A method for UAV logistics route planning in mountainous areas based on graph neural networks

By improving the ARMA graph neural network and combining cross-ridge event graphing with dynamic replanning of the safety backoff dependency domain, the problem of handling independent event states of cross-ridge behavior in UAV path planning in mountainous areas is solved, which improves path safety and backoff capability and enhances the efficiency of logistics path update under changing mountainous environmental conditions.

CN123113941APending Publication Date: 2026-11-17BEIJING GUOHUAN ZHILAN TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611580236.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing UAV path planning methods for mountainous areas lack independent event state processing for ridge crossing behavior. This makes it difficult for the remaining energy and safe retreat relationship of UAVs after crossing ridges to be directly involved in the recursive propagation control of graph neural networks. Furthermore, dynamic replanning lacks a safe retreat identification mechanism, which can easily lead to situations where some flight segments are still passable but the safe retreat chain fails. It is difficult to balance the forward accessibility of mountain logistics routes with the safe retreat capability after crossing ridges.

Method used

An improved ARMA graph neural network is adopted, which combines cross-ridge event mapping and safe retreat dependency domain dynamic replanning. By identifying terrain units, cross-ridge locations and safe retreat locations, a cross-ridge event map is constructed. The improved ARMA graph neural network is used to perform cross-ridge event Laplace reconstruction, reversible recursive gating and retreat anchor moving average backinjection to generate event open states and subsequent associated states, generate candidate logistics paths and generate safe retreat dependency domains along the safe retreat relationship, and perform dynamic replanning.

Benefits of technology

It improves the path safety and retreat capability of mountain route planning, enhances the comprehensive characterization of continuous ridge crossing risks and residual energy, reduces global redundant calculations, and improves the efficiency and safety of logistics route updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113941A_ABST
    Figure CN123113941A_ABST
Patent Text Reader

Abstract

This invention discloses a method for UAV logistics path planning in mountainous areas based on graph neural networks, belonging to the field of UAV intelligent path planning technology. The method includes: collecting and preprocessing comprehensive logistics planning data to obtain standard path planning data; identifying terrain units, ridge crossing locations, and safe retreat locations to construct a ridge crossing event graph; determining the minimum retreat energy and retreat anchor state; inputting the ridge crossing event graph and retreat state into an improved ARMA graph neural network to generate event open states and subsequent associated states; generating a ridge crossing event sequence and connecting flight segments within terrain units to obtain candidate logistics paths; validating candidate logistics paths and generating a logistics path; executing the logistics path and determining changing ridge crossing events to generate a safe retreat dependency domain; retaining out-of-domain recursive states, updating the in-domain ridge crossing event graph and retreat anchor state, and re-inferring and generating an updated logistics path. This invention improves path adaptability and dynamic replanning efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent path planning technology for unmanned aerial vehicles (UAVs), and in particular to a method for planning logistics paths for UAVs in mountainous areas based on graph neural networks. Background Technology

[0002] With the increasing demand for drone logistics transportation in mountainous areas, drone path planning is gradually being applied to mountain delivery, material transportation, and emergency resupply scenarios. Existing methods mainly collect information on digital elevation, obstacles, wind fields, communication coverage, drone energy status, and logistics tasks. Flight paths are generated through grid maps, topology maps, heuristic search, or graph neural networks, and path selection is performed by combining energy, payload, and task time constraints.

[0003] Existing UAV path planning in mountainous areas mostly uses waypoints and flight segments as the basic planning objects, and evaluates the flight status before and after ridge crossing using distance, risk, or energy consumption weights. There is a lack of methods to establish independent event states for ridge-crossing behavior. After a UAV crosses a ridge, the relationship between the remaining energy and the energy required for safe retreat is difficult to directly participate in the recursive propagation control of the graph neural network, and the safe retreat history formed by continuous ridge crossings is also difficult to continuously retain within the model.

[0004] When drones execute logistics routes, wind fields, obstacle conditions, no-fly zones, and remaining energy will change. Existing dynamic replanning mostly updates local routes based on spatial proximity and lacks a mechanism to identify affected areas according to safety retreat relationships. This can easily lead to situations where some flight segments are still passable but the safety retreat chain has failed, making it difficult to balance the forward accessibility of logistics routes in mountainous areas with the safety retreat capability after crossing ridges.

[0005] Therefore, how to provide a method for planning logistics routes for drones in mountainous areas based on graph neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for planning logistics routes for unmanned aerial vehicles (UAVs) in mountainous areas based on graph neural networks. This invention employs an improved ARMA graph neural network, combining cross-ridge event graph construction with dynamic replanning of the safety backoff dependency domain to achieve the generation and updating of logistics routes in mountainous areas. It has the advantages of high route safety, strong backoff capability, and high dynamic planning efficiency.

[0007] A method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks, according to an embodiment of the present invention, includes: Collect comprehensive logistics planning data, perform preprocessing on the comprehensive logistics planning data, and obtain standard route planning data; Based on standard path planning data, terrain units, ridge crossing locations, and safe retreat locations are identified, and a ridge crossing event map is constructed. Determine the minimum retreat energy and retreat anchorage state based on the cross-ridge event map; The cross-ridge event graph, minimum backoff energy, and backoff anchor state are input into an improved ARMA graph neural network to perform cross-ridge event Laplace reconstruction, reversible recursive gating, and backoff anchor moving average backinjection. The recursive state is updated based on the expected remaining energy and minimum backoff energy to generate event open state and subsequent associated state. Based on the open state and subsequent associated state of the event, a cross-ridge event sequence is generated, connecting the segments within the terrain unit to generate candidate logistics routes; Based on drone status data and logistics task data, candidate logistics routes are validated and logistics routes are generated. Execute the logistics path and update the running status, determine the cross-ridge events of the status change, and generate the safety backoff dependency domain along the safety backoff relationship; Preserve the recursive state outside the safe backoff dependency domain, reconstruct the cross-ridge event graph and backoff anchor state within the safe backoff dependency domain, re-execute the improved ARMA graph neural network inference, and generate an updated logistics path.

[0008] Optionally, the comprehensive logistics planning data includes mountainous environment data, drone status data, and logistics task data.

[0009] Optionally, obtaining the standard route planning data includes: Perform coordinate unification, elevation benchmark unification, and anomaly data processing on mountain environmental data to generate standard mountain environmental data; The drone status data is aligned with standard mountainous environment data according to the time of collection, and the current location of the drone is mapped to the corresponding spatial location to generate standard drone status data. The logistics origin, logistics destination and delivery node in the logistics task data are mapped to standard mountainous environmental data to generate standard logistics task data. Standard route planning data is obtained by associating standard mountainous environmental data, standard drone status data, and standard logistics task data with spatial location.

[0010] Optionally, constructing the cross-ridge event graph includes: Digital elevation data, obstacle distribution data, and no-fly zone data are extracted from mountainous environmental data. Continuous elevation ridges are extracted and combined with obstacle distribution and no-fly zone data to divide terrain units. Extract flight altitude constraints from logistics task data, filter continuous elevation ridge positions between adjacent terrain units that meet flight altitude constraints and do not fall into obstacle distribution areas or no-fly zones, and determine the ridge crossing positions; The logistics start point, logistics end point, and delivery node are extracted from the logistics task data. Obstacle distribution data, communication coverage data, and no-fly zone data are extracted from the mountain environment data. The logistics start point, logistics end point, and delivery node that have communication coverage and have not fallen into the obstacle distribution area or no-fly zone are identified as safe retreat locations. The terrain units are mapped to terrain unit nodes. Cross-ridge event nodes are established for the two travel directions of adjacent terrain units, and directed connections are established according to the cross-ridge positions to construct a cross-ridge event graph.

[0011] Optionally, determining the minimum backoff energy and backoff anchor state based on the cross-ridge event map includes: Elevation and wind field data are extracted from mountainous environmental data, and the current payload is extracted from UAV status data. The flight distance, elevation change and wind direction effect are calculated based on the entry position, crossing position and exit position of the ridge crossing event. The flight energy of the ridge crossing event is calculated in combination with the current payload. Search for the retreat path from the exit position of the cross-ridge event to the safe retreat position, calculate the retreat flight energy based on flight distance, elevation change, wind field data and current load, determine the minimum value as the minimum retreat energy, and record the corresponding retreat path as the safe retreat relationship; The safe retreat position corresponding to the minimum retreat energy is determined as the retreat anchor point. The retreat anchor point, the minimum retreat energy, and the cross-ridge event sequence are associated to generate the retreat anchor track state.

[0012] Optionally, the event open state and subsequent associated state include: Based on the graph propagation structure, recursive propagation branch, input jump branch and recursive stack fusion structure of ARMA graph neural network, an improved ARMA graph neural network is constructed by setting up a cross-ridge event Laplace reconstruction layer, a cross-ridge reversible recursive gating layer, a backtracking anchor jump injection layer, an ARMA recursive fusion layer and a cross-ridge event sequence readout layer. The cross-ridge event graph is input into the cross-ridge event Laplace reconstruction layer. Based on the terrain unit nodes, cross-ridge event nodes and directed connectivity, an adjacency matrix and a node degree matrix are constructed. The node degree is symmetrically normalized on the adjacency matrix to generate the cross-ridge event propagation matrix. The cross-ridge event Laplace matrix is ​​generated by subtracting the cross-ridge event propagation matrix from the identity matrix. The remaining energy is extracted from the UAV state data. The flight energy of the cross-ridge event is subtracted from the remaining energy to obtain the expected remaining energy. The minimum backoff energy is subtracted from the expected remaining energy to obtain the cross-ridge reversible margin. When the cross-ridge reversible margin is greater than 0, the ratio of the cross-ridge reversible margin to the expected remaining energy is used as the recursive maintenance coefficient. When the cross-ridge reversible margin is not greater than 0, the recursive maintenance coefficient is set to 0, and the ARMA recursive propagation state corresponding to the cross-ridge event propagation matrix is ​​adjusted with the recursive maintenance coefficient. Input the retreat anchor state into the retreat anchor jump injection layer, perform feature mapping on the retreat anchor point, minimum retreat energy and cross-ridge event sequence to generate retreat anchor injection features, concatenate the retreat anchor injection features with the node input features, and inject them in each round of recursive propagation through ARMA input jump branches; In the ARMA recursive fusion layer, the node states in each parallel ARMA recursive stack are updated based on the cross-ridge event propagation matrix, recursive preservation coefficient, and node input features after back-injection. The final node states of each parallel ARMA recursive stack are fused to generate a cross-ridge event fusion representation. The cross-ridge event fusion representation and recursive preservation coefficient are input into the cross-ridge event sequence readout layer. Cross-ridge events with a recursive preservation coefficient of 0 are marked as closed, and cross-ridge events with a recursive preservation coefficient greater than 0 are marked as open. The open cross-ridge event fusion representation is concatenated with the fusion representation of directed adjacent open cross-ridge events. After mapping and normalization, the subsequent association value is generated to obtain the event open state and the subsequent association state.

[0013] Optionally, generating candidate logistics routes includes: Based on the open state of the event, search for open ridge-crossing event sequences that can connect to the logistics destination and do not repeatedly pass through ridge-crossing events from the terrain unit node where the logistics origin is located. The successor correlation values ​​of each open cross-ridge event sequence are summed, and the sequence with the largest cumulative value is determined as the cross-ridge event sequence; Based on the exit and entry positions of adjacent ridge crossing events, a route segment within the terrain unit is generated, connecting the logistics starting point, the route segment within the terrain unit, the ridge crossing position, and the logistics endpoint to generate candidate logistics routes.

[0014] Optionally, generating the logistics route includes: Extract remaining energy, maximum allowable payload, and flight speed from drone status data; extract mission payload and mission time limit from logistics mission data; and extract communication coverage status from mountainous environment data. Calculate the estimated flight energy and estimated flight duration of candidate logistics routes, and detect the continuity of communication coverage; Candidate logistics routes that meet the requirements of energy, payload, mission time limit, and communication continuity are identified as logistics routes.

