Low-altitude-highway infrastructure collaborative layout optimization method based on ai perception

Abstract

The application relates to the cross technical field of artificial intelligence and intelligent transportation systems, and discloses a low-altitude-highway infrastructure collaborative layout optimization method based on AI perception. The method deploys hierarchical composite infrastructure nodes, fuses multi-modal perception data, constructs a dynamic interaction relationship graph by using a lightweight space-time graph neural network, and realizes online collaborative optimization of calculation, communication and perception resources by combining a distributed task offloading algorithm; the hierarchical deployment of the composite infrastructure node system sinks the calculation, communication and perception capabilities to the network edge, thereby avoiding the long-distance transmission delay of centralized cloud computing; through the cooperation of the lightweight space-time graph neural network and the distributed task offloading decision algorithm, the utilization efficiency of infrastructure resources is dynamically optimized, the overall energy consumption and operation cost are reduced while the high-priority task response is ensured.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and intelligent transportation systems, specifically involving an AI-based method for the collaborative deployment and optimization of low-altitude and highway infrastructure. Background Technology

[0002] With the deep integration of the low-altitude economy and intelligent transportation systems, emerging application scenarios such as drone logistics, aerial inspection, and collaborative operations with ground vehicles are placing unprecedented demands on communication response speed and real-time decision-making. While traditional infrastructure architectures relying on centralized cloud computing possess powerful computing capabilities, their inherent high transmission latency, bandwidth bottlenecks, and network congestion make it difficult to support millisecond-level interaction requirements. Especially during emergencies, the centralized cloud processing model, due to long data return paths and slow dispatching instructions, is highly susceptible to delayed collaborative responses, weakening the overall resilience and efficiency of the system.

[0003] The collaborative deployment of low-altitude highway infrastructure, as a key component supporting integrated air-ground operations, aims to achieve efficient integration of sensing, communication, and control resources through the joint deployment of roadside units, UAV nesting stations, and communication nodes. The core objective is to build an edge-based collaborative network capable of dynamically sensing environmental changes, rapidly generating local decisions, and executing cross-domain scheduling to meet the operational requirements of high concurrency, low latency, and high reliability.

[0004] Existing technologies mostly employ static deployment strategies, allocating computing and communication resources to pre-defined nodes, lacking the ability to dynamically perceive traffic flow, airspace activity, and sudden events. Even with the introduction of edge computing nodes in some solutions, AI models are generally large and frequently updated, making efficient deployment in resource-constrained roadside equipment or drone pods difficult. Furthermore, existing systems often rely on periodic polling mechanisms for status acquisition and command issuance, failing to achieve real-time triggering and precise response to specific events. These shortcomings result in low infrastructure resource utilization and persistently high system response latency, particularly pronounced in highly dynamic and highly coupled air-ground collaborative scenarios, necessitating an optimized distributed collaborative deployment method based on AI perception, event-driven approaches, and support for lightweight model deployment. Summary of the Invention

[0005] This invention provides an AI-based method for the collaborative deployment optimization of low-altitude highway infrastructure, aiming to solve the technical problem that low-altitude drones and ground vehicles have extremely high requirements for low-latency communication and real-time decision-making, while centralized cloud computing is unable to meet the millisecond-level response requirements.

[0006] This invention provides an AI-based method for optimizing the collaborative deployment of low-altitude highway infrastructure, comprising: Deploy composite infrastructure nodes consisting of edge computing units, multi-band communication base stations, environmental perception sensor arrays, and energy supply modules in the pre-designed low-altitude flight corridor and high-grade highway intersection area. The environmental perception sensor array collects in real time historical flight trajectory data, current three-dimensional position coordinates, velocity vector, attitude angle information, obstacle distribution map of low-altitude UAVs, as well as lane position, driving speed, acceleration, steering status, distance information and road condition data of ground vehicles; Multi-source heterogeneous data is input into a spatiotemporal alignment engine deployed on edge computing units to perform timestamp synchronization, spatial coordinate normalization, linear interpolation filling of missing data values, and three sigma threshold filtering of outliers, generating a spatiotemporal aligned data stream in a unified format. A lightweight spatiotemporal graph neural network model is used to encode the features of the spatiotemporal aligned data stream and extract the dynamic interaction relationship graph between low-altitude and ground traffic entities. The graph nodes represent the state vector of a single traffic entity, and the graph edge weights are jointly determined by the relative distance, relative speed and communication link quality between entities. Based on the dynamic interaction relationship graph, a distributed task offloading decision algorithm is run to dynamically allocate computing tasks such as sensing data processing, path replanning, and conflict early warning generation according to the remaining computing power, communication bandwidth utilization, and power reserve status of each edge computing unit. Based on the task allocation results, the communication beam direction, computing resource slicing ratio, and sensor activation strategy of each composite infrastructure node are adjusted to achieve online optimization of infrastructure deployment parameters.

