Multi-UAV path planning method for transmission line inspection

Abstract

The present application discloses a multi-UAV path planning method for power transmission line inspection, which belongs to the field of path planning. The method combines mobile edge computing (MEC) with UAV-based intelligent decision-making: first, the MEC server synchronizes the cluster status and issues a global demand matrix; the UAV dynamically updates the global risk matrix through real-time environmental perception; each UAV-based airborne unit integrates its own real-time status, risk and demand matrix, and autonomously decides to generate the optimal short-term planning path. Before the path is officially executed, it will be reported to the MEC server for collaborative conflict detection. Only after confirming that there is no conflict can the UAV execute the flight. The present invention gives the UAV cluster the ability to autonomously perceive and quickly avoid dynamic unknown risks, making path planning a real-time evolving intelligent decision-making process, and ensuring the safety of cluster collaborative operations through centralized conflict detection, thereby significantly improving the efficiency, robustness and reliability of power transmission line inspection tasks.

Description

Technical Field

[0001] The present application relates to the technical field of drone path planning, and more specifically, to a multi-drone path planning method for power transmission line inspection. Background Art

[0002] With the continued advancement of my country's smart grid construction and the growing scale of its transmission line network, the demand for efficient and accurate transmission line inspections has rapidly increased. Traditional transmission line inspection methods primarily rely on manual climbing, ground-based telescope observations, or fixed online monitoring equipment. These methods not only face inherent bottlenecks in inspection efficiency, cost control, and coverage, but also pose severe safety challenges to inspectors in complex terrain (such as mountains, rivers, and lakes) or extreme weather conditions, potentially hindering their mission. Unmanned aerial vehicles (UAVs), or drones, have become an indispensable tool for transmission line inspection due to their superior maneuverability and efficient data collection capabilities. Equipped with specialized payloads such as high-definition cameras, infrared thermal imagers, and lidar, drones can efficiently perform key tasks such as defect identification, abnormal temperature rise monitoring, and insulator contamination analysis on transmission lines, improving inspection efficiency while significantly ensuring operational safety. However, the application of drones in large-scale transmission line inspections still faces a number of technical challenges, the most critical of which is path planning in complex environments. Especially in large-scale inspections, it is usually necessary to deploy multiple drones to work together, which places extremely high demands on the real-time, safety and collaborative efficiency of path planning.

[0003] Currently, most of the existing research on multi-UAV path planning focuses on the global optimal solution or local obstacle avoidance strategy in static environments. For example, Traditional path planning algorithms, such as the CNN algorithm or the rapidly expanding random tree (RRT), can find feasible paths given a static map. However, their core reliance on prior knowledge of the environment makes them incapable of responding to dynamic and sudden obstacles commonly encountered during power line inspections, such as temporary construction cranes, floating objects, or other aircraft. This makes real-time path replanning difficult, posing significant safety risks. Furthermore, existing multi-UAV systems often utilize centralized control architectures or pre-set fixed flight paths to perform tasks. While this approach is easy to manage, it leads to rigid task allocation and lagging flight strategies. If the environment ahead changes or a UAV malfunctions, the collaborative efficiency of the entire fleet is significantly reduced. This is especially true in close-quarters collaborative operations that require real-time interaction to avoid collisions. Neither the response speed nor the decision-making flexibility required to meet millisecond-level safety requirements.

[0004] Therefore, existing technologies generally have defects such as insufficient adaptability to dynamic environments, slow response of multi-UAV collaborative collision avoidance mechanisms, and lack of ability to quickly model and adjust countermeasures to real-time environmental changes. This directly limits the application efficiency and safety and reliability of UAV clusters in complex and changeable transmission line inspection scenarios. Summary of the Invention

[0005] In order to overcome the limitations of the existing technology, according to one aspect of the present application, a multi-UAV path planning method for power transmission line inspection is provided, which includes:

[0006] The first drone uploads its current drone state to the MEC server, and the MEC server synchronizes the current drone state of the first drone to other drones in the cluster;

[0007] Real-time environmental perception is performed through the sensors of each drone to obtain real-time environmental perception results;

[0008] In response to the real-time environment perception result being that an obstacle is detected, triggering a dynamic risk assessment to update a global risk matrix; and in response to the real-time environment perception result being that there are no accidents, loading and using the global risk matrix of the previous cycle;

[0009] The MEC server calculates a global demand matrix based on the distance between each UAV and the target point, and synchronizes the global demand matrix to the first UAV;

[0010] The onboard computing unit of the first UAV determines an optimal action based on the current UAV state, the global risk matrix, and the global demand matrix, and generates a short-term planning path based on the optimal action;

[0011] The first UAV broadcasts its short-term planned path to the MEC server, wherein the MEC server performs collaborative conflict detection based on the short-term planned paths of each UAV;

[0012] In response to the MEC server's collaborative conflict detection result indicating that there is no conflict, the first UAV executes the optimal action to fly along the short-term planned path.

[0013] Compared with the existing technology, the main purpose of the multi-UAV path planning method for transmission line inspection provided by this application is to overcome the defects of current UAV path planning technology that strongly relies on static environment models and has insufficient adaptability when facing sudden obstacles, and to solve the technical bottlenecks of traditional multi-UAV systems in collaborative operations, such as delayed collision avoidance response and low overall operating efficiency due to over-centralized control or rigid strategies.