[0015] Optionally, generating the safe backoff dependency domain along the safe backoff relation includes: During the execution of the logistics route, update the mountain environment data and drone status data, map the updated mountain environment data and drone status data to the ridge crossing event map, extract obstacle distribution status, no-fly status and wind field status from the updated mountain environment data, extract remaining energy from the updated drone status data, and determine the ridge crossing event of status change. Read the safety backoff relationship before the removal of the state change cross-ridge event, determine the terrain unit node to be judged through the state change cross-ridge event, and temporarily remove the state change cross-ridge event from the cross-ridge event map; Wind field status is extracted from the updated mountainous environment data, and remaining energy is extracted from the updated UAV status data. The expected flight energy is calculated based on the flight distance, elevation change and wind field status of the logistics path segment from the current location to each terrain unit node to be determined. The expected remaining energy of the terrain unit is obtained by subtracting the expected flight energy from the remaining energy. After removing state-change cross-ridge events, re-search the retreat path from the undetermined terrain unit node to the safe retreat position in the cross-ridge event graph and calculate the minimum retreat energy. Determine the undetermined terrain unit node that has no retreat path or whose minimum retreat energy is greater than the expected remaining energy of the terrain unit as a safe retreat dependency node. Collect the safe retreat dependency nodes, connected cross-ridge event nodes and directed connections to generate a safe retreat dependency domain.

[0016] Optionally, generating the updated logistics route includes: Based on the safety retreat dependency domain, the cross-ridge event graph is divided, and the external terrain unit nodes, cross-ridge event nodes, directed connection relationships, ARMA recursive states and retreat anchor states are retained. The ARMA recursive states of the domain boundary nodes are determined as the initial boundary states. Based on the updated mountainous environment data and UAV status data, delete the cross-ridge event nodes that have lost access conditions within the safe retreat dependency domain, update the directed connection relationships within the domain, and generate an updated cross-ridge event subgraph. Based on the safe retreat position, a new safe retreat path is searched in the updated cross-ridge event subgraph, the cross-ridge event flight energy and minimum retreat energy are recalculated, and an updated retreat anchor state is generated based on the new safe retreat relationship. The updated cross-ridge event subgraph, cross-ridge event flight energy, minimum backoff energy, updated backoff anchor state, and boundary initial state are input into the improved ARMA graph neural network. The intradomain recursive propagation is initiated with the boundary initial state to regenerate the event open state and subsequent associated state. Based on the regenerated event open state and subsequent associated state, an alternative cross-ridge event sequence and corresponding terrain unit intra-segment are generated, which are then connected with the logistics path retained outside the domain to generate an updated logistics path.

[0017] The beneficial effects of this invention are: This invention generates standard path planning data by collecting mountain environment data, UAV status data, and logistics task data. It identifies terrain units, ridge crossing locations, and safe retreat locations, and constructs a ridge crossing event diagram of the ridge crossing process. This allows path planning to no longer rely solely on ordinary waypoints and segment connections, but to express the flight energy, retreat path, and safety status of the UAV before and after crossing the ridge, thereby improving the completeness and relevance of path status expression under complex mountain terrain conditions.

[0018] This invention constructs an improved ARMA graph neural network, incorporates the cross-ridge event propagation structure into the graph propagation process, controls the ARMA recursive propagation state based on the expected remaining energy and minimum backoff energy, and retains the safe backoff history formed during continuous ridge crossings through backoff anchor jump injection. This suppresses path states that are locally passable but lack safe backoff capabilities after crossing, thereby improving the comprehensive characterization ability of logistics path selection on continuous ridge crossing risks, remaining energy, and safe backoff capabilities.

[0019] This invention identifies ridge-crossing events with changing states, determines the retreat reachability changes of terrain units after the failure of a ridge-crossing event based on safe retreat relationships, constructs a safe retreat dependency domain, and retains the ARMA recursive state outside the safe retreat dependency domain. Updates are performed only on the ridge-crossing event graph, retreat anchor state, and path sequence within the dependency domain, so that the dynamic replanning range is determined by the safe retreat topological dependency relationship, reducing global redundant calculations and improving the efficiency and safety of logistics path updates under changing mountainous environments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a mountainous drone logistics route planning method based on graph neural networks proposed in this invention. Figure 2 This is a schematic diagram of the improved ARMA graph neural network structure of the mountain drone logistics route planning method proposed in this invention; Figure 3 This is a schematic diagram illustrating the safe backoff dependency domain generation process of a mountainous UAV logistics path planning method based on graph neural networks proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1 , Figure 2 and Figure 3 A method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks includes: Collect comprehensive logistics planning data, perform preprocessing on the comprehensive logistics planning data, and obtain standard route planning data; Based on standard path planning data, terrain units, ridge crossing locations, and safe retreat locations are identified, and a ridge crossing event map is constructed. Determine the minimum retreat energy and retreat anchorage state based on the cross-ridge event map; The cross-ridge event graph, minimum backoff energy, and backoff anchor state are input into an improved ARMA graph neural network to perform cross-ridge event Laplace reconstruction, reversible recursive gating, and backoff anchor moving average backinjection. The recursive state is updated based on the expected remaining energy and minimum backoff energy to generate event open state and subsequent associated state. Based on the open state and subsequent associated state of the event, a cross-ridge event sequence is generated, connecting the segments within the terrain unit to generate candidate logistics routes; Based on drone status data and logistics task data, candidate logistics routes are validated and logistics routes are generated. Execute the logistics path and update the running status, determine the cross-ridge events of the status change, and generate the safety backoff dependency domain along the safety backoff relationship; Preserve the recursive state outside the safe backoff dependency domain, reconstruct the cross-ridge event graph and backoff anchor state within the safe backoff dependency domain, re-execute the improved ARMA graph neural network inference, and generate an updated logistics path.

[0023] In this embodiment, the comprehensive logistics planning data includes mountainous environment data, drone status data, and logistics task data.

[0024] In this embodiment, obtaining the standard route planning data includes: Perform coordinate unification, elevation benchmark unification, and anomaly data processing on mountain environmental data to generate standard mountain environmental data; The drone status data is aligned with standard mountainous environment data according to the time of collection, and the current location of the drone is mapped to the corresponding spatial location to generate standard drone status data. The logistics origin, logistics destination and delivery node in the logistics task data are mapped to standard mountainous environmental data to generate standard logistics task data. Standard route planning data is obtained by associating standard mountainous environmental data, standard drone status data, and standard logistics task data with spatial location.

[0025] In this embodiment, constructing the cross-ridge event graph includes: Digital elevation data, obstacle distribution data, and no-fly zone data are extracted from mountainous environmental data. Continuous elevation ridges are extracted and combined with obstacle distribution and no-fly zone data to divide terrain units. Extract flight altitude constraints from logistics task data, filter continuous elevation ridge positions between adjacent terrain units that meet flight altitude constraints and do not fall into obstacle distribution areas or no-fly zones, and determine the ridge crossing positions; The logistics start point, logistics end point, and delivery node are extracted from the logistics task data. Obstacle distribution data, communication coverage data, and no-fly zone data are extracted from the mountain environment data. The logistics start point, logistics end point, and delivery node that have communication coverage and have not fallen into the obstacle distribution area or no-fly zone are identified as safe retreat locations. Terrain units are mapped to terrain unit nodes. Cross-ridge event nodes are established for the two travel directions of adjacent terrain units, and directed connections are established according to the cross-ridge positions to construct a cross-ridge event graph, where: Mapping terrain units to terrain unit nodes is done as follows: Read the boundary coordinate sequence and digital elevation data corresponding to each terrain unit, arrange the boundary points according to the record order of the boundary coordinate sequence, calculate the mean of the abscissa and the mean of the ordinate of all boundary points respectively, and determine the position corresponding to the two means as the plane center position of the terrain unit. Map the plane center position to digital elevation data, read the corresponding elevation value, and combine the mean of the horizontal coordinate, the mean of the vertical coordinate, and the corresponding elevation value to generate the center coordinates of the terrain unit; All terrain units are arranged in ascending order of their center coordinates (x-coordinate). If the x-coordinates are the same, they are arranged in ascending order of their ordinates (y-coordinate). If the ordinates are the same, they are arranged in ascending order of their elevation values. Terrain unit node identifiers are generated based on the arrangement sequence number. Obstacle distribution data, communication coverage data, and no-fly zone data are extracted from mountainous environmental data. The terrain unit node identifier, center coordinates, boundary coordinate sequence, internal elevation data of the terrain unit, obstacle distribution status, communication coverage status, and no-fly status are written into the same node record to generate a terrain unit node. For each of the two travel directions of adjacent terrain units, cross-ridge event nodes are established, specifically as follows: Read the continuous elevation ridge line and corresponding ridge crossing positions shared by two adjacent terrain units, sort all ridge crossing positions on the same shared continuous elevation ridge line in ascending order according to the x-coordinate, y-coordinate and elevation value of the ridge crossing positions, and set a position number for each ridge crossing position; Read the sampling points of two adjacent ridges before and after the continuous elevation ridge where the ridge crosses the ridge position, and form a local ridge direction vector by the difference between the abscissa and ordinate of the two sampling points. Swap the two components of the local ridge direction vector and take the opposite number of one component to generate the local ridge normal vector. Calculate the horizontal vectors pointing to the center coordinates of the two terrain units at the ridge crossing position, and perform inner product calculations on the two horizontal vectors and the local ridge line normal vectors respectively. Determine the direction of the local ridge line normal vectors pointing to the two terrain units based on the sign of the inner product, and obtain the terrain units corresponding to the two sides of the ridge crossing position. A cross-ridge event node is established using the first terrain unit as the starting terrain unit and the second terrain unit as the target terrain unit. Another cross-ridge event node is established using the second terrain unit as the starting terrain unit and the first terrain unit as the target terrain unit. Move grid by grid from the ridge position along the direction of the starting terrain unit. When you encounter the center of the first grid that belongs to the starting terrain unit and has not fallen into the obstacle distribution area or the no-fly zone, determine the entry position. Use the same method to determine the exit position along the direction of the target terrain unit. A ridge-crossing event node identifier is generated by combining the starting terrain unit node identifier, the target terrain unit node identifier, and the ridge-crossing position sequence number. The entry position, ridge-crossing position, exit position, travel direction, starting terrain unit node identifier, and target terrain unit node identifier are written into the ridge-crossing event node record. A directed connection relationship is established based on the cross-ridge location, and a cross-ridge event graph is constructed, specifically as follows: For each ridge-crossing event node, read the identifier of the starting terrain unit node and the identifier of the target terrain unit node, establish an entry directed edge from the starting terrain unit node to the ridge-crossing event node, and establish an exit directed edge from the ridge-crossing event node to the target terrain unit node. The terrain unit nodes are arranged in ascending order according to their identifiers, and the ridge-crossing event nodes are arranged in ascending order according to their identifiers. The terrain unit nodes are placed before the ridge-crossing event nodes to form a unified node sequence. Assign matrix row and column indices to each node based on the unified node sequence. When there is an incoming or outgoing directed edge, write 1 in the corresponding starting row and ending column. When there is no directed connection, write 0 to generate a cross-ridge event adjacency matrix. A cross-ridge event graph is constructed by associating and storing the unified node sequence, the cross-ridge event adjacency matrix, the terrain unit node record, and the cross-ridge event node record.

