CN-122027652-A - Cloud edge end cooperative architecture and task unloading method for apron intelligent monitoring system

CN122027652ACN 122027652 ACN122027652 ACN 122027652ACN-122027652-A

Abstract

The invention discloses a cloud edge end cooperative framework and a task unloading method for an apron intelligent monitoring system, and relates to the technical field of cloud edge end cooperative computing. The task unloading method of the system under the cloud edge collaborative architecture comprises the steps of constructing a task model, a time delay model, an energy consumption model and an accuracy model of the system to form a multi-objective optimization problem of time delay, energy consumption and accuracy, modeling the optimization problem into a Markov decision process, designing a state space, an action space and a reward function required by deep reinforcement learning, designing a task unloading method based on multi-agent deep reinforcement learning, and finding an optimal task unloading strategy of the system. The method and the system can dynamically carry out scientific and reasonable allocation and collaborative scheduling on the computing tasks among the terminal, the edge and the cloud so as to realize comprehensive optimization of time delay, energy consumption and accuracy and improve the overall performance of the system.

Inventors

DING MENG
HUANG XINGYU
XU YIMING
HAO SHUYI
WANG JIAJUN
LIU HAODONG
ZHANG BO

Assignees

南京航空航天大学

Dates

Publication Date: 20260512
Application Date: 20260206

Claims (8)

