A method and system for edge-cloud collaborative artificial intelligence inference applied to unmanned aerial vehicles (UAVs)

By quantifying the amount of feature information and the contribution of tasks, the model segmentation point is determined, a multi-dimensional execution scheme is constructed, UAV operating status data is collected, evaluation scores are calculated, and the best execution scheme is selected. This solves the dynamic adaptation problem of UAV edge-cloud collaborative schemes, optimizes resource utilization and extends battery life.

CN123117328APending Publication Date: 2026-11-17XIE HYDROGEN (SHANGHAI) NEW ENERGY TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611281209.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing drone AI inference solutions face a dilemma between local and cloud execution, and edge-cloud collaborative solutions lack dynamic adaptive decision-making capabilities, resulting in insufficient resource utilization and environmental adaptability.

Method used

By quantifying the feature information and task contribution of each intermediate layer of the artificial intelligence model, and combining the preset segmentation criteria to determine the model segmentation point, multiple candidate execution schemes are constructed. Multidimensional operation status data of UAVs are collected, and the evaluation scores of each candidate scheme are calculated based on these data. The target execution scheme is then selected to achieve edge-cloud collaborative artificial intelligence reasoning.

Benefits of technology

It enables the dynamic selection of the optimal AI inference execution method based on the real-time changing comprehensive status of the UAV, thereby optimizing energy utilization, improving the effectiveness of model segmentation, enhancing system response speed and mission reliability, and extending the UAV's endurance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117328A_ABST
    Figure CN123117328A_ABST
Patent Text Reader

Abstract

This invention relates to the fields of artificial intelligence and edge computing technology, specifically to a method and system for edge-cloud collaborative artificial intelligence inference applied to unmanned aerial vehicles (UAVs). The method includes: quantifying the feature information and task contribution of each intermediate layer of the artificial intelligence model to determine semantically aware model segmentation points; constructing multiple candidate execution schemes, including a model segmentation execution scheme; real-time acquisition of multi-dimensional operational status data of the UAV, such as communication link status, onboard computing resource status, energy reserve status, and task constraints; calculating the evaluation score of each candidate execution scheme based on the multi-dimensional operational status data, and dynamically adjusting the energy consumption weight according to the energy reserve status; finally, selecting the optimal target execution scheme based on the evaluation score and executing it. This invention achieves optimal allocation of artificial intelligence inference tasks between the edge and cloud through dynamic and intelligent decision-making, significantly improving the resource utilization efficiency, task adaptability, and endurance of UAVs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and edge computing technology, specifically to an edge-cloud collaborative artificial intelligence inference method and system applied to unmanned aerial vehicles (UAVs). Background Technology

[0002] With the development of artificial intelligence technology, drones, especially industrial drones that perform long-duration, complex missions, are increasingly relying on onboard AI models to complete key functions such as autonomous navigation, target detection, and environmental perception. These AI models, especially high-precision models, typically require enormous computational resources.

[0003] Currently, deploying AI models on drones faces two main challenges. The first is performing inference entirely on the drone's onboard device. Due to limitations in drone size, payload, and energy supply, the computing power, memory, and power consumption of its onboard computing units are severely restricted. Running large, high-precision models generates enormous energy consumption, significantly shortening the drone's operational endurance, and may also lead to excessively high inference latency, failing to meet the stringent timeliness requirements of tasks such as real-time obstacle avoidance.

[0004] The second approach involves completely offloading the AI ​​inference task to a remote cloud server. The cloud possesses powerful computing resources, easily capable of running any complex model. However, this method heavily relies on the wireless communication link between the drone and the cloud. In many industrial applications, such as line inspection in remote areas or emergency rescue, unstable 4G / 5G network coverage, limited bandwidth, or high latency are common. Fluctuations in communication quality can directly lead to delays or even loss of inference results, posing significant risks to the reliability and security of the task.

[0005] To balance the advantages and disadvantages of the two approaches mentioned above, edge-cloud collaborative inference technology has emerged. Its basic idea is to distribute computing tasks between edge devices (drones) and the cloud. However, most existing edge-cloud collaborative solutions are relatively rudimentary, such as only supporting binary switching of the entire model between the edge and the cloud, or using fixed, preset rules to segment the model. These solutions generally lack real-time perception and adaptive decision-making capabilities regarding the multi-dimensional operational status of the drone itself, and cannot flexibly adjust collaborative strategies based on dynamically changing conditions such as communication quality, onboard computing power load, remaining energy, and task real-time requirements. Therefore, existing technologies still have significant room for improvement in terms of resource utilization and environmental adaptability. Summary of the Invention

[0006] To address the aforementioned technical challenges, namely the dilemma between local and cloud-based execution in existing UAV AI inference schemes and the lack of dynamic adaptive decision-making capabilities in edge-cloud collaborative schemes, this invention provides an edge-cloud collaborative AI inference method and system for UAVs.

[0007] This invention provides an edge-cloud collaborative artificial intelligence inference method for unmanned aerial vehicles (UAVs), comprising the following steps: quantifying the feature information output and task contribution of each intermediate layer of the artificial intelligence model; determining the model segmentation points of the artificial intelligence model based on the quantification results and a preset segmentation criterion; constructing multiple candidate execution schemes, including a model segmentation execution scheme generated based on the model segmentation points; collecting multi-dimensional operational status data of the UAV, including communication link status, onboard computing resource status, energy reserve status, and task constraints; calculating the evaluation score of each candidate execution scheme based on the multi-dimensional operational status data; selecting the target execution scheme based on the evaluation score; controlling the UAV to run the artificial intelligence model according to the target execution scheme to complete the edge-cloud collaborative artificial intelligence inference. This method can dynamically and intelligently select the optimal artificial intelligence inference execution mode based on the real-time changing comprehensive status of the UAV, thereby achieving the best balance between performance, power consumption, and reliability.

[0008] Furthermore, the steps for quantifying the feature information content and task contribution of each intermediate layer of the artificial intelligence model specifically include: analyzing the artificial intelligence model layer by layer, calculating the information entropy of the output feature map of each layer as a quantitative indicator of feature information content, and simultaneously calculating the task relevance corresponding to the features of each layer based on the gradient attribution algorithm as a quantitative indicator of task contribution. This quantification method provides an objective and quantifiable basis for finding the optimal model split point.

