An edge-end collaborative inference method and system for air-ground collaborative environment monitoring
By constructing a comprehensive performance index function and an optimization model using the Q-learning algorithm, the optimal splitting point and quantization bit depth are determined. This solves the problems of lightweighting and privacy protection of intermediate features in the air-ground collaborative environmental monitoring system, improves processing efficiency and privacy protection, and reduces data transmission latency and leakage risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG PROVINCIAL ECOLOGICAL ENVIRONMENT MONITORING CENT
- Filing Date
- 2026-02-14
- Publication Date
- 2026-06-09
Smart Images
Figure CN122174983A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of edge computing technology, and in particular to an edge-to-edge collaborative reasoning method and system for air-ground collaborative environmental monitoring. Background Technology
[0002] In remote air-ground collaborative environmental monitoring systems, the edge-edge collaborative inference architecture, comprised of UAVs and other end-user devices and ground-based edge servers, is a crucial support for achieving wide-area ecological perception and real-time disaster assessment. Its core mechanism involves rationally decomposing a deep learning model into end-user sub-models and edge sub-models. The UAV performs front-end inference and uploads intermediate features to the edge server to jointly complete the monitoring task. Therefore, determining the target model decomposition points is particularly important. If the decomposition points are too shallow, the original features transmitted by the UAV are easily stolen or reconstructed through inversion attacks, leading to the leakage of sensitive geographic images and ecological data. If the decomposition points are too deep, it increases the computational burden on the end-user devices, significantly increasing UAV energy consumption and affecting the real-time response of the task.
[0003] Existing splitting strategies generally lack lightweight and privacy-preserving processing for intermediate features. Directly transmitting high-dimensional features inferred from the UAV front end not only increases communication overhead and transmission latency, but also further reduces transmission efficiency under low bandwidth conditions, resulting in privacy issues and low processing efficiency for monitoring tasks. Summary of the Invention
[0004] This invention provides an edge-to-edge collaborative reasoning method and system for air-ground collaborative environmental monitoring, which can solve the problems of lack of lightweight and privacy processing of intermediate features in the prior art, resulting in low privacy and processing efficiency of monitoring tasks.
[0005] One embodiment of the present invention provides an edge-side collaborative reasoning method for air-ground collaborative environmental monitoring, applicable to UAV edge devices, including: It acquires its own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; among them, each preset inference parameter combination contains a split point, a quantization bit depth, and a dimensionality reduction dimension. Based on the above hardware parameters, server parameters, network parameters, and several monitoring data samples, energy consumption indicators, privacy indicators, and latency indicators are constructed. Based on the above energy consumption indicators, privacy indicators, latency indicators, and preset environmental monitoring scenario weights, a comprehensive performance indicator function is constructed. Based on the above hardware parameters and comprehensive performance index function, with the goal of maximizing comprehensive performance, an optimization model and corresponding constraints are constructed. Under the above constraints, the above optimization model is solved to obtain the optimal combination of inference parameters from several preset combinations of inference parameters. The optimal quantization bit depth and optimal split point in the optimal combination of inference parameters are transmitted to the ground edge server, so that the ground edge server can perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, optimal split point, and its own deployed model to be split.
[0006] Furthermore, the construction of the aforementioned latency metrics includes: Obtain the edge inference computation cost at each splitting point and the total computation cost of the above-mentioned model to be split; Based on the maximum computing power and inference computation amount on the edge side in the above hardware parameters, the edge computing latency of each split point is calculated. Based on the above network parameters and the fixed transmit power at the end of the hardware parameters, the characteristic transmission delay of each split point is calculated. Based on the server's maximum computing power, edge inference computing load, and total computing load in the above server parameters, the edge computing latency of each split point is calculated. Based on the aforementioned end-side computation latency, feature transmission latency, and edge computation latency, a latency index is constructed.
[0007] Furthermore, the construction of the aforementioned energy consumption indicators includes: Based on the edge computing power and edge computing latency in the above hardware parameters, the edge inference power consumption is calculated. The reciprocal of the above end-side inference energy consumption is used as the above energy consumption index.
[0008] Furthermore, the construction of the aforementioned privacy metrics includes: Obtain the reconstructed data samples corresponding to each monitoring data sample; wherein, the above reconstructed data samples are obtained by inversion based on the intermediate data features output by the model layer where each split point is located according to the corresponding monitoring data sample; Based on the above monitoring data samples and reconstructed data, privacy indicators were constructed.
[0009] Furthermore, the objective function of the above optimization model is: In the formula, The split point is indicated as A comprehensive performance index function with dimensionality reduction dimension d and quantization bit depth b. This indicates the weight of the preset environmental monitoring scenario corresponding to the energy consumption index. The split point is indicated as The energy consumption index when the dimensionality reduction dimension is d and the quantization bit number is b. This indicates the preset environmental monitoring scenario weights corresponding to the privacy metrics. The split point is indicated as Privacy metrics when the dimensionality reduction dimension is d and the quantization bit depth is b. This indicates the preset environmental monitoring scenario weights corresponding to the latency index. The split point is indicated as The latency metric when the dimensionality reduction dimension is d and the quantization bit number is b.
[0010] Furthermore, the above constraints are as follows: In the formula, The split point is indicated as End-side energy consumption at that time This refers to the end-side calculated power in the hardware parameters. This indicates the maximum power consumption limit in the hardware parameters. The split point is indicated as The time consumed by the compressed end-side calculation.
[0011] Furthermore, the optimization model is solved under the aforementioned constraints to obtain the optimal combination of inference parameters from several preset combinations of inference parameters, including: Each preset inference parameter combination is used as the action of the Q-learning algorithm, the currently selected preset inference combination is used as the state of the Q-learning algorithm, the above comprehensive performance index function is used as the reward function for each step in the Q-learning algorithm, and the cumulative number of steps of the above Q-learning algorithm in each iteration is set to n; where n is a positive integer. Based on the Q-learning algorithm, with the goal of maximizing the overall performance, each preset inference combination is solved iteratively, and the Q value after each iteration is calculated until all the current Q values converge, thus obtaining the final Q value of each preset inference combination. The preset inference combination corresponding to the largest Q value among all the final Q values that satisfy the above constraints is taken as the optimal inference parameter combination.
