A method and system for collaborative path planning of unmanned aerial vehicle (UAV) swarms for power line inspection

Abstract

This invention discloses a method and system for collaborative path planning of UAV swarms for power line inspection. The method includes the following steps: S01. Each UAV receives encoded data from a global complex information field and stores it locally; S02. When a UAV detects a threat target in real time, it extracts feature vectors and calculates disturbance parameters. For grids within the influence range of the threat target, it calculates the anisotropic repulsion field increment and updates the corresponding safety repulsion strength value, forming a threat update message and broadcasting it for synchronous updates; S03. It calculates the fusion utility value and selects the flight target for the next control cycle; S04. Whenever a UAV arrives at a designated task point and completes a designated inspection operation, it updates the corresponding task attraction strength value and broadcasts a completion message. This invention enables time-optimal path planning and collaborative operation of UAV swarms in dynamic environments, ensuring full inspection coverage while significantly reducing overall operation time.

Description

Technical Field

[0001] This invention relates to the field of power line inspection technology, and in particular to a method and system for collaborative path planning of unmanned aerial vehicle (UAV) swarms for power line inspection. Background Technology

[0002] Regular inspections of power lines are a mandatory task to ensure power grid safety. Their primary goal is to achieve 100% inspection of all critical equipment (such as insulators and fittings) to avoid omissions. While drone swarm collaboration is an effective means to improve inspection efficiency, in real-world dynamic environments, existing drone swarm collaborative path planning methods for power line inspection struggle to simultaneously meet the dual requirements of "full coverage" and "high efficiency." Specifically, the following problems exist: 1. It is difficult to ensure full coverage and there is a risk of missed detections.

[0003] In existing technologies, drone swarms typically employ two collaborative decision-making strategies: centralized allocation and distributed allocation. Centralized task allocation can plan coverage paths, but in power line inspection environments, there are various dynamic threats (such as birds and construction machinery). Traditional centralized task allocation methods are prone to interrupting task execution and resulting in missed detections when faced with sudden dynamic threats. Distributed methods, on the other hand, lack strict global state consensus, and task blind spots or duplications are prone to occur in multi-drone collaboration, which also poses a risk of missed detections.

[0004] 2. Lack of cluster collaboration, suboptimal obstacle avoidance paths for individual machines, and extended operation time.

[0005] Existing drone swarm obstacle avoidance strategies mostly employ a reactive approach. This means each drone independently calculates and outputs obstacle avoidance actions based on real-time local environmental information acquired by sensors (such as LiDAR, cameras, and ultrasonic sensors), and then returns to its original path after avoiding an obstacle. However, in reality, the path returned to the original task sequence after a drone avoids a threat is often not optimal, and the lack of coordinated replanning at the swarm level leads to a significant increase in overall operation time due to frequent and disordered obstacle avoidance behaviors.

[0006] 3. It is difficult to reconcile time optimization with safety constraints.

[0007] Traditional drone swarm control methods typically decouple path planning from real-time obstacle avoidance, that is, separate task planning from obstacle avoidance, or only consider the shortest static distance during optimization, which cannot achieve truly time-optimal operation in dynamically changing environments.

[0008] Therefore, there is an urgent need for a collaborative path planning method for UAV swarms in power line inspection, so as to unify time optimization and safety constraints, and enable UAV swarms to perform time-optimal path planning and collaborative operations in dynamic threat interference environments, so as to shorten the overall operation time as much as possible while ensuring full coverage. Summary of the Invention

[0009] The technical problem to be solved by the present invention is as follows: In view of the above-mentioned problems existing in the prior art, the present invention provides a method and system for collaborative path planning of UAV swarms for power inspection that is simple to implement, low in cost, highly efficient in control and with high coverage. It can realize time-optimal path planning and collaborative operation of UAV swarms in dynamic environments, and can significantly shorten the overall operation time while ensuring inspection coverage, thereby significantly improving the efficiency and economy of power inspection.

[0010] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for collaborative path planning of UAV swarms for power line inspection, comprising the following steps: Step S01. Each UAV in the cluster receives the encoded data of the global complex information field distributed by the ground control station and stores it locally. The global complex information field is formed by discretizing the three-dimensional space of the power line corridor into a set of grids, and storing each grid in the set. Use task attraction intensity value Safety Repulsion Strength Value Define complex information field value The task attraction strength value is constructed and formed. Used for quantizing grid position The security repulsion strength value is used as the priority of inspection targets to characterize the attractiveness of the task. Used to quantify the arrival and safe passage of drones through grid positions. The time cost required is used to characterize the environmental safety repulsion force; Step S02. When each UAV in the cluster detects a threat target in real time during flight, it extracts a feature vector from the detected threat target, calculates a disturbance parameter based on the extracted feature vector to assess the impact range of the threat target, calculates anisotropic repulsion field increment based on the disturbance parameter for the grid within the impact range of the threat target, updates the security repulsion strength value of the corresponding grid based on the anisotropic repulsion field increment, and broadcasts a threat update message formed by the feature vector, disturbance parameter, and the location of the threat target so that the other UAVs can synchronously update the corresponding security repulsion strength value. Step S03. Each UAV in the cluster determines, in real time, the set of neighboring grid cells that can be reached within the next control cycle based on the current state and dynamic constraints of the UAV. Then, based on each candidate grid cell in the neighboring grid set... The safety repulsion strength value and the mission attraction strength value are used to calculate the corresponding fusion utility value, and the candidate grid with the largest fusion utility value is selected as the flight target for the next control cycle. Step S04. Whenever a drone arrives at a designated task point and completes a designated inspection operation, update the corresponding task attraction intensity value to the preset value and broadcast a completion message to synchronize the task attraction intensity value of each drone until the task attraction intensity value of all drones in the cluster reaches the preset value, thus completing the inspection task.

[0011] Further, in step S01, the global complex information field is constructed according to the following steps: Step S101. Initialize the task attraction field: The task point set consists of the key equipment nodes to be inspected on the power line. M represents the number of critical points to be inspected. These critical equipment nodes include insulator strings, conductor joints, and key tower components. For the task point set... At each key point in the grid, the task attraction intensity value The initial configuration is set to the preset maximum task attraction intensity value of the inspection target. The guide path is formed by a continuous grid along the direction of the conductor. For the guidance path The grid on By grid The geometrical spatial distance between the nearest task point and the task attraction strength value decreases non-linearly: , For grid To the nearest mission point Euclidean distance, To control the spatial decay coefficient that causes the task attraction intensity to decrease exponentially with increasing spatial distance, the task attraction intensity values ​​of grids on non-inspection task points and non-guided paths are configured to be 0, forming an initial task attraction intensity field. ; Step S102. Initialize the safety repulsion field: Initialize the safety repulsion strength value of the grid set occupied by the known static obstacle body to a preset maximum penalty value or infinity to construct an absolute hard no-fly constraint. Initialize the safety repulsion strength value of the boundary buffer grid within a preset distance outside the static obstacle to decrease with distance. B (g)= ,in, This represents the basic safety repulsion strength value of a static obstacle. This represents the attenuation coefficient of the static obstacle boundary buffer zone. Indicates the current grid position The Euclidean distance to the outer physical boundary of the nearest static obstacle is used to initialize the safety repulsion strength value of the remaining absolutely safe free-space grid outside the buffer zone to 0, forming the initial safety repulsion strength field. ; Step S103. Based on the initial task attraction intensity field and safety repulsion strength field To construct and form a global complex information field .

