A Smart Collaborative Method and System for Emergency Inspection of Networked Unmanned Aerial Vehicle Clusters

Abstract

This invention provides an intelligent collaborative method and system for emergency inspection of networked unmanned aerial vehicle (UAV) swarms. The method constructs a 3D heterogeneous inspection scenario based on fault area information and generates an offline radio map. The scenario is modeled as a graph structure, and node embedding and graph embedding features are extracted. An Auto-K mechanism adaptively determines the optimal number of UAVs to deploy and merges this with the graph embedding to generate a global context vector. Inspection sequences are generated using a multi-UAV polling autoregressive approach, and the legality of actions is ensured by dynamic topology mask constraints. The time consumed by each UAV for obstacle avoidance flight, inspection operations, and data transmission is calculated, with the longest time taken as the completion time, and the overall inspection cost is calculated to generate a reward signal. The path generation and swarm size decision networks are jointly trained, ultimately outputting a collaborative scheduling scheme containing the optimal number of deployments and the complete inspection sequence. This invention achieves end-to-end collaborative optimization of swarm size and inspection paths, effectively reducing task completion time and overall cost.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) swarm control technology, specifically a networked UAV swarm emergency inspection intelligent collaborative method and system. Background Technology

[0002] The stable operation of power systems is crucial to modern society, and fault diagnosis and emergency inspection are essential to ensuring its stability. Reliability is a core aspect. Taking power system failures caused by extreme weather events such as typhoons and floods as an example, traditional ground inspection methods are often limited by complex terrain and traffic disruptions, making it difficult to complete fault location and information collection in a timely manner. In recent years, with the gradual opening of low-altitude airspace and the rapid development of the low-altitude economy, drones with rapid deployment capabilities and high mobility have gradually become a core technological means for power emergency inspection and auxiliary communication. Cellular-Connected UAVs (CC-UAVs), as important access terminals for cellular networks, can achieve highly reliable and wide-coverage data transmission by relying on existing communication infrastructure, demonstrating significant advantages in scenarios such as post-disaster emergency communication, power facility inspection, and network coverage optimization. Based on this, multi-drone collaborative inspection modes are gradually becoming an important means to improve emergency response efficiency.

[0003] However, existing multi-drone collaborative inspections in power grid scenarios still have the following shortcomings: 1. Insufficient 3D scene modeling capabilities: Existing methods mostly model inspection tasks based on 2D planar diagrams, simplifying the inspection targets into discrete nodes in Euclidean space. This makes it difficult to characterize physical constraints such as building occlusion and the physical shape of conductors in a confined 3D space. This often results in the planned paths being unexecutable in actual 3D space, failing to meet the real physical requirements of UAV-based follow-up operations and refined operations within towers.

[0004] 2. Fixed cluster size and lack of adaptive response capability: Existing methods typically preset the number of drones deployed as a fixed parameter, adopting a two-stage strategy of "allocation first, planning later," which makes it difficult to dynamically adjust the cluster deployment size according to the scale of the disaster and the task load. This results in wasted transport resources when the task scale is small, and insufficient transport capacity when the task scale is large, lacking the ability to adapt to different task loads.

[0005] 3. Incomplete time cost modeling: Existing methods typically only consider flight distance when establishing optimization cost models, without incorporating the time spent on detailed inspections within the tower or the communication offloading time generated by data transmission into the comprehensive cost model. This leads to a discrepancy between the optimization results and the actual comprehensive inspection costs, making it difficult to effectively minimize the real inspection costs.

[0006] In summary, existing methods still have significant limitations in handling 3D physical constraint modeling, adaptive cluster size decision-making, and comprehensive cost integrity. Therefore, there is an urgent need for an intelligent decision-making framework that can deeply integrate 3D physical environment modeling, adaptive cluster size decision-making, and comprehensive time cost modeling to achieve joint optimization of cluster size and multi-UAV collaborative inspection sequence in multi-UAV collaborative inspection scheduling. Summary of the Invention

[0007] This invention proposes an intelligent collaborative method and system for emergency inspection of networked unmanned aerial vehicle (UAV) swarms with adaptive sizing. When the UAV swarm performs emergency inspection tasks on heterogeneous power grid facilities, under the constraints of three-dimensional physical obstacle avoidance and power grid topology connectivity, it introduces an adaptive swarm size selection (Auto-K) mechanism and end-to-end collaborative task sequence generation based on deep reinforcement learning (DRL). This integrates the time consumption of three-dimensional obstacle avoidance flight, inspection operation, and communication offloading into the time cost for joint optimization, thereby minimizing the comprehensive inspection cost that includes time cost and scheduling cost.

[0008] To address the aforementioned technical problems, the present invention adopts the following technical solution: A smart collaborative method for emergency inspection of networked drone swarms includes the following steps: The system acquires the location information of transmission towers and line topology connections within the fault area, the location distribution information of available UAV nests and ground base stations, and the geometric information of building obstacles to construct a three-dimensional heterogeneous inspection scenario; and pre-calculates the data transmission rate of any point in the scenario to generate a three-dimensional offline radio map. The three-dimensional heterogeneous inspection scenario is modeled as a graph structure containing nest nodes, transmission tower nodes, and transmission line segment nodes. A one-time global feature extraction is performed on the graph structure to generate a node embedding matrix and a graph embedding vector. Based on the graph embedding vector, the optimal number of drones to be deployed, K, is inferred, and K is concatenated and fused with the graph embedding vector to generate a global context vector with cluster scale awareness capability. The global context vector is used as the initial guiding information, and the node embedding matrix is ​​used as the feature library of candidate nodes. These are input into the pointer decoder based on the attention mechanism. Through multi-machine polling autoregressive decoding, the joint inspection sequence of each UAV is generated alternately. A comprehensive physical assessment of the inspection cost is performed on the joint inspection sequence. The comprehensive time consumption, including flight obstacle avoidance, inspection tasks and data transmission, is calculated. The longest comprehensive time consumption among all UAVs is taken as the task completion time. The comprehensive inspection cost is calculated in combination with the number of UAVs deployed. Based on the comprehensive inspection cost, path generation reward and cluster size decision reward are generated respectively; the global context vector is input into the Critic network to estimate the baseline value; and joint parameters are updated based on the path generation reward, cluster size decision reward and the baseline value. Repeat the above steps for iterative training until the preset maximum number of iterations is reached, and output the current optimal collaborative inspection solution.

[0009] Furthermore, the transmission line segment node is the midpoint of the three-dimensional line connecting the exit point of the preceding tower and the entrance point of the subsequent tower; the entrance point serves as the obstacle avoidance navigation interface when the UAV approaches the tower, guiding the UAV to safely enter the tower's operating path; the exit point serves as the reference position pre-calculated for the optimal transmission point, guiding the UAV to complete data unloading near the tower.