[0012] Furthermore, the optimal quantization bit depth and optimal split point in the optimal combination of inference parameters are transmitted to the ground edge server, enabling the ground edge server to perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, optimal split point, and its own deployed model to be split, including: Based on the above-mentioned optimal splitting point, the above-mentioned model to be split is deployed at the above-mentioned UAV end-side device to obtain a first sub-model; wherein, the above-mentioned first sub-model is the model part in the above-mentioned model to be split that is before the model layer where the above-mentioned optimal splitting point is located, and includes the model part within the model layer where the above-mentioned optimal splitting point is located. The first sub-model processes the current environmental monitoring data it collects to obtain the first output feature. Then, it performs dimensionality reduction and quantization on the first output feature according to the optimal dimensionality reduction dimension in the optimal combination of the optimal quantization bits and the optimal inference parameters, and transmits it to the ground edge server. The optimal splitting point is transmitted to the ground edge server so that the ground edge server can split the model to be split based on the optimal splitting point to obtain a second sub-model; wherein the second sub-model is the remaining model part after the model layer where the optimal splitting point is located. The first output feature after dimensionality reduction and quantization is dequantized and dimensionality increased according to the optimal quantization bit depth. The first output feature after dequantization and dimensionality increase is then processed according to the second sub-model to obtain the output result of the model to be split. The output result is then fed back to the UAV end-side device.
[0013] Furthermore, before performing data processing on the first output feature after dequantization and dimensionality upscaling based on the aforementioned second sub-model, the following steps are also included: After obtaining several monitoring data samples and processing them under the corresponding first sub-model, the first output feature sample is obtained, and the second output feature sample is obtained by sequentially performing dimensionality reduction, quantization, dequantization and dimensionality increase on the first output feature sample. Based on the first output feature sample and the second output feature sample, a preset linear regression model is trained with the goal of minimizing the mean square error between the first output feature sample and the corresponding second output feature sample. Based on the aforementioned preset linear regression model, error correction is performed on the first output feature after dequantization.
[0014] Based on the above method embodiments, the present invention provides corresponding system embodiments; This invention provides an edge-to-edge collaborative reasoning system for air-to-ground collaborative environmental monitoring, comprising: Unmanned aerial vehicle (UAV) edge devices and ground-based edge servers; The aforementioned UAV end-side device is used to acquire its own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; wherein, each preset inference parameter combination contains a split point, a quantization bit depth, and a dimensionality reduction dimension. Based on the above hardware parameters, server parameters, network parameters, and several monitoring data samples, energy consumption indicators, privacy indicators, and latency indicators are constructed. Based on the above energy consumption indicators, privacy indicators, latency indicators, and preset environmental monitoring scenario weights, a comprehensive performance indicator function is constructed. Based on the above hardware parameters and comprehensive performance index function, with the goal of maximizing comprehensive performance, an optimization model and corresponding constraints are constructed. Under the above constraints, the above optimization model is solved to obtain the optimal combination of inference parameters from several preset combinations of inference parameters. The optimal quantization bit depth and optimal split point in the optimal combination of inference parameters are transmitted to the ground edge server. The aforementioned ground edge server is used to perform collaborative inference with the aforementioned UAV end-side device based on the optimal quantization bit depth, optimal splitting point, and the model to be split deployed by itself.
[0015] The embodiments of the present invention have the following beneficial effects: This invention provides an edge-side collaborative inference method and system for air-ground collaborative environment monitoring. The method is used on a UAV edge device and includes: acquiring its own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; constructing energy consumption indicators, privacy indicators, and latency indicators; and then constructing a comprehensive performance indicator function. Subsequently, based on the hardware parameters and the comprehensive performance indicator function, with the goal of maximizing comprehensive performance, an optimization model and corresponding constraints are constructed. The optimal inference parameter combination is obtained by solving the problem under the constraints. Finally, the optimal quantization bit depth and optimal splitting point from the optimal inference parameter combination are transmitted to a ground edge server, enabling the ground edge server to perform collaborative inference with the UAV edge device based on the optimal quantization bit depth, optimal splitting point, and its own deployed splitting model. Therefore, this invention solves the optimization model with the goal of maximizing the comprehensive performance indicator containing privacy indicators, obtaining the optimal splitting point, optimal quantization bit depth, and optimal dimensionality reduction dimension with privacy properties. Finally, collaborative inference is achieved based on the obtained optimal inference parameter combination, greatly improving the processing efficiency and privacy of the monitoring task during the entire edge-side collaborative inference process. Attached Figure Description
[0016] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0017] Figure 1 This is a flowchart illustrating an edge-to-edge collaborative reasoning method for air-to-ground collaborative environmental monitoring, provided by an embodiment of the present invention.
[0018] Figure 2 This is a schematic diagram of the PCA dimensionality reduction process provided in an embodiment of the present invention.
[0019] Figure 3 This is a schematic diagram of an integer quantization process provided in an embodiment of the present invention.
[0020] Figure 4 This is a schematic diagram comparing the performance indicators of different strategies provided in an embodiment of the present invention.
[0021] Figure 5 This is a schematic diagram of the structure of an edge-to-edge collaborative reasoning system for air-ground collaborative environmental monitoring, provided by an embodiment of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.
[0024] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.
[0025] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0026] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.
[0027] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).
[0028] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.