[0014] Specifically, the core concept of this system is to provide a distributed collaborative path planning framework that deeply integrates edge computing with drone-based intelligent decision-making. Specifically, it uses a mobile edge computing (MEC) server as the core hub for information exchange. Each drone first uses its own sensors to perceive its surrounding environment in real time, dynamically evaluating and updating a globally shared risk matrix. Simultaneously, the MEC server calculates and distributes a global demand matrix that quantifies the mission's value based on the inspection mission objectives and the real-time location of the drone cluster. Subsequently, each drone's onboard computing unit independently and in real time integrates its current state, the global risk matrix, and the global demand matrix to autonomously determine the optimal short-term flight maneuver and generate a planned path. Crucially, before formal execution, this planned path is submitted to the MEC server for unified collaborative conflict detection and is only executed after ensuring there is no collision risk.

[0015] By adopting the above-mentioned technical concept, the present invention can achieve significant technical effects: it gives the drone cluster the ability to autonomously perceive and quickly avoid unknown or dynamic risks in the inspection environment, so that path planning is no longer a rigid preset route, but an intelligent decision-making process that evolves in real time. At the same time, the MEC-based collaborative conflict detection mechanism ensures the safety of multiple drones operating in close proximity in complex airspace, greatly improving the collaborative efficiency and robustness of cluster operations. Ultimately, this method can fundamentally improve the safety and reliability of transmission line inspection tasks in changing and complex environments, and achieve a higher level of automated and intelligent inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0017] Figure 1 This is a flowchart of a multi-UAV path planning method for power transmission line inspection according to an embodiment of the present application.

[0018] Figure 2 This is another flowchart of the multi-UAV path planning method for transmission line inspection according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. While the drawings illustrate certain embodiments of the present disclosure, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0020] Therefore, in response to the technical defects revealed in the background technology, this application proposes a multi-UAV path planning method for transmission line inspection. Figure 1 This is a flowchart of a multi-UAV path planning method for power transmission line inspection according to an embodiment of the present application. Figure 2 This is another flowchart of the multi-UAV path planning method for transmission line inspection according to an embodiment of the present application.

[0021] like Figure 1 and Figure 2 As shown, according to the embodiment of the present application, the multi-UAV path planning method for power transmission line inspection includes: S110, the first UAV uploads its current UAV state to the MEC server, and the MEC server synchronizes the current UAV state of the first UAV to other UAVs in the cluster; S120, real-time environmental perception is performed through the sensors of each UAV to obtain a real-time environmental perception result; S130, in response to the real-time environmental perception result that an obstacle is detected, a dynamic risk assessment is triggered to update the global risk matrix; and in response to the real-time environmental perception result that there is no accident, the global risk matrix of the previous cycle is loaded and used; S140, the MEC server calculates the global demand matrix based on the distance between each UAV and the target point, and the global demand matrix is ​​updated. synchronized to the first UAV; S150, the onboard computing unit of the first UAV determines the optimal action based on the current UAV state, the global risk matrix and the global demand matrix and generates a short-term planning path based on the optimal action; S160, the first UAV broadcasts its short-term planning path to the MEC server, wherein the MEC server performs collaborative conflict detection based on the short-term planning paths of each UAV; S170, in response to the collaborative conflict detection result of the MEC server being that there is no conflict, the first UAV executes the optimal action to fly along the short-term planning path; and, S180, after the first UAV executes the optimal action, the BMS reports the updated remaining energy value and updates it to the current UAV state.

[0022] In step S110, the first drone uploads its current drone state to the MEC server, and the MEC server synchronizes the first drone's current drone state with the other drones in the cluster. In a multi-drone path planning task for power line inspection, the initial stage of system operation includes uploading the drones and their current drone states, which represent their conditions, to the mobile edge computing (MEC) server. Subsequently, after receiving the current drone state of each drone, the MEC server synchronizes it with the other drones in the cluster. Here, the process is illustrated using the first drone as an example. The first drone uploads its current drone state, which represents its own condition, to the mobile edge computing (MEC) server. Subsequently, after receiving this state, the MEC server synchronously distributes the first drone's current drone state to the other drones in the cluster.

[0023] Through this state upload and synchronization mechanism, each drone in the swarm can maintain real-time knowledge of key information such as the location and energy levels of its neighboring drones. This global state sharing is a prerequisite for efficient and safe swarm operations. It fundamentally avoids issues such as flight path conflicts and uncoordinated task allocation that can arise from information occlusion, ensuring that each drone has an accurate understanding of the current swarm layout when making subsequent decisions.

[0024] In this embodiment, the current drone status is a structured data set that specifically includes the drone's location, energy status, risk radius, and identity. The location status is typically the three-dimensional coordinates (x, y, z) acquired by the onboard Global Positioning System (GPS) or Real-Time Differential Kinematics (RTK) module; the energy status is the remaining energy b provided by the battery management system (BMS); the risk radius r is a parameter that dynamically adjusts based on the risk level of the surrounding environment; and the identity is a unique identifier assigned to each drone. These data components together constitute a complete status package that comprehensively describes the drone's current condition.

