An air-ground dynamic cooperative reasoning method and device for low-altitude intelligent networking

Abstract

The application discloses an air-ground dynamic cooperative reasoning method and device for low-altitude intelligent networking, comprising the following steps: deploying a trained small and lightweight deep neural network model and a large and complex deep neural network model on a single unmanned aerial vehicle and a ground station respectively; taking the unmanned aerial vehicle as an air node and the ground station as a ground node, after receiving a reasoning task request, the unmanned aerial vehicle performs statistical evaluation on its own state and available resources, and perceives the channel condition of air-ground transmission; selecting whether to perform reasoning only by the unmanned aerial vehicle using the small and lightweight deep neural network model, if not, dividing the reasoning task into two subtasks according to the layer structure and the number of layers of the large and complex deep neural network model, and distributing the subtasks to the unmanned aerial vehicle and the ground station for joint execution. The application can provide a flexible and efficient adaptive reasoning strategy for air-ground cooperative computing, and greatly improves the computing efficiency of low-altitude intelligent networking.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned aerial vehicle air-ground collaborative reasoning, in particular to an air-ground dynamic collaborative reasoning method and device for low-altitude intelligent networking. BACKGROUND

[0002] The development needs of low-altitude economy give rise to low-altitude intelligent networking, which is an important infrastructure for realizing the "man-machine-object" three-in-one fusion intelligent interconnection in low-altitude airspace (3km / 1km), and can help to realize the networked, digitized and intelligent operation of low-altitude business entity network space. As a key low-altitude new infrastructure, low-altitude intelligent networking is an essential basic information network infrastructure to support the development of low-altitude economy.

[0003] Unmanned aerial vehicles are usually used as the main application carrier of low-altitude intelligent networking due to their flexibility, real-time data collection and access to difficult-to-reach areas. Unmanned aerial vehicles equipped with cameras, infrared sensors and GPS (Global Positioning System) sensors are an ideal choice for performing low-altitude tasks in challenging environments. In order to handle some complex tasks in low-altitude, it is essential to use accurate and efficient deep neural network (DNN) models to process on-board sensory data in real time. However, since DNN tasks are usually computationally intensive, it is challenging to perform complex DNN reasoning on unmanned aerial vehicles with limited on-board resources. Although it is feasible to transmit a large amount of on-board sensory data directly to the ground station with abundant resources, considering the limited and past air-ground communication resources, this may result in significant transmission delay and high bandwidth consumption.

[0004] In order to solve the inevitable conflict between the high resource requirements of DNN and the limited on-board resources, and improve the computing efficiency in low-altitude intelligent networking, the existing works mainly exist the following two methods: one method is to compress large and complex DNNs into small and lightweight DNNs through pruning, quantization and knowledge distillation techniques; in this way, the compressed model can be directly executed on a single unmanned aerial vehicle, thereby reducing the execution delay at the cost of significant accuracy loss. The other method is to use unmanned aerial vehicles and ground stations to collaboratively execute complex DNN models to ensure high accuracy, while the air-ground communication involved may still introduce relatively large reasoning delay, especially considering the dynamic air-ground communication due to the movement of unmanned aerial vehicles. In summary, the above two methods cannot simultaneously maintain high accuracy and low delay.

[0005] The task reasoning method based on the collaborative action of complex neural network models and lightweight neural network models of unmanned aerial vehicle groups is disclosed in the inventions with publication numbers CN116451576A, CN117669741A, CN115934310A, etc. However, these studies focus on the unmanned aerial vehicle groups, and rarely involve air-ground collaborative computing.

