Air, land and sea unmanned vehicle logistics scheduling method and device

Abstract

The application provides a kind of sea land air unmanned vehicle logistics scheduling method and device, the task scheduling module receives logistics task request in the device, based on task attribute, geographical constraint and real-time hardware resource state Generation unmanned vehicle's cooperative scheduling scheme;Path planning module is used to generate the unified semantic map corresponding to cooperative scheduling scheme, based on unified semantic map and preset algorithm carries out path planning;Real-time monitoring and state management module collects the operating state data of unmanned vehicle corresponding to cooperative scheduling scheme, processes operating state data by rule engine and lightweight prediction model, carries out abnormal detection and dynamic scheduling according to processing result.The application can realize the unified access and scheduling of sea land air unmanned vehicle, effectively reduce the dependence on manual coordination, improve scheduling efficiency, and can combine dynamic environment to carry out path planning, improve the environmental adaptability of path and effectively deal with dynamic change and sudden abnormality, strengthen adaptability and robustness.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics system technology, and more specifically, to a method and apparatus for scheduling logistics using unmanned vehicles on land, sea, and air. Background Technology

[0002] With the rapid development of autonomous driving, the Internet of Things, and artificial intelligence technologies, unmanned transportation vehicles such as drones, unmanned vehicles, and unmanned ships have been initially applied in scenarios such as logistics distribution, regional transfer, and port operations. However, existing logistics scheduling devices still have the following technical limitations:

[0003] Modal fragmentation and lack of coordination: Existing devices are mostly designed for a single transportation mode, lacking a unified access and coordinated scheduling mechanism for unmanned vehicles at sea, on land, and in the air. Cross-modal transportation relies on manual coordination, which is inefficient and makes it difficult to achieve seamless task connection.

[0004] Static route planning and poor environmental adaptability: Existing route planning is mostly based on static maps and fails to fully integrate multi-source dynamic information such as real-time weather, traffic flow, and sudden obstacles. The planning results are often infeasible or risky in real complex environments.

[0005] Limited intelligence in decision-making: scheduling decisions rely heavily on preset rules, which cannot cope with dynamic changes and sudden anomalies, and the device has weak adaptability and robustness. Summary of the Invention

[0006] The embodiments of this application aim to provide a method and apparatus for logistics scheduling of unmanned vehicles for land, sea and air, so as to solve the problems mentioned in the background art, such as difficulty in modal coordination, poor environmental adaptability and insufficient intelligent decision-making.

[0007] To achieve the above objectives, one aspect of this application proposes a logistics scheduling device for unmanned land, sea, and air vehicles, comprising:

[0008] The task scheduling module is used to receive logistics task requests and generate a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints and real-time hardware resource status. The collaborative scheduling scheme satisfies the hardware constraints of the unmanned vehicles in terms of payload, endurance and computing power. The unmanned vehicles include drones, unmanned vehicles and unmanned boats.

[0009] The path planning module is used to integrate static geographic and dynamic environmental data to generate a unified semantic map corresponding to the collaborative scheduling scheme. Based on the unified semantic map and a preset algorithm, global path planning and dynamic local planning are performed to obtain the route information of the unmanned vehicle corresponding to the collaborative scheduling scheme. The algorithm includes multimodal A* algorithm, four-dimensional spatiotemporal RRT* algorithm, Floyd algorithm, and flow field guidance algorithm.

[0010] The real-time monitoring and status management module is used to collect the operating status data of the unmanned vehicles corresponding to the collaborative scheduling scheme at a preset frequency, process the operating status data through a rule engine and a lightweight prediction model, and perform anomaly detection and dynamic scheduling based on the processing results.

[0011] Furthermore, it includes an intelligent packing module, which generates a loading scheme corresponding to the logistics task request based on a three-dimensional packing algorithm and cargo attributes. The center of gravity distribution in the loading scheme is suggested to match the hardware motion characteristics of the unmanned vehicle.

[0012] Furthermore, the method for generating a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints, and real-time hardware resource status includes:

[0013] An initial scheduling scheme is generated through a rule engine and heuristic algorithm units;

[0014] An improved non-dominated sorting genetic algorithm is used to iterate the initial scheduling scheme to generate the cooperative scheduling scheme.

[0015] Furthermore, based on the unified semantic map and the preset algorithm, global path planning and dynamic local planning are performed, including:

[0016] The multimodal A* algorithm is used to perform cross-path search and generate initial paths in the unified semantic map;

[0017] The four-dimensional spatiotemporal RRT* algorithm is used to plan the obstacle avoidance trajectory of the UAV in a dynamic environment;

[0018] In a static road network, the shortest path between all nodes is pre-calculated using Floyd's algorithm unit;

[0019] Based on ocean current and wind field data, the flow field guidance algorithm is used to plan the routes of unmanned ships and drones;

[0020] Based on the initial path, obstacle avoidance trajectory, shortest path between nodes, and route planning results generated from the flight path.

[0021] Furthermore, it includes a multi-agent collaborative module equipped with a contract network protocol task allocation unit, a distributed consensus coordination unit, and a four-dimensional spatiotemporal conflict detection and arbitration unit.

[0022] The contract network protocol task allocation unit is used to perform distributed resource matching of sub-tasks through a bidding-evaluation mechanism after receiving bidding information. The bidding information is submitted by idle objects based on their own hardware status, remaining range, and transportation costs to assess the feasibility of the task. The idle objects include idle unmanned vehicles.

[0023] The distributed consistency coordination unit is used to perform consistency processing on the key collaborative instructions corresponding to the collaborative scheduling scheme when the network is abnormal.

[0024] The four-dimensional spatiotemporal conflict detection and arbitration unit is used to identify potential conflicts between unmanned vehicles through a trajectory prediction model, and to eliminate the potential conflicts based on the identification results.

[0025] Furthermore, it includes a monitoring and status management module equipped with a high-frequency data acquisition unit, a status caching and modeling unit, and a hybrid anomaly detection unit.

[0026] The high-frequency data acquisition unit is used to acquire the status message corresponding to the unmanned vehicle according to a preset frequency.

[0027] The state caching and modeling unit is used to cache the state messages and generate trajectory playback information and trend analysis information based on the cached state messages;

[0028] The hybrid anomaly detection unit is used to identify abnormal states based on the status message, preset threshold rules, and a lightweight LSTM prediction model.

[0029] Furthermore, it includes a data fusion and decision-making module equipped with a multi-source data spatiotemporal alignment unit, a hybrid prediction engine unit, and a decision generation unit.

[0030] The multi-source data spatiotemporal alignment unit is used to map the multi-source data corresponding to the unmanned vehicle to a unified grid reference according to the spatiotemporal kernel matching algorithm. The multi-source data includes the status message.

[0031] The hybrid prediction engine unit is used to perform static resource configuration optimization and dynamic delay risk prediction based on the multi-source data;

[0032] The decision generation unit uses the NSGA-II algorithm to find the Pareto optimal solution among cost, time, and security objectives, and outputs the final scheduling strategy through the utility function.

[0033] Furthermore, the unmanned vehicle is connected to the cloud via an edge unmanned vehicle. The unmanned vehicle is equipped with a navigation module, radar, and sensors, and an edge computing controller is also provided on the unmanned vehicle. The edge computing controller is used to preprocess the local data and sensor data of the unmanned vehicle.

[0034] The edge unmanned vehicle is equipped with an edge computing unit, the cloud is equipped with a GPU computing cluster, and the Raspberry Pi 4B edge computing controller runs a lightweight Debian device and is equipped with the ONNX Runtime / TensorFlow Lite lightweight inference framework to realize the localized operation of lightweight algorithm models and the preprocessing of hardware sensor data.

[0035] Furthermore, the step of planning the obstacle avoidance trajectory of the UAV using the four-dimensional spatiotemporal RRT* algorithm in a dynamic environment includes:

[0036] The navigation module is used to acquire the trajectory data of the UAV, and the path or resource conflict between UAVs is predicted based on the trajectory data and the four-dimensional spatiotemporal RRT* algorithm.

