Unmanned aerial vehicle assisted method for generating artificial intelligence layered reasoning in internet of vehicles

Abstract

The present application relates to the technical field of Internet of Vehicles, in particular to a kind of multi-layer collaborative inference method in unmanned aerial vehicle assisted Internet of Vehicles environment.The present application designs the generative artificial intelligence multi-layer collaborative inference algorithm based on semantic perception in unmanned aerial vehicle assisted Internet of Vehicles environment.In the Internet vehicle computing environment assisted by unmanned aerial vehicle, based on communication model, inference delay model, content quality model and energy consumption model, the generative artificial intelligence multi-layer collaborative inference optimization problem based on semantic perception in unmanned aerial vehicle assisted Internet of Vehicles environment is established, to realize efficient balance between delay, energy consumption and content quality.The ADMM distributed algorithm is used to solve the proposed joint optimization problem, each vehicle node only solves its own subtask, greatly saves solving resources and delay, and realizes the optimal balance of energy consumption, delay and content quality.

Description

Technical Field

[0001] This invention relates to the field of vehicle networking technology, and in particular to a generative artificial intelligence multi-layer collaborative reasoning method based on semantic perception in a drone-assisted vehicle networking environment. Background Technology

[0002] With the widespread application of generative artificial intelligence (GAI) in connected vehicles, the demand for personalized content generation and real-time response in vehicles is constantly growing. However, cloud-based GAI services often introduce high communication and inference latency when faced with unstable wireless links, making it difficult to meet the low-latency requirements of in-vehicle tasks. At the same time, personalized generation services typically rely on large amounts of user data and high-intensity computing, which contradicts the limited computing and storage resources on the vehicle side, further exacerbating the bottleneck in task processing.

[0003] Against this backdrop, drones, as computing nodes with high mobility and aerial coverage capabilities, offer new ideas for task offloading and intelligent inference in the Internet of Vehicles (IoV). However, due to the limited energy and service range of drones, as well as the high-speed movement of vehicles, a single drone cannot continuously provide stable services to the same vehicle, making it difficult to achieve continuous execution of inference tasks. Therefore, in a drone-assisted IoV environment, how to flexibly allocate inference tasks and rationally divide inference steps among multiple nodes (vehicles, drones, edge servers), while balancing latency and energy consumption while ensuring the quality of content generation, has become an urgent problem to be solved.

[0004] While the invention patent application with publication number CN121037812A has solved this problem to some extent, it still has the following issues.

[0005] First: It only considered the end-to-end system latency, without considering the balance between latency and energy consumption.

[0006] Second: By using a two-layer method of "outer layer genetic algorithm for relay selection and inner layer convex optimization for resource allocation", the complex mixed integer nonlinear programming problem is transformed into a combination problem of discrete search and continuous convex optimization that can be solved efficiently. However, the heuristic algorithm of this method cannot guarantee finding the global optimal solution, and the results of multiple runs may be inconsistent. In addition, it is necessary to concentrate the channel information and data volume of all vehicles at the central node for optimization, which puts a heavy computational burden on dense vehicle scenarios.

[0007] Therefore, it is necessary to develop a generative artificial intelligence multi-layer collaborative reasoning method based on semantic perception in the drone-assisted vehicle network environment to improve the task reasoning efficiency in the drone-assisted vehicle network. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing a generative artificial intelligence multi-layer collaborative reasoning method based on semantic awareness in a drone-assisted vehicle-to-everything (V2X) environment. This method can simultaneously consider the optimal partitioning points of task semantic similarity, shared reasoning, and local reasoning, as well as the overall reasoning strategy of the vehicle, thereby effectively improving the service efficiency of V2X and achieving an optimal balance between latency, energy consumption, and the quality of generated content.

[0009] The technical solution of this invention is:

[0010] A generative artificial intelligence hierarchical reasoning method for drone-assisted vehicle networking, characterized in that: the semantic-aware generative artificial intelligence multi-layer collaborative reasoning method includes the following steps:

[0011] 1) Based on the communication model, inference latency model, content quality model and energy consumption model in the UAV-assisted vehicle network environment, a generative artificial intelligence multi-layer collaborative reasoning optimization problem based on semantic perception is proposed in the UAV-assisted vehicle network environment to achieve an efficient balance between latency, energy consumption and content quality.

[0012] 2) This optimization problem is complex and non-convex. It is difficult to find the optimal solution in polynomial time. To address this complex problem, McCormick envelope, Taylor first-order expansion, and relaxation methods are used for reconstruction.

[0013] 3) The Alternating Direction Method of Multipliers (ADMM) distributed algorithm is used to solve the proposed joint optimization problem. Each vehicle node only solves its own subtask, which greatly saves solution resources and latency.

[0014] 3.1 Obtain the initial environment state and derive the initial actions;

[0015] 3.2 Calculate the inference latency, generated content quality, and corresponding energy consumption for each task;

[0016] 3.3 The global optimization variables are split into local variables of each vehicle node and global consistency variables. An augmented Lagrangian function is constructed, and the consistency constraints are relaxed by Lagrange multipliers and quadratic penalty terms.

[0017] 3.4 In each iteration, each vehicle node solves its corresponding local subproblem in parallel and updates its local decision variables; then, it updates the global consistency variables and Lagrange multipliers through a small amount of information exchange.

[0018] The optimization problem of the above steps is specifically manifested as follows: , ; st; , ; , ; , ; , ; , ; ;

[0019] Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. This represents the reasoning steps for vehicle sharing. This represents the maximum number of reasoning steps. Indicates drone Maximum GPU power consumption. Constraint (2) represents the binary variable used to indicate the inference location. Constraint (3) indicates that vehicles within the same group share intermediate results, thus the number of shared denoising steps remains consistent within the group. Constraint (4) limits the number of shared inference steps and the total number of inference steps to avoid content generation failure, where The maximum number of denoising steps. Constraints (5) and (6) indicate that the shared denoising steps and the total denoising steps are integers. Constraint (7) indicates the GPU power capacity limit of the UAV.

[0020] The advantages of this invention are:

[0021] Current work on UAV-assisted vehicle-to-everything (V2X) environments focuses only on latency or content quality in edge computing environments, neglecting energy consumption and semantic similarity issues. To address these problems, this invention designs a semantically aware generative AI multi-layer collaborative reasoning algorithm. In a UAV-assisted V2X computing environment, based on communication models, inference latency models, content quality models, and energy consumption models, a semantically aware generative AI multi-layer collaborative reasoning optimization problem is established to achieve an efficient balance between latency, energy consumption, and content quality. Then, this complex problem is reconstructed using McCormick envelope, Taylor first-order expansion, and relaxation methods. Finally, the ADMM algorithm is used to solve the proposed joint optimization problem. Attached Figure Description

[0022] Figure 1This is a flowchart of the generative artificial intelligence multi-layer collaborative reasoning method based on semantic perception in the UAV-assisted vehicle network environment of the present invention.

[0023] Figure 2 This invention presents a joint optimization method model based on ADMM.

[0024] Figure 3 This is an example diagram of an unmanned aerial vehicle (UAV) assisted vehicle networking environment according to the present invention;

[0025] Figure 4 This is an example of resource collaborative scheduling and reasoning for disaster assessment tasks in emergency rescue, as described in this invention.

