A channel map construction method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and provides a channel map construction method and device, electronic equipment and a storage medium. The method comprises the following steps: deploying N unmanned aerial vehicles (UAVs) to cooperatively perform a mapping task in multiple grids of a target area; inputting system running state information related to the mapping task into a joint decision model to output a joint decision result of each UAV; controlling each UAV to move according to a flight decision and to perceive environmental feature data of a grid to be mapped; performing reasoning based on the environmental feature data according to a mapping decision through the UAV and a server to generate a local channel map; and merging the local channel maps into a global channel map of the target area. The channel map construction method provided by the application adopts a UAV and server cooperative reasoning architecture, greatly reduces the computing load of the UAV end, realizes efficient cooperative work of multiple UAVs, and significantly shortens the overall mapping time delay under the premise of ensuring mapping accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a channel map construction method, apparatus, electronic device and storage medium. Background Technology

[0002] Sensing and modeling the wireless communication environment is a crucial foundation for the operation and network optimization of mobile communication systems. Radio maps, by describing the mapping relationship between spatial locations and wireless channel characteristic parameters (such as path loss, signal-to-noise ratio, or channel capacity), support network planning, interference management, resource allocation, and intelligent decision-making. Compared to traditional methods relying on ground measurements or fixed base station data collection, unmanned aerial vehicles (UAVs), due to their mobility, ease of deployment, and three-dimensional spatial coverage capabilities, are used for channel sensing and channel map construction in scenarios such as complex terrain, emergency communications, and temporary network deployments. This provides an effective means to improve the environmental adaptability and service assurance capabilities of wireless networks.

[0003] With the continuous expansion of the scale and application scenarios of wireless networks, the channel environment exhibits significant spatial heterogeneity and temporal dynamics, placing higher demands on the spatial resolution, update frequency, and modeling accuracy of channel maps. Early channel map construction methods were mostly based on spatial interpolation or empirical propagation models, estimating unknown regions using a limited number of measurement points. However, their modeling accuracy was limited under complex occlusion environments or non-uniform sampling conditions. In recent years, with the development of machine learning technology, learning-based methods have been gradually introduced into the field of channel modeling. For example, convolutional neural networks are used to recover high-resolution channel distributions from sparse observation data, improving mapping accuracy and generalization ability. Based on this, generative models, especially diffusion models, have been used for channel map reconstruction tasks due to their ability to model complex distributions and reconstruct details. Existing research shows that diffusion models can generate high-precision and high-resolution channel maps under limited observation sample conditions, helping to improve modeling performance in complex environments. However, these models typically require multiple rounds of denoising inference, resulting in high computational complexity, inference latency, and energy consumption, placing high demands on computing resources and limiting their direct deployment on computationally and energy-constrained platforms (such as drones). Summary of the Invention

[0004] This application provides a channel map construction method, apparatus, electronic device, and storage medium to solve the problems of high computational complexity, large inference latency and energy consumption caused by the need for multiple rounds of denoising inference in existing channel map construction technologies, which also have high requirements for computing resources and limit their direct deployment on platforms with limited computing power and energy.

[0005] Firstly, this application provides a method for constructing a channel map, including: The target area is divided into multiple grids, and N drones are deployed in the multiple grids to collaboratively perform mapping tasks. The system operation status information related to the mapping task is input into a joint decision-making model based on reinforcement learning to obtain the joint decision-making results of each UAV output by the joint decision-making model; the joint decision-making results include flight decisions and mapping decisions; Each drone is controlled to move according to the corresponding flight decision, reach the grid to be mapped, and perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the drone and the server, according to the corresponding mapping decision, based on the environmental feature data, cloud-edge collaborative execution of diffusion model inference is performed to generate a local channel map of the grid to be mapped. The local channel maps of each grid are merged to generate a global channel map of the target area.

[0006] In one embodiment, the system operating status information includes non-spatial scalar information and spatial tensor information; The non-spatial scalar information includes at least one of the following: server's computing queue status, task progress time, UAV computing latency statistics, channel status between UAV and server, and relative distance between UAVs. The spatial tensor information is a multi-channel raster map constructed based on the multiple grids. Each UAV corresponds to an independent multi-channel raster map. The multi-channel raster map is superimposed with at least one of the following semantic channels: UAV's own position, distribution of other UAVs, transmitter power and density, mapping completion status, and flight trajectory within a preset historical period.

[0007] In one embodiment, the flight decision includes: indicating the direction of movement for the drone's flight; The mapping decision includes: mapping on / off status for indicating the UAV's task execution, total number of denoising steps for controlling the inference complexity of the diffusion model, cloud-edge task partitioning points for dividing UAV and server computing tasks, and local offloading ratio for indicating mapping tasks within the grid.

[0008] In one embodiment, the grid to be mapped includes multiple transmitters, and the environmental feature data includes environmental structural features and transmitter locations; the process of generating a local channel map of the grid to be mapped by cloud-edge collaborative computing between the UAV and the server, according to the corresponding mapping decision, and based on the environmental feature data, by cloud-edge collaboratively executing diffusion model inference, includes: Based on the local offloading ratio, all transmitters within the grid to be built are divided into a local execution group and an offloading execution group; For each transmitter in the local execution group, the UAV performs diffusion model inference based on the environmental structure features and the transmitter's location, according to the total number of denoising steps, to generate a local sub-channel map of the transmitter within the grid to be built. For each transmitter in the unloading execution group, the UAV and the server, based on the environmental structure features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task division points, perform phased cloud-edge collaborative execution of diffusion model inference to generate a local sub-channel map of the transmitter within the grid to be built; The local sub-channel maps of each transmitter within the grid to be constructed are merged to generate the local channel map of the grid to be constructed.

[0009] In one embodiment, for each transmitter in the offloading execution group, the process involves generating a local sub-channel map of the transmitter within the grid to be built, using the UAV and the server, based on the environmental structural features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task partitioning points, through phased cloud-edge collaborative execution of the diffusion model inference, including: Based on the cloud-edge task partitioning points, the total number of denoising steps is divided into the first denoising steps performed by the drone and the second denoising steps performed by the server. For the first transmitter in the unloading execution group, the UAV performs diffusion model inference based on the environmental structure features and the location of the first transmitter, according to the first denoising step count, to generate intermediate latent variable features; the server performs diffusion model inference based on the intermediate latent variable features, according to the second denoising step count, to generate a local sub-channel map of the first transmitter. For the remaining transmitters in the unloading execution group, the server, based on the intermediate latent variable features, performs diffusion model inference according to the second denoising step number to generate local sub-channel maps for the remaining transmitters. The first transmitter is the most valuable transmitter among all transmitters in the offloading execution group; the remaining transmitters are the transmitters in the offloading execution group other than the first transmitter.

[0010] In one embodiment, the joint decision-making model is trained in the following manner: Non-spatial scalar samples and spatial tensor samples are encoded respectively to obtain the first embedding feature and the second embedding feature, and the first embedding feature and the second embedding feature are concatenated to construct the state vector; Based on flight decision samples and mapping decision samples, construct action vectors; A reward function is constructed based on the globally weighted mapping revenue and the total system energy consumption; the globally weighted mapping revenue is obtained by summing the single mapping revenue of each transmitter; the total system energy consumption is obtained by summing the computing energy consumption of the UAV, the flight energy consumption, and the computing energy consumption of the server. Based on the state vector, the action vector, and the reward function, a reinforcement learning algorithm is used to adaptively optimize and train the joint decision.

[0011] In one embodiment, the reinforcement learning algorithm is a proximal policy optimization algorithm.

[0012] In one embodiment, the drone and the server constitute a cloud-edge collaborative computing architecture. This architecture dynamically allocates computing tasks between the drone and the server using the cloud-edge task partitioning point and the local offloading ratio. The joint decision-making model adaptively adjusts the local offloading ratio based on the server's computing queue status. When the server's computing queue congestion increases, the joint decision-making model increases the local offloading ratio to reduce the number of mapping tasks offloaded to the server. When the server's computing queue congestion decreases, the joint decision-making model decreases the local offloading ratio to increase the number of mapping tasks offloaded to the server.

[0013] Secondly, this application also provides a channel map construction apparatus, comprising: The deployment module is used to divide the target area into multiple grids and deploy N drones in the multiple grids to collaboratively perform mapping tasks. The joint decision-making module is used to input system operation status information related to the mapping task into the reinforcement learning-based joint decision-making model to obtain the joint decision-making results of each UAV output by the joint decision-making model; the joint decision-making results include flight decisions and mapping decisions; The inference task execution module is used to control each UAV to move according to the corresponding flight decision, reach the grid to be mapped, and perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the UAV and the server, according to the corresponding mapping decision, based on the environmental feature data, cloud-edge collaborative execution of diffusion model inference is performed to generate a local channel map of the grid to be mapped. The channel map generation module is used to merge the local channel maps of each grid to generate a global channel map of the target area.

[0014] Thirdly, this application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the channel map construction methods described above.

[0015] Fourthly, this application also provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the channel map construction methods described above.

[0016] The channel map construction method, apparatus, electronic device, and storage medium provided in this application globally optimize the flight path planning and mapping execution strategy of UAVs through a joint decision-making model based on reinforcement learning. It can dynamically output the optimal decision for each UAV based on the real-time system operating status, and then execute the diffusion model inference process based on the optimal decision. This process adopts an architecture of collaborative inference between UAVs and backend servers, and rationally distributes the computational tasks of the diffusion model at both the cloud and edge ends. It makes full use of the powerful computing power of the server and the flexible mobile perception advantages of UAVs, which significantly reduces the computational load and energy consumption pressure on the UAV end, and realizes efficient collaborative operation of multiple UAVs. While ensuring the accuracy of channel map construction, it significantly shortens the overall mapping latency of large-scale complex areas, and can overcome the bottleneck that diffusion models are difficult to deploy on UAV platforms with limited computing power and energy. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the channel map construction method provided in this application.

[0019] Figure 2 This is a schematic diagram of the cloud-edge collaborative inference architecture provided in this application.