[0010] Furthermore, generating a 3D offline radio map includes: discretizing the mission airspace into a two-dimensional grid at equal intervals in the horizontal direction, taking multiple height layers in the vertical direction, calculating the instantaneous transmission rate of each grid center point in each height layer under all base station and all sector combinations through Monte Carlo sampling, taking the maximum value as the channel transmission rate of that point, and statically storing it as a key-value pair; during online scheduling, using the real-time 3D coordinates of the UAV as an index to look up a table to complete the channel rate query.

[0011] Further, performing a one-time global feature extraction on the graph structure to generate a node embedding matrix and a graph embedding vector includes: taking the original features of each node in the graph structure as input, calculating the attention coefficient between each node and its neighboring nodes through the multi-head attention mechanism of the graph attention network, and weighting and aggregating the features of the neighboring nodes based on the attention coefficients to update the feature representation of each node. After propagation through a multi-layer graph attention network, the final features of all nodes constitute the node embedding matrix; and aggregating the node embedding matrix into a single vector through a global pooling operation, which serves as the graph embedding vector.

[0012] Furthermore, the optimal number of drones to be deployed is derived based on the Auto-K mechanism, which consists of an Auto-K selector and a K-value fusion unit connected in series, and includes the following sub-steps: Using the graph embedding vector as the input to the Auto-K selector, the signal passes through two layers of perceptrons and a temperature-sensitive layer. The Softmax function outputs the probability distribution of the cluster deployment scale; For the scale of cluster deployment Obtained by class sampling A greedy strategy is adopted to directly take the maximum probability as the optimal cluster size decision; During the training phase, the probability distribution of the scramble size is sampled by category to obtain the scramble size for this round. During the inference phase, a greedy strategy is adopted to take the maximum probability as the optimal scramble size. The K-value fusion processor concatenates the normalized value of the cluster deployment scale with the graph embedding vector, and obtains a cluster scale-aware global context vector through linear transformation. Further, generating the joint inspection sequence for each UAV includes: Based on the optimal number of drones to be deployed, a corresponding number of drones are selected from the available drone nests to participate in the collaborative inspection; during the training phase, random sampling is performed from all nest combinations; during the inference phase, all combinations are traversed, and each combination is sampled and decoded multiple times, and the combination with the lowest overall inspection cost is selected as the final deployment plan. A multi-drone polling autoregressive decoding mode is adopted, and each UAV is alternately scheduled. In each decoding step, the current UAV's position is embedded and fused with the global context vector to generate a query vector. The attention score of each candidate node is calculated by combining the node embedding matrix with the scaling dot product attention mechanism. Construct a dynamic topology mask function based on whether each candidate node in the global node set is a valid candidate action; After masking, the attention score of the legal action is normalized to the action selection probability. During the training phase, the probability distribution is randomly sampled. During the inference phase, greedy deterministic selection is adopted. The global unfinished task set and UAV position are updated according to the selected action. The process is iterated until the global task is cleared and all UAVs return to their nests, and a joint inspection sequence is output.

[0013] Furthermore, the calculation of the overall inspection cost includes: For the total time spent by each UAV performing its joint inspection sequence, the following calculations are made: the three-dimensional flight time based on dynamic obstacle avoidance, the sum of the fixed operation time of each task node, and the communication time for data return based on preset unloading trigger conditions. The dynamic obstacle avoidance adaptively adjusts the flight trajectory based on collision detection results of poles, building obstacles, and power transmission lines during the flight segment. If the flight path is safe, the flight time is calculated according to the normal displacement. If obstacle avoidance is triggered, the UAV climbs to a safe height above the obstacle, crosses it, and then descends to the target node. The data return adopts a combination of flight accumulation and hovering transmission. Unloading is triggered when the onboard buffer exceeds its limit, the UAV reaches a high transmission rate node, or the task ends. The communication time is calculated based on the accumulated data volume and the transmission rate at the unloading position. The longest total time among all UAVs is taken as the task completion time. Combined with the optimal number of UAVs deployed, the comprehensive inspection cost is calculated, and this comprehensive inspection cost is used as the environmental feedback signal for reinforcement learning.

[0014] Furthermore, a near-end policy optimization algorithm is used to update the joint parameters of the collaborative decision-making network. Specifically, after the complete joint inspection sequence is generated, path generation rewards and cluster scale decision rewards are calculated and distributed. The evaluation is conducted in segments based on whether the trajectory meets the full coverage constraint. When there is a missed detection, a negative penalty related to the missed detection rate is applied. When the trajectory is valid, a normalized reward settlement is performed based on the advantages and disadvantages of the completion time or the overall inspection cost relative to other trajectories in the batch. For the path generation network, the global context vector is used as input, the Critic network outputs a baseline value estimate, and the Monte Carlo reward is combined to calculate the advantage estimate; for the cluster-scale decision network, the reward is directly processed by batch centralization to construct the advantage estimate. Based on two advantage estimates, path generation loss and scale decision loss are constructed using the truncated agent objective form of near-end policy optimization, and a joint optimization objective is constructed by combining the value mean square error loss of the Critic network and the policy information entropy regularization term. By calculating the gradient of the joint optimization objective and performing gradient descent, the parameters of the path generation network, cluster size decision network, and Critic network are simultaneously updated.

[0015] Furthermore, during the iterative training process, the average cost of feasible solutions in the current batch is monitored. The current network parameters are saved to the checkpoint only when full coverage is satisfied and the average cost is better than the historical best record. After training, the checkpoint is loaded as the final inference model. After entering the inference deployment stage, the collaborative decision network is inferred using a greedy strategy. The output is a multi-machine collaborative scheduling scheme that includes the optimal number of drones to be dispatched and the complete spatial node access sequence of each drone starting from its home nest, sequentially visiting the power transmission towers and power transmission lines it is responsible for, completing data unloading at the optimal transmission point, and returning to the nest.

[0016] On the other hand, the present invention provides a networked unmanned aerial vehicle (UAV) swarm emergency inspection intelligent collaborative system, comprising: The scene construction module is used to acquire the location information of transmission towers and line topology connections within the fault area, the location distribution information of available UAV nests and ground base stations, and the geometric information of building obstacles to construct a three-dimensional heterogeneous inspection scene; and to pre-calculate the data transmission rate of any point in the scene to generate a three-dimensional offline radio map. The graph encoding module is used to model the three-dimensional heterogeneous inspection scene as a graph structure containing nest nodes, transmission tower nodes and transmission line segment nodes, and to perform one-time global feature extraction on the graph structure to generate a node embedding matrix and a graph embedding vector. The cluster size decision module is used to infer the optimal number of drones to be dispatched K based on the graph embedding vector, and to concatenate and fuse K with the graph embedding vector to generate a global context vector with cluster size awareness capability. The sequence generation module is used to take the global context vector as initial guiding information and the node embedding matrix as a feature library of candidate nodes, and input them into the pointer decoder based on the attention mechanism. Through multi-machine polling autoregressive decoding, the joint inspection sequence of each UAV is generated alternately. The physical assessment module is used to perform a comprehensive physical assessment of the inspection cost of the joint inspection sequence, calculate the comprehensive time including flight obstacle avoidance, inspection tasks and data transmission, take the longest comprehensive time among all UAVs as the task completion time, and calculate the comprehensive inspection cost in combination with the number of UAVs deployed. The learning and training module is used to generate path generation rewards and cluster size decision rewards based on the comprehensive inspection cost; input the global context vector into the Critic network to estimate the baseline value; and perform joint parameter updates based on the path generation rewards, cluster size decision rewards, and the baseline value. The optimal solution output module is used to repeatedly execute the above steps for iterative training until the preset maximum number of iterations is reached, and then output the current optimal collaborative inspection solution.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Compared with the limitations of existing methods that are based on two-dimensional planar modeling and have difficulty in depicting three-dimensional physical constraints, this invention constructs a high-fidelity three-dimensional heterogeneous inspection scene that integrates the three-dimensional structure of the tower, the parabolic shape of the conductor, and building obstacles. Through two-stage collision detection and dynamic topology masking, it ensures that the planned path meets the physical feasibility of three-dimensional space, while also meeting the real needs of UAVs following the line and refined operations inside the tower.