[0009] Furthermore, the step of determining the model segmentation point of the artificial intelligence model specifically includes: identifying the layer in the intermediate layer that satisfies the first significant decrease in information entropy and the task relevance reaching a preset threshold as a semantic saturation point, and using the semantic saturation point as the model segmentation point to separate the model front-end and model back-end. Segmentation at this model segmentation point ensures that the intermediate features transmitted to the cloud are compressed while retaining information valuable to the final task to the greatest extent, thus improving the accuracy of the model segmentation execution scheme.

[0010] Furthermore, the energy reserve status refers to the remaining hydrogen fuel level of the drone; the multi-dimensional operational status data is periodically collected four-dimensional resource status data, which includes communication status, onboard computing power status, remaining hydrogen fuel level, and the real-time requirements of the current artificial intelligence task. This allows the decision-making process to comprehensively consider all key factors affecting drone operation, especially for energy-sensitive drone platforms, enabling more forward-looking resource planning.

[0011] Furthermore, the step of calculating the evaluation score of each candidate execution scheme based on the multi-dimensional operational status data specifically includes: calculating the comprehensive score of each candidate execution scheme using a weighted energy consumption product method based on the four-dimensional resource status data; the calculation process of the comprehensive score introduces an energy consumption weighting factor, which is inversely proportional to the current hydrogen fuel reserve of the UAV. This calculation method formalizes the multi-dimensional trade-off problem. When energy reserves are tight, the system will automatically tend to select the execution scheme with lower power consumption, thereby effectively extending the operation time.

[0012] Furthermore, the candidate execution scheme includes a local inference scheme where different model variants are run independently on the UAV's onboard unit; the method also includes: pre-processing the original artificial intelligence model using model compression technology to generate multiple model variants with different levels of differentiation; pre-loading all model variants into the cache of the onboard computing unit, so that the switching latency between model variants is less than a preset switching time threshold. This provides the UAV with a "hot standby" model pool, enabling it to quickly switch between different performance and power consumption modes to cope with sudden communication interruptions or computing power shortages.

[0013] Furthermore, the candidate execution scheme includes an edge-cloud segmentation inference scheme where the UAV-borne model front-end runs and the cloud-based model back-end runs. The method further includes configuring a lightweight feature compression module at the output of the model front-end, compressing the intermediate features to be transmitted using this module, and then transmitting the compressed feature data to the cloud. This effectively reduces the consumption of communication bandwidth, ensuring the model segmentation execution scheme remains feasible even under poor network conditions.

[0014] Furthermore, the step of preloading multiple model variants into the onboard computing unit's cache specifically includes: converting each model variant into a pre-compiled inference engine format and storing it resident in the onboard computing unit's video memory to achieve low-latency model switching. By pre-compiling and resident caching, the overhead of model switching is minimized, ensuring the continuity and real-time performance of task execution.

[0015] Furthermore, the method also includes a model pre-deployment step: based on the UAV's flight trajectory and the preset mission plan, predict the AI ​​model configuration required in the near future, and pre-load the corresponding model variants into the video memory or pre-start the corresponding inference pipeline. This predictive resource preparation method can eliminate the "cold start" delay caused by model loading and initialization, ensuring that the UAV is in optimal readiness when entering critical areas or performing critical tasks.

[0016] This invention also provides an edge-cloud collaborative artificial intelligence inference system for use in unmanned aerial vehicles, comprising: The system comprises the following components: an analysis unit for quantifying the feature information and task contribution of each intermediate layer of the artificial intelligence model; a determination unit for determining the model segmentation points of the artificial intelligence model based on the quantification results and a preset segmentation criterion; a construction unit for constructing multiple candidate execution schemes, including a model segmentation execution scheme generated based on the model segmentation points; an acquisition unit for collecting multi-dimensional operational status data of the UAV, including communication link status, onboard computing resource status, energy reserve status, and task constraints; a calculation unit for calculating the evaluation score of each candidate execution scheme based on the multi-dimensional operational status data; and a control unit for selecting the target execution scheme based on the evaluation score, controlling the UAV to run the artificial intelligence model according to the target execution scheme, and completing edge-cloud collaborative artificial intelligence inference.

[0017] The present invention also provides a drone equipped with the aforementioned edge-cloud collaborative artificial intelligence inference system for drones.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the edge-cloud collaborative artificial intelligence reasoning method for unmanned aerial vehicles as described in any of the preceding claims.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. Achieved dynamic intelligent decision-making based on multi-dimensional state perception. This invention incorporates real-time states of multiple dimensions such as communication, computing power, energy, and tasks into a unified decision-making framework. By calculating evaluation scores, it formalizes complex resource trade-offs, enabling UAVs to automatically select the optimal artificial intelligence inference execution scheme based on rapidly changing actual conditions, significantly improving resource utilization efficiency and environmental adaptability.

[0020] 2. Improved the effectiveness of model segmentation. By using a semantically aware model segmentation point determination method, the rationality of the segmentation position in terms of information theory and task contribution is ensured. This allows the transmitted intermediate features to retain more information useful to the task under the same communication bandwidth, thereby guaranteeing the final accuracy of edge-cloud collaborative inference.

[0021] 3. Enhanced system response speed and task reliability. By constructing a model variant hot standby pool and combining it with model pre-deployment technology, the switching latency between models with different performance and power consumption modes is extremely low, and preparations can be made in advance for upcoming tasks, eliminating the cold start delay of resource scheduling. This is of great significance for tasks with extremely high real-time requirements, such as obstacle avoidance.