[0020] Figure 3 This is a schematic diagram of the skip-step reasoning and intermediate feature reuse mechanism provided in this application.

[0021] Figure 4 This is a schematic diagram of the accuracy degradation curve of denoising step compression and partition point reuse provided in this application.

[0022] Figure 5 This is a schematic diagram of the heterogeneous collaborative reasoning and partition point reuse process provided in this application.

[0023] Figure 6 This is a schematic diagram of the three-transmitter heterogeneous collaboration and partition point multiplexing execution example provided in this application.

[0024] Figure 7 This is a schematic diagram of the PPO network architecture provided in this application.

[0025] Figure 8 This is a schematic diagram of the PPO training reward curve provided in this application.

[0026] Figure 9 This is a schematic diagram of the flight trajectory of the UAV provided in this application and the number of UAV tasks processed in each time slot.

[0027] Figure 10 This is a schematic diagram of the decision distribution of PPO under different environmental conditions provided in this application.

[0028] Figure 11 This is a schematic diagram of the channel map construction device provided in this application.

[0029] Figure 12 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein.

[0032] In recent years, with the development of wireless communication technology and intelligent sensing systems, channel modeling and channel map construction methods for complex environments have emerged continuously. Regarding the "UAV + diffusion model + cooperative offloading" problem addressed in this application, relevant technical solutions can be summarized from three aspects: channel map construction methods, diffusion model inference acceleration and cloud-edge segmentation, and multi-constraint joint offloading and decision-making.

[0033] In terms of channel map construction methods, existing methods are mainly divided into model-based methods and data-driven methods. Model-based methods, based on channel propagation theory, combine path loss models, shadow fading models, or ray tracing methods to model wireless channel characteristics, possessing a certain degree of physical interpretability. However, in complex urban scenarios or environments with occlusion and multipath effects, they rely on accurate environmental information, resulting in high modeling costs and limited adaptability to environmental differences. With the development of machine learning technology, data-driven methods utilize convolutional neural networks, graph neural networks, or generative models to learn the mapping relationship between spatial location and channel characteristics, predicting the channel state in unobserved areas, and demonstrating strong modeling capabilities in complex environments. Among these, diffusion models have been introduced into the field of channel map construction due to their advantages in complex distribution modeling and high-quality data reconstruction. A representative method is RadioDiff, which uses environmental structure and transmitter location as input conditions to perform high-fidelity reconstruction of channel distribution, significantly improving spatial continuity and prediction accuracy compared to previous discriminative models, and has gradually become one of the mainstream methods in this field. However, this type of method requires a large number of iterations and has high single inference latency under the original settings, making it unfriendly to UAV platforms with limited computing power. To address this overhead issue, several lightweight inference schemes have been proposed, with representative works such as RadioDiff-Turbo and RadioDiff-Flux reducing inference costs by decreasing the number of denoising iterations and reusing intermediate features across tasks, respectively, while maintaining accuracy. This application uses RadioDiff as the basic inference engine for channel map reconstruction and incorporates the aforementioned two acceleration mechanisms into the decision space of cloud-edge collaborative inference to support its efficient collaborative execution on UAV platforms.

[0034] In terms of accelerating diffusion model inference and cloud-edge segmentation, existing research can be summarized into four categories of methods: spatial segmentation, temporal segmentation, model segmentation, and task scheduling. Spatial partitioning methods (such as DistriFusion and PipeFusion) divide the input image into multiple blocks and distribute them to different devices for parallel denoising. The speedup increases nearly linearly with the number of devices, but they rely on the assumption of homogeneous devices and have a large amount of cross-block communication. Temporal partitioning methods (such as Hybrid SD and EC-Diff) partition the denoising step along the time dimension to different devices. They have less communication and naturally support heterogeneous device collaboration, but the selection of partitioning points directly affects the quality-latency tradeoff, and each device needs to deploy the model independently. Model partitioning methods (such as AsyncDiff) distribute the diffusion model parameters to different devices according to the inter-layer or intra-layer weight matrix, which can support the deployment of ultra-large models, but inter-layer partitioning is limited by non-linear topology such as skip connections, and intra-layer partitioning relies on high-bandwidth interconnects. Task scheduling methods (such as LAD-TS and EdgeAIGC) schedule the complete diffusion generation task among multiple servers without modifying the model itself and are compatible with any diffusion model architecture, but it is difficult to reduce the inference latency of a single task, and the scheduling decision depends on real-time perception of the system state. However, the above methods are mostly designed for homogeneous server clusters or image generation scenarios, lacking a "native denoising step compression" mechanism for restricted end devices, and failing to form a unified decision-making framework with the adaptive configuration of cloud-edge partitioning points. This makes it difficult to directly support the efficient diffusion inference needs of UAV platforms under computing power constraints.

[0035] In the area of ​​multi-constraint joint offloading and decision-making, edge computing and task offloading technologies have been widely applied to UAV-assisted wireless networks to address the limitations of computing resources and energy on UAV platforms. Related research typically models offloading decisions as a joint optimization problem with multiple constraints, jointly considering communication rates, computing power, task scale, and energy consumption models. It constructs optimization models with system latency, energy consumption, or task completion rate as objectives, and employs techniques such as greedy strategies, branch and bound, convex approximation, alternating optimization, or iterative solutions based on Lagrange decomposition to process discrete offloading decisions and continuous resource allocation variables step-by-step or with relaxation. Some studies also segment the computation graph through phased or modular task partitioning to achieve collaborative execution between terminals, edge, and the cloud. For example, combining channel map construction with computation offloading mechanisms reduces system latency and energy consumption by jointly optimizing device association, computing resource allocation, and task scheduling. For tasks with high computational complexity generated by diffusion models, collaborative computing between terminals, edge, and the cloud is achieved through model segmentation or piecemeal execution. However, the above methods are mainly aimed at homogeneous server clusters or image generation scenarios, and they generally face problems such as high computational complexity, strong dependence on model assumptions, and difficulty in balancing stability and multiple objectives under conditions of multiple drones, large-scale tasks and multiple constraint links, making it difficult to obtain a stable and effective global solution.

[0036] In summary, in the channel map construction task based on the diffusion model in the UAV scenario, how to combine the UAV's mobility characteristics, the hierarchical iterative characteristics of the diffusion model, and the multi-UAV collaborative structure to uniformly model the channel mapping accuracy, inference latency, and system energy consumption, and model the decision-making process as a multi-time-slot sequence optimization problem, still has the following three shortcomings: First, existing diffusion acceleration schemes (such as optional denoising steps and cross-task intermediate feature reuse) are mostly triggered at the model / graphics processing unit (GPU) level according to preset rules, without linkage with UAV mobility-perception-cloud-edge collaboration, making it difficult to adaptively select the inference granularity based on the task state on edge platforms with limited computing power; second, existing joint offloading and resource allocation methods are difficult to balance training stability and multiple objectives (accuracy-latency-energy consumption) tradeoffs under multi-UAV, large-scale concurrent tasks; third, existing deep reinforcement learning (Deep Reinforcement) methods are insufficient. Learning (DRL) offloading methods are mostly geared towards general image processing or single-task scenarios. They do not include the adjustable inference granularity of the diffusion model itself (optional denoising steps and cross-task intermediate feature reuse) as decision variables in the policy space, resulting in the neglect of a collaborative optimization dimension closely related to the structure of the diffusion model.

[0037] To address the aforementioned issues, this application focuses on a cloud-edge collaborative inference and decision-making scheme based on a diffusion model for constructing channel maps in unknown areas using multi-UAV collaboration. The system deploys multiple UAVs to perform mapping tasks along a grid within an unknown area. Each time slot perceives the environmental characteristics of its grid and the locations of transmitters within that grid, using these as inputs to the RadioDiff diffusion model. The diffusion model then infers and outputs the channel map for each transmitter within that grid, ultimately merging them to form a global channel map for the entire area. Diffusion model inference is essentially a multi-step iterative denoising process, inherently involving numerous steps and high latency per inference iteration. The UAVs have limited computing power, making it difficult to independently handle complete diffusion inference. While the backend server has sufficient computing power per machine, the number of available GPUs is limited, leading to significant queuing inference queues under concurrent requests from multiple UAVs. Furthermore, the intermediate latent variables of multiple transmitters within the same grid are highly similar, resulting in substantial redundant computations during independent inference. To address this, this application designs a diffusion model inference and decision-making scheme for multi-UAV-back-end server cloud-edge collaboration: at the diffusion model inference level, optional denoising steps and reuse of intermediate features from multiple transmitters in the same grid are introduced to reduce computational load; at the system decision level, the UAV flight path, diffusion model inference steps, cloud-edge task partitioning points, and communication bandwidth allocation are uniformly incorporated into the joint decision space, which is then solved online adaptively by a reinforcement learning agent, thereby reducing total energy consumption and end-to-end latency under multi-UAV collaboration conditions while ensuring mapping accuracy.

[0038] The following is combined Figures 1-10The various embodiments describe the channel map construction method provided in this application.

[0039] The channel map construction method provided in this application embodiment can be implemented based on a channel map construction device. Therefore, this application embodiment uses a channel map construction device as the execution subject to describe the channel map construction method.

[0040] Figure 1 This is a flowchart illustrating the channel map construction method provided in this application.

[0041] like Figure 1 As shown, the channel map construction method includes the following steps: Step 101: Divide the target area into multiple grids and deploy N drones in the multiple grids to collaboratively perform mapping tasks; Step 102: Input the system operation status information related to the mapping task into the joint decision-making model based on reinforcement learning to obtain the joint decision-making results of each UAV output by the joint decision-making model; the joint decision-making results include flight decisions and mapping decisions; Step 103: Control each UAV to move according to the corresponding flight decision, reach the grid to be mapped, and perceive the environmental feature data of the grid to be mapped; through the UAV and the server, according to the corresponding mapping decision, based on the environmental feature data, collaboratively execute diffusion model inference to generate a local channel map of the grid to be mapped; Step 104: Merge the local channel maps of each grid to generate a global channel map of the target area.