[0018] 2. Compared with the limitations of existing methods that preset the number of drones to a fixed parameter and lack adaptive response capabilities, the Auto-K mechanism proposed in this invention can dynamically determine the optimal number of drones to be deployed based on the task size and topology. This avoids wasting transport resources when the task size is small and ensures sufficient transport capacity when the task size is large, effectively balancing task timeliness and scheduling costs.

[0019] 3. Compared with existing methods that only consider flight distance and do not include the communication unloading time caused by image data transmission in the task completion time, this invention integrates the three-dimensional obstacle avoidance flight time, hovering operation time and communication unloading time into the comprehensive time cost calculation, so that the optimization target is consistent with the actual comprehensive inspection cost, and achieves effective minimization of the real inspection cost. Attached Figure Description

[0020] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is an example scene layout diagram of an embodiment of the present invention; Figure 3 This is a schematic diagram of a heterogeneous inspection diagram according to an embodiment of the present invention; Figure 4 This is a diagram illustrating the inspection route for a single-circuit straight-line wine glass tower in an embodiment of the present invention. Figure 5 This is a three-dimensional offline radio map according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the feature extraction process in an embodiment of the present invention; Figure 7 A schematic diagram illustrating the generation of global context vectors in an embodiment of the present invention; Figure 8 This is a schematic diagram of the multi-machine collaborative sequence generation process according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the physical assessment process for comprehensive inspection costs in an embodiment of the present invention; Figure 10 This is a system structure diagram of an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Example 1 like Figure 1 As shown, this invention provides an intelligent collaborative method for emergency inspection of networked drone swarms, comprising the following steps: Step 1: Obtain the location information of transmission towers and line topology connections within the fault area, the location distribution information of available UAV nests and ground base stations, and the geometric information of building obstacles to construct a three-dimensional heterogeneous inspection scenario; and pre-calculate the data transmission rate of any point in the scenario to generate a three-dimensional offline radio map; Step 2: Model the three-dimensional heterogeneous inspection scene as a graph structure containing machine nest nodes, transmission tower nodes, and transmission line segment nodes. Perform a one-time global feature extraction on the graph structure to generate a node embedding matrix and a graph embedding vector. Step 3: Based on the graph embedding vector, the optimal number of drones to be deployed, K, is inferred and then concatenated with the graph embedding vector to generate a global context vector with cluster scale awareness capability; Step 4: Use the global context vector as initial guiding information and the node embedding matrix as a feature library of candidate nodes, input them into the pointer decoder based on the attention mechanism, and generate joint inspection sequences of each UAV alternately through multi-machine polling autoregressive decoding; Step 5: Perform a comprehensive physical assessment of the inspection cost of the joint inspection sequence, calculate the comprehensive time including flight obstacle avoidance, inspection tasks and data transmission, take the longest comprehensive time among all UAVs as the task completion time, and calculate the comprehensive inspection cost in combination with the number of UAVs deployed. Step 6: Based on the comprehensive inspection cost, generate path generation reward and cluster size decision reward respectively; input the global context vector into the Critic network to estimate the baseline value; perform joint parameter update based on the path generation reward, cluster size decision reward and the baseline value; Step 7: Repeat the above steps for iterative training until the preset maximum number of iterations is reached, and output the current optimal collaborative inspection solution.

[0023] This invention aims to propose a networked drone-based emergency power grid inspection and dispatch system for complex and confined spaces. It utilizes an adaptive-scale networked drone swarm to perform full-coverage inspections and data transmission of all power grid facilities within the emergency area. Through end-to-end joint optimization, the system achieves shorter overall completion time, lower resource scheduling costs, and greater physical feasibility for inspection tasks.

[0024] Assuming in The faulty power grid area includes 5 available cell towers, 5 ground base stations, 25 transmission towers, and 3 transmission lines. Figure 2 This is the corresponding scene layout diagram. Based on discretization rules, the system equates any two adjacent power line segments between towers to a discrete power line node, thus generating a total of 52 task nodes. This system needs to jointly decide the number of drones deployed and the multi-drone collaborative path, completing image acquisition and data transmission for all the aforementioned towers and power line nodes while avoiding complex obstacles.

[0025] In step 1 of this embodiment, upon receiving a power grid fault alarm, the inspection scene within the alarm area is first perceived and information is collected. Specifically, the three-dimensional coordinates and orientation of the transmission towers, the topological connections of the transmission lines, the spatial distribution of building obstacles, and the location information of various ground base stations within the alarm area are read to construct a complete three-dimensional scene description. Based on this, a global spatial node set is defined. ,in, and The set of discrete spatial particles represents the drone nest and the power transmission tower, respectively. The drone adopts a "one drone, one nest" configuration. For a transmission line segment, a virtual working node is defined. Any line segment connecting the exit point of the preceding tower and the entrance point of the succeeding tower is used. The midpoint of the line connecting these two points in three-dimensional space is extracted as the virtual node representing that conductor segment. A topological edge set between nodes is constructed based on the physical connection relationships of the power grid. This is used to define the legal transfer relationships between drones and consists of three subsets:

[0026] in, This indicates the connection between the nest node and all pole nodes in the scene. This represents the fully interconnected structure between any two towers. This represents a chain-like topology formed by a wire node and its adjacent nodes. Any node not belonging to... There are no legitimate flight transfer paths between any of the node pairs. Figure 3 The heterogeneous inspection graph is shown, which consists of three types of nodes and three types of topological edge sets.