[0012] Further, in step S02, the threat target includes birds and drones, and the elements in the feature vector include the speed of the threat target relative to the drone. Current Euclidean distance between the threat target and the drone The unit vector of the threatening target's direction of movement. And the largest eigenvalue of the spatial distribution covariance matrix of the threat target point cloud. The disturbance parameters include disturbance intensity S and first influence radius. Second radius of influence and the time delay spatial decay sensitivity factor The first radius of influence The second radius of influence corresponds to the boundary of the core area adjacent to the threatening target. This corresponds to the boundary of the outer transition buffer zone.

[0013] Furthermore, in step S02, a local threat effect estimator composed of a three-layer fully connected feedforward neural network is used. The local threat effect evaluator calculates the perturbation parameters. It consists of an input layer, two hidden layers, and an output layer. The input layer takes as input the feature vector extracted from the threat target. The output layer outputs the corresponding perturbation parameters. , This represents the evaluator parameters.

[0014] Furthermore, in step S02, a directional gain function is defined based on the perturbation parameters. and the radius of influence of anisotropy : ,

[0015] in It is the backward gain factor; ,

[0016] Based on the disturbance parameters and directional gain function and the radius of influence of anisotropy The anisotropic repulsive field increment was calculated as follows: ,

[0017] in, Geometric vectors of relative position Unit vector of the threat's movement direction The angle between them, the geometric vector of their relative positions Defined as from the center of the threat target Pointing to the associated reference grid position in space The vector, Represents Euclidean distance. For the disturbance parameters, Indicates the intensity of the disturbance. Indicates the first radius of influence, Indicates the second radius of influence and the first radius of influence. Corresponding to the core area boundary adjacent to the threatening target, the second radius of influence This corresponds to the boundary of the outer transition buffer zone. This represents the spatial decay sensitivity factor of time delay.

[0018] Furthermore, in step S02, for those in a threatened position Grids within the affected area Update the corresponding safety repulsion strength value according to the following formula. : , in, Represents the raster at time t+1 Updated safety rejection strength value, Represents the raster at time t The safety rejection strength value; In each global synchronization cycle Security rejection strength for all grids Exponential decay: ,

[0019] in, Representing grids The security rejection strength value before and after the global synchronization cycle update. This is the attenuation coefficient.

[0020] Further, in step S03, candidate grids are calculated according to the following formula. Fusion utility value: , in, >0 indicates an adjustable weight parameter. Representing candidate grids The task attraction strength value and the safety repulsion strength value.

[0021] Furthermore, in step S04, each time the drone arrives at the mission point and completes the inspection operation, a lightweight field state update message is also generated. And broadcast a task completion message, the lightweight field state update message The message carries the following information: message type, message sequence, source UAV, target grid index, and updated task attraction intensity value; when the UAV receives the lightweight field state update message... At that time, check the message sequence number. If it is a processed message, discard it. If it is a task completion message, directly update the task attraction strength value of the corresponding grid in local storage. If it is a threat update message, use the same local threat effect evaluator. Recalculate the perturbation parameters and superimpose the same anisotropic repulsive field increment once. To update the corresponding safety rejection strength value.

[0022] Furthermore, during mission execution, each drone in the cluster collects data for... ,in Indicates the obstacle avoidance result. This represents the feature vector extracted from the threatening target. This represents the perturbation parameters, and after the task is completed, it also includes the use of the collected data to... Optimize evaluator parameters using supervised learning The loss function used is: , in, λ The regularization coefficient is . N Represents the total number of samples. i Indicates the sample index. Indicates using the current evaluator parameters For the first Feature variables of each sample The four-dimensional perturbation parameter prediction vector output after forward propagation prediction. This represents the preset parameter dimension balance matrix.

[0023] A drone swarm system includes multiple drones, each drone being connected to a wireless ad hoc network. Each drone is equipped with a processor and a memory, the memory being used to store computer programs, and the processor being used to execute the computer programs to perform the methods described above.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention encodes the task attraction strength, which represents the attractiveness of the task, and the security repulsion strength, which represents the safety / time cost, into a complex field. The task attraction strength is used as the real part, and the security repulsion strength is used as the imaginary part. This constructs a unified complex information field to represent the attractiveness of the inspection task and the environmental security repulsion force. This allows task attraction and security repulsion to be processed within the same computational framework, achieving a balance between time optimization and security constraints. At the same time, it adopts a distributed architecture, with each UAV maintaining a local field copy. Through a closed loop of local perception, local update, autonomous decision-making, and synchronous update, it can realize time-optimal path planning and collaborative operation of UAV swarms in dynamic environments.

[0025] 2. This invention enables strict inspection point status management within a complex information field framework, ensuring comprehensive coverage of inspection targets without omissions. It solves the problem of missed inspections caused by the separation of task planning and obstacle avoidance in traditional methods. The complex field allows for single-serialized transmission, which also reduces communication overhead. Thus, while ensuring 100% inspection coverage, the overall operation time is significantly shortened, significantly improving the efficiency and economy of power inspection.

[0026] 3. This invention further constructs an anisotropic threat field increment model, which enables the region in front of the threat movement to obtain a higher field increment and the region behind it to have a lower gain. This allows for more accurate time cost modeling, solves the problem of conservative obstacle avoidance paths caused by traditional symmetrical danger zone models, and avoids excessive detours.

[0027] 4. The present invention further adopts a nonlinear utility function based on time cost to decide the next flight target, so that the UAV automatically performs the minimization of the expected travel time cost at each decision moment, selects the locally time-optimal feasible path, realizes time-optimal path planning, and at the same time, the attractiveness of high time cost areas is drastically suppressed, so as to achieve better time efficiency in dynamic threat environments.

[0028] 5. Furthermore, by adopting an incremental synchronization mechanism, this invention can ensure synchronization between drones by broadcasting only lightweight feature parameters, without the need to transmit complete field data or conduct complex negotiations. This can significantly reduce communication overhead and achieve low-communication-overhead cluster collaboration. Attached Figure Description

[0029] Figure 1 This is a schematic diagram illustrating the implementation process of the UAV swarm collaborative path planning method for power line inspection in this embodiment.

[0030] Figure 2 This is a schematic diagram illustrating the complete process of drone swarm collaborative path planning in a specific application embodiment of the present invention. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present invention.

[0032] like Figure 1 As shown, the steps of the UAV swarm cooperative path planning method for power line inspection in this embodiment include: Step S01. Local Storage of Complex Information Field: Each UAV in the cluster receives the encoded data of the global complex information field distributed by the ground control station and stores it locally. The global complex information field is formed by discretizing the three-dimensional space of the power line corridor into a set of grids, and storing each grid in the set locally. Use task attraction intensity value Safety Repulsion Strength Value Define complex information field value The formation of task attraction intensity value Used to characterize grid position As a priority for inspection targets, the safety rejection strength value Used to quantify the arrival and safe passage of drones through grid positions. The time cost required.

[0033] In this embodiment, the three-dimensional space of the power line corridor is first obtained, which is the three-dimensional strip-shaped area defined on both sides and above the transmission line path to ensure the safe operation of the overhead transmission line. Specifically, the step of discretizing into a uniform grid set includes: First, using the center line of the overhead transmission line to be detected as a reference, and in combination with a preset safety distance, the axial length, horizontal width, and vertical height of the power line corridor are configured to determine the absolute physical boundary of the three-dimensional space; then, a three-dimensional rectangular coordinate system is established in the determined three-dimensional space, with any specified reference point in the space (such as the center foundation point of a specific tower) as the origin, and the X-axis, Y-axis in the horizontal plane and the Z-axis perpendicular to the horizontal plane are established respectively; then, the spatial step size of uniform discretization is set. According to the spatial step size The physical dimensions corresponding to the X, Y, and Z axes are divided at equal intervals, thereby dividing the entire three-dimensional strip region into uniformly sized cubic units. Each cubic unit is a uniform grid and is indexed by a corresponding coordinate. Unique identifier, among which These represent the X, Y, and Z axis coordinates, respectively, and are then used for each grid cell. Define complex information field value: (1) in, To determine the intensity of the task's attraction, quantify the priority of this location as an inspection target; The safety repulsion strength characterizes the level of obstacle or threat at a location; its numerical value quantifies the additional time cost required for a drone to arrive and safely pass through the grid. It is the imaginary unit.