[0042] Specifically, this application addresses the task of constructing channel maps for large-scale unknown regions. To accomplish this task, a system is deployed... A drone continuously flies within an unknown physical region and collaboratively performs environmental perception and mapping tasks. (In discrete time slots) Running on, the time slot length is In each time slot Within the system, the UAV senses environmental characteristics data within its local area as input, driving the diffusion model to reconstruct the channel map within that local area, thus providing support for subsequent communication planning, resource allocation, and network optimization.

[0043] Each drone is equipped with a computing unit and a data transmission and receiving unit with limited computing power, responsible for collecting environmental feature data within a local area; the back-end server has strong computing power to provide additional computing power to support inference calculations. The drones and the server communicate via an air-to-ground wireless link to offload and distribute the task workload.

[0044] Combination Figure 2 , Figure 2 This is a schematic diagram of the cloud-edge collaborative inference architecture provided in this application.

[0045] The large-scale location region of the map to be constructed is taken as the target region, and the target region is discretized into equal-sized regions. Grid, system deployment A drone flies along a grid within the target area to perform mapping tasks, with additional computing power provided by a back-end server, forming a multi-drone-server cloud-edge collaborative inference architecture.

[0046] Before executing the mapping task, collaborative decisions are needed to determine the flight path and computational task offloading scheme for each UAV to ensure that the multi-UAV system can efficiently and adaptively complete the global channel map construction. To this end, this application introduces a joint decision model based on reinforcement learning. This model performs end-to-end decision optimization based on the real-time system operating status in each decision slot, and outputs a joint decision result containing flight and mapping instructions for each UAV, realizing the global unified optimization of UAV movement strategy and mapping execution strategy.

[0047] Based on the output of the joint decision-making model, control commands are issued to each UAV. Each UAV, according to its corresponding flight decision, autonomously plans and executes flight maneuvers to move to the designated location on the mapping grid. (UAV) In the time slot The grid position is denoted as Its flight altitude is the preset flight altitude. Upon reaching the grid to be mapped, the UAV uses its onboard sensors to perceive environmental feature data of the grid area. Subsequently, the UAV establishes a communication connection with the backend server and, based on the perceived environmental feature data and the corresponding mapping decision, collaboratively executes the inference calculation process of the diffusion model to generate a local channel map corresponding to the grid to be mapped.

[0048] As multiple UAVs continuously fly and map the target area, the mapped grids gradually accumulate. The system terminates when all grids have been mapped, all UAVs have been offline, or the maximum number of time slots (T) has been reached. Finally, the mapping results from multiple grids are merged according to a preset spatial coordinate relationship to generate a global channel map covering the entire target area. This global channel map comprehensively characterizes core channel characteristics such as signal strength and path loss at any location within the target area, providing indispensable foundational data support for subsequent precise planning, dynamic resource scheduling, and intelligent interference management of wireless communication networks.

[0049] The channel map construction method, apparatus, electronic device, and storage medium provided in this application globally optimize the flight path planning and mapping execution strategy of UAVs through a joint decision-making model based on reinforcement learning. It can dynamically output the optimal decision for each UAV based on the real-time system operating status, and then execute the diffusion model inference process based on the optimal decision. This process adopts an architecture of collaborative inference between UAVs and backend servers, and rationally distributes the computational tasks of the diffusion model at both the cloud and edge ends. It makes full use of the powerful computing power of the server and the flexible mobile perception advantages of UAVs, which significantly reduces the computational load and energy consumption pressure on the UAV end, and realizes efficient collaborative operation of multiple UAVs. While ensuring the accuracy of channel map construction, it significantly shortens the overall mapping latency of large-scale complex areas, and can overcome the bottleneck of the difficulty in deploying diffusion models on UAV platforms with limited computing power and energy.

[0050] In one embodiment, the system operating status information includes non-spatial scalar information and spatial tensor information; The non-spatial scalar information includes at least one of the following: server's computing queue status, task progress time, UAV computing latency statistics, channel status between UAV and server, and relative distance between UAVs. The spatial tensor information is a multi-channel raster map constructed based on the multiple grids. Each UAV corresponds to an independent multi-channel raster map. The multi-channel raster map is superimposed with at least one of the following semantic channels: UAV's own position, distribution of other UAVs, transmitter power and density, mapping completion status, and flight trajectory within a preset historical period.

[0051] Specifically, system operating status information can be further divided into two major categories based on data structure characteristics: non-spatial scalar information and spatial tensor information.

[0052] Non-spatial scalar information refers to one-dimensional numerical data that lacks spatial distribution attributes. It is used to characterize various global state parameters and resource constraints during system operation. It includes at least one of the following: server computation queue status, i.e., the GPU queue length in the server, reflecting the server's current computing load and prioritizing task allocation to idle service nodes with shorter queue lengths; task progress time, using normalized time slot progress, reflecting the cumulative runtime since the self-graphing task started, providing a time benchmark for the model's latency-sensitive decisions; UAV computation latency statistics, characterizing the average time and computing load of a single UAV performing local inference; channel status between the UAV and the server, quantifying the transmission quality and available bandwidth of the air-to-ground wireless link, providing communication resource basis for task offloading decisions; and relative distance between UAVs, characterizing the spatial interval between multiple UAVs, assisting the model in collision avoidance and regional task allocation optimization.

[0053] Spatial tensor information is multi-dimensional structured data with spatial distribution attributes. It is constructed into a multi-channel raster map based on multiple grid divisions. The size of the raster map is consistent with the grid division size of the target area, and each raster cell corresponds to one grid. Each UAV corresponds to an independent multi-channel raster map. The multi-channel raster maps of different UAVs differ only in semantic channels related to their own positions. The other channels share globally unified spatial information, enabling the model to make distributed collaborative decisions from the local perspective of a single UAV while combining global spatial distribution information. The multi-channel raster map is overlaid with at least one of the following semantic channels: UAV's own position channel, used to mark the grid position of the current UAV; other UAV distribution channel, used to mark the grid positions of all UAVs other than the current UAV; transmitter power and density channel, used to characterize the sum or number distribution of the signal strength of the transmitters to be mapped in each grid; mapping completion status channel, used to mark whether each grid has completed the task of building all channel maps; and flight trajectory channel within a preset historical period, used to reflect the flight trajectory of the UAV in a recent period.

[0054] This application embodiment divides the system operation status information into two major categories according to data structure characteristics: non-spatial scalar information and spatial tensor information. Through the classification, representation and fusion of heterogeneous data, it provides comprehensive and accurate global state input for the joint decision-making model based on reinforcement learning, ensuring that the decision results can simultaneously adapt to system resource constraints and task progress requirements.

[0055] In one embodiment, the flight decision includes: indicating the direction of movement for the drone's flight; The mapping decision includes: mapping on / off status for indicating the UAV's task execution, total number of denoising steps for controlling the inference complexity of the diffusion model, cloud-edge task partitioning points for dividing UAV and server computing tasks, and local offloading ratio for indicating mapping tasks within the grid.

[0056] Specifically, the joint decision-making results for each UAV output by the joint decision-making model are divided into two core decisions: flight decision and mapping decision.

[0057] The flight decision core includes movement direction parameters that instruct the drone's flight. These parameters are defined based on a grid-space coordinate system and correspond to the drone's movement direction selection between adjacent grids. Based on these parameters, the drone can be controlled to move from its current grid to a designated grid to be mapped to perform its task. The model integrates information such as the global distribution of unmapped areas, the location of high-value transmitters, and the task allocation of other drones to output the optimal movement direction for each drone, achieving efficient and non-duplicative coverage of the target area while avoiding flight collisions and task conflicts between multiple drones.

[0058] Mapping decisions are a multi-dimensional composite decision set, containing four interrelated core control parameters that jointly determine the execution method and resource allocation scheme of the mapping task within the current grid: First, the mapping switch state, which instructs the UAV to perform the task. This parameter is a binary decision variable used to control whether the UAV initiates the mapping inference process in the current time slot. When the mapping switch state is on, all subsequent mapping-related decision parameters take effect, and the UAV will execute environmental perception and diffusion model inference tasks. When the mapping switch state is off, the UAV only completes flight maneuvers and then enters a low-power standby state, without performing any mapping-related operations. Second, the total number of denoising steps, which controls the inference complexity of the diffusion model. This parameter is a positive integer and is a core adjustment variable balancing mapping accuracy and inference efficiency. The value of the total number of denoising steps directly determines the level of detail and computation time of the channel map generated by the diffusion model. The model dynamically adjusts this parameter based on overall latency requirements, currently available computing resources, and other factors, minimizing computation time while ensuring mapping accuracy requirements. The system offers several advantages: 1) Low inference computational load; 2) A cloud-edge task partitioning point used to divide the computational tasks of drones and servers. This parameter is a positive integer less than or equal to the total number of denoising steps. It defines the boundary between the computational tasks of drones and backend servers during the diffusion model denoising process. This parameter allows the complete diffusion model inference process to be split into two stages, completed collaboratively by the drone and server respectively. This shifts some of the computational load from the computationally limited drone to the computationally powerful backend server, significantly reducing the computational pressure on the drone. 3) A local offloading ratio for mapping tasks within the grid. This parameter is a continuous variable ranging from 0 to 1. It determines the ratio of mapping tasks executed locally by the drone to those offloaded to the server within the current grid to be mapped. The model dynamically adjusts this ratio based on the drone's remaining computing power and the real-time load of the backend server, achieving optimal matching between local and cloud computing resources and further optimizing the overall inference latency and energy consumption of the system. In one embodiment, the model can adaptively adjust the local offloading ratio based on the server's computing queue status. When the server's computing queue congestion increases, the model increases the local offloading ratio to reduce the number of mapping tasks offloaded to the server; when the server's computing queue congestion decreases, the model decreases the local offloading ratio to increase the number of mapping tasks offloaded to the server. Based on this, the cloud-edge collaborative computing architecture dynamically allocates computing tasks between the drone and the server through cloud-edge task partitioning points and local offloading ratios.

[0059] This application's embodiments consider systemic issues in large-scale edge deployments, refining the joint decision-making results into two major decisions: flight decision and mapping decision. This comprehensively covers the entire process control logic of UAV movement from the airspace to the execution of mapping tasks, realizing integrated optimization of UAV movement strategy and mapping execution strategy, and improving overall computing efficiency.