[0026] Figure 5 This is the convergence graph for step 1 in Example 1;

[0027] Figure 6 This is the convergence graph from step 1 of Example 1;

[0028] Figure 7 The figure shows the experimental results of step 3 in Example 1. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] The generative artificial intelligence hierarchical reasoning method for UAV-assisted vehicle networking, specifically the semantic-aware generative artificial intelligence multi-layer collaborative reasoning optimization method in the UAV-assisted vehicle networking environment, includes the following steps:

[0031] 1) Based on the communication model, inference latency model, content quality model and energy consumption model in the UAV-assisted vehicle network environment, a generative artificial intelligence multi-layer collaborative reasoning optimization problem based on semantic perception is proposed in the UAV-assisted vehicle network environment to achieve an efficient balance between latency, energy consumption and content quality.

[0032] The specific steps are as follows:

[0033] 1.1 Establish uplink and downlink communication models between vehicles, between vehicles and drones, and between drones and MEC servers;

[0034] In UAV-to-ground communication scenarios, we consider the impact of both line-of-sight (LAS) and non-LAS links, employing a probabilistic path loss model. (UAV) With vehicles The channel power gain between them can be expressed as:

[0035]

[0036] in, This represents the probability of a line-of-sight link. and Representing vehicles and drones Channel gain between line-of-sight and non-line-of-sight links. Channel gain of line-of-sight links. Channel gain of non-line-of-sight links It can be calculated using the following formula.

[0037]

[0038]

[0039] in This indicates a small-scale attenuation in line-of-sight links. This indicates a small-scale attenuation in non-line-of-sight links. This represents the path loss of a line-of-sight link. This represents the path loss of a non-line-of-sight link. and Shadow fading of line-of-sight links and non-line-of-sight links respectively.

[0040] Channel gain and This study integrates small-scale fading, path loss, and shadowing fading components in both line-of-sight and non-line-of-sight links. The modeling methods for each component are as follows. The Nakagami-m fading distribution is used to model small-scale fading; this model was chosen because of its strong parameter scalability and high agreement with measured channel data. Specifically, and They respectively obey the fading parameters as and Nakagami distribution, small-scale fading in line-of-sight and non-line-of-sight links and The probability density function is expressed as follows:

[0041]

[0042]

[0043] in, This represents the average received power in the fading envelope. For gamma function, and This is an expression for small-scale fading in line-of-sight and non-line-of-sight links that follows a Nakagami distribution of parameters.

[0044] The path loss between the vehicle and the drone under line-of-sight (LAS) and non-LAS links can be expressed as follows:

[0045] ,

[0046] ,

[0047] in, Indicates the carrier frequency. It's the speed of light. For reference distance, This represents the actual distance between the vehicle and the drone. and These correspond to the path loss exponents for line-of-sight (LOS) and non-line-of-sight (NOS) links, respectively. Shadow fading characterizes the signal attenuation caused by obstacle blockage, and is modeled as a zero-mean Gaussian distributed random variable, with the specific expression as follows:

[0048]

[0049] ,

[0050] in, and These represent the standard deviations of shadow fading for line-of-sight links and non-line-of-sight links, respectively. and For line-of-sight links and non-line-of-sight links, this refers to shadow fading.

[0051] It should be noted that the communication channel remains static within each time slot. Therefore, according to Shannon's theorem, from the vehicle... To drones The transmission rate can be expressed as:

[0052] ,

[0053] ,

[0054] in, It is a vehicle Uplink bandwidth, It is a vehicle Transmission power, It is Gaussian white noise. For vehicles and drones The channel gain of the transmission channel.

[0055] Similarly, from drones To the vehicle The downlink transmission rate can be expressed as:

[0056] ,

[0057] in, Indicates drone Transmission power, drone The bandwidth is represented as , For vehicles and drones The channel gain of the transmission channel, It is Gaussian white noise.

[0058] In addition, from drones The uplink data transfer rate to the MEC server is

[0059] .

[0060] in, Indicates drone Transmission power, drone The bandwidth is represented as , For drones and edge servers The channel gain of the transmission channel, It is Gaussian white noise.

[0061] Accordingly, from MEC servers to drones The downlink data transmission rate is

[0062] ,

[0063] in, The transmission power of the MEC server is represented by [insert value here], and the bandwidth of the MEC server is represented by [insert value here]. , For drones and edge servers The channel gain of the transmission channel, It is Gaussian white noise.

[0064] 1.2 Establishing an inference delay model

[0065] Given that each vehicle is equipped with semantic extraction capabilities, which convert raw data into text prompts, significantly reducing the latency of raw data transmission. The semantic extraction time is defined as follows:

[0066] ,

[0067] in, Indicates vehicle The amount of computation required during semantic extraction. Indicates vehicle The computing resources available.

[0068] The generated text prompts will then be sent to their associated drones for further analysis. (From the vehicle) To drones The uplink transmission delay is expressed as

[0069] ,

[0070] in, Represents a word element Size edge node Computing resources. Indicates by vehicle Generate text prompts The number of lexical units contained in it It is a vehicle and drones The uplink transmission rate.

[0071] drones Task knowledge graphs (TKG) are used to identify intents and relationships in task requests. Tasks with similar intents are grouped together using a classification algorithm. Therefore, vehicles exhibiting similar intents can be aggregated using this classification algorithm. The semantic analysis latency for analyzing semantic relationships can be expressed as:

[0072] ,

[0073] in, Indicates drone Processing by vehicle Generated text prompts The amount of computation required. It is a drone Analyze the computational resources required for this prompt word.

[0074] Furthermore, intermediate results are generated during the inference process, and these results can be pre-cached. Since both the drone and the edge server are equipped with caching units, we introduce a binary indicator variable. To represent drones Is it already the first The vehicle group caches the corresponding inference content. If If yes, it means the drone has cached the corresponding shared inference results, and the intermediate results can be directly transmitted to the vehicle group within its coverage area; otherwise... Similarly, another binary indicator variable This is used to indicate whether the edge server has cached intermediate inference results for this vehicle group. If yes, it means the intermediate results can be obtained directly from the edge server; otherwise... .

[0075] when The latency for returning intermediate inference results is

[0076] .

[0077] in, Indicates drone Cache vehicles The required size of intermediate inference results, It is a vehicle and drones Downlink transmission rate.

[0078] Because vehicles within the same group share the same intermediate results, using Indicates that the drone is the first The intermediate shared results of the group vehicle cache. From this, we can obtain... This means that the intermediate results of vehicles within the same group are the same.

[0079] when and The vehicle retrieves intermediate results from the edge server (ES). At this point, the edge server sends the results to the vehicle. The downlink transmission delay for transmitting intermediate results can be expressed as:

[0080] .

[0081] in, Indicates drone Cache vehicles The required size of intermediate inference results, It is a vehicle and drones downlink transmission rate For drones and edge servers Downlink transmission rate.

[0082] When neither the drone nor the edge server caches intermediate results, i.e. and The vehicles will obtain shared inference results from the most suitable entity (ES or UAV). Specifically, the selection of the service node depends on the node's computation latency, and each vehicle group will connect to the node that provides the lowest service latency.

[0083] Based on the above analysis, we set up an indicator variable. This is used to indicate whether the shared inference process is executed on the drone. If yes, it means the drone is performing the shared inference process; otherwise, it means the shared inference process is performed on the edge server.

[0084] .