[0007] Preferably, the deployment of a composite infrastructure node consisting of an edge computing unit, a multi-band communication base station, an environmental sensing sensor array, and an energy supply module in the pre-defined intersection area of ​​the low-altitude flight corridor and the high-grade highway includes: Backbone nodes are deployed along the main highway at intervals of no more than 800 meters. Each backbone node is equipped with an edge computing unit with a computing power of no less than 16 TOPS, a base station that supports dual-band communication of centimeter wave and millimeter wave, and a holographic perception kit that includes lidar, infrared thermal imager, visible light high-definition camera and meteorological sensor. Enhancement nodes are deployed at intervals of no more than 300 meters in low-altitude dense flight areas, intersections, or terrain-obscured areas. The edge computing unit has a computing power of no less than 8 TOPS. The communication base station supports the centimeter wave frequency band. The perception kit includes LiDAR and visible light high-definition camera. In emergency response scenarios, mobile emergency nodes are deployed and carried by roadside emergency vehicles. They have rapid networking capabilities, and their edge computing units have a computing power of no less than 4 TOPS. The communication base station supports the centimeter wave frequency band, and the sensing kit includes a visible light high-definition camera and a weather sensor.

[0008] Preferably, the step of inputting multi-source heterogeneous data into a spatiotemporal alignment engine deployed on an edge computing unit to perform timestamp synchronization, spatial coordinate normalization, linear interpolation imputation of missing data values, and three-sigma threshold filtering of outliers includes: Based on the timing signal of the Global Navigation Satellite System, all sensor data are labeled with time tags with nanosecond-level precision. A unified geographic coordinate system is used to convert the drone's location into geodetic coordinates, and the vehicle's location is converted into a point in the same coordinate system through high-precision map matching; For missing data points within the sampling interval, linear fitting is performed using the adjacent valid data points before and after them. Calculate the historical mean and standard deviation for each data dimension, and remove data points that deviate from the mean by more than 3 times the standard deviation.

[0009] Preferably, the lightweight spatiotemporal graph neural network model includes an input embedding layer, a spatiotemporal attention encoding layer, a graph message passing layer, and an output projection layer; The input embedding layer maps the state vector of each traffic entity to a 128-dimensional embedding vector; The spatiotemporal attention coding layer employs a multi-head self-attention mechanism to calculate the state correlation of the same entity at different time steps, as well as the spatial correlation of different entities at the same time step, generating spatiotemporal context-enhanced node representations. The graph message passing layer performs two rounds of message aggregation. In each round, each node receives the embedding vector of its first-order neighbor node, performs information fusion through the gated loop unit, and updates its own node representation. The output projection layer projects the final node representation onto a 64-dimensional interactive feature space.

[0010] Preferably, the distributed task offloading decision algorithm adopts a convex optimization solver based on the Lagrange multiplier method. Its objective function is to minimize the weighted sum of task processing delays of all edge computing units. The constraints include: the computing load of each edge computing unit is not greater than 90% of its peak computing power, the bandwidth utilization rate of the communication link is not greater than 85% of the total bandwidth, and the power consumption rate is not greater than the continuous output power of the energy supply module.

[0011] Preferably, each edge computing unit exchanges local resource status information, including remaining computing power, available bandwidth, battery power, and current task queue length, through a regional broadcast protocol.

[0012] Preferably, the adjustment of the communication beam pointing, computing resource slicing ratio, and sensor activation strategy of each composite infrastructure node includes: Based on the azimuth and elevation angles of the drones and vehicles currently in service, the phase shifter phase of the phased array antenna is controlled to ensure that the main lobe beam is precisely aligned with the target user. The central processing unit and graphics processing unit resources of the edge computing unit are divided into perception processing slices, communication scheduling slices and emergency response slices according to task priority. During periods of low traffic density, some lidar and infrared thermal imagers are turned off to reduce power consumption. When high-priority tasks or emergencies are detected, all sensing sensors are fully activated and the sampling frequency is increased.