[0029] See Figure 1 To address the lack of lightweight and privacy-preserving processing for intermediate features in existing technologies, which leads to privacy concerns and low processing efficiency in monitoring tasks, this invention provides an edge-side collaborative inference method for air-to-ground collaborative environment monitoring. This method is applicable to UAV edge devices and includes: Step S101: Obtain your own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; wherein, each preset inference parameter combination contains a split point, a quantization bit depth, and a dimensionality reduction dimension. Specifically, the hardware parameters include: fixed transmit power at the terminal side, maximum computing power at the terminal side, maximum power consumption limit, and computing power at the terminal side; the server parameters include: maximum computing power of the server; and the network parameters include: uplink bandwidth, channel gain, and noise power.
[0030] Specifically, a target model suitable for air-ground collaborative environmental monitoring is pre-determined as the model to be split. Then, based on the hierarchical structure of the model, semantically separable and computationally separable layers are selected. Layers that the UAV end-side device cannot support are excluded by combining the hardware parameters of the UAV end-side device. Subsequently, the set of splitting points of the model to be split is determined, which can be represented as: In the formula, L represents the set of splitting points. This represents a set of determined splitting points, where n' represents the total number of splitting points.
[0031] Specifically, the aforementioned quantization bit depth is used to quantize the features, and the aforementioned dimensionality reduction dimension is used to reduce the dimensionality of the features.
[0032] Specifically, the aforementioned monitoring data samples consist of pre-collected base map samples and historical change samples. Each monitoring task scenario category has its own corresponding monitoring data samples and its own corresponding preset inference parameter combinations.
[0033] Step S102: Based on the above hardware parameters, server parameters, network parameters and several monitoring data samples, construct energy consumption indicators, privacy indicators and latency indicators, and construct a comprehensive performance indicator function based on the above energy consumption indicators, privacy indicators, latency indicators and preset environmental monitoring scenario weights. Specifically, the comprehensive performance index function built based on energy consumption, privacy, and latency indicators can achieve multi-dimensional performance evaluation.
[0034] In a preferred embodiment, the construction of the above-mentioned latency metric includes: Obtain the edge inference computation cost at each splitting point and the total computation cost of the above-mentioned model to be split; Based on the maximum computing power and inference computation amount on the edge side in the above hardware parameters, the edge computing latency of each split point is calculated. Specifically, the end-side computational delay is obtained using the following formula: In the formula, Indicates the split point The corresponding end-side computation delay, This indicates that the i-th layer in the model to be split is the split point. Indicates the split point The corresponding edge-side inference computation amount, This indicates the maximum computing power at the edge.
[0035] Based on the above network parameters and the fixed transmit power at the end of the hardware parameters, the characteristic transmission delay of each split point is calculated. Specifically, the characteristic transmission delay is calculated using the following formula: In the formula, Indicates the uplink transmission rate. Indicates uplink bandwidth. Indicates a fixed transmit power at the end. Indicates channel gain. Indicates noise power. Indicates the split point The size of the intermediate feature data inferred and output by the UAV end-side equipment.
[0036] Based on the server's maximum computing power, edge inference computing load, and total computing load in the above server parameters, the edge computing latency of each split point is calculated. Specifically, the edge computing latency is calculated using the following formula: In the formula, Indicates the split point The corresponding edge computing latency, This represents the total computational load. This indicates the server's maximum computing power.
[0037] Based on the aforementioned end-side computation latency, feature transmission latency, and edge computation latency, a latency index is constructed.
[0038] Specifically, the monitoring data samples are first subjected to dimensionality reduction and feature quantization to construct a compressed feature transmission latency index. Then, based on the compressed feature transmission latency index, the end-side computation latency, and the sum of the edge computation latency, the final latency index is obtained. For dimensionality reduction, PCA dimensionality reduction is used in this invention, and the PCA dimensionality reduction flowchart is shown below. Figure 2 As shown, based on the monitoring data samples under the current monitoring task scenario category, the split points are... The original feature matrix corresponding to the output sample feature data is Let m be the number of monitoring data samples, then the mean of its feature dimension can be expressed as: In the formula, This represents the mean of the feature dimension k. Indicates the number of monitoring data samples. This represents the k-th feature dimension of the i-th monitoring data sample. This represents the original feature dimension of the monitoring data sample, for each monitoring data sample. .
[0039] Subsequently, subtracting the expected matrix composed of the mean values of the aforementioned feature dimensions from the original feature matrix yields the centered matrix: In the formula, Represents a centered matrix. Let represent the above expectation matrix.
[0040] Then, the covariance matrix is calculated based on the centered matrix: In the formula, Represents the covariance matrix. This represents the transpose of the feature matrix after feature centering.
[0041] Perform eigenvalue decomposition on the above covariance matrix: In the formula, This represents the k-th eigenvector of the covariance matrix. The eigenvalues corresponding to the k-th eigenvector are represented as follows: The dimensions of the covariance matrix mentioned above are: .
[0042] Then, the cumulative variance percentage is calculated based on these eigenvalues: In the formula, This represents the preset variance threshold, which can be 0.9. d represents the dimension reduction dimension, which is a dimension reduction dimension to be determined.
[0043] Then, the optimal dimensionality reduction dimension can be calculated: In the formula, The above formula represents the final determined optimal dimensionality reduction dimension. It expresses the goal of finding the minimum value of 'd' such that the variance of the first 'd' principal components accounts for ≥ 1 / 2 the proportion of the original total variance. Then, this smallest d (i.e. () is the final optimal dimension reduction dimension.
[0044] Finally, based on the optimal dimensionality reduction dimension, the original feature matrix is reduced in dimensionality to obtain the dimensionality-reduced feature matrix: In the formula, This represents the feature matrix after dimensionality reduction. Indicates the preceding The projection matrix consists of eigenvectors.