[0060] In one embodiment, the grid to be mapped includes multiple transmitters, and the environmental feature data includes environmental structural features and transmitter locations; the step of generating a local channel map of the grid to be mapped by cloud-edge collaborative computing between the UAV and the server, according to the corresponding mapping decision, and based on the environmental feature data, performing diffusion model inference collaboratively via cloud-edge computing, includes: Based on the local offloading ratio, all transmitters within the grid to be built are divided into a local execution group and an offloading execution group; For each transmitter in the local execution group, the UAV performs diffusion model inference based on the environmental structure features and the transmitter's location, according to the total number of denoising steps, to generate a local sub-channel map of the transmitter within the grid to be built. For each transmitter in the unloading execution group, the UAV and the server, based on the environmental structure features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task division points, perform phased cloud-edge collaborative execution of diffusion model inference to generate a local sub-channel map of the transmitter within the grid to be built; The local sub-channel maps of each transmitter within the grid to be constructed are merged to generate the local channel map of the grid to be constructed.

[0061] Specifically, a large number of signal transmitters are distributed within the target area. After the target area is discretized into multiple grids, each UAV flies and stays in a grid to be mapped in each time slot. By sensing the environmental structure features of the grid and the location of the transmitters within the grid, a high-resolution channel map is generated for the transmitters located in that grid. As multiple UAVs continuously and collaboratively move and sense within the grid area, the mapped grids and transmitters are gradually accumulated. Finally, all local reconstruction results are merged to form a global channel map covering the entire target area.

[0062] In one embodiment, the diffusion model is trained for a preset receiving point height, and the output local channel map is a horizontal slice channel map corresponding to the preset receiving point height. For channel map construction requirements at different receiving point heights, diffusion models trained at corresponding heights can be deployed as the inference engine of this application. The cloud-edge collaborative inference architecture, joint decision model and intermediate feature reuse mechanism described in this application are independent of the specific receiving point height corresponding to the diffusion model.

[0063] Diffusion model inference is essentially an iterative denoising process. The classic RadioDiff model requires hundreds of steps to reconstruct a complete channel map from pure noise. The total computational cost of a single inference is approximately proportional to the number of steps, which is difficult to meet the real-time mapping requirements of UAV platforms.

[0064] The first type is skip-step inference, which reduces the number of denoising iterations. Figure 3 This is a schematic diagram of the skip-step reasoning and intermediate feature reuse mechanism provided in this application, such as... Figure 3 As shown, the RadioDiff native diffusion model performs multi-step denoising along a fixed discrete time axis: each step depends on the intermediate state of the previous step. Figure 3 The original denoising trajectory (consisting of circular nodes along the trajectory within the latent space) converges stepwise from pure noise to a clear channel distribution, often requiring hundreds of steps. Skip-step inference aims to convert several small denoising steps into a single large denoising step while maintaining the network structure, compressing the entire denoising trajectory onto fewer critical nodes, thus reducing the total FLOPs of a single inference by nearly an order of magnitude. However, the larger the denoising range of a single segment, the greater the fitting error to the original denoising trajectory, and the higher the accuracy loss introduced by each segment.

[0065] The second type involves the reuse of intermediate features among multiple transmitters within the same grid. In diffusion models, the intermediate latent variables are highly similar between adjacent denoising steps and between adjacent inputs with similar transmitter conditions during the initial denoising process—this part of the computation can be shared by multiple similar tasks. Specifically, when a UAV sequentially performs mapping on multiple transmitters within the same grid, the first transmitter is used as a reference to complete the denoising segment, and the resulting intermediate latent variables are then reused. Cache it; subsequent transmitters within this grid will directly use... The remaining denoising will begin from this starting point. For example... Figure 3 As shown, multiple similar tasks share the same set of intermediate states in the front end, and only perform independent denoising in the back end branches. However, the larger the number of segments skipped during reuse, and the further the conditions of the reused transmitter are from the reference transmitter, the more buffered the data becomes. The further the deviation from the actual conditions of the current transmitter, the more significant the output position offset becomes, and it is amplified step by step with subsequent denoising steps.

[0066] The accuracy degradation of both of the above mechanisms can be obtained by fitting measured data, see details below. Figure 4 , Figure 4 This is a schematic diagram illustrating the accuracy degradation curves of denoising step compression and partition point reuse provided in this application. For the transmitter using the grid as the feature reuse source (m=1), the mapping error is represented by the Normalized Mean Square Error (NMSE), which is mainly composed of the total number of denoising steps. The decision, its value The multiplexing characteristics are obtained from measured asynchronous counts. For the m-th (m≥2) transmitter within the grid, the multiplexing feature... Introducing additional multiplexing error, the increase of which is due to the number of denoising segments skipped during multiplexing. The decision is made to model it as follows: in, To control the overall degradation introduced by reuse; The coefficient is used to control the rate of degradation as a function of the reuse ratio; condition , This indicates that the degradation term is only processed for the second and subsequent transmitters within the grid, and that reuse inference is used for drones. That is, there is reusable This takes effect upon ( ). The number of skipped segments is reused. The larger the cache size, the more cached data is. The further the output deviates from the actual conditions of the transmitter, the more significant the output error.

[0067] Given that the above-mentioned skip-step reasoning framework is used at both ends, the total number of denoising steps in this application is [number missing]. Cloud edge task division points The mapping task is incorporated into the decision space to achieve an optimal trade-off between accuracy degradation and computational overhead. Each transmitter within the grid can choose between two modes: local execution or offloading to the server. Local tasks are completed independently by the UAV. Noise reduction is performed step-by-step, eliminating the need to transmit data to the server; unloading tasks are performed at a unified reuse point. The division, that is, the drone end completes the first step. Denoising and feature intermediate latent variables step by step Transmitted to the server via wireless link, and then from the server... Departure to complete remaining The channel map is obtained by segment denoising and decoding. Multiple offloading tasks within the same grid share the same set. The front end of the drone The step only needs to be executed once. Specific collaboration methods are as follows: Figure 5 As shown, Figure 5 This is a schematic diagram of the heterogeneous collaborative reasoning and partition point reuse process provided in this application.

[0068] The Actor branch output of the joint decision model in time slots drones The set of decision variables is: in, Indicates the direction of movement; The mapping switch is in operation. This represents the total number of noise reduction steps. Assign points to cloud-edge tasks; This represents the local uninstallation percentage. Instead of directly assigning a local / offload identifier to each transmitter, it is determined by the total number of transmitters in the current grid that are yet to be mapped. Calculate the number of local tasks Considering the varying importance of different transmitters within the region, transmitters of higher importance should be mapped earlier and with higher quality. Therefore, it is necessary to first sort them in descending order of transmitter value, and then select the top-ranked transmitters from this sorting result. One transmitter performs inference locally, while the remaining transmitters offload inference to a backend server.

[0069] Therefore, based on the local unloading ratio parameter, the task assignment of all transmitters to be mapped within the mapping grid can be determined, i.e., the previous... One transmitter was assigned to the local execution group, the remaining... Each transmitter is assigned to an offload execution group to ensure that local computing resources and cloud computing resources are utilized to the maximum extent.

[0070] For each transmitter in the local execution group, the UAV independently completes the entire diffusion model inference process. The UAV uses pre-perceived environmental structural features and the transmitter's location information as input conditions for the diffusion model, and executes... The noise reduction operation can be performed without data interaction with the backend server, and the local sub-channel map of the transmitter within the current grid to be built can be generated directly on the local machine.

[0071] For each transmitter in the offloading execution group, diffusion model inference is performed using a phased collaborative approach between the drone and the backend server. This is based on the total number of denoising steps. Cloud edge task division point The complete diffusion model denoising inference process is divided into two sequential computational stages. The first stage is executed locally by the UAV, while the second stage is executed by a backend server with more computing power. The UAV uses environmental structural features and the transmitter's location information as input. After completing the denoising calculation in the first stage, it then incorporates the intermediate latent variable features... Transmitted to the backend server via an air-to-ground wireless link; the backend server receives... Then, using this as the starting point for denoising, all denoising calculations for the remaining stages are completed, and a local sub-channel map of the transmitter within the current grid to be built is generated through decoding operations. This phased collaborative reasoning method shifts most of the computational load from the UAV, which is limited in both computing power and energy, to the backend server, significantly reducing the computational complexity and energy consumption of the UAV and extending the continuous operation time of the UAV.

[0072] After generating local sub-channel maps for all transmitters within the current grid to be mapped, the local sub-channel maps of all transmitters are merged according to a preset spatial coordinate relationship to generate a local channel map covering the entire grid to be mapped. This local channel map can comprehensively and accurately characterize the core channel characteristics such as signal strength, path loss, and delay spread of any spatial location within the current grid relative to each transmitter, and is the basic data unit for constructing a global channel map of the target area.

[0073] This application embodiment rationally splits and allocates the multi-transmitter mapping task within a single grid, adopting two task execution modes: independent operation of UAVs and collaborative operation of UAVs and servers. This fully leverages the edge perception and local computing advantages of UAVs while effectively utilizing the powerful computing resources of the backend servers, achieving an optimal balance between mapping accuracy, inference latency, and system energy consumption.

[0074] In one embodiment, for each transmitter in the offloading execution group, the process involves using the UAV and the server, based on the environmental structural features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task partitioning points, to perform phased cloud-edge collaborative execution of diffusion model inference to generate a local sub-channel map of the transmitter within the grid to be constructed, including: Based on the cloud-edge task partitioning points, the total number of denoising steps is divided into the first denoising steps performed by the drone and the second denoising steps performed by the server. For the first transmitter in the unloading execution group, the UAV performs diffusion model inference based on the environmental structure features and the location of the first transmitter, according to the first denoising step count, to generate intermediate latent variable features; the server performs diffusion model inference based on the intermediate latent variable features, according to the second denoising step count, to generate a local sub-channel map of the first transmitter. For the remaining transmitters in the unloading execution group, the server, based on the intermediate latent variable features, performs diffusion model inference according to the second denoising step number to generate local sub-channel maps for the remaining transmitters. The first transmitter is the most valuable transmitter among all transmitters in the offloading execution group; the remaining transmitters are the transmitters in the offloading execution group other than the first transmitter.