[0022] 4. Optimized energy utilization efficiency and extended drone endurance. The decision-making mechanism incorporates an energy consumption weighting factor negatively correlated with energy reserve status, enabling the system to automatically shift to a low-power mode when energy is scarce. This achieves more intelligent energy management and effectively extends the drone's operational endurance. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the edge-cloud collaborative artificial intelligence reasoning method in this invention. Figure 2 This is a schematic diagram illustrating the working principle of the semantically aware model segmentation point search in this invention. Figure 3 This is a schematic diagram of the edge-cloud collaborative artificial intelligence reasoning system in this invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0025] This invention provides an edge-cloud collaborative artificial intelligence inference method and system for energy-constrained edge devices such as hydrogen-powered drones. The core idea of ​​this method is to treat a complete artificial intelligence model as a dynamically reconfigurable computational pipeline, rather than statically switching between "edge operation" and "cloud operation." By deeply decomposable designing the artificial intelligence model and combining it with real-time perception of the drone's multi-dimensional operational status data, the system can dynamically decide the optimal inference execution scheme online, thereby maximizing energy efficiency and adapting to unstable network environments while ensuring mission performance. In the context of this invention, "model split point" refers to a specific intermediate layer position that divides a complete artificial intelligence model into a front-end and a back-end part; "multi-dimensional operational status data" refers to a data set reflecting the current comprehensive state of the drone across multiple dimensions, such as communication, computing power, energy, and mission requirements; and "model variants" are multiple versions of the original artificial intelligence model generated through model compression technology, differing in accuracy and power consumption. This invention aims to address the technical pain points of current hydrogen-powered drones when performing complex tasks, namely "high power consumption during airborne operation and reliance on communication for cloud operation". By introducing a mechanism of deep model decomposability and multi-dimensional resource perception and decision-making, it enables flexible and efficient collaboration of artificial intelligence inference tasks between the edge and the cloud.

[0026] The technical solution of the present invention will be described in detail below. In one aspect of the present invention, an edge-cloud collaborative artificial intelligence inference method applied to unmanned aerial vehicles (UAVs) is provided. For example... Figure 1 As shown, the method includes the following steps: S101. Quantify the amount of feature information output by each intermediate layer of the artificial intelligence model and its contribution to the task. S102. Based on the quantification results and in conjunction with the preset segmentation criteria, determine the model segmentation points of the artificial intelligence model; S103. Construct multiple candidate execution schemes, including a model segmentation execution scheme generated based on the model segmentation points; S104. Collect multi-dimensional operational status data of the UAV, including communication link status, onboard computing resource status, energy reserve status, and mission constraints. S105. Based on the multi-dimensional operating status data, calculate the evaluation score for each candidate execution scheme; S106. Based on the evaluation score, select the best target execution plan, control the drone to run the artificial intelligence model according to the target execution plan, and complete the edge-cloud collaborative artificial intelligence inference.

[0027] Specifically, this method first requires analyzing an AI model to be deployed, specifically quantifying the amount of feature information output by each intermediate layer of the AI ​​model and its contribution to the task. This step is fundamental to achieving intelligent model segmentation. Through this quantitative analysis, a deeper understanding of the role each layer plays in feature extraction and semantic reasoning can be achieved. Subsequently, based on the aforementioned quantitative results and combined with preset segmentation criteria, the model segmentation point of the AI ​​model is determined. This segmentation point is not arbitrarily chosen, but rather based on a profound insight into the information flow within the model, aiming to find a balance point that allows the segmented front-end portion to extract sufficiently rich information with relatively low computational cost.

[0028] After determining the model segmentation point, the system constructs multiple candidate execution schemes. These schemes include at least a model segmentation execution scheme generated based on the model segmentation point, i.e., a collaborative mode where the model front-end runs on the UAV's onboard terminal and the model back-end runs in the cloud. Furthermore, candidate schemes typically include purely onboard and purely cloud-based operation schemes to form a complete decision space. To make the optimal choice, it is necessary to collect multi-dimensional operational status data of the UAV in real time. This data comprehensively reflects the UAV's current operating environment and its own status, specifically including communication link status, onboard computing resource status, energy reserve status, and the constraints of the current artificial intelligence task on real-time performance, accuracy, etc.

[0029] Next, the system calculates an evaluation score for each candidate execution plan based on real-time collected multi-dimensional operational status data. This calculation process is a multi-objective optimization process, aiming to unify the status data from different dimensions into a quantitative evaluation system. Finally, the system selects the optimal target execution plan based on the evaluation scores of each candidate execution plan, and controls the UAV to run the artificial intelligence model according to the target execution plan, thereby completing one edge-cloud collaborative artificial intelligence inference. Through this series of steps, this invention realizes a dynamic and adaptive artificial intelligence inference load allocation mechanism, enabling the UAV to maintain optimal performance and energy efficiency under varying operating conditions.

[0030] Furthermore, the specific implementation method for quantifying the feature information content and task contribution of each intermediate layer of the artificial intelligence model includes: First, the target artificial intelligence model is analyzed layer by layer, traversing each computational layer in its network structure. For each layer, the information entropy of its output feature map is calculated, and this information entropy value is used as a quantitative indicator of the feature information content of that layer. Information entropy can effectively measure the uncertainty or richness of the information contained in the feature map. At the same time, gradient-based attribution algorithms, such as Gradient Weighted Class Activation Mapping (Grad-CAM) or similar techniques, are used to calculate the contribution or correlation of each layer's features to the final task output, and this calculation result is used as a quantitative indicator of task contribution. This design ensures that the evaluation of the model's internal state considers both the amount of information and the degree of correlation between the information and the final task, providing a more comprehensive basis for determining subsequent segmentation points.

[0031] As a preferred implementation, the specific process for determining the model segmentation points of the artificial intelligence model is as follows: After obtaining quantified data on feature information (information entropy) and task contribution (task relevance) layer by layer, the system analyzes the trend of these data changing with network depth. For example... Figure 2 As shown, in a typical neural network model 201, as the number of intermediate layers increases, the information entropy change curve 202 typically undergoes a process from rising to stabilizing and then falling, while the task relevance change curve 203 usually increases monotonically or grows rapidly before saturating. This invention identifies the intermediate layer that first satisfies the condition of "a significant decrease in information entropy and task relevance reaching a preset threshold (e.g., 90% of the saturation value)" as the semantic saturation point 204. This semantic saturation point 204 is selected as the model segmentation point for separating the front-end and back-end of the model. The technical advantage of choosing this point as the segmentation point is that the intermediate layers before this point mainly complete the extraction of abstract features from the original input, and their output intermediate features already contain most of the key information needed to complete the task. The intermediate layers after this point perform more high-level semantic combination and reasoning, which involves a large computational load but limited information gain. Therefore, segmenting at this point allows the intermediate data transmitted to the cloud to achieve an ideal balance between information completeness and data volume.