[0027] For any transmission tower Parametric 3D modeling is performed, and the corresponding data acquisition points are statically fixed in the local coordinate system of the tower to form a complete standardized operation path within the tower, thus standardizing the precise inspection actions of the UAV. The first data acquisition point on the operation path within the tower is called the entry point, and the last data acquisition point is called the exit point. Figure 4 This demonstrates an inspection path within a tower using a single-circuit, straight-line, glass-shaped AC transmission tower as an example. The entry point serves as the obstacle avoidance navigation interface when the drone approaches the tower, guiding it safely into the tower's inspection path. The exit point serves as the pre-calculated reference position for the optimal transmission point, guiding the drone to unload data near the tower. The global coordinates of the entry and exit points of the tower's inspection path are calculated as follows:

[0028] in, This is the entry point for the work path inside the tower. This is the exit point of the work path inside the tower. For towers The center's global coordinates, The spatial orientation angle of the tower on the horizontal plane. and These represent the lateral, depth, and height offsets of the entry and exit points in the local coordinate system of the tower. To prevent collisions between drones and the tower's metal truss, an axis-aligned bounding box (AABB) is used to enclose the tower's safety exclusion zone. The tower's safety half-width and height uniquely determine the tower's collision avoidance zone. Similarly, for any architectural obstacle in the scene, AABB is used for 3D spatial occupancy modeling, and its obstructing area is determined. It is determined by the coordinates of the building's center, the length of its base side, and its vertical height.

[0029] For transmission line segments, since the conductors exhibit a catenary shape under their own gravity, a high-precision approximate model is performed using a parabolic equation, and normalized path parameters are introduced. The three-dimensional continuity equation of the conductor Defined as

[0030] in For the maximum sag, The horizontal span between the two towers This represents the maximum sag of the conductor. , They are connecting towers and The three-dimensional coordinates of the hanging points at both ends of the transmission line segment.

[0031] Finally, a three-dimensional offline radio map is constructed to quantify the channel transmission rate at any spatial location. This is achieved by integrating the 3GPP three-dimensional antenna pattern model, the ITU air-to-ground path loss model, and the Rice / Rayleigh small-scale fading model, combined with building obstruction areas. Calculate any coordinate point in computation space The letter of the place is more than Therefore, the maximum theoretical transmission rate can be obtained by Shannon's theorem:

[0032] in, For any coordinate point in space The maximum theoretical transmission rate at that location; Available channel bandwidth for drone data backhaul.

[0033] Pre-set the length to Width The rectangular task airspace is in the horizontal direction with Discretize at equal intervals as A two-dimensional grid, taking in the vertical direction Each height layer, each height layer The coordinates of the center points of each grid on the grid are These elements together form a three-dimensional discrete grid. For each grid center point, the instantaneous transmission rate under all base station and all sector combinations is calculated using Monte Carlo sampling, and the maximum value is taken as the channel transmission rate at that point. Static storage as key-value pairs During online scheduling, it is only necessary to look up the table using the real-time 3D coordinates of the UAV as an index. The channel rate query is completed in a time complexity of [time complexity]. Figure 5 A 3D offline radio map of each altitude level within the example task area is displayed.

[0034] In step 2 of this embodiment, since the multi-UAV collaborative inspection and scheduling problem is essentially a mixed-integer non-linear programming (MINLP) problem, a collaborative decision network is invoked to generate a multi-UAV inspection and scheduling scheme. The collaborative decision network adopts an encoder-decoder architecture, consisting of three modules connected in series: a graph attention encoder, an Auto-K mechanism, and a pointer network decoder.

[0035] During the decision initialization phase, the system first extracts the global spatial node set. The static characteristics of all nodes in the array. For any node... Its initial eigenvector It includes four dimensions:

[0036] Spatial location characteristics: The normalized three-dimensional physical coordinates are used, where the tower node takes the global coordinates of its entry point, the nest node takes the actual position coordinates of the nest center, and the wire node takes the coordinates of the virtual work point.

[0037] Node type characteristics: Use One-Hot encoding to identify whether the node is a nest, pole, or wire node; Topological prior features: The normalized ID of the line to which it belongs. This is the topological position number of the node in the line, used to assist the network in implicitly identifying the backbone structure of the power grid corridor; Communication sensing characteristics: The normalized channel transmission rate is the optimal transmission point for each node. Wire nodes do not participate in data transmission; therefore, their communication awareness characteristic is always 0, distinguishing them from physical nodes with clearly defined transmission points. The system is configured to transmit data back only at the optimal transmission points near the exit point of tower nodes and near the nest nodes. At wire nodes, the UAV is in a low-speed following flight state and is not used as an unloading trigger position. Due to factors such as building obstruction, the physical location of a node is often not the optimal communication location. Therefore, the system pre-calculates the optimal transmission points around each tower node and nest node offline. For tower nodes, a two-stage strategy of "coarse search + fine search" is adopted. First, a coarse-grained search is performed on multiple Fibonacci spheres centered on the exit point, and then a fine search is performed using a 5 × 5 × 5 grid at the optimal point of the coarse search. For nest nodes, combined sampling is performed at multiple horizontal radii and vertical heights, similarly excluding building collision points. The height of all candidate points is limited to 10 m. Within a range of 100 meters, and with a distance of no more than 90 meters from the original exit point, excluding points that fall inside buildings, the location with the highest transmission rate is finally identified on the offline radio map and recorded as the optimal transmission point. The obtained optimal transmission rate After normalization, the 9th dimension of the feature vector of the above nodes is assigned to the encoder to perceive the potential communication backhaul capability of each node.

[0038] The GAT encoder first maps the original feature vectors of each node to a learnable linear projection layer. Dimensional embedding space, then through Multi-head graph attention propagation enables each node to adaptively aggregate features from its neighboring nodes through multiple attention heads, capture spatial topological dependencies between nodes, globally extract node features from the task graph, and output a node embedding matrix. With graph embedding vectors ;in The first Each node passes through The embedding vector after attention propagation in the layered graph. The graph embedding vector is obtained by average pooling of the embeddings of all nodes, i.e.

[0039] in, For the i-th node Embedded vectors after layer graph attention propagation For the dimensions of the embedded space, This represents the total number of nodes in the global space node set.

[0040] like Figure 6 As shown, in this embodiment, the specific execution flow of step 2 is as follows: Step 201: Process begins; Step 202: Construct a 9-dimensional static feature vector for all nodes (nest / tower / wire), including normalized 3D spatial location, node type, topological priors, and communication-aware features; map the original node features from 9D to 1D using a linear projection layer. Dimensional embedding dimension; Step 203: via Layer GAT processes and aggregates node neighborhood information to generate a high-dimensional node embedding matrix. and average pooling full-image embedding features ; Step 204: Process ends.