[0075] Specifically, for each transmitter in the unloading execution group, based on the cloud-edge task partitioning point, the total number of denoising steps corresponding to this mapping task is divided into a first denoising step executed by the UAV and a second denoising step executed by the server. The first denoising step is equal to the value of the cloud-edge task partitioning point, and the second denoising step is equal to the difference between the total denoising steps and the first denoising step. The UAV executes the initial denoising process corresponding to the first denoising step before the server executes the subsequent denoising process corresponding to the second denoising step, and the subsequent denoising process uses the intermediate latent variable features generated by the initial denoising process as its sole input.

[0076] For the first transmitter in the offloading execution group, a complete cloud-edge collaborative inference process is executed. The first transmitter is the most valuable transmitter in the group; prioritizing high-value transmitters ensures the quality and timeliness of mapping in the core area. The UAV uses pre-perceived environmental structural features and the transmitter's location information as input to the diffusion model. Following the first denoising step, it performs front-end denoising inference to generate intermediate latent variable features containing global environmental semantic information, which are then cached in the UAV's local memory. Simultaneously, the UAV transmits these features to the back-end server via an air-to-ground wireless link. Upon receiving these features, the back-end server uses them as the starting point for denoising, performing the remaining back-end denoising inference according to the second denoising step. It then uses decoding to generate a local sub-channel map of the high-value transmitter within the current mapping grid.

[0077] Finally, for the remaining transmitters in the offloading execution group, a feature-reuse collaborative inference process is executed. These remaining transmitters are all those in the offloading execution group excluding the first high-value transmitter, and they are also denoised according to the principle of building maps for high-value transmitters first. Since all transmitters within the same grid share the same environmental structure features, and the front-end denoising process of the diffusion model is mainly used to encode the global semantic information of the environment, the remaining transmitters do not need to repeat the front-end denoising calculation. The back-end server uses the intermediate latent variable features as the starting point for denoising, executes the back-end denoising inference according to the second denoising step, and then generates the local sub-channel maps of the remaining transmitters through decoding operations.

[0078] As an example, consider a drone within a certain time slot. After entering the grid to be mapped, a typical scenario of three transmitters to be mapped was detected. The decision for this time slot is... , For the first transmitter: the UAV first completes the first two stages of noise reduction and outputs the intermediate latent variable features. ,Will Transmitted to the server via an air-to-ground wireless link; the server from After completing the remaining three denoising segments, the channel map for the first transmitter is obtained through decoding. For the latter two transmitters: the UAV end does not repeat the first two denoising segments, but directly uses the same... As the starting point, the server will complete the remaining three segments. Figure 6 This is a schematic diagram illustrating the example of heterogeneous collaboration and partition point multiplexing execution provided in this application for three transmitters. Figure 6 The Gantt chart in the image shows the time allocation between the drone and server sides for this time slot—the drone side only needs to complete the first two denoising segments once, while the server side, after receiving the intermediate latent variable features, uses these as a starting point to process the three subsequent denoising segments. The entire process is as follows: Figure 6 As shown.

[0079] This application embodiment adopts a highly efficient collaborative reasoning architecture of "one front-end calculation and multiple back-end reuse" for each transmitter in the unloading execution group. While ensuring the accuracy of mapping, it eliminates redundant calculations of mapping tasks for multiple transmitters in the same grid to the greatest extent, and further reduces the computing load on the UAV.

[0080] In one embodiment, the joint decision-making model is trained in the following manner: Non-spatial scalar samples and spatial tensor samples are encoded respectively to obtain the first embedding feature and the second embedding feature, and the first embedding feature and the second embedding feature are concatenated to construct the state vector; Based on flight decision samples and mapping decision samples, construct action vectors; A reward function is constructed based on the globally weighted mapping revenue and the total system energy consumption; the globally weighted mapping revenue is obtained by summing the single mapping revenue of each transmitter; the total system energy consumption is obtained by summing the computing energy consumption of the UAV, the flight energy consumption, and the computing energy consumption of the server. Based on the state vector, the action vector, and the reward function, a reinforcement learning algorithm is used to adaptively optimize and train the joint decision.

[0081] Specifically, during the training phase, the proportion of continuous local unloading will be... After sorting the transmitters within the grid in descending order of their value, the projection is the local / offload binary identifier for each transmitter. (0 indicates local execution, 1 indicates uninstall execution), and simultaneously set the mapping switch status. Integrating into the direction of movement (That is, the "offline" action corresponds to the graph construction being closed). Therefore, the equivalent action vector used for training is: in, Actions are defined in five discrete directions (up, down, left, right, and down line; "down line" indicates that the drone actively stops flying to reduce energy consumption). This refers to the level of the total number of noise reduction steps; The cloud edge task partitioning point is the reuse point where tasks are unloaded locally to the server; For each transmitter within the grid The local / uninstall binary identifier (0 indicates local execution, 1 indicates uninstall execution). It is a grid Internal transmitter assembly.

[0082] The system optimization objective is to maximize the globally weighted mapping benefits while constraining the total system energy consumption. Where β is the energy consumption penalty coefficient; For drones In the time slot Mapping benefits for all transmitters within the grid; For drones In the time slot The total energy consumption, that is, the total energy consumption of the system; The number of time slots; This refers to the number of drones.

[0083] The following section describes the calculation methods for mapping benefits and total energy consumption.

[0084] To represent the mapping progress of each transmitter, a binary identifier is introduced. Initially, the identifier for all transmitters is 0. It is set to 1 after a drone successfully maps the corresponding grid, and no further benefits are generated. The channel map construction uses the RadioDiff diffusion model as its inference framework. The model takes the environmental structure features within the grid and the transmitter's location as input to reconstruct the channel distribution of that transmitter within the grid area. (Grid) The set of transmitters within is denoted as .

[0085] The benefit of a single mapping operation is determined by three factors: the importance of the transmitter itself, the mapping accuracy (determined by the NMSE model mentioned above), and the timing of mapping completion (characterized by a monotonically decreasing delay discount factor). (UAV) In the time slot For the transmitter The revenue from a single map building operation is defined as: in, Assigning importance weights to the transmitter itself; The normalized mean square error for this mapping is given by the NMSE model mentioned above; The maximum tolerance NMSE preset for the system is factored That is, the accuracy factor of this mapping; For transmitter The total physical time from the start to the completion of this mapping project. Let be the delay attenuation constant. The above product form incorporates transmitter importance, mapping accuracy, and completion delay into the single mapping gain, thereby guiding the strategy to prioritize coverage of important transmitters while also considering mapping accuracy and completion delay.

[0086] Finally, by summing the single-cycle mapping gains of all transmitters within the grid, the drone can be obtained. In the time slot Mapping benefits for all transmitters within the grid.

[0087] To support optimization modeling, the overhead expressions for end-to-end latency, communication link, and system energy consumption are provided. Transmitter In the time slot The end-to-end build latency consists of four parts, corresponding to UAV local computation, intermediate result transmission (i.e., intermediate latent variable feature transmission), server queuing, and server computation: in, and These are the single-step denoising inference latency for the drone and server sides, respectively, which are determined by the corresponding hardware performance. For intermediate result transmission delay; This refers to the delay while waiting for the server task queue.

[0088] Server configuration Each GPU maintains an independent task queue, which uses a first-come, first-served (FIFO) approach. For time slots At the beginning The sum of the remaining processing time of unfinished tasks on each GPU, and the allocation of newly arriving tasks to the GPU with the lowest current load. .

[0089] The drone and the server use an air-to-ground channel model with channel gain. The location is determined by both the drone's current grid position and the server's ground position. Each drone uses a fixed transmission power. Transmission, total system bandwidth The time slot is evenly divided among the drones with unloading tasks in the current time slot, assuming the drones... The allocated bandwidth is The transmission rate is modeled according to the Shannon capacity formula: in, Let be the noise power spectral density. When... At that time, the drone needs to The transmission to the server has an intermediate result transmission latency of [time value missing]. , This represents the size of the intermediate latent variable.

[0090] The total energy consumption of the system is calculated by the drone terminal. Server-side computing power consumption Energy consumption of drone flight It consists of three parts, which are related to energy consumption, operating time, and power. Specifically: in, and These are the computing power of the drone-side and server-side computing devices, respectively, which are determined by the actual hardware performance. For the flight power of the drone; and These represent the actual computation time occupied by the drone and the server within that time slot, respectively, influenced by factors such as task concurrency within the grid, local / offload partitioning, and server GPU queue scheduling. (Drone) In the time slot Total energy consumption is .

[0091] When solving for the optimization objective, the set of decision variables must satisfy the following constraints: (1) Mesh boundary constraints: the grid position of each UAV in each time slot It must fall within the defined G×G target region, that is ; (2) Constraints on the set of legal discrete actions: direction of movement ∈{top, bottom, left, right, bottom line}, mapping switch status Total number of noise reduction steps Cloud edge task allocation point Local / Uninstallation identifier .

[0092] This joint optimization problem involves discrete decision coupling in all dimensions: flight decisions are aware of the transmitter set and channel quality; local / offload flags... Determine the computational allocation; total number of denoising steps Cloud edge task division point The accuracy model indirectly affects the gains of subsequent transmitters within the same grid. In the time dimension, the mapping status... Monotonous updates occur; mapped transmitters no longer contribute revenue. The system terminates when all transmitters have been mapped, all drones have been taken offline, or the maximum number of time slots (T) has been reached.

