Layered decoupling optimization method for in-transit collaborative distribution of vehicle-mounted unmanned aerial vehicle

Abstract

The invention relates to the field of multi-mode collaborative transportation and logistics, in particular to a collaborative distribution optimization method and system for taking off, landing and synchronous recovery of a vehicle-mounted unmanned aerial vehicle on the way (in the advancing process). According to the method, a layered decoupling framework is adopted; in the first layer, subsection division is performed on a road network, taking-off and landing positions are continuously optimized in candidate taking-off and landing subsection pairs based on a Lagrange multiplier method and geometric synchronization constraint, and a client cluster candidate set is formed; the second layer performs joint optimization on selection of a single take-off and landing sub-segment pair in each client cluster and a global path of a ground vehicle to construct a generalized traveling salesman variant, and designs a sub-segment distance measurement and an overlapping deduction strategy to deal with any continuous take-off and landing distance calculation and multi-unmanned aerial vehicle cooperation problem. Experiments show that compared with a fixed stop point or discretization in-transit strategy, the method can significantly reduce the total delivery cost and the completion time, and has robustness and expandability for customer scale, distribution, unmanned aerial vehicle speed and road network density change.

Description

Technical Field

[0001] This invention belongs to the field of smart logistics technology, specifically relating to an optimization method for vehicle-mounted drone collaborative delivery, particularly a layered decoupling optimization method for synchronous delivery en route. Background Technology

[0002] With the rapid development of drone technology, drones are increasingly being used in logistics and delivery. Traditional ground vehicle delivery systems have bottlenecks, such as traffic congestion and geographical limitations, while drones have the advantages of high mobility and rapid response. However, they are limited by their range and payload capacity, making it impossible for them to independently complete long-distance deliveries. Therefore, vehicle-mounted drone collaborative delivery systems have emerged. These systems utilize ground vehicles as mobile platforms to achieve coordinated optimization of drone takeoff and landing. Existing research shows that this system can significantly reduce delivery costs and improve efficiency, and has broad application prospects in logistics and emergency response.

[0003] Existing research often assumes that ground vehicles pause at predefined stops to allow drones to take off and land. This approach is typically modeled as a bilevel optimization or mixed-integer programming problem to determine the scheduling paths for vehicles and drones. However, this model has significant limitations: vehicles need to detour to the stops, increasing travel distance and traffic disruption; vehicles and drones wait for each other at the meeting point, leading to "synchronous idling," resulting in energy waste and reduced efficiency; fixed stops are difficult to adapt to dynamic traffic or weather conditions, limiting the system's flexibility and robustness.

[0004] To address these issues, a few studies in recent years have proposed on-the-go vehicle-mounted drone delivery systems, allowing drones to take off and land arbitrarily while the vehicle is in motion, achieving spatiotemporal synchronization. This system can reduce parking costs and traffic impact, minimize synchronization delays, and improve scheduling flexibility and environmental adaptability.

[0005] However, this introduces modeling and optimization challenges: theoretically, there are an infinite number of feasible stopping points along the vehicle path; the drone and vehicle need to collaborate in different spaces (road network and Euclidean space); and path optimization requires dynamically combining the results from both spaces. The problem is influenced by nonlinear geometric constraints and strong spatiotemporal synchronization coupling, resulting in a non-convex problem that is a hybrid continuous-discrete optimization. Traditional methods struggle to solve this efficiently. Existing simplification methods, such as restricting drone operations to predefined discrete positions or geometric approximations, reduce the search space but sacrifice optimality or incur high computational costs. Furthermore, incorporating spatiotemporal synchronization constraints into the model increases nonlinear distance calculations and path decision coupling, making it impossible to find a globally optimal solution.

[0006] The shortcomings of the existing technology are: (1) Fixed docking points limit flexibility, resulting in extra detours and waiting; (2) Continuous space modeling is complex, discretization or approximation methods have heavy computational burden and low solution quality; (3) Multi-UAV coordination and global path optimization lack efficient strategies and cannot handle arbitrary docking distances and concurrent operations. Summary of the Invention

[0007] The purpose of this invention is to provide a hierarchical decoupling optimization method for on-the-go collaborative delivery by vehicle-mounted drones, overcoming the limitations of fixed docking points and the complexity of continuous spatial optimization in existing technologies. This method decouples local synchronization and global path optimization through a hierarchical structure, supports concurrent operation of multiple drones, achieves spatiotemporal synchronization, and significantly reduces delivery costs and time.