[0085] As described earlier, determining the optimal allocation between shared and local denoising steps for a vehicle group is crucial. With a fixed total number of denoising steps, allocating too many steps to the shared inference stage, while improving content quality, is only a slight and limited improvement, effectively wasting significant computational resources. Conversely, insufficient shared inference steps necessitates compensation from local processes, potentially leading to decreased output quality and unacceptable latency for users. Furthermore, since each vehicle may require a different number of denoising steps due to customized needs, tailoring an appropriate total number of denoising steps for each user is equally important.

[0086] Therefore, we introduce Used for vehicle groups Shared denoising steps Indicates vehicle The total number of denoising steps. Additionally, since inference results are shared within the same vehicle group, therefore... For ease of representation, drones The next The shared reasoning process of each group is denoted as: The local reasoning process is denoted as .

[0087] when , and At this time, the UAV will transmit the analyzed prompts to the edge server for further shared inference. Therefore, the uplink transmission latency of uploading these results from the UAV to the ES for a certain vehicle group can be expressed as:

[0088] ,

[0089] in, Indicates in drones Vehicle groups within the service area The size of the analysis results It is a drone and edge servers The uplink transmission rate.

[0090] The time required to perform shared inference on a drone can be expressed as:

[0091] ,

[0092] in, Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. Used for vehicle groups The number of shared denoising steps. A binary indicator variable. To represent drones Is it already the first The vehicle group caches the corresponding inference content. Another binary indicator variable. This is used to indicate whether the edge server has cached intermediate inference results for this vehicle group.

[0093] when , and At that time, the edge server Elasticsearch performs the shared inference process, and the computation latency on the edge server is:

[0094] ,

[0095] in, This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result. Similarly, another binary indicator variable... This indicates whether the edge server has cached intermediate inference results for this vehicle group. Used for vehicle groups Shared denoising steps.

[0096] After the final shared inference is completed, subsequent denoising steps are performed locally by each vehicle. The latency incurred by the vehicle performing local inference can be expressed as:

[0097] ,

[0098] in, This represents the average computational cost required for each local inference step. and Representing vehicle groups Shared noise reduction steps and vehicles Total number of denoising steps, Indicates vehicle The computing resources available.

[0099] In summary, the vehicle group The reasoning delay can be derived as follows:

[0100] .

[0101] in, A binary variable indicating whether the drone has cached the corresponding shared inference result. Similarly, another binary indicator variable... Indicator variable used to indicate whether the edge server has cached intermediate inference results for this vehicle group. This is used to indicate whether the shared inference process is executed on the drone. and These represent the time required to perform shared inference on the drone and the time required to perform shared inference on the edge server, respectively. This indicates that semantically extracted content from vehicle groups is uploaded to the drone. The uplink transmission latency.

[0102] For vehicles The total computation delay is:

[0103] .

[0104] in, This indicates the latency of vehicle semantic extraction. Indicates from vehicle Upload to drone upload latency, Indicates the latency of semantic analysis. Indicates vehicle For vehicle clusters Local inference latency, a binary indicator variable Used to represent edge servers Has this vehicle group been... Caching intermediate inference results, indicator variables This is used to indicate whether the shared reasoning process occurs in the drone. Execute on, Indicates the time when the intermediate inference results are transmitted down. This indicates that intermediate inference results are downloaded from the edge server to the vehicle. The download latency.

[0105] Therefore, the inference delay for all vehicles can be deduced as follows:

[0106] .

[0107] in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total delay of inference.

[0108] 1.3 For each vehicle reasoning task, establish a content quality model.

[0109] In the vehicle During the denoising inference process, a series of intermediate points will be generated. Since the number of local denoising steps affects the quality of the generated content, local inference needs to be appropriately divided. Furthermore, considering the differences in needs between vehicles, the number of denoising steps required varies for different vehicles; therefore, the total number of steps must be tailored to different users. We use the following function to describe the relationship between inference steps and the quality of generated content. This relationship can be expressed as:

[0110] ,

[0111] in, Indicates standard deviation, and Representing vehicles Shared noise reduction steps and vehicles Total number of denoising steps.

[0112] The total content quality of all cars is:

[0113]

[0114] in, The content quality function representing vehicle reasoning.

[0115] 1.4 Energy efficiency is crucial for resource-constrained equipment such as drones and vehicles. Therefore, energy consumption models should be established for drones and vehicles.

[0116] Performing computationally intensive AIGC tasks results in significant energy consumption, making energy consumption a critical factor affecting system performance. These tasks are distributed across a heterogeneous computing architecture between the UAV and the vehicle, and are processed using GPUs.

[0117] In this article, we do not consider the static power consumption of the GPU, but only its dynamic power consumption. Furthermore, the operating power consumption of the GPU can be expressed by the following formula:

[0118] ,

[0119] in, and These represent the GPU's core voltage, core clock speed, and memory clock speed, respectively. (Coefficient) and The values ​​are fixed and depend on the hardware configuration and specific application. These coefficients respectively characterize the impact of memory frequency adjustment and processor core voltage / frequency adjustment on power consumption.

[0120] Will Designed as a performance function to represent the performance of GPU-accelerated applications, its expression can be represented by the following formula:

[0121] ,

[0122] in, This represents the fixed time overhead of the GPU in executing a task. and These represent the GPU's core frequency and memory frequency, respectively. (Coefficient) This indicates the application's sensitivity to changes in GPU memory frequency, while This indicates its sensitivity to changes in GPU core frequency. By adjusting... , and This model can simulate the impact of dynamic voltage and frequency regulation (DVFS) on various applications. It should be noted that... and They are mutually coupled. For a given Maximum core frequency Depend on The decision, and its relationship, can be represented by a function: ,in For function parameters, and These represent the GPU's core voltage and core frequency, respectively.

[0123] Based on the above analysis, vehicles GPU energy consumption within a specific time interval can be expressed as:

[0124] .

[0125] in, Indicates vehicle The GPU's power consumption function. Indicates vehicle GPU performance functions and Representing vehicles The GPU's power supply core voltage, core frequency, and memory frequency.

[0126] Similarly, for a specific drone In drones The power consumption of the GPU under dynamic workload can be expressed as:

[0127] .

[0128] in, and They represent drones The GPU's core voltage, core frequency, and memory frequency. (Coefficient) and The values ​​are fixed and depend on the hardware configuration and specific application. These coefficients respectively characterize the impact of memory frequency adjustment and processor core voltage / frequency adjustment on power consumption.

[0129] drones GPU performance function is

[0130] ,

[0131] in, , and They represent drones The GPU's core frequency and memory frequency. (Coefficient) This indicates the application's sensitivity to changes in GPU memory frequency, while This indicates its sensitivity to changes in GPU core frequency. By adjusting... , and This model can simulate the impact of dynamic voltage and frequency regulation (DVFS) on various applications. It should be noted that... and They are mutually coupled. For a given Maximum core frequency Depend on The decision, and its relationship, can be represented by a function: ,in For function parameters, and These represent the GPU's core voltage and core frequency, respectively.

[0132] Therefore, drones Total energy consumption can be expressed as

[0133] .

[0134] in, Indicates drone The GPU's power consumption function. Indicates drone GPU performance functions and They represent drones The GPU's power supply core voltage, core frequency, and memory frequency.

[0135] The energy consumption of all vehicles and drones can be expressed as:

[0136] ;

[0137] in, and Representing vehicles and drones Energy consumption.