[0041] In step 3 of this embodiment, an Auto-K mechanism is introduced between the encoder and decoder to achieve automated end-to-end decision-making for the cluster deployment size K. The Auto-K mechanism consists of two modules connected in series: an Auto-K selector and a K-value fusion unit. The Auto-K selector uses the graph embedding vector output in step 2. As input, it passes through two layers of sensors and includes temperature parameters. The Softmax function outputs the probability distribution of the cluster deployment size. ,Right now

[0042] in , Here are the weight matrix and bias terms for the first fully connected layer. , The weight matrix and bias terms of the second fully connected layer, and the temperature parameters. Used to control the balance between exploration and exploitation. During the training phase... The scale of this round of deployment was obtained by category sampling. During the inference phase, a greedy strategy is adopted to directly take the maximum probability as the optimal swarm deployment size. Subsequently, the K-value fusion unit will... Normalized values ​​and graph embedding The concatenation and linear transformation yield a cluster-scale-aware global context vector. ,Right now

[0043] in This is the normalized cluster size value. This represents a vector concatenation operation. and Here are the learnable weight matrix and bias terms for the K-value fusion machine.

[0044] like Figure 7As shown, in this embodiment, the specific execution flow of step 3 is as follows: Step 301: Process begins; Step 302: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require Input the Auto-K selector module in the Auto-K mechanism and output the probability distribution of the optimal number of drones K; Step 303: Based on probability distribution sampling, obtain the number of deployments K in this round, normalize it using a K-value fusion tool, and then embed it with the full-image features. Integration; Step 304: Output a global context vector containing cluster size information. .

[0045] Step 305: Process ends.

[0046] In step 4 of this embodiment, after the K value is determined, it is necessary to select from all elements within the scene. K drones were selected from the existing drone nests, and one drone was assigned to each nest to participate in collaborative inspections. During the training phase, drones were selected from all the nests... One species is randomly sampled from the nest combination to increase the diversity of training samples; during the inference phase, all are traversed. For each of the following combinations, multiple sampling and decoding operations are performed, and the combination with the lowest overall inspection cost is selected as the final deployment scheme. Subsequently, the pointer network decoder uses the output from step 3... Given the condition, a multi-machine polling autoregressive model is used to generate task sequences for K UAVs sequentially. At each decoding time step... System scheduling To meet the needs of drones Using the embedding vector of the node where the drone is currently located and The query vector is obtained by concatenating and then performing a linear transformation. Calculate its relationship with the global node embedding matrix. Attention score between scaled dot products:

[0047] in, For query vector With global node embedding matrix Attention score between scaled dot products; To query the learnable linear projective weight matrix of the key, Scaling factor to prevent gradient vanishing.

[0048] set up The set of inspection task nodes that have not yet been completed at any given time is , For the first The home nest of the drone. For the tower The set of adjacent wire nodes, the tower completion determination function is defined as follows:

[0049] The system is based on the global node set Each candidate node in To determine if an action is legal, the dynamic topology mask function is constructed as follows:

[0050] Masking conditions (a) and (b) enable differentiated task completion status management, ensuring that power line nodes are masked upon task completion, while tower nodes are only masked after their own and adjacent tower nodes have completed their tasks, preserving the flexibility of tower nodes as aerial transit hubs. Masking condition (c) forces all UAV movements to strictly adhere to the predefined power grid physical topology edge set. This ensures a "tower-line-tower" spatial connectivity structure, significantly reducing the space available for ineffective exploration. Masking condition (d) mandates that each UAV must and can only return to its assigned home base after the mission is completed. Masking condition (e) prohibits any UAV from returning early before the global mission is finished, preventing the premature withdrawal of some low-altitude transport capacity and resulting resource idleness. The system also provides a deadlock prevention mechanism: when a UAV's physical topology restricts all forward legal actions, resulting in an empty candidate action set, the system will proactively lift the return restriction for that UAV, allowing it to return to its home base early. This ensures the absolute smoothness and physical security of the joint decoding process.

[0051] After rigorous filtering by the dynamic masking mechanism, the model uses the Softmax function to normalize the attention scores of legitimate candidate actions, generating an action probability distribution pointing to the next task node. :

[0052] in For temperature parameters, for The dynamic topological mask vector of all candidate nodes at time 1, candidate nodes The dynamic topology mask is composed of The calculated value is: The probability of a valid node is 0, and after Softmax, the probability of an invalid node becomes zero, thus strictly limiting the action space to the set of valid candidate nodes. Similarly, during the training phase, the probability distribution is adjusted. Perform random sampling of categories and select actions To maintain the exploratory nature of the strategy and avoid getting trapped in local optima; during the reasoning phase, greedy decoding is used. Obtain a deterministic optimal scheduling scheme. Based on the selected actions. The system updates the global set of incomplete tasks in real time. and the The current spatial location of the drone.

[0053] After the current state and mask are updated, proceed to the next time step. Afterwards, the system scheduling authority continues to be determined based on the mapping function. The task automatically transfers to the next drone. This autoregressive decoding process will be executed alternately until the global task set is cleared and all drones have returned to their home nests. Only then can the system terminate decoding and finally output a joint inspection sequence that satisfies the physical topology constraints. This information is then passed to the execution evaluation layer for comprehensive cost accounting.

[0054] like Figure 8 As shown, in this embodiment, the specific execution flow of step 4 is as follows: Step 401: Process begins; Step 402: Initialization Phase: Start the pointer network decoder, and... Each drone initializes at its dedicated nesting node and initializes its time step. And the global set of unfinished task nodes; Step 403: Determine the current decision-making entity: at the current time step It adopts a multi-machine polling autoregressive mode, based on the formula Determined by the first Using drones to make action decisions; Step 404: Generate dynamic mask. Invoke the dynamic topology masking mechanism, based on the physical connectivity constraints of the power grid "tower-line-tower", task leakage prevention and nest affiliation constraints, force the probability of illegal action nodes to zero, and generate the current set of legal candidate nodes; Step 405: Node Sampling and State Update: Decoder Combines Node Embedding With global context Calculate the attention score, generate a probability distribution for the legitimate candidate nodes and sample them, and determine the first... The next access node for the drone is determined, and the global set of unfinished task nodes and the drone's position status are updated simultaneously. Step 406: Determine if decoding has terminated: Check if the total number of unfinished task nodes is 0, and if all... Have all the drones chosen to return to their respective home nests? If not, let time step... Return to step 403 and continue the alternating loop; if so, end the autoregressive decoding and proceed to step 407. Step 407: Output A complete and independent collaborative inspection path sequence for each drone; Step 408: Process ends.

[0055] In step 5 of this embodiment, the drone swarm follows the joint inspection sequence output in step 4. To calculate the overall inspection cost, the time spent on each drone during the mission was calculated individually. As environmental feedback signals for reinforcement learning. For the first drones along the sequence Total flight time The time required for three-dimensional obstacle avoidance flight Inspection operation time Communication offloading time It consists of three parts.