[0026] In this embodiment, determining the minimum retreat energy and retreat anchor state based on the cross-ridge event graph includes: Elevation and wind field data are extracted from mountainous environmental data, and the current payload is extracted from UAV status data. The flight distance, elevation change and wind direction effect are calculated based on the entry position, crossing position and exit position of the ridge crossing event. The flight energy of the ridge crossing event is calculated in combination with the current payload. The retreat path is searched from the exit position of the cross-ridge event to the safe retreat position. The retreat flight energy is calculated based on flight distance, elevation change, wind field data, and current load. The minimum value is determined as the minimum retreat energy, and the corresponding retreat path is recorded as a safe retreat relationship. Where: The retreat path from the exit position of the cross-ridge event to the safe retreat position is as follows: Read the exit position and target terrain unit node identifier from the cross-ridge event node record, extract digital elevation data, obstacle distribution data and no-fly zone data from the mountain environment data, and take the digital elevation raster center within the target terrain unit and the connected terrain unit range as candidate retreat nodes. Eliminate candidate retreat nodes that fall into obstacle distribution areas and no-fly zones, and sort the remaining candidate retreat nodes in ascending order of horizontal coordinate, vertical coordinate, and elevation value, and assign retreat node identifiers; Within the same terrain unit, establish backoff connections for horizontally adjacent, vertically adjacent, and diagonally adjacent candidate backoff nodes. Check the digital elevation grid through which the line connecting the centers of two candidate backoff nodes passes. Delete the corresponding backoff connection when the line passes through an obstacle distribution area or a no-fly zone. Connect the exit position of the cross-ridge event to the candidate retreat node with the smallest spatial distance, and connect each safe retreat position to the candidate retreat node with the smallest spatial distance within the corresponding terrain unit. Different terrain units are connected only through the cross-ridge event node to generate a retreat search map. Using the backoff node corresponding to the exit position of the cross-ridge event as the search starting point, the cumulative backoff energy of the search starting point is initialized to 0, and the remaining backoff nodes are set to unvisited state. In each round, the unvisited node with the smallest cumulative backoff energy is selected as the current node. Read all passable connections of the current node, calculate the cumulative backoff energy of the current node after passing through each passable connection to reach the adjacent node, and update the cumulative backoff energy of the adjacent node and record the current node as the predecessor node when the new cumulative backoff energy is less than the cumulative backoff energy already recorded by the adjacent node. Repeatedly execute node selection and cumulative backoff energy update until all nodes corresponding to reachable safe backoff positions have been visited. Backtrack from each safe backoff position to the search starting point in reverse order of the predecessor node, and then reverse the order of the results to generate backoff paths from the ridge event exit position to each safe backoff position. The retreat flight energy is calculated based on flight distance, elevation change, wind field data, and current load, specifically as follows: The three-dimensional coordinates and wind field data corresponding to each retreat node of the retreat path are extracted from the mountainous environment data. The current payload, UAV empty mass, cruise airspeed, reference level flight power and propulsion efficiency are extracted from the UAV status data. The UAV empty mass is added to the current payload to obtain the current flight mass. Read adjacent retreat nodes sequentially according to the order in which they pass through the retreat path nodes, calculate the difference in horizontal coordinate, vertical coordinate and elevation, square the three differences, sum them and take the square root to obtain the three-dimensional flight distance between adjacent retreat nodes; The difference between the horizontal and vertical coordinates is used to form a horizontal heading vector. The horizontal heading vector is divided by the horizontal heading vector magnitude to obtain the horizontal heading unit vector. Read the horizontal components of the wind field corresponding to two adjacent retreat nodes and calculate their averages to obtain the horizontal wind field vector of the flight segment. Perform an inner product calculation between the horizontal wind field vector of the flight segment and the horizontal heading unit vector to obtain the wind speed component along the heading. The cruising airspeed is added to the wind speed component along the heading to obtain the segment ground speed. When the segment ground speed is not greater than 0, the corresponding retreat connection is marked as impassable. When the segment ground speed is greater than 0, the three-dimensional flight distance is divided by the segment ground speed to obtain the segment flight time. Extract the baseline flight mass from the UAV status data, divide the current flight mass by the baseline flight mass to obtain the load correction ratio, multiply the baseline level flight power by the load correction ratio to obtain the load-corrected level flight power, and multiply the load-corrected level flight power by the segment flight time to obtain the segment basic flight energy. When the elevation difference is greater than 0, the current flight mass, elevation difference and gravitational acceleration 9.81 are multiplied in sequence and then divided by the propulsion efficiency to obtain the additional energy for the climb segment. When the elevation difference is not greater than 0, the additional energy for the climb segment is set to 0. The basic flight energy of the flight segment and the additional energy for climb of the flight segment are added together to obtain the flight segment retreat energy. The flight segment retreat energy is accumulated in the order of the retreat path nodes. When the flight segment passes through the cross-ridge event node, the corresponding cross-ridge event flight energy is added to obtain the retreat flight energy corresponding to the retreat path. The minimum value is determined as the minimum backoff energy, and the corresponding backoff path is recorded as a safe backoff relationship, specifically: Collect all retreat paths and corresponding retreat flight energies from the exit position of the same cross-ridge event to each safe retreat position, and delete retreat paths containing impassable retreat connections; The remaining retreat paths are arranged in ascending order of retreat flight energy, and the retreat flight energy at the top of the list is determined as the minimum retreat energy. When there are multiple retreat paths with the same retreat flight energy, the three-dimensional flight distances within each retreat path are accumulated, and the paths are sorted in ascending order of accumulated three-dimensional flight distances. If the accumulated three-dimensional flight distances are still the same, the paths are sorted in ascending order of terrain unit node identifiers corresponding to the safe retreat position, and the retreat path at the top of the sort is selected. Record the retreat node identifiers, terrain unit node identifiers, ridge crossing event node identifiers, and safe retreat positions sequentially along the selected retreat path according to the actual retreat direction, and associate them with the minimum retreat energy to form a safe retreat relationship; When there is no safe retreat path to the safe retreat position, the corresponding ridge crossing event is recorded as having no safe retreat relationship, no minimum retreat energy value is generated, and 0 is written to the retreat reachability state field. The safe backoff position corresponding to the minimum backoff energy is determined as the backoff anchor point. The backoff anchor state is generated by associating the cross-ridge event node identifier, the backoff anchor point, the minimum backoff energy, and the cross-ridge event sequence. Specifically, the safe backoff position corresponding to the minimum backoff energy is determined as follows: Read the safe backoff position and minimum backoff energy in the safe backoff relationship, and determine the safe backoff position corresponding to the minimum backoff energy as the backoff anchor point; Read the spatial coordinates, elevation values, and topographic unit node identifiers of the setback anchor points from the standard path planning data, and write the spatial coordinates, elevation values, and topographic unit node identifiers into the setback anchor point record; When there are multiple safe retreat positions with the same minimum retreat energy and cumulative three-dimensional flight distance for the same cross-ridge event, they are arranged in ascending order of horizontal coordinate, vertical coordinate, and elevation value, and the first safe retreat position in the arrangement is selected as the retreat anchor point. In this embodiment, the event open state and subsequent associated state include: Based on the graph propagation structure, recursive propagation branch, input jump branch, and recursive stack fusion structure of the ARMA graph neural network, an improved ARMA graph neural network is constructed by setting up a cross-ridge event Laplace reconstruction layer, a cross-ridge reversible recursive gating layer, a backtracking anchor jump injection layer, an ARMA recursive fusion layer, and a cross-ridge event sequence readout layer. Specifically, the construction of the improved ARMA graph neural network involves: An improved ARMA graph neural network is designed with five functional layers: a cross-ridge event Laplace reconstruction layer, a cross-ridge reversible recursive gating layer, a backtracking anchor jump injection layer, an ARMA recursive fusion layer, and a cross-ridge event sequence readout layer. These five functional layers are connected in the order of cross-ridge event graph structure reconstruction, recursive propagation constraints, backtracking history injection, recursive state fusion, and event sequence readout. The ARMA recursive fusion layer contains four parallel ARMA recursive stacks. Each recursive stack performs four rounds of recursive propagation. The four rounds of recursive propagation within the same recursive stack share recursive transformation parameters, while different recursive stacks have independent recursive transformation parameters. The four recursive stacks share node input mapping parameters and input jump mapping parameters. Let N be the total number of terrain unit nodes and cross-ridge event nodes in the cross-ridge event graph corresponding to a single logistics planning sample, M be the number of cross-ridge event nodes, and E be the number of directed connections. All nodes are arranged into a unified node sequence according to the ascending order of terrain unit node identifiers and cross-ridge event node identifiers. The start index and end index of each directed connection are arranged into two rows to form a 2×E directed connection index. The node attributes are arranged into an N×32 node input feature matrix according to the unified node sequence. The cross-ridge event flight energy, cross-ridge event arrival flight energy, minimum retreat energy, and the current remaining energy of the UAV are used as the state input of the cross-ridge reversible recursive gating layer. The retreat anchor state is input into the retreat anchor jump injection layer according to the ascending order of cross-ridge event node identifiers. The continuous fields in the N×32 node input feature matrix are derived from the spatial coordinates, elevation status, wind field status, and communication coverage status in the standard path planning data. The continuous fields are normalized to 0 to 1 according to the minimum and maximum values ​​of the corresponding fields in the training set. The discrete fields are derived from the node type, obstacle status, no-fly status, and cross-ridge travel direction. The discrete fields are encoded using fixed one-hot encoding. Each field is connected in a fixed order of node type, spatial coordinates, elevation status, wind field status, obstacle status, communication coverage status, no-fly status, and travel direction. When the encoding dimension is less than 32 dimensions, zeros are added at the end. When the encoding dimension reaches 32 dimensions, no more fields are added, so that no additional trainable embedding parameters are introduced into the node input encoding process. The retreat anchor jump injection layer receives the 36-dimensional original retreat anchor features corresponding to each ridge-crossing event. The 36-dimensional original retreat anchor features consist of 3-dimensional relative coordinates of retreat anchor points, 1-dimensional minimum retreat energy state, and 32-dimensional ridge-crossing event sequence features. After 36-dimensional to 32-dimensional feature mapping, 32-dimensional retreat anchor injection features are obtained. The corresponding positions of terrain unit nodes are filled with 32-dimensional all-zero vectors, and the corresponding positions of ridge-crossing event nodes are filled with the corresponding retreat anchor injection features, thus forming an N×32 retreat anchor injection feature matrix that corresponds row by row to the N×32 node input feature matrix. The node input feature dimension is set to 32, and the ARMA recursive hidden state dimension is set to 64. The 32-dimensional node input feature is converted into a 64-dimensional initial hidden feature through a shared input mapping. The 32-dimensional node input feature is connected with the 32-dimensional backtracking anchor injection feature to form a 64-dimensional jump input. The 64-dimensional jump input is entered into each parallel ARMA recursive stack through a shared jump mapping. After the ARMA recursive fusion layer completes the final round state fusion of the four recursive stacks, it forms an N×64 node fusion representation. From this, an M×64 cross-ridge event fusion representation is extracted according to the cross-ridge event node index. The cross-ridge event sequence readout layer reads the 64-dimensional fusion representation of the current cross-ridge event, the 64-dimensional fusion representation of the candidate successor cross-ridge event, the 1-dimensional recursive preservation coefficient of the current cross-ridge event, and the 1-dimensional recursive preservation coefficient of the candidate successor cross-ridge event for each cross-ridge event pair with a directed successor relationship. The fusion representation is then concatenated in the order of the current event representation, the candidate successor event representation, the current recursive preservation coefficient, and the candidate successor recursive preservation coefficient to form a 130-dimensional event pair feature. The 130-dimensional event pair feature is then used to generate a successor association score through a 64-dimensional hidden mapping and a 1-dimensional output mapping. The final output of the improved ARMA graph neural network includes an M×1 event open state vector and an M×M successor association state matrix. The M×1 event open state vector arranges the open and closed states of each cross-ridge event according to the cross-ridge event node identifier. The row index of the M×M successor association state matrix corresponds to the current cross-ridge event node and the column index corresponds to the candidate successor cross-ridge event node. The corresponding successor association value is stored at the position where there is a directed successor relationship, and 0 is written at the position where there is no directed successor relationship. The trainable parameters of the model consist of 2112 parameters from the 32- to 64-dimensional shared input mapping, 1184 parameters from the 36- to 32-dimensional backtracking anchor mapping, 4160 parameters from the 64- to 64-dimensional shared jump mapping, 16640 parameters from four independent 64- to 64-dimensional recursive transformations, and 8449 parameters from the 130- to 64-dimensional and then to 1-dimensional cross-ridge event sequence readout structure. The cross-ridge event Laplace reconstruction layer and the cross-ridge reversible recursive gating layer use graph structures and energy states to perform deterministic calculations without setting trainable parameters. According to the above parameter sharing method, the total number of trainable parameters of the model is 32545. The training data sources are limited to historical mission records of mountainous environment data, UAV status data, and logistics mission data. The mountainous environment data is used to read the elevation, wind field, obstacle distribution, communication coverage, and no-fly status during the corresponding historical mission execution period. The UAV status data is used to read the time-stamped position trajectory, remaining energy, and payload status. The logistics mission data is used to read the logistics start point, logistics end point, delivery node, mission payload, and mission completion record. The three types of data are associated according to the logistics mission identifier and timestamp. The historical UAV position trajectory is mapped to the corresponding ridge crossing position in sequence. The real ridge crossing event sequence is generated according to the actual time sequence of passing the ridge crossing position. Training supervision labels are established item by item according to the real cross-ridge event sequence. The next cross-ridge event that actually occurs after the current cross-ridge event in the sequence is taken as the real successor event. Cross-ridge events with directed connection conditions but not actually selected are taken as candidate successor events. This makes the training labels directly correspond to the successor association state output by the cross-ridge event sequence readout layer. The event open state is determined by the cross-ridge reversible recursive gating layer based on energy constraints without setting additional manual classification labels. Using complete logistics tasks as indivisible sample units, historical tasks are divided into training, validation, and test sets in an 8:1:1 ratio. All node states, cross-ridge events, and event transition records corresponding to the same logistics task are only included in one data set to avoid data leakage caused by adjacent states in the same flight trajectory entering different data sets at the same time. The model is trained using the Adam optimization method, with a batch size of 32, an initial learning rate of 0.001, a weight decay coefficient of 0.0001, and a maximum number of training epochs of 200. For each event transition in the real cross-ridge event sequence, the normalized successor correlation value corresponding to the real successor event is read and the negative logarithm is calculated. The path sequence loss is obtained by summing the negative logarithms of all event transitions on the same path and dividing by the actual number of event transitions. The average path sequence loss of all path sequences in the training batch is then calculated and backpropagation is performed. Training is stopped when the path sequence loss on the validation set does not decrease after 12 consecutive epochs, and the model parameters corresponding to the minimum path sequence loss on the validation set are saved. Compared with existing ARMA graph neural networks, the improved ARMA graph neural network retains the basic structure of parallel ARMA recursive stack, recursive state propagation, input jump branch, and stack fusion. It transforms the general graph propagation structure into a cross-ridge event propagation structure for terrain unit nodes and directional cross-ridge event nodes. It directly applies the reversible constraint formed by the available energy of the UAV after reaching the cross-ridge event and the energy required for safe retreat to the ARMA recursive state maintenance process. It writes the retreat anchor information formed by continuous cross-ridges into the ARMA input jump branch and transforms the general node representation readout into event open state and subsequent associated state. This allows the same terrain location to form different recursive propagation results under different cross-ridge histories, different remaining energies, and different safe retreat conditions. It suppresses the event propagation that lacks safe retreat capability after completing the cross-ridge from within the model and provides a graph representation with safe retreat constraints for the generation of cross-ridge event sequences. The cross-ridge event graph is input into the cross-ridge event Laplacian reconstruction layer. An adjacency matrix and a node degree matrix are constructed based on terrain unit nodes, cross-ridge event nodes, and directed connectivity. Node degree symmetric normalization is performed on the adjacency matrix to generate the cross-ridge event propagation matrix. Finally, the cross-ridge event Laplacian matrix is ​​generated by subtracting the cross-ridge event propagation matrix from the identity matrix. Where: The adjacency matrix