[0025] That is, in the embodiment of the present application, the state of each drone at time slot t is represented by a six-tuple, where the three-dimensional coordinates (x, y, z) accurately describe the real-time position of the drone in the transmission line inspection space. The risk radius r is a dynamically changing parameter that automatically adjusts according to the risk level of the surrounding environment. The remaining energy b is the core state parameter that directly determines the drone's ability to continue operating. Assume that the system contains N drones, and the state of each drone at time slot t is represented as: .in represents three-dimensional coordinates, represents the risk radius, Indicates the remaining energy.

[0026] The specific execution process of this state interaction includes two actions: upload and synchronization. The upload process is initiated by the drone. At the beginning of each decision cycle, its onboard computing unit collects data such as location, energy, risk radius, and identity from various subsystems, integrates it into a standard data packet, and sends it to the MEC server via wireless communication modules such as 5G / 6G. The synchronization process is led by the MEC server. After receiving the status data packet, it updates the internally maintained cluster state table and immediately distributes the information containing the latest status to all other drones in the cluster via broadcast or multicast. Leveraging the edge deployment characteristics of MEC, this ensures low latency and near-real-time transmission of status information.

[0027] To facilitate understanding, here's a scenario: Suppose Drone A and Drone B are performing a joint mission. At time t0, after Drone A completes initialization, it collects its current drone status: {ID: "Drone A", Position Status: (120.1, 30.2, 150.0), Energy Status: 99%, Risk Radius: 5 meters}. Drone A then uploads this status packet to the MEC server, which immediately synchronizes this information with Drone B. Therefore, Drone B, during its own path planning calculations, already knows Drone A's precise position and can treat it as a dynamic avoidance target in subsequent decisions, ensuring the effectiveness of collaborative collision avoidance from the very beginning.

[0028] In step S120, each drone's sensors perform real-time environmental perception to obtain real-time environmental perception results. This step is a prerequisite for dynamic obstacle avoidance and path adjustment, and it utilizes each drone's sensors to perform real-time environmental perception. The drone's onboard sensors act as the system's perceptual endpoints, ensuring that the environmental information underlying decision-making accurately and instantly reflects the actual physical world.

[0029] It's understandable that in complex inspection scenarios like power transmission lines, numerous unpredictable, time-varying dynamic factors exist, such as the sudden appearance of construction machinery, flocks of birds, or other flying objects. To cope with these emergencies, drones must be able to switch from a static, pre-installed map-based planning mode to an intelligent operating mode with dynamic, real-time response capabilities. Through high-frequency, real-time environmental perception, drones can promptly capture all environmental changes beyond the preset 3D model, providing the most direct and accurate raw data input for subsequent dynamic risk assessment and route replanning. In the embodiments of this application, the real-time environmental perception result is a structured judgment conclusion obtained after processing and comparison. It can be mainly divided into two types: one is obstacle detection, and the other is no accidents. To obtain this result, the drone utilizes onboard sensors such as cameras and LiDAR. The camera is used to capture texture and color information of the environment, while the LiDAR is used to accurately measure the three-dimensional distance to surrounding objects, generating real-time point cloud data. The complementary advantages of these two provide powerful hardware support for environmental perception.

[0030] This perception process primarily involves two steps: environmental scanning and change detection. During environmental scanning, the drone's onboard camera and lidar continuously scan its surroundings, particularly the space ahead, generating real-time image streams and point cloud data. During change detection, the onboard computing unit aligns and compares the acquired real-time data with a preloaded, static, high-precision 3D digital map containing all known fixed objects (such as towers and wires), thereby identifying unmatched "incremental objects." If such incremental objects are present, the perception result indicates an obstacle has been detected; otherwise, it is considered normal.

[0031] This process can be understood through the following scenario: At time t1, a drone's perception result is clear because the point cloud data it scanned perfectly matches the pre-stored tower model. However, at time t2, as it approaches the next tower, the new frame of point cloud data it acquires reveals, in addition to the tower's outline, a cluster of points never before recorded in the static map, such as a crane in operation. The comparison algorithm immediately identifies this significant discrepancy and determines the real-time environmental perception result at that moment as an obstacle, along with its location and size. This result immediately triggers the subsequent risk assessment and path adjustment process.

[0032] In step S130, in response to the real-time environmental perception result being that an obstacle is detected, a dynamic risk assessment is triggered to update the global risk matrix; and in response to the real-time environmental perception result being that there are no accidents, the global risk matrix of the previous cycle is loaded and used. It should be understood that in the embodiment of the present application, conditional judgment and response are made according to the real-time situation, and the global risk matrix is ​​selectively updated based on the real-time environmental perception result obtained in the previous link. This step connects the "perception" and "decision-making" links, and its processing logic has clear branches: when the real-time environmental perception result is that an obstacle is detected, the system will trigger a dynamic risk assessment to update the global risk matrix; and when the result is that there are no accidents, the global risk matrix of the previous cycle can be directly loaded and used.