[0093] To solve the aforementioned non-convex integer stochastic optimization problem, this application models it as a finite-step Markov Decision Process (MDP), using tuples... Describe the system dynamics. State Includes all observable information describing the current global operational status of the system; actions Aggregate decision variables for all UAVs in the current time slot, including movement, mapping, total number of denoising steps, reuse points, and local / unload flags; state transitions are determined by the post-flight grid position and mapping status. The update depends on the transmitter corresponding to the drone's location; Rewards Taken from the instantaneous value of the joint optimization objective in the current time slot; The discount factor is used. The modeling objective is to learn a stochastic policy. This strategy increases the expected cumulative discount return. This maximizes the ability to adaptively coordinate the flight, mapping, and inference decisions of multiple transmitters under different transmitter distributions, channel conditions, and server loads.

[0094] Typical methods of deep reinforcement learning include Deep Q-Network (DQN) based on value functions, Deep Deterministic Policy Gradient (DDPG) for continuous control, and Actor-Critic methods. DQN is only applicable to discrete action spaces and struggles to handle continuous resource allocation variables. While DDPG supports continuous control, it is sensitive to hyperparameters and suffers from poor training stability. Standard Actor-Critic methods are prone to performance collapse when policy update steps are too large, and convergence is difficult to guarantee in complex multi-objective scenarios. Proximal Policy Optimization (PPO), as an online policy method based on policy gradients, constrains the magnitude of each policy update by pruning the objective function, thus avoiding policy collapse caused by excessively large single-step updates. Compared to methods like DDPG, it is less sensitive to hyperparameters, has a more stable training process, and naturally supports mixed discrete and continuous action spaces, making it one of the mainstream reinforcement learning algorithms in complex scenarios such as multivariate joint decision-making.

[0095] Therefore, for the above MDP, the Proximal Policy Optimization (PPO) algorithm is adopted as the underlying solution algorithm: the policy update magnitude is constrained by truncating the importance sampling ratio to ensure training stability, and the policy gradient variance is reduced by combining generalized advantage estimation (GAE). Figure 7 This is a schematic diagram of the PPO network architecture provided in this application. The overall network architecture is as follows: Figure 7 As shown, the state The input is divided into two paths: a non-spatial scalar and a spatial tensor, which are encoded separately. The two paths are then concatenated and fed into the Actor and Critic branches to output the action in parallel. With state value .

[0096] State encoding: Non-spatial scalar samples (server GPU queue length, normalized time slot progress, computational latency statistics of each UAV, air-to-ground channel gain, and pairwise distances between UAVs, etc.) are compressed into 64-dimensional first embedding features through two fully connected layers + ReLU; Spatial tensor samples are multi-channel raster images of N UAVs (using a G×G grid, superimposed with F semantic channels such as their own position, distribution of other UAVs, transmitter power and density, mapping completion status, and recent flight trajectories), which are mapped step by step through three layers of convolutional neural network (CNN) to the second embedding features with F→64→128 channels.

[0097] The Actor branch takes the state vector, obtained by concatenating the first and second embedded features, as input. It then passes through a three-layer fully connected network (256 dimensions + ReLU per layer) to extract shared hidden features. Finally, five independent action heads are connected in parallel at the end to sample the movement direction. Mapping switch status Total number of noise reduction steps Cloud edge task division points and the local uninstallation ratio .

[0098] The Critic branch takes the same set of state vectors as input and outputs scalar state values ​​through a three-layer fully connected network (each layer has 256 dimensions + ReLU). During the training phase, Critic stitches together the state features of all UAVs for centralized valuation. During the execution phase, each UAV makes action decisions based solely on its own local observations.

[0099] Training employs on-policy PPO: within each update cycle, the agent follows the current policy. After several rounds of interaction with the environment, the advantage is estimated using GAE. Then use PPO to cut off the proxy target Update the Actor, and regress the loss using the value function in smooth L1 form. Update the Critic; to prevent excessive deviation in a single update, calculate the approximate KL divergence at the end of each epoch and set a threshold for early stopping. The Actor and Critic are trained using independent Adam optimizers, and the entire process is implemented in Python + PyTorch.

[0100] Through continuous interaction with the environment, the agent can adaptively select flight direction, mapping switch, total number of denoising steps, reuse points and local offloading ratio under different transmitter distributions, channel conditions, server queue status and mapping completion conditions. In complex decision-making scenarios involving multi-machine collaboration, it can stably approach the above-defined goal of maximizing the global weighted mapping benefits.

[0101] This application's embodiments achieve simultaneous and accurate perception of system resource constraints and spatial task distribution through heterogeneous data encoding fusion; integrated optimization of flight and mapping decisions is realized based on unified action vectors, avoiding the traditional single decision-making mode; a dual-objective reward function based on mapping benefits and system energy consumption is adopted to guide the model to automatically balance mapping quality and operating costs; combined with a stable and efficient near-end policy optimization algorithm, the model's environmental adaptability and decision robustness are significantly improved, providing reliable technical support for the construction of low-power, high-precision channel maps in large-scale complex areas.

[0102] To verify the effectiveness and stability of the proposed PPO-based heterogeneous collaborative reasoning and partitioning point reuse method in multi-UAV wireless edge computing scenarios, this application conducted a systematic experimental evaluation based on a simulation platform, including training process convergence analysis, performance comparison with various baseline strategies, and sensitivity analysis for key environmental parameters.

[0103] The experimental environment of this application is based on a discrete grid scenario with multiple transmitters: five UAVs collaboratively perform channel mapping tasks, the target area is discretized into an 8×8 grid, and 180 fixed transmitters are distributed within the area; the GPU computing power of the UAVs and the back-end server is heterogeneous. Data transmission between the two ends is carried out through an air-to-ground wireless link to complete collaborative inference. The PPO agent is trained for 10,000 episodes under this configuration. Detailed system, hardware, accuracy, reward, and training parameters are listed in Table 1 below.

[0104] Table 1 Default System and Simulation Parameter Settings

[0105] Based on the above configuration, the PPO agent was trained over a long period of 10,000 episodes. The overall reward evolved with each episode as follows: Figure 8 As shown, Figure 8This is a schematic diagram of the PPO training reward curve provided in this application. It can be observed that: in the early stages of training, the reward is at a low level of approximately 50 with a large variance, corresponding to the high-variance exploration stage; as training progresses, the reward continues to rise monotonically, reaching the 200-250 range around episodes 4000-5000; after approximately 7000 episodes, the curve enters a stable plateau region of approximately 300, and the moving average curve of 200 episodes shows almost no significant fluctuations at the end. Throughout training, the best training reward is 367.51 (episode 9542), and the best validation set reward is 307.24 (episode 8899), with the episode reward stabilizing at approximately 326 at the end of training. This convergence trajectory indicates that the proposed bi-branch Actor-Critic network can converge stably in multi-type discrete action spaces and has strong robustness to perturbations such as random seeds.

[0106] To quantitatively evaluate the performance of the proposed method, the PPO agent of this application was first compared with four typical fixed baselines (fixed denoising step size, fully local computation, fully server offloading, and no reuse) in the same environment, in terms of comprehensive benefits, latency, energy consumption, and accuracy, as shown in Table 2 below. Figure 9 and Figure 10 , Figure 9 This is a schematic diagram showing the flight trajectory of the UAV provided in this application and the number of UAV tasks processed in each time slot. Figure 10 This is a schematic diagram illustrating the decision distribution of PPO under different environmental conditions provided in this application. The following section discusses multi-machine collaboration and path planning capabilities (see reference). Figure 9 ) and environmental adaptability (reference) Figure 10 We analyze the specific decision-making behaviors learned by PPO from two perspectives and provide a mechanistic explanation for its performance advantages.

[0107] Table 2. Overall performance comparison of PPO with four typical baselines.

[0108] Table 2 shows that PPO ranks first or second in all three key metrics: total revenue, latency, and energy consumption. In terms of total revenue, PPO achieves 289.85, exceeding the four baselines (fixed denoising step size, fully local computation, fully server offload, and no reuse) by 9.8%, 25.2%, 7.9%, and 9.1%, respectively. Furthermore, no single baseline surpasses PPO in any single metric. Regarding latency, PPO achieves 262.93s, reducing latency by approximately 73% and 50% compared to fully local and fixed denoising step sizes, respectively. This reflects its ability to dynamically select the appropriate total denoising step size and cloud-edge task partitioning point based on the number of transmitters and server queue status within the grid, avoiding the accumulated queuing latency in the baseline schemes. In terms of energy consumption, PPO achieves 58,181, only about 3.5% higher than fully server offload and about 18% lower than fully local, indicating that PPO maintains low latency without sacrificing additional local energy consumption at the UAV end. In terms of accuracy, PPO's NMSE is 0.00785, placing it in the middle of all baselines. This reflects its proactive sacrifice of some accuracy in exchange for latency and energy savings under multi-objective coupling. Overall, PPO outperforms static heuristic baseline methods in terms of total gain, latency, and energy consumption through coordinated adaptive adjustments to the total number of denoising steps, cloud-edge task partitioning points, and local offloading ratio.

[0109] On the other hand, to further characterize the decision-making behavior actually learned by the PPO agent after training, the analysis is carried out from two perspectives: first, whether it has learned reasonable path planning and load distribution in multi-UAV collaborative mapping tasks (refer to...). Figure 9 Secondly, whether key decision variables can be reasonably adaptively adjusted under different environmental conditions (see reference). Figure 10 ).

[0110] Figure 9 The flight trajectories of 5 drones in a typical episode and the number of tasks each drone handles in each time slot are given. Figure 9 (Left image) shows that five drones depart from the same takeoff point (box S) and disperse in different directions, respectively covering five transmitter clusters in the upper left, upper right, center, lower left and lower right of the area. The flight paths hardly overlap, showing a collaborative strategy of "spontaneous division of labor according to area". When a drone has covered all the high-value transmitters in its area of ​​responsibility, it actively chooses to go offline to save energy.

[0111] Figure 9(Right figure) shows the number of per-UAV tasks processed in each time slot within the same episode (based on UAV stacking). It can be observed that: in the early stage of the episode (approximately time slots 0–10), the workload of multi-UAV parallel processing surges, reaching over 20 transmitters per step, corresponding to the initial stage of flying towards high-density clusters; in the middle stage (approximately time slots 10–20), the workload gradually decreases, reflecting the decreasing number of remaining unmapped transmitters in the area as mapping progresses; in the final stage (after approximately time slot 20), most UAVs complete their tasks in their respective responsibility areas and actively go offline, with only a small number of UAVs processing scattered transmitters in the edge areas. The above trajectory and load evolution together demonstrate that the PPO agent can adaptively learn a multi-UAV cooperative strategy of "division of labor coverage + stop upon completion" solely through reward signals, without explicit spatial allocation rules, exhibiting good path planning capabilities.