and node degree matrix are constructed based on terrain unit nodes, cross-ridge event nodes, and directed connectivity relationships, as follows: Read all terrain unit nodes and cross-ridge event nodes in the cross-ridge event graph, sort the terrain unit nodes in ascending order according to their identifiers, and sort the cross-ridge event nodes in ascending order according to their identifiers. Connect the two types of nodes in sequence to form a unified node sequence, and assign matrix indices starting from 0 according to their arrangement positions. Count the number of nodes in a unified node sequence, construct an adjacency matrix with the number of rows and columns equal to the number of nodes, and initialize all elements of the adjacency matrix to 0; Read the directed connections in the cross-ridge event graph one by one, use the matrix index corresponding to the starting point of the directed connection as the row index, use the matrix index corresponding to the ending point of the directed connection as the column index, write 1 in the corresponding matrix position, and keep 0 in the matrix position where no directed connection has been established. The number of outgoing connections to a node is obtained by counting the number of non-zero elements in the adjacency matrix row by row, and the number of incoming connections to a node is obtained by counting the number of non-zero elements in the adjacency matrix column by column. The node degree is obtained by adding the number of outgoing connections to the number of incoming connections. Construct a diagonal matrix with the same dimension as the adjacency matrix, write the degree of each node into the corresponding diagonal position according to a unified node sequence, and write 0 into the remaining positions to generate a node degree matrix. Perform node degree symmetric normalization on the adjacency matrix to generate the cross-ridge event propagation matrix, specifically: Read the node degree at the diagonal position of the node degree matrix one by one according to the unified node sequence, take the square root of the node degree and take the reciprocal. When the node degree is 0, set the corresponding reciprocal value to 0. Write it to the diagonal position according to the original node index to generate the inverse square root matrix of node degree. Multiply the inverse square root matrix of node degree by the adjacency matrix, and then multiply the product by the inverse square root matrix of node degree, so that the matrix element corresponding to each directed connection is converted from the original value of 1 to the reciprocal of the square root of the product of the starting node degree and the ending node degree. Keeping the original directed connection directions in the adjacency matrix unchanged, the normalized non-zero elements are used as the propagation coefficients of the corresponding directed connections. All propagation coefficients are written into the matrix according to a unified node sequence to generate a cross-ridge event propagation matrix. The Laplace matrix for cross-ridge events is generated by subtracting the cross-ridge event propagation matrix from the identity matrix, specifically as follows: Create an identity matrix with the same dimension as the cross-ridge event propagation matrix, write 1 in the main diagonal position and 0 in the non-main diagonal position; Read the identity matrix and the cross-ridge event propagation matrix element by element according to the same row and column indices, subtract the corresponding cross-ridge event propagation matrix element from the identity matrix element, write the calculation result into the same row and column position, and generate the cross-ridge event Laplace matrix. Establish an index association between the cross-ridge event propagation matrix, the cross-ridge event Laplace matrix, and the unified node sequence, so that the i-th row and i-th column of the matrix always correspond to the terrain unit node or cross-ridge event node with index i in the unified node sequence. The remaining energy is extracted from the UAV state data. The flight energy from the ridge-crossing event is subtracted from the remaining energy to obtain the expected remaining energy. The minimum backoff energy is then subtracted from the expected remaining energy to obtain the ridge-crossing reversible margin. When the ridge-crossing reversible margin is greater than 0, the ratio of the ridge-crossing reversible margin to the expected remaining energy is used as the recursive maintenance coefficient. When the ridge-crossing reversible margin is not greater than 0, the recursive maintenance coefficient is set to 0. The ARMA recursive propagation state corresponding to the ridge-crossing event propagation matrix is ​​adjusted using the recursive maintenance coefficient. Where: The estimated remaining energy is obtained by subtracting the flight energy from the ridge event from the remaining energy. Extract the current remaining energy from the UAV status data, and read the cross-ridge event flight energy corresponding to each cross-ridge event in ascending order according to the cross-ridge event node identifier; For each ridge crossing event, subtract the corresponding ridge crossing event flight energy from the current remaining energy, and write the difference into the expected remaining energy field of the corresponding ridge crossing event node; When the expected remaining energy is less than 0, the corresponding expected remaining energy will be retained as a negative value and the cross-ridge event will be marked as an energy shortage state; when the expected remaining energy is not less than 0, the actual calculated value will be retained. The cross-ridge reversible margin is obtained by subtracting the minimum retreat energy from the expected remaining energy, specifically: Read the estimated remaining energy and the corresponding minimum retreat energy according to the cross-ridge event node identifier, so that the two values ​​correspond one-to-one with the same cross-ridge event node; The difference between the expected remaining energy and the minimum retreat energy is used as the cross-ridge reversible margin. A cross-ridge reversible margin greater than 0 indicates that there is still an energy margin that meets the minimum retreat energy after the cross-ridge event is completed. A cross-ridge reversible margin not greater than 0 indicates that the minimum retreat energy is not met after the cross-ridge event is completed. For ridge crossing events that do not form a safe retreat relationship and do not have a minimum retreat energy record, the ridge crossing reversible margin state is directly recorded as an unavoidable state. The ratio of the transridge reversible margin to the expected remaining energy is used as the recursive preservation coefficient, specifically: For ridge-crossing events with a ridge-crossing reversible margin greater than 0, the ridge-crossing reversible margin is divided by the expected remaining energy to obtain the corresponding recursive maintenance coefficient. Since the minimum backoff energy is not less than 0, the recursive maintenance coefficient is kept within the range of greater than 0 and not greater than 1. For ridge crossing events with a reversible margin of no more than 0, the recursive hold coefficient is written to 0; for ridge crossing events where no safe backoff relationship is formed, the recursive hold coefficient is written to 0. Establish a recursive preservation vector according to a unified node sequence, write 1 at the corresponding position of the terrain unit node, write the corresponding recursive preservation coefficient at the corresponding position of the cross-ridge event node, and write each element of the recursive preservation vector into a diagonal matrix of the same dimension to generate a recursive preservation matrix. The ARMA recursive propagation state corresponding to the cross-ridge event propagation matrix is ​​adjusted by the recursive preservation coefficient, specifically as follows: Read the node state matrix formed by the previous round of ARMA recursive propagation. Each row of the node state matrix is ​​arranged according to a unified node sequence. Multiply the recursive preservation matrix with the node state matrix to keep the terrain unit node state in its original value and scale the cross-ridge event node state according to the corresponding recursive preservation coefficient. According to the directed propagation relationship in the cross-ridge event propagation matrix, all incoming source nodes are read for each target node. The scaled state vector of the source node is multiplied by the corresponding propagation coefficient and accumulated element by element to obtain the gated aggregation state of the target node. Cross-ridge event nodes with a recursion retention coefficient of 0 form an all-zero recursion state after gating, preventing them from continuing to pass the previous recursion state to the target terrain unit node. Cross-ridge event nodes with a recursion retention coefficient greater than 0 pass the recursion state to the target terrain unit node according to the corresponding retention ratio. The retreat anchor state is input into the retreat anchor jump injection layer. Feature mapping is performed on the retreat anchor point, minimum retreat energy, and cross-ridge event sequence to generate retreat anchor injection features. These features are concatenated with the node input features and injected back into the recursive propagation in each round via the ARMA input jump branch. Feature mapping is performed on the retreat anchor point, minimum retreat energy, and cross-ridge event sequence to generate retreat anchor back-injection features, specifically: Read the retreat anchor state corresponding to each ridge crossing event, and extract the three-dimensional coordinates of the retreat anchor point, the node identifier of the terrain unit to which the retreat anchor point belongs, the minimum retreat energy, and the ridge crossing event sequence arranged according to the retreat direction from the retreat anchor state; Subtract the three-dimensional coordinates corresponding to the current ridge-crossing event exit position from the three-dimensional coordinates of the retreat anchor point to obtain the relative coordinates of the retreat anchor point. Divide the three relative coordinates by the maximum three-dimensional distance between all terrain unit nodes to obtain three normalized retreat anchor point coordinate components. Read the minimum backoff energy corresponding to all cross-ridge events, divide the current minimum backoff energy by the maximum value among all minimum backoff energies to obtain the normalized minimum backoff energy, and write the normalized minimum backoff energy to 0 when the maximum value is 0. Read the node input features of each cross-ridge event node in the sequence according to the order of the cross-ridge event sequence, set the position weight of the first cross-ridge event to 1, set the position weight of the second cross-ridge event to half, set the position weight to the reciprocal of the corresponding number according to the increasing position in the sequence, and then divide all position weights by the sum of position weights so that the sum of all position weights is 1. The input features of each cross-ridge event node are multiplied by the corresponding normalized position weights and summed element by element to generate the retreat event sequence features. The backoff anchor point coordinate components, normalized minimum backoff energy, and backoff event sequence features are spliced ​​together in the order of normalized backoff anchor point coordinate components, normalized minimum backoff energy, and backoff event sequence features to generate the original features of the backoff anchor track. The original features of the retreat anchor are input into the linear mapping matrix. The output dimension of the linear mapping matrix is ​​set to the dimension of the node input features. The upper and lower bounds of the mapping matrix are initialized by taking the square root of 6 divided by the sum of the input and output dimensions and the corresponding negative number. The mapping bias is initialized to 0. The retreat anchor back-injection features are obtained through linear mapping. The retreat anchor back-injection features are concatenated with the node input features, specifically as follows: Establish a retreat anchor feature matrix according to a unified node sequence. Write a zero vector with the same dimension as the retreat anchor back-injection feature in the row corresponding to the terrain unit node. Write the corresponding retreat anchor back-injection feature in the row corresponding to the ridge event node. Read the node input feature matrix according to the unified node sequence, and concatenate the node input features of each node with the backoff anchor features in the same row along the feature dimension to generate the concatenated input features; The concatenated input features are input into the jump mapping matrix. The output dimension of the jump mapping matrix is ​​set to the dimension of the ARMA recursive hidden state. The jump mapping matrix adopts the same initialization method as the backtracking anchor feature mapping matrix, and the bias is initialized to 0 to generate the node input features after back-injection. The ARMA input jump branch is used for back-injection in each round of recursive propagation, specifically as follows: The node input features after back-injection are saved as fixed jump inputs shared by each parallel ARMA recursive stack, and the node input features after back-injection remain unchanged during the same round of model inference. Each round of ARMA recursive propagation is executed, and the current round of gated aggregation state is input into the recursive transformation process of the corresponding stack. The node input features after back-injection are mapped to the same hidden state dimension through the input jump branch. The recursive transformation result and the input jump branch mapping result are added element by element according to the corresponding nodes, and then the activation process is performed to obtain the current round update node state, which is used as the input node state for the next round of recursive propagation. In the ARMA recursive fusion layer, the node states in each parallel ARMA recursive stack are updated based on the cross-ridge event propagation matrix, recursion preservation coefficients, and node input features after back-injection. The final round node states of each parallel ARMA recursive stack are fused to generate a cross-ridge event fusion representation, where: In the ARMA recursive fusion layer, the node state in each parallel ARMA recursive stack is updated based on the cross-ridge event propagation matrix, recursion preservation coefficients, and node input features after back-injection. Specifically: This implementation sets up 4 parallel ARMA recursion stacks, each recursion stack performs 4 rounds of recursive propagation, the hidden state dimension of the node is set to 64, different recursion stacks use independent recursion transformation matrices, and the same recursion stack shares the corresponding recursion transformation matrix in each round. In the first round of each recursive stack, the node input features corresponding to the unified node sequence are read, the node input features are mapped to a 64-dimensional hidden state, the hidden state of the cross-ridge event node is scaled according to the recursive preservation coefficient, and the nodes are aggregated to the directed adjacent nodes according to the cross-ridge event propagation matrix. The gated recursive state obtained by aggregation is input into the recursive transformation matrix of the corresponding recursive stack. The node input features after back-injection are mapped to 64 dimensions through the input jump branch. The two results are added element by element and linear rectified activation is performed to obtain the node state of the current round. The current round node state is used as the next round recursive state. The recursive maintenance coefficient scaling, cross-ridge event propagation matrix aggregation, recursive transformation and jump back injection are repeatedly executed until 4 rounds of recursive propagation are completed, and the final round node state of the corresponding recursive stack is obtained. By fusing the final node states of each parallel ARMA recursive stack, a cross-ridge event fusion representation is generated, specifically as follows: Read the final node state of the four parallel ARMA recursive stacks according to a unified node sequence, so that the row index of the 64-dimensional state vector of the same node in the four recursive stacks remains consistent. For each node, add the elements of the same dimension of the state vectors corresponding to the four recursive stacks and divide by 4 to obtain a 64-dimensional fused state vector. Extract the fusion state vectors corresponding to cross-ridge event nodes in the unified node sequence, arrange them in ascending order according to the cross-ridge event node identifiers, and generate cross-ridge event fusion representations; The cross-ridge event fusion representation and recursive preservation coefficient are input into the cross-ridge event sequence readout layer. Cross-ridge events with a recursive preservation coefficient of 0 are marked as closed, and cross-ridge events with a recursive preservation coefficient greater than 0 are marked as open. The fusion representation of open cross-ridge events is concatenated with the fusion representation of directed adjacent open cross-ridge events. After mapping and normalization, subsequent association values ​​are generated to obtain the event open state and subsequent association state, where: Cross-ridge events with a recursive hold coefficient of 0 are marked as closed, and cross-ridge events with a recursive hold coefficient greater than 0 are marked as open. Specifically: Read the recursive retention coefficient according to the cross-ridge event node identifier, write the cross-ridge event node with the recursive retention coefficient equal to 0 into the closed state identifier, and write the cross-ridge event node with the recursive retention coefficient greater than 0 into the open state identifier. Establish a one-to-one correspondence between the cross-ridge event node identifier, the recursive preservation coefficient, and the corresponding open state identifier to form an event open record; The fusion representation of open cross-ridge events is concatenated with the fusion representation of directed adjacent open cross-ridge events, and the subsequent association value is generated through mapping and normalization. Specifically: For each open ridge crossing event, read the target terrain unit node identifier, search for ridge crossing events in the ridge crossing event graph where the starting terrain unit node identifier is consistent with the target terrain unit node identifier and is in an open state, and determine the found ridge crossing events as directed adjacent open ridge crossing events. Read the 64-dimensional fusion state vector of the current open cross-ridge event and the 64-dimensional fusion state vector of each directed adjacent open cross-ridge event, and read the recursive preservation coefficients corresponding to the two cross-ridge events. Concatenate them in the order of current event fusion state, adjacent event fusion state, current recursive preservation coefficient, and adjacent recursive preservation coefficient to generate event pair feature vectors. The event pair feature vector is input into a two-level mapping structure. The first level maps the event pair feature vector to a 64-dimensional associated feature and performs linear rectified activation. The second level maps the 64-dimensional associated feature to a single associated score. The weights of the two mapping layers are determined by dividing 6 by the sum of the input and output dimensions, taking the square root, to determine the positive and negative initialization boundaries and performing uniform initialization. The bias is initialized to 0. For all directed adjacent open cross-ridge events corresponding to the same current open cross-ridge event, take the index value of each association score, divide each index value by the sum of all candidate index values, so that the sum of all subsequent association values ​​corresponding to the same current open cross-ridge event is 1, and obtain the subsequent association value of each directed adjacent open cross-ridge event. If there are no directed adjacent open cross-ridge events, the set of successor associated values ​​corresponding to the current open cross-ridge event is recorded as an empty set; The event's open state and subsequent associated state are obtained, specifically: The event open records are arranged in ascending order according to the cross-ridge event node identifier, and the cross-ridge event node identifier, recursive preservation coefficient, and open status identifier are combined to form the event open status; Arrange all subsequent associated values ​​in the order of the current open cross-ridge event node identifier and the directed adjacent open cross-ridge event node identifiers, and form a subsequent associated record by combining the current event node identifier, the subsequent event node identifier, and the subsequent associated value. Collect all subsequent related records and establish an index association with the event open status according to the cross-ridge event node identifier to obtain the subsequent related status.