[0034] In order to give the complex information field a clear physical quantization boundary, in this embodiment, the task attraction intensity With safety repulsion strength The initial spatial field distributions are constructed in the following ways:

[0035] Regarding the intensity of task attraction The task points are first composed of the key equipment nodes on the power line to be inspected (including insulator strings, conductor joints, and key components of the towers). ,in This indicates the total number of critical points to be inspected; simultaneously, the guide path for the UAV is formed by a continuous grid along the direction of the overhead power line, which can specifically be an overhead phase line, split conductor, or lightning protection wire used for transmitting electrical energy in a high-voltage transmission line. For any grid within the guide path... Its initial task attraction intensity value Follow the grid to the nearest task point The geometric spatial distance is configured in a non-linear decreasing manner, and the specific calculation formula is as follows: ;in, This represents the maximum task attraction intensity value of the preset inspection target. For grid To the nearest mission point Euclidean distance, This is a spatial attenuation coefficient used to control the exponential decay rate of task attraction intensity with increasing spatial distance; for a grid located at a task inspection point, the task attraction intensity value... Initial configuration as For the remaining safe airspace grid cells that are neither inspection task points nor guidance paths, their initial task attraction intensity value is... The initial configuration is then set to 0.

[0036] For safety rejection strength It is used to quantify the arrival and safe passage of drones through grid positions. The required time cost is used to characterize the environmental safety repulsion force, thereby finely quantifying the risk cost posed by static environmental obstacles and surrounding safety boundaries to UAV flight. Specifically, the safety repulsion strength value of the grid set occupied by the known static obstacle body is initialized to a preset maximum penalty value or infinity to construct an absolute hard no-fly constraint; while for the boundary buffer grid within a preset distance outside the static obstacle, its safety repulsion strength value... The initial configuration is set to exhibit exponential slip decay with distance, and the specific calculation formula is as follows: B (g)= ;in, This represents the basic safety repulsion strength value of a static obstacle. This represents the attenuation coefficient of the static obstacle boundary buffer zone. Indicates the current grid position The Euclidean distance to the physical boundary outside the nearest static obstacle; the initial safety repulsion strength value for the remaining absolutely safe free-space grid outside the buffer zone. If the value is set to 0, it means that passing through this area will not incur any additional security time costs.

[0037] This embodiment uses the task attraction intensity, which characterizes the attractiveness of a task, to represent that task appeal. With characterization of security / time cost Safety repulsion strength is uniformly encoded as a complex field. This can provide a natural mathematical basis for subsequent fusion decision functions, enabling task attraction and safety repulsion to be processed within the same computational framework. At the same time, complex fields can be serialized and broadcast as a single data structure, reducing communication overhead. This allows for unified encoding and efficient transmission of task attraction and safety repulsion fields.

[0038] Specifically, the detailed steps for constructing the global complex information field are as follows: Step S101. Initialize the task attraction field: The task point set consists of the key points to be inspected on the power line. Task point set Task attraction intensity values ​​of grids at each key point in the middle Initial configuration as The guide path is formed by continuous grids along the conductor direction in the three-dimensional grid set of the power line corridor. For the guidance path The upper grid is assigned decreasing task attraction strength values ​​based on its distance from the task point: And configure the task attraction intensity value of grids on non-inspection task points and non-guided paths to 0 to form the initial task attraction intensity field. .

[0039] Specifically, guiding path It consists of a continuous grid along the overhead power line, where the power line specifically refers to overhead phase lines, split conductors, and lightning protection wires used for transmitting electrical energy in high-voltage transmission lines. The three-dimensional space of the power line corridor is obtained through the following methods: before the drone swarm takes off, an airborne lidar is used to perform a pre-inspection scan of the power line corridor to obtain three-dimensional point cloud data of the corridor, and the central axis trajectory of the conductor is extracted using a three-dimensional reconstruction algorithm; or the spatial coordinates of the conductor are directly imported from the 3D model of the transmission line's geographic information system or building information model. Physically, the conductor and the power line to be inspected have an inclusive relationship; that is, the conductor is the core physical structure constituting the main body of the entire power line's spatial span, while the key points to be inspected are located at specific spatial positions on the conductor or tower. All pre-set inspection points are completed. After initialization, the values ​​form a unique list of tasks that ensures completeness, providing a foundational data guarantee for achieving 100% coverage.

[0040] Due to the set of task points This involves multiple independent inspection targets. To accurately quantify the geometric proximity relationship between each spatial grid and multiple task points, this embodiment employs a global minimum retrieval strategy to determine the distance between the current grid and multiple task points. Specifically, for any grid in the space, the distance to all grid points in the set is calculated. The Euclidean distances between each task point are calculated, and the minimum value among them is selected as the Euclidean distance from that grid point to the nearest task point. This ensures that each grid is only subjected to the strong attraction radiation of the task point closest to it, avoiding field distortion caused by the superposition of attraction between multiple task points.

[0041] The underlying physical principle behind constructing the initial task attraction field in this embodiment lies in building an adaptively iterative distributed virtual ledger using the real part of a complex field. In the initial state, each key point to be inspected is assigned a preset maximum task attraction strength value. The purpose is to forcibly anchor the hard boundary of the inspection task in the digital space, ensuring that all key points have the highest access priority in the initial state and eliminating the risk of missed inspections. Furthermore, the guiding path... The grid on the grid forms a nonlinear decreasing gradient guidance field based on the direction of the conductor. Its core function is to provide a long-range potential field traction mechanism for UAVs flying over long distances: when the UAV is far away from all task points and cannot directly perceive the high concentration of field strength at the task points, as long as the UAV enters the conductor buffer zone of the power corridor, the gradient guidance field can provide deterministic flight control guidance for the UAV through the local field strength gradient direction. This guides the UAV to smoothly and efficiently approach the next nearest inspection target along the conductor direction, avoiding disordered flight caused by blindly searching in the airspace. At the underlying level, the complex global three-dimensional full-coverage path planning problem is cleverly transformed into local autonomous decision-making control of the UAV as it ascends along the field strength gradient, fundamentally ensuring the optimality of the overall operation time of the cluster.

[0042] Unlike area coverage tasks, power line inspection points (insulators, fittings) are linearly distributed along the conductor. In this embodiment, the task attraction intensity is set along the conductor's direction. The gradient value forms a "pipeline-like" guiding field, allowing the drone to fly naturally along the route and avoid ineffective searches. Furthermore, the task attraction strength value... As a distributed task ledger, by employing a method of setting unchecked high values ​​and checked values ​​to zero, the task attraction intensity values ​​of all drones are... When all values ​​are 0, the inspection task can be considered complete, thus effectively ensuring comprehensive testing and no omissions.

[0043] The power corridor is densely populated with static obstacles such as iron towers and high-voltage power line towers. In this embodiment, the task point grid is initialized to high. The system sets up a value zone and an obstacle buffer zone for attenuation, enabling the drone to detect and avoid obstacles in advance, thus avoiding emergency braking.