[0006] The invention disclosed in the invention with publication number CN118865148A discloses a satellite internet-oriented target identification method, including a satellite internet end-edge-cloud collaborative target identification system architecture, a UAV cluster cooperative identification method, and a satellite-assisted edge cooperative identification method. Among them, the end-edge-cloud collaborative target identification architecture is divided into three parts: the terminal layer, the satellite edge layer and the ground cloud layer. The UAV cluster cooperative identification method improves the cooperative reasoning performance in terms of delay and energy consumption. The satellite-assisted edge cooperative identification method selects a satellite with the largest elevation angle to the terminal for access according to the maximum elevation angle criterion. The invention can meet the demand for video analysis in unmanned areas, and effectively utilize the communication and computing resources of UAVs and satellites while maintaining detection accuracy. However, the focus of the invention is to utilize the resources of satellites, UAVs and ground stations for collaborative target identification, emphasizing communication access to the satellite with the largest elevation angle to the terminal to improve the cooperative efficiency of target identification, and not involving the efficient use of computing resources in air-ground collaborative reasoning. SUMMARY

[0007] The purpose of the present application is to disclose an air-ground dynamic cooperative reasoning method and device for low-altitude intelligent networking, which can provide flexible and efficient adaptive reasoning strategies for air-ground collaborative computing, greatly improving the computing efficiency of low-altitude intelligent networking. The UAV equipped with the dynamic cooperative reasoning method can not only realize fast and efficient air-ground collaborative computing in low-altitude intelligent networking, but also is expected to provide more computing solutions for various industry applications of low-altitude economy.

[0008] To achieve the above technical purpose, the technical scheme adopted by the present application is:

[0009] In a first aspect, an air-ground dynamic cooperative reasoning method for low-altitude intelligent networking is provided, the method comprising the following steps:

[0010] A small and lightweight deep neural network model and a large and complex deep neural network model are respectively deployed on a single UAV and a ground station.

[0011] The UAV is used as an air node and the ground station is used as a ground node. After receiving a reasoning task request, the UAV evaluates its own state and available computing resources, selects a time to execute the next task through the arrival of the task and the time to complete the current task, and calculates the channel transmission rate by considering the distance from the UAV to the ground station, thereby completing the perception of the channel condition for air-ground transmission.

[0012] In combination with the statistical evaluation result and the perception result, the inference task is searched for an optimized decision by a deep reinforcement learning algorithm according to the model layer and the dynamic task flow, and the total delay of the entire task flow under different decision sets is evaluated under the channel condition at the current position. The DQN network of the deep reinforcement learning is trained by the reward value corresponding to the total delay under different value conditions, and finally converges to adjust the node type participating and the model calculation amount allocated according to the decision set of all tasks. If the local inference of the UAV is selected, a small and lightweight deep neural network model previously deployed by the UAV is used for inference alone. If the ground station is selected for collaborative processing, the inference task is divided into two subtasks according to the decision according to the layer, and the subtasks are allocated to the UAV and the ground station for joint execution.

[0013] Further, when the small and lightweight deep neural network model is selected for task inference alone, the sum of the calculation delay and the transmission delay of the UAV executing the inference task is calculated according to the following formula:

[0014]

[0015] wherein t small is the sum of the inference task delay from the first layer to the L smax th layer received by the UAV executing n tasks alone, and the L smax th layer is the maximum layer of the small and lightweight model; C sj is the calculation amount required by the UAV executing the j a th layer inference task; P

[0016] Further, when the UAV and the ground station are selected for joint execution of the inference task, the target layer of the deep learning model is the smallest subtask execution unit, and the sum of the calculation delay and the transmission delay of the UAV executing the target layer inference task is calculated according to the following formula:

[0017]

[0018] wherein t large is the sum of the inference task delay of the UAV executing the β k th layer and the output data delay of the ground station receiving the β k +1 layer under the condition of air-ground collaborative execution of the large and complex model inference, the β k th layer is the target layer; C lj is the calculation amount required by the UAV executing the j a th layer inference task; P lj is the calculation capability of the UAV; O k is the data size of the output of the β g th layer inference task; and ρThe signal transmission rate of the unmanned aerial vehicle; and respectively represent the computing delay of the unmanned aerial vehicle performing the inference task and the transmission delay of the ground station receiving the input data; in the formula of calculating p g , B g represents the transmission bandwidth of the air-ground channel; D is the distance between the unmanned aerial vehicle and the ground station, a' is the path loss index, g0 is the channel gain, and N0 is the noise power.