[0013] Preferably, this also includes the offline planning phase of the infrastructure deployment scheme: A digital twin simulation environment is constructed based on historical traffic flow data, low-altitude flight permit area maps, terrain elevation models, and electromagnetic environment mapping data. In the digital twin simulation environment, the number of composite infrastructure nodes deployed is used as the decision variable, and the average communication latency, sensing coverage and construction cost of the whole domain are used as optimization objectives. A multi-objective genetic algorithm is run to generate a Pareto optimal solution set. From the Pareto optimal solution set, a specific deployment scheme is selected based on the actual fiscal budget and performance requirements; The node location coordinates, equipment configuration list, and initial parameter configuration of the selected deployment scheme are sent to the on-site construction system to complete the physical deployment.

[0014] Preferably, the individual coding of the spatiotemporal attention coding layer adopts a hybrid coding method, with the first half being the latitude and longitude coordinate sequence of the backbone node and the second half being the latitude and longitude coordinate sequence of the augmentation node; the fitness function is composed of the weighted average communication latency, sensing coverage and construction cost across the entire domain. Perception coverage is defined as the proportion of low-altitude-ground transportation entities that are at least covered by the perception range of composite infrastructure nodes at any given time.

[0015] Preferably, a safety redundancy mechanism is also included: A dedicated fiber optic direct link is established between any two adjacent backbone nodes to transmit critical control commands when wireless communication is interrupted. Each edge computing unit has a built-in dual power supply module that automatically switches to the backup power supply when the main power supply fails. All sensing data is stored locally in encrypted form, using the national cryptographic SM4 algorithm, with keys dynamically distributed by the regional key management center.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a layered and deployed composite infrastructure node system, computing, communication and sensing capabilities are pushed down to the network edge, avoiding the long-distance transmission latency of centralized cloud computing; 2. By combining a lightweight spatiotemporal graph neural network with a distributed task offloading decision algorithm, the utilization efficiency of infrastructure resources is dynamically optimized, ensuring the response of high-priority tasks while reducing overall energy consumption and operating costs. 3. Through a two-stage mechanism of offline digital twin planning and online dynamic optimization, the infrastructure deployment plan is ensured to meet both macro-level coverage requirements and micro-level scenario adaptability. 4. Through multiple safety redundancy designs, the reliability and availability of the system under extreme operating conditions are improved, providing a solid technical foundation for the deep integration of low-altitude economy and intelligent transportation. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the lightweight spatiotemporal graph neural network and the distributed task offloading decision collaboration mechanism in this invention; Figure 3 This is a logical flowchart of the hierarchical deployment of composite infrastructure nodes and the fusion of multimodal perception in this invention; Figure 4 This is a logical flowchart of the spatiotemporal alignment engine and dynamic interaction relationship graph construction in this invention; Figure 5 This is a logical flowchart of the two-stage collaborative process of offline digital twin planning and online dynamic optimization in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between low-altitude and highway infrastructure in this invention. Detailed Implementation

[0018] refer to Figures 1 to 6 This invention provides an AI-based method for the coordinated deployment and optimization of low-altitude highway infrastructure, aiming to address the technical challenge of low-latency communication and real-time decision-making requirements for low-altitude drones and ground vehicles, where centralized cloud computing struggles to meet millisecond-level response demands. This method deploys composite infrastructure nodes in physical space, possessing edge computing, multi-band communication, and multimodal perception capabilities. At the logical level, it constructs a collaborative mechanism for spatiotemporal alignment, graph neural network modeling, and distributed task offloading, achieving system performance targets of less than 10 milliseconds of communication latency, a perception update frequency of no less than 20 Hz, and a task scheduling response time of no more than 5 milliseconds.

