Space-air-ground integrated collaborative semantic communication and resource management method

Abstract

The application particularly relates to a space-air-ground integrated cooperative semantic communication and resource management method, which comprises the following steps: constructing a space-air-ground integrated perception-oriented cooperative semantic communication system model; establishing a probability semantic compression and recovery model, a satellite visibility and cross-layer routing model, a communication model, an execution time delay and energy consumption model and a perception utility model, and constructing a joint optimization objective function; converting the joint optimization problem into a partially observable Markov decision process; adopting a hierarchical graph guided multi-agent joint decision algorithm to jointly decide on a vehicle-unmanned aerial vehicle association, a semantic descriptor selection, an execution mode, a satellite access, a cross-layer routing and a communication and computing resource allocation, so as to obtain an optimal resource management strategy. The application solves the problems of model fragmentation and high-dimensional action solution instability in the prior art, improves the cooperative perception communication efficiency of the Internet of Vehicles, and reduces the energy consumption of the whole network.

Description

Technical Field

[0001] This invention relates to the fields of Internet of Things, wireless communication, and artificial intelligence, and specifically to a method for integrated air-space-ground collaborative semantic communication and resource management. Background Technology

[0002] As intelligent transportation systems evolve from basic vehicle-to-everything (V2X) connectivity to collaborative perception, real-time intelligent decision-making, and large-scale autonomous collaboration, single-vehicle perception, limited by sensing distance, occlusion effects, local perspectives, and rapid changes in road topology, struggles to independently obtain complete, reliable, and up-to-date environmental information. Therefore, collaborative perception and task-level information sharing between vehicles and between vehicles and external service nodes have become crucial supporting technologies for next-generation intelligent transportation systems. Meanwhile, in scenarios such as remote highways, sparsely populated suburban roads, and disaster-stricken areas, terrestrial communication infrastructure is often unstable or even absent, making it difficult to continuously provide low-latency, high-reliability perception services solely through traditional ground-based edge servers. Against this backdrop, an integrated air-space-ground network comprised of low-Earth orbit satellites, drones, and vehicles can leverage the wide-area coverage of satellites and the flexible deployment capabilities of drones to provide cross-layer communication and computing support for vehicle collaborative perception.

[0003] Existing technologies primarily employ the following service methods for collaborative sensing tasks: one is based on traditional bit-level communication for uploading and edge processing of sensing data, transmitting the raw sensing data as a whole to edge or remote nodes for processing; another is a vehicle-edge-cloud collaborative computing method based on task offloading, reducing the processing pressure on a single vehicle by selecting local or edge execution; and yet another involves introducing UAVs or satellite relays into the air-space-ground integrated network to expand the coverage and service continuity of sensing services. However, most of these existing methods address semantic compression, task offloading, cross-layer routing, or resource slicing issues separately, and have not yet formed a unified technical solution that can simultaneously consider semantic information selection, cross-layer execution paths, and multi-dimensional resource constraints for heterogeneous air-space-ground networks.

[0004] Furthermore, with the continuous growth of high-dimensional semantic information such as images, targets, and scene relationships in collaborative perception tasks, traditional bit-level transmission methods suffer from heavy communication burdens and low link resource utilization efficiency. While semantic communication can transmit only task-related information, thus reducing the transmission burden, there is a strong coupling relationship between semantic compression ratio, semantic recovery quality, and the timeliness of perception task completion: higher compression ratios result in smaller link occupancy, but also increase the reasoning difficulty required for the receiving end to recover the compressed semantics; conversely, transmitting too many semantic descriptors to ensure semantic integrity significantly increases upload latency and system energy consumption. Therefore, in integrated air-space-ground networks, achieving joint optimization among semantic fidelity, perception freshness, execution location selection, and cross-layer resource consumption has become a crucial challenge for existing technologies. The existing technology still has the following shortcomings: (1) Most existing collaborative sensing and offloading schemes are based on bit-level transmission mechanisms, which process the original sensing data as a unified data stream. They lack the ability to distinguish the importance of different semantic information and make it difficult to adaptively select the key semantic information to be sent according to the task requirements. This results in low communication efficiency in long-distance, low-bandwidth, and energy-limited air-space-ground links.

[0005] (2) Existing methods typically design semantic compression, execution location selection, satellite access, cross-layer routing, and communication and computing resource allocation separately, lacking unified modeling. Especially in scenarios where the visibility window of low-orbit satellites is limited, the access relationship between UAVs and vehicles changes dynamically, and the result return link is asymmetrical, existing technologies struggle to simultaneously address geometric visibility, latency constraints, semantic quality constraints, and resource capacity constraints.

[0006] (3) Most existing air-space-ground integrated resource management methods are optimized for single-layer or two-layer networks. They do not adequately consider the complete cross-layer path of “vehicle-UAV-access satellite-execution satellite-UAV-vehicle”, especially lacking a unified description of the separation of access satellite and execution satellite, inter-satellite forwarding and the visibility of execution satellite return links. Therefore, they are difficult to apply to dynamic collaborative sensing services in heterogeneous air-space-ground networks.

[0007] (4) Most existing resource optimization methods based on reinforcement learning adopt flat decision structures or ordinary Gaussian strategies, which are difficult to effectively handle the strong coupling relationship between discrete decisions such as execution mode, correlation, and cross-layer routing and continuous resource decisions such as bandwidth, power, and computing frequency. At the same time, for continuous resource variables that meet capacity constraints, ordinary continuous strategies are difficult to guarantee feasibility naturally, which can easily lead to training instability and difficulties in online deployment.

[0008] Therefore, there is an urgent need for a perception-oriented collaborative semantic communication and resource management method for integrated air-space-vehicle networks. This method should meet the requirements of geometric visibility, link capacity, computing resources, latency, and semantic quality, and achieve joint optimization of semantic compression, execution mode selection, cross-layer routing, and multi-dimensional resource allocation. This would improve the overall perception effectiveness of the system and reduce system energy consumption.

[0009] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] The purpose of this invention is to provide an integrated air-space-ground collaborative semantic communication and resource management method to solve the problems of semantic compression, execution mode, cross-layer routing, and communication and computing resource allocation being fragmented and difficult to optimize in the prior art. In particular, it addresses the difficulties in joint optimization of topology and resources caused by limited satellite visibility windows, asymmetric return links, and dynamic changes in cross-layer paths in heterogeneous air-space-ground networks, as well as the problems of low efficiency of traditional flat reinforcement learning in high-dimensional hybrid action spaces and difficulty in meeting strict capacity constraints for continuous resource actions.

[0011] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0012] According to a first aspect of the present invention, a method for integrated air-space-ground collaborative semantic communication and resource management is provided, the method comprising: Step S1: Construct an integrated air-space-ground perception-guided collaborative semantic communication system model; wherein, the integrated air-space-ground perception-guided collaborative semantic communication system model consists of a three-layer network structure composed of low-orbit satellites, UAVs and vehicles, the low-orbit satellites are used to provide wide-area connectivity and remote computing support, the UAVs are used to provide relay and edge-assisted services, and the vehicles are used to generate collaborative perception tasks and initiate semantic communication requests. Step S2: Based on the integrated air-space-ground perception-guided collaborative semantic communication system model in Step S1, establish a probabilistic semantic compression and recovery model, a satellite visibility and cross-layer routing model, a communication model, an execution latency and energy consumption model, and a perception utility model, and construct a joint optimization objective function; Step S3: Transform the joint optimization problem in step S2 into a partially observable Markov decision process; Step S4: For the partially observable Markov decision process established in Step S3, a multi-agent joint decision-making algorithm guided by a hierarchical graph is adopted to jointly make decisions on vehicle-UAV association, semantic descriptor selection, execution mode, satellite access, cross-layer routing, and communication and computing resource allocation to obtain the optimal resource management strategy.