[0032] The determination of a significant decrease in information entropy adopts the following quantification rules: For the output feature map of the i-th layer, calculate its information entropy E_i (unit: bits). Define the first-order difference of information entropy between adjacent layers as ΔE_i = E_{i+1}-E_i, and the second-order difference as Δ²E_i = ΔE_{i+1}-ΔE_i. When the following three conditions are met simultaneously, the i-th layer is determined to be the starting layer of "the first significant decrease in information entropy": (1) ΔE_i < -ε, where ε is the preset decrease sensitivity threshold (the default value is ε = 0.05 × max(E), which is 5% of the maximum information entropy value); (2) Δ²E_i < 0, indicating that the downward trend is accelerating; (3) Starting from the i-th layer, the information entropy of the three consecutive layers shows a monotonically decreasing trend (i.e., ΔE_j < 0, j = i, i+1, i+2), so as to eliminate misjudgment caused by single-layer fluctuations. Taking the ResNet-50 test on the ImageNet dataset as an example, the information entropy of the 22nd layer is 6.8 bits, and the 23rd layer drops to 6.1 bits (ΔE=-0.7, exceeding the threshold of ε=0.4), and the second difference is negative. After that, the 24th and 25th layers continue to decrease. Therefore, the 22nd layer is judged as the starting layer of "first significant decrease".

[0033] To verify the effectiveness of the semantically aware segmentation point selection method described above, this embodiment uses the test results of the YOLOv5 object detection model on the public dataset VisDrone as an example. Three different intermediate layers were selected as segmentation points for comparative experiments: layer 10 (information entropy has not yet decreased, task relevance is approximately 65%), layer 15 (semantic saturation point, information entropy decreases significantly for the first time, task relevance reaches 95%), and layer 20 (information entropy has continued to decrease, task relevance reaches 98%). Experimental results show that when segmenting at layer 15 (semantic saturation point), the amount of intermediate feature data transmitted to the cloud is 42% of that of the layer 10 segmentation scheme, while the final detection accuracy (mAP@0.5) of the cloud model backend only decreases by 1.2%. In contrast, although the data volume of the layer 20 segmentation scheme is further reduced to 28% of that of layer 10, the accuracy decreases by 4.7%. These experimental data demonstrate that segmentation at the semantic saturation point achieves the optimal balance between data compression efficiency and task accuracy maintenance.

[0034] In the specific application scenario of this invention, the drone is a hydrogen-powered industrial drone, and its energy reserve status is the drone's remaining hydrogen fuel status. In this case, the multi-dimensional operational status data is specifically defined as periodically collected four-dimensional resource status data in real time. This collection process can be executed at a fixed frequency, such as once per second, to ensure the timeliness of decision-making. This four-dimensional resource status data specifically includes: communication status, quantified by real-time measurement of network round-trip time (RTT), packet loss rate, and effective bandwidth, for example, through WebRTC statistical interfaces; onboard computing power status, assessed by monitoring the utilization rate, memory usage, and chip temperature of onboard computing units (such as GPUs / CPUs); hydrogen fuel remaining status, comprehensively judged by reading the percentage of remaining hydrogen calculated from hydrogen cylinder pressure sensor data, the current output power of the fuel cell stack, and the expected hydrogen consumption in the next 10 minutes output by a hydrogen consumption prediction model; and the real-time requirements of the current artificial intelligence task, i.e., task constraints, for example, for obstacle avoidance tasks, the maximum allowable latency must be less than 50ms, while for target detection tasks, the latency can be relaxed to 200ms. This refined four-dimensional perception provides a solid data foundation for subsequent accurate decision-making.

[0035] Based on the aforementioned four-dimensional resource status data, the specific method for calculating the evaluation score of each candidate execution scheme is as follows: a weighted energy consumption product method is used to calculate the comprehensive score of each candidate execution scheme. This comprehensive score calculation process introduces a key energy consumption weighting factor, which is inversely proportional to the current hydrogen fuel reserve of the UAV. Specifically, the evaluation score (comprehensive score) can be calculated using the following formula: Comprehensive Score = (1 / Latency) × Accuracy × Energy Consumption Weighting Factor. Here, "Latency" is the estimated end-to-end time for executing the scheme, "Accuracy" is the estimated task completion quality for executing the scheme, and "Energy Consumption Weighting Factor" is a dynamically adjusted coefficient. When the hydrogen fuel reserve is sufficient, this factor is smaller, and the decision is more inclined to select a high-performance (high accuracy, low latency) scheme; while when the hydrogen fuel reserve is tight, this factor will increase significantly, making the decision more inclined to select a low-power scheme.

[0036] The definitions and calculation methods for the parameters of the overall score are as follows: (I) Calculation method of latency T. For the local inference scheme, latency T_local is estimated as follows: T_local = T_compute(M_v, S_load), where T_compute is the inference time of the onboard computing unit executing model variant M_v, which is obtained by averaging 100 inference tests on each model variant on the target hardware in advance, in milliseconds (ms); S_load is the current onboard computing load coefficient (value range 0~1, calculated by combining GPU utilization and memory usage), and the actual latency T_local × (1 + 0.3 × S_load) is corrected. For the cloud inference scheme, latency T_cloud = T_trans(RTT, loss) + T_cloud_compute, where T_trans is the data transmission time, estimated based on the current measured round-trip time RTT (in ms) and packet loss rate loss; T_cloud_compute is the inference time of the cloud server (assuming sufficient cloud computing resources, a fixed value is taken). For the split execution scheme, the delay T_split = T_front_compute + T_compress + T_trans_feature + T_back_compute, where each component corresponds to the front-end inference time, feature compression time, feature transmission time, and back-end inference time, respectively.

[0037] (II) Calculation Method of Accuracy (Acc). Accuracy (Acc) represents the expected task completion quality of each execution scheme, expressed as a normalized value (range 0~1). For local model variant schemes, Acc is the test accuracy of the model variant on the validation set (e.g., mAP for object detection), obtained through pre-testing. For cloud inference schemes, Acc is the test accuracy of the original high-precision model (normalized to 1). For segmentation execution schemes, Acc = Acc_front_back × (1 - L_compression), where Acc_front_back is the theoretical accuracy of complete inference with the model front-end and back-end, and L_compression is the accuracy loss coefficient due to feature compression (determined by the compression ratio, range 0~0.05).