[0138] The primary objective is to achieve a trade-off between inference latency, energy consumption (for vehicles and drones), and generated image quality. To this end, we introduce the following decision variable: the optimal split point between shared and local inference. Total number of noise reduction steps per vehicle Reasoning position allocation To eliminate dimensions and the problem of dimensionality, the utility function is transformed into the following form. Therefore, the utility function... Defined as

[0139] ,

[0140] in, It is the weighting coefficient. , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. , and These are the actual baseline latency, energy consumption, and generated quality values, respectively.

[0141] 2) This optimization problem is complex and non-convex. It is difficult to find the optimal solution in polynomial time. To address this complex problem, McCormick envelope, Taylor first-order expansion, and relaxation methods are used for reconstruction.

[0142] 3) The Alternating Direction Method of Multipliers (ADMM) distributed algorithm is used to solve the proposed joint optimization problem. Each vehicle node only solves its own subtask, which greatly saves solution resources and latency.

[0143] 3.1 Obtain the initial environment state and derive the initial actions;

[0144] 3.2 Calculate the inference latency, generated content quality, and corresponding energy consumption for each task;

[0145] 3.3 The global optimization variables are split into local variables of each vehicle node and global consistency variables. An augmented Lagrangian function is constructed, and the consistency constraints are relaxed by Lagrange multipliers and quadratic penalty terms.

[0146] 3.4 In each iteration, each vehicle node solves its corresponding local subproblem in parallel and updates its local decision variables; then, it updates the global consistency variables and Lagrange multipliers through a small amount of information exchange.

[0147] The optimization problem of the above steps is specifically manifested as follows: , ; st; , ; , ; , ; , ; , ; . ;

[0148] Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. This represents the reasoning steps for vehicle sharing. This represents the maximum number of reasoning steps. Indicates drone Maximum GPU power consumption. Constraints. Represents a binary variable used to indicate the location of inference. Constraints. This indicates that vehicles within the same group share intermediate results, therefore the number of shared denoising steps remains consistent within the group. Constraints The number of shared inference steps and the total number of inference steps are limited to avoid content generation failures. The maximum number of denoising steps. Constraints. and constraints This indicates that the shared denoising steps and the total denoising steps are integers. Constraints This indicates the GPU power capacity limit of the drone.

[0149] The optimization problem (1) is a nonlinear computational programming problem and is nonconvex, involving integer and binary variables. Furthermore, there are coupling relationships between the variables. Specifically, constraints (2)-(6), i.e., formulas (2)-(6), are nonlinear and discrete constraints; constraint (7), i.e., formula (7), contains a product term of two variables. Content quality item The problem is complex and non-convex. Therefore, it is difficult to find the optimal solution in polynomial time. Thus, the problem needs to be restructured.

[0150] Specifically, for two variables and The product of these variables introduces a new variable. Therefore, the newly added constraint can be expressed as:

[0151] ,

[0152] Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. Used for vehicle groups Shared denoising steps.

[0153] It is worth noting that, since vehicles within the same cluster can share intermediate inference results, therefore... Therefore, the utility function can be expressed as:

[0154]

[0155] in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total delay of inference. and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. The content quality function representing vehicle reasoning. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result. Similarly, another binary indicator variable... This is used to indicate whether the edge server has cached intermediate inference results for this vehicle group.

[0156] Subsequently, the McCormick relation method is used to relax constraint (8), and its relaxation form can be expressed as: , ; , ; , ; . ;

[0157] Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. This represents the reasoning steps for vehicle sharing. This represents the maximum number of reasoning steps. Furthermore, .

[0158] Constraints - With constraints Strict equivalence. For complex items related to content quality. The term is replaced by a first-order Taylor expansion, and its expression can be represented as:

[0159]

[0160] in, and Indicates the initial point. It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps. This represents the value of the function at the initial point. and This represents the distance from the initial point.

[0161] Therefore, the optimization problem P1 has been transformed into the following problem P2: ; st; - , - ;

[0162] in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total delay of inference. and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result. Similarly, another binary indicator variable... This indicates whether the edge server has cached intermediate inference results for this vehicle group. and Indicates at the initial point, It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps. This represents the value of the function at the initial point. and This represents the distance from the initial point.

[0163] Optimization problem P2 contains binary variables. and and integer variables and This remains a non-convex problem. To facilitate further solutions, these variables need to be relaxed. The relaxed form can be expressed as: ; st; - , , - ; . ;

[0164] in, Indicator variables This is used to indicate whether the shared inference process is executed on the drone.

[0165] Now, the optimization problem P2 is about the variables. , , and The problem is a linear programming problem with linear constraints. Therefore, the optimization problem P2 is convex.

[0166] 3) Using the Alternating Direction Method of Multipliers (ADMM) (see the instruction manual appendix) Figure 2 A distributed algorithm is used to solve the proposed joint optimization problem. Each vehicle node only solves its own subtask, which greatly saves solution resources and latency.

[0167] 3.1 Obtain the initial environment state and derive the initial actions;

[0168] 3.2 Calculate the inference latency, generated content quality, and corresponding energy consumption for each task;

[0169] 3.3 The global optimization variables are split into local variables of each vehicle node and global consistency variables. An augmented Lagrangian function is constructed, and the consistency constraints are relaxed by Lagrange multipliers and quadratic penalty terms.

[0170] 3.4 In each iteration, each vehicle node solves its corresponding local subproblem in parallel and updates its local decision variables; then, it updates the global consistency variables and Lagrange multipliers through a small amount of information exchange.

[0171] This paper presents a generative AI multi-layer collaborative reasoning method based on semantic awareness in a drone-assisted vehicle-to-everything (V2X) environment. This method is based on task semantic similarity classification and considers shared reasoning, reasoning offloading, and the overall reasoning steps, effectively improving the accuracy and efficiency of model training. For example... Figure 1 As shown, this method includes the following steps: ; st; ; ; ; ; ; (7);

[0172] in, . Indicates vehicle end Total inference latency, Indicates vehicle cluster Total delay of inference. and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result. Similarly, another binary indicator variable... This indicates whether the edge server has cached intermediate inference results for this vehicle group. and Indicates at the initial point, It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps. This represents the value of the function at the initial point. and Indicates the distance from the initial point, an indicator variable. This is used to indicate whether the shared inference process is executed on the drone. Indicates drone Maximum GPU power consumption.

[0173] Constraint (2) represents the binary variable used to indicate the inference location. Constraint (3) indicates that vehicles within the same group share intermediate results, thus the number of shared denoising steps remains consistent within the group. Constraint (4) limits the number of shared inference steps and the total number of inference steps to avoid content generation failure, where The maximum number of denoising steps. Constraints (5) and (6) indicate that the shared denoising steps and the total denoising steps are integers. Constraint (7) indicates the GPU power capacity limit of the UAV.

[0174] The research process of this invention is described in detail below:

[0175] The semantic-aware generative AI multi-layer collaborative reasoning method in the UAV-assisted vehicle-to-everything (V2X) environment must include the parameters and initial states of each node in the environment. The parameters of each node include the number of UAVs and vehicles, the available storage space and computing resources of each node, the task size, the uplink and downlink data transmission rates between vehicles and UAVs, and the uplink and downlink data transmission rates between UAVs and edge servers. These are key to the semantic-aware generative AI multi-layer collaborative reasoning method in the UAV-assisted V2X environment. This method constructs an optimization problem for semantic-aware generative AI multi-layer collaborative reasoning in the UAV-assisted V2X environment using communication models, inference latency models, content quality models, and energy consumption models, achieving an efficient balance between latency, energy consumption, and content quality. Then, the complex problem is reconstructed using McCormick envelope, Taylor first-order expansion, and relaxation methods. Finally, the ADMM distributed algorithm is used to solve the proposed joint optimization problem.