[0019] Further, the process of adjusting the participating node type and the allocated model computing amount according to the obtained decision set of all tasks includes the following steps:

[0020] Two decision variables, α and β, are designed; α is used to decide whether to perform the task inference alone, and β is used to decide the layer of the model at the split point; when a small lightweight deep neural network model is selected to perform the task inference alone, α = 0 and β = 0; when the unmanned aerial vehicle and the ground station jointly perform the inference task, α = 1 and β ≠ 0, and β represents the layer of the model at the split point; the result of the inference decision of the system under the dynamic environment is represented by the value set (α, β) of the two decision variables;

[0021] Real-time perception of the state of the unmanned aerial vehicle, for different task arrival times in the task flow, by substituting the position of the unmanned aerial vehicle, evaluating the arrival and completion of the previous task, and considering the memory and precision constraints, the task delay of different decision sets (α, β) under the channel conditions at the current position is evaluated; the DQN network of the deep reinforcement learning is trained by the reward value corresponding to the total delay under different value conditions, and finally it converges; the participating node type α and the allocated model computing amount β are adjusted according to the obtained decision set of all tasks, and the low-delay collaborative inference is performed under the precision constraint.

[0022] Further, the method further includes:

[0023] For the arrival of n tasks, the total system delay of the inference is represented as t = {t1, t2,..., t k ..., t n}; for the kth task, the task completion time T k , the task arrival time F k , and the computing delay C k are represented by the following relationship:

[0024]

[0025] With the optimization objective of minimizing the total delay of multiple task executions, the objective function of the optimization problem is defined as:

[0026]

[0027] wherein, alpha k and beta k represent two decision variables for the kth task, total task delay t k is calculated by the computation delay, task arrival time and computation delay;

[0028] The constraint condition is constructed as follows:

[0029]

[0030] wherein, A k is the precision of the kth task, A set is the lowest threshold of the precision of all tasks; m j is the memory requirement of the jth layer reasoning task, m a is the memory capacity available to the system.

[0031] In a second aspect, the application discloses an air-ground dynamic collaborative reasoning device for low-altitude intelligent networking, which comprises:

[0032] a model deployment module, configured to deploy a small and lightweight deep neural network model and a large and complex deep neural network model trained on a single unmanned aerial vehicle and a ground station respectively;

[0033] a strategy evaluation module, configured to detect and collect low-altitude reasoning tasks, and evaluate the system state, available resources and channel conditions of air-ground transmission after receiving a task request;

[0034] a task decision module, configured to select a model reasoning mode according to the evaluation result, wherein the model reasoning mode comprises a mode in which the unmanned aerial vehicle performs reasoning by using the small and lightweight deep neural network model alone, and a mode in which the task is assigned to the unmanned aerial vehicle and the ground station for joint execution;

[0035] a task execution module, configured to execute a task reasoning process according to the selected model reasoning mode.

[0036] Compared with the prior art, the application has the following beneficial effects:

[0037] Firstly, the air-ground dynamic collaborative reasoning method and device for low-altitude intelligent networking of the application are different from the traditional data backhaul reasoning mode and the single-machine local reasoning mode, the method proposed by the application improves the task execution efficiency by deploying deep learning models in the air-ground system and performing air-ground collaborative reasoning.

[0038] Second, the air-ground dynamic collaborative reasoning method and device for low-altitude intelligent networks of the present invention, based on real-time perception of UAV and task processing status, adjusts the participating node types and allocated model layer computation through deep reinforcement learning algorithms, performs optimization decisions under accuracy constraints, and finally achieves low-latency collaborative reasoning, enabling the system to complete reasoning tasks more efficiently. Attached Figure Description

[0039] Figure 1 A flowchart illustrating an air-ground dynamic collaborative reasoning method for low-altitude intelligent networks, provided as an embodiment of the present invention;

[0040] Figure 2 This invention provides a scenario diagram of air-ground dynamic collaborative reasoning for low-altitude intelligent networks.

[0041] Figure 3 The graph shows the performance comparison results of the algorithm using the AlexNet and VGG16 models on devices with different computing capabilities, as provided in the embodiments of the present invention.

[0042] Figure 4 The graph shows the performance comparison results of the algorithm using the AlexNet and ResNet152 models on devices with different computing capabilities, as provided in the embodiments of the present invention.