[0037] The task urgency level of the drone is obtained, the passage priority of the drone is determined based on the prediction results and the task urgency level, and the obstacle avoidance trajectory of the drone is planned based on the passage priority.

[0038] One aspect of this application proposes a sea-land-air unmanned logistics collaborative scheduling method for the sea-land-air unmanned vehicle logistics scheduling device described above. The method includes:

[0039] The task scheduling module receives logistics task requests and generates a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints, and real-time hardware resource status.

[0040] Static geographic and dynamic environmental data generate a unified semantic map corresponding to the collaborative scheduling scheme. Based on the unified semantic map and a preset algorithm, global path planning and dynamic local planning are performed to obtain the route information of the unmanned vehicle corresponding to the collaborative scheduling scheme. The algorithm includes multimodal A* algorithm, four-dimensional spatiotemporal RRT* algorithm, Floyd algorithm, and flow field guidance algorithm.

[0041] Operational status data is collected at a preset frequency, processed by a rule engine and a lightweight prediction model, and anomaly detection and dynamic scheduling are performed based on the processing results.

[0042] Compared to existing technologies, the land, sea, and air unmanned vehicle logistics scheduling device of the present invention includes: a task scheduling module, used to receive logistics task requests, generate a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints, and real-time hardware resource status, the collaborative scheduling scheme meeting the hardware constraints of payload, endurance, and computing power of the unmanned vehicles, the unmanned vehicles including drones, unmanned vehicles, and unmanned ships; a path planning module, used to fuse static geographic and dynamic environmental data to generate a unified semantic map corresponding to the collaborative scheduling scheme, and perform global path planning and dynamic local planning based on the unified semantic map and preset algorithms to obtain the route information of the unmanned vehicles corresponding to the collaborative scheduling scheme, the algorithms including multimodal A* algorithm, four-dimensional spatiotemporal RRT* algorithm, Floyd algorithm, and flow field guidance algorithm; and a real-time monitoring and status management module, used to collect operating status data at a preset frequency, process the operating status data through a rule engine and a lightweight prediction model, and perform anomaly detection and dynamic scheduling based on the processing results. This invention enables unified access and scheduling of unmanned vehicles on land, sea, and air, effectively reducing reliance on manual coordination, improving scheduling efficiency, and combining dynamic environment for path planning, thereby enhancing the environmental adaptability of the path and effectively responding to dynamic changes and sudden anomalies, thus strengthening adaptability and robustness. Attached Figure Description

[0043] Figure 1 A structural diagram of the unmanned land, sea and air logistics scheduling device provided in the embodiments of this application;

[0044] Figure 2 A flowchart illustrating the workflow of the land, sea, and air unmanned vehicle logistics scheduling device provided in this application embodiment;

[0045] Figure 3 A working sequence diagram of the unmanned land, sea and air logistics scheduling device provided in the embodiments of this application;

[0046] Figure 4 Example diagram of the visualization interface of Wanshan Hailian Intelligent Dispatch Center provided in the embodiments of this application;

[0047] Figure 5 A flowchart of a logistics scheduling method for unmanned land, sea and air vehicles provided in an embodiment of this application. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description is provided in conjunction with the accompanying drawings. Figures 1-5 The present invention will be further described in detail with reference to specific embodiments. The implementation of the unmanned land, sea and air logistics scheduling device of the present invention includes the hardware deployment environment, software architecture implementation and the specific development and configuration of each core module.

[0049] In this embodiment, the hardware layer of the scheduling device of the present invention uses unmanned vehicles such as drones, unmanned vehicles, and unmanned ships as the basic motion carriers. A Raspberry Pi 4B edge computing controller is standardized and integrated on the basic motion carriers. For drones, unmanned vehicles, and unmanned ships, Beidou-3 / GPS / IMU integrated navigation modules, lidar, millimeter-wave radar, AIS ship automatic identification system and logistics and scene-specific sensors are deployed differently. At the same time, a three-level hardware architecture of end-edge-cloud is constructed. The end side is the unmanned vehicle integrating the above-mentioned devices. The edge side is equipped with Intel NUC edge computing units, FPGA heterogeneous acceleration cards, 5G / Wi-Fi 6 communication base stations, unmanned vehicle battery swapping / charging connection platforms and cross-modal cargo handover platforms. The cloud side is equipped with GPU computing power clusters, distributed storage servers and spatiotemporal big data processing servers. The Raspberry Pi 4B collects the device motion status and basic sensing data from the original control board, merges it with the data from the newly added sensors, performs preprocessing, and uploads it to the edge or cloud. At the same time, it converts the scheduling and path planning instructions issued by the edge or cloud into PWM / serial port signals that the original control board can recognize, driving the basic motion carrier to perform corresponding actions.

[0050] In this embodiment, the land, sea and air unmanned vehicle logistics scheduling device may include, at the software level, a task scheduling module, an intelligent packing module, a path planning module, a multi-agent collaboration module, a real-time monitoring and status management module, a data fusion and decision support module, a data storage module, a communication interaction module, and a safety fault tolerance management module. These modules enable unmanned vehicle scheduling, path planning, communication, monitoring, and management operations.

[0051] The cloud-based system can deploy a high-performance Linux server cluster. Computationally intensive modules such as task scheduling, data fusion, and decision support can be deployed in the cloud (e.g., using Alibaba Cloud ECS GPU computing instances, with each node configured with 8 NVIDIA A100 40GB GPUs, interconnected via NVLink, and nodes interconnected using 100Gbps InfiniBand). Distributed data storage can utilize a Ceph architecture, configured with SSD caching (2TB NVMe SSD per node) and an HDD capacity pool (12TB HDD per node), with up to three data replicas to ensure high availability. The spatiotemporal big data processing server can use a Huawei FusionServer Pro 2288H V6 (dual-socket Intel Xeon Platinum 8380, 40 cores, 80 threads, 512GB memory), equipped with PostGIS and RedisGraph for spatiotemporal indexing acceleration.

[0052] Deploy 5G private networks or high-speed Wi-Fi 6 access points at key logistics nodes such as warehouses, ports, and transfer points to ensure high-bandwidth, low-latency communication between unmanned vehicles and devices. Microservices within the devices are interconnected via gigabit / 10-gigabit internal networks.

[0053] In this embodiment, the hardware configuration of the edge computing node can be:

[0054] (1) The edge computing unit uses an Intel NUC 11 Pro Kit (model NUC11TNKi7), equipped with a Core i7-1165G7 processor (4 cores and 8 threads, up to 4.7GHz), 32GB DDR4 memory, 1TB NVMe SSD, running UbuntuServer 20.04, and deploying Docker containerized microservices.

[0055] (2) The FPGA heterogeneous acceleration card connects to a Xilinx Alveo U250 acceleration card (or Intel Arria 10 GX FPGA) via a PCIe 3.0 x4 interface for hardware-accelerated computationally intensive tasks. The FPGA internally consists of multiple parallel pipeline units, each responsible for:

[0056] Collision detection using the 4D spatiotemporal RRT* algorithm: After rasterizing the obstacles, the spatiotemporal intersection of thousands of sampling points with the obstacles is detected in parallel with a latency of <5ms;

[0057] Heuristic computation of the multimodal A* algorithm: heuristic cost of parallel computation of multiple nodes;

[0058] Real-time point cloud preprocessing: Filtering, downsampling, and coordinate transformation of the lidar point cloud.

[0059] (3) Communication base stations: Deploy 5G indoor micro base stations (such as Huawei AtomCell) and Wi-Fi 6 access points (such as TP-Link AX11000). The base stations are connected to the edge computing units via optical fiber, providing high-bandwidth (5G downlink >1Gbps) and low-latency (<10ms) communication with a coverage radius of 200 meters. The base stations support MIMO and beamforming to ensure concurrent access by multiple unmanned vehicles.