[0176] The generative artificial intelligence multi-layer collaborative reasoning method based on semantic perception proposed in this invention for UAV-assisted vehicle-to-everything (V2X) environments mainly consists of three parts: (1) calculating the wireless data transmission rate between vehicles, the uplink and downlink data transmission rates between vehicles and UAVs, and the uplink and downlink data transmission rates between UAVs and edge servers to establish a communication model. (2) calculating the inference latency of the task to establish an inference latency model. (3) establishing a content quality model through inference steps. (4) establishing an energy consumption model by calculating the energy consumption of vehicles and UAVs. (5) constructing a generative artificial intelligence multi-layer collaborative reasoning optimization problem based on semantic perception to achieve an efficient balance between latency, energy consumption, and content quality. Then, the complex problem is reconstructed using McCormick envelope, Taylor first-order expansion, and relaxation methods. The ADMM distributed algorithm is used to solve the proposed joint optimization problem.

[0177] The pseudocode description of a semantic-aware generative artificial intelligence multi-layer collaborative reasoning method in a drone-assisted vehicle-to-everything (V2X) environment is as follows: ; ;

[0178] As shown in the algorithm's pseudocode, the initial environment state is first obtained and the initial actions are determined. Then, the inference latency, generated content quality, and corresponding energy consumption are calculated for each task. Based on this, the global optimization variable is decomposed into local variables for each vehicle node and a global consistency variable. An augmented Lagrangian function is constructed, and the consistency constraint is relaxed using Lagrange multipliers and a quadratic penalty term. In each iteration, each vehicle node solves its local subproblems in parallel to update its local decision variables, and then updates the global consistency variable and Lagrange multipliers through a small amount of information exchange.

[0179] To address the challenges of resource constraints and semantic overlap in many tasks within UAV-assisted vehicle-to-everything (V2X) applications, which hinders meeting the diverse and low-latency service demands of vehicles, this paper proposes a semantic-aware joint reasoning framework for UAV-assisted V2X. A semantic-aware collaborative reasoning method is also proposed to jointly optimize the total inference steps, shared and local inference partitioning points, and inference locations. Furthermore, a quality quantification model is constructed, and a joint optimization problem for the total inference steps, shared and local inference partitioning points, and inference locations is established to achieve an optimal trade-off between latency, energy consumption, and inference quality. To address the non-convexity of this problem, McCormick envelope, first-order Taylor expansion, and variable relaxation techniques are employed to transform it into a solvable form. A joint optimization algorithm based on semantic awareness for shared inference, inference offloading, and the total inference steps is designed. The alternating direction multiplier method is used to obtain the corresponding collaborative reasoning optimization strategy, significantly reducing inference latency while ensuring user inference quality.

[0180] Example 1: To verify the practicality of this application, the applicant applied it in a specific way;

[0181] Using disaster assessment tasks in emergency rescue as a specific case, this paper discusses in detail the reasoning process at each stage. An example of reasoning for disaster assessment tasks in emergency rescue is provided. Figure 4 As shown.

[0182] This scenario employs a multi-layered structure, including base stations, drones, edge servers, cloud centers, and vehicle nodes, forming a two-tiered collaborative system covering both the air and the ground. Based on real-time disaster conditions and computational and energy resource availability, resources at each layer can be dynamically orchestrated to achieve coordinated emergency data collection, intelligent processing, and command and dispatch.

[0183] In collaborative reasoning computation, taking disaster assessment as an example, a combined model of centralized coordination and distributed execution is adopted. The drone first collects high-resolution images and video data of the disaster area, and performs preliminary target detection and damage identification through edge nodes, such as identifying damaged buildings, blocked roads, and stranded vehicles. Nodes within the cluster can share inference results, reducing redundant computation overhead. The drone can also cache frequently accessed high-precision map fragments and real-time traffic conditions, among other key data. When the requested content is cached, intermediate inference results can be directly distributed, thereby reducing service latency on the vehicle side. Vehicles, based on their own task requirements, complete local inference to generate joint disaster assessment results, providing rescue personnel with fast and accurate decision-making information, such as recommending optimal rescue routes or assessing the safety level of the area.

[0184] When roads are blocked or ground communications are paralyzed, drones can quickly establish aerial emergency communication links and collaborate with edge nodes to perform real-time analysis and reasoning of disaster images, enabling rapid assessment of the extent of damage and road accessibility in disaster areas. In multi-point concurrent rescue scenarios, task offloading and computing power allocation enable parallel processing and priority scheduling of multiple tasks, thereby providing timely and reliable decision support for rescue personnel, demonstrating high practical application value and promising prospects for widespread adoption.

[0185] The reasoning steps for disaster assessment tasks in emergency rescue using the technical solution of this application are as follows:

[0186] 1) Based on the communication model, inference latency model, content quality model and energy consumption model in the UAV-assisted vehicle network environment, a semantic-aware generative artificial intelligence multi-layer collaborative reasoning optimization problem is proposed to achieve an efficient balance between latency, energy consumption and content quality.

[0187] Achieving an efficient balance between latency, energy consumption, and content quality, the relevant convergence graph is shown below (see attached manual). Figure 5-6 ).

[0188] This demonstrates that the proposed algorithm exhibits good convergence performance and stability in multi-node collaborative reasoning scenarios, effectively improving the overall quality of the generation task, reducing latency and energy consumption, and achieving an effective balance between latency, energy consumption, and quality.

[0189] 2) This optimization problem is complex and non-convex; therefore, it is difficult to find the optimal solution in polynomial time. The McCormick envelope, Taylor first-order expansion, and relaxation method are used to reconstruct this complex problem.

[0190] 3) The Alternating Direction Method of Multipliers (ADMM) distributed algorithm is used to solve the proposed joint optimization problem. Each vehicle node only solves its own subtask, which greatly saves solution resources and latency.

[0191] The algorithm is based on four benchmarks: Non-Clustered Inference Algorithm (NCI), Fully UAV Inference Algorithm (FUI), Fully Edge Server Inference Algorithm (FEI), and Fully Local Inference Algorithm (FLI). In the Non-Clustered Inference Algorithm, no vehicle clustering is performed; all inference tasks (including those with significantly different semantic features) are processed directly without clustering. Shared inference can be executed on the UAV or edge server, while each vehicle independently performs local inference. In the Fully UAV Inference Algorithm, all shared inference processes are handled entirely by the UAV. The vehicle is only responsible for semantic extraction, transmitting and receiving data, and performing local inference; it does not participate in any shared inference process. The edge server is only responsible for sending back inference results when it has cached results. In the Fully Edge Server Inference Algorithm, all shared inference tasks are executed by the edge server, relying on the powerful computing resources of the edge server to ensure inference quality, while the vehicle is responsible for local inference. Furthermore, the UAV is only equipped with a caching module and does not have a computing module. When the drone has intermediate inference cache results, it can directly feed the cached results back to the vehicle, thus avoiding repeated inference. However, this approach introduces long-distance data transmission, leading to increased communication latency and energy consumption. In the all-local inference algorithm, all inference tasks are executed entirely locally by the vehicle; the drone and edge server do not participate in actual computation, but are only used as cache nodes. When the drone or edge server has cached inference results, it can directly transmit the results back to the vehicle. Each vehicle independently completes the entire inference process and obtains the results locally. Although this decentralized approach eliminates communication overhead, it often results in high latency and energy consumption due to the limited computing power of the vehicle. Compared to these comparative algorithms, the proposed algorithm allows each vehicle node to solve only its own subtask, significantly saving solution resources and latency. Experimental results are shown in the figure (see the appendix in the specification). Figure 7 ).