1. The cloud edge end cooperative architecture for the apron intelligent monitoring system is characterized by comprising a terminal equipment layer, an edge server layer and a cloud server layer, wherein all the layers are connected through a wired or wireless network to form a distributed intelligent sensing and reasoning system; The method comprises the steps of deploying a plurality of high-definition monitoring cameras on a terminal equipment layer, uniformly covering a main operation area of an apron, and acquiring dynamic video streams of an airplane, a ground service vehicle and personnel in real time, wherein each camera is used as an independent data source, and uploading original video data to a corresponding edge server through a local area network; the edge server is provided with a high-performance GPU, can run a lightweight multi-mode large model to perform real-time video understanding tasks, comprises interaction behaviors, state identification and abnormal judgment, and meanwhile, the edge nodes are interconnected through a high-speed local area network to support cross-node task migration and load balancing; In the cloud server layer, a high-performance cloud server is deployed in a data center and is used for processing complex scenes which cannot be determined by edge nodes, providing higher-precision semantic understanding and decision support, and in addition, the cloud is also responsible for model training, version updating, historical data analysis and system management and monitoring.
2. The task offloading method of a cloud edge end collaborative architecture for an apron-oriented intelligent monitoring system of claim 1, comprising the steps of: The method comprises the steps that 1, original video stream data collected by a high-definition monitoring camera deployed on an airport parking apron are firstly transmitted to a corresponding edge server, the edge server is called a local edge server, a preprocessing module in the edge server detects video frame by using a target detection model, a frame extraction rate is determined according to a detection result, resolution adjustment is carried out on extracted image frames, then continuous image frames form a task, and an unloading decision module is waited to distribute an execution node; the unloading decision-making module selects proper calculation nodes according to the data size of the task, the semantic complexity and the real-time state of the current system, wherein the specific selection is determined by an intelligent body based on deep reinforcement learning, and the intelligent body does not adopt fixed rules, but continuously learns and optimizes the strategy by interacting with the environment so as to obtain dynamic balance among time delay, system energy consumption and accuracy; And 3, after determining the execution node, transmitting the task to corresponding equipment to perform reasoning of the visual language large model, transmitting the output reasoning and analysis result to an apron control personnel for reference, and uploading the processing result of the edge server to a cloud for archiving and storing data.
3. The task offloading method of a cloud-edge co-architecture for an apron-oriented intelligent monitoring system of claim 2, wherein the preprocessing module comprises: The preprocessing module specifically needs to complete three tasks, namely, dynamically adjusting the resolution of an image, downsampling high-resolution input to low resolution as required, realizing intelligent frame extraction through a frame rate controller, judging scene liveness and complexity based on a target detection model, reducing sampling frequency when no scene key target appears, maintaining higher frame rate in a contour dynamic working stage, ensuring the integrity of a behavior time sequence, judging the semantic complexity of the task according to the result of target detection, specifically, considering the task as a simple task if no airplane or operation and maintenance vehicle is in a detected image frame, considering the task as a general task if only the airplane or operation and maintenance vehicle is in the detected image frame, and considering the task as a difficult task if the airplane or operation and maintenance vehicle is in the detected image frame.
4. The task offloading method of a cloud edge end collaborative architecture for an apron-oriented intelligent monitoring system according to claim 2, wherein the construction of the offloading decision module comprises the steps of: S1, constructing a system model, wherein the system model comprises a task model, a time delay model, an energy consumption model and an accuracy model, so as to form a multi-objective optimization problem of time delay, energy consumption and accuracy; S2, modeling an optimization problem into a Markov decision process, and designing a state space, an action space and a reward function required by deep reinforcement learning; S3, designing a task unloading method based on multi-agent deep reinforcement learning, and finding out an optimal task unloading strategy of the system; And S4, training the deep reinforcement learning intelligent agent, and disposing the deep reinforcement learning intelligent agent on each edge server to carry out unloading decision.
5. The task offloading method of a cloud-edge co-architecture for an apron-oriented intelligent monitoring system of claim 4, wherein S1 comprises: S1.1, constructing a task model, abstracting a continuous video frame into a calculation task M k , and representing the calculation task M3834 by a triplet M k ={v k , d k , c k , wherein v k represents task data volume and affects transmission delay, d k represents calculation load and affects reasoning delay, and c k represents task complexity, represents understanding difficulty of video content and affects accuracy of model reasoning; The method comprises the steps of S1.2, constructing a time delay model, wherein time delay of task processing mainly comprises four parts, namely (1) reasoning time delay, namely time required by the task to be processed by the large model, (2) waiting time delay, namely time required by the task to be queued on corresponding equipment for large model reasoning, (3) transmission time delay, namely time required by the task to be unloaded from a local edge server to other edge servers or cloud servers and data to be transmitted to the corresponding equipment, (4) network time delay, namely fixed time delay on a physical link in the data transmission process, wherein a specific calculation formula is as follows: ; t local represents the time delay of processing at the local edge server, T edge represents the time delay of processing unloaded to other edge servers, and T cloud represents the time delay of processing unloaded to the cloud server; Respectively representing the reasoning time delay of the task on the local edge server, other edge servers and the cloud server; respectively representing the waiting time delay of tasks on a local edge server, other edge servers and a cloud server; respectively representing the transmission delay of the task unloaded from the local edge server to other edge servers or cloud servers; network delays respectively representing offloading of tasks from a local edge server to other edge servers or cloud servers; The method comprises the steps of S1.3, building an energy consumption model, wherein the energy consumption of a system mainly comprises two parts, namely (1) reasoning the energy consumption, namely, the energy consumption of a server during the task processing of a large model, (2) transmitting the energy consumption, namely, the energy consumed by data transmitted to corresponding equipment in the process of unloading the task from a local edge server to other edge servers or cloud servers, wherein the specific calculation formula is as follows: ; E local represents the energy consumption processed at the local edge server, E edge represents the energy consumption unloaded to other edge server processes, and E cloud represents the energy consumption unloaded to the cloud server processes; respectively representing the reasoning energy consumption of the task on the local edge server, other edge servers and the cloud server; respectively representing the energy consumption in the data transmission process when the tasks are unloaded from the local edge server to other edge servers or cloud servers; S1.4, constructing an accuracy model, adding a certain correction amount on a reference value of the accuracy depending on the accuracy of model reasoning on different devices, and relating to the complexity c k of a task, wherein a specific calculation formula is as follows: ; Acc local represents the accuracy rate