[0060] (4) Battery swapping / charging connection platform:

[0061] Drone battery swapping platform: Equipped with a six-degree-of-freedom robotic arm (such as UR10e), a visual positioning camera (Keyence CV-X series) and a lidar (SICK LMS series), with a positioning accuracy of ±2cm, enabling drones to swap batteries in 5 minutes;

[0062] Unmanned vehicle / boat fast charging platform: adopts high-power DC fast charging (400V voltage, 200A current), the charging interface complies with GB / T 20234 standard, and is equipped with mechanical locking and waterproof design;

[0063] Cross-modal cargo handover platform: integrates high-precision UWB positioning tags (Decawave DWM1000) and electronic locks, supports precise cargo handover between drones and unmanned vehicles, and between unmanned ships and drones, with a positioning accuracy of ±2cm. The handover process is coordinated and controlled by the edge computing unit.

[0064] Unmanned vehicle adaptation: Equip drones (such as DJI Matrice 300 RTK), unmanned vehicles (such as Baidu Apollo chassis), and unmanned boats (such as Yunzhou Intelligent M80) with standardized airborne control units (OBU). This unit integrates a 4G / 5G communication module, a high-precision GPS / BeiDou positioning module, and pre-installs agent software compatible with the communication interaction module.

[0065] Hardware configuration for unmanned vehicles: Each type of unmanned vehicle is equipped with a unified edge computing controller (such as a Raspberry Pi 4B, running a lightweight Debian system, with a main frequency of 1.5GHz and 4GB of LPDDR4 memory), and unmanned vehicle agents and ONNXRuntime / TensorFlow Lite lightweight inference engines are deployed.

[0066] In this embodiment, the sensor configuration and connection method of the unmanned vehicle are as follows:

[0067] (1) Positioning: It integrates a Beidou-3 / GPS dual-mode positioning module (such as the Hexin Xingtong UM980) and an industrial-grade IMU (such as the ADI ADIS16470), connects to the Raspberry Pi through the UART interface, and uses RTK differential technology to achieve centimeter-level positioning with a data update rate of 20Hz.

[0068] (2) Environmental perception:

[0069] A forward-facing 16-line LiDAR (such as the Velodyne VLP-16) is connected to the side via an Ethernet interface, with a point cloud data rate of approximately 300k points / second, and is used for obstacle detection and SLAM mapping.

[0070] A binocular vision camera (such as the Intel RealSense D435), connected via USB 3.0 interface, operates at 30fps and is used for target recognition and obstacle avoidance;

[0071] Millimeter-wave radar (such as Continental ARS 408) connects via a CAN interface, with a detection range of 0.2~250m, and is used for auxiliary sensing in adverse weather conditions.

[0072] (3) Status monitoring:

[0073] Current / voltage sensors (such as LEM LV 25-P) are connected via an I²C interface with a sampling rate of 1kHz.

[0074] Motor encoders (incremental or absolute) are connected via orthogonal encoding interfaces or SSI interfaces;

[0075] Temperature sensors (such as DS18B20) are connected via a 1-Wire interface.

[0076] (4) Communication: Integrates 4G / 5G modules (such as Quectel RM500Q) and Wi-Fi 6 modules (such as Intel AX210), transmits data through PCIe or USB 3.0 interfaces, and supports multi-link redundancy and automatic switching.

[0077] (5) Power management: A dual-redundant power supply scheme (main battery + supercapacitor backup) is adopted. The power management chip (such as TI BQ40Z50) monitors the power, voltage and temperature and reports to the Raspberry Pi through I²C. Hot-swappable power replacement is supported.

[0078] All hardware unmanned vehicles connect to the edge computing controller via CAN bus (1Mbps) or EtherCAT (100Mbps / 1Gbps) to achieve low-latency data acquisition and command issuance. The Raspberry Pi communicates bidirectionally with the original flight controller / chassis control board through a UART / CAN bridge chip (such as MCP2515) to convert high-level commands into PWM / CAN signals to drive the actuators.

[0079] 2. Software Architecture and Microservice Implementation

[0080] Development Framework and Languages: The backend is primarily developed using C++ (for high-performance core algorithms such as path planning and real-time monitoring) and Python (for upper-layer business logic and machine learning model services). A microservice framework based on Flask and FastAPI is used to build RESTful APIs or gRPC interfaces for each module.

[0081] Containerization and Orchestration: All microservices are packaged as Docker images and managed using Kubernetes. Deployment configurations are defined using Helm Charts to achieve automatic scaling, rolling updates, and high availability. For example, the task scheduling module can be configured with multiple replicas (ReplicaSet) to handle peak task request volumes.

[0082] Service discovery and communication: Integrate Consul or Nacos as the service registration and discovery center. Synchronous calls between microservices are conducted via HTTP / gRPC, while asynchronous event-driven communication is implemented using Apache Kafka message queues. Events such as task status changes and exception alerts are published and subscribed to via Kafka Topics.

[0083] In this embodiment, the task scheduling module may have a task receiving and parsing interface. During operation, the task scheduling module can receive task requests in JSON format. The Pydantic library is used to verify and parse the task request data to obtain task-related data. Furthermore, it integrates with the Gaode Map / Baidu Map API service to convert the text address in the task request into latitude and longitude coordinates (geocode), thereby obtaining the start and end points of the logistics task.

[0084] In this embodiment, a collaborative scheduling scheme for unmanned vehicles is generated based on task attributes, geographical constraints, and real-time hardware resource status. This includes: generating an initial scheduling scheme using a rule engine and heuristic algorithm units; and iterating the initial scheduling scheme using an improved non-dominated sorting genetic algorithm to generate a collaborative scheduling scheme. This collaborative scheduling scheme satisfies the hard constraints of the basic motion vehicle's load, range, and computing power hardware.

[0085] In one embodiment, such as Figure 3 As shown, after receiving a logistics task request, the task scheduling module can send a resource query request to the basic motion vehicle, receive information from the basic motion vehicle based on the request, and generate a collaborative scheduling scheme based on this information. The task scheduling module can adopt a two-stage hybrid solution strategy. Specifically, it quickly generates an initial scheduling scheme through a rule engine and a heuristic algorithm unit. The initial scheduling scheme filters out solutions that do not meet the hardware constraints of the basic motion vehicle. Then, an improved non-dominated sorting genetic algorithm is used to iteratively optimize the initial scheduling scheme. Its objective function is to minimize the weighted total cost: F = α·time + β·energy consumption + γ·risk, where F is the total cost, and the weights α, β, and γ can be dynamically configured according to the urgency of the task. Specifically, in emergency medical tasks, α = 0.7, β = 0.2, and γ = 0.1; in ordinary material tasks, α = 0.3, β = 0.4, and γ = 0.3. Simultaneously, the XGBoost machine learning model is integrated to predict the completion time and failure probability of candidate scheduling schemes for scheme ranking and screening. The population size of the improved non-dominated sorting genetic algorithm can be set to 100, the number of iterations to 200, the crossover probability to 0.8, and the mutation probability to 0.1. The algorithm takes less than 50ms per iteration on a cloud-side GPU computing cluster.

[0086] Optionally, the rule engine configuration can be: using the Drools rule engine to write DRL rule files for modal filtering. For example, define the rule: "When the cargo type is 'dangerous goods' and the mode of transport includes 'drone', remove this modality from the candidate set."

[0087] In this embodiment, during task scheduling, the resource status of unmanned vehicles can be queried to obtain unmanned vehicles available for scheduling. Specifically, the gRPC client can be used to call the unmanned vehicle status query interface provided by the real-time monitoring and status management module to obtain a DeviceStatus snapshot of all unmanned vehicles. This DeviceStatus snapshot includes the device... i d, type, position, battery l evel, task s tatus, estimated f ree t Fields such as ime are used to obtain the current status information of the unmanned vehicle.

[0088] In one embodiment, a combined solution using an improved non-dominated sorting genetic algorithm (NSGA-II) and an ant colony algorithm can be employed. Specifically, decision variables can be defined as binary variables, representing whether a segment of a task is performed by a particular unmanned vehicle. There are three optimization objectives: minimizing total transportation time, minimizing total energy consumption, and minimizing the comprehensive risk score. The total transportation time is calculated from the path length and estimated speed of each segment; total energy consumption is calculated based on the unmanned vehicle power consumption model (including operational power consumption and additional energy consumption for takeoff and landing); and the comprehensive risk score is a weighted sum of weather risk, traffic congestion index, and unmanned vehicle health.