[0044] Step S102. Initialize the safety repulsion field: Initialize the safety repulsion strength value of the grid set occupied by the known static obstacle body to a preset maximum penalty value or infinity, and initialize the safety repulsion strength value of the boundary buffer grid within a preset distance of the static obstacle to decrease with distance: B (g)= The safety repulsion strength values ​​of the remaining absolutely safe free space grids outside the buffer zone are initialized to 0, forming the initial safety repulsion strength field. .

[0045] As an optional implementation, the buffer zone grid around obstacles can be determined through the following spatial expansion mechanism: First, based on the obtained three-dimensional geometric outline of the static obstacles (including high-voltage towers, tension towers, surrounding vegetation, and intersecting buildings), the grids occupied by these obstacles in the three-dimensional grid space are marked as the core obstacle grid set. Then, using the outer boundary space of the core obstacle grid set as a reference, equidistant three-dimensional spatial expansion is performed outwards. The characteristic radius of the expansion is determined by a combination of the safe flight distance threshold of the UAV and the positioning error of the onboard sensors. Thus, all the spatial grids within this equidistant expansion radius that are not occupied by the obstacles themselves are defined as the buffer zone grid around the obstacles. Within the buffer zone grid, the safety repulsion strength value is based on an exponential decay formula. B (g)= The safety repulsion strength decreases non-linearly with distance; that is, as the physical distance between the spatial grid and the obstacle increases, its safety repulsion strength decreases exponentially with the natural logarithm. The base value decreases exponentially, thus creating a soft safety barrier in the digital space that is stronger and weaker from the inside out for static obstacles.

[0046] In the aforementioned attenuation mechanism and calculation formula, the Euclidean distance... Specifically, it represents a three-dimensional spatial grid representing the current safety repulsion strength to be calculated. The central three-dimensional three-axis spatial coordinates, and the distance from the grid in the core obstacle grid set. The three-dimensional geometric straight-line distance between the spatial coordinates of the center of the nearest static obstacle boundary grid. To achieve accurate obstacle avoidance planning in complex spatial environments with multiple obstacles, this embodiment calculates... The local nearest neighbor search strategy is adopted. That is, for any grid cell to be searched in the buffer, the geometric Euclidean distance from it to all grid cells on the outer surface of static obstacles in space is calculated, and only the minimum value is selected as the final distance input for that grid cell. The above measurement method not only accurately defines the physical spatial proximity relationship between the grid and obstacles, but also ensures at the underlying control mechanism that the safety repulsion field can adaptively and smoothly wrap the boundaries of complex obstacles of any shape, avoiding the drone from making false obstacle avoidance actions in the feasible airspace due to the blind superposition of repulsive forces between multiple static obstacles.

[0047] Specifically, the detailed steps for initializing the safety repulsion field are as follows: First, the initial safety repulsion strength value of the entire three-dimensional spatial grid set of the power line corridor is uniformly initialized to 0, to characterize that the safe free space grid does not have any additional safety risks or time costs in the initial state; then, the safety repulsion strength value of the core grid set actually occupied by known static obstacles (including high-voltage towers, buildings, and mountain vegetation) is forcibly rewritten and configured to the preset maximum penalty value. This is to construct an absolutely insurmountable hard no-fly zone boundary; finally, the safety repulsion strength value is determined on the buffer zone grid extending from the perimeter of the static obstacle. Then, depending on the geometric distance from the current grid position to the nearest static obstacle surface... As the number of elements increases, it exhibits exponential, non-linear sliding decay: B (g)= Through this dynamic configuration of the spatial field, the originally discrete and isolated static obstacle physical boundaries are transformed into a continuous risk warning gradient field that decreases in strength from the inside out in the three-dimensional digital space. This ensures that the UAV can make smooth obstacle avoidance path adjustments in advance by sensing the gradually increasing repulsive field strength before approaching dangerous areas such as high-voltage towers, thus fundamentally eliminating emergency braking and mechanical oscillations caused by sudden changes in field strength when approaching obstacles.

[0048] Step S103. Based on the initial task attraction intensity field and safety repulsion strength field To construct and form a global complex information field .

[0049] Initialize the task attraction intensity field and safety repulsion strength field Then, according to This allows the initial global complex information field to be constructed. The ground control station will construct a global complex information field. Encoded as a binary stream, it is distributed to all inspection drones via a reliable broadcast protocol. Each drone receives and decodes the data, then stores a copy in its local storage. Furthermore, to accommodate heterogeneous drones, different resolution versions of field data can be provided during distribution based on the drone's computing capabilities.

[0050] Step S02. Threat Effect Assessment and Local Security Field Update: When each UAV in the cluster detects a threat target in real time during flight, it extracts a feature vector from the detected threat target, calculates a disturbance parameter based on the extracted feature vector to assess the impact range of the threat target, calculates anisotropic repulsion field increment based on the disturbance parameter for the grid within the impact range of the threat target, updates the security repulsion strength value of the corresponding grid based on the anisotropic repulsion field increment, and broadcasts the feature vector, disturbance parameter, and the location of the threat target to form a threat update message so that the other UAVs can synchronously update the corresponding security repulsion strength value.

[0051] In this embodiment, the threat targets include birds and drones. The feature vector extracted from the detected threat targets includes elements such as the speed of the threat relative to the drone. Current Euclidean distance between threats and drones Unit vector of the threat's direction of movement And the largest eigenvalue of the spatial distribution covariance matrix of the threat point cloud. , by the largest eigenvalue It can characterize the spatial expansion uncertainty of the target.

[0052] As an optional implementation, each drone With a fixed period Perform the following sensing process: Step S201. Acquire local point cloud and image data using airborne millimeter-wave radar and visual sensors; Step S202. Fuse multi-sensor data for dynamic target detection and tracking, and identify a set of moving threat targets. Specifically, improvements can be adopted. Clustering algorithms are used to adapt to the accuracy differences of different sensors.

[0053] Step S203. For each threat target Extracting feature vectors .

[0054] Each drone in the cluster assesses the local threat effect in real time based on the extracted feature vectors of threat targets, and calculates a perturbation parameter as a dynamic adjustment coefficient for the threat impact range. Then, based on the assessment results, the location of the threat target is determined, and the perturbation parameter is multiplied by a preset base threat radius to accurately determine the dynamic impact range of the threat target's location. For grids within the threat target's impact range, anisotropic repulsion field increments are calculated, and the security repulsion strength stored locally on each drone is updated based on these increments.

[0055] As an optional implementation, a local threat effect estimator consisting of a three-layer fully connected feedforward neural network is used. Achieve local threat effect assessment, calculate perturbation parameters, and use a local threat effect evaluator. It consists of an input layer, two hidden layers, and an output layer. The input layer takes the feature vector as input. The output layer outputs the corresponding perturbation parameters. Specifically, this is expressed as the disturbance intensity First radius of influence Second radius of influence and the time delay spatial decay sensitivity factor The constructed four-dimensional vector: Among them, the disturbance intensity The first impact radius is positively correlated with the expected detour time delays that the threat may cause. With the second radius of influence They jointly define the spatial range in which time delays may occur, and satisfy... Geometric constraints.

[0056] Specifically, the first radius of influence The corresponding core area boundary is located adjacent to the threat target. This boundary characterizes the high-risk core repulsion area boundary adjacent to the threat source. Within this radius, the repulsion field strength remains at its maximum or slowly decays, forcing the UAV to perform large-angle safe evasion. The second influence radius... This corresponds to the boundary of the outer transition buffer zone, used to characterize the boundary of the outer low-risk transition buffer zone, within the first influence radius. To the second radius of influence Within the spatial region, the repulsive field strength undergoes anisotropic diffusion and rapidly decays to zero as the spatial distance from the threat source increases. Thus, a multi-level risk transition boundary, ranging from strong to weak, is constructed in space through the first and second influence radii, which together define the segmented decay interval of the anisotropic diffusion of the repulsive field as the spatial distance increases. This represents the internal parameters of the evaluator, specifically the parameter set of the feedforward neural network, which consists of the connection weights and biases between neurons in each layer. Parameter set Mathematically, it is determined from the input feature vector Output disturbance parameters The nonlinear mapping relationship, i.e. .