[0192] Figure 7The image illustrates how changes in vehicle computing power significantly impact system inference performance. As vehicle computing power increases, overall system computing resources are enhanced, improving task processing efficiency and affecting performance metrics such as system quality, utility, latency, and energy consumption to varying degrees. Overall, inference quality changes relatively steadily, while system utility gradually improves, and total latency and total energy consumption generally show a downward trend. This indicates that increased vehicle computing power effectively alleviates the load pressure caused by insufficient computing resources, thereby improving system operating efficiency.

[0193] It should be understood by those skilled in the art that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A generative artificial intelligence hierarchical reasoning method for unmanned aerial vehicle-assisted vehicle networks, characterized in that: In the context of drone-assisted vehicle-to-everything (V2X) environments, a generative AI-based multi-layer collaborative reasoning method based on semantic awareness includes the following steps: 1) Based on the communication model, inference latency model, content quality model and energy consumption model in the UAV-assisted vehicle network environment, a semantic-aware generative artificial intelligence multi-layer collaborative reasoning optimization problem is proposed to achieve an efficient balance between latency, energy consumption and content quality. 2) The optimization problem is complex and non-convex; it is difficult to find the optimal solution in polynomial time; McCormick envelope, Taylor first-order expansion and relaxation method are used to reconstruct the complex problem. 3) The proposed joint optimization problem is solved by using the alternating direction multiplier method distributed algorithm. Each vehicle node only solves its own subtask, which greatly saves solution resources and time latency. 3.1 Obtain the initial environment state and derive the initial actions; 3.2 Calculate the inference latency, generated content quality, and corresponding energy consumption for each task; 3.3 The global optimization variables are split into local variables of each vehicle node and global consistency variables. An augmented Lagrangian function is constructed, and the consistency constraints are relaxed through Lagrange multipliers and quadratic penalty terms. 3.4 In each iteration, each vehicle node solves its corresponding local subproblem in parallel and updates its local decision variables; then, it updates the global consistency variables and Lagrange multipliers through a small amount of information exchange.

2. The generative artificial intelligence hierarchical reasoning method in UAV-assisted vehicle networking as described in claim 1, characterized in that: The specific steps of step 1) are as follows: 1.1 Establish uplink and downlink communication models between vehicles, between vehicles and drones, and between drones and MEC servers: 1.2 Establishing an inference delay model: 1.3 For each vehicle reasoning task, establish a content quality model; In the vehicle During the denoising inference process, a series of intermediate points will be generated. Since the number of local denoising steps affects the quality of the generated content, local inference needs to be appropriately divided. Furthermore, considering the differences in needs between vehicles, the number of denoising steps required varies for different vehicles; therefore, the total number of steps must be tailored to different users. The following function describes the relationship between the inference steps and the quality of the generated content; this relationship can be expressed as: ； in, Indicates standard deviation, and Representing vehicles Shared noise reduction steps and vehicles Total number of denoising steps; The total content quality of all cars is: ； in, The content quality function representing vehicle reasoning; 1.4 Establish energy consumption models for drones and vehicles; Performing computationally intensive AIGC tasks leads to significant energy consumption, making energy consumption a critical factor affecting system performance; these tasks are distributed across a heterogeneous computing architecture between the UAV and the vehicle, and are processed using GPUs. Ignoring the GPU's static power consumption, and only considering its dynamic power consumption; furthermore, the GPU's operating power consumption can be expressed by the following formula: ； in, and These represent the GPU's core voltage, core frequency, and memory frequency, respectively; coefficients and The values ​​are fixed and depend on the hardware configuration and specific application; these coefficients respectively characterize the impact of memory frequency adjustment and processor core voltage / frequency adjustment on power consumption. Will Designed as a performance function to represent the performance of GPU-accelerated applications, its expression can be represented by the following formula: ； in, This represents the fixed time overhead of the GPU in executing a task. and These represent the GPU's core frequency and memory frequency, respectively; coefficients This indicates the application's sensitivity to changes in GPU memory frequency, while This indicates its sensitivity to changes in GPU core frequency; by adjusting... , and This model can simulate the impact of dynamic voltage and frequency regulation on different applications; and They are mutually coupled; for a given Maximum core frequency Depend on The decision, and its relationship, can be represented by a function: ,in For function parameters, and These represent the GPU's core voltage and core frequency, respectively. vehicle GPU energy consumption within a specific time interval can be expressed as: ； in, Indicates vehicle The GPU's power consumption function. Indicates vehicle GPU performance functions and Representing vehicles The GPU's power supply core voltage, core frequency, and memory frequency; Similarly, for a specific drone In drones The power consumption of a GPU under dynamic workload can be expressed as: ； in, and They represent drones The GPU's power supply core voltage, core frequency, and memory frequency; coefficients. and The values ​​are fixed and depend on the hardware configuration and specific application; these coefficients respectively characterize the impact of memory frequency adjustment and processor core voltage / frequency adjustment on power consumption. drones The GPU performance function is as follows: ； in, , and They represent drones GPU core frequency and memory frequency; coefficient This indicates the application's sensitivity to changes in GPU memory frequency, while This indicates its sensitivity to changes in GPU core frequency; by adjusting... , and This model can simulate the impact of dynamic voltage and frequency regulation on different applications; it should be noted that... and They are mutually coupled; for a given Maximum core frequency Depend on The decision, and its relationship, can be represented by a function: ,in For function parameters, and These represent the GPU's core voltage and core frequency, respectively. Therefore, drones Total energy consumption can be expressed as ； in, Indicates drone The GPU's power consumption function. Indicates drone GPU performance functions and They represent drones The GPU's power supply core voltage, core frequency, and memory frequency; The energy consumption of all vehicles and drones can be expressed as: ； in, and Representing vehicles and drones Energy consumption; The primary objective is to achieve a balance between inference latency, energy consumption, and generated image quality; to this end, the following decision variable is introduced: the optimal split point between shared and local inference. Total number of noise reduction steps per vehicle Reasoning position allocation To eliminate dimensions and the problem of dimensionality, the utility function is transformed into the following form; therefore, the utility function Defined as: ； in, These are weighting coefficients; , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. , and These are the actual baseline latency, energy consumption, and generated quality values, respectively. The goal is to maximize the utility function to achieve the optimal trade-off between inference latency, inference energy consumption, and content quality.