[0089] The constraints include:

[0090] Each mission segment must be assigned to one and only one unmanned vehicle;

[0091] The load of unmanned vehicles must not exceed their rated capacity;

[0092] The cumulative operating time of unmanned vehicles must not exceed their range (mileage or time limit).

[0093] The connection between task segments must meet the minimum transfer time window;

[0094] Emergency missions can prioritize resource acquisition.

[0095] For the improved NSGA-II algorithm, integer encoding can be used, with chromosome length equal to the total number of task segments and each gene locus representing the unmanned vehicle ID. Initialization: The population size is 100, with 50% of individuals generated heuristically (e.g., proximity-based allocation) and 50% randomly generated. Non-dominated sorting: Individuals are stratified based on Pareto dominance relationships. Crowding calculation: Within the same non-dominated layer, the crowding distance for each individual in each target direction is calculated; a larger distance is better.

[0096] Selection: A tournament selection (scale 2) is used, prioritizing individuals with lower levels and higher crowding.

[0097] Crossover: Simulated binary crossover is used, with a crossover probability of 0.8 and a distribution index of 20.

[0098] Mutation: Multinomial mutation was used, with a mutation probability of 0.1 and a distribution index of 20.

[0099] Improvements: Introduce adaptive crossover and mutation probabilities, which are dynamically adjusted according to the number of generations: the crossover probability decreases linearly from 0.9 to 0.5, and the mutation probability increases linearly from 0.1 to 0.2, in order to balance global exploration and local development.

[0100] Termination condition: The maximum number of iterations is 200, or there is no improvement in the Pareto front for 20 consecutive generations.

[0101] In large-scale scenarios with more than 50 tasks or more than 30 unmanned vehicles, the solution time of NSGA-II increases exponentially. In such cases, the algorithm can be automatically switched to ant colony optimization (ACO) for global near-optimal solution search. ACO complements NSGA-II and is suitable for scenarios with a large number of tasks and an explosion of combinations. ACO models the task-unmanned vehicle allocation problem as a bipartite graph, where nodes represent tasks and unmanned vehicles, edges represent allocations, and pheromones are stored on the edges. For heuristic information: the distance from the unmanned vehicle to the task starting point, estimated energy consumption, and the unmanned vehicle's health index are considered, and their reciprocals are used as heuristic values. Transition probability: the probability of an ant choosing an unmanned vehicle from a task is determined by the product of pheromone concentration and heuristic information, and is normalized. Pheromone update: after each iteration, the pheromone on the optimal solution path is enhanced, while the pheromone on all paths is volatilized, with a volatilization factor set to 0.2. Cooperative mechanism: ACO runs 10 times, generating 10 near-optimal solutions as the initial population for NSGA-II, accelerating convergence.

[0102] In one embodiment, the ant colony algorithm iterative process can be as follows: Each task-unmanned vehicle combination is treated as a path, and the pheromone concentration represents the historical success rate of that combination; heuristic information includes the distance from the unmanned vehicle to the starting point, remaining battery power, and historical on-time rate; pheromone update: after each iteration, the pheromone on the optimal scheduling path is enhanced, while pheromone on all paths is evaporated to avoid getting trapped in local optima. Output: A set of near-optimal scheduling schemes is generated as the initial population for NSGA-II, accelerating convergence.

[0103] Optionally, after obtaining the initial scheduling scheme after iteration, it is determined as a candidate scheduling scheme and evaluated using the XGBoost machine learning model. The XGBoost model service receives features of the candidate scheduling schemes (including path length, weather coding, unmanned vehicle load rate, etc.) and outputs the predicted delay probability and confidence level, which are used to rank the schemes (schemes with low predicted delay probability and high confidence level are ranked first). The top two schemes can be determined as the collaborative scheduling schemes.

[0104] Optionally, after determining the collaborative scheduling scheme, information about the unmanned vehicles in the scheme can be obtained, and a loading scheme can be generated based on this information. Specifically, the intelligent packing module generates a loading scheme corresponding to the logistics task request based on a 3D packing algorithm and cargo attributes. The center of gravity distribution in the loading scheme is suggested to match the hardware motion characteristics of the unmanned vehicles.

[0105] In this embodiment, the intelligent packing module integrates heuristic algorithms and deep reinforcement learning. Combining cargo attributes such as cargo size, weight, fragility, and stacking limitations, it generates the optimal placement scheme for cargo in the cargo hold of the unmanned vehicle. The module outputs a loading scheme that includes cargo placement position, loading order, and center of gravity distribution suggestions. The center of gravity distribution suggestions meet hardware constraints such as UAV center of gravity offset ≤5%, unmanned vehicle center of gravity height ≤ 1 / 2 of vehicle height, and unmanned vessel draft depth ≤ 80% of equipment rated draft depth, ensuring transportation safety and maximizing space utilization. The packing scheme is sent as a dynamic constraint to the path planning module, which adjusts the path planning strategy according to the constraints.

[0106] In this embodiment, the intelligent packing module uses a C++ implementation of a 3D packing algorithm library, supporting various packing strategies (such as first-fit decreasing, depth-first search, and reinforcement learning). Cargo attributes include length, width, height, weight, stacking limits, and fragile markings. This module can output the 3D coordinates and orientation of the cargo in the cargo hold and display the packing effect through a visualization module.

[0107] In this embodiment, the linkage mechanism between intelligent packing and route planning can be as follows:

[0108] (1) The intelligent packing module outputs the three-dimensional coordinates and center of gravity position of the cargo in the cargo hold;

[0109] (2) This information is sent to the path planning module via Kafka and added to the planning model as a "dynamic constraint":

[0110] For drones: center of gravity shift will affect flight attitude stability, and path planning needs to limit turning radius and speed;

[0111] For autonomous vehicles: center of gravity height affects the risk of rollover, and path planning needs to avoid steep slopes and sharp bends;

[0112] For unmanned vessels: cargo distribution affects draft and stability, and path planning needs to take into account shallow water areas.

[0113] (3) If the packing scheme makes a certain path infeasible (e.g., the center of gravity is too high to cross the bridge), trigger the "repacking-replanning" iterative optimization until a feasible solution is found.

[0114] In this embodiment, global path planning and dynamic local planning are performed based on a unified semantic map and a preset algorithm, including: using a multimodal A* algorithm to perform cross-path search in the unified semantic map to generate an initial path; using the four-dimensional spatiotemporal RRT* algorithm to plan the obstacle avoidance trajectory of the UAV in a dynamic environment; using Floyd algorithm units to pre-calculate the shortest path between all nodes in a static road network; planning the routes of the unmanned surface vessel and the UAV based on ocean current, wind field data, and flow field guidance algorithms; and generating route planning results based on the initial path, obstacle avoidance trajectory, shortest path between nodes, and route. Among them, the multimodal A* algorithm is used to support cross-domain path search of three modes—land roads, air routes, and sea routes—and their switching nodes in a unified semantic map, and bidirectional A* search is used to accelerate the algorithm. The four-dimensional spatiotemporal RRT* algorithm is used to plan obstacle avoidance trajectories for UAVs in dynamic environments, introducing a time dimension to avoid moving obstacles and areas with sudden weather changes. The sampling step size for UAVs is 0.5-1m, and the sampling step size for unmanned surface vessels is 2-3m. When the lidar detects an obstacle at a distance of <5m, the algorithm triggers dynamic replanning. The Floyd algorithm is used to pre-calculate the shortest path between all nodes in a static road network as a basic reference for dynamic planning. The calculation results are stored on a distributed storage server on the cloud side. The flow field guidance algorithm is designed for unmanned surface vessels and UAVs. Combining ocean current and wind field data, it transforms the effects of hydrodynamics into path costs to achieve energy-saving route planning. The algorithm parameters are dynamically adjusted according to the hardware rotation speed characteristics of the UAV propellers.

