An end-side collaborative intelligent task processing method, device and equipment in a mobile scenario and a storage medium

Abstract

An end-edge collaborative intelligent task processing method and device in a mobile scenario, equipment and a storage medium, the method is for the end-edge collaborative intelligent task processing in a mobile scenario based on model segmentation during the switching process of a mobile intelligent terminal in the coverage range of multiple edge nodes, three task processing schemes of original service node cooperation before switching, node cooperation after switching and local processing are proposed, the power resources of the terminal equipment and the selected edge node, the communication quality between them, the mobile position of the terminal equipment and other environmental information are combined to dynamically optimize the intelligent task processing scheme and model segmentation, solve the dynamic decision challenge caused by the dynamic change of the device environment and resources in the mobile scenario, and the cold start problem caused by the task migration between edge nodes, and reduce the processing delay of the end-edge collaborative intelligent task based on model segmentation during the node switching period of the terminal side equipment.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and specifically relates to a method, device, equipment and storage medium for end-edge collaborative intelligent task processing in a mobile scenario. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, large models such as ChatGPT have received widespread attention and research. However, the further increase in the number of AI model parameters has led to an explosive growth in the computing power required for AI model training and inference. Furthermore, with the development of mobile smart terminals, the number of smart terminals has increased dramatically, and the demand for smart services on mobile terminals has become more diverse, and the demand for service quality has further increased. This will bring massive amounts of smart task requests and data transmission on the mobile terminal side.

[0003] The traditional cloud computing offload paradigm places computing and storage in a centralized data center. However, the linear growth of computing equipment in cloud computing centers makes it difficult to support the massive processing requests of compute-intensive intelligent applications. Furthermore, the massive amount of data transmitted at the edge further increases bandwidth load, making it difficult to guarantee the real-time performance of AI applications. Therefore, to address the excessive centralized processing load, existing research has considered offloading computing tasks from the cloud to the device and edge, proposing edge computing architectures and combining them with AI to develop the concept of edge intelligence. To further improve the computing performance of AI tasks and fully utilize the available computing power and storage resources of device-side devices, a model-sharded device-edge collaborative intelligence paradigm has been proposed.

[0004] However, due to the dynamic changes in the device's communication environment and available computing nodes in mobile scenarios, as well as the coupled nature of model partitioning and task processing decisions, dynamic joint optimization of model partitioning and task processing decisions for mobile devices is an urgent challenge. Furthermore, due to the varying cache probabilities of different edge nodes, when a mobile terminal switches between edge nodes, computing tasks migrating between them may require deploying corresponding cached content, leading to cold starts and resulting in latency overhead for cache deployment.

[0005] It is crucial to propose a dynamic and efficient intelligent task processing solution for the end-edge collaborative architecture based on model segmentation in mobile scenarios. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to propose a method, device, equipment and storage medium for end-edge collaborative intelligent task processing in a mobile scenario to solve or partially solve the problems mentioned in the background technology.

[0007] In order to achieve the above object, in a first aspect, the present application provides a method for end-edge collaborative intelligent task processing in a mobile scenario, comprising:

[0008] A central server receives a task processing request initiated by a mobile intelligent terminal when generating an intelligent task, wherein the intelligent task includes an AI workflow task of model inference and model training;

[0009] The central server senses available resources and environmental information of devices in a current edge network to obtain sensing data, wherein the devices in the edge network include the mobile intelligent terminal and edge nodes;

[0010] The central server uses an intelligent task processing decision algorithm to jointly optimize and decide on a processing scheme and model splitting based on the sensing data, with the aim of minimizing the delay of end-edge collaborative processing of intelligent tasks based on model splitting during node switching, and then issues a joint optimization and decision instruction to the mobile intelligent terminal and the edge nodes;

[0011] The mobile intelligent terminal and the edge nodes collaboratively execute intelligent tasks according to the joint optimization and decision of the processing scheme and the model splitting, and return the execution results to the mobile intelligent terminal, thereby completing the task processing request initiated by the mobile intelligent terminal.

[0012] As a preferred solution of the method for end-edge collaborative intelligent task processing in a mobile scenario, if the intelligent task is determined by the intelligent task processing decision algorithm to be processed collaboratively by the original service edge node before switching, the current belonging edge node and the neighboring nodes of the belonging edge node are used as relays to establish a connection with the original service edge node, and model splitting and collaborative processing are performed according to the model splitting decision; the original service edge node before switching has deployed an AI model required by the intelligent task.

[0013] As a preferred solution of the method for end-edge collaborative intelligent task processing in a mobile scenario, if the intelligent task is determined by the intelligent task processing decision algorithm to be processed collaboratively by the belonging edge node after switching, it is determined whether the belonging edge node has a cache of the corresponding intelligent task model, and if not, the model is requested from a data center; when the belonging edge node has a cache of the corresponding intelligent task model, model splitting and collaborative processing are performed according to the model splitting decision.

[0014] As a preferred solution of the method for end-edge collaborative intelligent task processing in a mobile scenario, if the intelligent task is determined by the intelligent task processing decision algorithm to be processed locally, the end-side device completes all intelligent task processing.

[0015] As a preferred solution for the device-edge collaborative intelligent task processing method in mobile scenarios, the mobile intelligent terminal deploys the AI ​​model required for the intelligent task in advance;

[0016] The edge node randomly deploys the AI ​​model required for the intelligent task of the end-side device with a preset probability. If the edge node currently selected as the collaborative node has not deployed the AI ​​model in advance, the edge node currently selected as the collaborative node requests the data center equipped with the required model to download the model and complete the collaborative intelligent task processing;

[0017] The data center is a macro base station and a centralized network node of a cloud computing center to provide the AI ​​model cache required by the mobile smart terminal.