[0045] Next, the reduced-dimensionality feature matrix needs to be integer quantized. A schematic diagram of the integer quantization process is shown below. Figure 3 As shown, from Figure 3 As can be seen from this, firstly, the global extremum of the features is calculated based on the reduced-dimensional feature matrix: In the formula, This represents the minimum value of all elements in the feature matrix after dimensionality reduction. This represents the maximum value of all elements in the feature matrix after dimensionality reduction. This represents the element in the i-th row and j-th column of the feature matrix after dimensionality reduction.
[0046] Subsequently, b-bit integer quantization mapping is performed based on the global extremum of the feature: In the formula, This represents the element in the i-th row and j-th column of the feature matrix after dimensionality reduction and quantization with b bits. This indicates the rounding operation.
[0047] Subsequently, after integer quantization of the reduced feature dimensions, the total size of the compressed sample feature data and the corresponding b-bit quantized integer feature matrix can be represented as: In the formula, Indicates the split point And the number of quantization bits is the total size of the sample feature data corresponding to b. Indicates the split point The feature dimensions after dimensionality reduction by dimension reduction dimension d. This represents the integer feature matrix after b-bit quantization.
[0048] It is worth noting that the preset inference parameter combination includes pre-set split points, quantization bits, and dimensionality reduction dimensions. Since the preset inference parameter combination corresponds to a monitoring task scenario category, if the corresponding monitoring task scenario category requires high-dimensional data processing, the dimensionality reduction dimension is determined according to the aforementioned PCA dimensionality reduction algorithm. If the corresponding monitoring task scenario category requires low-dimensional data processing, the dimensionality reduction dimension d is set to the original dimension of the current environmental monitoring data collected by the UAV device itself during the actual inference process. Preferably, in this invention, data with more than 1024 dimensions is considered high-dimensional data; otherwise, it is considered low-dimensional data.
[0049] Specifically, in obtaining Then, latency metrics can be constructed based on this data: In the formula, Indicates the split point The corresponding latency metrics, Indicates the split point The corresponding compressed transmission delay.
[0050] In this preferred embodiment, a latency index is constructed based on the edge inference computation, total computation, and monitoring data samples at each split point.
[0051] In another preferred embodiment, the construction of the above-mentioned energy consumption index includes: Based on the edge computing power and edge computing latency in the above hardware parameters, the edge inference power consumption is calculated. The reciprocal of the above end-side inference energy consumption is used as the above energy consumption index.
[0052] Specifically, the energy consumption indicators mentioned above refer to the energy consumption after compression, and are therefore expressed by the following formula: In the formula, Indicates the split point The corresponding energy consumption indicators Indicates the split point The corresponding edge-side inference energy consumption, Indicates the calculated power at the end side. The split point is indicated as The time consumed by the compressed end-side calculation.
[0053] In this preferred embodiment, the energy consumption index is obtained based on the reciprocal of the edge-side inference energy consumption.
[0054] In another preferred embodiment, the construction of the above privacy metrics includes: Obtain the reconstructed data samples corresponding to each monitoring data sample; wherein, the above reconstructed data samples are obtained by inversion based on the intermediate data features output by the model layer where each split point is located according to the corresponding monitoring data sample; Specifically, the reconstructed data sample is obtained by inputting the corresponding monitoring data sample into the first sub-model after splitting at the corresponding split point, and then reconstructing the intermediate features.
[0055] Based on the above monitoring data samples and reconstructed data, privacy indicators were constructed.
[0056] Specifically, the privacy metric uses KL divergence to reconstruct differences. First, a first privacy metric is constructed based on the monitoring and reconstructed data. Then, based on the first privacy metric and the enhancement value of the KL divergence by compression, the aforementioned privacy metric is obtained. The calculation formula for the privacy metric is as follows: In the formula, Indicates the split point The corresponding first privacy metric, D, represents the set of all monitored data samples. This represents the value of the monitoring data sample x in the k-th dimension. This represents the reconstructed data corresponding to monitoring data sample x. This represents the improvement in KL divergence after data compression by reducing the dimension d and the number of quantization bits b.
[0057] In this preferred embodiment, a privacy metric is constructed based on the monitoring data sample and the corresponding reconstructed data sample.
[0058] Specifically, after obtaining the energy consumption, privacy, and latency metrics, the comprehensive performance index function is derived using the following formula: In the formula, The split point is indicated as A comprehensive performance index function with dimensionality reduction dimension d and quantization bit depth b. This indicates the weight of the preset environmental monitoring scenario corresponding to the energy consumption index. The split point is indicated as The energy consumption index when the dimensionality reduction dimension is d and the quantization bit number is b. This indicates the preset environmental monitoring scenario weights corresponding to the privacy metrics. The split point is indicated as Privacy metrics when the dimensionality reduction dimension is d and the quantization bit depth is b. This indicates the preset environmental monitoring scenario weights corresponding to the latency index. The split point is indicated as The latency metric when the dimensionality reduction dimension is d and the quantization bit number is b.
[0059] Step S103: Based on the above hardware parameters and comprehensive performance index function, construct an optimization model and corresponding constraints with the goal of maximizing comprehensive performance; In a preferred embodiment, the objective function of the above optimization model is: In the formula, The split point is indicated as A comprehensive performance index function with dimensionality reduction dimension d and quantization bit depth b. This indicates the weight of the preset environmental monitoring scenario corresponding to the energy consumption index. The split point is indicated as The energy consumption index when the dimensionality reduction dimension is d and the quantization bit number is b. This indicates the preset environmental monitoring scenario weights corresponding to the privacy metrics. The split point is indicated as Privacy metrics when the dimensionality reduction dimension is d and the quantization bit depth is b. This indicates the preset environmental monitoring scenario weights corresponding to the latency index. The split point is indicated as The latency metric when the dimensionality reduction dimension is d and the quantization bit number is b.
[0060] Specifically, the preset environmental monitoring scenario weights are determined based on the specific scenario. For example, privacy should be given priority in monitoring ecologically sensitive areas, and time delay should be given priority in disaster emergency monitoring, which means increasing the corresponding weight values.