[0019] The method includes the following steps: First, in the pre-designed area where the low-altitude flight corridor intersects with the high-grade highway, deploy composite infrastructure nodes consisting of edge computing units, multi-band communication base stations, environmental perception sensor arrays, and energy supply modules. Secondly, the environmental perception sensor array collects in real time historical flight trajectory data, current three-dimensional position coordinates, velocity vector, attitude angle information, obstacle distribution map of low-altitude UAVs, as well as lane position, driving speed, acceleration, steering status, distance information and road condition data of ground vehicles; Next, the aforementioned multi-source heterogeneous data is input into the spatiotemporal alignment engine deployed on the edge computing unit to perform timestamp synchronization, spatial coordinate normalization, linear interpolation filling of missing data values, and three sigma threshold filtering of outliers, generating a spatiotemporal aligned data stream in a unified format. Subsequently, a lightweight spatiotemporal graph neural network model is used to encode the spatiotemporal aligned data stream and extract the dynamic interaction relationship graph between low-altitude and ground traffic entities. The graph nodes represent the state vector of a single traffic entity, and the graph edge weights are jointly determined by the relative distance, relative speed and communication link quality between entities. Next, based on the dynamic interaction relationship graph, a distributed task offloading decision algorithm is run to dynamically allocate computing tasks such as sensing data processing, path replanning, and conflict early warning generation according to the remaining computing power, communication bandwidth utilization, and power reserve status of each edge computing unit. Finally, based on the task allocation results, the communication beam direction, computing resource slicing ratio, and sensor activation strategy of each composite infrastructure node are adjusted to achieve online optimization of infrastructure deployment parameters.

[0020] In the method, the steps of deploying composite infrastructure nodes specifically include: deploying backbone nodes along the main highway at intervals of no more than 800 meters, each backbone node being equipped with an edge computing unit with a computing power of no less than 16 TOPS, a base station supporting dual-band communication of centimeter wave and millimeter wave, and a holographic sensing kit including lidar, infrared thermal imager, visible light high-definition camera and meteorological sensor. In areas with high concentrations of low-altitude flights, intersections, or terrain-obscured areas, augmentation nodes are deployed at intervals of no more than 300 meters. The edge computing units have a computing power of no less than 8 TOPS. The communication base station supports the centimeter-wave frequency band. The sensing kit includes lidar and visible light high-definition cameras. In emergency response scenarios, mobile emergency nodes are deployed, carried by roadside emergency vehicles, and possess rapid network deployment capabilities. Their edge computing units have a computing power of at least 4 TOPS, and the communication base stations support centimeter-wave frequency bands. The sensing suite includes visible light high-definition cameras and weather sensors. All nodes are connected to the regional edge cloud platform via fiber optic or microwave backhaul links, forming a layered, complementary infrastructure network.

[0021] In the method described above, the step of acquiring multi-source heterogeneous data through an environmental sensing sensor array specifically includes: The lidar scans the surrounding 120-degree horizontal field of view and 30-degree vertical field of view at a frequency of 20 frames per second, and outputs point cloud data to construct a three-dimensional spatial distribution map of obstacles. Infrared thermal imagers acquire temperature distribution images at a frequency of 15 frames per second to identify moving targets at night or in low visibility conditions. Visible light high-definition cameras capture 1080-line resolution video streams at a frequency of 30 frames per second for license plate recognition, lane detection, and behavior analysis. The meteorological sensor monitors wind speed, wind direction, temperature, humidity, air pressure and precipitation intensity in real time, with a sampling frequency of once per second; Global Navigation Satellite System (GNSS) receivers provide nanosecond-level timing signals and centimeter-level positioning accuracy for the unification of spatiotemporal references; The inertial measurement unit outputs triaxial acceleration, angular velocity, and attitude angle, with an update frequency of 100 Hz. All sensor data is transmitted to the input buffer of the edge computing unit via a high-speed serial bus, awaiting further processing.

[0022] In the method, the spatiotemporal alignment engine performs data preprocessing steps, specifically including: Using the timing signal from the Global Navigation Satellite System as a reference, all sensor data are time-stamped with nanosecond-level precision to ensure that data from different sampling frequencies are aligned on the time axis. By adopting a unified geographic coordinate system, the location of the UAV is converted into geodetic coordinates through geocentric coordinates, and the location of the vehicle is projected to a two-dimensional plane point under the same coordinate system through a high-precision map matching algorithm, thereby achieving spatial coordinate normalization. For missing data points within the sampling interval, linear fitting is performed on the adjacent valid data points before and after them to complete the linear interpolation filling of missing values, and the interpolation error is controlled within the original accuracy range of the sensor. Calculate the historical mean and standard deviation for each data dimension, remove data points that deviate from the mean by more than 3 times the standard deviation, and complete the outlier three-sigma threshold filtering.