[0115] In this embodiment, a unified semantic map can be constructed based on a PostGIS spatial database and a Redis cache. This map stores vector road networks, electronic nautical charts, and airspace structures (GeoJSON format). Dynamic layers related to the unified semantic map (such as real-time wind field raster data and traffic event points) can be retrieved from external APIs (such as meteorological bureaus and traffic data centers) via scheduled tasks and updated in Redis.

[0116] Optionally, upon device startup, after map updates (such as the addition of new roads / waterways), or during the early morning off-peak hours, the Floyd-Warshall algorithm is run based on the static road network to calculate the shortest path (weighted by distance / time / energy consumption) between all node pairs. The calculation results are stored in a Redis cache for quick lookup of reference paths during dynamic programming, reducing the real-time computation burden. Specifically, when calculating the shortest path, the road network can be abstracted as a weighted directed graph G=(V,E), where nodes are intersections / waypoints, edges are road segments / waterways, and weights are the weighted sum of distance / time / energy consumption; the Floyd-Warshall algorithm is run to calculate the shortest path between all node pairs, and the results are stored in the Redis shortest cache. p In the hash table `ath:all`, the key is (start) i d,end i d), the value is the path node sequence and total cost; when dynamic programming (such as A, RRT) needs quick reference, this table can be directly queried as heuristic information or alternative paths. Advantage: Reduces the online computation complexity from O(n3) to O(1) query, significantly improving response speed.

[0117] In this embodiment, the implementation of the multimodal A* algorithm can be abstracted as edges in a weighted directed graph, with connection points (such as take-off and landing fields and docks) abstracted as special nodes supporting mode switching. The edge weight cost = base d istance * (1 + weather f actor + congestion f (Actor). Algorithm acceleration: A bidirectional A* search is employed, and a priority queue is implemented using Python's heapq. For large-scale regions, map pre-segmentation and hierarchical path planning based on QuadTree are implemented.

[0118] In this embodiment, the four-dimensional spatiotemporal RRT* algorithm is executed using hardware on the edge computing unit. When planning obstacle avoidance trajectories, the state space is defined as (x, y, z, t), and obstacles (such as moving birds or temporary no-fly zones) are represented as cylinders or cubes in the spatiotemporal volume. Sampling strategy: A combination of target bias sampling and Gaussian sampling is used, introducing a dynamic step size adjustment mechanism to reduce the step size in narrow regions to improve accuracy. Optimization and output: Path smoothing uses B-spline curves. The final output is a list of waypoints with timestamps [{"lat": xx,"lon": xx,"alt": xx,"t": xxx}, ...]. This algorithm is specifically designed for planning obstacle avoidance trajectories for UAVs in dynamic environments. Its runtime is an Intel NUC series edge computing unit deployed at key nodes, which integrates an FPGA heterogeneous acceleration card for hardware acceleration.

[0119] Specifically, the state consists of three-dimensional spatial coordinates and a time dimension, i.e., the state space is defined as (x, y, z, t). The control input is a velocity vector, which must meet the device's maximum velocity constraint. The spatiotemporal information of dynamic obstacles comes from two sources: first, point cloud data uploaded after real-time scanning by the lidar of the edge-side UAV and preprocessing by the Raspberry Pi; second, time-stamped abrupt meteorological regional grid data (such as wind shear and rainstorm clouds) from external meteorological services. These data are fused at the edge cells to construct a four-dimensional spatiotemporal obstacle model (e.g., moving obstacles are represented as four-dimensional cylinders). Sampling strategy: combining target bias sampling and Gaussian sampling. Sampling is performed in the target area with a certain probability, and Gaussian sampling centered on the current node is performed in the global space with the remaining probability, and the variance decreases with the number of iterations. The sampling step size of the algorithm can be dynamically adjusted according to the hardware dynamic characteristics of the UAV receiving the command. For example, for drones with high maximum speed and strong maneuverability (such as the DJI Matrice series), the sampling step size is set to 1m in open areas to quickly explore paths; for drones with heavy payloads, the step size is set to 0.5m to improve planning accuracy. This dynamic adjustment mechanism ensures the feasibility of the planned trajectory at the hardware execution level. For the most time-consuming collision detection and progressive optimization parts of the algorithm, parallel pipeline processing can be performed through FPGA heterogeneous acceleration cards, so that the obstacle avoidance response latency is strictly controlled within 200ms. When the edge-side LiDAR detects an unknown static obstacle within 5 meters ahead, or receives a warning of a moving obstacle ahead through the communication module, the Raspberry Pi controller immediately initiates a dynamic replanning request to the edge-side path planning service. This request carries the latest obstacle point cloud data, triggering the four-dimensional spatiotemporal RRT* algorithm to regenerate a smooth local obstacle avoidance path on the edge-side unit with the current pose as the starting point within 200ms, and then sends the new waypoint command to the drone flight control device through the 5G network.

[0120] In one embodiment, when planning obstacle avoidance trajectories using the four-dimensional spatiotemporal RRT* algorithm, obstacles are represented in four-dimensional space as the spatiotemporal occupied region of the bounding box of the moving object. The separation axis theorem is used to detect whether a newly generated path segment intersects with the obstacle's spatiotemporal volume: for each obstacle, if the projection intervals on any of the three spatial axes and the time axis do not overlap, it is determined to be a collision-free obstacle. This calculation is parallelized on the FPGA: the obstacle is meshed into a four-dimensional grid (resolution 0.5m × 0.5m × 0.5m × 0.1s), and each grid stores an occupancy flag. The FPGA internally consists of multiple parallel comparison units that simultaneously detect the collision state of multiple sampling points, with a single-point detection latency of less than 100 nanoseconds. Incremental optimization: after each insertion of a new node, nodes within the nearby radius are rewired, and the parent node with the lowest cost is selected. The cost function comprehensively considers time and energy consumption; the time cost is directly determined by the path length, and the energy cost is obtained by the square integral of the velocity. Dynamic replanning trigger: When the edge-side LiDAR detects an obstacle within 5 meters or receives a weather warning, the Raspberry Pi sends a replanning request to the edge-side path planning service via MQTT, carrying the latest obstacle point cloud. The FPGA-accelerated RRT* on the edge unit completes the replanning within 200 milliseconds, generating a smooth B-spline trajectory, and sends it to the device via 5G.

[0121] For the flow field guidance algorithm implementation: integrate ocean current / wind field grid data (NetCDF format) from ECMWF or local meteorological agencies. When planning the unmanned vessel path, the flow velocity vector v c The current is incorporated into the motion model, and the speed v e effective = v s hip + v c urrent, path cost and |v e The inverse correlation of ffective|. The downstream course is generated using the potential field method or numerical integration.

[0122] In this embodiment, the multi-agent collaboration module includes a contract network protocol task allocation unit, a distributed consistency coordination unit, and a four-dimensional spatiotemporal conflict detection and arbitration unit. The contract network protocol task allocation unit, upon receiving bidding information, performs distributed resource matching of sub-tasks through a bidding-evaluation mechanism. Bidding information is submitted by idle objects based on their hardware status, remaining range, and transportation costs to assess task feasibility; idle objects include idle unmanned vehicles. The distributed consistency coordination unit handles the consistency of key collaborative instructions corresponding to the collaborative scheduling scheme in case of network anomalies. The four-dimensional spatiotemporal conflict detection and arbitration unit identifies potential conflicts between unmanned vehicles using a trajectory prediction model and eliminates these conflicts based on the identification results. Specifically, the distributed consistency coordination unit ensures the consistency of key collaborative instructions during partial network anomalies based on a Raft protocol variant. The four-dimensional spatiotemporal conflict detection and arbitration unit identifies potential conflicts between unmanned vehicles using a trajectory prediction model and resolves them using priority scheduling, speed adjustment, or path reallocation strategies. Furthermore, the synchronization latency of collaborative communication between heterogeneous devices is ≤50ms, and conflict adjustment instructions are directly sent to the original control board of the basic motion carrier.