[0013] In some exemplary embodiments, the probabilistic semantic compression and recovery model in step S2 includes: Extract a set of semantic descriptors from the raw vehicle perception data, share a probabilistic semantic graph between the transmitting and receiving ends, and carry a truth probability for each semantic rule; Set a 0-1 semantic sending indicator variable, calculate the task semantic compression rate based on the number of sending variables, and derive the amount of semantic data after compression. For unsent semantic descriptors, the probability recovery value of a single descriptor is calculated based on the shared semantic graph. The effective fidelity of a single descriptor is obtained by fusing the sending indicator variable and then taking a weighted average to obtain the overall semantic fidelity of the task. Establish a piecewise affine function to characterize the mapping relationship between semantic compression rate and vehicle coding workload.

[0014] In some exemplary embodiments, the satellite visibility and cross-layer routing model in step S2 includes: Set up satellite-drone visibility 0-1 indicator, vehicle-drone association 0-1 variable, and drone-satellite access 0-1 variable; The design incorporates a four-dimensional cross-layer routing 0-1 indicator to distinguish between access satellites and execution satellites, and supports modeling of inter-satellite forwarding links. Set up binary instructions for local execution and drone execution, and constrain the three execution modes to be mutually exclusive and one of them must be selected; Add topology consistency constraints and satellite return link visibility constraints, and enable the corresponding cross-layer routing path only when the geometric visibility and access association conditions are met.

[0015] In some exemplary embodiments, the communication model described in step S2 includes three types of communication links: Vehicle-to-UAV access link: Uplink and downlink transmission rates are modeled using Rician fading channel, and orthogonal pre-allocated feedback channel is used for downlink; UAV-satellite backhaul link: Introducing a zero-order Bessel function to construct a channel aging coefficient, characterizing the time-varying channel error caused by Doppler frequency shift; Inter-satellite relay links: The transmission rate of inter-satellite optical links is calculated based on the free-space optical communication model, and the rate is derived by integrating optical gain, noise temperature, and carrier loss; the one-way propagation delay of each link is defined for the calculation of the total end-to-end delay.

[0016] In some exemplary embodiments, the execution latency and energy consumption model in step S2 includes: Four types of latency and energy consumption sub-models are constructed for local execution, vehicle-side semantic encoding, UAV-side offloading execution, and satellite-side offloading execution, respectively. Calculate the energy consumption of each node based on the DVFS dynamic voltage-frequency model; The complete end-to-end total latency and total energy consumption of both drone and satellite execution modes are derived for different scenarios. The execution indicator variable is used to achieve a unified expression of latency and energy consumption for multiple modes.

[0017] In some exemplary embodiments, the perceived utility model in step S2 includes: Perceived freshness is defined based on the exponential decay function; the greater the time delay, the lower the freshness. The fusion execution mode corrects the effective semantic fidelity; By introducing task weights and urgency coefficients, we can construct single-task perceived utility. The higher the urgency coefficient, the more the task focuses on timeliness, and vice versa, the more it focuses on semantic integrity. The total perceived utility of the system is obtained by summing the utility of all tasks. Construct a global long-term joint optimization objective function with energy consumption penalty, latency default penalty, and semantic quality default penalty.

[0018] In some exemplary embodiments, the joint optimization problem described in step S3 is transformed into a partially observable Markov decision process, specifically: The intelligent agent ensemble consists of all drones and satellites; local observations of intelligent agents include four types of information: link status, local resources, geometric visibility, and local task profile; the hybrid action space is divided into high-level discrete actions and low-level continuous actions. High-level discrete actions: UAV intelligent agent outputs vehicle--UAV association, semantic descriptor sending, and task execution mode decision-making; satellite intelligent agent outputs satellite access, cross-layer routing decision-making, and masks infeasible discrete actions based on geometric visibility masking; Low-level continuous actions: The UAV outputs vehicle transmit power, encoding frequency, local computing power, vehicle bandwidth, and UAV computing power; the satellite outputs return power, inter-satellite transmit power, return bandwidth, inter-satellite bandwidth, and satellite computing frequency; the bandwidth class is bounded and summed using Dirichlet distribution modeling, and the power and frequency range resources are modeled using Beta distribution modeling, which naturally satisfies the resource capacity constraint. The global instant reward consists of five parts: system perceived utility, system energy consumption penalty, latency exceeding the standard penalty, semantic fidelity insufficient penalty, and downlink duration violation penalty.

[0019] In some exemplary embodiments, the hierarchical graph-guided multi-agent proximal policy optimization algorithm in step S4 comprises five core modules: Edge-conditional heterogeneous graph module: Vehicles, drones, and satellites are used as graph nodes. Node features include observation information, and edge features include node distance, line-of-sight probability, and visibility. An edge-conditional attention mechanism is used to aggregate cross-layer node messages and generate node spatial embeddings. Temporal Memory GRU Module: Embeds a single time-slot graph into the input gated loop unit and fuses historical time-slot states to obtain the agent's hidden state with temporal memory; Hierarchical Actor Policy Network: High-level discrete Actors output masks and make topology decisions, while low-level continuous Actors output Dirichlet and Beta distribution parameters based on discrete action conditions, and sample to obtain legal resource allocation values; Centralized Critic Value Network: Input global state and output value estimate; calculate the advantage function through temporal difference residuals and generalized advantage estimation (GAE). Multi-agent MAPPO policy update module: Constructs Actor pruning loss and Critic mean square loss respectively, adds policy entropy regularization term to improve environment exploration capability, and outputs global joint resource management policy after iterating until policy convergence.

[0020] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the integrated air-space-ground collaborative semantic communication and resource management method described in the first aspect.

[0021] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the integrated air-space-ground collaborative semantic communication and resource management method described in the first aspect above.

[0022] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the air-space-ground integrated collaborative semantic communication and resource management method described in the first aspect by executing the executable instructions.

[0023] Compared with the prior art, the present invention has the following advantages: 1. The present invention provides an integrated air-space-ground collaborative semantic communication and resource management method, which establishes a three-layer air-space-ground collaborative semantic communication and resource management model, and realizes unified modeling and joint optimization of semantic compression, semantic recovery, execution mode selection, cross-layer routing and communication and computing resource allocation.

[0024] 2. The present invention provides an integrated air-space-ground collaborative semantic communication and resource management method, which transforms the joint optimization problem into a partially observable Markov decision process, and constructs a hierarchical graph-guided multi-agent reinforcement learning algorithm to obtain the optimal collaborative semantic communication and resource management strategy that satisfies the constraints.

[0025] 3. The present invention provides an integrated air-space-ground collaborative semantic communication and resource management method that combines graph structure representation learning with multi-agent reinforcement learning, thereby improving the state representation capability, joint decision-making capability, and policy solution efficiency in complex air-space-ground heterogeneous network scenarios.