[0033] The logic behind this conditional branch is designed to efficiently balance the system's computational overhead with the real-time nature of decision-making. Path planning is a high-frequency, repetitive decision-making process. If the risk of the entire space is indiscriminately recalculated in every cycle, it would waste significant computing resources and hinder decision-making response speed. Therefore, the computationally intensive risk assessment process is only initiated when there is a substantial change in the environment (i.e., an obstacle is detected). This on-demand computation mechanism significantly improves the system's overall operational efficiency while ensuring the highest level of safety response.

[0034] The global risk matrix is ​​a three-dimensional array formed by discretizing the inspection airspace into a grid. Each element of this array corresponds to a spatial voxel, and its value represents the risk level of that voxel. Dynamic risk assessment, on the other hand, is an event-driven algorithmic process. Upon sensing a new obstacle, the system initiates a calculation to quantify the increased risk posed by the obstacle and integrate it into the existing global risk matrix.

[0035] The specific logic of this step depends on the results of real-time environmental perception. If the result is no accidents, the system determines that the existing risk matrix is ​​still valid. At this time, the drone's onboard computing unit will directly reference the global risk matrix stored in memory and used in the previous decision cycle when performing subsequent path planning, without any update calculations. Conversely, if the result is an obstacle detection, the system immediately initiates a dynamic risk assessment process to update the global risk matrix, ensuring that subsequent decisions can take into account the newly discovered hazards.

[0036] When dynamic risk assessment is required, the process is broken down into several steps. First, the location of the obstacle is obtained based on the perception results. Then, based on this location information, a Gaussian distribution risk model is applied to the obstacle to calculate a new independent risk field that only includes the impact range of this new obstacle. This model can scientifically quantify the risk intensity of the obstacle and its attenuation in space. Specifically, this application uses a Gaussian distribution to quantify obstacle risk. For K obstacles, the risk exposure of drone i to obstacle k can be expressed as:

[0037]

[0038] in represents the Euclidean distance, The value represents the maximum intensity of the risk, Control the impact range of risk, distance squared term Ensure that the near-field risk rises rapidly.

[0039] Then, the system will deeply fuse and superimpose this new risk field with the original global risk matrix. In order to avoid distortion of the results caused by simple numerical addition, this method uses the probability product operation method for fusion, and finally generates a brand new global risk matrix that includes the latest hazard sources. Specifically, its logic is to first convert the risk value into a safety probability (1-risk value), and then multiply the safety probability of the original global risk field at a certain point with the safety probability of the new risk field at that point to obtain a joint safety probability, and finally subtract the joint safety probability from 1 to obtain the updated fusion risk value of the point. This process can be expressed as: the global risk matrix realizes the superposition calculation of multi-obstacle risks through probability product operations, which is more in line with actual physical scenarios. The global risk matrix is ​​expressed as:

[0040]

[0041] Through this method, the system accurately maps the new risk field onto the original global risk map, generating a new global risk matrix that includes the latest hazard sources for use in the current decision-making cycle.

[0042] For example, a drone is flying, and its initial global risk matrix only includes fixed transmission tower and wire risks. When its sensors detect an unregistered construction crane on the route ahead, a dynamic risk assessment is triggered. The MEC server obtains the crane's three-dimensional coordinates and generates an independent risk field for it with a Gaussian distribution. The MEC server then performs a probability product operation, integrating this "crane risk field" with the system's existing global risk matrix. This creates a new global risk matrix that includes both fixed risks and newly added dynamic risks. This matrix is ​​then synchronized with the drone, ensuring it can proactively avoid the crane.

[0043] In step S140, the MEC server calculates a global demand matrix based on the distances between each drone and the target point, and synchronizes the global demand matrix with the first drone. It should be understood that the core of the method of the present invention, in order to guide the drone cluster to efficiently perform tasks, lies in the calculation and synchronization of the global demand matrix by the MEC server.

[0044] Specifically, the MEC server calculates a global demand matrix based on the real-time distances between each drone in the cluster and all inspection targets, and synchronizes this matrix to every drone in the cluster. This step effectively creates a dynamic "gravity map" for the drones in three-dimensional space, clearly oriented to the mission's value. This is crucial for transforming abstract inspection tasks into quantifiable reward signals that intelligent agents can optimize. Without a clear objective to guide them, even if a drone can avoid all risks, its flight trajectory will inevitably be blind and unable to efficiently cover all critical areas requiring inspection. By constructing a global demand matrix that reflects the real-time attractiveness of each inspection target, the drone's reinforcement learning decision-making model can be provided with clear, positive incentives, driving it to proactively and strategically fly toward targets with more urgent needs and higher value, thereby ensuring the efficiency of the inspection mission.

[0045] In this embodiment of the present application, the global demand matrix is ​​a three-dimensional discretized data structure corresponding to the inspection airspace. In this three-dimensional array, the vast majority of voxels are zero. Only those voxels corresponding to the locations of target points to be inspected (such as insulator strings, shock absorbers, etc.) are assigned a non-zero positive value, which is the demand value. The demand value is a dynamically changing numerical score that quantifies the "benefit" or "value" that a specific drone can obtain by flying to this target point for inspection at the current moment. Its value is generally related to the inherent priority of the target point and the drone's distance from it.