[0043] Figure 5 A comparison chart of algorithm performance under different numbers of tasks when using AlexNet and VGG16 models, provided for embodiments of the present invention;

[0044] Figure 6 A comparison chart of algorithm performance with different numbers of tasks when using AlexNet and ResNet152 models, provided in an embodiment of the present invention;

[0045] Figure 7 A comparison chart of algorithm performance under different precision constraints when using AlexNet and VGG16 models, provided for embodiments of the present invention;

[0046] Figure 8 A comparison chart of algorithm performance under different precision constraints when using AlexNet and ResNet152 models, provided for embodiments of the present invention;

[0047] Figure 9 This is a structural diagram of an air-ground dynamic collaborative reasoning system for low-altitude intelligent networks provided in an embodiment of the present invention. Detailed Implementation

[0048] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0049] In an embodiment, as shown in Figure 1 The embodiment of the present application provides an air-ground dynamic collaborative inference method for low-altitude intelligent networking, which comprises the following steps:

[0050] Step 101, a large and a small trained deep learning model are respectively deployed on the unmanned aerial vehicle and the ground station in the air-ground system.

[0051] As shown in Figure 2 The embodiment of the present application is composed of an unmanned aerial vehicle and a ground station. The unmanned aerial vehicle processes on-board computing, while the ground station supports more complex tasks, both of which use DNN models to perform inference. The system supports two types of models: a small, lightweight model for small, lightweight tasks requiring lower accuracy, and a large, complex model for higher accuracy but with increased latency. We pre-deploy a small model and a large model on the unmanned aerial vehicle. The small model runs independently on the unmanned aerial vehicle, while the large model is partitioned and cooperatively executed by the unmanned aerial vehicle and the ground station according to task requirements.

[0052] Step 102, after receiving the inference task request, the unmanned aerial vehicle evaluates the system state, available resources and channel conditions of air-ground transmission.

[0053] Specifically, taking the unmanned aerial vehicle as an air node and the ground station as a ground node, after receiving the inference task request, the unmanned aerial vehicle evaluates its own state and available computing resources, selects the time to execute the next task through the arrival of the task and the time to complete the current task, and calculates the channel transmission rate by substituting the distance from the unmanned aerial vehicle to the ground station, thereby completing the perception of the channel conditions of air-ground transmission.

[0054] Step 103, according to the evaluation result, it is selected whether to use the small lightweight deep neural network model alone for inference by the unmanned aerial vehicle or to distribute the task to the unmanned aerial vehicle and the ground station for cooperative execution. In step 103, the deep learning model is a deep neural network, which includes an input layer, a hidden layer and an output layer; the layer in this embodiment refers to the layer structure in the deep neural network.

[0055] Combining statistical evaluation results and perception results, the inference task is optimized by searching using a deep reinforcement learning algorithm based on the number of model layers and the dynamic task flow. The total task latency of the entire task flow under different decision sets is evaluated under the current channel conditions. The reward values ​​corresponding to the total latency under different values ​​are used to train the deep reinforcement learning DQN network until it converges. The participating node types and the allocated model computation are adjusted according to the decision sets of all tasks obtained in this way. If local inference is selected by the UAV, model partitioning is not required. If collaborative processing with the ground station is required, the inference task is divided into two sub-tasks according to the number of layers based on the decision, and the sub-tasks are assigned to the UAV and the ground station for joint execution.

[0056] For ease of explanation, this invention designs two decision variables: α and β. When a small, lightweight deep neural network model is chosen for task inference alone, α = 0. In this case, inference using a large, complex model is not involved, so β ​​also takes the value 0. When a UAV and a ground station are chosen to jointly perform the inference task, α = 1, and β ≠ 0. β represents the layer of the model at the split point. The set of values ​​(α, β) for these two decision variables represents the system's inference decision-making outcome under dynamic conditions.

[0057] Real-time perception of UAV status is achieved. For different task arrival times in the task flow, the arrival and completion status of the previous task are evaluated by substituting the UAV's position, considering the limitations of memory and accuracy, etc., to assess the task latency of different decision sets (α, β) under the current location and channel conditions. The reward value corresponding to the total latency under different values ​​is used to train a deep reinforcement learning DQN network, which eventually converges. The participating node types (α) and the allocated model computation (β) are adjusted according to the decision sets of all tasks obtained in this way, and low-latency collaborative inference is performed under accuracy constraints.