[0026] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0028] Figure 1 The flowchart of the integrated air-space-ground collaborative semantic communication and resource management method provided by this invention; Figure 2 This invention provides a model for an integrated air-space-ground perception-guided collaborative semantic communication system. Figure 3 This invention provides a schematic diagram of a partially observable Markov decision process. Detailed Implementation

[0029] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0030] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0031] To address the shortcomings and deficiencies of existing technologies, this example implementation provides a method for integrated air-space-ground collaborative semantic communication and resource management. By establishing a three-layer collaborative semantic communication system model, semantic descriptor selection, probabilistic semantic recovery, perception freshness, execution location selection, satellite access, cross-layer routing, and bandwidth, power, and computing frequency allocation are explicitly coupled. A multi-agent joint decision-making method guided by a hierarchical graph is used to solve the problem, thereby maximizing perception utility and minimizing system energy consumption while satisfying constraints on geometric visibility, link capacity, latency, and semantic fidelity.

[0032] refer to Figure 1 As shown, the specific steps may include: Step S1: Construct a model for an integrated air-space-ground perception-guided collaborative semantic communication system; Step S2: Establish probabilistic semantic compression and recovery model, satellite visibility and cross-layer routing model, communication model, execution latency and energy consumption model, and perception utility model, and construct a joint optimization objective function; Step S3: Transform the joint optimization problem in step S2 into a partially observable Markov decision process; Step S4: Employ a multi-agent joint decision-making algorithm guided by a hierarchical graph to jointly decide on vehicle-UAV association, semantic descriptor selection, execution mode, satellite access, cross-layer routing, and communication and computing resource allocation to obtain the optimal resource management strategy.

[0033] The steps in this exemplary embodiment will now be described in more detail with reference to the accompanying drawings and embodiments.

[0034] Step S1: Construct an integrated air-space-ground perception-guided collaborative semantic communication system model This invention considers a three-layer integrated space-air-ground network composed of low-Earth orbit satellites, drones, and vehicles. The space layer provides wide-area connectivity and remote computing support; the air layer provides flexible relay and edge-assisted services; and the vehicle layer generates collaborative perception tasks and initiates semantic communication requests. This system is particularly suitable for intelligent transportation scenarios with insufficient terrestrial network coverage, such as remote highways, sparsely populated suburban roads, and areas with damaged infrastructure after disasters. Specifically, as shown below... Figure 2 As shown.

[0035] Let the low-orbit satellite set be

[0036] in, This represents the set of low-Earth orbit satellites in the system; Represents the total number of low-Earth orbit satellites; elements in the set It refers to a specific satellite.

[0037] Let the set of drones be

[0038] in, This represents the collection of drones in the system; Represents the total number of drones; elements in the set This refers to a specific drone. Drones are used to establish flexible airborne access and relay services between vehicles and satellites.

[0039] Let the set of vehicles be

[0040] in, Represents the set of vehicles in the system; Represents the total number of vehicles; elements in the set This refers to a specific vehicle. To avoid ambiguity, all subsequent references to this symbol are omitted in this invention. The indicated workload refers to "vehicles". In the time slot The corresponding physical quantity of the "generated task".

[0041] The system time is divided into multiple equal-length time slots, denoted as...

[0042] in, Represents a set of discrete-time time slots; Indicates the total number of time slots; Indicates the first Each time slot has a length denoted as . Its unit can be seconds.

[0043] In any time slot Inside, each vehicle Generate a perception-guided task, the task profile of which is defined as follows:

[0044] in, Indicates vehicle In the time slot The generated perception task profile; This indicates the size of the raw sensor data corresponding to this task; This represents the computational intensity per unit bit in the processing of raw sensor data, and the unit can be CPU cycles / bit. This represents the unit bit computational strength of semantic feature processing after compression. This indicates that the task is in a time slot. Maximum tolerable delay; This indicates the importance weight of the task; This indicates the urgency level of the task. Among them, The larger the value, the more the task emphasizes the freshness of information, that is, the more it focuses on the timeliness of processing. The smaller the value, the more the task emphasizes semantic integrity, that is, it pays more attention to the fidelity of the restored semantic results.

[0045] In this invention, each task can select one of the following three execution modes: vehicle local execution mode: the vehicle processes the raw perception data directly on the local machine; UAV-side execution mode: the vehicle first completes semantic compression, and then uploads the semantic data to the UAV for execution; satellite-side execution mode: the vehicle first completes semantic compression, and then uploads the data via the UAV and satellite backhaul link before execution on the satellite side.

[0046] Therefore, all subsequent models in this invention, including the semantic compression model, association and routing model, communication model, latency model, energy consumption model, and reinforcement learning action model, revolve around the above three execution modes. Furthermore, to describe the spatial relationships between nodes in the system, the time slots for vehicles, drones, and satellites are defined. The three-dimensional positions are respectively

[0047] in, Indicates vehicle In the time slot The position vector; Indicates drone In the time slot The position vector; Indicates satellite In the time slot The position vector. The position vector can be represented by two-dimensional coordinates plus height, or directly by three-dimensional rectangular coordinates.

[0048] The Euclidean distances between vehicles and drones, drones and satellites, and satellites can be further defined based on the locations of each node. For example:

[0049] in, Indicates vehicle With drones In the time slot The distance; Indicates satellite With drones In the time slot The distance; Indicates satellite With satellite In the time slot The distance; This represents the Euclidean norm. The aforementioned distance quantities provide the foundation for subsequent geometric visibility determination, propagation delay calculation, and link rate modeling.

[0050] In the system described in this invention, since low-Earth orbit satellites move at high speeds and vehicles cannot stably and directly access the satellites, unmanned aerial vehicles (UAVs) serve as relay and edge service nodes between vehicles and satellites. Within each time slot, one UAV can serve multiple vehicles, and one UAV can access at most one satellite within the same time slot. This system configuration ensures that subsequent access associations, execution mode selection, and cross-layer routing planning have clear physical meaning and a solid foundation for implementation.

[0051] In summary, step S1 has completed the basic modeling of the integrated air-space-ground perception-guided collaborative semantic communication system, clarifying the node set, time slot structure, task profile, and spatial location representation of the three-layer network. This provides a unified system foundation for the probabilistic semantic compression, semantic recovery, satellite visibility, cross-layer routing, communication, latency, energy consumption, and perception utility modeling in step S2.

[0052] S2: Construct probabilistic semantic compression, semantic recovery, communication, latency, energy consumption, and perceived utility models. Based on the integrated air-space-ground perception-guided collaborative semantic communication system model established in step S1, this step further establishes a probabilistic semantic compression model, a semantic recovery model, a satellite visibility and cross-layer routing model, a communication model, an execution latency and system energy consumption model, and a perception utility model, thereby forming the mathematical description upon which subsequent joint optimization and reinforcement learning decisions are based. To avoid symbolic ambiguity, all subsequent symbols with subscripts will be used... The workload is represented by "vehicle" In the time slot The corresponding physical quantity of the "generated task".

[0053] S21. Constructing a probabilistic semantic compression and semantic recovery model For vehicles In the time slot The generated task extracts a set of semantic descriptors from the original perceptual stream, defined as follows:

[0054] in, Indicates vehicle In the time slot Generate a set of semantic descriptors corresponding to the task; Indicates the first element in the set. One semantic descriptor; Indicates time slot Inner vehicle The total number of semantic descriptors extracted by the task; The semantic descriptors are numbered. These semantic descriptors represent task-related semantic information such as targets, states, attributes, and contextual relationships.