[0038] (III) Dynamic Adjustment Rules for Energy Consumption Weighting Factor W_energy. The energy consumption weighting factor W_energy is inversely proportional to the current hydrogen fuel reserve H_remain (percentage, ranging from 0 to 100%) of the UAV. The specific calculation formula is: W_energy = 1 + α × (100 - H_remain) / 100, where α is the preset energy sensitivity coefficient (ranging from 1 to 5, with a default value of 3). When the hydrogen fuel reserve is 90%, W_energy = 1 + 3 × 0.1 = 1.3; when the hydrogen fuel reserve drops to 20%, W_energy = 1 + 3 × 0.8 = 3.4. This factor is used as a product term in the comprehensive score formula, where T is the aforementioned delay value (in ms), and Acc is the accuracy normalization value (0 to 1). The physical meaning of this formula is: when energy is sufficient, the system tends to choose the low-latency, high-precision scheme; when energy is scarce, the energy consumption weighting factor increases significantly, allowing the low-power scheme to obtain a higher comprehensive score.

[0039] In this way, the formula makes explicit and model the trade-offs between the four dimensions of communication, computing power, energy, and tasks. Compared with fixed heuristic rules that rely on manual settings, it can make more intelligent and fuller use of rapidly changing resource conditions to achieve global optimization.

[0040] To further expand the decision space, the candidate execution schemes described in this invention, in addition to purely airborne, purely cloud-based, and segmented execution schemes, also include a local inference scheme where different model variants are run independently on the UAV's airborne terminal. To this end, the method also includes a preprocessing step: the original artificial intelligence model is pre-processed using model compression techniques, such as network pruning, weight quantization, and knowledge distillation, to generate multiple differentiated model variants. For example, a high-precision model at level 0 (FP16 precision, 20W power consumption), a balanced model at level 1 (INT8 quantization, 12W power consumption), an energy-efficient model at level 2 (knowledge distillation, 6W power consumption), and an emergency model at level 3 (extreme compression, 3W power consumption) can be generated. These model variants form a gradient in terms of precision, power consumption, and computational cost, providing the decision-maker with a richer selection.

[0041] To achieve efficient switching between these model variants, this invention preloads all generated model variants into the cache of the onboard computing unit, such as the GPU's video memory. This "hot standby" approach allows for direct pointer or context switching within the cache when a model switch is needed, without reloading the model from slow storage (such as flash memory). This design ensures that the switching latency between model variants is less than a preset switching time threshold, such as 50 milliseconds. This low-latency switching capability is crucial for real-time-critical tasks, such as rapidly switching from a general perception model to a high-precision obstacle avoidance model during flight based on environmental changes, thereby avoiding safety risks caused by model loading issues.

[0042] To achieve low-latency switching of the aforementioned model variants, a preferred approach to preloading multiple model variants into the onboard computing unit's cache is to pre-convert each model variant, whether the original model or a compressed model of different levels, into a pre-compiled inference engine format optimized for the target hardware. For example, when using platforms such as NVIDIA Jetson, the model can be converted to the TensorRT engine format. These pre-compiled engine files are all loaded and resident in the onboard computing unit's video memory during system initialization or before the task begins. When the decision-maker instructs a model switch, the control unit only needs to call the different engine to perform inference, avoiding time-consuming operations such as model parsing, network construction, and weight loading. This strategy of maintaining model variants in video memory in a "hot standby" state is a key technical guarantee for achieving a switching latency of less than 50ms.

[0043] To achieve low-latency switching between model variants, this embodiment further provides specific implementation methods and test data. Taking the NVIDIA Jetson AGX Orin platform as an example, each model variant (level 0 to level 3) is converted into a pre-compiled inference engine format (.engine file) by the TensorRT optimizer and uniformly loaded into the GPU memory during system initialization. The engine context (ExecutionContext) corresponding to each model variant is maintained in the form of a pointer array. During switching, it is only necessary to point the current execution pointer of the inference session to the context of the target model variant and reset the input and output tensor buffers. Test data shows that, under the above implementation method, the switching latency between level 0 and level 1 is 12ms, and the switching latency between level 0 and level 3 is 18ms, both of which are far better than the preset switching time threshold of 50ms. In comparison, if the model is dynamically loaded from flash memory, the latency of the same switching operation is about 800ms to 1200ms. By using the pre-compiled engine resident in GPU memory and the context pointer switching mechanism, the latency caused by model loading and initialization is effectively eliminated, ensuring the response speed of real-time tasks such as obstacle avoidance.

[0044] In the implementation of the model segmentation execution scheme, namely the edge-cloud segmentation inference scheme in which the UAV onboard terminal runs the model front-end and the cloud runs the model back-end, the method of the present invention further includes an optimization step to further reduce the consumption of communication bandwidth. Specifically, at the output end of the model front-end, i.e., after the model segmentation point, a lightweight feature compression module is configured. This module can be a small bottleneck layer structure of an autoencoder or other efficient feature compression algorithms. Through this module, the amount of intermediate feature data to be transmitted can be compressed to 15% to 25% of the original size. At the same time, since the segmentation point is selected based on semantic saturation, key information is preserved, so the impact on the final task accuracy is minimal. This design significantly reduces the communication overhead in the edge-cloud collaborative mode, making it more feasible in bandwidth-constrained scenarios.

[0045] To further clarify the specific implementation of the lightweight feature compression module, the present invention provides the following preferred solutions: (I) Module Network Structure. The lightweight feature compression module adopts a symmetrical encoder-decoder structure. The encoder consists of three convolutional blocks: the first convolutional block is a 3×3 convolution (input channel C_in, output channel C_in / 2) + batch normalization + ReLU, with a stride of 2; the second convolutional block is a 3×3 convolution (C_in / 2→C_in / 4) + batch normalization + ReLU, with a stride of 2; the third convolutional block is a 1×1 convolution (C_in / 4→C_bottleneck) + batch normalization, with no activation function, where C_bottleneck is the number of bottleneck layer channels (usually C_in / 4 to C_in / 8). The decoder consists of three deconvolutional blocks, with a structure symmetrical to the encoder, restoring the bottleneck features to their original spatial dimensions. The total number of parameters in the entire module is controlled within 5% of the number of parameters in the front end of the original model to ensure its lightweight characteristics.