[0027] In this embodiment, generating candidate logistics routes includes: Based on the open state of the event, search for open ridge-crossing event sequences that can connect to the logistics destination and do not repeatedly pass through ridge-crossing events from the terrain unit node where the logistics origin is located. The successor correlation values ​​of each open cross-ridge event sequence are summed, and the sequence with the largest cumulative value is determined as the cross-ridge event sequence; Based on the exit and entry positions of adjacent ridge crossing events, a route segment within the terrain unit is generated, connecting the logistics starting point, the route segment within the terrain unit, the ridge crossing position, and the logistics endpoint to generate candidate logistics routes.

[0028] In this embodiment, generating the logistics path includes: Extract remaining energy, maximum allowable payload, and flight speed from drone status data; extract mission payload and mission time limit from logistics mission data; and extract communication coverage status from mountainous environment data. Calculate the estimated flight energy and estimated flight duration of candidate logistics routes, and detect the continuity of communication coverage; Candidate logistics routes that meet the requirements of energy, payload, mission time limit, and communication continuity are identified as logistics routes.

[0029] In this embodiment, generating the safety backoff dependency domain along the safety backoff relation includes: During the execution of the logistics route, update the mountain environment data and drone status data, map the updated mountain environment data and drone status data to the ridge crossing event map, extract obstacle distribution status, no-fly status and wind field status from the updated mountain environment data, extract remaining energy from the updated drone status data, and determine the ridge crossing event of status change. Read the safety backoff relationship before the removal of the state change cross-ridge event, determine the terrain unit node to be judged through the state change cross-ridge event, and temporarily remove the state change cross-ridge event from the cross-ridge event map; Wind field status is extracted from the updated mountainous environment data, and remaining energy is extracted from the updated UAV status data. The expected flight energy is calculated based on the flight distance, elevation change and wind field status of the logistics path segment from the current location to each terrain unit node to be determined. The expected remaining energy of the terrain unit is obtained by subtracting the expected flight energy from the remaining energy. After removing state-change cross-ridge events, a new retreat path is searched in the cross-ridge event graph for the terrain unit node to be determined to a safe retreat position, and the minimum retreat energy is calculated. Terrain unit nodes with no retreat path or whose minimum retreat energy is greater than the expected remaining energy of the terrain unit are identified as safe retreat dependent nodes. Safe retreat dependent nodes, connected cross-ridge event nodes, and directed connections are aggregated to generate a safe retreat dependency domain, where: After removing state-change cross-ridge events, the retreat path from the terrain unit node to the safe retreat position is re-searched in the cross-ridge event graph, and the minimum retreat energy is calculated. Specifically: Read the ridge event graph after temporarily removing the ridge event that changed the state. Sort the terrain unit nodes to be determined in ascending order according to the terrain unit node identifier. Sort the safe retreat positions in ascending order according to the horizontal coordinate, vertical coordinate and elevation value. For each terrain unit node to be determined, the corresponding terrain unit node is set as the backoff search starting point, the cumulative backoff energy of the backoff search starting point is set to 0, the cumulative backoff energy of the remaining reachable terrain unit nodes and cross-ridge event nodes is initialized to infinity, and all nodes are set to unvisited state. In each round, the node with the smallest cumulative backoff energy among the unvisited nodes is selected as the current search node. When the current search node is a terrain unit node, the directed connection from the terrain unit node to the cross-ridge event node is read. When the current search node is a cross-ridge event node, the directed connection from the cross-ridge event node to the target terrain unit node is read. When a terrain unit node points to a ridge-crossing event node, the corresponding ridge-crossing event flight energy is used as the connection increment energy. When a ridge-crossing event node points to a target terrain unit node, the connection increment energy is set to 0. The cumulative backoff energy of the current search node is added to the connection increment energy to obtain the candidate cumulative backoff energy. When the candidate cumulative backoff energy is less than the target node's recorded cumulative backoff energy, the candidate cumulative backoff energy replaces the original cumulative backoff energy, and the current search node is identified as the predecessor node of the target node. When the current search node corresponds to the terrain unit node where the safe retreat position is located, read the flight segment within the terrain unit from the terrain unit node to the safe retreat position, extract the elevation data and wind field data of the corresponding flight segment from the mountain environment data, extract the current payload from the UAV status data, calculate the retreat energy within the terrain unit according to the flight distance, elevation change and wind field along the flight direction component, and add it to the cumulative retreat energy of the current search node to obtain the retreat flight energy to reach the corresponding safe retreat position; After completing the search for all reachable safe retreat locations, the search results are sorted in ascending order of retreat flight energy. The minimum retreat flight energy is determined as the minimum retreat energy of the terrain unit node to be judged. Then, the results are backtracked from the corresponding safe retreat location to the terrain unit node to be judged in reverse order along the predecessor node identifier. Finally, the backtracking results are arranged in reverse order to generate the corresponding safe retreat path. When all safe retreat positions are unreachable, write 0 to the retreat reachability status of the terrain unit node to be determined, and record the minimum retreat energy field as null. The terrain unit nodes to be determined that have no retreat path and whose minimum retreat energy is greater than the expected remaining energy of the terrain unit are identified as safe retreat dependent nodes. Specifically: Read the retreat reachability, minimum retreat energy and expected remaining energy of the terrain unit according to the node identifier of the terrain unit to be determined, so that the three data items correspond one-to-one according to the same node index; When the reachability of the backoff is 0, the corresponding terrain unit node to be determined will be directly marked as a safe backoff dependent node, and the dependency reason will be backoff path failure; When the retreat reachability state is 1, the minimum retreat energy is compared with the expected remaining energy of the terrain unit. If the minimum retreat energy is greater than the expected remaining energy of the terrain unit, the corresponding terrain unit node to be determined is marked as a safe retreat dependent node, and the dependency reason is insufficient retreat energy. When the retreat reachability state is 1 and the minimum retreat energy is not greater than the expected remaining energy of the terrain unit, the corresponding terrain unit node to be determined will be marked as a non-safe retreat dependent node. All safe backoff dependent nodes are sorted in ascending order according to the terrain unit node identifier. The node identifier, dependency reason, expected remaining energy of the terrain unit, and minimum backoff energy obtained by re-search are recorded to generate a safe backoff dependent node set. By aggregating secure backoff dependency nodes, connected cross-ridge event nodes, and directed connections, a secure backoff dependency domain is generated, specifically: Using the terrain unit nodes in the set of safe retreat dependent nodes as the core nodes in the domain, read the directed edges that are directly connected to the core nodes in the cross-ridge event graph, and read the corresponding cross-ridge event nodes. For the read cross-ridge event node, if the cross-ridge event node connects to two safe backoff dependent nodes, then the cross-ridge event node and all directed connections on both sides are included in the safe backoff dependent domain. If the cross-ridge event node connects to only one safe backoff dependent node, then the cross-ridge event node is included in the safe backoff dependent domain and the non-safe backoff dependent terrain unit node on the other side is recorded as the domain boundary node. Continue reading connected cross-ridge event nodes from the newly added safe backoff dependency node. Continue expanding outward only when the newly added node belongs to the set of safe backoff dependency nodes. Stop expanding in the corresponding direction when encountering a non-safe backoff dependency node. The topographic unit nodes in the domain are arranged in ascending order according to the topographic unit node identifier, and the cross-ridge event nodes in the domain are arranged in ascending order according to the cross-ridge event node identifier. The directed connection relationship in the domain, the domain boundary node identifier, the dependency reason, and the cross-ridge event node identifier of the state change are recorded respectively. The set of terrain unit nodes, the set of cross-ridge event nodes, the set of directed connections, and the set of boundary nodes within the domain are associated and stored to generate a safe backoff dependency domain.

[0030] In this embodiment, generating the updated logistics route includes: Based on the safety backoff dependency domain, the cross-ridge event graph is divided, retaining the external terrain unit nodes, cross-ridge event nodes, directed connections, ARMA recursive states, and backoff anchor states. The ARMA recursive states of the domain boundary nodes are determined as the initial boundary states, where: The cross-ridge event graph is partitioned based on the safe backoff dependency domain, specifically as follows: Read the unified node sequence, directed connection relationships, and domain terrain unit node set, domain cross-ridge event node set, and domain boundary node set in the cross-ridge event graph; The unified node sequence is compared with the set of nodes in the safe backoff dependency domain one by one according to the node identifier. Nodes belonging to the safe backoff dependency domain are written into the node sequence within the domain, and nodes that do not belong to the safe backoff dependency domain are written into the node sequence outside the domain. Read the directed connections in the cross-ridge event graph one by one, write the directed connections whose start and end points both belong to the intra-domain node sequence into the intra-domain connection set, write the directed connections whose start and end points both belong to the extra-domain node sequence into the extra-domain connection set, and write the directed connections whose one end belongs to the intra-domain node sequence and the other end belongs to the extra-domain node sequence into the boundary connection set. A subgraph corresponding to the safe backoff dependency domain is formed according to the intra-domain node sequence and intra-domain connection set, and an extra-domain reserved subgraph is formed according to the extra-domain node sequence and extra-domain connection set. The connection position between the two subgraphs is saved through the boundary connection set. Preserve the following: external terrain unit nodes, cross-ridge event nodes, directed connectivity, ARMA recursive state, and backoff anchor state. Read the state of the last node of each parallel ARMA recursive stack saved when the path reasoning was completed before the state change occurred, and match each row of the node state matrix with the node identifier according to the unified node sequence. Extract the node records, directed connections, and final round states of each parallel ARMA recursive stack corresponding to the external terrain unit nodes and external cross-ridge event nodes based on the external node sequence, without re-executing the recursive calculation of the external nodes; Read the corresponding backoff anchor state according to the off-site cross-ridge event node identifier, keep the backoff anchor point, minimum backoff energy, cross-ridge event inclusion vector and cross-ridge event order vector unchanged, and when the cross-ridge event with state change belongs to the event in the corresponding backoff anchor, mark the corresponding backoff anchor as pending update state instead of as an off-site fixed backoff anchor. The retained out-of-domain node records, directed connection relationships, ARMA recursive states, and unaffected backoff anchor states are indexed according to node identifiers to form the out-of-domain retained states; The ARMA recursive state of the domain boundary node is determined as the initial boundary state, specifically: Read each directed connection in the boundary connection set, determine the end belonging to the sequence of nodes outside the domain as the boundary node outside the domain, and determine the end belonging to the sequence of nodes inside the domain as the receiving node inside the domain; Extract the corresponding 64-dimensional state vector from the final node state of each parallel ARMA recursive stack according to the external boundary node identifier, and arrange them in ascending order according to the parallel ARMA recursive stack number. When a receiving node within the same domain connects to multiple boundary nodes outside the domain, the state vectors of each boundary node outside the domain are weighted according to the propagation coefficients corresponding to the