[0123] In one embodiment, the communication middleware deployment of the multi-agent collaboration module adopts ROS 2 Galactic version, with Fast DDS as the middleware. Standard ROS 2 message types (e.g., DeviceState.msg, TaskAssignment.msg) and service types (e.g., BidForTask.srv) are defined for each type of unmanned vehicle. The Contract Network protocol implementation includes:

[0124] (1) Task Issuance: Coordinator Node n ode) publishes subtasks (such as "transport from point P1 to P2") to / task a The topic of announcement.

[0125] (2) Bidding: Idle unmanned vehicle nodes subscribe to this topic and calculate the bidding cost based on their own location and power. c ost = distance t o p pickup / remaining b attery, and via / bid s Submit a bid using the ubmission service.

[0126] (3) Bid Evaluation and Award: The coordinating node collects all bids, selects the lowest cost bidder, and awards the contract via / task. a The ward topic sends award messages.

[0127] (4) For critical states such as "handover completed", the distributed consensus coordination unit uses ROS2 based on Raft. r The AFT library implements a consensus group consisting of 3 nodes, ensuring that state changes take effect only after a majority of nodes have confirmed them.

[0128] Optionally, the four-dimensional spatiotemporal conflict detection and arbitration unit implements an independent conflict d etection s Service (Conflict Detection Service). The input is the predicted trajectories (4D point set) of all unmanned vehicles over a future period. The 4D spatiotemporal conflict detection and arbitration unit uses the Separating Axis Theorem (SAT) to perform spatiotemporal bounding box intersection detection. If a conflict is detected, an arbitration strategy (such as a rule base based on task priority) is invoked to generate adjustment suggestions (e.g., "UAV2 takeoff delayed by 10 seconds"), and these suggestions are communicated via / velocity. a Send the "djust" topic to the corresponding unmanned vehicle.

[0129] In this embodiment, the monitoring and status management module includes a high-frequency data acquisition unit, a status caching and modeling unit, and a hybrid anomaly detection unit. The high-frequency data acquisition unit is used to acquire status messages corresponding to the unmanned vehicle according to a preset frequency. The status caching and modeling unit is used to cache status messages and generate trajectory playback information and trend analysis information based on the cached status messages. The hybrid anomaly detection unit is used to identify abnormal states based on the status messages, preset threshold rules, and a lightweight LSTM prediction model. Specifically, the high-frequency data acquisition unit receives status messages at a frequency of 1Hz–10Hz through a heartbeat mechanism. The acquisition frequency is dynamically adjusted according to the sensor accuracy of the unmanned vehicle. The status caching and modeling unit uses a circular buffer to maintain the short-term historical status sequence of the unmanned vehicle, supporting real-time trajectory playback and trend analysis. The cached data comes from various sensor hardware on the edge side. The hybrid anomaly detection unit combines threshold rules and a lightweight LSTM prediction model to achieve real-time identification of multiple anomalies such as yaw, insufficient power, communication interruption, performance degradation, and hardware failure. The lightweight LSTM prediction model is integrated into the Raspberry Pi 4B edge computing controller through ONNX Runtime, with a single-step inference time of <20ms.

[0130] In one embodiment, a lightweight C++ agent runs on the unmanned vehicle's OBU, sending heartbeat packets to a designated port of the monitoring server via UDP at a configurable frequency (default 1Hz). The data packets are in TLV (Type-Length-Value) format and undergo simple XOR verification. The Raspberry Pi 4B edge computing controller on the edge collects sensor data (GPS, IMU, LiDAR, cameras, etc. on the unmanned vehicle) via the CAN bus. After processing by the edge computing unit, the data is packaged into DeviceTelemetry messages. These messages are uploaded to the cloud communication module via MQTT and simultaneously stored in a local cache. The real-time monitoring module subscribes to these messages for anomaly detection and writes the processed status data to Redis and InfluxDB. The task scheduling module, path planning module, and multi-agent collaboration module retrieve the latest unmanned vehicle status data from Redis and use this data for scheduling decisions, path updates, and conflict detection. If the edge computing unit detects an emergency obstacle, it can directly send an obstacle avoidance command (hard real-time response) to the unmanned vehicle's controller via the CAN bus and simultaneously report it to the cloud. The monitoring and status management module can also use C++ and Boost.Asio libraries to develop high-performance UDP servers. A thread pool is used to handle concurrent connections. The parsed data is converted into an internal DeviceStatus structure and performs two operations simultaneously: (1) writing to the InfluxDB device. m (2) Etrics measurement in progress; (3) Device published to Redis s tatus s Tream channel.

[0131] In one embodiment, the hybrid anomaly detection unit can integrate the open-source rule engine Esper or a self-developed rule interpreter. An example rule is: SELECT * FROM DeviceStatus WHERE battery l`evel < 20 AND type = 'UAV'` triggers a low battery alarm. Hardware-level anomaly prediction can also be performed based on a lightweight LSTM prediction model on the Raspberry Pi edge controller. To achieve rapid self-response to hardware anomalies, the lightweight LSTM prediction model is directly integrated into the Raspberry Pi 4B edge controller via the ONNX Runtime framework. The input features of the lightweight LSTM prediction model come from hardware sensors connected to the Raspberry Pi 4B edge controller via a CAN bus. A 60-dimensional circular buffer is constructed from the state cache and modeling unit to store historical data. This historical data can include: vibration signals collected by the motor encoder (0-500Hz frequency band energy features extracted by FFT transformation, 10 dimensions in total); motor winding temperature and battery temperature collected by the temperature sensor (2 dimensions); power sequence collected by the current / voltage sensor (1 dimension); and triaxial acceleration variance collected by the IMU (characterizing vibration intensity, 1 dimension). The total input feature dimension is 14, with a time step of 60, forming a tensor (60, 14).

[0132] Specifically, the LSTM network structure of the lightweight LSTM prediction model is as follows:

[0133] First layer: LSTM units, 32 hidden units, returns a sequence, dropout=0.2;

[0134] Second layer: LSTM unit, 16 hidden units, returns sequence, dropout=0.2;

[0135] The third layer is a fully connected layer with an output dimension of 5 and an activation function of sigmoid (corresponding to the failure probability in the next 5 time steps).

[0136] The total number of parameters is approximately 12,000, and the model size is approximately 150KB, making it suitable for deployment on Raspberry Pi.

[0137] For the training data of this prediction model:

[0138] Historical fault data (including cases of motor overheating, battery degradation, and communication interruption) and simulated degradation trajectories can be used to construct training samples using the sliding window method. The label is whether a fault will occur within the next 5 seconds (binarized). The ratio of training set to validation set is 8:2, and the loss function is binary cross-entropy.

[0139] Reasoning and threshold setting:

[0140] The Raspberry Pi performs inference every second, taking the features (hardware health indicator sequence) of the last 60 seconds as input and outputting the failure probability sequence for the next 5 seconds. A dual threshold is determined based on the ROC curve.

[0141] Threshold 1 (alarm threshold): 0.5. When the probability of any future time step is greater than 0.5, a "performance degradation" alarm is reported.

[0142] Threshold 2 (Emergency Threshold): 0.9. When the probability of any future time step is greater than 0.9, the Raspberry Pi immediately sends a power limit or emergency landing command to the original control board via the CAN bus and reports it to the cloud.

[0143] Single-step inference takes less than 20 milliseconds (Raspberry Pi 4B CPU), meeting real-time requirements.

[0144] Upon detecting an anomaly, a response executor can be used to respond. Depending on the anomaly level, the response executor responds through different channels. This response can include log alerts (written to the ELK stack), UI notifications (pushed to the front end via WebSocket), and automatic command delivery (by calling the ` / api / v1 / task / reschedule` interface of the task scheduling module).