[0046] (II) Training Method. The feature compression module adopts a self-supervised training strategy. The training dataset consists of image data (no less than 10,000 images) collected by UAVs in typical operational scenarios. During training, the complete artificial intelligence model (taking the object detection model as an example) is run on the dataset, and the original intermediate feature maps at the segmentation points (semantic saturation points) are extracted as the supervision target. The compression module takes the original intermediate feature maps as input, compresses them through the encoder, and then reconstructs them through the decoder. The training loss function is the sum of the reconstruction loss and the semantic loss: L_total = L_recon + β × L_semantic. Where L_recon is the mean squared error loss, which measures the pixel-level difference between the reconstructed feature map and the original feature map; L_semantic is the task loss, which calculates the difference between the reconstructed feature map input to the model backend and the task output (such as detection boxes and categories) obtained by inputting the original feature map into the model backend (for object detection tasks, CIoU loss and cross-entropy loss are used); β is the balance coefficient (default is 0.1). By jointly optimizing the reconstruction loss and the semantic loss, it is ensured that the compressed features retain semantic information while minimizing distortion.

[0047] (III) Semantic Integrity Guarantee Mechanism. During the inference phase, the compression module only uses the encoder portion (the decoder is only used for supervised learning during the training phase). To further ensure the semantic integrity of the compressed features, a feature importance reweighting mechanism is introduced after the bottleneck layer of the encoder: based on the gradient attribution algorithm, the contribution of each channel in the bottleneck layer to the final task output is calculated. Channels with a contribution value lower than a preset threshold (default is 20% of the average contribution value) are pruned, while channels with high contribution values ​​are enhanced. This mechanism ensures that the compressed features transmitted to the cloud focus on information channels valuable to the task without increasing the computational burden. Experimental data shows that after adopting the above compression module, the target detection accuracy (mAP) at a compression ratio of 16:1 is controlled to decrease by less than 2%, which is better than the control group without semantic loss supervision (accuracy decrease of about 5%).

[0048] Furthermore, to address the dynamic changes in mission requirements across time and space, the method of this invention may also include a model pre-deployment step. The core idea of ​​this step is to utilize predictive information to prepare computing resources in advance. Specifically, the system predicts the AI ​​tasks and required model configurations that the UAV is most likely to perform within a short time window (e.g., the next 30 seconds) based on the UAV's current flight trajectory and pre-set mission plan. For example, when the UAV's GPS trajectory prediction indicates that it will enter a densely populated power tower inspection area in 40 seconds, the system triggers a predictive model preloading module. This module preloads a "high-precision defect detection model" (e.g., a level 0 model variant) for power tower defect detection from flash memory into video memory, or, if the video memory is full, replaces the currently running but soon-to-be-unnecessary model (e.g., an "energy-efficient general perception model" for open-area flight). In this way, it ensures that the required core mission models are already in a "hot standby" state and can be immediately invoked when the UAV arrives at the critical mission area, thereby eliminating the cold start delay that may be caused by model switching and ensuring the continuity of mission execution and the timeliness of response.

[0049] In another aspect of the invention, an edge-cloud collaborative artificial intelligence inference system for use in unmanned aerial vehicles (UAVs) is also provided. This system serves as a hardware or hardware-software integration implementation of the aforementioned methods. The system can be integrated into the UAV's onboard computing platform (such as NVIDIA Jetson). Figure 3 As shown, the system specifically includes: an analysis unit 31, used to quantify the amount of feature information and task contribution of each intermediate layer of the artificial intelligence model; a determination unit 32, used to determine the model segmentation point of the artificial intelligence model based on the quantification results and in combination with preset segmentation criteria; a construction unit 33, used to construct multiple candidate execution schemes, including a model segmentation execution scheme generated based on the model segmentation point; an acquisition unit 34, used to collect multi-dimensional operating status data of the UAV, including communication link status, onboard computing resource status, energy reserve status, and task constraints; a calculation unit 35, used to calculate the evaluation score of each candidate execution scheme based on the multi-dimensional operating status data; and a control unit 36, used to select the target execution scheme based on the evaluation score, control the UAV to run the artificial intelligence model according to the target execution scheme, and complete edge-cloud collaborative artificial intelligence inference.

[0050] The acquisition unit acquires multi-dimensional operational status data of the UAV in real time, including communication link status, onboard computing resource status, energy reserve status (such as hydrogen fuel reserves), and current task constraints. The computing unit, the "brain" of the entire system, receives the data collected by the acquisition unit and calculates an evaluation score for each candidate execution scheme, including a model segmentation execution scheme, based on the multi-dimensional operational status data. Notably, the built-in evaluation score weighting logic of this computing unit explicitly stipulates that the influence weight of energy reserve status is negatively correlated with the amount of energy reserves; that is, the less energy available, the higher the weight of energy consumption factors in the decision-making process. Finally, the control unit, based on the evaluation scores of each candidate execution scheme output by the computing unit, selects the optimal target execution scheme and issues instructions to control the UAV's AI task scheduler to run the AI ​​model according to the target execution scheme, such as initiating local inference, loading specific model variants, or sending intermediate feature data to a cloud server.

[0051] The present invention also provides a drone equipped with the aforementioned edge-cloud collaborative artificial intelligence inference system for drones. This drone is preferably a hydrogen-powered industrial drone, whose energy system's unique characteristics enable the energy efficiency optimization strategy of the present invention to achieve significant beneficial effects.

[0052] Furthermore, the present invention provides a computer-readable storage medium on which a computer program is stored. When the instructions of the computer program are executed by a processor, they can implement the edge-cloud collaborative artificial intelligence inference method for UAVs described in the foregoing embodiments. The storage medium can be any type of non-volatile or volatile memory, such as ROM, RAM, flash memory, hard disk, etc.

[0053] To more clearly illustrate the technical solution of the present invention, a detailed description is provided below with reference to specific embodiments. Consider the following scenario: a hydrogen-powered drone is performing a complex power line inspection task. In this embodiment, before mission deployment, we analyze an artificial intelligence model for target detection required for the inspection task, quantifying the amount of feature information output by each convolutional layer and its contribution to the final detection box localization and classification. Based on this quantified data, we determine a model segmentation point according to a simple criterion (e.g., when the computational cost reaches 40% of the total computational cost).