3. The generative artificial intelligence hierarchical reasoning method in UAV-assisted vehicle networking as described in claim 2, characterized in that: The specific steps of step 1.1) are as follows: In UAV-to-ground communication scenarios, considering both line-of-sight (LAS) and non-LAS links, a probabilistic path loss model is adopted; UAV With vehicles The channel power gain between them can be expressed as: ； in, This represents the probability of a line-of-sight link. and Representing vehicles and drones Channel gain between line-of-sight and non-line-of-sight links; channel gain between line-of-sight and non-line-of-sight links and Calculate according to the following formula, ； ； in, This indicates a small-scale attenuation in line-of-sight links. This indicates a small-scale attenuation in non-line-of-sight links; This represents the path loss of a line-of-sight link. This represents the path loss of a non-line-of-sight link. This represents the shadow fading component of a line-of-sight link. This represents the shadow fading component of a non-line-of-sight link; Channel gain and This study integrates small-scale fading, path loss, and shadowing fading components in both line-of-sight and non-line-of-sight links. The modeling methods for each component are as follows: The Nakagami-m fading distribution is used to model small-scale fading; this model was chosen because of its strong parameter scalability and high consistency with measured channel data. Specifically... and They respectively obey the fading parameters as and The Nakagami distribution has the following probability density function expression: ； ； in, This represents the average received power in the fading envelope. For gamma function, and This is an expression for small-scale fading in line-of-sight and non-line-of-sight links that follows a parametric Nakagami distribution; The path loss between the vehicle and the drone under line-of-sight (LAS) and non-LAS links can be expressed as follows: ， , in, Indicates the carrier frequency. It's the speed of light. For reference distance, This represents the actual distance between the vehicle and the drone. and These correspond to the path loss exponents for line-of-sight (LOS) and non-line-of-sight (NOS) links, respectively. Shadow fading characterizes the signal attenuation caused by obstacle blockage, and is modeled as a zero-mean Gaussian distributed random variable, with the specific expression as follows: ； ； in, and These represent the standard deviations of shadow fading for line-of-sight links and non-line-of-sight links, respectively. and For line-of-sight links and non-line-of-sight links, shadow fading occurs; The communication channel remains static within each time slot; therefore, according to Shannon's theorem, from the vehicle To drones The transmission rate can be expressed as: ； in, It is a vehicle Uplink bandwidth, It is a vehicle Transmission power, It is Gaussian white noise. For vehicles and drones The channel gain of the transmission channel; Similarly, from drones To the vehicle The downlink transmission rate can be expressed as: ； in, Indicates drone Transmission power, drone The bandwidth is expressed as , For vehicles and drones The channel gain of the transmission channel, It is Gaussian white noise; In addition, from drones The uplink data transfer rate to the MEC server is ； in, Indicates drone Transmission power, drone The bandwidth is expressed as , For drones and edge servers The channel gain of the transmission channel, It is Gaussian white noise; Accordingly, from MEC servers to drones The downlink data transmission rate is ； in, The transmission power of the MEC server is represented by [insert value here], and the bandwidth of the MEC server is represented by [insert value here]. , For drones and edge servers The channel gain of the transmission channel, It is Gaussian white noise.

4. The generative artificial intelligence hierarchical reasoning method in UAV-assisted vehicle networking as described in claim 2, characterized in that: The steps in step 1.3 are as follows: Considering that each vehicle is equipped with semantic extraction capabilities, which can convert raw data into text prompts, thereby significantly reducing the latency of raw data transmission; the definition of semantic extraction time is as follows: ； in, Indicates vehicle The computational cost required during semantic extraction; Indicates vehicle The computing resources available; The generated text prompts will then be sent to their associated drones for further analysis; from the vehicle To drones The uplink transmission delay is expressed as ； in, Represents a word element Size edge node Computing resources; Indicates by vehicle Generate text prompts The number of lexical units contained in it It is a vehicle and drones Uplink transmission rate; drones Task knowledge graphs are used to identify intents and relationships in task requests; tasks with similar intents are grouped together using a classification algorithm; therefore, vehicles exhibiting similar intents can be aggregated using this classification algorithm; the semantic analysis latency for analyzing semantic associations is represented as follows: ； in, Indicates drone Processing by vehicle Generated text prompts The amount of computation required; It is a drone Analyze the computational resources required for this prompt word; Furthermore, intermediate results are generated during the inference process, and these results can be pre-cached; since both the drone and the edge server are equipped with caching units, a binary indicator variable is introduced. To represent drones Is it already the first The vehicle group caches the corresponding inference content; if If yes, it means the drone has cached the corresponding shared inference results, and the intermediate results can be directly transmitted to the vehicle group within its coverage area; otherwise... Similarly, another binary indicator variable Used to indicate whether the edge server has cached intermediate inference results for this vehicle group; when If yes, it means the intermediate results can be obtained directly from the edge server; otherwise... ; when The latency for returning intermediate inference results is ； in, Indicates drone Cache vehicles The required size of intermediate inference results, It is a vehicle and drones Downlink transmission rate; Since vehicles within the same group share the same intermediate results, therefore... Indicates that the drone is the first The intermediate shared results of the group vehicle cache; from this, we can obtain This means that the intermediate results of vehicles within the same group are the same; when and The vehicle retrieves intermediate results from the edge server; at this point, the edge server sends the results to the vehicle. The downlink transmission delay for transmitting intermediate results can be expressed as: ； in, Indicates drone Cache vehicles The required size of intermediate inference results, It is a vehicle and drones downlink transmission rate For drones and edge servers Downlink transmission rate; When neither the drone nor the edge server caches intermediate results, i.e. and The vehicles will obtain the shared inference results from the most suitable entity; the selection of the service node depends on the node's computation latency, and each vehicle group will connect to the node that provides the lowest service latency. Based on the above analysis, an indicator variable is set. This is used to indicate whether the shared inference process is executed on the drone; if If yes, it means the drone is performing the shared inference process; otherwise, it means the shared inference process is performed on the edge server. ； Determining the optimal allocation between shared and local denoising steps for a vehicle group is crucial. Given a fixed total number of denoising steps, allocating too many steps to the shared inference stage, while improving content quality, is only a minor and limited improvement, effectively wasting significant computational resources. Conversely, insufficient shared inference steps necessitates compensation from local processes, leading to decreased output quality and unacceptable latency for users. Furthermore, since each vehicle requires a different number of denoising steps due to customized needs, tailoring an appropriate total number of denoising steps for each user is equally important. Therefore, the introduction Used for vehicle groups Shared denoising steps Indicates vehicle The total number of denoising steps; in addition, since the inference results are shared within the same vehicle group, therefore For ease of representation, drones The next The shared reasoning process of each group is denoted as: The local reasoning process is denoted as ; when , and At that time, the UAV will transmit the analyzed prompts to the edge server for further shared inference; therefore, the uplink transmission latency of uploading these results from the UAV to the ES for a certain vehicle group can be expressed as: ； in, Indicates in drones Vehicle groups within the service area The size of the analysis results It is a drone and edge servers Uplink transmission rate; The time required to perform shared inference on a drone can be expressed as: ； Among them, binary indicator variables To represent drones Is it already the first The vehicle group caches the corresponding inference content; another binary indicator variable. Used to indicate whether the edge server has cached intermediate inference results for this vehicle group; Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. Used for vehicle groups Shared denoising steps; when , and At that time, the edge server Elasticsearch performs the shared inference process, and the computation latency on the edge server is: ； Among them, binary indicator variables To represent drones Is it already the first The vehicle group caches the corresponding inference content; another binary indicator variable. Used to indicate whether the edge server has cached intermediate inference results for this vehicle group; This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. Used for vehicle groups Shared denoising steps; After the final shared inference is completed, the subsequent denoising steps are performed locally by each vehicle; the latency incurred by the vehicle performing local inference can be expressed as: ； in, This represents the average computational cost required for each local inference step. and Representing vehicle groups Shared noise reduction steps and vehicles Total number of denoising steps, Indicates vehicle The computing resources available; In summary, the vehicle group The inference delay can be derived as follows: ； in, A binary variable indicating whether the drone has cached the corresponding shared inference result; similarly, another binary indicator variable. Indicator variable used to indicate whether the edge server has cached intermediate inference results for this vehicle group. This is used to indicate whether the shared inference process is executed on the drone. and These represent the time required to perform shared inference on the drone and the time required to perform shared inference on the edge server, respectively. This indicates that semantically extracted content from vehicle groups is uploaded to the drone. Uplink transmission latency; For vehicles The total computation delay is: ； in, This indicates the latency of vehicle semantic extraction. Indicates from vehicle Upload to drone upload latency, Indicates the latency of semantic analysis. Indicates vehicle For vehicle clusters Local inference latency, a binary indicator variable Used to represent edge servers Has this vehicle group been identified? Caching intermediate inference results, indicator variables This is used to indicate whether the shared reasoning process occurs in the drone. Execute on, Indicates the time when the intermediate inference results are transmitted down. This indicates that intermediate inference results are downloaded from the edge server to the vehicle. Download latency; Therefore, the inference delay for all vehicles can be deduced as follows: ； in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total delay of inference.