[0008] This invention employs a layered decoupling framework:

[0009] The first layer (local layer) performs synchronous optimization on the continuous take-off and landing positions on the road sub-segment pairs to generate customer candidate clusters;

[0010] The second layer (global layer) selects a scheme from each customer candidate cluster and jointly optimizes the vehicle access order, constructing a GTSP variant. It then calculates the effective driving length of the vehicle using segment-level distance decomposition and path overlap reduction. Finally, it uses the ALNS metaheuristic to obtain an feasible approximate optimal delivery scheme.

[0011] Key symbols and parameter definitions:

[0012] G = (V, E) is a road network graph, where V is the set of nodes and E is the set of road edges.

[0013] L e The length of the road edge e, e∈E

[0014] The i-th road sub-segment, L i sub-segment length

[0015] A i B i sub-segment r i The starting and ending coordinates

[0016] P i (s) sub-segment r i The above parameter is the continuous position of s

[0017] κ i Upper bound of subsegment curvature

[0018] ε Linear approximation error bound

[0019] c j The coordinates of the j-th customer point, and the customer set C = {c1, ..., cj} n}

[0020] S Warehouse / Distribution Center Node

[0021] s and t are the continuous position parameters on the takeoff and landing segments, respectively.

[0022] d G (u,v) is the shortest path length from node u to v on the road network.

[0023] f(s,t) represents the Euclidean flight distance of the UAV.

[0024] h(s,t) represents the distance the vehicle travels on the road network.

[0025] v D drone speed

[0026] v T vehicle speed

[0027] α = v D / v T speed ratio

[0028] Δt Vehicle-UAV synchronization tolerance time

[0029] E max Maximum flight range of drones

[0030] Q j Candidate clusters of customer j

[0031] q j,l The l-th candidate in the candidate cluster

[0032] x j,l Candidate selection variables

[0033] π Customer visit order

[0034] Vehicle serial travel length

[0035] D T Considering the effective driving length of the vehicle after parallel reduction

[0036] K Number of drones

[0037] R e Number of overlapping directions of road edge e

[0038] Θ Reduction

[0039] λ D ,λ T Cost weighting of drones and vehicles

[0040] The Big-M constant is used to link constraints and selection variables.

[0041] GTSP Generalized Traveling Salesman Problem

[0042] ALNS Adaptive Large Neighborhood Search Metaheuristic

[0043] To achieve the above objectives, the technical solution of the present invention is as follows: a hierarchical decoupling optimization method for vehicle-mounted unmanned aerial vehicle (UAV) in-transit collaborative delivery, the method comprising the following steps.

[0044] Step 1: Road Network Segmentation and Candidate Generation. The road network G=(V,E) is segmented based on curvature κ. e Subdivided into controllable granularity subsets Each sub-segment r i Satisfy κ i L i ≤ε (ε is the linear approximation error bound), point parameterization is s∈[0,L i A i B i For endpoints.

[0045] The problem of having an infinite number of continuously feasible take-off and landing locations is transformed into continuous parameter locations on sub-segments, which not only preserves the accuracy of continuous optimization but also provides a structured foundation for subsequent calculations. Compared with restricting take-off and landing to discrete docking points or dense sampling, this step significantly reduces the burden of candidate space and distance calculations while ensuring that geometric errors are controllable.

[0046] Step 2: Local synchronization optimization. For each customer and sub-segment pairs (r i ,r k For each )∈R×R, the path distance of the UAV is defined as f(s,t)=‖P i (s)-c j ‖+‖c j -P k (t)‖, vehicle path distance Where d G (u,v) represents the shortest path length in the network. p is the edge sequence from u to v.