[0057] For example, each drone is equipped with a local threat effects assessor. This evaluator enables local threat effect assessment. A three-layer fully connected feedforward neural network is used, and its structure is as follows: Input layer: 4 dimensions, corresponding to feature vectors ; Hidden layer 1: Dimension 8, using Activation function; Hidden layer 2: Dimension 4, using Activation function; Output layer: Dimension 4, corresponding to perturbation parameters A linear activation function is used. Evaluator parameters. Maintain consistency within the cluster; acquired through offline learning.

[0058] Specifically, the local threat effects assessor Network parameter set The learning process can be achieved through the following steps: During the offline preparation phase, a training set containing a large number of "threat scenarios - detour costs" samples is first constructed based on historical UAV operation datasets and a power corridor simulation environment. The input of each training sample is a feature vector of a specific threat target, and the ground truth value at the output is quantitatively calculated as follows: After setting the threat target in the simulation environment, the actual flight time required for the UAV to perfectly detour around the threat is calculated using a global optimal path planning algorithm. This time is then subtracted from the theoretical flight time under no-threat conditions, and the difference is defined as the expected detour time delay. Based on this expected detour time delay, the corresponding disturbance intensity is quantitatively calculated according to the positive correlation function. The true value, combined with the requirements for safe isolation distance, is used to quantitatively determine the first radius of influence. Second radius of influence and sensitivity factors The true values ​​of these parameters, combined together, constitute the perturbation parameters. The ground truth labels are used to establish the input. With output The deterministic correspondence between the samples is established. Subsequently, the error backpropagation algorithm is used, with the mean squared error as the loss function, to iteratively train the network on the sample training set, and the parameter set inside the network is continuously corrected by gradient descent. This process continues until the loss function converges, thus determining and solidifying the network parameters. In subsequent actual online flights, the solidified network parameter set... Keep it constant, thus ensuring that when the same threat feature vector is input... At that time, the local threat effect assessor can calculate and output the corresponding perturbation parameters in real time, stably and deterministically. .

[0059] As an optional implementation, the anisotropic repulsive field increment is calculated according to the following steps:

[0060] First, define the directional gain function based on the disturbance parameters. and the radius of influence of anisotropy for: (2) in This is the backward gain factor, and its specific value range can be 0.2 to 0.5; (3) Based on the disturbance parameters and directional gain function and the radius of influence of anisotropy The formula for calculating the increment of the anisotropic repulsive field is as follows: (4) in, Indicates associated reference grid The anisotropic repulsive field increment at that location, Geometric vectors of relative position Unit vector of the threat's movement direction The angle between them, the geometric vector of their relative positions Defined as from the center of the threat target Position points to the associated reference grid in space The position vector is used to ensure the determinism and geometric explanatoryness of the anisotropic repulsive field increment calculation. This represents the Euclidean distance.

[0061] Specifically, relative position geometric vector Used to accurately characterize the relative spatial orientation and physical distance of the currently assessed spatial grid point relative to the center of a static or dynamic threat source in three-dimensional space, where the vector middle Represents the center grid coordinates of the threat target (i.e., the physical geometric center of the threat target). The evaluation grid represents the increment of the repulsive field to be calculated. Associated reference grid coordinates located on the same threat impact surface or equipotential surface. The coordinates of the center grid of the threat target can be determined as follows: Starting from the current direction of movement of the threatening target, the unit vector... Extend, and the extension length is equal to the grid cell to be evaluated. to center grid coordinates The distance, i.e., its expression is: Using the aforementioned geometric mapping rules, the coordinate difference between the two constructs the relative position geometric vector. (Threat movement direction unit vector) The physical essence of the threat is the dimensionless unit direction vector obtained by normalizing the velocity vector of the threat target in real time in three-dimensional space. The threat movement direction specifically represents the real-time displacement velocity vector direction of the threat target relative to the three-dimensional global coordinate system of the power corridor at the current sampling time, which is used to characterize the real-time movement trend of dynamic threats causing risk spillover to the surrounding airspace.

[0062] As shown in equation (4), the included angle is... Corresponding perturbation parameters output by the evaluator It has explicit control relationships, including the angle. Together with the corresponding perturbation parameters, these serve as core update variables and are jointly input into the subsequent formula for calculating the anisotropic repulsive field increment; among them, the perturbation intensity output by the neural network estimator... First radius of influence Second radius of influence and spatial attenuation sensitivity factor Together, they determine the fundamental energy levels of the repulsive field and the geometric boundary of the anisotropic space, while the included angle... Then, as a key azimuth control variable, it is used through... The trigonometric function transformation dynamically assigns weights to the repulsion field intensity increments of grids with different spatial orientations, thereby providing asymmetric risk weights when updating the local safety repulsion intensity of the UAV.

[0063] Furthermore, for those in a threatening position Grids within the affected area The safety repulsion strength value stored locally on the drone can be updated according to the following formula. : (5) in, This represents the updated safety repulsion strength value of the drone at time t+1. This represents the safety repulsion strength value of the drone at time t.

[0064] Dynamic threats such as flocks of birds and construction machinery frequently occur along power lines. Traditional potential field methods typically treat threats as circular, symmetrical danger zones, failing to distinguish the direction of threat movement, leading to conservative or inefficient obstacle avoidance paths. This embodiment constructs the aforementioned anisotropic threat field incremental model and extracts threat feature vectors. Conduct local threat effect assessment, and then use the directional gain function. and the radius of influence of anisotropy Calculate the increment of the anisotropic repulsive field This allows the area in front of the threat to receive a higher field increment (i.e., a higher time cost), while the area behind it receives a lower gain. This aligns with the physical intuition that "oncoming threats need to be avoided earlier." It can balance safety and efficiency to achieve more accurate time cost modeling, enabling drones to avoid oncoming threats in advance and reduce unnecessary detours when threats approach from the side or rear. This solves the problem of conservative obstacle avoidance paths caused by traditional symmetrical danger zone models, avoids excessive detours, and thus achieves low-communication-overhead cluster collaboration.

[0065] This embodiment also includes introducing a field attenuation mechanism to prevent the infinite accumulation of the impact of historical threats. Specifically, in each global synchronization cycle... For all grids Safety Rejection Strength Exponential decay: (6) in, Representing grids The security rejection strength value before and after the global synchronization cycle update. This is the attenuation coefficient, with a value ranging from 0.8 to 0.95. Furthermore, it can be configured to completely remove the corresponding field disturbance within the next three synchronization cycles after threat tracking is lost. This achieves periodic exponential attenuation of the security repulsion strength. Field values ​​can avoid conservative path planning caused by the permanent remnants of historical threats.

[0066] In this embodiment, whenever the drone detects a threatening target, it also includes extracting the feature vector. Calculated disturbance parameters and the location of the threat target Threat update messages are generated and broadcast, and the recipients then independently recalculate using the same evaluator. By superimposing and updating the security rejection strength value, the synchronization results of each drone can be kept consistent, achieving cluster state consistency. Furthermore, this is achieved by broadcasting only lightweight feature parameters. It eliminates the need to transmit complete field data and solves the problem of high communication overhead caused by the traditional method of transmitting complete field data, thus reducing communication overhead and making it particularly suitable for the harsh communication environment of power inspection.