[0112] To verify the adaptive decision-making ability of PPO under different environmental conditions, the key decision variables of the trained policy on a large-scale evaluation set were averaged according to "contextual conditions," and the results are as follows: Figure 10 As shown, the three panels respectively examine the impact of server queue congestion on offloading and reuse decisions, and the impact of the highest transmitter value within the grid on the selection of the total number of denoising steps.

[0113] Figure 10 The results of the training policy on a large-scale evaluation set, with key decision variables averaged according to context conditions, are presented to reflect the adaptive decision-making ability of PPO under different environmental conditions. Figure 10 (a) Displays the local processing ratio As the server queue congestion level increased, the value rose from 0.224 to 0.331, an increase of approximately 48%, indicating that PPO will proactively reduce offloading and increase local execution after sensing server queuing pressure to avoid tasks queuing on the server side. Figure 10 (b) Display the reuse ratio Similarly, as the server queue congestion level monotonically increased from 0.606 to 0.667, it indicates that PPO increases the depth of intermediate feature reuse among multiple transmitters in the same grid under congested conditions, and retains more front-end computation to the UAV end shared cache to reduce the cost of congestion. Figure 10 (c) shows the average number of denoising steps when the highest transmitter value within the grid is low. It tends to complete low-value tasks in fewer steps; when a high-value transmitter appears in the grid... The score improved to approximately 3.70, an increase of about 18%, indicating that PPO automatically allocates more denoising steps to high-value transmitters to ensure mapping accuracy. The three bucketing methods collectively demonstrate that PPO has learned an intuitive conditional response strategy—adaptively adjusting offloading, reuse, and accuracy—based on environmental contexts such as server load and transmitter value, reflecting a fine-grained awareness and trade-off capability regarding system state.

[0114] The aforementioned experiments demonstrate that the proposed PPO-based joint decision-making method can achieve stable convergence in complex multi-machine collaborative environments (see reference). Figure 8 It can learn finer-grained decision-making behaviors (see reference). Figure 9 and Figure 10 At the multi-drone collaboration level, the agent spontaneously learns a collaborative pattern of "regional division of labor + stopping upon completion," without explicitly specifying the responsibility area or termination condition for each drone. At the decision-making trade-off level, the agent can dynamically adjust the offloading ratio, reuse depth, and denoising steps based on different environmental signals such as server queue status, channel quality, and transmitter value, adaptively balancing accuracy, latency, and energy consumption—a capability that fixed-rule heuristics or offline optimization methods struggle to achieve simultaneously. This adaptive characteristic gives the proposed method good robustness and scalability to changes in environment and task load, allowing it to be used as a unified solution framework under different transmitter distributions, mesh sizes, and hardware heterogeneity conditions.

[0115] This application proposes a joint optimization scheme for cloud-edge hierarchical inference and intermediate feature reuse, based on a diffusion model, for channel map construction tasks in multi-UAV collaborative mapping scenarios. Its beneficial effects are mainly reflected in the following four aspects.

[0116] First, a denoising step compression mechanism based on skip-step inference is adopted, which compresses the original long iterative denoising process of hundreds of steps by more than an order of magnitude. While maintaining the accuracy, it significantly reduces the computational load of a single mapping, providing a feasible technical path for UAV platforms with limited computing power to perform diffusion model inference.

[0117] Second, a cloud-edge collaborative hierarchical inference architecture is proposed, which integrates the UAV terminal and the back-end server, and introduces cloud-edge task partitioning points. Compared to local uninstallation ratio As two types of adjustment variables for cloud-edge computing allocation: when the server queue is congested, the offloading should be reduced and the local processing ratio should be increased; when the server load conditions are good, the offloading should be done to the server as much as possible to utilize its higher computing power, so as to achieve dynamic balance between onboard computing power and server resources.

[0118] Third, for the typical situation of multiple transmitters in the same region, the characteristics of the intermediate latent variables in the front end are introduced for the first time. The cross-transmitter multiplexing mechanism is incorporated into the joint decision-making framework, significantly reducing the redundant computational overhead of multi-transmitter mapping within the grid; simultaneously, the bias degradation introduced by multiplexing is explicitly characterized in the accuracy model, in relation to the cloud-edge task partitioning points. The choice leads to coupling optimization.

[0119] Fourth, taking the transmitter's mapping state as the core state variable, the flight, inference segmentation, reuse, and unloading decisions are uniformly modeled as a multi-slot Markov decision process, and solved online adaptively using a PPO reinforcement learning framework. Experimental results show that the proposed method has significant advantages over various typical fixed and heuristic baselines in terms of comprehensive reward indicators. Furthermore, it can spontaneously learn multi-machine collaborative strategies with division of labor coverage and fine-grained environmental adaptability based on server load, transmitter distribution, and value structure, demonstrating high engineering practicality and application promotion value.

[0120] Figure 11 This is a schematic diagram of the channel map construction device provided in this application.

[0121] like Figure 11 As shown, the channel map construction device includes: Deployment module 1110 is used to divide the target area into multiple grids and deploy N UAVs in the multiple grids to collaboratively perform mapping tasks. The joint decision module 1120 is used to input system operation status information related to the mapping task into the joint decision model based on reinforcement learning, and obtain the joint decision results of each UAV output by the joint decision model; the joint decision results include flight decisions and mapping decisions; The inference task execution module 1130 is used to control each UAV to move according to the corresponding flight decision, reach the grid to be mapped, and perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the UAV and the server, according to the corresponding mapping decision, based on the environmental feature data, cloud-edge collaborative execution of diffusion model inference, and generating a local channel map of the grid to be mapped. The channel map generation module 1140 is used to merge the local channel maps of each grid to generate a global channel map of the target area.

[0122] The channel map construction device provided in this application globally optimizes the flight path planning and mapping execution strategy of UAVs through a joint decision-making model based on reinforcement learning. It can dynamically output the optimal decision for each UAV based on the real-time system operation status, and then execute the diffusion model inference process based on the optimal decision. This process adopts an architecture of collaborative inference between UAVs and backend servers, and rationally distributes the computational tasks of the diffusion model at both the cloud and edge ends. It makes full use of the powerful computing power of the server and the flexible mobile perception advantages of UAVs, which greatly reduces the computational load and energy consumption pressure on the UAV end, and realizes efficient collaborative operation of multiple UAVs. While ensuring the accuracy of channel map construction, it significantly shortens the overall mapping latency of large-scale complex areas, and can overcome the bottleneck that diffusion models are difficult to deploy on UAV platforms with limited computing power and energy.

[0123] In one embodiment, the system operating status information includes non-spatial scalar information and spatial tensor information; The non-spatial scalar information includes at least one of the following: server's computing queue status, task progress time, UAV computing latency statistics, channel status between UAV and server, and relative distance between UAVs. The spatial tensor information is a multi-channel raster map constructed based on the multiple grids. Each UAV corresponds to an independent multi-channel raster map. The multi-channel raster map is superimposed with at least one of the following semantic channels: UAV's own position, distribution of other UAVs, transmitter power and density, mapping completion status, and flight trajectory within a preset historical period.

[0124] In one embodiment, the flight decision includes: indicating the direction of movement for the drone's flight; The mapping decision includes: mapping on / off status for indicating the UAV's task execution, total number of denoising steps for controlling the inference complexity of the diffusion model, cloud-edge task partitioning points for dividing UAV and server computing tasks, and local offloading ratio for indicating mapping tasks within the grid.

[0125] In one embodiment, the grid to be mapped includes multiple transmitters, and the environmental feature data includes environmental structural features and transmitter locations; the process of generating a local channel map of the grid to be mapped by cloud-edge collaborative computing between the UAV and the server, according to the corresponding mapping decision, and based on the environmental feature data, by cloud-edge collaboratively executing diffusion model inference, includes: Based on the local offloading ratio, all transmitters within the grid to be built are divided into a local execution group and an offloading execution group; For each transmitter in the local execution group, the UAV performs diffusion model inference based on the environmental structure features and the transmitter's location, according to the total number of denoising steps, to generate a local sub-channel map of the transmitter within the grid to be built. For each transmitter in the unloading execution group, the UAV and the server, based on the environmental structure features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task division points, perform phased cloud-edge collaborative execution of diffusion model inference to generate a local sub-channel map of the transmitter within the grid to be built; The local sub-channel maps of each transmitter within the grid to be constructed are merged to generate the local channel map of the grid to be constructed.

[0126] In one embodiment, for each transmitter in the offloading execution group, the process involves generating a local sub-channel map of the transmitter within the grid to be built, using the UAV and the server, based on the environmental structural features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task partitioning points, through phased cloud-edge collaborative execution of the diffusion model inference, including: Based on the cloud-edge task partitioning points, the total number of denoising steps is divided into the first denoising steps performed by the drone and the second denoising steps performed by the server. For the first transmitter in the unloading execution group, the UAV performs diffusion model inference based on the environmental structure features and the location of the first transmitter, according to the first denoising step count, to generate intermediate latent variable features; the server performs diffusion model inference based on the intermediate latent variable features, according to the second denoising step count, to generate a local sub-channel map of the first transmitter. For the remaining transmitters in the unloading execution group, the server, based on the intermediate latent variable features, performs diffusion model inference according to the second denoising step number to generate local sub-channel maps for the remaining transmitters. The first transmitter is the most valuable transmitter among all transmitters in the offloading execution group; the remaining transmitters are the transmitters in the offloading execution group other than the first transmitter.