[0047] The optimization problem is:

[0048]

[0049] st

[0050] |g(s,t)|≤Δtv D

[0051] f(s,t)≤E max

[0052] 0≤s≤L i ,0≤t≤L k

[0053] Where |g(s,t)|=|f(s,t)-αh(s,t)|, α=v D / v T >1 represents the speed ratio, v T v is the vehicle speed. D For drone speed, E max Δt represents the maximum battery life, and Δt represents the synchronization time tolerance.

[0054] Pre-screening using pruning strategies:

[0055] (P2) Range Limit: If and Then exclude;

[0056] (P2) Existence of the zero solution: If Then exclude;

[0057] (P3) Directional Consistency: If the sub-segment direction vector d i ·d k >0 and (A) k -B i )·d i If the value is less than 0, then it is excluded.

[0058] For optimization using segment pairs, sequential least squares programming (SLSQP) or other nonlinear constraints, record the synchronous optimal solution (s) obtained for each pair of segment combinations. * ,t * ).

[0059] Generate each customer cluster T j,l = (u,v)∈E is the set of path segments for vehicles.

[0060] Traditional fixed-stop-point methods require vehicles to detour and wait at designated points; this step allows take-off and landing at any location along the way, directly reducing detours and waiting; compared to approximate methods that discretize continuous take-off and landing points, this step directly optimizes take-off and landing points in the continuous domain, improving solution quality; the pruning strategy eliminates a large number of unsolvable / inferior solution combinations before entering nonlinear optimization, significantly reducing computational load and improving scalability.

[0061] Step 3: Global path optimization.

[0062] Model the problem as a GTSP variant: from each Choose one q j,l (x j,l =1), connecting them into a global vehicle path. Decision variable: x j,l ∈{0,1},yj,j′ ∈{0,1} (cluster connection), u j ∈{1,2,…,n}(sequence), π=(π1,π2,…,π n )(arrangement).

[0063] Vehicle serial path length Where S represents the warehouse.

[0064] Multiple drone overlap merging: Θ=∑ e∈E min{max(0,R e -1),(K-1)}·2L e ,in

[0065] Actual path length

[0066] Optimization goal:

[0067] The constraints include:

[0068] Customer-only service constraint:

[0069] Start and end point constraints:

[0070] Unique in / out constraint:

[0071] Self-loop prohibition constraint:

[0072] Sub-ring constraint removal:

[0073] Parallel number constraint of drones: R e ≤K, Θ≥0

[0074] Synchronization constraints:

[0075] Battery life constraint: f j,l ≤E max +M(1-x j,l )

[0076] Range of values ​​for the variable: x j,l ,y j,j′ ∈{0,1},u j ∈{1,2,…,n},R e ≥0,D T ≥0;

[0077] Traditional methods tightly couple the solution of continuous synchronization positions and global paths, leading to an explosion in scale. This invention generates high-quality candidate clusters through hierarchical decoupling before performing combinatorial optimization, significantly reducing the difficulty. Segment-level distance decomposition avoids complex recalculation of continuous shortest paths in each global search, improving solution efficiency. Overlap reduction explicitly incorporates the vehicle segment reuse structure caused by multiple UAVs in parallel into vehicle mileage calculation, making cost assessment more consistent with physical facts and improving solution quality in multi-UAV scenarios.

[0078] Step 4: In the global path optimization process, since this problem is NP-hard, traditional exact algorithms struggle to obtain effective solutions in large-scale networks. This invention employs a cluster-based adaptive large neighborhood search (ALNS) metaheuristic method, specifically including:

[0079] Step 4-1, Candidate Graph Construction: Convert the customer's candidate take-off and landing points into a candidate graph, eliminate arcs that do not meet the synchronization and direction constraints, and narrow the search space;

[0080] Step 4-2, Initial Solution Generation: Using greedy insertion, vehicle paths are inserted sequentially based on the distance between the customer and the warehouse or the candidate cost. For each inserted customer, the candidate solution that minimizes the COST increment is selected, forming an initial feasible solution.

[0081] Step 4-3, Iterative optimization: Under the ALNS framework, destruction and repair operations are performed alternately. The destruction operation removes some candidate solutions, and the repair operation re-inserts them according to the comprehensive cost criterion. The algorithm adaptively adjusts the operator selection probability.