[0056] Specifically, when the drone departs from the node fly to node ( When this happens, its transfer trajectory can be abstracted as a straight line segment in three-dimensional space. The flight segment requires a two-stage collision detection process: First, based on the AABB collision detection model between the tower and the building, the ray intersection algorithm (Slab Method) is used to check the line segment. Does it penetrate any environmental building bounding box? or tower enclosure box Secondly, the continuous parabolic space equation in the transmission line segment. Based on this, a piecewise linear approximation strategy is adopted to approximate the catenary of each conductor along the path parameters. Discretize at equal intervals A series of tiny straight line segments that are connected end to end, denoted as the first... Each micro-segment is The flight trajectory of the UAV is solved sequentially. With each conductor micro segment The shortest spatial distance between them. Check if the minimum distance is lower than the set physical isolation threshold. ,Right now If both of the above two stages of verification pass, the flight path is deemed absolutely safe, and the flight time is calculated by adding up the time taken for horizontal and vertical displacement. Once any obstacle avoidance condition is triggered, the drone must respond with a vertical speed... Climb to a safe height at the top of the obstacle and then move horizontally. The aircraft flew horizontally across the target area and finally descended to land there. to The actual physical obstacle avoidance transfer time between them is denoted as Then the first drones along the sequence The cumulative flight time is calculated as follows:

[0057] in To transfer decision variables, the first... The drones that were deployed Is it leaving the node? Afterwards, it will fly directly to the node. .

[0058] After arriving at each task node, the drone performs the corresponding image acquisition actions. The time required for each tower operation is [not specified]. The same fixed value is applied to all towers, and the operation time for each virtual node of the power line is calculated. The time for a single machine to perform a cumulative inspection is uniformly calibrated as a constant.

[0059] in, For nodes The time consumed by a single inspection operation, when the node When it is a tower node, take a fixed value. When node When used as a virtual node for wires, a fixed value is used. :

[0060] in, The fixed time required for drones to perform hovering and detailed inspection operations at pole nodes. The fixed time for a drone to perform follow-up shooting operations by passing through a virtual node on a power line.

[0061] For data backhaul, a "flight-accumulation-hover transmission" operation mode is adopted. When the UAV meets any of the following trigger conditions at the nest node or pole node, the indicator variable... Set to 1, the drone will yaw and fly towards the optimal transmission point. Perform data unloading: 1. Passive overflow trigger: When the onboard cache is overflowed. Approaching the capacity limit At that time, force uninstallation will be triggered; 2. Proactive Opportunity Trigger: When a drone arrives at a certain tower node and the optimal transmission rate of that tower is among the top 20% of all towers in the scene, at the same time... At that time, it actively triggers offloading to preempt high-quality channels; 3. Mission endpoint trigger: If there is still residual data in the onboard cache at the end of the inspection sequence, yaw to the optimal transmission point near the hive to clear and unload the data before returning to the hive.

[0062] The cumulative communication offloading time for a single machine is calculated as follows:

[0063] in For drones at nodes The amount of unreturned data accumulated in the cache when the uninstallation is triggered. This represents the channel transmission rate at the unloading location.

[0064] when After all drones have completed their physical assessments, the longest completion time among all drones will be taken as the mission completion time. Combined with the scale of deployment Substitute into the comprehensive inspection cost calculation formula:

[0065] in , This is the time cost weighting coefficient. This is the weighting coefficient for the fixed capacity cost of a single drone deployment. This cost value serves as an environmental feedback signal for reinforcement learning, transmitted to the PPO training module, and driving the parameters of the collaborative decision-making network to continuously update and converge towards the optimal scheduling direction.

[0066] like Figure 9 As shown, the specific steps of step 5 are as follows: Step 501: Process begins; Step 502: Initialize the evaluation loop: Obtain the current UAV's collaborative inspection task sequence, and traverse the adjacent node pairs in the trajectory sequence in turn; Step 503: Perform flight transfer and obstacle avoidance judgment on the current adjacent node pair: calculate the actual flight time of the current segment by combining the collision detection of the two stages; Step 504: Execute the task and accumulate data: When the UAV arrives at the target node, the hovering data acquisition time at that node is recorded, and the newly generated image data is added to the current onboard cache. Step 505: Data offloading condition judgment: Determine whether the current state meets one of the following conditions: passive overflow trigger, active opportunity trigger, and task endpoint trigger; if it meets the condition, the time spent flying to the optimal hovering transmission point and data return is included, and the onboard cache is cleared; if it does not meet the condition, proceed directly to step 506. Step 506: Single-drone trajectory traversal completion judgment: Determine whether the entire trajectory of the current drone has been traversed. If not, update the current position and return to step 503 to process the next flight segment; if yes, proceed to step 507; Step 507: Add up the time consumed by all the aforementioned flights, operations, and unloading to obtain the result. The final completion time of the drone; Step 508: Global evaluation ends: Determine all Have all drones been evaluated? If not, select the next drone and return to step 502; if yes, proceed to step 509. Step 509: Process ends.

[0067] In step 6 of this embodiment, the PPO algorithm is used to perform end-to-end joint training on the collaborative decision-making network. This invention employs a round-based sparse reward mechanism, i.e., a complete joint trajectory sequence. After generation, the physics evaluation module calculates and issues reward signals in one go, generating path rewards. Monte Carlo reward estimation as a path generation network Cluster size reward Monte Carlo return estimation for Auto-K networks Path generation rewards Cluster size decision rewards All are evaluated in segments based on whether the trajectory meets the full coverage constraint, i.e.

[0068]

[0069] in, This represents the number of task nodes missed in the trajectory. This represents the total number of global task nodes. , , , , The penalty coefficient is... , The base reward coefficient for legal trajectories. and These are the mean and standard deviation of the completion times for all valid trajectories within the current batch, respectively. A normalized scaling scalar that matches the scene size.

[0070] Rewards generated along the path In the process, when there are missed detections in the trajectory ( When the trajectory is valid, a negative reward strongly correlated with the missed detection rate and the number of missed nodes is applied, forcing the network to prioritize satisfying the full coverage constraint; when the trajectory is valid ( When the current trajectory's completion time is normalized and compared to the batch average, a positive reward is given if the time is lower than the average, and a negative reward is given if it is higher, thus guiding the strategy to continuously optimize towards lower time costs. This applies to cluster-scale reward decisions. In the process, a fixed negative penalty is imposed for missed inspections, while the penalty is applied based on the overall inspection cost when the trajectory is valid. Normalized reward settlement is implemented so that the Auto-K selector can perceive the combined impact of completion time and deployment scale on total cost, and reasonably control the number of drones deployed under the premise of meeting timeliness standards.

[0071] For path generation networks, the Critic network uses a global context vector that incorporates cluster-scale information. As input, output the baseline value estimate of the current graph state, i.e.

[0072] The advantage calculation of the path generation network is performed using the benchmark value provided by the Critic network.

[0073] For Auto-K networks, the reward is directly related to... Intra-batch centralization is adopted, and the batch mean is subtracted. Construct its own unique advantage estimate, namely

[0074] Right now, s represents the baseline value of the current graph state, and s represents the current environmental state. The baseline value of the current graph state is used to estimate the advantage of the path generation network. For the estimation of the advantage of Auto-K networks; Subsequently, path generation loss Losses related to scale decision-making All adopt the PPO truncated proxy target form, that is

[0075] in The importance sampling ratio for the path generation strategy. Let be the system state at time t. The action decision at time t; The sampling ratio for the cluster size strategy; The current learnable parameters of the collaborative decision-making network. and This is the old strategy from before the last parameter update; To truncate hyperparameters.