[0046] The generation and distribution of this demand matrix can be divided into three steps: calculation, integration, and synchronization. During the calculation phase, the MEC server iterates over all inspection target points and, for each target point, uses a modified Sigmoid demand function to calculate its current demand value based on its inherent task priority and its distance from the current drone swarm. This function effectively simulates the characteristic that the demand attractiveness of a target point increases significantly only when drones are positioned close enough to ensure inspection service quality.

[0047] More specifically, in this step, the MEC server first needs to traverse all the target points that need to be inspected in the system. Then, for each target point j in the list, the MEC server will independently calculate its demand value at the current time t based on the distance between each drone in the current cluster and the target point. This application uses an innovative improved sigmoid demand function to perform this calculation. The principle and process are as follows: The improved sigmoid demand function has the following advantages over the traditional linear model: 1) Parameters Different settings can be made based on the priority of different transmission equipment; 2) The exponential term can better simulate the nonlinear relationship between signal strength and demand; 3) The adjustable parameters η and β make the model flexible and adaptable to various inspection scenarios. The demand function of the end user is designed as:

[0048]

[0049] in Indicates the basic demand level, Indicates signal strength, It is an adjustable parameter.

[0050] Specifically, Represents the signal strength. In practice, it is a function that is strongly related to the distance from the drone to the target point j. The closer the distance, the stronger the signal strength. The higher it is, the better the inspection service quality is. and are two adjustable parameters, Plays the role of service quality threshold, This controls the steepness of the demand value curve as the service quality changes. Through this function, only when the position of the drone is good enough (i.e. ), the demand value of the target point will increase significantly, thus accurately simulating the demand characteristics of effective inspection.

[0051] After calculating the current demand values ​​for all target points, the MEC server performs an integration operation. It fills the voxel position corresponding to the target point's 3D coordinates with the calculated demand value for each target point in a 3D array initially set to all zeros. After this operation is completed, the 3D array constitutes the global demand matrix valid for the current decision cycle. The global demand matrix can be expressed as:

[0052]

[0053] As shown in the formula, the demand matrix adopts a threshold judgment mechanism: when the signal strength Reach service radius When a task is completed, it is included in the service demand; otherwise, the demand is ignored. This design effectively avoids wasting energy on ineffective services. Finally, in the synchronization phase, the MEC server immediately sends this matrix data containing the global task score distribution to all drones in the cluster via the wireless network.

[0054] For example, suppose an inspection mission includes two target points: Insulator A (high priority) and Hammer B (low priority). At time t, the MEC server, based on the drone's real-time location, determines that it is close to Insulator A. Based on the demand function, it calculates a high demand value of 95.8 for Point A. However, Hammer B is still farther away, resulting in a demand value close to 0. The MEC server then generates a global demand matrix in which only the spatial location corresponding to Insulator A is assigned a high value of 95.8. When this "heat map" highlighting Point A is synchronized to the drone, its decision-making algorithm clearly perceives the potential for significant reward by flying toward Point A, and thus tends to generate a flight path toward that point.

[0055] In step S150, the onboard computing unit of the first drone determines the optimal action based on the current drone state, the global risk matrix, and the global demand matrix, and generates a short-term planned path based on the optimal action. Specifically, the onboard computing unit uses the collected current drone state, the global risk matrix, and the global demand matrix as inputs to comprehensively determine an optimal action and then generates a short-term planned path based on this optimal action. This step constitutes the "decision-making brain" of the individual drone, capable of making real-time trade-offs between risk avoidance, mission pursuit, and consideration of its own state, outputting an optimal action instruction.

[0056] This onboard autonomous decision-making capability integrates all previously collected multi-dimensional situational information, completing the crucial transition from environmental perception to action decision-making. In the previous steps, the drone acquired three types of information: its own status, environmental risks, and mission value. The goal of this step is to use its onboard intelligent decision-making model to comprehensively weigh these three types of information and calculate an optimal short-term action strategy that balances safety, efficiency, and energy consumption. This gives each drone true autonomy and is the foundation for highly dynamic, intelligent, and collaborative flight planning.

[0057] It's worth noting that in this application's embodiments, the optimal action typically refers to the action, selected from the set of discrete actions the drone can perform in its current state (for example, flying a fixed distance in eight directions or hovering in place), that the value assessment model determines will yield the greatest long-term cumulative reward. The short-term planned path is directly defined by this optimal action, i.e., the trajectory from the drone's current location to the next location it will reach after executing that action.

[0058] This process is implemented on the drone's terminal through a computational flow that includes input packaging, model inference, and action selection. First, the drone's onboard computing unit packages its current state data, the received global risk matrix, and the global demand matrix into a unified state tensor format that meets the input requirements of the subsequent neural network model. This step ensures that all decision-making factors are taken into account. The onboard computing unit then inputs this packaged state tensor into a locally deployed, fully trained, lightweight Deep Q-Network (DQN) model. This DQN model, acting as a value evaluation function, rapidly infers and calculates the long-term expected reward (Q-value) of each possible action under the current state. The output is a Q-value vector, such as [Q(S, a1), Q(S, a2), ..., Q(S, an)], where each entry corresponds to the value evaluation of a potential action.