[0067] Step S03. Real-time autonomous planning based on complex information field: Each UAV in the cluster determines the set of neighboring grids that can be reached in the next control cycle based on the current state of the UAV and dynamic constraints. Then, based on each candidate grid in the neighboring grid set... The safety repulsion strength value and the mission attraction strength value are used to calculate the corresponding fusion utility value, and the candidate grid with the largest fusion utility value is selected as the flight target for the next control cycle.

[0068] Specifically, based on the current state and dynamic constraints of the UAV, the set of neighboring grids that can be reached within the next control cycle is determined. The specific screening mechanism is as follows: Within the current control cycle, the current state information of the UAV is first acquired. The current state information includes at least the real-time position coordinates, real-time velocity vector, and real-time heading angle of the UAV in the three-dimensional global coordinate system at the current sampling time. At the same time, the inherent dynamic constraint parameters of the UAV are retrieved. The dynamic constraint parameters include at least the preset maximum flight speed, maximum acceleration, maximum overload coefficient, and maximum turning climb angle. Subsequently, based on the real-time position coordinates and maximum flight speed, combined with the step size of the current control cycle, all candidate three-dimensional grids within the spatial sphere envelope centered on the current position coordinates and with the maximum feed distance as the radius are initially traversed. Based on this, the reachability of the candidate grids initially traversed is screened using dynamic constraint parameters: First, using the maximum acceleration and maximum overload coefficient, the minimum turning radius required for the UAV to transition from its current real-time velocity vector to the direction pointing to each candidate grid is calculated, eliminating reachable grids that cannot be safely turned in the next control cycle due to excessive curvature; Second, the pitch angle difference and horizontal yaw angle difference between the current real-time heading angle and the relative azimuth angle pointing to the candidate grid are calculated, and these differences are compared with the maximum turning climb angle and the maximum dynamic turning angle, respectively, eliminating grids that exceed the UAV's maneuver rate limit. Finally, the remaining candidate grids that simultaneously meet the above speed, acceleration, turning radius, and angular maneuver constraints are retained, forming a set of neighboring grids that can be reached in the next control cycle.

[0069] In this embodiment, based on the current state (location) of the drone ,speed ) and dynamic constraints (maximum speed) Maximum acceleration Determine the set of neighboring grid cells that can be reached within the next control cycle. Then, the neighborhood grid set Each grid cell within the neighborhood is considered a candidate grid cell. The size of the neighborhood can be determined based on the drone's performance; high-performance drones can explore larger neighborhoods. This is then determined based on the set of neighborhood grid cells. Each candidate grid The safety repulsion strength value and the task attraction strength value are used to calculate the fusion utility value. The candidate grid with the largest fusion utility value is selected as the flight target for the next control cycle. This can achieve cluster collaboration and make the recovery path after obstacle avoidance optimal, thus avoiding the extension of operation time caused by the UAV due to dynamic obstacle avoidance.

[0070] As an optional implementation, for each candidate grid Calculate the candidate grid according to the following formula. Fusion utility value: (7) in, >0 indicates an adjustable weight parameter, for example, it can be set to 0. =0.5, =0.5, Representing candidate grids The task attraction strength value and the safety repulsion strength value.

[0071] Then, the grid with the highest utility value is selected as the flight target for the next control cycle: (8) If multiple grid cells have the same utility, the candidate grid cell that is closest to the current grid cell is selected first.

[0072] As shown in equation (7) above, from the first term Driving drones to unchecked locations, index items Ensure the time cost of the path to that point If the value is too high, its attractiveness will be drastically suppressed (the degree of suppression is proportional to the square of the time cost), allowing for better time efficiency in dynamic threat environments; the second term This provides a direct, low-time-cost region bias. By employing the aforementioned nonlinear utility function aimed at minimizing time cost, the UAV maximizes the fused utility value in real time. This enables the drone to automatically minimize the expected travel time cost at each decision moment. Local decisions naturally serve the global time optimization, thereby selecting the feasible path with the local time optimization and achieving time-optimal path planning.

[0073] Step S04. Whenever a drone arrives at a designated task point and completes a designated inspection operation, the corresponding task attraction intensity value is updated to the preset value, and a completion message is broadcast to synchronize the task attraction intensity value of each drone until the task attraction intensity value of all drones in the cluster reaches the preset value, thus completing the inspection task.

[0074] In this embodiment, when the drone Arrive at the mission point After completing the inspection operations (such as image acquisition and infrared temperature measurement), update the task attraction intensity value of that point in the local field copy: Understandably, the task attraction intensity value can also be set to other values ​​to indicate the completion of the inspection, depending on actual needs.

[0075] In this embodiment, when the drone Arrive at the mission point After completing the inspection operations (such as image acquisition and infrared temperature measurement), it also includes generating a lightweight field state update message. This lightweight field status update message It can be a structure that carries any one or more of the following information: message type, message sequence (the message sequence is specifically timestamp information or an incrementing integer count used to characterize the order in which the lightweight field state update messages are sent, for the receiving UAV to perform multi-broadcast deduplication and timing consistency verification), source UAV, target grid index, and updated task attraction intensity value (task field value).

[0076] As a preferred implementation method, the drones in the cluster use an incremental field synchronization method to synchronize information, the steps of which include: drones Lightweight field status update messages are broadcast via wireless ad hoc network. And attach a sequence number to each message when the drone Upon receiving a message, check its sequence number. If it is a processed message, discard it. If it is a task completion message, directly update the task attraction intensity value of the corresponding raster in local storage. If it is a threat update message, use the same evaluator. The perturbation model is recalculated and superimposed with the same anisotropic repulsive field increment. To update the corresponding security rejection strength value, i.e., using the local threat effect assessor. After calculating the disturbance parameters, the anisotropic repulsion field increment is calculated according to equation (4), and then the anisotropic repulsion field increment is superimposed on the current safe repulsion strength value according to equation (5). The updated security rejection strength value is obtained. This is achieved by employing the aforementioned incremental synchronization mechanism and a deterministic recalculation synchronization mechanism (meaning that after each UAV receives a lightweight field state update message, it utilizes the same local threat effect evaluator). The corresponding perturbation parameters and anisotropic repulsion field increments are recalculated autonomously. It can ensure synchronization between drones by broadcasting only lightweight feature parameters, without the need to transmit complete field data or conduct complex negotiations. This can significantly reduce communication overhead and achieve low-communication-overhead cluster collaboration. At the same time, combined with the sequence number and timestamp mechanism, it can also handle practical problems such as packet loss, delay, and conflict, and ensure the eventual consistency of the cluster state.

[0077] Furthermore, for delayed messages, their validity can be determined based on the message timestamp: if the threat corresponding to the message has disappeared or the task point has been updated, the message is ignored. When a conflict update is received (e.g., different task attraction intensities for the same grid),... The value is updated using the latest timestamped message, and a correction message is broadcast to the cluster. Combining this mechanism with the sequence number mechanism, the status of inspection tasks within the cluster can be ensured under various communication conditions. The final consistency of the field (site) further ensures that no detection is missed.

[0078] In this embodiment, the global task is considered complete when the following conditions are met: 1. Task attraction intensity of all task point grids The value is 0 across the entire cluster; 2. Continuous No new threat updates or mission completion messages were generated within the specified time.

[0079] Condition 1 ensures 100% inspection coverage, while condition 2 ensures the system enters a stable and complete state.