[0082] Step 4-4, Local Search: Swap 2-opt and 3-opt for vehicle access order to reduce vehicle length; perform "candidate replacement" and "neighborhood candidate fine-tuning" on candidate selections to reduce UAV flight distance or improve synchronization margin; trigger feasibility repair if necessary;

[0083] Steps 4-5, Weight Update: Set a weight w for each destruction / repair operator; in each iteration, score and update w based on the improvement brought by the operator, the feasible solution rate, etc., so that more effective operators are selected more frequently.

[0084] Steps 4-6, Acceptance criteria: Use threshold acceptance or simulated annealing criteria: if the new solution is better, accept it; if it is slightly worse, accept it with a temperature-related probability to escape the local optimum.

[0085] Steps 4-7, Termination Output: If the maximum number of iterations is reached or there is no improvement for L consecutive rounds, output the historical best solution, including vehicle path, customer access order, take-off and landing sub-segments for each customer and consecutive take-off and landing positions, and UAV task allocation.

[0086] Compared to the exact algorithm, ALNS can quickly obtain high-quality feasible solutions in large-scale instances; compared to local search in a single neighborhood, ALNS is more likely to escape local optima through a large neighborhood destruction-repair mechanism; adaptive weight updates make the algorithm more robust to changes in scenarios such as customer distribution and road network density.

[0087] Compared with the prior art, the present invention has at least the following technical effects:

[0088] (1) By segmenting the road and using continuous parameterization modeling, the theoretically infinite number of take-off and landing positions along the way are transformed into a controllable continuous optimization problem, which reduces computational complexity while ensuring geometric accuracy.

[0089] (2) By using local synchronous continuous optimization, the optimal take-off and landing positions are directly solved on the sub-segments, avoiding the synchronization error and suboptimal solutions introduced by the discretization method;

[0090] (3) The number of infeasible sub-segments is significantly reduced by pruning strategy, the number of nonlinear optimization calls is reduced, and the overall solution stability is improved;

[0091] (4) By decomposing vehicle distances at the sub-segment level and reducing the overlap of multiple UAV paths, the benefits of parallel execution can be accurately characterized.

[0092] (5) Through the adaptive large neighborhood search algorithm, a high-quality approximate optimal solution can still be obtained under the conditions of large-scale customers and complex road networks;

[0093] (6) The method shows good robustness and scalability for different customer sizes, customer distributions, number of drones, speed ratios and road restrictions. Attached Figure Description

[0094] To more clearly illustrate the technical solution of the present invention, only some of the accompanying drawings are briefly described below:

[0095] Figure 1 This is a schematic diagram illustrating the optimized process for in-transit vehicle-mounted drone delivery.

[0096] Figure 2 A schematic diagram illustrating a layered decoupling optimization method for on-the-go vehicle-mounted drone delivery.

[0097] Figure 3 This is a schematic diagram illustrating the breakdown and overlap reduction of vehicle routes for multi-drone delivery.

[0098] Figure 4 This is a sensitivity analysis chart of key parameters on system performance.

[0099] Figure 5 This is a route map for vehicle-mounted drone delivery under different takeoff restrictions in real-world scenarios. Detailed Implementation

[0100] To enhance understanding of the present invention, the embodiments will be described in detail below with reference to the accompanying drawings.

[0101] Example 1: Benchmark Dataset Experiment

[0102] This embodiment is used to verify the effectiveness of the hierarchical decoupling optimization method for on-the-go collaborative delivery of vehicle-mounted drones described in this invention.

[0103] This embodiment constructs a typical experimental scenario based on the Solomon benchmark dataset, where customer point distribution uses two instances: R101 (random distribution) and C101 (centralized distribution). During the experiment construction, a gridded road network is constructed based on a unified two-dimensional coordinate system, with grid spacing set to 6, 12, and 18 units to simulate urban road structures of different densities. The warehouse location is fixed at (40, 50), and customer nodes are only retained at non-overlapping points across all road structures to ensure experimental consistency.