[0076] Value mean squared loss combined with Critic network Together with the policy information entropy regularization term, they constitute the final joint optimization objective:

[0077] in and These are the information entropy values ​​for the path and scale action probability distributions, respectively. , Here is the corresponding regularization coefficient. This is the value loss weighting coefficient. (Based on...) Calculate the gradient and perform gradient descent, while simultaneously updating the parameters of the path generation network, Auto-K network, and Critic network, driving the collaborative decision-making network to continuously converge toward the optimal scheduling direction.

[0078] In step 7 of this embodiment, to address the performance regression issue caused by policy oscillations in the later stages of reinforcement learning, this invention further introduces a peak weight rollback strategy based on average stable performance. During training, the average cost of feasible solutions in the current batch is continuously monitored. Only when the batch achieves full coverage of inspection nodes and the average cost breaks through the historical best record is all current network parameters saved to a temporary checkpoint. After all training is completed, this peak checkpoint is uniformly loaded as the final inference model.

[0079] Once the system has completed training and converged, it enters the inference and deployment phase. The system uses a greedy strategy to infer the collaborative decision-making network, ultimately outputting a complete multi-machine collaborative scheduling scheme, specifically including the following two levels of decision results: Swarm Scale: The optimal number of drones to be deployed is adaptively determined by the Auto-K mechanism. ; Multi-drone collaborative inspection sequence: Each drone takes off from its assigned nest, visits its assigned power transmission tower and power line segment in sequence, unloads data at the optimal transmission point, and finally returns to the nest, completing the complete spatial node access sequence. .

[0080] Example 2 like Figure 10 As shown, this embodiment provides a networked drone swarm emergency inspection intelligent collaborative system, including: The scene construction module is used to acquire the location information of transmission towers and line topology connections within the fault area, the location distribution information of available UAV nests and ground base stations, and the geometric information of building obstacles to construct a three-dimensional heterogeneous inspection scene; and to pre-calculate the data transmission rate of any point in the scene to generate a three-dimensional offline radio map. The graph encoding module is used to model the three-dimensional heterogeneous inspection scene as a graph structure containing nest nodes, transmission tower nodes and transmission line segment nodes, and to perform one-time global feature extraction on the graph structure to generate a node embedding matrix and a graph embedding vector. The cluster size decision module is used to infer the optimal number of drones to be dispatched K based on the graph embedding vector, and to concatenate and fuse K with the graph embedding vector to generate a global context vector with cluster size awareness capability. The sequence generation module is used to take the global context vector as initial guiding information and the node embedding matrix as a feature library of candidate nodes, and input them into the pointer decoder based on the attention mechanism. Through multi-machine polling autoregressive decoding, the joint inspection sequence of each UAV is generated alternately. The physical assessment module is used to perform a comprehensive physical assessment of the inspection cost of the joint inspection sequence, calculate the comprehensive time including flight obstacle avoidance, inspection tasks and data transmission, take the longest comprehensive time among all UAVs as the task completion time, and calculate the comprehensive inspection cost in combination with the number of UAVs deployed. The learning and training module is used to generate path generation rewards and cluster size decision rewards based on the comprehensive inspection cost; input the global context vector into the Critic network to estimate the baseline value; and perform joint parameter updates based on the path generation rewards, cluster size decision rewards, and the baseline value. The optimal solution output module is used to repeatedly execute the above steps for iterative training until the preset maximum number of iterations is reached, and then output the current optimal collaborative inspection solution.

[0081] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0082] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations to the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.

[0083] All other parts not described in detail are existing technologies.

Claims

1. A method for intelligent collaborative emergency inspection of networked unmanned aerial vehicle (UAV) swarms, characterized in that, Includes the following steps: Acquire the location information of transmission towers and line topology connections within the fault area, the location distribution information of available UAV nests and ground base stations, and the geometric information of building obstacles to construct a three-dimensional heterogeneous inspection scenario. It also pre-calculates the data transmission rate at any point in the space within the scenario and generates a three-dimensional offline radio map; The three-dimensional heterogeneous inspection scenario is modeled as a graph structure containing nest nodes, transmission tower nodes, and transmission line segment nodes. A one-time global feature extraction is performed on the graph structure to generate a node embedding matrix and a graph embedding vector. Based on the graph embedding vector, the optimal number of drones to be deployed, K, is inferred, and K is concatenated and fused with the graph embedding vector to generate a global context vector with cluster scale awareness capability. The global context vector is used as the initial guiding information, and the node embedding matrix is ​​used as the feature library of candidate nodes. These are input into the pointer decoder based on the attention mechanism. Through multi-machine polling autoregressive decoding, the joint inspection sequence of each UAV is generated alternately. A comprehensive physical assessment of the inspection cost is performed on the joint inspection sequence. The comprehensive time consumption, including flight obstacle avoidance, inspection tasks and data transmission, is calculated. The longest comprehensive time consumption among all UAVs is taken as the task completion time. The comprehensive inspection cost is calculated in combination with the number of UAVs deployed. Based on the comprehensive inspection cost, path generation rewards and cluster size decision rewards are generated respectively. The global context vector is input into the Critic network to estimate the baseline value; based on the path-generated reward, the cluster-size-determined reward, and the baseline value, joint parameters are updated; Repeat the above steps for iterative training until the preset maximum number of iterations is reached, and output the current optimal collaborative inspection solution.

2. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, The transmission line segment node is the midpoint of the three-dimensional line connecting the exit point of the preceding tower and the entrance point of the subsequent tower; the entrance point serves as the obstacle avoidance navigation interface when the UAV approaches the tower, guiding the UAV safely into the tower's operating path. The exit point serves as the pre-calculated reference position for the optimal transmission point, guiding the drone to unload data near the tower.

3. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, Generating a 3D offline radio map involves: discretizing the mission airspace into a 2D grid at equal intervals in the horizontal direction, taking multiple height layers in the vertical direction, calculating the instantaneous transmission rate of each grid center point in each height layer under all base station and all sector combinations using Monte Carlo sampling, taking the maximum value as the channel transmission rate of that point, and statically storing it as a key-value pair; during online scheduling, using the real-time 3D coordinates of the UAV as an index to look up a table and complete the channel rate query.

4. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, The process of performing a one-time global feature extraction on the graph structure to generate a node embedding matrix and a graph embedding vector includes: taking the original features of each node in the graph structure as input, calculating the attention coefficient between each node and its neighboring nodes through the multi-head attention mechanism of the graph attention network, and weighting and aggregating the features of the neighboring nodes based on the attention coefficients to update the feature representation of each node. After propagation through a multi-layer graph attention network, the final features of all nodes constitute the node embedding matrix; and aggregating the node embedding matrix into a single vector through a global pooling operation, which serves as the graph embedding vector.

5. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 4, characterized in that, The optimal number of drones to be deployed is derived based on the Auto-K mechanism, which consists of an Auto-K selector and a K-value fusion unit connected in series, and includes the following sub-steps: Using the graph embedding vector as the input to the Auto-K selector, the signal passes through two layers of perceptrons and a temperature-sensitive layer. The Softmax function outputs the probability distribution of the cluster deployment scale; During the training phase, the probability distribution of the scramble size is sampled by category to obtain the scramble size for this round. During the inference phase, a greedy strategy is adopted to take the maximum probability as the optimal scramble size. The K-value fusion processor concatenates the normalized value of the cluster deployment scale with the graph embedding vector, and then performs a linear transformation to obtain a cluster scale-aware global context vector.

6. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, The generated joint inspection sequence for each UAV includes: Based on the optimal number of drones to be deployed, a corresponding number of drones are selected from the available drone nests to participate in the collaborative inspection; during the training phase, random sampling is performed from all nest combinations; during the inference phase, all combinations are traversed, and each combination is sampled and decoded multiple times, and the combination with the lowest overall inspection cost is selected as the final deployment plan. A multi-drone polling autoregressive decoding mode is adopted, and each UAV is alternately scheduled. In each decoding step, the current UAV's position is embedded and fused with the global context vector to generate a query vector. The attention score of each candidate node is calculated by combining the node embedding matrix with the scaling dot product attention mechanism. Construct a dynamic topology mask function based on whether each candidate node in the global node set is a valid candidate action; After masking, the attention score of the legal action is normalized to the action selection probability. During the training phase, the probability distribution is randomly sampled. During the inference phase, greedy deterministic selection is adopted. The global unfinished task set and UAV position are updated according to the selected action. The process is iterated until the global task is cleared and all UAVs return to their nests, and a joint inspection sequence is output.

7. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, The calculation of comprehensive inspection costs includes: For the total time spent by each UAV performing its joint inspection sequence, the following calculations are made: the three-dimensional flight time based on dynamic obstacle avoidance, the sum of the fixed operation time of each task node, and the communication time for data return based on preset unloading trigger conditions. The dynamic obstacle avoidance adaptively adjusts the flight trajectory based on collision detection results of poles, building obstacles, and power transmission lines during the flight segment. If the flight path is safe, the flight time is calculated according to the normal displacement. If obstacle avoidance is triggered, the UAV climbs to a safe height above the obstacle, crosses it, and then descends to the target node. The data return adopts a combination of flight accumulation and hovering transmission. Unloading is triggered when the onboard buffer exceeds its limit, the UAV reaches a high transmission rate node, or the task ends. The communication time is calculated based on the accumulated data volume and the transmission rate at the unloading position. The longest total time among all UAVs is taken as the task completion time. Combined with the optimal number of UAVs deployed, the comprehensive inspection cost is calculated, and this comprehensive inspection cost is used as the environmental feedback signal for reinforcement learning.

8. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, The near-end policy optimization algorithm is used to update the joint parameters of the collaborative decision-making network. Specifically, after the complete joint inspection sequence is generated, the path generation reward and the cluster scale decision reward are calculated and distributed. The evaluation is conducted in segments based on whether the trajectory meets the full coverage constraint. When there is a missed detection, a negative penalty related to the missed detection rate is applied. When the trajectory is valid, a normalized reward settlement is performed based on the advantages and disadvantages of the completion time or the overall inspection cost relative to other trajectories in the batch. For the path generation network, the global context vector is used as input, the Critic network outputs a baseline value estimate, and the Monte Carlo reward is combined to calculate the advantage estimate; for the cluster-scale decision network, the reward is directly processed by batch centralization to construct the advantage estimate. Based on two advantage estimates, path generation loss and scale decision loss are constructed using the truncated agent objective form of near-end policy optimization, and a joint optimization objective is constructed by combining the value mean square error loss of the Critic network and the policy information entropy regularization term. By calculating the gradient of the joint optimization objective and performing gradient descent, the parameters of the path generation network, cluster size decision network, and Critic network are simultaneously updated.

9. The intelligent collaborative method for emergency inspection of a networked unmanned aerial vehicle (UAV) swarm according to claim 1, characterized in that, During the iterative training process, the average cost of feasible solutions in the current batch is monitored. The current network parameters are saved to the checkpoint only when full coverage is achieved and the average cost is better than the historical best record. After training, the checkpoint is loaded as the final inference model. After entering the inference and deployment phase, the collaborative decision-making network is inferred using a greedy strategy. The output is a multi-drone collaborative scheduling scheme that includes the optimal number of drones to be dispatched and a complete spatial node access sequence for each drone starting from its home nest, sequentially visiting the power transmission towers and power transmission lines it is responsible for, completing data unloading at the optimal transmission point, and returning to the nest.

10. A networked unmanned aerial vehicle (UAV) swarm emergency inspection intelligent collaborative system, characterized in that, include: The scene construction module is used to acquire the location information of transmission towers and line topology connections within the fault area, the location distribution information of available drone nests and ground base stations, and the geometric information of building obstacles to construct a three-dimensional heterogeneous inspection scene. It also pre-calculates the data transmission rate at any point in the space within the scenario and generates a three-dimensional offline radio map; The graph encoding module is used to model the three-dimensional heterogeneous inspection scene as a graph structure containing nest nodes, transmission tower nodes and transmission line segment nodes, and to perform one-time global feature extraction on the graph structure to generate a node embedding matrix and a graph embedding vector. The cluster size decision module is used to infer the optimal number of drones to be dispatched K based on the graph embedding vector, and to concatenate and fuse K with the graph embedding vector to generate a global context vector with cluster size awareness capability. The sequence generation module is used to take the global context vector as initial guiding information and the node embedding matrix as a feature library of candidate nodes, and input them into the pointer decoder based on the attention mechanism. Through multi-machine polling autoregressive decoding, the joint inspection sequence of each UAV is generated alternately. The physical assessment module is used to perform a comprehensive physical assessment of the inspection cost of the joint inspection sequence, calculate the comprehensive time including flight obstacle avoidance, inspection tasks and data transmission, take the longest comprehensive time among all UAVs as the task completion time, and calculate the comprehensive inspection cost in combination with the number of UAVs deployed. The learning and training module is used to generate path generation rewards and cluster size decision rewards based on the comprehensive inspection cost. The global context vector is input into the Critic network to estimate the baseline value; based on the path-generated reward, the cluster-size-determined reward, and the baseline value, joint parameters are updated; The optimal solution output module is used to repeatedly execute the above steps for iterative training until the preset maximum number of iterations is reached, and then outputs the current optimal collaborative inspection solution. The intelligent collaborative system for emergency inspection of connected drone swarms is used to execute the steps in the intelligent collaborative method for emergency inspection of connected drone swarms as described in any one of claims 1-9.

Images