[0080] In this embodiment, each drone in the cluster collects data during mission execution. ,in Indicates the obstacle avoidance result, for example The result indicates obstacle avoidance (1 for success, 0 for failure). After the task is completed, the collected data is used to optimize the evaluator parameters using supervised learning. The loss function is: (9) in, λ The regularization coefficient is . N Represents the total number of samples. i Indicates the sample index. Indicates using the current evaluator parameters For the first Feature variables of each sample The four-dimensional perturbation parameter prediction vector output after forward propagation prediction. This represents the preset parameter dimension balance matrix.

[0081] It should be noted that, due to input features and label perturbation parameters Each of these parameters contains multiple sub-parameters with different physical meanings and vastly different dimensions and orders of magnitude (e.g., perturbation intensity). First radius of influence Second radius of influence and spatial attenuation sensitivity factor (Since the dimensions and value ranges of the parameters are different), in order to eliminate the undesirable effects of different dimensions on the gradient optimization direction and convergence speed, this embodiment uses a parameter dimension balancing matrix before substituting the data into formula (9) for loss calculation and parameter optimization. Weighted standardization mapping is applied to the residuals of each dimension, transforming sub-parameters with different dimensions into dimensionless normalized variables within the same order of magnitude. This ensures that each sub-parameter has a balanced optimization weight when directly calculating the Euclidean distance, guaranteeing the optimal parameters of the evaluator network. It can converge quickly and accurately.

[0082] Optimized parameters The data will be updated to all drones via ground stations, improving threat response accuracy for subsequent missions and further optimizing the accuracy of time cost predictions. Obstacle avoidance data will be collected to... Supervised learning is used to optimize the evaluator parameters. This enables the system to evolve itself and continuously improve the accuracy of time cost prediction, thereby enhancing the system's robustness and adapting it to different power inspection scenarios such as mountainous areas and plains.

[0083] In summary, this embodiment uses the above method to characterize the task attraction intensity, which represents the attractiveness of the task. With respect to the strength of security rejection that characterizes the security / time cost The unified encoding is a complex field, which determines the task attraction intensity. As the real part, the safety repulsion strength As the imaginary part, constructing a unified complex information field to represent the attractiveness of inspection tasks and the repulsion of environmental safety allows task attraction and safety repulsion to be processed within the same computational framework, achieving a balance between time optimization and safety constraints. Simultaneously, employing a distributed architecture, with each drone maintaining a local field copy, and through a closed loop of local perception-local update-autonomous decision-making-synchronous update, enables time-optimal path planning and collaborative operation of drone swarms in dynamic environments. Within the complex information field framework, strict inspection point status management can also be achieved, ensuring comprehensive coverage of inspection targets without omissions. This solves the problem of missed inspections caused by the separation of task planning and obstacle avoidance in traditional methods. Furthermore, the complex field can be transmitted in a single serialized manner, reducing communication overhead. Thus, while ensuring 100% inspection coverage, the overall operation time is significantly shortened, significantly improving the efficiency and economy of power line inspection.

[0084] like Figure 2 As shown, in a specific application embodiment, the above-described method of the present invention is used to implement path planning and control of a drone swarm. The detailed steps are as follows: Step 1: Initialize the mission attraction field and safety repulsion field at the ground control station, and construct a global complex information field. And distribute it to each drone in the cluster; Step 2: Each UAV loads the global complex information field and returns to the perception and planning loop, using sensors to perceive environmental changes in real time; Step 3: Each UAV uses the local neural network evaluator to output perturbation parameters, recalculates the anisotropic repulsion field increment, and updates the local safety repulsion field in real time; Step 4: Each UAV determines the set of neighboring grids that can be reached in the next cycle based on the current state and dynamic constraints, and selects the candidate grid with the largest fusion utility value as the flight target; Step 5: The drone maneuvers towards the target and determines whether it has reached the inspection point: if it has not reached it, it continues flying and planning; if it has reached it, it completes the inspection operation and resets the attraction field intensity at that point locally. And generate a status update message; Step 6: The UAV broadcasts a lightweight update message through a wireless ad hoc network, and the receiving UAV performs deterministic recalculation using the same local model to synchronize the local complex information field. Step 7: Determine if the system meets the termination condition (i.e. (All values ​​are 0 and the system is stable): If this condition is not met, return to step 2 and continue the loop. If the condition is met, collect full lifecycle data to perform supervised learning optimization on the evaluator parameters and end the process.

[0085] To verify the effectiveness of the present invention, the present invention was applied to a certain 220 Inspection and testing were conducted on a mountainous section of a V-type high-voltage transmission line. The line is approximately 15 kilometers long and includes 45 towers, requiring detailed inspection of insulators, conductor joints, and other components. The inspection environment presented dynamic interference such as gusts of wind and bird activity. The cluster consisted of three heterogeneous drones. (Long battery life, moderate computing power) (High-performance, with strong computing power) (Standard type, general computing power). The specific implementation process includes:

[0086] During the initialization process, a complex information field is constructed based on the three-dimensional model of the line, and all 45 insulator points (a total of 135) of the towers are mapped to grids. Set the value to 10.0 to ensure the completeness of the task list.

[0087] During the obstacle avoidance decision-making process, at t=125 seconds, A flock of birds was detected near tower #23, and the evaluator outputs disturbance parameters. ,in The value reflects the level of detour time delay that the threat may cause. Calculate the affected raster and update local Post-match broadcast threat update.

[0088] After receiving the update, synchronize locally. field. Grids along the original planned path The value increased from 1.2 to 4.5, and its utility value... The value plummeted from 3.5 to approximately 0.1. This is based on the decision-making principle of maximizing fusion utility. Automatically re-evaluating the neighborhood grid set and autonomously selecting a detour path that avoids high-risk exclusion areas and maximizes the overall fusion utility value is the embodiment of local time optimal obstacle avoidance based on complex information fields.

[0089] exist The inspection of insulator #15 on tower was completed at 310 seconds, and the local system was reset to zero. Value and broadcast the task completion message #89.

[0090] State synchronization and consensus to ensure no missed detections: and After receiving message #89, synchronize the corresponding grid. The value is set to zero. Through this incremental synchronization and conflict handling mechanism, the entire cluster maintains a consistent understanding of the "completed point".

[0091] Task completion and effect verification: When all 135 insulator points The system determines the task is complete when the average value for all drones in the cluster is confirmed to be 0 in the local field. This consensus state is the final procedural proof of "no missed inspections". In other words, this invention can significantly shorten the overall operation time while ensuring 100% inspection coverage, thus significantly improving the efficiency and economy of power line inspection.

[0092] This embodiment further provides a drone swarm system, including multiple drones, each drone being connected to a wireless ad hoc network, and each drone being equipped with a processor and a memory. The memory is used to store computer programs, and the processor is used to execute the computer programs to perform the methods described above.