[0127] In one embodiment, the joint decision-making model is trained in the following manner: Non-spatial scalar samples and spatial tensor samples are encoded respectively to obtain the first embedding feature and the second embedding feature, and the first embedding feature and the second embedding feature are concatenated to construct the state vector; Based on flight decision samples and mapping decision samples, construct action vectors; A reward function is constructed based on the globally weighted mapping revenue and the total system energy consumption; the globally weighted mapping revenue is obtained by summing the single mapping revenue of each transmitter; the total system energy consumption is obtained by summing the computing energy consumption of the UAV, the flight energy consumption, and the computing energy consumption of the server. Based on the state vector, the action vector, and the reward function, a reinforcement learning algorithm is used to adaptively optimize and train the joint decision.

[0128] In one embodiment, the reinforcement learning algorithm is a proximal policy optimization algorithm.

[0129] In one embodiment, the drone and the server constitute a cloud-edge collaborative computing architecture. This architecture dynamically allocates computing tasks between the drone and the server using the cloud-edge task partitioning point and the local offloading ratio. The joint decision-making model adaptively adjusts the local offloading ratio based on the server's computing queue status. When the server's computing queue congestion increases, the joint decision-making model increases the local offloading ratio to reduce the number of mapping tasks offloaded to the server. When the server's computing queue congestion decreases, the joint decision-making model decreases the local offloading ratio to increase the number of mapping tasks offloaded to the server.

[0130] It should be noted that the channel map construction apparatus provided in this application can execute the channel map construction method described in any of the above embodiments during specific operation, and this embodiment will not elaborate on this.

[0131] Figure 12 This is a schematic diagram of the structure of the electronic device provided in this application, such as... Figure 12As shown, the electronic device may include: a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240, wherein the processor 1210, the communications interface 1220, and the memory 1230 communicate with each other through the communication bus 1240. The processor 1210 can call logical instructions in the memory 1230 to execute a channel map construction method. This method includes: dividing the target area into multiple grids; deploying N drones in the multiple grids to collaboratively perform mapping tasks; inputting system operating status information related to the mapping task into a reinforcement learning-based joint decision model to obtain the joint decision result of each drone output by the joint decision model; the joint decision result includes flight decisions and mapping decisions; controlling each drone to move according to the corresponding flight decision to reach the grid to be mapped and perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the drones and the server, according to the corresponding mapping decisions, performing diffusion model inference based on the environmental feature data to generate a local channel map of the grid to be mapped; merging the local channel maps of each grid to generate a global channel map of the target area.

[0132] Furthermore, the logical instructions in the aforementioned memory 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] On the other hand, this application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the channel map construction method provided in the above embodiments. The method includes: dividing a target area into multiple grids, deploying N drones in the multiple grids to collaboratively perform a mapping task; inputting system operation status information related to the mapping task into a joint decision model based on reinforcement learning to obtain the joint decision result of each drone output by the joint decision model; the joint decision result includes flight decision and mapping decision; controlling each drone to move according to the corresponding flight decision to reach the grid to be mapped and to perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the drones and the server, according to the corresponding mapping decision, performing diffusion model inference based on the environmental feature data to generate a local channel map of the grid to be mapped; merging the local channel maps of each grid to generate a global channel map of the target area.

[0134] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the channel map construction method provided in the above embodiments. The method includes: dividing a target area into multiple grids and deploying N drones in the multiple grids to collaboratively perform a mapping task; inputting system operation status information related to the mapping task into a joint decision model based on reinforcement learning to obtain the joint decision result of each drone output by the joint decision model; the joint decision result includes flight decision and mapping decision; controlling each drone to move according to the corresponding flight decision to reach the grid to be mapped and to perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the drones and the server, according to the corresponding mapping decision, performing diffusion model inference based on the environmental feature data to generate a local channel map of the grid to be mapped; merging the local channel maps of each grid to generate a global channel map of the target area.

[0135] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0136] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for constructing a channel map, characterized in that, The channel map construction method includes: The target area is divided into multiple grids, and N drones are deployed in the multiple grids to collaboratively perform mapping tasks. The system operation status information related to the mapping task is input into a joint decision-making model based on reinforcement learning to obtain the joint decision-making results of each UAV output by the joint decision-making model; the joint decision-making results include flight decisions and mapping decisions; Each drone is controlled to move according to the corresponding flight decision, reach the grid to be mapped, and perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the drone and the server, according to the corresponding mapping decision, based on the environmental feature data, cloud-edge collaborative execution of diffusion model inference is performed to generate a local channel map of the grid to be mapped. The local channel maps of each grid are merged to generate a global channel map of the target area.

2. The channel map construction method according to claim 1, characterized in that, The system operating status information includes non-spatial scalar information and spatial tensor information; The non-spatial scalar information includes at least one of the following: server's computing queue status, task progress time, UAV computing latency statistics, channel status between UAV and server, and relative distance between UAVs. The spatial tensor information is a multi-channel raster map constructed based on the multiple grids. Each UAV corresponds to an independent multi-channel raster map. The multi-channel raster map is superimposed with at least one of the following semantic channels: UAV's own position, distribution of other UAVs, transmitter power and density, mapping completion status, and flight trajectory within a preset historical period.

3. The channel map construction method according to claim 2, characterized in that, The flight decision includes: a direction of movement to indicate the flight direction of the UAV; The mapping decision includes: mapping on / off status for indicating the UAV's task execution, total number of denoising steps for controlling the inference complexity of the diffusion model, cloud-edge task partitioning points for dividing UAV and server computing tasks, and local offloading ratio for indicating mapping tasks within the grid.

4. The channel map construction method according to claim 3, characterized in that, The grid to be mapped includes multiple transmitters, and the environmental feature data includes environmental structural features and transmitter locations. The process of generating a local channel map of the grid to be mapped, through cloud-edge collaborative computing between the UAV and the server, and according to the corresponding mapping decision, based on the environmental feature data and performing cloud-edge collaborative diffusion model inference, includes: Based on the local offloading ratio, all transmitters within the grid to be built are divided into a local execution group and an offloading execution group; For each transmitter in the local execution group, the UAV performs diffusion model inference based on the environmental structure features and the transmitter's location, according to the total number of denoising steps, to generate a local sub-channel map of the transmitter within the grid to be built. For each transmitter in the unloading execution group, the UAV and the server, based on the environmental structure features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task division points, perform phased cloud-edge collaborative execution of diffusion model inference to generate a local sub-channel map of the transmitter within the grid to be built; The local sub-channel maps of each transmitter within the grid to be constructed are merged to generate the local channel map of the grid to be constructed.

5. The channel map construction method according to claim 4, characterized in that, For each transmitter in the unloading execution group, the process involves using the UAV and the server, based on the environmental structure features and the transmitter's location, and according to the total number of denoising steps and the cloud-edge task partitioning points, to perform phased cloud-edge collaborative execution of diffusion model inference to generate a local sub-channel map of the transmitter within the grid to be built, including: Based on the cloud-edge task partitioning points, the total number of denoising steps is divided into the first denoising steps performed by the drone and the second denoising steps performed by the server. For the first transmitter in the unloading execution group, the UAV performs diffusion model inference based on the environmental structure features and the location of the first transmitter, according to the first denoising step count, to generate intermediate latent variable features; the server performs diffusion model inference based on the intermediate latent variable features, according to the second denoising step count, to generate a local sub-channel map of the first transmitter. For the remaining transmitters in the unloading execution group, the server, based on the intermediate latent variable features, performs diffusion model inference according to the second denoising step number to generate local sub-channel maps for the remaining transmitters. The first transmitter is the most valuable transmitter among all transmitters in the offloading execution group; the remaining transmitters are the transmitters in the offloading execution group other than the first transmitter.

6. The channel map construction method according to claim 4, characterized in that, The joint decision-making model is trained in the following way: Non-spatial scalar samples and spatial tensor samples are encoded respectively to obtain the first embedding feature and the second embedding feature, and the first embedding feature and the second embedding feature are concatenated to construct the state vector; Based on flight decision samples and mapping decision samples, construct action vectors; A reward function is constructed based on the globally weighted mapping revenue and the total system energy consumption; the globally weighted mapping revenue is obtained by summing the single mapping revenue of each transmitter; the total system energy consumption is obtained by summing the computing energy consumption of the UAV, the flight energy consumption, and the computing energy consumption of the server. Based on the state vector, the action vector, and the reward function, a reinforcement learning algorithm is used to adaptively optimize and train the joint decision.

7. The channel map construction method according to claim 6, characterized in that, The reinforcement learning algorithm is a near-end policy optimization algorithm.

8. The channel map construction method according to claim 4, characterized in that, The drone and the server constitute a cloud-edge collaborative computing architecture. This architecture dynamically allocates computing tasks between the drone and the server using the cloud-edge task partitioning point and the local offloading ratio. The joint decision-making model adaptively adjusts the local offloading ratio based on the server's computing queue status. When the server's computing queue congestion increases, the joint decision-making model increases the local offloading ratio to reduce the number of mapping tasks offloaded to the server; when the server's computing queue congestion decreases, the joint decision-making model decreases the local offloading ratio to increase the number of mapping tasks offloaded to the server.

9. A channel map construction device, characterized in that, The channel map construction device includes: The deployment module is used to divide the target area into multiple grids and deploy N drones in the multiple grids to collaboratively perform mapping tasks. The joint decision-making module is used to input system operation status information related to the mapping task into the reinforcement learning-based joint decision-making model to obtain the joint decision-making results of each UAV output by the joint decision-making model; the joint decision-making results include flight decisions and mapping decisions; The inference task execution module is used to control each UAV to move according to the corresponding flight decision, reach the grid to be mapped, and perceive the environmental feature data of the grid to be mapped; through cloud-edge collaborative computing between the UAV and the server, according to the corresponding mapping decision, based on the environmental feature data, cloud-edge collaborative execution of diffusion model inference is performed to generate a local channel map of the grid to be mapped. The channel map generation module is used to merge the local channel maps of each grid to generate a global channel map of the target area.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the channel map construction method as described in any one of claims 1 to 8.

11. A non-transitory computer-readable storage medium, wherein a computer program is stored on the non-transitory computer-readable storage medium, characterized in that, When the computer program is executed by a processor, it implements the steps of the channel map construction method as described in any one of claims 1 to 8.

Images