updated cross-ridge event propagation matrix, and the weighted state vectors are summed element by element to obtain the boundary input state corresponding to the receiving node within the domain. For each of the four parallel ARMA recursive stacks, a boundary input state is formed. The domain receiving node identifier, the parallel ARMA recursive stack number and the corresponding 64-dimensional boundary input state are associated and recorded to generate the boundary initial state. When the safe backoff depends on the domain covering all nodes of the cross-ridge event graph and there are no external boundary nodes, the initial state of the boundary is recorded as empty, and all recursive states are initialized by the input features of the nodes within the domain. Based on the updated mountainous environment data and UAV status data, delete the cross-ridge event nodes that have lost access conditions within the safe retreat dependency domain, update the directed connection relationships within the domain, and generate an updated cross-ridge event subgraph. Based on the safe backoff position, a new safe backoff path is searched in the updated cross-ridge event subgraph. The cross-ridge event flight energy and minimum backoff energy are recalculated, and an updated backoff anchor state is generated based on the new safe backoff relationship, where: Based on the safe backoff location, a new safe backoff path is searched in the updated cross-ridge event subgraph, specifically as follows: Read the terrain unit nodes, ridge event nodes, and directed connections in the updated ridge event subgraph, read the previously determined safe retreat positions, and establish connections between the terrain unit nodes where the safe retreat positions are located and the updated ridge event subgraph. For each cross-ridge event node in the updated cross-ridge event subgraph, the exit position of the cross-ridge event is used as the starting point for the retreat search. The terrain unit nodes and cross-ridge event nodes are expanded in order of increasing cumulative retreat flight energy. Each time the target node's cumulative retreat flight energy and the precursor node's identifier are expanded, the new cumulative retreat flight energy is less than the previously recorded value, replacing the original record and updating the precursor node's identifier. After finding the terrain unit node where the safe retreat position is located, add the terrain unit node to the retreat flight energy of the flight segment within the corresponding terrain unit of the safe retreat position, and record the complete retreat flight energy. Starting from the safe retreat position with the minimum retreat flight energy, trace back along the predecessor node marker to the exit position of the cross-ridge event and arrange them in reverse order to obtain the updated safe retreat path and safe retreat relationship; The flight energy and minimum retreat energy for the cross-ridge event are recalculated as follows: Elevation data and wind field data corresponding to each ridge location in the updated ridge-crossing event submap are extracted from the updated mountainous environmental data. Current payload, UAV empty mass, cruise airspeed, baseline level flight power, and propulsion efficiency are extracted from the updated UAV status data. The three-dimensional flight distance and elevation difference between adjacent positions are calculated in the order of entry position, crossing position and exit position of the ridge crossing event. The inner product of the corresponding horizontal heading unit vector and the horizontal wind field vector of the adjacent positions is performed to obtain the wind speed component along the heading. The ground speed is calculated by the sum of the cruising airspeed and the wind speed component along the heading. The flight time is obtained by dividing the three-dimensional flight distance by the ground speed. When the ground speed is not greater than 0, the corresponding ridge crossing event is marked as a loss of passage conditions. The basic flight energy is calculated based on the current flight quality, baseline level flight power, and flight time. When the elevation difference is greater than 0, the additional climb energy is calculated based on the current flight quality, elevation difference, 9.81, and propulsion efficiency. The basic flight energy of each segment is added to the additional climb energy to obtain the updated cross-ridge event flight energy. The retreat flight energy within the terrain unit and the updated cross-ridge event flight energy are accumulated segment by segment according to the updated safe retreat path. The minimum value of the retreat flight energy corresponding to all reachable safe retreat positions is determined as the updated minimum retreat energy. The updated backoff anchor state is generated based on the new safe backoff relationship, specifically as follows: Read the new safe backoff relationship and the corresponding safe backoff position, and determine the safe backoff position corresponding to the minimum backoff energy as the updated backoff anchor point; Read the cross-ridge event node identifiers according to the actual sequence of the new safety backoff relationship, generate and update the cross-ridge event sequence, and regenerate the cross-ridge event containment vector and cross-ridge event sequence vector according to the global cross-ridge event fixed event index; Write the cross-ridge event node identifier, the updated 3D coordinates of the retreat anchor point, the updated terrain unit node identifier to which the retreat anchor point belongs, the updated minimum retreat energy, the updated cross-ridge event sequence length, the cross-ridge event inclusion vector, and the cross-ridge event sequence vector into the updated retreat anchor track record; All updated backoff anchor records in the cross-ridge event subgraph are updated in ascending order according to the cross-ridge event node identifier. Cross-ridge event nodes that have not formed a safe backoff path will write the backoff reachability status to 0 and record the minimum backoff energy field as an empty value, and generate updated backoff anchor status. The updated cross-ridge event subgraph, cross-ridge event flight energy, minimum backoff energy, updated backoff anchor state, and boundary initial state are input into the improved ARMA graph neural network. Intradomain recursive propagation is initiated with the boundary initial state to regenerate the event open state and subsequent associated states, where: The improved ARMA graph neural network is input by updating the cross-ridge event subgraph, cross-ridge event flight energy, minimum backoff energy, updated backoff anchor state, and boundary initial state. Read the terrain unit nodes and cross-ridge event nodes in the updated cross-ridge event subgraph, sort the terrain unit nodes in ascending order according to the terrain unit node identifier, and sort the cross-ridge event nodes in ascending order according to the cross-ridge event node identifier to generate a unified node sequence within the domain. Construct an intra-domain adjacency matrix and a node degree matrix according to a unified intra-domain node sequence. Perform node degree symmetric normalization on the intra-domain adjacency matrix to generate an intra-domain cross-ridge event propagation matrix. Subtract the intra-domain cross-ridge event propagation matrix from the identity matrix to generate an intra-domain cross-ridge event Laplace matrix. Extract the current remaining energy from the updated UAV status data. For each cross-ridge event node in the domain, subtract the corresponding cross-ridge event flight energy from the current remaining energy to obtain the expected remaining energy. Then, subtract the updated minimum backoff energy from the expected remaining energy to obtain the cross-ridge reversible margin. When the reversible margin across the ridge is greater than 0, the recursive maintenance coefficient is obtained by dividing the reversible margin across the ridge by the expected remaining energy. When the reversible margin across the ridge is not greater than 0, when the cross-ridge event loses its passage conditions, or when no safe retreat path is formed, the recursive maintenance coefficient is set to 0. Establish a domain recursive preservation matrix according to the unified node sequence within the domain. Write 1 to the diagonal position of the terrain unit node and write the corresponding recursive preservation coefficient to the diagonal position of the cross-ridge event node. Read the updated backoff anchor state, construct the original backoff anchor features in the order of the relative coordinates of the backoff anchor point, the normalized minimum backoff energy, and the weighted features of the updated cross-ridge event sequence, and generate the updated backoff anchor injection features using the parameters of the trained backoff anchor jump injection layer. The updated backtracking anchor backfeed features and the intra-domain node input features are concatenated according to the same node index, and the intra-domain backfeeding node input features are generated using the trained input jump mapping parameters. Initiating intra-domain recursive propagation with the initial boundary state is as follows: Keep the four parallel ARMA recursive stacks, the number of recursions in each stack, the 64-dimensional hidden state, and the recursive transformation matrices unchanged after training. Only replace the domain propagation matrix, recursion preservation coefficients, backtracking anchor back-injection features, and node input features. For nodes within the domain that are not part of the domain boundary receiving node, the input features of the nodes within the domain are mapped to 64 dimensions to obtain the first round of recursive state. For nodes within the domain that are part of the domain boundary receiving node, the 64-dimensional boundary initial state of the corresponding parallel ARMA recursive stack is read. The initial state of the boundary is added element by element to the input feature mapping result of the corresponding node in the domain to obtain the first round of recursive state of the receiving node at the domain boundary, so that the historical propagation state retained outside the domain for safe backoff continues to be input into the domain from the original boundary position; At the start of each round of recursive propagation, the recursive preservation matrix within the domain is multiplied by the current round node state, so that the cross-ridge event node state is scaled according to the updated recursive preservation coefficient, while the terrain unit node state remains unchanged. The scaled node state is aggregated by using the cross-ridge event propagation matrix within the domain. The aggregation result is input into the recursive transformation matrix of the corresponding parallel ARMA recursive stack. The input features of the back-injection nodes within the domain are mapped to 64 dimensions through the input jump branch. The recursive transformation result, the input jump branch result, and the boundary input state corresponding to the current round are added element by element and linear rectification activation is performed to obtain the node state in the current round domain. The boundary input state is only written in the first round of recursive propagation and is not added again from the second round onwards. The current state of the node in the domain is used as the input for the next round of recursion. The recursive scaling, directed propagation, recursive transformation and backtracking anchor jump back-injection are repeatedly executed. After completing 4 rounds of recursive propagation, the state of the node in the last round of each parallel ARMA recursive stack is obtained. When there are no domain boundary nodes, all nodes within the domain initialize their first round of recursive states with the updated node input features and complete the intra-domain propagation according to the same recursive process. Regenerate the event open state and subsequent associated states, specifically as follows: Read the final node state of each parallel ARMA recursive stack according to the unified node sequence within the domain, sum the same-dimensional state values ​​of the same node in the four parallel ARMA recursive stacks and divide by 4 to obtain the domain-fused state. Extract the 64-dimensional fusion state corresponding to the cross-ridge event nodes within the domain, sort them in ascending order according to the cross-ridge event node identifiers, and generate and update the cross-ridge event fusion representation. Read the recursive retention coefficients corresponding to cross-ridge events in each domain, write cross-ridge events with a recursive retention coefficient of 0 to the closed state, and write cross-ridge events with a recursive retention coefficient greater than 0 to the open state, thus forming the open state of the update event; For each open ridge crossing event, read the target terrain unit node identifier, find intra-domain ridge crossing events that start with the target terrain unit node and are in an open state, and read open ridge crossing events that can continue to enter the reserved path outside the domain through boundary connection relationship, forming a candidate successor event set; The current open cross-ridge event fusion state, candidate successor event fusion state, current recursive preservation coefficient and candidate successor event recursive preservation coefficient are concatenated in a predetermined order. The association score is calculated using the readout layer parameters of the trained cross-ridge event sequence. Exponential normalization is then performed within the same candidate successor event set to obtain the updated successor association value. The successor association value is updated according to the current cross-ridge event node identifier and the candidate successor cross-ridge event node identifier. An index association is established between the updated event open status and the updated successor association value to obtain the regenerated event open status and successor association status. Based on the regenerated event open state and subsequent associated state, an alternative cross-ridge event sequence and corresponding terrain unit intra-segment are generated, which are then connected with the logistics path retained outside the domain to generate an updated logistics path.