[0054] Subsequently, we constructed three candidate execution schemes for the model: Scheme A (pure airborne operation), Scheme B (pure cloud-based operation), and Scheme C (model-split execution, i.e., the front-end runs on the airborne system and the back-end runs in the cloud). During the UAV's mission, the airborne system periodically collects multi-dimensional operational status data, which we simplify to four discrete states: communication quality (good / medium / poor), computing load (high / low), energy reserves (sufficient / scarce), and mission latency requirements (high / low). The system scores the three schemes based on these states. For example, when the communication quality is "poor," the scores of Schemes B and C will drop significantly; when the energy reserves are "scarce," the score of Scheme A will decrease due to its high power consumption. Finally, the system selects the scheme with the highest current score and executes it immediately. For example, when flying in open areas with good communication and sufficient energy, the system may choose Scheme B or C to save airborne computing power; when entering mountainous areas with no signal, the system will automatically switch to Scheme A to ensure uninterrupted mission execution. This basic embodiment demonstrates the core logical closed loop of the dynamic decision-making of this invention.

[0055] Based on the above embodiments, a specific implementation of the present invention provides a detailed definition of the model segmentation process to achieve more intelligent segmentation. Taking a YOLOv5-based object detection model as an example, as follows... Figure 2 As shown, each intermediate layer of the neural network model 201 is analyzed. The Shannon entropy of the output feature map is calculated layer by layer, resulting in the entropy change curve 202. Simultaneously, the contribution of each layer's features to the final detection result is calculated using the gradient attribution method, resulting in the task relevance change curve 203. In a certain intermediate layer of the network (e.g., layer 15), the entropy first shows a significant decrease, while the task relevance has reached 95% of its saturation value. We determine this layer as the semantic saturation point 204 and use it as the model segmentation point. This point is chosen because the model has already fully extracted the low-level and mid-level features of the image, and subsequent intermediate layers mainly perform logical reasoning; therefore, the intermediate features transmitted after segmentation have the highest cost-effectiveness.

[0056] To address data transmission issues in the model segmentation execution scheme, a lightweight feature compression module was deployed at the segmentation point. This module employs a small autoencoder structure, pre-trained, capable of compressing an intermediate feature map of size [256, 40, 40] generated by front-end inference into a compact representation of size [64, 20, 20], achieving a compression ratio of 16:1, or approximately 6.25% of the original size, far superior to the 15%-25% compression of typical targets. The complete data flow is as follows: images are captured by the UAV camera; the onboard computing platform runs the model front-end for inference, obtaining an intermediate feature map; this feature map is then fed into the feature compression module; the compressed features are wirelessly transmitted to the cloud server via a 4G / 5G network; and the cloud server then runs the model back-end to complete the final target detection. Through this semantically aware segmentation and feature compression, complex cloud-based collaborative inference can be completed with minimal accuracy loss, even under limited bandwidth conditions.

[0057] In another specific implementation, the decision-making mechanism and management of model variants are further elaborated. Specifically, the onboard four-dimensional resource awareness decision-maker executes the decision-making process once per second. During the acquisition of four-dimensional resource status, precise four-dimensional data is obtained, such as: communication status (RTT=80ms, packet loss rate=1%), onboard computing power status (GPU utilization=65%, temperature=72℃), hydrogen fuel reserve status (remaining hydrogen=60%), and task real-time requirements (currently performing an obstacle avoidance task, requiring latency <50ms). During the traversal of candidate execution schemes, the decision-maker evaluates all feasible schemes.

[0058] Specifically, the candidate solutions at this point include not only purely onboard, purely cloud-based, and segmented execution solutions, but also multiple local solutions from a model compression hot-standby pool. This hot-standby pool is located in the onboard GPU memory and pre-configures four levels of model variants for the obstacle avoidance task: Model Variant (Level 0, High Precision), using FP16 precision, 20W power consumption; Model Variant (Level 1, Balanced), quantized with INT8, 12W power consumption; Model Variant (Level 2, Energy Saving), knowledge distilled, 6W power consumption; Model Variant (Level 3, Emergency), extremely compressed, 3W power consumption. All these variants have been compiled into the TensorRT engine format, with a switching latency of less than 50ms. In calculating the weighted energy efficiency product score, the decision-maker uses the formula: Overall Score = (1 / Latency) × Precision × Energy Consumption Weighting Factor, to score each solution.

[0059] For example, the decision-making process is simulated under two scenarios. Scenario 1: Sufficient hydrogen fuel (90%), good communication (RTT < 50ms). In this scenario, the energy consumption weighting factor is small, and the decision-maker will prioritize the option with the best accuracy and latency performance, such as the "model segmentation execution option" or the "local operating level 0 model variant," to achieve the best obstacle avoidance effect. Scenario 2: Low hydrogen fuel (20%), communication interrupted. In this scenario, the energy consumption weighting factor is extremely large, and the score for cloud-based options is zero due to infinite latency. The decision-maker will choose from local model variants. Due to its extreme sensitivity to energy consumption, it will ultimately choose the "local operating level 2 or level 3 model variant," sacrificing some accuracy to maximize flight time. This refined decision-making and model variant management gives the UAV strong environmental adaptability and mission endurance.

[0060] To further reduce the latency caused by task switching, another embodiment of the present invention introduces a predictive model pre-deployment mechanism. For example, a specific flight mission is set: a hydrogen-powered drone takes off from a base and patrols a 5-kilometer-long power transmission corridor along a planned route. The last kilometer section has exceptionally dense power towers and vegetation, requiring high-precision obstacle avoidance and defect detection capabilities. For the first 4 kilometers after takeoff, the drone is in an open area and primarily uses an "energy-efficient general perception model" (Level 2) for basic environmental perception and navigation.

[0061] The onboard predictive model preloading module continuously monitors the UAV's GPS position and mission plan. When the system predicts, based on the UAV's flight speed and remaining distance, that it will enter the final kilometer of a complex inspection area within the next 30 seconds, it triggers a model pre-deployment operation. Specifically, the control unit instructs the loading of the "high-precision defect detection model" (level 0) and "high-precision obstacle avoidance model" (level 0) used for inspection from flash memory into the onboard GPU memory, replacing the previously used "energy-saving general perception model." This loading process is completed in the background and does not affect the current flight mission. When the UAV actually arrives at the entrance of the complex area, the required high-performance model is fully ready and in a hot standby state, and can be immediately invoked by the decision-maker. This predictive loading mechanism effectively eliminates the "second-level" latency caused by model switching and loading, ensuring seamless connection of computing resources at critical mission nodes, and greatly improving flight safety and operational efficiency.