[0023] After the above processing, a spatiotemporally aligned data stream with a unified format is generated. Its data structure is a sequence of tuples containing timestamps, entity identifiers, position coordinates, velocity vectors, attitude angles, perception confidence, and sensor type identifiers. This data is stored in the shared memory pool of the edge computing unit for subsequent model calls.

[0024] In the method described above, the step of feature encoding using a lightweight spatiotemporal graph neural network model specifically includes: The input embedding layer maps the state vector of each traffic entity to a 128-dimensional embedding vector. The state vector includes position coordinates, velocity vector, acceleration, attitude angle, historical trajectory fragments, and perception confidence. The spatiotemporal attention encoding layer adopts a multi-head self-attention mechanism to calculate the state correlation of the same entity over 10 consecutive time steps, as well as the spatial correlation of different entities at the same time step, generating spatiotemporal context-enhanced node representations. The number of attention heads is 8, and the dimension of each head is 16. The graph message passing layer performs two rounds of message aggregation. In each round, each node receives the embedding vector of its first-order neighbor node. The neighbor is defined as other entities with an Euclidean distance of less than 50 meters. The information is fused through the gated recurrent unit to update its own node representation. The hidden state dimension of the gated recurrent unit is 128. The output projection layer projects the final node representation onto a 64-dimensional interactive feature space through a fully connected layer for subsequent task decision-making.

[0025] The total number of parameters of the lightweight spatiotemporal graph neural network model is controlled within 1.5 megabytes, the inference latency is no more than 3 milliseconds, and it runs continuously at a frequency of 20 times per second on the edge computing unit.

[0026] In the method, the steps of running the distributed task offloading decision algorithm specifically include: constructing an optimization problem, the objective function of which is to minimize the weighted sum of task processing delays of all edge computing units, with weights set according to task type, the weight of perception data processing being 0.6, the weight of path replanning being 0.3, and the weight of conflict warning generation being 0.1; The constraints include: the computing load of each edge computing unit is no more than 90% of its peak computing power, the bandwidth utilization of the communication link is no more than 85% of the total bandwidth, and the power consumption rate is no more than the continuous output power of the energy supply module; a convex optimization solver based on the Lagrange multiplier method is used, which iterates once every 200 milliseconds and outputs the target unloading node identifier of each computing task.

[0027] During the solution process, each edge computing unit exchanges local resource status information, including remaining computing power, available bandwidth, battery level, and current task queue length, through a regional broadcast protocol to ensure global consistency in decision-making. Task unloading commands are sent to the target node via a control channel, with an execution latency of no more than 5 milliseconds.

[0028] In the method, the steps of adjusting the infrastructure deployment parameters specifically include: controlling the phase shifter phase of the phased array antenna according to the azimuth and elevation angles of the currently serving drones and vehicles, so that the main lobe beam is accurately aligned with the target user, the side lobe suppression ratio is not less than 20 dB, and the beam switching time is not greater than 1 millisecond; The central processing unit and graphics processing unit resources of the edge computing unit are divided into perception processing slices, communication scheduling slices and emergency response slices according to task priority. The perception processing slice accounts for no less than 60% and is used to run graph neural networks and obstacle detection algorithms. The communication scheduling slice accounts for no less than 25% and is used to perform task offloading and beam management. The emergency response slice reserves no less than 15% for handling sudden conflict warnings or emergency obstacle avoidance commands. During periods of low traffic density, some lidar and infrared thermal imagers are turned off to reduce power consumption, leaving only visible light cameras and weather sensors running, with the sampling frequency reduced to 10 Hz. When a high-priority task or emergency event is detected, all sensing sensors are fully activated, and the sampling frequency is increased to 30 Hz, with a response delay of no more than 10 milliseconds.