[0059] After obtaining a vector containing the Q values ​​of all possible actions, the onboard computing unit ultimately determines the optimal action using a greedy strategy. It directly selects the action with the highest Q value as the optimal decision for the current state. For example, if the inference results show that the action "fly forward" has the highest Q value, it is selected as the optimal action. The system then generates a short-term planned path from the current point to the target point based on this action. It is important to note that at this stage, the path is only in the "planning" state and has not yet been executed.

[0060] More specifically, the underlying principles of the process are explained at a more substantive level. This step models the drone path planning problem as a Markov decision process (MDP), enabling intelligent decision-making in dynamic environments. The core of this approach lies in the design of a comprehensive state representation model that incorporates information such as the drone's location coordinates, risk radius, and remaining energy. It also employs a modified sigmoid demand function to accurately quantify the service requirements of end users (TUs). Using a dual-network architecture based on a Deep Q-Network (DQN), combined with experience replay and priority sampling, the algorithm efficiently learns optimal path strategies while simultaneously enabling information sharing and collision avoidance through multi-UAV collaboration.

[0061] The state space completely contains the drone state, environmental requirements and risk information. It can be expressed as:

[0062]

[0063] The action space is designed as discrete end-user service selection to facilitate algorithm implementation. It can be expressed as:

[0064]

[0065] The reward function innovatively integrates three key factors, including service demand satisfaction, risk aversion, and energy consumption cost. It can be expressed as:

[0066]

[0067] This application uses a dual network architecture to effectively solve the overestimation problem in Q learning, where the main network is responsible for action selection, the target network provides a stable learning target, and the soft update coefficient Control parameter update speed. The Q function update rule is expressed as:

[0068]

[0069] The target network parameters θ^- are updated periodically:

[0070]

[0071] Strategy optimization considers two aspects: maximizing long-term rewards and ensuring that total energy consumption does not exceed battery capacity. This can be transformed into an unconstrained optimization problem using the Lagrange multiplier method. Therefore, it can be expressed as:

[0072]

[0073]

[0074] The implementation of its path planning algorithm is as follows:

[0075] First is the cost matrix calculation:

[0076] Calculate the cost matrix using a piecewise linear model :

[0077]

[0078] in,

[0079] Each training episode contains a complete state-action-reward sequence. The experience replay mechanism improves learning efficiency by breaking temporal correlation, improving data utilization, and stabilizing the training process. The specific process is as follows:

[0080] First, initialize the cost matrix and memory matrix ;

[0081] Then, for each episode:

[0082] a. Observe the current state ;

[0083] b. Select an action ;

[0084] c. Perform actions and receive rewards and the new state ;

[0085] d. Storage To Memory ;

[0086] e. Sample minibatch from D to update Q network;

[0087] Finally, the target network is updated every C steps.

[0088] Imagine this scenario: The first drone's onboard computing unit, fed with its collected state, a risk matrix indicating danger to the right ahead, and a demand matrix indicating a high-value inspection point directly ahead, infers a Q-value vector using its local deep Q-network. This vector likely indicates a negative Q-value for "fly right" and a high positive Q-value for "fly forward." The drone then selects "fly forward one meter" as the optimal action and generates a short-term planned path from its current position to the one-meter point ahead, ready for the next stage of collaborative conflict detection.

[0089] In step S160, the first UAV broadcasts its short-term planned path to the MEC server, wherein the MEC server performs collaborative conflict detection based on the short-term planned paths of each UAV.

[0090] That is, after a drone determines its optimal short-term planned path, the method of the present invention also incorporates a critical safety verification step to ensure the safety of swarm flight. This involves first broadcasting its generated short-term planned path to the MEC server, which then performs unified collaborative conflict detection based on all short-term planned paths collected from each drone in the swarm. This centralized safety arbitration mechanism, executed by a central node, is the final line of defense for ensuring the safety of multiple drones operating in close coordination in complex airspace.

[0091] This step is necessary to address potential path conflicts that can arise from a fully distributed decision-making architecture. Although each drone considers the current positions of other drones when making decisions, because the decision-making process is nearly simultaneous and independent within a very short period of time, it is entirely possible that drone A decides to turn right while nearby drone B decides to turn left, resulting in a fatal intersection on the future path. Therefore, before each drone officially executes its flight maneuver, the MEC server must centrally review the flight intentions of all drones to predict and prevent these potential collisions.

[0092] It should be understood that in the embodiments of this application, collaborative conflict detection is a predictive geometric calculation process led by the MEC server. Its essence is to simulate whether the physical volumes (including a safety radius) of all drones will overlap in space in the near future, assuming that all drones fly according to their submitted planned paths. This conflict detection is not a simple geometric path intersection judgment, but a four-dimensional spatiotemporal interference check of the drones' flight envelopes, including the time dimension.

[0093] This process, centered around the MEC server, includes two steps: path collection and conflict prediction. First, path collection involves each drone in the cluster generating its own short-term planned path and immediately submitting it to the MEC server. The MEC server aggregates the path plans submitted by all drones within the same decision cycle. Next, conflict prediction involves the MEC server traversing all combinations of short-term planned paths, for example, pairwise, and calculating the shortest distance between each pair of drones as they fly along their respective paths.