[0055] To achieve probabilistic semantic compression, the sender and receiver share a probabilistic semantic graph. In this semantic graph, each recoverable semantic rule is represented as follows:

[0056] in, Indicates the relationship with the first The semantic descriptor related to the first Probabilistic semantic rules; This represents the truth probability of the rule; This represents the antecedent semantic concept of the rule; The semantic concept of the consequent of a rule; the symbol " The symbol “” indicates a semantic reasoning relationship from the antecedent to the consequent; the symbol “::” indicates that the correspondence has a probability weight. By sharing the semantic graph, the receiver can perform probabilistic recovery of the unsent semantics.

[0057] Define the semantic descriptor sending indicator variable as

[0058] in, Indicates vehicle In the time slot The first task to generate One semantic descriptor is explicitly sent; This indicates that the descriptor will not be sent, but will be recovered by the receiving end based on the shared semantic graph.

[0059] Assuming that all semantic descriptors have the same length after normalization, the semantic compression ratio of the task is defined as follows:

[0060] in, Indicates vehicle In the time slot The semantic compression rate of the generated task; This indicates the minimum compression ratio allowed for this task; This indicates the number of semantic descriptors explicitly sent. Therefore, the compression ratio is essentially determined directly by the semantic descriptor selection strategy.

[0061] Furthermore, the compressed semantic data volume is defined as follows:

[0062] in, Indicates vehicle In the time slot The amount of data generated after semantic compression; represents the original perceived data volume of the task. This formula indicates that the selection of semantic descriptors not only affects the semantic quality, but also directly affects the subsequent upload delay, calculation delay and system energy consumption.

[0063] For unsent semantic descriptors, the receiver performs probability recovery based on the shared probabilistic semantic graph. Let the set of semantic rules related to the descriptor be , then under the independent inference evidence assumption, the graph recovery probability of the th descriptor is defined as wherein, represents the recovery probability of the th semantic descriptor at time slot ; represents the set of semantic rules related to this descriptor; represents the rule number; represents the product operation. The meaning of this formula is that as long as at least one of several semantic rules related to the descriptor is successfully inferred, the semantics can be recovered.

[0064] When a descriptor is explicitly sent, its fidelity contribution is 1; when a descriptor is not sent, its fidelity contribution is determined by the graph recovery probability. Therefore, the effective recoverability of a descriptor is defined as

[0065] wherein, represents the effective fidelity contribution of the th semantic descriptor; the first term represents the corresponding fidelity contribution when explicitly sent; the second term represents the fidelity contribution brought by graph recovery when not explicitly sent.

[0066] Accordingly, the task-level semantic fidelity is defined as

[0067] wherein, represents the overall semantic fidelity of the task generated by the vehicle at time slot ; represents the effective fidelity contribution of the th semantic descriptor. This formula shows that the overall semantic quality of the task is jointly determined by the explicit transmission and probabilistic recovery of all semantic descriptors.

[0068] On the vehicle side, semantic encoding includes front-end semantic extraction and subsequent probabilistic compression. Define the semantic encoding workload as

[0069] wherein, Indicates vehicle In the time slot The semantic encoding workload; and For the first Fitting parameters for piecewise affine functions; Indicates the first Section and the Compression ratio boundary between segments; This represents the total number of segments. This model is used to characterize the impact of changes in compression ratio on coding complexity.

[0070] S22. Construct satellite visibility, correlation, and cross-layer routing models. Due to the high-speed motion of low-Earth orbit satellites, the satellite-UAV geometric visibility indicator is defined as follows:

[0071] in, Indicates in time slot Inner satellite For drones visible; This indicates that the satellite is not visible. This is typically determined by whether the satellite's elevation angle relative to the drone exceeds a preset threshold.

[0072] Define the vehicle-drone access indication as

[0073] in, Indicates vehicle In the time slot Access to drones ; This indicates that no connection has been established. Any vehicle can connect to at most one drone in any time slot, therefore satisfying the condition...

[0074] in, Represents a collection of drones; symbol " This indicates that the constraint applies to all vehicles and all time slots.

[0075] Define the UAV-satellite access correlation quantity as

[0076] in, Indicates drone In the time slot Access satellite ; This indicates no connection. To ensure geometric feasibility and unique access, the following conditions must be met:

[0077] in, This represents the set of satellites. The first formula above indicates that invisible satellites cannot be selected as access satellites; the second formula above indicates that each UAV can associate with at most one access satellite in any time slot.

[0078] To distinguish between access satellites and execution satellites, the four-dimensional cross-layer routing indication is defined as follows:

[0079] in, Indicates vehicle The mission was carried out by drones Upload to access satellite And ultimately in the execution of the satellite Upgrade processing; This indicates that the path was not used. This indicates that the mission is executed directly on the access satellite; if This indicates that the mission needs to be relayed from the access satellite to the executing satellite via an inter-satellite link.

[0080] Define the local execution instruction and the UAV execution instruction as follows:

[0081] in, Indicates vehicle The task is executed locally; Indicates vehicle The mission in drones Execution is performed locally. Therefore, any task can only be executed in one of three modes: "local execution, drone execution, or satellite execution," thus satisfying the requirement...

[0082] In this formula, the first term corresponds to local execution, the second term to drone execution, and the third term to satellite execution. The sum of the three equals 1, indicating that the three modes are mutually exclusive and one must be chosen.

[0083] To ensure consistency in execution mode, access relationships, and routing, it is also necessary to satisfy...

[0084] The first formula above indicates that only when the vehicle Access to drones Only then is its mission permitted on the drone. The above formula indicates that the satellite execution path must be established on the premise that the vehicle has been connected to the drone; the above formula indicates that the satellite execution path must be established on the premise that the drone has been connected to the satellite.

[0085] To ensure the satellite is executed Capable of sending drones Returning results directly must also satisfy the return link visibility constraint:

[0086] in, Indicates the execution of satellite With drones In the time slot The return link visibility indicator. This formula ensures that the selected satellite execution path is also physically feasible during the result return phase.

[0087] S23. Constructing a communication link model Vehicle records drones and satellite In the time slot The three-dimensional coordinates are respectively

[0088] in, , , Representing vehicles drones and satellite In the time slot The spatial position vector. Correspondingly, the vehicle-to-drone distance is denoted as... The distance between the satellite and the drone is denoted as Satellite-to-satellite distance is denoted as The return link distance from the satellite to the drone is defined as...

[0089] in, This represents the Euclidean norm.

[0090] (1) Vehicle-UAV access link The vehicle-to-UAV air-to-ground link is modeled as a Rician fading channel with an uplink transmission rate of

[0091] in, Indicates vehicle To drones The access link transmission rate; Indicates access bandwidth; Indicates vehicle The transmission power; Represents the complex channel coefficients of the vehicle-to-drone link; Indicates channel gain; Indicates the average background interference power; Indicates the noise power at the drone end; This represents the logarithmic function with base 2.

[0092] As a result, the downlink feedback link uses a pre-allocated orthogonal control channel, and its transmission rate is defined as follows:

[0093] in, Indicates drone To the vehicle The resulting downlink transmission rate; This indicates the pre-allocation of downlink bandwidth; Indicates the downlink transmission power of the drone; This indicates the noise power at the vehicle end.

[0094] (2) UAV-satellite backhaul link Considering the impact of channel aging, the effective channel for UAV-satellite communication is defined as follows:

[0095] in, Indicates time slot Internal drones With satellite Effective channel coefficients between; This indicates the estimated channel coefficients; This represents the random error term that follows a complex Gaussian distribution; This represents the channel aging coefficient.