[0029] The method also includes an offline planning phase for infrastructure deployment: First, based on historical traffic flow data, low-altitude flight permit area maps, terrain elevation models, and electromagnetic environment mapping data, a digital twin simulation environment is constructed. The simulation environment includes three-dimensional geographic information, electromagnetic propagation models, traffic flow models, and a UAV flight rule base. Secondly, in the digital twin simulation environment, the number of composite infrastructure nodes deployed is used as the decision variable, and the average communication latency, perception coverage and construction cost of the whole domain are used as the optimization objectives. A multi-objective genetic algorithm is run to generate a Pareto optimal solution set. Thirdly, from the Pareto optimal solution set, a specific deployment scheme is selected according to the actual fiscal budget and performance requirements. Finally, the node location coordinates, equipment configuration list, and initial parameter configuration of the selected deployment scheme are sent to the on-site construction system to complete the physical deployment.

[0030] During the offline planning phase, the population size of the multi-objective genetic algorithm is 500, the maximum number of iterations is 200, the crossover probability is 0.9, and the mutation probability is 0.1. Individual coding adopts a hybrid coding method, with the first half being the latitude and longitude coordinate sequence of the backbone node and the second half being the latitude and longitude coordinate sequence of the augmentation node. Each coordinate is represented by a double-precision floating-point number. The fitness function is composed of three weighted sub-objectives: the weight of the average communication latency across the entire domain is 0.5, the weight of the perception coverage rate is 0.3, and the weight of the construction cost is 0.2. Perception coverage is defined as the proportion of low-altitude-ground traffic entities that are at least covered by the perception range of composite infrastructure nodes at any given time. It is calculated using Monte Carlo simulation with a simulation duration of 72 hours and a time step of 1 second.

[0031] After the algorithm converges, it outputs 30 candidate solutions on the Pareto front for decision-makers to choose from.

[0032] The method also includes a security redundancy mechanism: a dedicated optical fiber direct link with a bandwidth of not less than 10 gigabits per second is established between any two adjacent backbone nodes to transmit critical control commands when wireless communication is interrupted, with a switching delay of not more than 5 milliseconds. Each edge computing unit has a built-in dual power supply module. The main power supply is connected to the mains, and the backup power supply is a lithium iron phosphate battery pack with a capacity of not less than 5,000 watt-hours. When the main power supply fails, it will automatically switch to the backup power supply to ensure continuous operation for at least 30 minutes. All sensing data is stored locally in encrypted form, using the national cryptographic algorithm SM4, with a key length of 128 bits. The key is dynamically distributed by the regional key management center, with an update cycle of 24 hours. Key distribution is completed through a quantum key distribution channel or a physical secure medium to ensure secure key transmission.

[0033] During the operation of the method, the system continuously monitors the health status of each node, including temperature, voltage, fan speed, storage utilization, and network packet loss rate. When any indicator exceeds a preset threshold, a self-healing mechanism is triggered: if the computing unit is overheating, the resource allocation ratio of non-critical tasks is reduced and the cooling fan is activated; if the communication link quality deteriorates, the system switches to a backup frequency band or a nearby node relays the signal; if the sensing sensor fails, redundant sensors are enabled or state inference compensation is performed through a graph neural network.

[0034] All abnormal events are logged locally and uploaded to the regional operations and maintenance center for subsequent preventative maintenance.

[0035] Through the above methods, this invention achieves dynamic coordination of low-altitude highway infrastructure in terms of spatial layout, communication resource allocation, and computing task scheduling, effectively solving the technical bottleneck that centralized cloud computing cannot meet millisecond-level response requirements. In actual testing, the system's end-to-end communication latency is stable within 8 milliseconds, the sensing update frequency reaches 25 Hz, and the task scheduling response time is 4 milliseconds, meeting the high real-time requirements of low-altitude UAVs and ground vehicles.

[0036] Meanwhile, through layered deployment and dynamic resource scheduling, overall energy consumption is reduced by 20%, construction costs are saved by 15%, and the utilization efficiency and economy of infrastructure are significantly improved.

[0037] At the system level, the composite infrastructure nodes upon which the method relies consist of edge computing units, multi-band communication base stations, environmental sensing sensor arrays, and energy supply modules.

[0038] The edge computing unit adopts a heterogeneous computing architecture, including a central processing unit, a graphics processing unit, and a neural network accelerator, and supports containerized task deployment and a real-time operating system; The multi-band communication base station integrates centimeter wave and millimeter wave radio frequency front-ends, supports beamforming and multi-user multiple-input multiple-output technology, and has a downlink peak rate of not less than 2000 megabits per second; The environmental sensing sensor array is connected to the edge computing unit via industrial-grade Ethernet, supporting plug-and-play and remote firmware upgrades. The power supply module includes an AC-DC converter, a battery management system, and a solar charging interface, supporting seamless switching between multiple power supply modes. All modules are enclosed in a chassis with a protection rating of at least IP65, enabling stable operation in environments ranging from -40 degrees Celsius to 70 degrees Celsius.