[0058] Specifically, for the arrival of n tasks, the total system delay for inference is represented as t = {t1, t2, ..., tn}. k ..., t n For the k-th task, the task completion time T k Mission arrival time F k and computational delay C k This can be represented by the following relationship:

[0059]

[0060] The objective function of the optimization problem is defined as follows:

[0061]

[0062] This formula represents the objective function in the optimization problem, aiming to minimize the total delay of multiple task executions. The total delay t k of tasks is calculated by synthesizing the computation delay, task arrival time, and computation delay. The goal of this formula is to minimize this delay among all tasks while making efficient decisions according to the requirements of the tasks. The constraints are constructed as follows:

[0063] The first constraint:

[0064] The second constraint:

[0065] The first constraint ensures that the overall accuracy of the system is maintained above a certain threshold. Specifically, the formula sums up the accuracy A k of all tasks and takes the average, and forces this sum to meet the minimum acceptable value A set , thereby ensuring that the accuracy of the system does not drop excessively due to delay optimization.

[0066] The second constraint ensures that the system has enough memory to handle the computational requirements of all possible model inference scenarios. The sum of the memory requirements m j of all inference tasks cannot exceed the available memory capacity m a of the system.

[0067] First case: when a small lightweight model is selected for individual task inference, and the inference task is executed by the UAV alone, the calculation of the sum of the computation delay and transmission delay of the UAV executing the inference task is calculated according to the following formula:

[0068]

[0069] where t small is the sum of the inference task delay from the first layer to the L smax th layer received by the UAV executing n tasks (k is the number of tasks) in the case of the UAV executing the inference task alone, and the L smax th layer is the maximum number of layers of the small lightweight model; C sj is the computation required for the UAV to execute the jth layer of inference task; and p a is the computing power of the UAV (number of multiply-add operations per second).

[0070] Second case: when the accuracy of the UAV alone participating in model inference cannot meet the requirements, the inference task needs to be divided into two sub-tasks according to the layer structure and number of layers of the deep learning model, and the sub-tasks are assigned to the UAV and the ground station for joint execution, including:

[0071] When the target layer of the deep learning model is the smallest subtask execution unit, the sum of the computation latency and transmission latency of the UAV performing the target layer inference task is calculated according to the following formula:

[0072]

[0073] Among them, t large In the context of air-ground collaborative execution of large and complex model inference, the UAV executes the βth... k Layer inference task latency and ground station reception of the βth layer k The sum of the output data delays of +1 layer, the βth layer k Layer C is the target layer. lj The computational cost required for the UAV to perform the j-th level inference task; P a The computing power of the drone (number of addition and multiplication operations per second); O lj To execute the βth k The size of the data output by the layered inference task; ρ g The signal transmission rate of the drone; and These represent the computational latency of the UAV performing the inference task and the transmission latency of the ground station receiving the input data, respectively. In calculating ρ... g In the formula, B g The vector is expressed as the transmission bandwidth of the air-to-ground channel; D is the distance between the UAV and the ground station; α is the path loss exponent; g0 is the channel gain; and N0 is the noise power. In this embodiment, only one of the first and second scenarios will occur simultaneously.

[0074] Figure 3 , Figure 4 This is a graph showing the performance comparison results of different computing devices and different algorithms in the embodiments of the present invention. The horizontal axis represents the device type, and the vertical axis represents the total inference latency. Figure 3 , Figure 4 Three benchmark algorithms were used to perform inference tasks on air-to-ground systems involving UAVs with three different computing capabilities. First, AlexNet and VGG-16 were used collaboratively for inference tasks, followed by AlexNet and ResNet152. Figure 3 It can be seen that, although the three neural network models have different complexities and computational costs, the embodiments of the present invention consistently achieve the lowest latency compared to the four benchmark algorithms. In particular, when using a Raspberry Pi 3B+, the latency is reduced by 34.19% and 29.03% compared to other algorithms, highlighting the benefits of higher computational power. Figure 4 The total processing time of tasks on the three devices and Figure 2Similar, while the large model VGG16 always requires more total time to process the same number of tasks as ResNet152, this is primarily due to its larger number of layers and more complex architecture.