5. The generative artificial intelligence hierarchical reasoning method in UAV-assisted vehicle networking according to claim 1, characterized in that: The optimization problem in step 1) is specifically manifested as follows: , ； st; , ； , ； , ； , ； , ； . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. Indicates drone energy consumption and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps. This represents the maximum number of reasoning steps. Indicates drone The highest GPU power consumption; Constraints Represents a binary variable used to indicate the location of reasoning; constraints. This indicates that vehicles within the same group share intermediate results, therefore the number of shared denoising steps remains consistent within the group; constraints The number of shared inference steps and the total number of inference steps are limited to avoid content generation failures. Maximum number of denoising steps; constraints and constraints Indicates that the shared denoising steps and the total denoising steps are integers; constraints This indicates the GPU power capacity limit of the drone; Optimization problem This is a nonlinear computational programming problem that is also nonconvex, involving integer and binary variables; furthermore, there are coupling relationships between the variables; specifically, the constraints... - The constraint is nonlinear and discrete; constraint conditions A product term containing two variables Content quality item The problem is complex and non-convex; therefore, it is difficult to find the optimal solution in polynomial time. The problem needs to be restructured. Specifically, for two variables and The product of the two variables introduces a new variable. Therefore, the newly added constraint is expressed as: . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. Used for vehicle groups Shared denoising steps; It is worth noting that, since vehicles within the same cluster can share intermediate inference results, therefore... Therefore, the utility function can be expressed as: ； in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total inference latency; and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. The content quality function representing vehicle reasoning. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result; similarly, another binary indicator variable. Used to indicate whether the edge server has cached intermediate inference results for this vehicle group; Subsequently, the McCormick relation method was used to apply the constraints. Relaxation can be expressed as follows: , ； , ； , ； . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. This represents the reasoning steps for vehicle sharing. This represents the maximum number of reasoning steps. Furthermore, ; Constraints - With constraints Strict equivalence; for complex items related to content quality The term is replaced by a first-order Taylor expansion, and its expression can be represented as: ； in, and This represents the value at the initial point. It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps. This represents the value of the function at the initial point. and This represents the distance from the initial point; Therefore, the optimization problem P1 has been transformed into the following problem P2: ； st; - , - ； in, , Indicates vehicle end Total inference latency, Indicates vehicle cluster Total inference latency; and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result; similarly, another binary indicator variable. This indicates whether the edge server has cached intermediate inference results for this vehicle group. and Indicates the initial point. It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps; This represents the value of the function at the initial point. and This represents the distance from the initial point; Optimization problem P2 contains binary variables. and and integer variables and It remains a non-convex problem; we relax these variables; the relaxed form can be expressed as: ； st; - , ， - ； . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone; Now, the optimization problem P2 is about the variables. , , and The linear programming problem P2 has linear constraints; therefore, the optimization problem P2 is convex.

6. A generative artificial intelligence hierarchical reasoning method for UAV-assisted vehicle networking as described in claim 5, characterized in that: Step 2 includes the following steps: For complex problems, McCormick envelope, Taylor first-order expansion, and relaxation methods are used for reconstruction. Optimization problem That is, formula (1) is a nonlinear computational programming problem and is nonconvex, involving integer variables and binary variables; in addition, there are coupling relationships between the variables; specifically, the constraints... - That is, formula - It is a nonlinear and discrete constraint; constraint condition (7), i.e., formula (7), contains a product term of two variables. Content quality item The problem is complex and non-convex; therefore, it is difficult to find the optimal solution in polynomial time. Refactor the problem: Optimization problem That is, formula (1) is a nonlinear computational programming problem and is nonconvex, involving integer variables and binary variables; in addition, there are coupling relationships between the variables; specifically, the constraints... - That is, formula - It is a nonlinear and discrete constraint; constraint condition (7), i.e., formula (7), contains a product term of two variables. Content quality item The problem is complex and non-convex; therefore, it is difficult to find the optimal solution in polynomial time. The problem needs to be restructured. Specifically, for two variables and The product of the two variables introduces a new variable. Therefore, the newly added constraint can be expressed as: . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. Used for vehicle groups Shared denoising steps; It is worth noting that, since vehicles within the same cluster can share intermediate inference results, therefore... Therefore, the utility function can be expressed as: ； in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total inference latency; and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. The content quality function representing vehicle reasoning. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result; similarly, another binary indicator variable. Used to indicate whether the edge server has cached intermediate inference results for this vehicle group; Subsequently, the McCormick relation method is used to relax constraint (8), and its relaxation form can be expressed as: , ； , ； , ； . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone. This represents the reasoning steps for vehicle sharing. This represents the maximum number of reasoning steps. Furthermore, ; Constraints - With constraints Strict equivalence; for complex items related to content quality The term is replaced by a first-order Taylor expansion, and its expression can be represented as: ； in, and Indicates the initial point. It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps. This represents the value of the function at the initial point. and This represents the distance from the initial point; Therefore, the optimization problem P1 has been transformed into the following problem P2: ； st; - ， - ； in, Indicates vehicle end Total inference latency, Indicates vehicle cluster Total inference latency; and Representing vehicles and drones energy consumption , and These are the baseline latency, energy consumption, and generated quality values ​​measured under fully local inference, respectively. Indicates in drones Computing resources This represents the computational cost required for each step of reasoning on the drone. This represents the average computational cost required for each local inference step. Indicates vehicle The computing resources available This indicates the computing resources used by Elasticsearch to perform shared inference. This represents the computational cost required for each step of inference in Elasticsearch. A binary variable indicating whether the drone has cached the corresponding shared inference result; similarly, another binary indicator variable. This indicates whether the edge server has cached intermediate inference results for this vehicle group. and Indicates the initial point. It is Gaussian noise. and Indicates vehicle The number of shared reasoning steps and the total number of reasoning steps; Optimization problem P2 contains binary variables. and and integer variables and It remains a non-convex problem; we relax these variables; the relaxed form can be expressed as: ； st; - , , - ； . ； Among them, indicator variables This is used to indicate whether the shared inference process is executed on the drone; Now, the optimization problem P2 is about the variables. , , and The linear programming problem P2 has linear constraints; therefore, the optimization problem P2 is convex.

Images