[0104] The delivery system consists of one ground vehicle and several drones. The vehicle can only travel along the road network, while the drones can take off and land at any sub-segment of the road, subject to endurance and spatiotemporal synchronization constraints. In this embodiment, the vehicle speed is set to 1 unit distance / unit time, the drone speed is 1 to 3 times the vehicle speed (default is 1.5 times), the maximum drone endurance is 20 units, the unit cost of the vehicle is 1, and the unit cost of the drone is 0.1; the number of drones is set to 1 to 4 (default is 2). The synchronization time tolerance is set to ±0.001 units of time.

[0105] In the experiment, multi-scale task scenarios were constructed based on different numbers of customers (10, 20, 40), and multiple sets of variables such as drone speed (1 to 3 times that of vehicles), number of drones, road network spacing and sub-segment division granularity were set to test the stability and adaptability of the algorithm under multi-factor conditions.

[0106] In the experiment, according to Figure 2 The process involves first dividing the road network into several segments that satisfy the linear approximation condition, and then enumerating feasible drone take-off and landing segments based on customer locations. The local optimization method proposed in this invention is then used to calculate the synchronously optimal set of take-off and landing points. Subsequently, each customer candidate set is input into the global optimization model, and an adaptive large neighborhood search algorithm is used for iterative solution to finally obtain the globally optimal delivery route.

[0107] in, Figure 3This diagram illustrates the serial travel path decomposition of a vehicle serving two adjacent customer cluster tasks. The vehicle path consists of sub-paths such as "entry segment to takeoff sub-segment, takeoff sub-segment travel segment, shortest road network connection segment from takeoff sub-segment to landing sub-segment, landing sub-segment travel segment, and connection segment from landing sub-segment to the next task's takeoff sub-segment". The overlapping edge count R of the vehicle recovery / meeting paths corresponding to adjacent customer cluster tasks on the same-direction road edges is calculated. e In relation to the reduction amount Θ, when multiple drones are operating in parallel, overlapping road sections in the same direction can support multiple drone missions simultaneously with a single vehicle passage, thereby reducing the effective driving length of the vehicle.

[0108] Experimental results show that the method described in this invention significantly reduces total delivery costs, shortens task completion time, eliminates synchronous waiting, and improves path connectivity. Compared with traditional fixed-stop delivery methods, the method of this invention reduces delivery costs by approximately 20% and shortens delivery time by approximately 30% under the same conditions, and effectively eliminates waiting time between vehicles and drones. Compared with the in-transit optimization method using discrete points, the method of this invention can significantly reduce computation time and provides a superior solution.

[0109] Further sensitivity analysis showed (e.g.) Figure 4 As shown, when the speed of the drone increases to twice that of the vehicle, the delivery cost can be reduced by about 18%; when the number of drones increases to 3, the cost decreases by about 12%, but shows a diminishing marginal benefit trend; when the road network is denser, the cost decreases by about 10%, but the computation time increases by about 2 times; when the segment granularity is reduced, the accuracy of the solution is improved, but the computation time increases significantly.

[0110] Therefore, the method proposed in this invention can operate stably under various customer scales, distribution patterns, drone speeds, and road network conditions, demonstrating strong robustness and scalability.

[0111] Example 2: Real Case Study

[0112] To verify the practicality and robustness of the proposed vehicle-mounted drone-based collaborative delivery hierarchical decoupling optimization method in a real-world environment, a case study was conducted based on actual express delivery data from Tonglu County, Hangzhou City, Zhejiang Province, China. This region features a mixed urban and rural road network and complex geographical and transportation conditions, including rivers and highways.

[0113] In this embodiment, 20 customer delivery points are selected and aggregated from historical delivery data, with the warehouse serving as a unified distribution center. The system consists of one ground vehicle and two drones. The vehicle operates along the road network, and the drones can take off, land, and complete customer deliveries while the vehicle is in motion, returning to the vehicle or warehouse after completing their tasks.

[0114] The parameters are set as follows: vehicle speed is 40km / h, and unit transportation cost is $0.8 / km; drone flight speed is 60km / h, maximum range is 20km, and unit cost is $0.079 / km; the road network is segmented at a maximum interval of 300m, taking into account road curvature constraints.