of processing at the local edge server, acc edge represents the accuracy rate of processing offloaded to other edge servers, and Acc cloud represents the accuracy rate of processing offloaded to the cloud server; representing the accuracy of the task on a local edge server or other edge servers and on a cloud server respectively; Representing the influence of the complexity of the task on the reasoning accuracy of the large model; S1.5, constructing an optimized objective function, and constructing a utility function U i,t for a task M i,t generated by an ith camera at any time t, wherein the calculation formula is as follows: ; Wherein w Acc ,w Lat ,w Eng is a non-negative weight coefficient of accuracy, time delay and energy consumption respectively and is used for balancing the importance of different performance indexes, acc i,t 、T i,t 、E i,t represents the accuracy, time delay and energy consumption of the task M i,t after normalization respectively; Since a plurality of tasks are simultaneously generated at the same time, a total utility function U (i, t) of the system at the time t is also constructed, and the calculation formula is as follows: ; wherein N represents the number of cameras in the system, M represents the number of edge servers in the system, lambda and eta represent non-negative weight coefficients, Representing the load of the mth edge server at time t i.e. the queuing latency, Representing the average load of all edge servers at time t; The overall optimization goal of the system is to find an optimal joint strategy under the condition that constraint conditions are met To maximize the long-term discount-accumulating utility of the system, namely: ; Wherein: Acc min represents the minimum allowable accuracy for each task and T max represents the maximum allowable delay for each task for the discounting factor.
6. The task offloading method of a cloud-edge co-architecture for an apron-oriented intelligent monitoring system of claim 5, wherein S2 comprises: S2.1, converting the multi-objective optimization problem constructed in the step S1.5 into a Markov decision process, wherein the Markov decision process is divided into a high-level decision and a low-level decision, the high-level decision decides whether a task is executed at an unloading cloud server layer or an edge layer, and if the high-level decision distributes the task to be executed at the edge layer, the low-level decision needs to be triggered, and the task is executed at a specific edge server; s2.2, constructing a state space, an action space and a reward function required by deep reinforcement learning, and specifically: the high-level agent focuses on the global resource view, and is used for determining a calculation level, and the state space is as follows: ; Wherein c k 、d k represents the complexity and computational load of the task respectively, The predicted latency for local edge, cloud, and edge layer averages, respectively; b E2C 、B E2E is the bandwidth between edge clouds and edge to edge respectively; The low-level agent focuses on the edge layer resource view, and is used for deciding which edge server is specifically allocated, and the state space is as follows: ; Wherein: Is the expected latency of the ith edge server, except the local edge server; Is the queue length of the corresponding ith edge server; B E2E is the bandwidth between edge servers; The actions of the high-level intelligent agent are discrete binary selection, and the action space is as follows: ; Wherein: Selecting a cloud server on behalf of a user; representing a select edge server; The action of the low-level intelligent agent is to select one of M edge servers, and the action space is as follows: ; Wherein: Representing a select edge server e i ; the higher and lower agents share a global rewarding function: 。
7. The task offloading method of a cloud edge end collaborative architecture for an apron-oriented intelligent monitoring system according to claim 6, wherein the S3 design task offloading method based on multi-agent deep reinforcement learning comprises: Selecting a multi-agent near-end strategy optimization algorithm to solve a constructed Markov decision process according to the task requirement of the apron intelligent monitoring system, and finding an optimal task unloading strategy of the system, wherein a specific network architecture is as follows: The high-level Actor network is a multi-layer perceptron, the input is a high-level state space, the output is a two-class action probability distribution corresponding to the high-level actions, the low-level Actor network is a multi-layer perceptron, the input is a low-level space, the output is an M-dimensional class action probability distribution corresponding to the low-level actions, and the Critic network adopts a centralized design and outputs state values V(s) through the multi-layer perceptron.
8. The tarmac-oriented cloud edge co-architecture task offloading method of claim 4, wherein the S4 training deep reinforcement learning agent comprises: The method adopts a centralized training to remove a centralized execution architecture, introduces a mask updating mechanism on the basis of MAPPO algorithm, and comprises the following specific steps: s4.1, data sampling, wherein each agent interacts with the environment, the collected track data is stored in an experience playback pool, and each piece of data comprises: S is the state of a system and a task observed by an agent at the current moment, a high and a low are discrete action indexes output by a high-level strategy network and a low-level strategy network respectively, log pi high and log pi low are old strategy log probability values corresponding to the actions, which are generated, and are used for calculating an importance sampling ratio subsequently, r is an immediate rewarding value of environmental feedback, v is a value evaluation value of a Critic network to the current state, and is used for calculating an advantage function; S4.2, dominance estimation, namely calculating dominance function of each moment t by using generalized dominance estimation GAE algorithm The specific calculation process comprises the following two sub-steps: S4.2.1 calculating time sequence difference residual error TD, calculating TD error at current moment according to state value estimation output by Critic network : ; R t is an immediate rewarding value of environmental feedback, V (s t ) is state value estimation of a Critic network to a t moment state s t , gamma is a discount factor, weight for balancing immediate rewarding and future long-term rewarding, d t is a task ending mark at the t moment, and if the task is ended or a segment is ended, the value is 1, otherwise, the value is 0; S4.2.2 recursively calculating the dominance function GAE, the formula is: ; wherein λ is a GAE smoothing factor used to trade-off between bias and variance; s4.3, updating the high-level strategy, and calculating the ratio of the high-level strategy The clip loss function of the PPO is used for updating the parameters of the higher-layer network, and the loss function formula of the higher-layer strategy is as follows: ; Wherein: The cutoff parameter is a super parameter in the PPO algorithm and is used for limiting the variation range of the ratio high ; S4.4, updating the mask of the lower-layer strategy, wherein when the upper layer selects 'cloud', the lower-layer action is invalid, noise is introduced in direct updating, and a mask tensor K is constructed firstly: ; Then calculate the original PPO loss of the low-level policy The loss function formula of the low-level strategy is: ; Wherein: A minute amount to prevent zero removal; S4.5, entropy regularization, wherein an entropy regularization term is introduced into the total loss function updated by the Actor network The entropy regularization loss function is calculated as follows: ; Where c ent is the entropy coefficient used to adjust the regularization strength, B is the batch size, i.e. the total number of samples used for each gradient update, Is the information entropy, and the calculation formula is , A minute amount to prevent zero removal; S4.6, calculating a total loss function updated by the Actor network, wherein the calculation formula is as follows: ; S4.7, calculating a total loss function of the Actor network update by the Critic network update, wherein the calculation formula is as follows: ; wherein c vf is the value loss factor used to balance the Actor loss and Critic loss, Is an estimate of the current value, The actual discount report corresponding to the sample is usually obtained by back-pushing the dominance function; The trained agent's Actor model will be deployed on each edge server to make offloading decisions.