[0075] Figure 5 、 Figure 6 The comparison chart of algorithm performance under different number of tasks provided by the embodiment of the application. In order to ensure the consistency of experimental conditions, the requirements for equipment and precision are consistent to ensure that all algorithms are uniformly evaluated. Figure 5 The results of the scenario using the highly complex model VGG16 are shown in the following table. Figure 6 The results of the scenario using the complex model ResNet152 are shown in the following table. In both cases, as the number of tasks increases, the total task latency also increases; however, the latency of the embodiment of the application remains stable and is the lowest among all algorithms. Specifically, in the following table, compared with the three benchmark algorithms, the total latency of the embodiment of the application is reduced by 42.21%, 57.46%, and 63.65%, respectively. Figure 5 In the following table, the reduction rates are 19.3%, 39.63%, and 30.3%, respectively. Figure 6

[0076] Figure 7 、 Figure 8 The comparison chart of algorithm performance under different precision constraints provided by the embodiment of the application. The results show that higher precision requirements will lead to an increase in task latency due to the need for more complex models. Due to the small number of tasks, the latency optimization effect of GC is slightly obvious. Overall, the embodiment of the application consistently identifies favorable decisions at different precision levels while maintaining stable processing efficiency.

[0077] In step 104, the task inference process is performed according to the selected model inference mode.

[0078] The UAV air-ground collaborative inference method based on adaptive switching of lightweight and complex models proposed in the embodiment of the application selects the model inference mode by evaluating the resources and communication states of the UAV and the ground station, selects lightweight model inference alone or divides the high-precision complex neural network model and collaborates with the ground station for inference, reduces the problems of high processing delay of complex models and insufficient precision of lightweight models, and improves the efficiency of the system in executing inference tasks. The method provided by the embodiment of the application also uses a deep reinforcement learning algorithm, which can quickly collect optimal solutions, so that the air-ground system can dynamically select appropriate strategies to perform computing tasks according to its own state and application requirements, thereby realizing high energy efficiency and low latency collaborative inference.

[0079] ​Based on the same inventive concept, the embodiment of the present application also provides an air-ground dynamic collaborative reasoning method system for low-altitude intelligent networking. Since the system solves problems by the same principle as the air-ground dynamic collaborative reasoning method for low-altitude intelligent networking, the implementation of the system can refer to the implementation of the UAV air-ground collaborative reasoning method based on the adaptive switching of lightweight and complex models, and the repeated parts will not be described here.

[0080] In another embodiment, the air-ground dynamic collaborative reasoning method system for low-altitude intelligent networking provided by the embodiment of the present application comprises: Figure 9

[0081] a model deployment module 10 configured to deploy the same two specified deep learning models trained on the UAV and the ground station in the air-ground system;

[0082] a strategy evaluation module 20 configured to detect and collect low-altitude reasoning tasks by the UAV as a task source, and evaluate the system state, available resources and channel conditions of air-ground transmission after receiving a task request;

[0083] a task decision module 30 configured to select whether to perform reasoning by the UAV using a small lightweight deep neural network model alone or to assign the task to the UAV and the ground station for joint execution according to the evaluation result;

[0084] a task execution module 40 configured to perform a task reasoning process according to the selected model reasoning mode.

[0085] The task division module comprises:

[0086] In the first case, when a small lightweight model is selected to perform a task reasoning alone, the UAV performs the reasoning task alone, and the sum of the calculation time delay and the transmission time delay of the UAV performing the reasoning task is calculated according to the following formula:

[0087]

[0088] wherein t small is the sum of the reasoning task time delay from the first layer to the L smax th layer received by the UAV in the case of performing the reasoning task alone by the UAV, the L smax th layer is the maximum number of layers of the small lightweight model; C sj is the calculation amount required by the UAV to perform the jth layer reasoning task; and P a is the calculation capability (the number of multiplication and addition operations per second) of the UAV.