[0096] Channel aging factor is defined as

[0097] in, Represents a zero-order Bessel function of the first kind; Indicates the maximum Doppler frequency shift; Indicates drone With satellite One-way propagation delay between them.

[0098] vehicle Mission via drone Back to satellite The transmission rate is

[0099] in, Indicates the backhaul link rate; Indicates the backhaul bandwidth; Indicates the transmitted power. This indicates the satellite-side noise power.

[0100] (3) Inter-satellite links when At that time, one-hop free-space optical communication inter-satellite forwarding is used, and its transmission rate is

[0101] in, Indicates task In satellite With satellite Inter-satellite transmission rates; Indicates inter-satellite bandwidth; Indicates inter-satellite transmission power; This represents the combined optical transceiver efficiency and antenna gain. Represents the Boltzmann constant; Indicates system noise temperature; Indicates the inter-satellite link carrier frequency; It represents the speed of light.

[0102] S24. Construct a model of execution latency and system energy consumption. (1) Local execution latency and energy consumption If the task is executed locally on the vehicle, the local processing latency is defined as follows:

[0103] in, Indicates vehicle Local execution latency; Indicates the original data volume; This represents the computational intensity per unit bit of the original data; Indicates vehicle The local processing frequency.

[0104] According to the DVFS model, local execution energy consumption is defined as...

[0105] in, Indicates vehicle Local execution energy consumption; This represents the effective capacitance coefficient of the vehicle's local processing chip.

[0106] (2) Vehicle-side semantic coding delay and energy consumption If the task uses offloading, semantic encoding is performed first. Its encoding latency is defined as...

[0107] in, Indicates semantic encoding delay; This indicates the amount of semantic encoding work; Indicates vehicle The semantic encoding frequency.

[0108] The corresponding coding energy consumption is defined as

[0109] in, Indicates vehicle The energy consumption of semantic encoding; This represents the effective capacitance coefficient of the semantic coding hardware.

[0110] (3) Basic latency and energy consumption during the execution phase of the UAV If vehicle Mission selection in drones If executed at this location, the vehicle will upload compressed semantic data to the drone. The upload latency is defined as...

[0111] in, Indicates vehicle To drones semantic upload latency; This indicates the amount of semantic data after compression; This indicates the access link speed.

[0112] The computational latency of a drone is defined as

[0113] in, Indicates the mission is in the drone computational latency; The unit bit computation strength representing the compressed semantic features; Indicates drone For the task The allocated calculation frequency.

[0114] Vehicle upload energy consumption is defined as

[0115] in, Indicates vehicle To drones The energy consumed in uploading semantic data.

[0116] The computing power consumption of drones is defined as

[0117] in, Indicates drone For the task The computational energy consumed in semantic processing; This represents the effective capacitance coefficient of the UAV-side computing chip. It should be noted that this section presents the basic uplink and computational overhead under UAV execution mode, while vehicle-side semantic encoding, downlink return transmission of results, and propagation latency will be uniformly included in the subsequent end-to-end latency model.

[0118] (4) Basic latency and energy consumption during the satellite-side execution phase If the mission is on a satellite If executed at the location, the satellite computation delay is defined as follows:

[0119] in, Indicates the execution of satellite For the task The time delay for performing calculations; Indicates the execution of satellite For the task The allocated calculation frequency.

[0120] The energy consumption for drone-to-satellite backhaul is defined as

[0121] in, Indicates task via drone Back to satellite The energy consumed.

[0122] If inter-satellite relay exists, the energy consumption for inter-satellite relay is defined as follows:

[0123] in, Indicates task From satellite Forwarded to satellite Energy consumption.

[0124] Satellite computing energy consumption is defined as

[0125] in, Indicates the execution of satellite For the task The energy consumed in performing the calculations; This represents the effective capacitance coefficient of the satellite-side computing chip.

[0126] (5) Propagation delay and result feedback The one-way propagation delay of each link is defined as follows:

[0127] in, Indicates vehicle With drones Propagation delay between; Indicates drone With satellite Propagation delay between; Indicates satellite With satellite Propagation delay between; Indicates the execution of satellite With drones Return propagation delay between; It represents the speed of light.

[0128] Let the amount of data returned be... The downlink energy consumption from the drone to the vehicle is then...

[0129] in, Indicates drone To the vehicle The downlink power consumption of the returned result; Indicates the amount of data in the result.

[0130] The energy consumption for satellite-to-drone result feedback is

[0131] in, Indicates the execution of satellite To drones The downlink power consumption of the returned result; This indicates the satellite-to-drone downlink rate.

[0132] (6) End-to-end total delay and total energy consumption If the task is executed on the drone side, its end-to-end total latency is defined as follows:

[0133] in, Indicates task The total end-to-end latency in UAV execution mode consists of five parts: semantic encoding latency, upload latency, result return latency, UAV computation latency, and bidirectional propagation latency.

[0134] If the mission is performed on the satellite side, its end-to-end total delay is defined as follows:

[0135] in, Indicates task via drone Upload to access satellite In executing satellite The total end-to-end latency for processing and returning data; Indicates an indicator function, when It is set to 1 if the signal is active and 0 otherwise, and is used to indicate whether inter-satellite relay has occurred.

[0136] Task The total end-to-end latency is uniformly written as

[0137] in, Indicates task The final total latency; the first item corresponds to the local execution latency, the second item corresponds to the UAV execution mode latency, and the third item corresponds to the satellite execution mode latency. The three parts are weighted and combined using execution mode indication and routing indication.

[0138] Task Total energy consumption is uniformly written as

[0139] in, Indicates task The total energy consumption is as follows: the first item is the local execution energy consumption; the second item is the semantic encoding energy consumption; the third item is the upload energy consumption; the fourth item is the UAV computation and result return energy consumption in UAV execution mode; and the fifth item is the backhaul, inter-satellite relay, satellite computation, and two-stage result return energy consumption in satellite execution mode.

[0140] S25. Constructing perceived novelty, perceived utility, and joint optimization objectives To describe the timeliness of perceived services, a task-level freshness function is defined as follows:

[0141] in, Indicates task Perceived novelty; Indicates task The time delay sensitivity coefficient; Indicates task The total latency. The lower the total latency of a task, the higher its freshness.

[0142] Since there is no semantic compression distortion during local execution, the effective semantic fidelity is defined as follows:

[0143] in, Indicates task Effective semantic fidelity; when the task is executed locally. At this point, the effective fidelity is 1; when the task is not executed locally, the effective fidelity is determined by the aforementioned task-level semantic fidelity. Decide.

[0144] Furthermore, the utility of perception-oriented services is defined as...

[0145] in, Indicates task The perceived service utility; Indicates the weight of task importance; Indicates the urgency level coefficient; Indicates semantic quality terms; This represents the timeliness term. This utility function reflects the task's combined preference for semantic fidelity and information freshness.

[0146] The total perceived utility of the system is

[0147] in, Indicates time slot The overall perceived utility of the system across all tasks.

[0148] Ultimately, the joint optimization objective is defined as follows: under the constraints of topology consistency, geometric visibility, link capacity, computational resources, latency, and semantic quality, to maximize the total perceived utility and minimize the total energy consumption of the system by jointly optimizing semantic descriptor selection, execution mode, satellite access, cross-layer routing, and communication / computing resource allocation. Its objective function can be written as follows:

[0149] in, Represents the set of all decision variables; Represents a set of time slots; This represents the energy consumption penalty coefficient; Indicates task Total energy consumption.