Description

Cloud edge end cooperative architecture and task unloading method for apron intelligent monitoring system Technical Field The invention relates to the field of cloud edge end cooperative computing, in particular to a cloud edge end cooperative architecture and a task unloading method for an apron intelligent monitoring system. Background With the rapid development of civil aviation transportation industry and the deep promotion of intelligent airport construction, the operating environment of an apron is increasingly complex, and higher requirements on safety, efficiency and intelligent level are provided. The apron intelligent monitoring system is used as a key infrastructure for guaranteeing the ground operation safety of an aircraft and the resource scheduling efficiency of an elevator apron, and needs to process mass data generated by a multi-source sensor in real time and rapidly complete high-computation-intensive tasks such as target detection, behavior recognition, conflict early warning and the like. The traditional apron monitoring system mostly adopts a centralized processing architecture, namely all perceived data are uploaded to a central cloud platform for unified analysis and decision. However, the mode has obvious limitations that on one hand, the volume of original data such as video streams in an airport apron scene is huge, the network bandwidth pressure is greatly increased due to overall conduction, the service requirements of low time delay and high reliability are difficult to meet, on the other hand, the cloud centralized processing introduces obvious communication delay, response is delayed when the emergency is handled, and potential safety hazards exist. To alleviate the above problems, edge computing technology is introduced in the field of apron monitoring, to implement localized processing of part of the computing tasks by deploying computing nodes at the edges of the network near the data sources. However, existing edge solutions typically employ static task allocation strategies, lacking the ability to adapt to dynamic apron environments. When edge node resources are limited or suffer from sudden high load, problems such as task queuing, processing timeout and even service interruption can still occur. In addition, most current systems fail to effectively build a cloud-edge three-level collaboration mechanism. The terminal equipment only bears the data acquisition function, and an efficient task unloading scheduling strategy is also lacking between the edge layer and the cloud end, so that the overall resource utilization rate is low, the energy efficiency ratio is low, and the comprehensive requirements of the apron monitoring system on instantaneity, robustness and expandability are difficult to support. Therefore, a novel cloud edge end cooperative architecture for an intelligent monitoring scene of an apron is needed to be provided, and a dynamic task unloading method is combined, so that the task execution position is intelligently decided according to multidimensional factors such as task characteristics, equipment loads, network states and the like, the elastic scheduling and global optimization of computing resources are realized, and the overall performance and the operation efficiency of a system are improved on the premise of guaranteeing the monitoring precision and timeliness. Disclosure of Invention The invention aims to provide a cloud edge end cooperative architecture and a task unloading method for an apron intelligent monitoring system, which can intelligently decide and calculate the execution position of a task according to real-time operation environment and task requirements, and realize the apron intelligent monitoring service with low time delay, high reliability and high energy efficiency. In order to achieve the above object, the present invention provides the following solutions: the first aspect provides a cloud edge end cooperative architecture for an apron intelligent monitoring system: The architecture consists of a terminal equipment layer, an edge server layer and a cloud server layer, wherein all the layers are connected through a wired or wireless network to form a distributed intelligent perception and reasoning system. And a plurality of high-definition monitoring cameras are deployed at the terminal equipment layer, and the high-definition monitoring cameras uniformly cover the main operation area of the parking apron and are used for collecting dynamic video streams of aircrafts, ground service vehicles and personnel in real time. And each camera is used as an independent data source, and the original video data is uploaded to the corresponding edge server through the local area network. The terminal equipment also comprises an apron control personnel workbench which is used for receiving the alarm information and checking the analysis result. At the edge server level, a plurality of edge computing nodes are configured, each edge server