[0061] Preferably, by adjusting the weights of the three scenarios mentioned above, different scenario requirements can be flexibly adapted. Furthermore, the multi-dimensional comprehensive performance indicator function can avoid performance imbalances caused by single-dimensional optimization, achieving a balance among the three.
[0062] In this preferred embodiment, the objective function of the optimization model is constructed by integrating the performance index function.
[0063] In another preferred embodiment, the above constraints are: In the formula, The split point is indicated as End-side energy consumption at that time This refers to the end-side calculated power in the hardware parameters. This indicates the maximum power consumption limit in the hardware parameters. The split point is indicated as The time delay is calculated at the compressed end side.
[0064] Specifically, the constraints of the optimization model are the end-side energy consumption constraints.
[0065] In this preferred embodiment, constraints are constructed based on the edge computing power, maximum energy consumption limit, and edge computing latency in the hardware parameters.
[0066] Step S104: Solve the above optimization model under the above constraints, and obtain the optimal combination of inference parameters from several preset combinations of inference parameters; Specifically, the optimal combination of inference parameters is obtained by using a multi-dimensional reward Q-learning algorithm adapted to edge-cooperative inference. This algorithm's core includes two improvements: first, it uses a comprehensive performance indicator function that fuses three dimensions as the reward signal, with specific weights... The settings are determined based on the specific scenario; secondly, there is an n-step multi-dimensional reward weighted accumulation mechanism. The specific value of the cumulative steps is set based on the communication latency between the drone and the ground edge server; a larger value is used when the communication latency is high. To improve convergence, based on the above two improvements, all preset inference parameter combinations are traversed to obtain the globally optimal solution that satisfies the constraints, i.e., the optimal inference parameter combination.
[0067] In a preferred embodiment, solving the optimization model under the aforementioned constraints to obtain the optimal inference parameter combination from a plurality of preset inference parameter combinations includes: Each preset inference parameter combination is used as the action of the Q-learning algorithm, the currently selected preset inference combination is used as the state of the Q-learning algorithm, the above comprehensive performance index function is used as the reward function for each step in the Q-learning algorithm, and the cumulative number of steps of the above Q-learning algorithm in each iteration is set to n; where n is a positive integer. Based on the Q-learning algorithm, with the goal of maximizing the overall performance, each preset inference combination is solved iteratively, and the Q value after each iteration is calculated until all the current Q values converge, thus obtaining the final Q value of each preset inference combination. The preset inference combination corresponding to the largest Q value among all the final Q values that satisfy the above constraints is taken as the optimal inference parameter combination.
[0068] Specifically, the currently selected preset inference parameter combination is defined as a state. State transition involves selecting another preset inference parameter combination as the next action from the state corresponding to the currently selected preset inference parameter combination, and switching to the new state. Subsequently, a preset reward decay coefficient is set to weaken the weight of long-term rewards, learning rate, and cumulative steps over n steps.
[0069] Subsequently, starting from the initial state, which is "no preset parameter combination selected", a preset inference parameter combination is selected as the first action. This preset inference parameter combination is determined based on relevant experience gained from previous inference tasks, completing the first state transition. Following this logic, n state transitions are executed, that is, n consecutive action selections are performed in the manner of "randomly selecting an action → switching states", forming an exploration path containing n state-action pairs (e.g., initial state → preset inference parameter combination A → preset inference parameter combination B → preset inference parameter combination C, n=3).
[0070] Subsequently, for each action in the path (i.e., each preset combination of inference parameters), its single-step reward value is calculated. In this invention, the value of the corresponding comprehensive performance index function is used as its single-step reward value. Then, based on the current exploration path, the reward value is calculated for each action. Step to the first Step corresponding The single-step reward value of a given combination of preset inference parameters is used to calculate the cumulative reward for n steps along this path: In the formula, This represents the cumulative reward over n steps. This represents the single-step reward value at step t. This represents the preset reward decay coefficient. This represents the single-step reward value at step t+1. Indicates the first The single-step reward value corresponding to each step.
[0071] Subsequently, for each state-action pair in the current exploration path, the Q-value is updated one by one: In the formula, Indicates the state at step t is The action is The corresponding state-action pair and the updated Q value. Indicates the learning rate. This represents the single-step reward value at step t. This represents the preset reward decay coefficient, with a value range of (0.8, 0.95). This represents the single-step reward value at step t+1. ' represents the discount factor of the Q-learning algorithm, Indicates the optimal target state value.
[0072] Preferably, the target state is optimal. The value is a preset value, determined based on the task requirements of air-ground collaborative environment detection, with the core basis being the task's emphasis on long-term rewards. Generally, if the scenario emphasizes the immediate benefits of real-time inference, a value of 0.8-0.9 is used; if the scenario emphasizes the long-term optimal combination of inference parameters, a value of 0.9-0.95 is used. In practice, the optimal value is selected by combining the Q-value convergence speed and collaborative inference accuracy in the algorithm. In this paper, a value range of (0.8, 0.95) is recommended.
[0073] This completes one iteration. The path exploration and Q-value update process is then repeated as described above: First, return to the initial state. Based on historical task experience or system-recommended parameter combinations, use this as the starting action for the initial iteration to generate a new exploration path (e.g., initial state → preset inference parameter combination B → preset inference parameter combination C → preset inference parameter combination A). Then, repeat the above process. After several iterations, the Q-value changes for all state-action pairs converge, indicating that the Q-value has fully absorbed the feedback from different paths and tends to stabilize. This yields the final Q-value for each preset inference parameter combination in the initial state (i.e., the final Q-values for "initial state → preset inference parameter combination B", "initial state → preset inference parameter combination A", etc.). Finally, the preset inference parameter combination corresponding to the maximum Q-value among all final Q-values that satisfies the constraints is taken as the optimal inference parameter combination.