[0039] The data flow architecture of the system is as follows: the sensed data flows from the sensor array into the input buffer of the edge computing unit, is processed by the spatiotemporal alignment engine and then enters the feature encoding module. The generated interactive feature vector is sent to the task unloading decision module, and the decision result is sent to the communication base station and the sensor array through the control bus to form a closed-loop optimization.

[0040] The system employs a publish-subscribe communication model, with modules exchanging data via message queues to ensure loose coupling and high reliability. All data processing is completed at the edge, with only aggregated statistics and anomaly alarms uploaded to the cloud, minimizing uplink bandwidth consumption.

[0041] In summary, this invention, through the deep integration of methods and systems, constructs a low-altitude-highway collaborative infrastructure system with high real-time performance, high reliability, and high energy efficiency, providing key technical support for the large-scale development of the future low-altitude economy and intelligent transportation.

[0042] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based method for optimizing the collaborative deployment of low-altitude highway infrastructure, characterized in that: include: Deploy composite infrastructure nodes consisting of edge computing units, multi-band communication base stations, environmental perception sensor arrays, and energy supply modules in the pre-designed low-altitude flight corridor and high-grade highway intersection area. The environmental perception sensor array collects in real time historical flight trajectory data, current three-dimensional position coordinates, velocity vector, attitude angle information, obstacle distribution map of low-altitude UAVs, as well as lane position, driving speed, acceleration, steering status, distance information and road condition data of ground vehicles; Multi-source heterogeneous data is input into a spatiotemporal alignment engine deployed on edge computing units to perform timestamp synchronization, spatial coordinate normalization, linear interpolation filling of missing data values, and three sigma threshold filtering of outliers, generating a spatiotemporal aligned data stream in a unified format. A lightweight spatiotemporal graph neural network model is used to encode the features of the spatiotemporal aligned data stream, and a dynamic interaction relationship graph between low-altitude and ground traffic entities is extracted. Graph nodes represent the state vector of a single traffic entity, and the graph edge weights are jointly determined by the relative distance, relative speed and communication link quality between entities. Based on the dynamic interaction relationship graph, a distributed task offloading decision algorithm is run to dynamically allocate computing tasks such as sensing data processing, path replanning, and conflict early warning generation according to the remaining computing power, communication bandwidth utilization, and power reserve status of each edge computing unit. Based on the task allocation results, the communication beam direction, computing resource slicing ratio, and sensor activation strategy of each composite infrastructure node are adjusted to achieve online optimization of infrastructure deployment parameters.

2. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 1, characterized in that, The deployment of a composite infrastructure node consisting of edge computing units, multi-band communication base stations, environmental sensing sensor arrays, and energy supply modules in the pre-designated intersection area of ​​low-altitude flight corridors and high-grade highways includes: Backbone nodes are deployed along the main highway at intervals of no more than 800 meters. Each backbone node is equipped with an edge computing unit with a computing power of no less than 16 TOPS, a base station that supports dual-band communication of centimeter wave and millimeter wave, and a holographic perception kit that includes lidar, infrared thermal imager, visible light high-definition camera and meteorological sensor. Enhancement nodes are deployed at intervals of no more than 300 meters in low-altitude dense flight areas, intersections, or terrain-obscured areas. The edge computing unit has a computing power of no less than 8 TOPS. The communication base station supports the centimeter wave frequency band. The perception kit includes LiDAR and visible light high-definition camera. In emergency response scenarios, mobile emergency nodes are deployed and carried by roadside emergency vehicles. They have rapid networking capabilities, and their edge computing units have a computing power of no less than 4 TOPS. The communication base station supports the centimeter wave frequency band, and the sensing kit includes a visible light high-definition camera and a weather sensor.

3. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 2, characterized in that, The process of inputting multi-source heterogeneous data into a spatiotemporal alignment engine deployed on an edge computing unit to perform timestamp synchronization, spatial coordinate normalization, linear interpolation imputation of missing data values, and three-sigma threshold filtering of outliers includes: Based on the timing signal of the Global Navigation Satellite System, all sensor data are labeled with time tags with nanosecond-level precision. A unified geographic coordinate system is used to convert the drone's location into geodetic coordinates, and the vehicle's location is converted into a point in the same coordinate system through high-precision map matching; For missing data points within the sampling interval, linear fitting is performed using the adjacent valid data points before and after them. Calculate the historical mean and standard deviation for each data dimension, and remove data points that deviate from the mean by more than 3 times the standard deviation.

4. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 3, characterized in that, The lightweight spatiotemporal graph neural network model includes an input embedding layer, a spatiotemporal attention encoding layer, a graph message passing layer, and an output projection layer. The input embedding layer maps the state vector of each traffic entity to a 128-dimensional embedding vector; The spatiotemporal attention coding layer employs a multi-head self-attention mechanism to calculate the state correlation of the same entity at different time steps, as well as the spatial correlation of different entities at the same time step, generating spatiotemporal context-enhanced node representations. The graph message passing layer performs two rounds of message aggregation. In each round, each node receives the embedding vector of its first-order neighbor node, performs information fusion through the gated loop unit, and updates its own node representation. The output projection layer projects the final node representation onto a 64-dimensional interactive feature space.

5. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 4, characterized in that, The distributed task offloading decision algorithm adopts a convex optimization solver based on the Lagrange multiplier method. Its objective function is to minimize the weighted sum of task processing delays of all edge computing units. The constraints include: the computing load of each edge computing unit is no more than 90% of its peak computing power, the bandwidth utilization rate of the communication link is no more than 85% of the total bandwidth, and the power consumption rate is no more than the continuous output power of the energy supply module.

6. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 5, characterized in that, Each edge computing unit exchanges local resource status information through the regional broadcast protocol, including remaining computing power, available bandwidth, battery power, and current task queue length.

7. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 6, characterized in that, The adjustment of the communication beam pointing, computing resource slicing ratio, and sensor activation strategy for each composite infrastructure node includes: Based on the azimuth and elevation angles of the drones and vehicles currently in service, the phase shifter phase of the phased array antenna is controlled to ensure that the main lobe beam is precisely aligned with the target user. The central processing unit and graphics processing unit resources of the edge computing unit are divided into perception processing slices, communication scheduling slices and emergency response slices according to task priority. During periods of low traffic density, some lidar and infrared thermal imagers are turned off to reduce power consumption. When high-priority tasks or emergencies are detected, all sensing sensors are fully activated and the sampling frequency is increased.

8. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 7, characterized in that, It also includes the offline planning phase of infrastructure deployment: A digital twin simulation environment is constructed based on historical traffic flow data, low-altitude flight permit area maps, terrain elevation models, and electromagnetic environment mapping data. In the digital twin simulation environment, the number of composite infrastructure nodes deployed is used as the decision variable, and the average communication latency, sensing coverage and construction cost of the whole domain are used as optimization objectives. A multi-objective genetic algorithm is run to generate a Pareto optimal solution set. From the Pareto optimal solution set, a specific deployment scheme is selected based on the actual fiscal budget and performance requirements; The node location coordinates, equipment configuration list, and initial parameter configuration of the selected deployment scheme are sent to the on-site construction system to complete the physical deployment.

9. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure as described in claim 8, characterized in that, The spatiotemporal attention coding layer uses a hybrid coding method for individual coding, with the first half being the latitude and longitude coordinate sequence of the backbone nodes and the second half being the latitude and longitude coordinate sequence of the augmentation nodes; the fitness function is composed of a weighted average of global average communication latency, sensing coverage, and construction cost. Perception coverage is defined as the proportion of low-altitude-ground transportation entities that are at least covered by the perception range of composite infrastructure nodes at any given time.

10. The AI-based perception-based collaborative deployment optimization method for low-altitude highway infrastructure according to claim 9, characterized in that, It also includes a safety redundancy mechanism: A dedicated fiber optic direct link is established between any two adjacent backbone nodes to transmit critical control commands when wireless communication is interrupted. Each edge computing unit has a built-in dual power supply module that automatically switches to the backup power supply when the main power supply fails. All sensing data is stored locally in encrypted form, using the national cryptographic SM4 algorithm, with keys dynamically distributed by the regional key management center.

Images