[0062] Using a multi-rotor hydrogen-powered drone as a carrier, the core of which is a high-performance airborne computing platform (such as NVIDIA Jetson AGX Orin). This platform runs all the software modules of the edge-cloud collaborative artificial intelligence inference system described in this invention and connects to a cloud server via a 4G / 5G network. The entire workflow forms a complete closed loop from model design, resource awareness, dynamic decision-making to predictive execution.

[0063] During the offline phase, developers use the analysis and determination units to analyze all task-specific AI models (such as obstacle avoidance, object detection, and scene segmentation), determine the optimal semantic saturation point for each model, and generate model variants with multiple levels. These processed models and their segmentation information are then deployed to the UAV. At the start of the flight mission, the acquisition unit starts as a background service, continuously collecting four-dimensional resource data through system APIs and sensor interfaces. Simultaneously, the predictive model preloading module loads the models required for the initial phase into the model compression hot-standby pool according to the mission plan.

[0064] During flight, the computing unit evaluates all candidate execution schemes (including local model variants at different levels, segmentation execution schemes, etc.) every second based on the latest data collected by the acquisition unit, and calculates a weighted energy efficiency product score. For example, when the drone has sufficient power and a strong network signal, the decision-maker may choose to execute computationally intensive scene segmentation tasks in the cloud; when the network signal weakens, it will automatically switch to running a medium-precision model variant locally; when hydrogen fuel is detected to be below the warning level, it will force all non-critical tasks to switch to the lowest power emergency model variant. After receiving the optimal scheme output by the decision-maker, the control unit immediately executes the scheduling: either calling the corresponding pre-compiled engine in the model compression hot standby pool, or packaging the intermediate features of the front-end inference and sending them to the cloud server. The entire process is seamless and automatic, making the drone seem to have an intelligent brain that can "deliberate" and "adapt to changing circumstances," allowing it to complete tasks with ease in complex environments.

Claims

1. A method for edge-cloud collaborative artificial intelligence reasoning applied to unmanned aerial vehicles (UAVs), characterized in that, include: Quantify the amount of feature information output by each intermediate layer of the artificial intelligence model and its contribution to the task; Based on the quantification results and in conjunction with preset segmentation criteria, the model segmentation points of the artificial intelligence model are determined; Construct multiple candidate execution schemes, including a model segmentation execution scheme generated based on the model segmentation points; Collect multi-dimensional operational status data of the UAV, including communication link status, onboard computing resource status, energy reserve status, and mission constraints. Based on the multidimensional operational status data, the evaluation score of each candidate execution scheme is calculated. Based on the evaluation score, the optimal target execution plan is selected, and the drone is controlled to run the artificial intelligence model according to the target execution plan to complete the edge-cloud collaborative artificial intelligence inference.

2. The method according to claim 1, characterized in that, Quantify the feature information content and task contribution of each intermediate layer of the artificial intelligence model, including: The artificial intelligence model is analyzed layer by layer, and the information entropy of the output feature map of each layer is calculated as a quantitative indicator of the amount of feature information. At the same time, the task relevance of each layer's features is calculated based on the gradient attribution algorithm as a quantitative indicator of task contribution.

3. The method according to claim 2, characterized in that, Determining the model segmentation points of the artificial intelligence model includes: The intermediate layer is defined as the semantic saturation point when the information entropy decreases significantly for the first time and the task relevance reaches a preset threshold. The semantic saturation point is then used as the model segmentation point between the front end and the back end of the model.

4. The method according to claim 1, characterized in that, The energy reserve status refers to the remaining hydrogen fuel status of the drone; the multi-dimensional operational status data refers to the periodically collected four-dimensional resource status data, which includes communication status, onboard computing power status, remaining hydrogen fuel status, and the real-time requirements of the current artificial intelligence task.

5. The method according to claim 4, characterized in that, The evaluation score for each candidate execution scheme is calculated based on the multidimensional runtime status data, including: Based on the four-dimensional resource status data, a weighted energy consumption product method is used to calculate the comprehensive score of each candidate execution scheme; the calculation process of the comprehensive score introduces an energy consumption weight factor, which is inversely proportional to the current hydrogen fuel reserve of the UAV.

6. The method according to claim 1, characterized in that, The candidate execution schemes include local inference schemes where different model variants are run independently on the UAV's onboard terminal; the method further includes: The original artificial intelligence model is pre-processed using model compression technology to generate multiple model variants with different levels of differentiation; All model variants are preloaded into the cache of the onboard computing unit, so that the switching latency between model variants is less than a preset switching time threshold.

7. The method according to claim 6, characterized in that, The candidate execution schemes include edge-cloud segmentation inference schemes that combine the UAV onboard operating model front-end with the cloud-based operating model back-end; the method further includes: A lightweight feature compression module is configured at the output end of the model front end. The intermediate features to be transmitted are compressed by the lightweight feature compression module, and then the compressed feature data is transmitted to the cloud.

8. The method according to claim 6, characterized in that, Multiple model variants are preloaded into the onboard computing unit cache, including: Each model variant is converted into a pre-compiled inference engine format and stored resident in the GPU memory of the onboard computing unit to achieve low-latency model switching.

9. The method according to claim 1, characterized in that, The method also includes a model pre-deployment step: based on the drone flight trajectory and the preset mission plan, predict the artificial intelligence model configuration required in the near future, and preload the corresponding model variants into the video memory or pre-start the corresponding inference pipeline.

10. An edge-cloud collaborative artificial intelligence inference system for use in unmanned aerial vehicles (UAVs), characterized in that, include: The analysis unit is used to quantify the amount of feature information output by each intermediate layer of the artificial intelligence model and its contribution to the task. The determining unit is used to determine the model segmentation points of the artificial intelligence model based on the quantization results and in combination with the preset segmentation criteria; A construction unit is used to construct multiple candidate execution schemes, including a model segmentation execution scheme generated based on the model segmentation points. The acquisition unit is used to collect multi-dimensional operational status data of the UAV, including communication link status, onboard computing resource status, energy reserve status, and mission constraints. The calculation unit is used to calculate the evaluation score of each candidate execution scheme based on the multidimensional running status data; The control unit is used to select the optimal target execution plan based on the evaluation score, control the drone to run the artificial intelligence model according to the target execution plan, and complete the edge-cloud collaborative artificial intelligence inference.