[0089] ​The second scenario involves situations where the accuracy of model inference using a drone alone is insufficient. In this case, the inference task needs to be divided into two sub-tasks based on the layer structure and number of layers of the deep learning model. These sub-tasks are then assigned to both the drone and the ground station for joint execution, including:

[0090] When the target layer of the deep learning model is the smallest subtask execution unit, the sum of the computation latency and transmission latency of the UAV performing the target layer inference task is calculated according to the following formula:

[0091]

[0092] Among them, t large In the context of air-ground collaborative execution of large and complex model inference, the UAV executes the βth... k Layer inference task latency and ground station reception of the βth layer k The sum of the output data delays of +1 layer, the βth layer k Layer C is the target layer. lj The computational cost required for the UAV to perform the j-th level inference task; P a The computing power of the drone (number of addition and multiplication operations per second); O lj To execute the βth k The size of the data output by the layered inference task; ρ g The signal transmission rate of the drone; and These represent the computational latency of the UAV performing the inference task and the transmission latency of the ground station receiving the input data, respectively. In this embodiment, only one of the first and second scenarios will occur; that is, the first and second scenarios will not occur simultaneously.

[0093] For more detailed information on the working process of each of the above modules, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0094] In another embodiment, the present invention provides a computer device including a processor and a memory; wherein the processor executes a computer program stored in the memory to implement the steps of the above-described model segmentation-based UAV swarm collaborative reasoning method.

[0095] For more detailed information on the above methods, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.

[0096] In another embodiment, the present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the above-described air-ground dynamic collaborative reasoning method for low-altitude intelligent networks.

[0097] Those skilled in the art will appreciate that embodiments of the application can be readily used as software, hardware, or a combination of software and hardware. In one embodiment, the application can be implemented in software and / or firmware. In addition, the software implementation can be implemented by one or more computer programs.

[0098] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing each of the flowchart blocks or the functions noted in the blocks

[0099] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing each of the flowchart blocks or the functions noted in the blocks

[0100] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing each of the flowchart blocks or the functions noted in the blocks

[0101] While the preferred embodiments of the application have been described, additional variations and modifications can be employed. Therefore, the terms and expressions

[0102] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A dynamic collaborative reasoning method for air-to-ground communication for low-altitude intelligent networks, characterized in that, The method includes the following steps: A small, lightweight deep neural network model and a large, complex deep neural network model are deployed on a single drone and a ground station, respectively. Using the UAV as an air node and the ground station as a ground node, after receiving the inference task request, the UAV assesses its own status and available computing resources. Based on the arrival and completion times of the current task, it selects the time to execute the next task. At the same time, it calculates the channel transmission rate by taking into account the distance between the UAV and the ground station, thus completing the perception of the channel conditions for air-to-ground transmission. Combining statistical evaluation results and perception results, the inference task is optimized by searching using a deep reinforcement learning algorithm based on the number of model layers and the dynamic task flow. The total task latency of the entire task flow under different decision sets is evaluated under the current channel conditions. The reward values ​​corresponding to the total latency under different values ​​are used to train the deep reinforcement learning DQN network until it converges. The participating node types and the allocated model computation are adjusted according to the decision sets of all tasks. If local inference is selected from the UAV, the small, lightweight deep neural network model pre-deployed on the UAV performs inference independently. If collaborative processing with the ground station is selected, the inference task is divided into two sub-tasks according to the number of layers, and the sub-tasks are assigned to the UAV and the ground station for joint execution.

2. The air-ground dynamic collaborative reasoning method for low-altitude intelligent networks according to claim 1, characterized in that, When a small, lightweight deep neural network model is selected for task inference alone, the sum of the computation latency and transmission latency of the UAV performing the inference task is calculated according to the following formula: Among them, t small In the case where a drone performs inference tasks independently, and the drone performs n tasks, it receives data from layer 1 to layer L. smax The sum of the latency of the Lth layer reasoning task smax The number of layers is the maximum number of layers in a small, lightweight model; The computational cost required for the UAV to perform the j-th level inference task; P a Let k represent the computing power of the drone, where k = 1, 2, ..., n.