[0150] For ease of writing, the set of all decision variables is defined as follows:

[0151] in, This represents the set of all vehicle-drone related variables; Represents the set of all UAV-satellite access variables; Represents the complete set of cross-layer routing variables; Represents the set of execution mode variables; This indicates the set of variables sent by the semantic descriptor; Represents the set of power allocation variables; Represents the set of bandwidth allocation variables; This represents the set of variables used to calculate frequency allocation.

[0152] As can be seen from the above steps, the present invention has completed the complete mathematical modeling from task semantic description, probabilistic semantic compression and recovery, geometric visibility and cross-layer routing, to communication, latency, energy consumption and perception utility, laying the foundation for the abstraction of partially observable Markov decision processes in the subsequent step S3 and the hierarchical graph-guided multi-agent proximal policy optimization solution in step S4.

[0153] S3: Transform the joint optimization problem into a partially observable Markov decision process. In step S2, a joint optimization problem has been established, comprising semantic compression, execution mode, satellite access, cross-layer routing, and communication and computing resource allocation. Since this problem includes both discrete and continuous decision variables, and the network topology, link states, satellite visibility, and mission arrival process all change dynamically over time, it exhibits characteristics such as non-convexity, a high-dimensional state space, and a high-dimensional mixed action space. To achieve online, adaptive joint decision-making, this invention further transforms the joint optimization problem into a partially observable Markov decision process, as detailed in the algorithm below. Figure 3 As shown.

[0154] S31. Definition of Agent Set and Dec-POMDP: To avoid the significant signaling overhead and training complexity of treating each vehicle as an independent learning agent, this invention defines drones and satellites as active decision-making agents, while vehicles serve as contextual task nodes. Let the set of agents be... ,in, Indicates a collection of drones. Represents a collection of satellites. This represents the set of all proactive decision-making agents. The vehicle, acting as a context task node, performs semantic encoding and uploads based on instructions issued by the service drone. The corresponding partially observable Markov decision process is defined as... .in, For the global state space, For local observation space, For the action space, Let be the state transition probability. For the reward function, This is the discount factor.

[0155] It should be noted that in this invention, each UAV agent is responsible for vehicle access control, semantic descriptor selection, and execution mode decision-making within its coverage area; each satellite agent is responsible for satellite access selection, cross-layer routing control, and space layer resource allocation. Vehicles do not directly learn policies, but instead execute semantic encoding, semantic uploading, and result reception according to control commands issued by the serving UAV.

[0156] S32. State Space and Local Observation Space In any time slot, the global state of the system is denoted as:

[0157] in, Indicates time slot The overall status of the entire integrated air-space-ground network. This represents the global state space. The global state typically contains mission, link, visibility, and resource information for all vehicles, drones, and satellites.

[0158] Since each agent can only access locally available information and cannot directly observe the global state of the system, therefore, for any agent... In the time slot Local observation is defined as

[0159] in, Represents intelligent agents In the time slot The local observation vector, Represents intelligent agents The local observation space; Indicates local link status and interference characteristics; Indicates the status of local resources; Represents geometric visibility information; This indicates partially visible task profile information.

[0160] S33. Motion Space In this invention, the action space of the intelligent agent is a hybrid action space, including both discrete topology decisions and continuous resource allocation decisions. For any intelligent agent... In the time slot Action definition

[0161] in, Represents intelligent agents In the time slot Joint actions taken; Represents intelligent agents The action space; Represents a subset of discrete actions; This represents a subset of consecutive actions.

[0162] (1) High-level discrete actions For drone intelligent agents Its high-level discrete actions Includes: Vehicle-Drone Correlation Semantic descriptor sends indication quantity ; Execution mode variables, including local execution instructions and drones execute instructions When satellite execution is selected, a satellite execution request is initiated to the space layer. For satellite intelligent agents... Its high-level discrete actions Includes: satellite access correlation data Cross-layer routing volume . in, Indicates vehicle Whether to connect drones ; Indicates the first Whether a semantic descriptor is sent; Indicates whether the task is executed locally on the vehicle; Indicates whether the mission is in the drone Execution at the location; Indicates drone Whether to connect to satellite ; Indicates vehicle Is the mission carried out via drone? Upload to access satellite and in executing satellite The above processing.

[0163] (2) Low-level continuous motion Given that the discrete actions at higher levels are already determined, the continuous actions at lower levels... Used to generate communication and computing resource allocations.

[0164] For drone intelligent agents Its low-level continuous actions include:

[0165] in, Indicates vehicle The transmit power of the access link; Indicates vehicle semantic encoding frequency; Indicates vehicle Local processing frequency; Indicates vehicle To drones Access bandwidth; Indicates drone For vehicles The calculation frequency of task allocation.

[0166] For satellite intelligent agents Its low-level continuous actions include:

[0167] in, Indicates task By drone Back to satellite The transmit power at that time; Indicates task In satellite With satellite Inter-satellite transmission power during relay; Indicates the backhaul bandwidth; Indicates inter-satellite bandwidth; Indicates the execution of satellite For the task The allocated calculation frequency.

[0168] S34. Reward Function In order to align the long-term optimization objective of reinforcement learning with the joint optimization objective in step S2, this invention defines time slots. The global instant reward is

[0169]

[0170] in, Indicates time slot Instant rewards; This represents the total perceived utility of the system. This represents the energy consumption penalty coefficient; Indicates vehicle Total energy consumption throughout the entire task processing process; , , This is the penalty coefficient; Indicates task Total end-to-end latency; Indicates task Maximum tolerable delay; Indicates task Minimum semantic fidelity requirement; Indicates task Effective semantic fidelity; This represents the residual violation of the fixed downlink duration constraint.

[0171] The physical meanings of each term in the above formula are as follows: This is a positive benefit, used to encourage the system to achieve higher overall perceived utility; This is an energy consumption penalty term used to suppress system resource consumption; This is a penalty for delays, used to prevent tasks from exceeding their deadlines. This is a semantic quality violation penalty item, used to suppress situations where the semantic fidelity of the task falls below the requirement; Fixed downlink feedback duration penalty item for default. This reward function allows the agent to maximize long-term cumulative rewards while simultaneously considering perceptual utility, energy consumption, latency, and semantic quality.

[0172] S4: Solve the optimal policy using a hierarchical graph-guided multi-agent proximal policy optimization method. In step S3, the joint optimization problem has been transformed into a partially observable Markov decision process. To simultaneously handle discrete topology decisions and continuous resource allocation decisions in a high-dimensional hybrid action space, this invention employs a hierarchical graph-guided multi-agent proximal policy optimization method. This method operates under a Centralized Training-Distributed Execution (CTDE) paradigm and consists of a side-conditional heterogeneous graph representation module, a temporal memory module, a high-level discrete policy module, a low-level continuous policy module, and a centralized value evaluation module.

[0173] S41. Constructing a heterogeneous graph representation To characterize the diverse topological relationships between vehicles, drones, and satellites, in each time slot Constructing a contextualized heterogeneous graph

[0174] in, Indicates time slot Heterogeneous graphs; This represents the set of nodes in the graph; Let represent the set of edges in the graph.

[0175] A node set is defined as

[0176] in, Indicates a collection of vehicles. Indicates a collection of drones. This represents a set of satellites. In other words, the nodes in the heterogeneous graph include vehicle nodes, drone nodes, and satellite nodes. Vehicle nodes, acting as contextual vehicle nodes, do not directly output control actions, but they carry mission profiles, local link states, and resource states, providing contextual information for joint decision-making by drones and satellites.