[0094] If the calculated shortest distance is less than a preset safety threshold (this threshold is greater than the drone's physical size plus a certain safety margin), the MEC server determines that there is a coordination conflict between the two paths. Conversely, if after checking all path combinations, no shortest distance is found to be less than the safety threshold, the entire cluster's action plan is considered to be conflict-free. The MEC server will then feed back its global coordination conflict detection results to the drone cluster.

[0095] For example, consider two drones, A and B, flying towards each other in a narrow corridor. At the same moment, A plans a path 2 meters forward, and B also plans a path 2 meters forward. They both report these paths to the MEC server. The server aggregates this information and performs simulations, calculating that the shortest distance between their paths is 1 meter, far less than the preset safety threshold of 5 meters. Therefore, the MEC server immediately determines that there is an imminent risk of collision between A and B. Accordingly, the coordinated conflict detection result is "conflict exists," and the MEC sends evasive maneuvers to both drones. For the other drone, C, located far away from the conflict zone, a "no conflict" result is returned.

[0096] In the step S170, in response to the MEC server's collaborative conflict detection result that there is no conflict, the first UAV executes the optimal action to fly along the short-term planned path. As the final closed loop of the entire decision-making cycle, the method of the present invention clearly points out that the UAV will only actually perform the actual flight action after receiving and confirming the conflict-free feedback from the MEC server. This link can be specifically described as: in response to the MEC server's collaborative conflict detection result that there is no conflict, the first UAV will execute its optimal action previously determined by calculation to fly along the short-term planned path it generated. This step sets a necessary safety prerequisite for all physical actions.

[0097] This "review before flight" mechanism transforms decisions reached through observation, reflection, and questioning into physical action. Without this step, all prior perception and calculation would be meaningless. More importantly, by strictly binding actual flight execution to the "no conflict" test result, this method establishes a strong safety barrier, eliminating any unverified actions that could cause collisions, thus ensuring the safety of swarm operations at the physical execution level.

[0098] The "collaborative conflict detection result is no conflict" here is reflected on the drone side as a logically true "safe passage" or "execution permission" signal issued by the MEC server. And "executing the optimal action" is the process by which the onboard computing unit formally transmits its selected digital instructions, such as "fly forward 1 meter", to the drone's underlying flight control system. Finally, "flying along the short-term planned path" refers to the actual performance of the drone's physical body completing the displacement precisely according to the line segment trajectory under the drive of the flight controller.

[0099] The specific implementation process is a "wait for confirmation before execution" process. After the drone uploads the planned path, it briefly waits for the coordination conflict detection results returned from the MEC server. Once the communication module receives the "execution permission" command, indicating that there is no conflict, its onboard computing unit immediately sends the temporarily stored optimal action (for example, "fly forward 1 meter") to the flight controller. After receiving this high-level command, the flight controller calculates a series of precise motor control commands to drive the drone to complete the flight segment.

[0100] Continuing with the previous example, when the MEC server broadcasts the "permit to execute" command, only drone C will receive a "no conflict" result for itself, while drones A and B will receive a "conflict exists" result and an adjustment command. Consequently, only drone C will send its planned maneuver to the flight control system for execution. Drones A and B, based on the adjustment command, will return to the decision-making step and regenerate a safe short-term planned path that avoids the other drone. The reporting, verification, and execution process will then repeat until a safe maneuver combination is found for all drones.

[0101] In another embodiment of the present application, multiple inspection drones share geometric information through the MEC server to ensure spatial consistency; at the same time, they share demand information to achieve task collaboration; and they also share risk information to ensure flight safety. At the same time, a collaborative collision avoidance strategy is adopted: the collision risk is calculated using the maximum value to ensure that individual risks are not underestimated and the system safety is maximized. When other drones are detected, update:

[0102]

[0103] in is the collision risk threshold.

[0104] In step S180, after the first drone completes the optimal action, the BMS reports the updated remaining energy value and updates it to the current drone state. After a complete decision and action are executed, the system also performs a critical feedback update. Specifically, after the first drone completes the flight corresponding to its optimal action, the onboard battery management system (BMS) accurately calculates the energy consumption caused by this action and generates an updated remaining energy value.

[0105] This step is crucial to maintaining the operational safety of the entire system. Energy is a drone's most constrained resource, and any consumption directly impacts subsequent mission planning and flight endurance. If the system's true energy consumption is not promptly reflected in the system's status, subsequent decision-making models will be based on outdated, inflated battery values. This could potentially lead the drone to plan paths that exceed its return-to-home capabilities, introducing serious safety risks. Therefore, promptly updating energy status information after each action is essential to ensuring consistently effective decision-making and flight safety.

[0106] The battery management system (BMS) is a sophisticated electronic module embedded in the drone's battery pack. It monitors battery voltage, current, and other data in real time, accurately estimating the remaining energy. The updated remaining energy value output by this system is the most authoritative source of energy status. The drone's onboard computing unit then writes this latest value, obtained from the BMS, into the energy status field of the current drone status data structure maintained in memory, overwriting it.

