: The specific process of extracting global cooperative state features of dynamic heterogeneous graphs using a pre-trained graph neural network and driving the central scheduler to generate real-time control commands is as follows: First, the dynamic heterogeneous graph is input into a pre-trained graph neural network, and a node embedding vector is generated for each drone node through multi-layer message passing, which serves as a global collaborative state feature. Next, the central scheduler generates real-time control commands based on the node embedding vectors of all UAV nodes through a pre-built policy network trained by multi-agent reinforcement learning. The policy network is trained by optimizing a multi-objective reward function, which is a weighted combination of task progress reward, corridor violation penalty and cooperation penalty. The collaborative penalty term is obtained by nonlinearly amplifying and integrating the interaction results of the spatial risk field functions of each pair of UAVs in the overlapping region, and is used to drive the policy network to generate fine-tuning actions to reduce the overlap of high-risk areas. The generated raw motion commands are processed by secure projection to ensure that the expected state of the UAV after application falls within its spatiotemporal corridor constraints, resulting in fine-tuning of flight speed and slight translation of the spraying start and stop sequence.

6. A drone-assisted precision spraying maintenance system for cable maintenance according to claim 5, characterized in that: The specific process of inputting real-time control commands and measured data into the digital twin unit for forward-looking simulation and predicting the collaborative operation status within a short time window is as follows: While real-time control commands are being issued, the digital twin unit receives real-time control commands, measured position and speed data of the UAV from airborne sensors, as well as dynamic operation influence envelope model and spatiotemporal corridor constraints. The digital twin unit maintains a high-fidelity simulation environment, which embeds a real-time calculation engine for the dynamic operation influence envelope model and stores all UAV state copies, environmental wind field estimates, and bridge structure models. Based on the input, whenever a new set of real-time control commands is generated, the digital twin unit immediately starts a multi-agent cooperative dynamics simulation that runs faster than the actual time, using the latest measured UAV position and speed as the initial state and the real-time control commands as the control input, to deduce the state evolution of all UAVs and their interactions over a preset period of time in the future. During the multi-agent cooperative dynamics simulation, the dynamic operation influence envelope model is synchronously invoked to calculate the spatial risk field function value of each UAV and its distribution in three-dimensional space at each simulation step. This yields the predicted risk field distribution of the entire operation airspace over time within a short future time window. The simulation process generates a spatial position for each UAV at each moment, which is called the simulation prediction position.

7. A drone-assisted precision spraying maintenance system for cable maintenance according to claim 6, characterized in that: The specific process of calculating the deviation from the spatiotemporal corridor and triggering emergency compensation control is as follows: Based on the predicted risk field distribution obtained from forward-looking simulation, a deviation called the spatiotemporal consistency index is calculated. The calculation process of the deviation is as follows: within the short future time window of the simulation, a continuous time interval is selected, and for each discrete time node in this time interval, the risk intensity compliance judgment and statistics of each spatial point in the three-dimensional space are performed in the entire range of the preset three-dimensional work space. After the statistics are completed, the proportion of the total volume of the three-dimensional space points that are judged to be compliant within the entire work space at that time point is calculated to obtain the space compliance rate at that time point. Next, the spatial compliance rates of all discrete time nodes within the selected time interval are averaged to obtain a spatiotemporal compliance rate; finally, the spatiotemporal compliance rate is subtracted from one to obtain the deviation; the calculated deviation is compared with a preset safety threshold between zero and one. If the calculated deviation is less than or equal to the safety threshold, no intervention will be taken; If the calculated deviation exceeds the safety threshold, emergency compensation control is triggered; an interrupt signal is sent to the online collaborative scheduling module. Upon receiving an interruption signal, the online collaborative scheduling module recalculates and generates a new set of compensatory real-time control instructions based on the latest emergency status within its policy network framework.

8. A drone-assisted precision spraying maintenance system for cable maintenance according to claim 7, characterized in that: The specific process of using the difference between the predicted data from the forward simulation and the measured data for periodic optimization is as follows: We continuously collect and buffer two types of discrepancy data: the first type is called trajectory tracking discrepancy; the second type is called risk perception discrepancy. The two types of differential data together constitute the risk perception differential data used for optimization. Risk perception discrepancy data is buffered and used to perform the following two periodic optimization processes: The first optimization is to optimize the dynamic operation influence envelope model. Specifically, the collected risk perception difference data is used as training samples, and an online learning algorithm is used to periodically adjust the key physical parameters in the dynamic operation influence envelope model so that the prediction output of the dynamic operation influence envelope model for the future risk field gradually approaches the real physical effect. The second optimization is to optimize the online collaborative scheduling module. Specifically, the experience tuples generated during the actual operation, consisting of the UAV status, real-time control commands issued, subsequent status, and actual operation cost, are stored as new experience samples in the experience replay buffer maintained by the policy network within the online collaborative scheduling module. This buffer is used to store historical experience data for training purposes. During a preset period when computing resources are idle, offline fine-tuning training of the policy network is triggered. New experience samples from the real world are used to update the internal parameters of the policy network, so that the cooperative policies generated by the policy network can better adapt to the actual dynamic environment.

Images