[0177] For any node Its original input features are denoted as And mapped to the initial node embedding through a type-dependent feature encoding network:

[0178] in, Represents a node The initial embedding vector; This represents a multilayer perceptron associated with node type. Represents a node The type flag can be a vehicle, drone, or satellite. Represents a node In the time slot The original observational features.

[0179] Furthermore, the edge feature vector is defined as...

[0180] in, Represents a node With nodes Edge features between them; This indicates that the two nodes are in the time slot. The physical distance; This represents the probability of the line-of-sight distance between two nodes; This indicates the geometric visibility between two nodes.

[0181] For nodes any neighboring node Define the edge conditional attention score as

[0182] in, Represents a node For nodes The unnormalized attention score; This represents the linear rectification activation function with leakage coefficient; This represents a trainable attention parameter vector; and These represent the nodes respectively. and nodes Linear projection matrix of type; This represents the projection matrix for the edge type; This represents a vector concatenation operation; This indicates the transpose operation.

[0183] For nodes All neighbors are subjected to softmax normalization to obtain normalized attention weights.

[0184] in, Represents a node For nodes The normalized attention coefficient; Represents an exponential function; Represents a node The set of neighbors.

[0185] After obtaining the attention coefficient, the node The graph embedding is updated to

[0186] in, Represents a node Embedded update after a message pass; Represents a non-linear activation function; Represents the message passing matrix; Representing neighboring nodes The initial embedding; This represents the edge features. This formula illustrates that edge features not only participate in attention calculation but also directly participate in message aggregation, thereby improving the ability to model the state of heterogeneous links between air, space, and ground.

[0187] S42. Constructing a time memory module Considering that satellite visibility window, link status, and mission arrival process all have significant temporal correlations, a gated cyclic unit is introduced to perform temporal aggregation on graph embeddings, defining an intelligent agent. In the time slot The hidden state is

[0188] in, Represents intelligent agents In the time slot Temporal memory state; Indicates a gated loop unit; This indicates the structural embedding obtained from the heterogeneous graph for the current time slot; This represents the hidden state of the previous time slot. Using this formula, the current decision depends not only on the current graph structure but also on historical time-series information.

[0189] S43. Constructing a High-Level Discrete Strategy Based on hidden state Constructing a high-level discrete policy network Its output is a high-level discrete action.

[0190] in, Represents intelligent agents In the time slot High-level discrete actions; Represents a high-level discrete strategy network; Represents the parameters of a high-level discrete strategy network; symbol " "" indicates that sampling is performed according to the distribution of this strategy.

[0191] For drone intelligent agents High-level discrete actions mainly include: vehicle-drone correlation. Semantic descriptor sends indication quantity ; Execution mode selection quantity and . For satellite intelligent agents High-level discrete actions mainly include: satellite access selection Cross-layer routing volume . To ensure that discrete actions satisfy geometric visibility constraints, a mask constraint is applied to the discrete logits before sampling. Let the discrete logits before masking be... After masking, logits are defined as follows:

[0192] in, Represents the discrete logits after the mask; Represents the original discrete logits; Represents a mask vector for geometric visibility and topological consistency; "" indicates element-wise multiplication. If a candidate discrete action does not satisfy access visibility. Or return link visibility If the corresponding mask component is zero, then the action is prevented from being sampled.

[0193] S44. Constructing a low-level contiguous resource allocation strategy Given that the actions of the high-level discrete topology are already determined, construct a low-level continuous policy network. This is used to output continuous resource allocation actions. Low-level continuous actions are defined as follows:

[0194] in, Represents intelligent agents In the time slot Continuous resource actions; This represents a low-level continuous policy network; Indicates the parameters of a continuous policy network; It represents high-level discrete actions and is used as a conditional input for continuous resource allocation.

[0195] Because variables such as access bandwidth, backhaul bandwidth, inter-satellite bandwidth, and UAV / satellite computing frequency need to meet capacity constraints, these resources are modeled using a Dirichlet distribution. Taking UAVs as an example... Taking the access bandwidth allocation as an example, let its currently active task set be... The Dirichlet parameters are then output by the continuous policy network.

[0196] in, Indicates drone The access bandwidth Dirichlet parameter vector; Indicates assignment to task Bandwidth parameters; Indicates in time slot Access to drones The set of activation tasks.

[0197] Then sample the normalized bandwidth share.

[0198] in, Represents the access bandwidth share vector; This represents the Dirichlet distribution.

[0199] The final access bandwidth allocation is

[0200] in, Indicates task To drones Access bandwidth; This indicates the corresponding normalized bandwidth share; Indicates drone Total available access bandwidth budget. Since the sum of samples in the Dirichlet distribution is 1, this formula naturally satisfies bandwidth and capacity constraints.

[0201] A Beta distribution is used to model the local frequency and power ratio variables within a limited range. (Taking vehicles as an example.) Taking the local processing frequency as an example, define

[0202] in, This represents the proportion of normalized local processing frequencies obtained from sampling by the Beta distribution; Represents the Beta distribution; and Indicates the shape parameter of the Beta distribution; Indicates vehicle In the time slot Local processing frequency; Indicates the local execution instruction quantity; Indicates vehicle The maximum computational frequency budget. This formula guarantees that when the task is not executed locally, the corresponding local frequency will automatically be zero.

[0203] S45. Construct a centralized value assessment module In the centrally trained and distributed execution paradigm, a centralized value network is used to estimate the value of the global state. The centralized critic is defined as...

[0204] in, This represents a centralized value network; Indicates parameters of a centralized value network; Indicates time slot The global state. This module is used to mitigate environmental non-stationarity during multi-agent parallel training.

[0205] After each round of interaction, the temporal difference residual is first calculated.

[0206] in, Indicates time slot The timing difference error; Indicates an immediate reward; Indicates the discount factor; This indicates the global state of the next time slot.

[0207] Based on the temporal difference residuals, generalized dominance estimation is used to obtain...

[0208] in, Indicates time slot Advantage estimation; The smoothing parameter represents the advantage estimation; This represents the trajectory length or round length. This formula is used to balance bias and variance, thereby improving the stability of policy updates.

[0209] S46. Employ the MAPPO pruning agent target update strategy network For any intelligent agent The likelihood ratio between the old and new strategies is defined as follows:

[0210] in, Represents intelligent agents In the time slot The strategy likelihood ratio; Indicates the current strategy; This indicates the old strategy before the update; Indicates time slot Joint actions taken; This indicates the corresponding local observation.

[0211] The corresponding actor update target function is:

[0212] in, Represents intelligent agents The actor's target function; This represents the expectation operation; Represents the clipping function; Indicates the clipping threshold; Represents the entropy regularity coefficient; This represents the entropy of the current strategy. The first term in the formula is used to limit the magnitude of a single strategy update, and the second term is used to enhance exploration capabilities.

[0213] The update objective function of the centralized critic is:

[0214] in, Represents the loss function of a centralized value network; Indicates time slot The target return; This represents the current value estimate. The formula uses mean squared error to train the value network.

[0215] After multiple rounds of environmental interaction and policy updates, the high-level discrete policies and low-level continuous policies of each UAV and satellite agent gradually converge, ultimately obtaining the optimal policy for joint control semantic compression, execution mode, satellite access, cross-layer routing, and resource allocation.