[0093] It is understood that the method described in this embodiment can be executed by a single device, such as a computer or server, or it can be applied to a distributed scenario where multiple devices cooperate to complete the task. In a distributed scenario, one of the multiple devices may execute only one or more steps of the method described in this embodiment, and the multiple devices interact to complete the method. The processor can be implemented using a general-purpose CPU, microprocessor, application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the method described in this embodiment. The memory can be implemented using read-only memory (ROM), random access memory (RAM), static storage devices, and dynamic storage devices. The memory can store the operating system and other applications. When the method described in this embodiment is implemented through software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0094] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for collaborative path planning of unmanned aerial vehicle (UAV) swarms for power line inspection, characterized by the following steps: include: Step S01. Each UAV in the cluster receives the encoded data of the global complex information field distributed by the ground control station and stores it locally. The global complex information field is formed by discretizing the three-dimensional space of the power line corridor into a set of grids, and storing each grid in the set. Use task attraction intensity value Safety Repulsion Strength Value Define complex information field value The task attraction strength value is constructed and formed. Used for quantizing grid position The security repulsion strength value is used as the priority of inspection targets to characterize the attractiveness of the task. Used to quantify the arrival and safe passage of drones through grid positions. The time cost required is used to characterize the environmental safety repulsion force; Step S02. When each UAV in the cluster detects a threat target in real time during flight, it extracts a feature vector from the detected threat target, calculates a disturbance parameter based on the extracted feature vector to assess the impact range of the threat target, calculates anisotropic repulsion field increment based on the disturbance parameter for the grid within the impact range of the threat target, updates the security repulsion strength value of the corresponding grid based on the anisotropic repulsion field increment, and broadcasts a threat update message formed by the feature vector, disturbance parameter, and the location of the threat target so that the other UAVs can synchronously update the corresponding security repulsion strength value. Step S03. Each UAV in the cluster determines, in real time, the set of neighboring grid cells that can be reached within the next control cycle based on the current state and dynamic constraints of the UAV. Then, based on each candidate grid cell in the neighboring grid set... The safety repulsion strength value and the mission attraction strength value are used to calculate the corresponding fusion utility value, and the candidate grid with the largest fusion utility value is selected as the flight target for the next control cycle. Step S04. Whenever a drone arrives at a designated task point and completes a designated inspection operation, update the corresponding task attraction intensity value to the preset value and broadcast a completion message to synchronize the task attraction intensity value of each drone until the task attraction intensity value of all drones in the cluster reaches the preset value, thus completing the inspection task.

2. The UAV swarm cooperative path planning method for power line inspection according to claim 1, characterized in that, In step S01, the global complex information field is constructed according to the following steps: Step S101. Initialize the task attraction field: The task point set consists of the key equipment nodes to be inspected on the power line. M represents the number of critical points to be inspected. These critical equipment nodes include insulator strings, conductor joints, and key tower components. For the task point set... At each key point in the grid, the task attraction intensity value The initial configuration is set to the preset maximum task attraction intensity value of the inspection target. The guide path is formed by a continuous grid along the direction of the conductor. For the guidance path The grid on By grid The geometrical spatial distance between the nearest task point and the task attraction strength value decreases non-linearly: , For grid To the nearest mission point Euclidean distance, To control the spatial decay coefficient that causes the task attraction intensity to decrease exponentially with increasing spatial distance, the task attraction intensity values ​​of grids on non-inspection task points and non-guided paths are configured to be 0, forming an initial task attraction intensity field. ; Step S102. Initialize the safety repulsion field: Initialize the safety repulsion strength value of the grid set occupied by the known static obstacle body to a preset maximum penalty value or infinity to construct an absolute hard no-fly constraint. Initialize the safety repulsion strength value of the boundary buffer grid within a preset distance outside the static obstacle to decrease with distance. B (g)= ,in, This represents the basic safety repulsion strength value of a static obstacle. This represents the attenuation coefficient of the static obstacle boundary buffer zone. Indicates the current grid position The Euclidean distance to the outer physical boundary of the nearest static obstacle is used to initialize the safety repulsion strength value of the remaining absolutely safe free-space grid outside the buffer zone to 0, forming the initial safety repulsion strength field. ; Step S103. Based on the initial task attraction intensity field and safety repulsion strength field To construct and form a global complex information field .

3. The UAV swarm cooperative path planning method for power line inspection according to claim 1, characterized in that, In step S02, the threat targets include birds and drones, and the elements in the feature vector include the speed of the threat target relative to the drone. Current Euclidean distance between the threat target and the drone The unit vector of the threatening target's direction of movement. And the largest eigenvalue of the spatial distribution covariance matrix of the threat target point cloud. The disturbance parameters include disturbance intensity S and first influence radius. Second radius of influence and the time delay spatial decay sensitivity factor The first radius of influence The second radius of influence corresponds to the boundary of the core area adjacent to the threatening target. This corresponds to the boundary of the outer transition buffer zone.

4. The UAV swarm cooperative path planning method for power line inspection according to claim 3, characterized in that, In step S02, a local threat effect estimator composed of a three-layer fully connected feedforward neural network is used. The local threat effect evaluator calculates the perturbation parameters. It consists of an input layer, two hidden layers, and an output layer. The input layer takes as input the feature vector extracted from the threat target. The output layer outputs the corresponding perturbation parameters. , This represents the evaluator parameters.

5. The UAV swarm cooperative path planning method for power line inspection according to claim 1, characterized in that, In step S02, the directional gain function is defined based on the perturbation parameters. and the radius of influence of anisotropy : ， in It is the backward gain factor; ， Based on the disturbance parameters and directional gain function and the radius of influence of anisotropy The anisotropic repulsive field increment was calculated as follows: ， in, Geometric vectors of relative position Unit vector of the threat's movement direction The angle between them, the geometric vector of their relative positions Defined as from the center of the threat target Pointing to the location of the associated reference grid in space The vector, Represents Euclidean distance. For disturbance parameters, Indicates the intensity of the disturbance. Indicates the first radius of influence, The second radius of influence is indicated by the first radius of influence. The second radius of influence corresponds to the boundary of the core area adjacent to the threatening target. This corresponds to the boundary of the outer transition buffer zone. This represents the spatial decay sensitivity factor of time delay.

6. The UAV swarm cooperative path planning method for power line inspection according to claim 5, characterized in that, In step S02, for those in a threatened position Grids within the affected area Update the corresponding safety repulsion strength value according to the following formula: ， in, Represents the raster at time t+1 Updated safety rejection strength value, Represents the raster at time t The safety rejection strength value; In each global synchronization cycle Security rejection strength for all grids Exponential decay: ， in, Representing grids The security rejection strength value before and after the global synchronization cycle update. This is the attenuation coefficient.

7. The UAV swarm cooperative path planning method for power line inspection according to any one of claims 1 to 6, characterized in that, In step S03, candidate grids are calculated according to the following formula. Fusion utility value: ， in, >0 indicates an adjustable weight parameter. Representing candidate grids The task attraction strength value and the safety repulsion strength value.

8. The UAV swarm cooperative path planning method for power line inspection according to any one of claims 1 to 6, characterized in that, In step S04, after each time the UAV arrives at the mission point and completes the inspection operation, a lightweight field state update message is also generated. And broadcast a task completion message, the lightweight field state update message The message carries the following information: message type, message sequence, source UAV, target grid index, and updated task attraction intensity value; when the UAV receives the lightweight field state update message... At that time, check the message sequence number. If it is a processed message, discard it. If it is a task completion message, directly update the task attraction strength value of the corresponding grid in local storage. If it is a threat update message, recalculate the perturbation parameters using the same local threat effect estimator and superimpose the same anisotropic repulsion field increment. To update the corresponding safety rejection strength value.

9. The UAV swarm cooperative path planning method for power line inspection according to any one of claims 1 to 6, characterized in that, During the mission execution, each drone in the cluster collects data... ,in Indicates the obstacle avoidance result. This represents the feature vector extracted from the threatening target. This represents the perturbation parameters, and after the task is completed, it also includes the use of the collected data to... Optimize evaluator parameters using supervised learning The loss function used is: ， in, λ The regularization coefficient is . N Represents the total number of samples. i Indicates the sample index. Indicates using the current evaluator parameters For the Feature variables of each sample The four-dimensional perturbation parameter prediction vector output after forward propagation prediction. This represents the preset parameter dimension balance matrix.

10. A drone swarm system comprising multiple drones, each drone being communicatively connected to a wireless ad hoc network, each drone being equipped with a processor and a memory, the memory being used to store computer programs, characterized in that... The processor is used to execute the computer program to perform the method as described in any one of claims 1 to 9.

Images