[0031] Example 1: In a simulated drone logistics delivery scenario in a mountainous area, the planned area is 18.9 square kilometers, the digital elevation grid resolution is 10 meters, resulting in 186,402 effective grids. The maximum elevation difference in the area is 612 meters. There are 37 obstacle areas, 4 no-fly zones, and 11 areas with weak communication coverage. The logistics task requires the drone to transport 3.2 kg of goods. The maximum allowable payload of the drone is 6 kg, the available battery energy is 360 Wh, and the remaining energy at the start of the task is 92%, equivalent to 331.2 Wh. The allowed task completion time is 28 minutes.

[0032] The system reads mountainous environmental data, drone status data, and logistics task data, unifying coordinates, elevation benchmarks, and data collection times. It then maps the drone's current location, logistics start point, logistics destination, and delivery node to a digital elevation grid. Based on continuous elevation ridges, obstacle distribution, and no-fly zones, 23 terrain units are identified. Thirty-one ridge-crossing locations are selected between adjacent terrain units, and ridge-crossing event nodes are established for each of the two travel directions at each ridge-crossing location, resulting in 62 ridge-crossing event nodes. These 23 terrain unit nodes and 62 ridge-crossing event nodes together constitute 85 graph nodes and 124 directed connections.

[0033] Calculate the arrival flight energy, the flight energy of the crossing event, and the minimum backoff energy for each ridge crossing event. For one candidate ridge crossing event, the arrival flight energy is 41.8 Wh, the crossing event flight energy is 18.6 Wh, the expected residual energy after crossing is 270.8 Wh, the minimum backoff energy is 83.7 Wh, the ridge crossing reversibility margin is 187.1 Wh, and the recursive preservation coefficient is 0.691. For another candidate sequence, the expected residual energy of the distant ridge crossing event is 122.6 Wh, the minimum backoff energy is 127.9 Wh, the ridge crossing reversibility margin is -5.3 Wh, the recursive preservation coefficient is set to 0, and it is marked as off.

[0034] The retreat anchor point for an open cross-ridge event is 2.46 km from the exit position, with a minimum retreat energy of 64.9 Wh. The safe retreat path passes through three cross-ridge events sequentially. The system reads the corresponding 32-dimensional node input features in retreat order, determines the weights based on the reciprocal of the sequence position, normalizes them, and obtains a weighted 32-dimensional cross-ridge event sequence feature. This feature is then connected with the 3-dimensional relative coordinates of the retreat anchor point and the 1-dimensional minimum retreat energy to form a 36-dimensional retreat anchor track original feature. This is mapped to a 32-dimensional retreat anchor track back-injection feature, which is then concatenated with the 32-dimensional node input features to form a 64-dimensional jump input.

[0035] After reasoning, 44 open events and 18 closed events were obtained from 62 ridge-crossing events, forming 7 feasible ridge-crossing event sequences. The selected path length is 12.8 kilometers, the estimated flight time is 19.6 minutes, the estimated total energy consumption is 214.3 watt-hours, the communication coverage length accounts for 97.8%, the maximum continuous communication interruption distance is 34 meters, the corresponding ground speed of the flight segment is 10.7 meters per second, and the maximum continuous loss of contact time is 3.18 seconds, which is less than the allowable value of 5 seconds.

[0036] During execution, a new impassable obstacle appeared in the area where an originally open cross-ridge event was located. Traditional local replanning uses an update radius of 1.5 km, which is the 90th percentile of the distance corresponding to the training samples, updating a total of 9 terrain units and 21 cross-ridge event nodes. This invention reads the original safe backoff relationship and temporarily removes the changed cross-ridge event, re-searches for backoff paths, and identifies 5 safe backoff dependent nodes and 12 connected cross-ridge event nodes. Among them, 2 terrain units are more than 1.5 km away from the changed location, with minimum backoff energies of 91.6 Wh and 96.4 Wh, respectively, which are higher than the corresponding expected remaining energies of 84.1 Wh and 88.5 Wh, and are therefore included in the safe backoff dependent domain.

[0037] The safe backoff dependency domain contains 17 nodes, with 68 nodes retained outside the domain. The system retains the nodes outside the domain, directed connections, the final recursive state of ARMA, and the backoff anchor state. It reads the 64-dimensional final state of the four boundary nodes outside the domain as the initial boundary state and reconstructs only the 12 cross-ridge event nodes within the domain and their related connections. After inference, two cross-ridge events in the original path change from open to closed, generating a 4.1-kilometer alternative path segment and connecting it with the path outside the domain. The total path length after the update is 13.2 kilometers, and the replanning time is 0.89 seconds.

[0038] The model training used 4800 historical logistics task samples, including 3840 training samples, 480 validation samples, and 480 test samples. The samples covered elevation differences of 180 to 760 meters, task loads of 0.8 to 5.5 kg, initial remaining energy of 58% to 100%, and 2 to 11 ridge-crossing events. One training sample contained 27 terrain units and 72 directional ridge-crossing events. The actual trajectory passed through E08, E19, E27, and E41 sequentially. E19 had an estimated remaining energy of 118.5 Wh and a minimum backoff energy of 73.2 Wh. E27 was the actual successor event to E19. In another training sample, E33 had no reachable safe backoff path, a recursive preservation coefficient of 0, and E33 did not enter the actual execution sequence.