[0145] In this embodiment, the data fusion and decision-making module includes a multi-source data spatiotemporal alignment unit, a hybrid prediction engine unit, and a decision generation unit. The multi-source data spatiotemporal alignment unit maps multi-source data corresponding to the unmanned vehicle to a unified grid benchmark according to the spatiotemporal kernel matching algorithm. The multi-source data includes status messages. The hybrid prediction engine unit is used to perform static resource configuration optimization and dynamic delay risk prediction based on the multi-source data. The decision generation unit uses the NSGA-II algorithm to find the Pareto optimal solution among cost, time, and safety objectives, and outputs the final scheduling strategy through a utility function. Specifically, the multi-source data spatiotemporal alignment unit uses the spatiotemporal kernel matching algorithm to map hardware sensor data and third-party platform data with different resolutions and sampling frequencies to a unified grid benchmark. The hybrid prediction engine unit integrates XGBoost and temporal fusion Transformer models, which are used for static resource configuration optimization and dynamic delay risk prediction, respectively. The decision generation unit uses the NSGA-II algorithm to find the Pareto optimal solution among cost, time, and safety objectives, and outputs the final scheduling strategy through a utility function. Moreover, the decision generation unit's single global decision-making time on the cloud-side GPU computing cluster is less than 1 second.

[0146] Specifically, the data fusion and decision support module uses Apache NiFi or Flink as the data stream orchestration tool. Multiple data source processors are configured to acquire data (GIS data, meteorological data, unmanned vehicle status, etc.). For GIS data, incremental changes are periodically pulled from PostGIS. Meteorological data is acquired by subscribing to MQTT topics (such as weather / storm / warning) from meteorological service providers. For spatiotemporal alignment, a spatiotemporal grid service is developed to divide the global region into a standard grid (such as GeoHash). All incoming data is accompanied by timestamps and geographic coordinates, and the device maps them to corresponding grid cells and time slots (such as 5-minute intervals).

[0147] The hybrid prediction engine unit uses the PyTorch Forecasting library to train a Temporal FusionTransformer model. The model's features include historical transport times, weather conditions, holidays, and the number of active unmanned vehicles in the area. After the model is service-oriented, it provides the ` / predict` function. d The `elayAPI` takes task features as input and returns a delay probability distribution. It can also optimize the model online based on prediction results. The decision generation unit runs the NSGA-II algorithm periodically (e.g., every minute). It retrieves the current queue of tasks to be scheduled and the state of the unmanned vehicles from the database, calls the Temporal Fusion Transformer model for rapid scenario evaluation, and solves for the Pareto front. The final solution is selected through a utility function (e.g., a weighted sum) and written to the planning queue.

[0148] In one embodiment, the decision generation unit is configured to seek the Pareto optimal solution among cost, time, and safety based on the NSGA-II algorithm, output a set of candidate scheduling strategies, and then combine the current task queue status and device hardware availability to evaluate the comprehensive score of each strategy through a utility function, select the optimal solution (the optimal unmanned vehicle scheduling solution), and encapsulate it into a control command for issuance. The control command is transmitted to the end-side hardware via the MQTT / UDP protocol, and the command issuance latency is ≤100ms.

[0149] In this embodiment, the communication module is deployed using MQTT Broker (for issuing commands and reporting status of unmanned vehicles) based on Eclipse Mosquitto. Critical control commands (such as emergency braking) use TCP to ensure reliability. Large data transmissions, such as video streams, use UDP.

[0150] In a preferred embodiment, the communication interaction module adopts a lightweight communication protocol to realize bidirectional data transmission between the device and various unmanned vehicles, supports edge computing node access, ensures reliable command issuance and status feedback in high-latency environments, and has encryption and authentication mechanisms for key communication links. The device end of the unmanned vehicle achieves authentication through a pre-installed TLS client certificate, and data transmission is encrypted using TLS. Furthermore, the Raspberry Pi 4B edge computing controller is connected to various sensor hardware on the edge side through CAN bus / EtherCAT, with a data transmission rate of ≥1Mbps, to achieve low-latency acquisition of sensor data.

[0151] In this embodiment, the security fault-tolerance management module provides identity authentication, access control, data encryption, and system-level fault tolerance and degradation mechanisms, covering security protection and fault handling across the three-tier hardware architecture of end-edge-cloud. Specifically, the security mechanism of the logistics device of this invention includes:

[0152] Unmanned Vehicle Authentication: Each unmanned vehicle comes pre-installed with a TLS client certificate. MQTT connections use two-way TLS authentication. For user authentication, the device front-end uses OAuth 2.0+JWT. After user login, subsequent API requests must include a Bearer Token in the header. Access Control: Based on the RBAC model, roles, permissions, and users are defined in the database. r The `oles` table. Use decorators to perform permission checks at the API layer.

[0153] The fault-tolerant design adopted includes:

[0154] Watchdog: Deploy an independent watchdog container for each critical microservice to periodically check service health and restart the Pod via the Kubernetes API if it fails.

[0155] Degradation strategy: Configure degradation rules in application.yml. For example, when the path planning service response time exceeds 2 seconds, the task scheduling module automatically degrades to using a distance-based greedy algorithm.

[0156] Transactions and Rollback: Database transactions are managed using Spring's `@Transactional` annotation. For multi-step scheduled operations, detailed operation logs are recorded. l OG supports manually triggering compensation transactions for rollback.

[0157] In one specific embodiment, the logistics scheduling device is used in the Wanshan Hailian intelligent scheduling center, and its specific workflow includes:

[0158] 1. Task Initiation and Resolution: The front-end Vue.js application calls the task scheduling module's submission API to send a task request. After back-end verification, an asynchronous task preprocessing workflow (Airflow DAG) is immediately triggered, which performs geocoding, initial risk assessment, and resource availability pre-check in parallel.

[0159] 2. Cooperative Scheduling Scheme Generation: In the objective function of the NSGA-II algorithm, α (time weight) for this task is dynamically set to 0.7 (extremely high), β (energy consumption weight) to 0.2, and γ (risk weight) to 0.1. During the XGBoost model evaluation, the feature vector includes a Boolean feature of "current sea wind speed greater than 10 m / s", which slightly increases the estimated probability of delay.

[0160] 3. Dynamic Path Planning: During path planning, the four-dimensional spatiotemporal RRT* algorithm avoids real-time wind shear areas (represented as obstacles in the spatiotemporal model) obtained from the meteorological API during sampling. The final path is optimized in the altitude dimension, selecting a flight altitude with lower wind resistance.

[0161] 4. The flow field guidance algorithm read 6-hour updated global ocean current reanalysis data (0.1° resolution) from the local cache. The algorithm calculation found that the original shortest path was against the current, while going 5 nautical miles eastward to cut into the downstream current could save 18% of the sailing time, although the distance increased. Therefore, the detour option was adopted.

[0162] 5. Multi-Agent Collaborative Execution: At the docking point P1, the ROS 2 nodes of the UAV and the unmanned surface vessel (USV) interact via a predefined HandoverAction. The UAV broadcasts its impending arrival information 500 meters before arrival, and the USV uses this information to obtain the information it is preparing to receive. After precise positioning, the UAV executes the "delivery" action via the Action Server, and the USV reports a "successful reception" result via the Action Client. The status changes throughout the process are published in real-time on a predefined topic (e.g., / handover / status), and are subscribed to by the real-time monitoring and status management module.

[0163] 6. Real-time Monitoring and Dynamic Response: The rule engine of the real-time monitoring and state management module contains a specific rule: "For unmanned vehicles such as unmanned vessels, if a high wind and wave warning is received and the navigation matching degree of its current path is less than 85%, the LEVEL2_RESCHE event is triggered." Once this event is triggered, the response executor not only replans the path but also sends an EnvironmentalDisruptionEvent message to the data fusion module via Kafka to update the risk prediction model.

[0164] 7. Task Completion and Closed-Loop Optimization: After the task is completed, the data fusion and decision support module initiates an offline reinforcement learning model training job. This job treats all states, decisions, and result sequences from "solution generation" to "execution completion" as a single stage to update the parameters of the policy network. This enables the device to generate a better initial solution when facing similar scenarios such as "crossing the sea + emergency + strong wind risk" in the future.