[0115] To study the impact of road restrictions on the feasibility of UAV take-off and landing, four scenarios were set up: (1) No no-fly zone scenario ( Figure 5 (1) Drones can take off and land on all road sections; (2) Real-world no-fly zones ( Figure 5 (2) Exclude sections of road such as bridges and tunnels that are unsuitable for takeoff and landing; (3) Random no-fly zone scenario I Figure 5 (3) Randomly select 25% of the route sections where takeoffs and landings are prohibited; (4) Random no-fly zone scenario II Figure 5 (4)): Randomly select 50% of the road sections to prohibit take-off and landing.

[0116] The results show that in unrestricted scenarios, vehicles and drones form a compact delivery network, and drones can flexibly cross obstacles such as rivers, significantly reducing the detour distance of vehicles. In scenarios with real-world and random restrictions, although some take-off and landing opportunities disappear and delivery costs and times increase, overall efficiency is only slightly affected; even when take-off and landing are restricted on 50% of the road sections, the system still retains a significant efficiency advantage.

[0117] The results further demonstrate that the method of this invention can dynamically adjust the take-off and landing positions of drones under road segment constraints, and redistribute them at the sub-segment level through a hierarchical optimization framework, thereby ensuring the continuity and overall efficiency of the delivery network. Simultaneously, the ability of drones to complete deliveries across rivers effectively avoids the increased costs associated with vehicle detours, highlighting the adaptability and advantages of this invention in complex geographical environments.

[0118] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A hierarchical decoupling optimization method for vehicle-mounted drone collaborative delivery, characterized in that, Includes the following steps: Step S1: Subdivide the road edges of the road network into several parameterizable road segments according to curvature and length thresholds to obtain a set of segments and the endpoints and parameterized position expressions of each segment. Step S2: For each customer point and several candidate take-off and landing sub-segment pairs, firstly, infeasible sub-segment pairs are screened out according to the pruning strategy, and then a local synchronization optimization model with take-off position parameters and landing position parameters as continuous variables is established. Under the conditions of satisfying the UAV endurance constraint and the vehicle-mounted UAV spatiotemporal synchronization tolerance constraint, the flight distance of the UAV is minimized, and the candidate take-off and landing sub-segment pairs and corresponding optimal take-off and landing positions of the customer are obtained, and the candidate cluster of the customer is formed. Step S3: Select a candidate solution from each customer candidate cluster and jointly optimize it with the customer visit order to construct a global joint optimization model as a variant of the generalized traveling salesman problem; calculate the vehicle distance using segment-level distance decomposition and obtain the effective vehicle travel length using a path overlap reduction strategy under the condition of multiple UAVs in parallel. Step S4: The global joint optimization model is solved using a metaheuristic algorithm, and the following results are output: customer access order, selected take-off and landing sub-segments and continuous take-off and landing positions for each customer, vehicle travel path and UAV task allocation results.

2. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 1, characterized in that, In step S1, each road segment r i Satisfying the linear approximation error condition κ i L i ≤ε, where κ i L represents the maximum curvature or upper bound of the curvature of the subsegment. i Let ε be the length of the sub-segment, and ε be the linear approximation error bound; P be the continuous position on the sub-segment. i (s) is derived from the parameterized function Where A i With B i Let s be the coordinates of the two endpoints of the sub-segment, s∈[0,L] i [] represents the segment parameter, where i represents the takeoff segment. Similarly, for the landing segment k, P has the same definition. k (t), representing the descent segment r k Parametric representation of the upper point.

3. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 1, characterized in that, In step S2, the spatiotemporal synchronization constraint of the vehicle-mounted UAV is expressed as: ∣f(s,t)-αh(s,t)∣≤Δt·v D , Where f(s,t) is the Euclidean flight distance of the UAV, h(s,t) is the shortest travel distance of the vehicle in the road network, and α = v D / v T >1 represents the speed ratio, v D For the drone's speed, v T Let Δt be the vehicle speed, and Δt be the synchronization time tolerance. In step S2, the drone's flight distance is defined as... f(s,t)=||P i (s)-c j ||+||c j -P k (t)||, Where c j Let (i, k) be the coordinates of the j-th customer point, and (i, k) be the takeoff and landing sub-segments; the vehicle travel distance is defined as... Where d G (u,v) represents the shortest path length from node u to node v on the road network.

4. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 1, characterized in that, In step S2, the local synchronization optimization model is as follows: s.t.∣f(s,t)-αh(s,t)∣≤Δt·v D ,f(s,t)≤E max Where E max This represents the maximum flight range of the drone.

5. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 2, characterized in that, In step S2, before solving the local synchronous optimization model, candidate segment pairs are pre-screened using at least the following three types of pruning strategies: (1) Reduce the remaining range by pruning: If d i +d k >E max Then remove the corresponding sub-segment pair; (2) Synchronous feasibility pruning: Define g(s,t)=f(s,t)-αh(s,t), if min g(s,t)·max g(s,t)>0 (min and maximum values ​​have the same sign), then remove the sub-segment pair; (3) Directional Consistency Pruning: If the takeoff segment direction vector d i = (B i -A i ) and the landing segment direction vector d k = (B k -A k If two sub-segments are in the same direction and their relative positions cause the vehicle to need to turn back, then that sub-segment pair is removed.

6. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 1, characterized in that, In step S2, the local synchronization optimization is solved using sequential least squares programming (SLSQP) or an equivalent nonlinear constrained optimization method, and the optimal synchronization solution (s) corresponding to each pair of sub-segments is recorded. * ,t * ), to form a customer candidate cluster.

7. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 1, characterized in that, In step S3, the vehicle's serial travel length is calculated through segment-level distance decomposition. Where S is a warehouse node, and π = (π1, ..., π) n () represents the customer access order; for customer π o The selected candidate schemes are denoted as follows: the takeoff segment and the landing segment are respectively denoted as The corresponding endpoint is denoted as and The corresponding sub-segment length is In step S3, the vehicle path overlap reduction strategy under multi-UAV parallel operation is implemented. Where Θ=∑ e∈E min{max(0,R e -1),(K-1)}·2L e Where E is the set of road edges, L e Let e ​​be the length of the roadside, K be the number of drones, and R be the length of the roadside. e This represents the number of times the vehicle paths of adjacent customers overlap on the same-direction road edge e.

8. The hierarchical decoupling optimization method for vehicle-mounted drone in-transit collaborative delivery according to claim 1, characterized in that, In step S3, for each customer candidate cluster Q j Select one candidate q from the options j,l The customer access order π = (π1, π2, ..., π) n Together, they form a global combinatorial optimization model, constructed as a variant of the Generalized Traveling Salesman Problem (GTSP), with the optimization objective being: Where f j,l For candidate q j,l The corresponding drone flight distance, x j,l ∈{0,1} is the selection variable, satisfying λ D and λ T Cost weighting.

9. The hierarchical decoupling optimization method for vehicle-mounted drone collaborative delivery in transit according to claim 1, characterized in that, The constraints in the global joint optimization model include: (1) Customer unique service constraint: each customer selects only one candidate solution; (2) Origin and destination constraint: the path must start from the warehouse and return; (3) Sub-loop elimination constraint: sub-loops are eliminated using the Miller-Tucker-Zemlin method; (4) UAV concurrency constraint: the number of times the same road segment is reused does not exceed the number of UAVs K; (5) Synchronization constraint: the difference between the UAV flight time and the vehicle travel time does not exceed Δt; (6) Endurance constraint: the flight distance does not exceed the maximum endurance flight distance E of the UAV. max .

10. The hierarchical decoupling optimization method for vehicle-mounted drone collaborative delivery in transit according to claim 1, characterized in that, In step S4, the adaptive large neighborhood search (ALNS) metaheuristic algorithm is used to solve the problem, which includes: candidate graph construction, initial solution generation, iterative destruction-repair, candidate replacement, local search and termination output; the destruction operator includes at least one or more of random removal, worst-cost removal and relevance removal; the repair operator includes at least one or more of greedy insertion and regret value insertion; the local search includes at least 2-opt and 3-opt; and the operator is selected through an adaptive update mechanism of operator score and probability to obtain a globally approximate optimal solution; the restricted flight zone is geometrically pruned or constrained in the form of polygonal obstacles in step S2.

Images