[0039] Under conditions of 480 identical test tasks and an 8-core CPU, 16GB of RAM, and an 8GB GPU, this invention is compared with the traditional ARMA graph neural network path planning method. Both methods use the same training samples, node input features, four parallel recursive stacks, and 64-dimensional hidden states. The traditional method does not implement cross-ridge reversible recursive gating, backoff anchor jump injection, or safe backoff dependency domain replanning. The traditional method achieves a task success rate of 91.5%, while this invention achieves 96.9%. The percentages of unreachable tasks after crossing ridges are 8.8% and 2.1%, respectively. The average path lengths are 12.9 km and 13.1 km, respectively, and the average task energy consumption is 222.8 Wh and 217.4 Wh, respectively.

[0040] In 162 test tasks involving state changes, the traditional method recalculated an average of 31.4 nodes, while the present invention recalculated 14.7 nodes, with average replanning times of 1.93 seconds and 0.88 seconds, respectively. The traditional method experienced secondary replanning in 20 tasks (12.3%), while the present invention experienced it in 4 (2.5%). The traditional method satisfied the safe backoff condition in 145 tasks (89.5%), while the present invention satisfied it in 158 tasks (97.5%). The results show that the present invention can utilize the cross-ridge reversible margin to control ARMA recursive propagation and determine the local replanning range based on safe backoff topological dependencies. It reduces task energy consumption and unsafe backoff events while only increasing the path length by about 1.6%, and also reduces the computational cost of dynamic replanning, verifying the feasibility and effectiveness of the method.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks, characterized in that, include: Collect comprehensive logistics planning data, perform preprocessing on the comprehensive logistics planning data, and obtain standard route planning data; Based on standard path planning data, terrain units, ridge crossing locations, and safe retreat locations are identified, and a ridge crossing event map is constructed. Determine the minimum retreat energy and retreat anchorage state based on the cross-ridge event map; The cross-ridge event graph, minimum backoff energy, and backoff anchor state are input into an improved ARMA graph neural network to perform cross-ridge event Laplace reconstruction, reversible recursive gating, and backoff anchor moving average backinjection. The recursive state is updated based on the expected remaining energy and minimum backoff energy to generate event open state and subsequent associated state. Based on the open state and subsequent associated state of the event, a cross-ridge event sequence is generated, connecting the segments within the terrain unit to generate candidate logistics routes; Based on drone status data and logistics task data, candidate logistics routes are validated and logistics routes are generated. Execute the logistics path and update the running status, determine the cross-ridge events of the status change, and generate the safety backoff dependency domain along the safety backoff relationship; Preserve the recursive state outside the safe backoff dependency domain, reconstruct the cross-ridge event graph and backoff anchor state within the safe backoff dependency domain, re-execute the improved ARMA graph neural network inference, and generate an updated logistics path.

2. The method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks according to claim 1, characterized in that, The comprehensive logistics planning data includes mountainous environment data, drone status data, and logistics task data.

3. The method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks according to claim 1, characterized in that, The obtained standard path planning data includes: Perform coordinate unification, elevation benchmark unification, and anomaly data processing on mountain environmental data to generate standard mountain environmental data; The drone status data is aligned with standard mountainous environment data according to the time of collection, and the current location of the drone is mapped to the corresponding spatial location to generate standard drone status data. The logistics origin, logistics destination and delivery node in the logistics task data are mapped to standard mountainous environmental data to generate standard logistics task data. Standard route planning data is obtained by associating standard mountainous environmental data, standard drone status data, and standard logistics task data with spatial location.

4. The method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks according to claim 1, characterized in that, The construction of the trans-ridge event graph includes: Digital elevation data, obstacle distribution data, and no-fly zone data are extracted from mountainous environmental data. Continuous elevation ridges are extracted and combined with obstacle distribution and no-fly zone data to divide terrain units. Extract flight altitude constraints from logistics task data, filter continuous elevation ridge positions between adjacent terrain units that meet flight altitude constraints and do not fall into obstacle distribution areas or no-fly zones, and determine the ridge crossing positions; The logistics start point, logistics end point, and delivery node are extracted from the logistics task data. Obstacle distribution data, communication coverage data, and no-fly zone data are extracted from the mountain environment data. The logistics start point, logistics end point, and delivery node that have communication coverage and have not fallen into the obstacle distribution area or no-fly zone are identified as safe retreat locations. The terrain units are mapped to terrain unit nodes. Cross-ridge event nodes are established for the two travel directions of adjacent terrain units, and directed connections are established according to the cross-ridge positions to construct a cross-ridge event graph.

5. The method for planning logistics routes for unmanned aerial vehicles (UAVs) in mountainous areas based on graph neural networks according to claim 1, characterized in that, The determination of minimum retreat energy and retreat anchor state based on the cross-ridge event graph includes: Elevation and wind field data are extracted from mountainous environmental data, and the current payload is extracted from UAV status data. The flight distance, elevation change and wind direction effect are calculated based on the entry position, crossing position and exit position of the ridge crossing event. The flight energy of the ridge crossing event is calculated in combination with the current payload. Search for the retreat path from the exit position of the cross-ridge event to the safe retreat position, calculate the retreat flight energy based on flight distance, elevation change, wind field data and current load, determine the minimum value as the minimum retreat energy, and record the corresponding retreat path as the safe retreat relationship; The safe retreat position corresponding to the minimum retreat energy is determined as the retreat anchor point. The retreat anchor point, the minimum retreat energy, and the cross-ridge event sequence are associated to generate the retreat anchor track state.

6. The method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks according to claim 1, characterized in that, The open state and subsequent associated state of the generated event include: Based on the graph propagation structure, recursive propagation branch, input jump branch and recursive stack fusion structure of ARMA graph neural network, an improved ARMA graph neural network is constructed by setting up a cross-ridge event Laplace reconstruction layer, a cross-ridge reversible recursive gating layer, a backtracking anchor jump injection layer, an ARMA recursive fusion layer and a cross-ridge event sequence readout layer. The cross-ridge event graph is input into the cross-ridge event Laplace reconstruction layer. Based on the terrain unit nodes, cross-ridge event nodes and directed connectivity, an adjacency matrix and a node degree matrix are constructed. The node degree is symmetrically normalized on the adjacency matrix to generate the cross-ridge event propagation matrix. The cross-ridge event Laplace matrix is ​​generated by subtracting the cross-ridge event propagation matrix from the identity matrix. The remaining energy is extracted from the UAV state data. The flight energy of the cross-ridge event is subtracted from the remaining energy to obtain the expected remaining energy. The minimum backoff energy is subtracted from the expected remaining energy to obtain the cross-ridge reversible margin. When the cross-ridge reversible margin is greater than 0, the ratio of the cross-ridge reversible margin to the expected remaining energy is used as the recursive maintenance coefficient. When the cross-ridge reversible margin is not greater than 0, the recursive maintenance coefficient is set to 0, and the ARMA recursive propagation state corresponding to the cross-ridge event propagation matrix is ​​adjusted with the recursive maintenance coefficient. Input the retreat anchor state into the retreat anchor jump injection layer, perform feature mapping on the retreat anchor point, minimum retreat energy and cross-ridge event sequence to generate retreat anchor injection features, concatenate the retreat anchor injection features with the node input features, and inject them in each round of recursive propagation through ARMA input jump branches; In the ARMA recursive fusion layer, the node states in each parallel ARMA recursive stack are updated based on the cross-ridge event propagation matrix, recursive preservation coefficient, and node input features after back-injection. The final node states of each parallel ARMA recursive stack are fused to generate a cross-ridge event fusion representation. The cross-ridge event fusion representation and recursive preservation coefficient are input into the cross-ridge event sequence readout layer. Cross-ridge events with a recursive preservation coefficient of 0 are marked as closed, and cross-ridge events with a recursive preservation coefficient greater than 0 are marked as open. The open cross-ridge event fusion representation is concatenated with the fusion representation of directed adjacent open cross-ridge events. After mapping and normalization, the subsequent association value is generated to obtain the event open state and the subsequent association state.

7. The method for planning unmanned aerial vehicle (UAV) logistics routes in mountainous areas based on graph neural networks according to claim 1, characterized in that, The generation of candidate logistics routes includes: Based on the open state of the event, search for open ridge-crossing event sequences that can connect to the logistics destination and do not repeatedly pass through ridge-crossing events from the terrain unit node where the logistics origin is located. The successor correlation values ​​of each open cross-ridge event sequence are summed, and the sequence with the largest cumulative value is determined as the cross-ridge event sequence; Based on the exit and entry positions of adjacent ridge crossing events, a route segment within the terrain unit is generated, connecting the logistics starting point, the route segment within the terrain unit, the ridge crossing position, and the logistics endpoint to generate candidate logistics routes.

8. The method for planning logistics routes for unmanned aerial vehicles (UAVs) in mountainous areas based on graph neural networks according to claim 1, characterized in that, The generated logistics path includes: Extract remaining energy, maximum allowable payload, and flight speed from drone status data; extract mission payload and mission time limit from logistics mission data; and extract communication coverage status from mountainous environment data. Calculate the estimated flight energy and estimated flight duration of candidate logistics routes, and detect the continuity of communication coverage; Candidate logistics routes that meet the requirements of energy, payload, mission time limit, and communication continuity are identified as logistics routes.

9. A method for planning logistics routes for unmanned aerial vehicles (UAVs) in mountainous areas based on graph neural networks according to claim 1, characterized in that, The generation of the safe backoff dependency domain along the safe backoff relation includes: During the execution of the logistics route, update the mountain environment data and drone status data, map the updated mountain environment data and drone status data to the ridge crossing event map, extract obstacle distribution status, no-fly status and wind field status from the updated mountain environment data, extract remaining energy from the updated drone status data, and determine the ridge crossing event of status change. Read the safety backoff relationship before the removal of the state change cross-ridge event, determine the terrain unit node to be judged through the state change cross-ridge event, and temporarily remove the state change cross-ridge event from the cross-ridge event map; Wind field status is extracted from the updated mountainous environment data, and remaining energy is extracted from the updated UAV status data. The expected flight energy is calculated based on the flight distance, elevation change and wind field status of the logistics path segment from the current location to each terrain unit node to be determined. The expected remaining energy of the terrain unit is obtained by subtracting the expected flight energy from the remaining energy. After removing state-change cross-ridge events, re-search the retreat path from the undetermined terrain unit node to the safe retreat position in the cross-ridge event graph and calculate the minimum retreat energy. Determine the undetermined terrain unit node that has no retreat path or whose minimum retreat energy is greater than the expected remaining energy of the terrain unit as a safe retreat dependency node. Collect the safe retreat dependency nodes, connected cross-ridge event nodes and directed connections to generate a safe retreat dependency domain.

10. A method for planning logistics routes for unmanned aerial vehicles (UAVs) in mountainous areas based on graph neural networks according to claim 1, characterized in that, The generation and updating of the logistics route includes: Based on the safety retreat dependency domain, the cross-ridge event graph is divided, and the external terrain unit nodes, cross-ridge event nodes, directed connection relationships, ARMA recursive states and retreat anchor states are retained. The ARMA recursive states of the domain boundary nodes are determined as the initial boundary states. Based on the updated mountainous environment data and UAV status data, delete the cross-ridge event nodes that have lost access conditions within the safe retreat dependency domain, update the directed connection relationships within the domain, and generate an updated cross-ridge event subgraph. Based on the safe retreat position, a new safe retreat path is searched in the updated cross-ridge event subgraph, the cross-ridge event flight energy and minimum retreat energy are recalculated, and an updated retreat anchor state is generated based on the new safe retreat relationship. The updated cross-ridge event subgraph, cross-ridge event flight energy, minimum backoff energy, updated backoff anchor state, and boundary initial state are input into the improved ARMA graph neural network. The intradomain recursive propagation is initiated with the boundary initial state to regenerate the event open state and subsequent associated state. Based on the regenerated event open state and subsequent associated state, an alternative cross-ridge event sequence and corresponding terrain unit intra-segment are generated, which are then connected with the logistics path retained outside the domain to generate an updated logistics path.