[0165] Based on the same inventive concept, this invention also proposes a sea-land-air unmanned logistics collaborative scheduling method, which is used in the sea-land-air unmanned vehicle logistics scheduling device as described in the above embodiments, such as... Figure 5 As shown, the method includes:

[0166] S101: Receives logistics task requests through the task scheduling module and generates a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints, and real-time hardware resource status.

[0167] S102: Static geographic and dynamic environmental data generate a unified semantic map corresponding to the collaborative scheduling scheme. Based on the unified semantic map and a preset algorithm, global path planning and dynamic local planning are performed to obtain the route information of the unmanned vehicle corresponding to the collaborative scheduling scheme. The algorithm includes multimodal A* algorithm, four-dimensional spatiotemporal RRT* algorithm, Floyd algorithm, and flow field guidance algorithm.

[0168] S103: Collects operational status data based on a preset frequency, processes the operational status data through a rule engine and a lightweight prediction model, and performs anomaly detection and dynamic scheduling based on the processing results.

[0169] Through the detailed description of the above embodiments, those skilled in the art can clearly and completely understand the technical solution of the present invention and its specific implementation method, and can realize the unmanned land, sea and air logistics scheduling method and device of the present invention on the basis of the prior art.

Claims

1. A logistics scheduling device for unmanned land, sea, and air vehicles, characterized in that, include: The task scheduling module is used to receive logistics task requests and generate a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints and real-time hardware resource status. The collaborative scheduling scheme satisfies the hardware constraints of the unmanned vehicles in terms of payload, endurance and computing power. The unmanned vehicles include drones, unmanned vehicles and unmanned boats. The path planning module is used to integrate static geographic and dynamic environmental data to generate a unified semantic map corresponding to the collaborative scheduling scheme. Based on the unified semantic map and a preset algorithm, global path planning and dynamic local planning are performed to obtain the route information of the unmanned vehicle corresponding to the collaborative scheduling scheme. The algorithm includes multimodal A* algorithm, four-dimensional spatiotemporal RRT* algorithm, Floyd algorithm, and flow field guidance algorithm. The real-time monitoring and status management module is used to collect the operating status data of the unmanned vehicles corresponding to the collaborative scheduling scheme at a preset frequency, process the operating status data through a rule engine and a lightweight prediction model, and perform anomaly detection and dynamic scheduling based on the processing results.

2. The unmanned land, sea, and air logistics scheduling device according to claim 1, characterized in that, It includes an intelligent packing module, which generates a loading plan corresponding to the logistics task request based on a three-dimensional packing algorithm and cargo attributes. The center of gravity distribution in the loading plan is suggested to match the hardware motion characteristics of the unmanned vehicle.

3. The unmanned land, sea, and air logistics scheduling device according to claim 1, characterized in that, The unmanned vehicle collaborative scheduling scheme generated based on task attributes, geographical constraints, and real-time hardware resource status includes: An initial scheduling scheme is generated through a rule engine and heuristic algorithm units; An improved non-dominated sorting genetic algorithm is used to iterate the initial scheduling scheme to generate the cooperative scheduling scheme.

4. The unmanned land, sea, and air logistics scheduling device according to claim 1, characterized in that, Global path planning and dynamic local planning based on the unified semantic map and preset algorithms include: The multimodal A* algorithm is used to perform cross-path search and generate initial paths in the unified semantic map; The four-dimensional spatiotemporal RRT* algorithm is used to plan the obstacle avoidance trajectory of the UAV in a dynamic environment; In a static road network, the shortest path between all nodes is pre-calculated using Floyd's algorithm unit; Based on ocean current and wind field data, the flow field guidance algorithm is used to plan the routes of unmanned ships and drones; Based on the initial path, obstacle avoidance trajectory, shortest path between nodes, and route planning results generated from the flight path.

5. The unmanned land, sea, and air logistics scheduling device according to claim 1, characterized in that, It includes a multi-agent collaborative module equipped with a contract network protocol task allocation unit, a distributed consistency coordination unit, and a four-dimensional spatiotemporal conflict detection and arbitration unit. The contract network protocol task allocation unit is used to perform distributed resource matching of sub-tasks through a bidding-evaluation mechanism after receiving bidding information. The bidding information is submitted by idle objects based on their own hardware status, remaining range, and transportation costs to assess the feasibility of the task. The idle objects include idle unmanned vehicles. The distributed consistency coordination unit is used to perform consistency processing on the key collaborative instructions corresponding to the collaborative scheduling scheme when the network is abnormal. The four-dimensional spatiotemporal conflict detection and arbitration unit is used to identify potential conflicts between unmanned vehicles through a trajectory prediction model, and to eliminate the potential conflicts based on the identification results.

6. The unmanned land, sea, and air logistics scheduling device according to claim 1, characterized in that, It includes a monitoring and status management module equipped with a high-frequency data acquisition unit, a status caching and modeling unit, and a hybrid anomaly detection unit. The high-frequency data acquisition unit is used to acquire the status message corresponding to the unmanned vehicle according to a preset frequency. The state caching and modeling unit is used to cache the state messages and generate trajectory playback information and trend analysis information based on the cached state messages; The hybrid anomaly detection unit is used to identify abnormal states based on the status message, preset threshold rules, and a lightweight LSTM prediction model.

7. The unmanned land, sea, and air logistics scheduling device according to claim 6, characterized in that, This includes a data fusion and decision-making module equipped with a multi-source data spatiotemporal alignment unit, a hybrid prediction engine unit, and a decision generation unit. The multi-source data spatiotemporal alignment unit is used to map the multi-source data corresponding to the unmanned vehicle to a unified grid reference according to the spatiotemporal kernel matching algorithm. The multi-source data includes the status message. The hybrid prediction engine unit is used to perform static resource configuration optimization and dynamic delay risk prediction based on the multi-source data; The decision generation unit uses the NSGA-II algorithm to find the Pareto optimal solution among cost, time, and security objectives, and outputs the final scheduling strategy through the utility function.

8. The unmanned land, sea, and air logistics scheduling device according to claim 4, characterized in that, The unmanned vehicle is connected to the cloud via an edge unmanned vehicle. The unmanned vehicle is equipped with a navigation module, radar, and sensors, and also has an edge computing controller. The edge computing controller is used to preprocess the local data and sensor data of the unmanned vehicle. The edge unmanned vehicle is equipped with an edge computing unit, the cloud is equipped with a GPU computing cluster, and the Raspberry Pi 4B edge computing controller runs a lightweight Debian device and is equipped with the ONNX Runtime / TensorFlow Lite lightweight inference framework to realize the localized operation of lightweight algorithm models and the preprocessing of hardware sensor data.

9. The unmanned land, sea, and air vehicle logistics scheduling device according to claim 8, characterized in that, The method of planning the obstacle avoidance trajectory of a UAV using the four-dimensional spatiotemporal RRT* algorithm in a dynamic environment includes: The navigation module is used to acquire the trajectory data of the UAV, and the path or resource conflict between UAVs is predicted based on the trajectory data and the four-dimensional spatiotemporal RRT* algorithm. The task urgency level of the drone is obtained, the passage priority of the drone is determined based on the prediction results and the task urgency level, and the obstacle avoidance trajectory of the drone is planned based on the passage priority.

10. A method for coordinated scheduling of unmanned logistics across land, sea, and air, characterized in that, The method for the land, sea, and air unmanned vehicle logistics scheduling device as described in any one of claims 1-8 includes: The task scheduling module receives logistics task requests and generates a collaborative scheduling scheme for unmanned vehicles based on task attributes, geographical constraints, and real-time hardware resource status. Static geographic and dynamic environmental data generate a unified semantic map corresponding to the collaborative scheduling scheme. Based on the unified semantic map and a preset algorithm, global path planning and dynamic local planning are performed to obtain the route information of the unmanned vehicle corresponding to the collaborative scheduling scheme. The algorithm includes multimodal A* algorithm, four-dimensional spatiotemporal RRT* algorithm, Floyd algorithm, and flow field guidance algorithm. Operational status data is collected based on a preset frequency, and the operational status data is processed by a rule engine and a lightweight prediction model. Anomaly detection and dynamic scheduling are performed based on the processing results.

Images