[0216] As can be seen from the above steps, the present invention extracts cross-layer structural information through edge-conditional heterogeneous graphs, models the dynamic environment through time memory modules, processes association, semantic selection and execution mode through high-level discrete strategies, processes bandwidth, power and computing frequency allocation through low-level continuous strategies, and achieves stable training through MAPPO pruning of agent targets, thereby obtaining a joint resource management strategy that satisfies the constraints of air-space-ground networks.

[0217] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0218] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A method for integrated air-space-ground collaborative semantic communication and resource management, characterized in that, The method includes: Step S1: Construct an integrated air-space-ground perception-guided collaborative semantic communication system model; wherein, the integrated air-space-ground perception-guided collaborative semantic communication system model consists of a three-layer network structure composed of low-orbit satellites, UAVs and vehicles, the low-orbit satellites are used to provide wide-area connectivity and remote computing support, the UAVs are used to provide relay and edge-assisted services, and the vehicles are used to generate collaborative perception tasks and initiate semantic communication requests. Step S2: Based on the integrated air-space-ground perception-guided collaborative semantic communication system model in Step S1, establish a probabilistic semantic compression and recovery model, a satellite visibility and cross-layer routing model, a communication model, an execution latency and energy consumption model, and a perception utility model, and construct a joint optimization objective function; Step S3: Transform the joint optimization problem in step S2 into a partially observable Markov decision process; Step S4: For the partially observable Markov decision process established in Step S3, a multi-agent joint decision-making algorithm guided by a hierarchical graph is adopted to jointly make decisions on vehicle-UAV association, semantic descriptor selection, execution mode, satellite access, cross-layer routing, and communication and computing resource allocation to obtain the optimal resource management strategy.

2. The method according to claim 1, characterized in that, The probabilistic semantic compression and recovery model mentioned in step S2 includes: Extract a set of semantic descriptors from the raw vehicle perception data, share a probabilistic semantic graph between the transmitting and receiving ends, and carry a truth probability for each semantic rule; Set a 0-1 semantic sending indicator variable, calculate the task semantic compression rate based on the number of sending variables, and derive the amount of semantic data after compression. For unsent semantic descriptors, the probability recovery value of a single descriptor is calculated based on the shared semantic graph. The effective fidelity of a single descriptor is obtained by fusing the sending indicator variable and then taking a weighted average to obtain the overall semantic fidelity of the task. Establish a piecewise affine function to characterize the mapping relationship between semantic compression rate and vehicle coding workload.

3. The method according to claim 1, characterized in that, The satellite visibility and cross-layer routing model mentioned in step S2 includes: Set up satellite-drone visibility 0-1 indicator, vehicle-drone association 0-1 variable, and drone-satellite access 0-1 variable; The design incorporates a four-dimensional cross-layer routing 0-1 indicator to distinguish between access satellites and execution satellites, and supports modeling of inter-satellite forwarding links. Set up binary instructions for local execution and drone execution, and constrain the three execution modes to be mutually exclusive and one of them must be selected; Add topology consistency constraints and satellite return link visibility constraints, and enable the corresponding cross-layer routing path only when the geometric visibility and access association conditions are met.

4. The method according to claim 1, characterized in that, The communication model described in step S2 includes three types of communication links: Vehicle-to-UAV access link: Uplink and downlink transmission rates are modeled using Rician fading channel, and orthogonal pre-allocated feedback channel is used for downlink; UAV-satellite backhaul link: A zero-order Bessel function is introduced to construct a channel aging coefficient, which characterizes the time-varying channel error caused by Doppler frequency shift; Inter-satellite relay links: The transmission rate of inter-satellite optical links is calculated based on the free-space optical communication model, and the rate is derived by integrating optical gain, noise temperature, and carrier loss; the one-way propagation delay of each link is defined and used for the end-to-end total delay superposition calculation.

5. The method according to claim 1, characterized in that, The execution latency and energy consumption model mentioned in step S2 includes: Four types of latency and energy consumption sub-models are constructed for local execution, vehicle-side semantic encoding, UAV-side offloading execution, and satellite-side offloading execution, respectively. Calculate the energy consumption of each node based on the DVFS dynamic voltage-frequency model; The complete end-to-end total latency and total energy consumption of both drone and satellite execution modes are derived for different scenarios. The execution indicator variable is used to achieve a unified expression of latency and energy consumption for multiple modes.

6. The method according to claim 1, characterized in that, The perceived utility model mentioned in step S2 includes: Perceived freshness is defined based on the exponential decay function; the greater the time delay, the lower the freshness. The fusion execution mode corrects the effective semantic fidelity. By introducing task weights and urgency coefficients, we can construct single-task perceived utility. The higher the urgency coefficient, the more the task focuses on timeliness, and vice versa, the more it focuses on semantic integrity. The total perceived utility of the system is obtained by summing the utility of all tasks. Construct a global long-term joint optimization objective function with energy consumption penalty, latency default penalty, and semantic quality default penalty.

7. The method according to claim 1, characterized in that, The joint optimization problem described in step S3 is transformed into a partially observable Markov decision process, specifically as follows: The intelligent agent ensemble consists of all drones and satellites; local observations of intelligent agents include four types of information: link status, local resources, geometric visibility, and local task profile; the hybrid action space is divided into high-level discrete actions and low-level continuous actions. High-level discrete actions: UAV intelligent agent outputs vehicle--UAV association, semantic descriptor transmission, and task execution mode decision-making; satellite intelligent agent outputs satellite access, cross-layer routing decision-making, and masks infeasible discrete actions based on geometric visibility masking; Low-level continuous actions: The UAV outputs the vehicle's transmission power, encoding frequency, local computing power, vehicle-to-vehicle bandwidth, and UAV computing power; Satellite output return power, inter-satellite transmit power, return bandwidth, inter-satellite bandwidth, and satellite calculation frequency; Dirichlet distribution is used to model the bandwidth class bounded summation resources, and Beta distribution is used to model the power and frequency range resources, which naturally satisfies the resource capacity constraints; The global instant reward consists of five parts: system perceived utility, system energy consumption penalty, latency exceeding the standard penalty, semantic fidelity insufficient penalty, and downlink duration violation penalty.

8. The method according to claim 1, characterized in that, Step S4 describes a hierarchical graph-guided multi-agent proximal policy optimization algorithm, which comprises five core modules: Edge-conditional heterogeneous graph module: Vehicles, drones, and satellites are used as graph nodes. Node features include observation information, and edge features include node distance, line-of-sight probability, and visibility. An edge-conditional attention mechanism is used to aggregate cross-layer node messages and generate node spatial embeddings. Temporal Memory GRU Module: Embeds a single time-slot graph into the input gated loop unit and fuses historical time-slot states to obtain the agent's hidden state with temporal memory; Hierarchical Actor Policy Network: High-level discrete Actors output masks and make topology decisions, while low-level continuous Actors output Dirichlet and Beta distribution parameters based on discrete action conditions, and sample to obtain legal resource allocation values; Centralized Critic Value Network: Input global state and output value estimate; calculate the advantage function through temporal difference residuals and generalized advantage estimation (GAE). Multi-agent MAPPO policy update module: Constructs Actor pruning loss and Critic mean square loss respectively, adds policy entropy regularization term to improve environment exploration capability, and outputs global joint resource management policy after iterating until policy convergence.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the integrated air-space-ground collaborative semantic communication and resource management method as described in any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the integrated air-space-ground collaborative semantic communication and resource management method as described in any one of claims 1 to 8.

Images