[0074] The preferred design employs an adaptive dimensionality reduction and quantization compression strategy, achieving a dual improvement in efficiency and privacy: feature compression after splitting reduces the amount of transmitted data and consequently lowers transmission latency; at the same time, the entropy of compressed feature information is reduced, significantly decreasing reconstructability, which reduces the accuracy of attackers reconstructing the original data using inversion algorithms, thus balancing data transmission efficiency and privacy protection.
[0075] Preferably, for scenarios where focusing only on single-objective performance leads to an imbalance of multiple objectives, this invention achieves a globally optimal balance of time, energy consumption, and privacy through a multi-dimensional reward Q-learning algorithm, combined with an n-step reward weighted accumulation mechanism, under a balanced configuration of scenario weights, highlighting the technical advantages of the present invention.
[0076] In this preferred embodiment, the optimal combination of inference parameters is obtained based on the Q-learning algorithm.
[0077] Step S105: Transmit the optimal quantization bit depth and optimal split point from the optimal combination of inference parameters to the ground edge server, so that the ground edge server can perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, optimal split point, and the model to be split deployed by itself.
[0078] Specifically, the aforementioned model to be split is deployed simultaneously on both the drone's edge device and the ground-based edge server.
[0079] Preferably, in this invention, the process of determining the optimal combination of inference parameters is primarily executed autonomously by the UAV end-side device, but can also be switched to be executed by the host computer according to the actual scenario requirements (for example, when the UAV end-side device detects that it has acquired a lot of abnormal data, and the UAV end-side device alone cannot process the data in a timely and effective manner, the host computer is activated), with the core objective of efficiently processing abnormal data. Furthermore, during execution by the host computer, after determining the optimal combination of inference parameters, the optimal quantization bit depth and optimal splitting point are transmitted to the ground edge server. The optimal inference parameter combination is also transmitted to the UAV end-device. Since the UAV end-device and the ground edge server have already deployed the model to be split, upon receiving the optimal inference parameter combination, the UAV end-device splits the model to be split. Based on the first sub-model, it processes the current environmental monitoring data it has collected to obtain intermediate features. Then, based on the optimal quantization bit depth and optimal dimensionality reduction dimension, it performs dimensionality reduction and quantization processing on the intermediate features before transmitting them to the ground edge server. On the ground edge server, based on the aforementioned optimal splitting point, it splits the deployed model to be split to obtain the second sub-model. Finally, based on the optimal quantization bit depth, it dequantizes and upscales the dimensionality-reduced and quantized intermediate features. Based on the second sub-model, it processes the dequantized and upscaled intermediate features to obtain the final output result, which is then fed back to the UAV end-device, completing the collaborative inference.
[0080] In a preferred embodiment, the optimal quantization bit depth and optimal split point in the optimal inference parameter combination are transmitted to a ground edge server, so that the ground edge server can perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, optimal split point, and its own deployed model to be split, including: Based on the above-mentioned optimal splitting point, the above-mentioned model to be split is deployed at the above-mentioned UAV end-side device to obtain a first sub-model; wherein, the above-mentioned first sub-model is the model part in the above-mentioned model to be split that is before the model layer where the above-mentioned optimal splitting point is located, and includes the model part within the model layer where the above-mentioned optimal splitting point is located. Specifically, after splitting the model to be split based on the optimal split point, the model is divided into two parts, one of which is the model layer containing the optimal split point. Including the former The layer sub-model (i.e., the first sub-model mentioned above).
[0081] The first sub-model processes the current environmental monitoring data it collects to obtain the first output feature. Then, it performs dimensionality reduction and quantization on the first output feature according to the optimal dimensionality reduction dimension in the optimal combination of the optimal quantization bits and the optimal inference parameters, and transmits it to the ground edge server. Specifically, in the UAV terminal device, the environmental monitoring data collected by itself is input into the first sub-model to obtain the first output feature. Then, it is dimensionality-reduced and compressed according to the optimal quantization bit depth and the optimal dimensionality reduction dimension, and then transmitted to the ground edge server for further processing.
[0082] The optimal splitting point is transmitted to the ground edge server so that the ground edge server can split the model to be split based on the optimal splitting point to obtain a second sub-model; wherein the second sub-model is the remaining model part after the model layer where the optimal splitting point is located. Specifically, after splitting the model to be split based on the optimal splitting point, the model parts other than the first sub-model mentioned above are taken as the second sub-model mentioned above.
[0083] The first output feature after dimensionality reduction and quantization is dequantized and dimensionality increased according to the optimal quantization bit depth. The first output feature after dequantization and dimensionality increase is then processed according to the second sub-model to obtain the output result of the model to be split. The output result is then fed back to the UAV end-side device.
[0084] Specifically, the ground edge server dequantizes the received dimensionality-reduced and compressed first output feature according to the optimal quantization bit depth, then upscales it and inputs it into the second sub-model for data processing to obtain the final model output result, which is then fed back to the UAV end-side device to achieve collaborative reasoning.
[0085] In this preferred embodiment, collaborative reasoning is achieved by combining the feature data processing operations of the first sub-model and the second sub-model with the UAV end-side device and the ground edge server, respectively.
[0086] In another preferred embodiment, before performing data processing on the dequantized and up-dimensionalized first output features according to the second sub-model described above, the method further includes: After obtaining several monitoring data samples and processing them under the corresponding first sub-model, the first output feature sample is obtained, and the second output feature sample is obtained by sequentially performing dimensionality reduction, quantization, dequantization and dimensionality increase on the first output feature sample. Based on the first output feature sample and the second output feature sample, a preset linear regression model is trained with the goal of minimizing the mean square error between the first output feature sample and the corresponding second output feature sample. Based on the aforementioned preset linear regression model, error correction is performed on the first output feature after dequantization.