3. The air-ground dynamic collaborative reasoning method for low-altitude intelligent networks according to claim 1, characterized in that, When both a drone and a ground station are selected to perform the inference task, the target layer of the deep learning model is the smallest subtask execution unit. The sum of the computation latency and transmission latency of the drone performing the target layer inference task is calculated according to the following formula: Among them, t large In the context of air-ground collaborative execution of large and complex model inference, the UAV executes the βth... k Layer inference task latency and ground station reception of the βth layer k The sum of the output data delays of +1 layer, the βth layer k The layer is the target layer; The computational cost required for the UAV to perform the j-th level inference task; P a The computing power of the drone; To execute the βth k The size of the data output by the layered inference task; ρ g The signal transmission rate of the drone; and These represent the computational latency of the UAV performing the inference task and the transmission latency of the ground station receiving the input data, respectively; in calculating ρ g In the formula, B g The vector is expressed as the transmission bandwidth of the air-to-ground channel; D is the distance between the UAV and the ground station; α' is the path loss exponent; g0 is the channel gain; and N0 is the noise power.

4. The air-ground dynamic collaborative reasoning method for low-altitude intelligent networks according to claim 1, characterized in that, The process of adjusting the participating node types and allocated model computation based on the obtained decision set of all tasks includes the following steps: Two decision variables are designed: α and β. α is used to decide whether to perform task inference independently, and β is the layer of the model at the decision split point. When a small, lightweight deep neural network model is selected for task inference independently, α = 0 and β = 0. When a UAV and a ground station are selected to jointly perform the inference task, α = 1, β ≠ 0, and β represents the layer of the model at the split point. The set of values ​​(α, β) of the two decision variables represents the system's inference decision in a dynamic environment. The system perceives the drone's status in real time. For different task arrival times in the task flow, it evaluates the task latency of different decision sets (α, β) under the current location and channel conditions by substituting the drone's position, assessing the arrival and completion status of the previous task, and considering memory and accuracy constraints. The reward value corresponding to the total latency under different values ​​is used to train a deep reinforcement learning DQN network, which eventually converges. The participating node type α and the allocated model computation β are adjusted according to the obtained decision sets of all tasks to perform low-latency collaborative inference under accuracy constraints.

5. The air-ground dynamic collaborative reasoning method for low-altitude intelligent networks according to claim 4, characterized in that, The method further includes: For the arrival of n tasks, the total system delay for inference is represented as t = {t1, t2, ..., tn}. k ..., t n For the k-th task, the task completion time T k Mission arrival time F k and computational delay C k This can be represented by the following relationship: With the goal of minimizing the total delay in the completion of all tasks in the entire task flow, the objective function of the optimization problem is defined as follows: Where, α k and β k Let t represent two decision variables for the k-th task, and the total task delay t. k It is calculated by combining computation latency, task arrival time, and computation delay. Construct constraints: In the formula, A k It is the precision of the k-th task, A set It is the lowest threshold of accuracy for all tasks; m j This is the memory requirement for the j-th level inference task, m a This is the available memory capacity of the system.

6. A dynamic collaborative reasoning device for low-altitude intelligent networks based on the method of any one of claims 1-5, characterized in that, The device includes: The model deployment module is used to deploy a small, lightweight deep neural network model and a large, complex deep neural network model, respectively, on a single drone and a ground station. The strategy evaluation module is used to detect and collect low-altitude inference tasks, and to evaluate the system status, available resources, and air-to-ground transmission channel conditions after receiving a push task request. The task decision module selects the model inference method based on the evaluation results. The model inference methods include two types: the UAV using a small and lightweight deep neural network model to perform inference alone, and the task being assigned to both the UAV and the ground station for joint execution. The task execution module executes the task reasoning process based on the selected model reasoning method.

7. A computer device, comprising a processor and a memory; characterized in that, When the processor executes the computer program stored in the memory, it implements the steps of the air-ground dynamic collaborative reasoning method for low-altitude intelligent networks as described in any one of claims 1-5.

8. A computer-readable storage medium for storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the steps of the air-ground dynamic collaborative reasoning method for low-altitude intelligent networks as described in any one of claims 1-5.

Images