[0107] In a specific embodiment, the battery management system uses the following energy consumption model to perform energy measurement, and its principle is as follows:

[0108] Energy Model The three main energy consumption aspects of drone operations are fully considered:

[0109]

[0110] The flight energy consumption is proportional to the square of the distance traveled and inversely proportional to the speed. It can be expressed as:

[0111]

[0112] The hovering energy consumption depends on the communication quality and service requirements, which can be expressed as:

[0113]

[0114] The computational energy consumption is quadratically related to the complexity of the processing task, which can be expressed as:

[0115]

[0116] The mechanism of this status update can be illustrated by a running example. When a flight energy-consuming action is completed, the onboard computing unit of the first UAV will query the BMS for the latest state of charge through the internal communication bus. For example, the energy of the UAV before the action was 80%, and the action consumed power. After measurement, the BMS obtains a new state of charge of 79.8%. The onboard computing unit reads this 79.8% value and uses it to update the energy item in the "current UAV state" maintained by itself. Therefore, at the beginning of the next decision cycle, the UAV uploads status information containing the latest energy value of 79.8% to the MEC server, ensuring that all subsequent decisions of the entire collaborative system are based on the most accurate endurance parameters.

[0117] The foregoing description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification of the technical solution of the present invention by a person skilled in the art, by means of equivalent substitution or equivalent transformation, without departing from the overall technical content of the technical solution of the present invention, shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A multi-UAV path planning method for power transmission line inspection, comprising: The first drone uploads its current drone state to the MEC server, and the MEC server synchronizes the current drone state of the first drone to other drones in the cluster; Real-time environmental perception is performed through the sensors of each drone to obtain real-time environmental perception results; In response to the real-time environment perception result being that an obstacle is detected, triggering a dynamic risk assessment to update a global risk matrix; In response to the real-time environmental perception result being that there are no accidents, loading and using the global risk matrix of the previous cycle; The MEC server calculates a global demand matrix based on the distance between each UAV and the target point, and synchronizes the global demand matrix to the first UAV; The onboard computing unit of the first UAV determines an optimal action based on the current UAV state, the global risk matrix, and the global demand matrix, and generates a short-term planning path based on the optimal action; The first UAV broadcasts its short-term planned path to the MEC server, wherein the MEC server performs collaborative conflict detection based on the short-term planned paths of each UAV; In response to the MEC server detecting that the coordination conflict does not exist, the first UAV executes the optimal action to fly along the short-term planned path; in: The MEC server calculates a global demand matrix based on the distance between each UAV and the target point, and synchronizes the global demand matrix to the first UAV, including: The MEC server traverses all target points that need to be inspected and calculates the current demand value for each target point based on the distance between each drone and the target point; Integrating the demand values ​​of all target points to obtain the global demand matrix; and Synchronizing the global demand matrix to the first UAV; The onboard computing unit of the first UAV determines an optimal action based on the current UAV state, the global risk matrix, and the global demand matrix, and generates a short-term planning path based on the optimal action, including: Inputting the current drone state, the global risk matrix, and the global demand matrix into a locally deployed deep Q-network to obtain the optimal action; and A short-term planning path is generated based on the optimal action.

2. The multi-UAV path planning method for power transmission line inspection according to claim 1 is characterized in that: The current drone status includes location status, energy status, risk radius and identity.

3. The multi-UAV path planning method for power transmission line inspection according to claim 1 is characterized in that: Real-time environmental perception is performed through the sensors of each drone to obtain real-time environmental perception results, including: Scan the surrounding environment using the drone's camera and lidar to obtain real-time point cloud data and real-time images; The real-time point cloud data and the real-time image are compared with a preloaded static environment model to obtain the real-time environment perception result.

4. The multi-UAV path planning method for power transmission line inspection according to claim 3 is characterized in that: In response to the real-time environment perception result being that an obstacle is detected, triggering a dynamic risk assessment to update a global risk matrix, including: Obtaining location information of the obstacle; Applying a Gaussian distribution risk model to the obstacle based on the position information of the obstacle to obtain a new risk field; The new risk field is fused and superimposed with the original global risk matrix to update the global risk matrix.

5. The multi-UAV path planning method for power transmission line inspection according to claim 1 is characterized in that: Inputting the current drone state, the global risk matrix, and the global demand matrix into a deep Q-network deployed on an onboard computing unit to obtain the optimal action includes: The deep Q network performs forward propagation reasoning based on the current drone state, the global risk matrix, and the global demand matrix to obtain a vector containing multiple Q values; An executable action corresponding to the maximum Q value is selected from the vector containing multiple Q values ​​as the optimal action.

6. The multi-UAV path planning method for power transmission line inspection according to claim 1 is characterized in that: The first UAV broadcasts its short-term planned path to the MEC server, wherein the MEC server performs collaborative conflict detection based on the short-term planned paths of each UAV, including: The MEC server determines whether there is a short-term planned path of another UAV that intersects with the spatial position of the first UAV and the distance is less than a safety threshold.

7. The multi-UAV path planning method for power transmission line inspection according to any one of claims 1 to 6, characterized in that: The multi-UAV path planning method for power transmission line inspection further includes the following steps: After the first drone completes the optimal action, the BMS reports the updated remaining energy value and updates it to the current drone state.

Images