[0087] Specifically, the results obtained by the ground edge server after dequantization and dimensionality upscaling may contain errors. Therefore, based on the monitoring data samples, a pre-trained linear regression model is developed to minimize the mean square error between a first output feature sample without any quantization or dimensionality reduction and a second output feature sample after dimensionality reduction, quantization, dequantization, and dimensionality upscaling. This model corrects the errors in the dequantized features from the ground edge server, ensuring that the inference accuracy meets the requirements. The inference accuracy index is as follows: In the formula, Indicates the split point The corresponding inference precision, where K represents the number of target categories. This represents the average precision of the k-th class. This indicates the preset inference accuracy threshold.
[0088] Specifically, the above target categories are the specific monitoring object categories that need to be inferred and identified in the air-ground collaborative environmental monitoring task.
[0089] Specifically, the training scheme is as follows: First, the training data consists of several monitoring data samples directly processed by the uncompressed first output feature samples from the corresponding first sub-model, and the second output feature samples after dimensionality reduction, quantization, dequantization, and dimensionality increase. Second, the training objective is to minimize the mean squared error (MSE) of the feature dimension. Finally, the model structure of the linear regression model is preset to have an input layer dimension equal to the feature dimension after PCA dimensionality reduction, an output layer dimension consistent with the input layer dimension, no hidden layers, and weights optimized through gradient descent. This correction scheme ensures that the inference accuracy is not less than a preset threshold. The corrected features are then input into the second sub-model to complete the inference, and finally, the inference results are fed back to the UAV's edge device.
[0090] In this preferred embodiment, the first output feature after dequantization is corrected based on a preset linear regression model trained on monitoring data samples.
[0091] This paper illustrates a simulation verification of an edge-device collaborative inference scenario based on UAV ecological monitoring. Under the constraints of limited computing power, energy sensitivity, and privacy protection requirements of UAV devices, the inference task is image target detection. The experiment uses a subset of Mini-ImageNet, selecting relevant images for air-to-ground collaborative environmental monitoring as the dataset, containing 100 image classes. The image resolution is uniformly adjusted to 640×360 to suit the deployment requirements of the edge-device collaborative inference model. Finally, a comparison diagram of the performance metrics of different strategies based on the inference results is shown below. Figure 4As shown, there are four strategies: optimizing only latency, optimizing only energy consumption, optimizing only privacy, and the solution proposed in this invention. From the perspective of latency cost alone, the strategies of optimizing only latency and optimizing only energy consumption exhibit the best latency performance. The solution proposed in this invention maintains superior real-time performance under multi-objective balance, meeting the real-time inference requirements of UAV ecosystem monitoring. From the perspective of total cost as a comprehensive evaluation indicator, the strategy of optimizing only privacy has the highest overall cost, while the solution proposed in this invention achieves globally optimal overall performance, effectively solving the problem of multi-objective imbalance in edge-edge collaborative inference.
[0092] Based on the above method embodiments, the present invention provides corresponding system embodiments.
[0093] like Figure 5 As shown, one embodiment of the present invention provides an edge-to-edge collaborative reasoning system for air-to-ground collaborative environmental monitoring, comprising: Unmanned aerial vehicle (UAV) edge devices and ground-based edge servers; The aforementioned UAV end-side device is used to acquire its own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; wherein, each preset inference parameter combination contains a split point, a quantization bit depth, and a dimensionality reduction dimension. Based on the above hardware parameters, server parameters, network parameters, and several monitoring data samples, energy consumption indicators, privacy indicators, and latency indicators are constructed. Based on the above energy consumption indicators, privacy indicators, latency indicators, and preset environmental monitoring scenario weights, a comprehensive performance indicator function is constructed. Based on the above hardware parameters and comprehensive performance index function, with the goal of maximizing comprehensive performance, an optimization model and corresponding constraints are constructed. Under the above constraints, the above optimization model is solved to obtain the optimal combination of inference parameters from several preset combinations of inference parameters. The optimal quantization bit depth and optimal split point in the optimal combination of inference parameters are transmitted to the ground edge server. The aforementioned ground edge server is used to perform collaborative inference with the aforementioned UAV end-side device based on the optimal quantization bit depth, optimal splitting point, and the model to be split deployed by itself.
[0094] It should be noted that the system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort. The above schematic diagrams are merely examples of an edge-end collaborative inference system for air-to-ground collaborative environmental monitoring and do not constitute a limitation on an edge-end collaborative inference system for air-to-ground collaborative environmental monitoring. It may include more or fewer components than illustrated, or combine certain components, or use different components.
Claims
1. An edge-to-edge collaborative reasoning method for air-to-ground collaborative environmental monitoring, characterized in that, Suitable for drone end-side devices, including: It acquires its own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; among them, each preset inference parameter combination contains a split point, a quantization bit depth, and a dimensionality reduction dimension. Based on the hardware parameters, server parameters, network parameters, and several monitoring data samples, energy consumption indicators, privacy indicators, and latency indicators are constructed. Based on the energy consumption indicators, privacy indicators, latency indicators, and preset environmental monitoring scenario weights, a comprehensive performance indicator function is constructed. Based on the hardware parameters and the comprehensive performance index function, with the goal of maximizing the comprehensive performance, an optimization model and corresponding constraints are constructed. The optimization model is solved under the constraints to obtain the optimal combination of inference parameters from a number of preset combinations of inference parameters. The optimal quantization bit depth and optimal split point in the optimal combination of inference parameters are transmitted to the ground edge server, so that the ground edge server can perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, optimal split point, and its own deployed model to be split.
2. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 1, characterized in that, The construction of the latency metric includes: Obtain the edge-side inference computation amount at each splitting point and the total computation amount of the model to be split; Based on the maximum computing power and inference computation amount on the edge side in the hardware parameters, the edge computing latency of each split point is calculated. Based on the network parameters and the fixed transmit power at the end of the hardware parameters, the characteristic transmission delay of each split point is calculated. Based on the server's maximum computing power, edge inference computing load, and total computing load in the server parameters, the edge computing latency of each split point is calculated. A latency index is constructed based on the end-side computation latency, feature transmission latency, and edge computation latency.
3. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 2, characterized in that, The construction of the energy consumption index includes: The edge computing power and edge computing latency in the hardware parameters are used to calculate the edge inference power consumption. The reciprocal of the end-side inference energy consumption is used as the energy consumption index.
4. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 3, characterized in that, The construction of the privacy metrics includes: Obtain the reconstructed data sample corresponding to each monitoring data sample; wherein, the reconstructed data sample is obtained by inversion based on the intermediate data features output by the model layer where each split point is located according to the corresponding monitoring data sample; Privacy metrics are constructed based on the monitoring data samples and reconstructed data.
5. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 3, characterized in that, The objective function of the optimization model is: In the formula, The split point is indicated as A comprehensive performance index function with dimensionality reduction dimension d and quantization bit depth b. This indicates the weight of the preset environmental monitoring scenario corresponding to the energy consumption index. The split point is indicated as The energy consumption index when the dimensionality reduction dimension is d and the quantization bit number is b. This indicates the preset environmental monitoring scenario weights corresponding to the privacy metrics. The split point is indicated as Privacy metrics when the dimensionality reduction dimension is d and the quantization bit depth is b. This indicates the preset environmental monitoring scenario weights corresponding to the latency index. The split point is indicated as The latency metric when the dimensionality reduction dimension is d and the quantization bit number is b.
6. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 5, characterized in that, The constraints are as follows: In the formula, The split point is indicated as End-side energy consumption at that time This refers to the end-side calculated power in the hardware parameters. This indicates the maximum power consumption limit in the hardware parameters. The split point is indicated as The time delay is calculated at the compressed end side.
7. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 6, characterized in that, Solving the optimization model under the constraints to obtain the optimal inference parameter combination from several preset inference parameter combinations includes: Each preset inference parameter combination is used as the action of the Q-learning algorithm, the currently selected preset inference combination is used as the state of the Q-learning algorithm, the comprehensive performance index function is used as the reward function for each step in the Q-learning algorithm, and the cumulative number of steps of the Q-learning algorithm in each iteration is set to n; where n is a positive integer. Based on the Q-learning algorithm, with the goal of maximizing the overall performance, each preset inference combination is solved iteratively, and the Q value after each iteration is calculated until all the current Q values converge, thus obtaining the final Q value of each preset inference combination. The preset inference combination corresponding to the largest Q value among all final Q values that satisfy the constraints is taken as the optimal inference parameter combination.
8. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 7, characterized in that, The step of transmitting the optimal quantization bit depth and optimal split point from the optimal inference parameter combination to the ground edge server, so that the ground edge server can perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, optimal split point, and its own deployed model to be split, includes: The model to be split is deployed at the UAV terminal device according to the optimal split point to obtain a first sub-model; wherein, the first sub-model is the model part of the model to be split that is before the model layer where the optimal split point is located and includes the model part within the model layer where the optimal split point is located. The first sub-model processes the current environmental monitoring data it collects to obtain the first output feature. Then, it performs dimensionality reduction and quantization processing on the first output feature according to the optimal dimensionality reduction dimension in the optimal combination of the optimal quantization bit and the optimal inference parameter. The result is then transmitted to the ground edge server. The optimal splitting point is transmitted to the ground edge server, so that the ground edge server splits the model to be split based on the optimal splitting point to obtain a second sub-model; wherein, the second sub-model is the remaining model part after the model layer where the optimal splitting point is located; The first output feature after dimensionality reduction and quantization is dequantized and dimensionality increased according to the optimal quantization bit depth. The first output feature after dequantization and dimensionality increase is then processed according to the second sub-model to obtain the output result of the model to be split. The output result is then fed back to the UAV end-side device.
9. The edge-end collaborative reasoning method for air-ground collaborative environmental monitoring according to claim 8, characterized in that, Before performing data processing on the first output feature after dequantization and dimensionality upscaling based on the second sub-model, the process also includes: After acquiring several monitoring data samples and processing them under the corresponding first sub-model, a first output feature sample is obtained, and a second output feature sample is obtained by sequentially performing dimensionality reduction, quantization, dequantization and dimensionality increase on the first output feature sample. Based on the first output feature sample and the second output feature sample, a preset linear regression model is trained with the goal of minimizing the mean square error between the first output feature sample and the corresponding second output feature sample. Based on the preset linear regression model, error correction is performed on the first output feature after dequantization.
10. An edge-to-edge collaborative reasoning system for air-to-ground collaborative environmental monitoring, characterized in that, include: Unmanned aerial vehicle (UAV) edge devices and ground-based edge servers; The UAV end-side device is used to acquire its own hardware parameters, server parameters, network parameters, several preset inference parameter combinations under the current monitoring task scenario category, and several monitoring data samples; wherein, each preset inference parameter combination contains a split point, a quantization bit depth, and a dimensionality reduction dimension. Based on the hardware parameters, server parameters, network parameters, and several monitoring data samples, energy consumption indicators, privacy indicators, and latency indicators are constructed. Based on the energy consumption indicators, privacy indicators, latency indicators, and preset environmental monitoring scenario weights, a comprehensive performance indicator function is constructed. Based on the hardware parameters and the comprehensive performance index function, with the goal of maximizing the comprehensive performance, an optimization model and corresponding constraints are constructed. The optimization model is solved under the constraints to obtain the optimal combination of inference parameters from a number of preset combinations of inference parameters. The optimal quantization bit depth and optimal split point in the optimal combination of inference parameters are transmitted to the ground edge server. The ground edge server is used to perform collaborative inference with the UAV end-side device based on the optimal quantization bit depth, the optimal splitting point, and the model to be split deployed by itself.