[0018] As a preferred solution for device-edge collaborative intelligent task processing in mobile scenarios, the available resources and environmental information of devices in the current edge network include:

[0019] The available computing resources of the mobile intelligent terminal and the edge node, the location information of the mobile intelligent terminal, the communication environment between the mobile intelligent terminal and the edge node, and the cache status of the edge node;

[0020] The model segmentation regards the AI ​​model corresponding to the intelligent task initiated by the mobile intelligent terminal as a hierarchical structure, and each layer of the structure is regarded as a segmentation point; the segmentation point of the AI ​​model is obtained according to the model segmentation decision, and the AI ​​model is divided into two parts. The first half is processed on the terminal side device, and the intermediate feature data is transmitted to the selected collaborative node, and the selected node completes the model processing of the second half.

[0021] In a second aspect, the present invention provides a device-edge collaborative intelligent task processing apparatus in a mobile scenario, comprising:

[0022] A task processing request receiving module is used for the central server to receive task processing requests, wherein the task processing requests are initiated to the central server when the mobile intelligent terminal generates an intelligent task, and the intelligent task includes an AI workflow task of model reasoning and model training;

[0023] A perception data acquisition module is used by the central server to perceive the available resources and environmental information of devices in the current edge network to obtain perception data; the devices in the edge network include mobile smart terminals and edge nodes;

[0024] A joint optimization decision module is configured for the central server to make joint optimization decisions on the processing scheme and model segmentation based on the perception data and using an intelligent task processing decision algorithm, with the goal of minimizing the latency of the intelligent task end-edge collaborative processing based on model segmentation during node switching, and to issue joint optimization decision instructions to the mobile intelligent terminal and the edge node;

[0025] A processing result returning module is used for the mobile intelligent terminal and the edge node to collaboratively execute intelligent tasks according to the joint optimization decision of the processing scheme and the model segmentation, and return the execution result to the mobile intelligent terminal to complete the task processing request initiated by the mobile intelligent terminal.

[0026] As a preferred solution for the device-edge collaborative intelligent task processing device in mobile scenarios, in the joint optimization decision module:

[0027] If the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the original service edge node before the switch, the current edge node and the adjacent nodes of the edge node are used as relays to establish a connection with the original service edge node, and model segmentation and collaborative processing are performed according to the model segmentation decision; the original service edge node is deployed with the AI ​​model required for the intelligent task before the switch;

[0028] If the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the edge node after the handover, it is determined whether the current edge node has a cache of the corresponding intelligent task model. If not, the model is requested from the data center. If the edge node has a cache of the corresponding intelligent task model, model segmentation and collaborative processing are performed according to the model segmentation decision.

[0029] If the intelligent task is determined by the intelligent task processing decision algorithm to be processed locally, all intelligent task processing is completed on the terminal side device.

[0030] As a preferred solution for the device-edge collaborative intelligent task processing device in mobile scenarios, the mobile intelligent terminal deploys the AI ​​model required for the intelligent task in advance;

[0031] The edge node randomly deploys the AI ​​model required for the intelligent task of the end-side device with a preset probability. If the edge node currently selected as the collaborative node has not deployed the AI ​​model in advance, the edge node currently selected as the collaborative node requests the data center equipped with the required model to download the model and complete the collaborative intelligent task processing;

[0032] The data center is a macro base station and a centralized network node of a cloud computing center to provide the AI ​​model cache required by the mobile smart terminal.

[0033] As a preferred solution for edge-device collaborative intelligent task processing in mobile scenarios, the available resources and environmental information of devices in the current edge network include:

[0034] The available computing resources of the mobile intelligent terminal and the edge node, the location information of the mobile intelligent terminal, the communication environment between the mobile intelligent terminal and the edge node, and the cache status of the edge node;

[0035] In the joint optimization decision module, the model segmentation regards the AI ​​model corresponding to the intelligent task initiated by the mobile intelligent terminal as a hierarchical structure, and each layer of the structure is regarded as a segmentation point; the segmentation point of the AI ​​model is obtained according to the model segmentation decision, and the AI ​​model is divided into two parts. The first half is processed on the terminal side device, and the intermediate feature data is transmitted to the selected collaborative node, and the selected node completes the model processing of the second half.

[0036] According to a third aspect, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for edge-to-edge collaborative intelligent task processing in a mobile scenario according to the first aspect or any possible implementation thereof is implemented.

[0037] In a fourth aspect, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to enable the computer to execute the end-edge collaborative intelligent task processing method in a mobile scenario of the first aspect or any possible implementation thereof.

[0038] From the above description, it can be seen that the technical solution provided by the present invention is aimed at the end-edge collaborative intelligent task processing based on model segmentation in the mobile scenario when the mobile intelligent terminal switches between the coverage areas of multiple edge nodes. It proposes three task processing schemes: selecting the original service node for collaboration before switching, selecting the belonging node for collaboration after switching, and local processing. Combined with the computing power resources of the terminal device and the selected edge node, the communication quality between the two, the mobile location of the terminal device and other environmental information, a dynamic joint optimization decision is made on the intelligent task processing scheme and model segmentation, which solves the dynamic decision-making challenges brought about by the dynamic changes of the device environment and resources in the mobile scenario, as well as the cold start problem caused by the task migration between edge nodes, and reduces the processing delay of the end-edge collaborative intelligent task based on model segmentation during the node switching period of the end-side device. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the present invention or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1This is a schematic diagram of edge-end collaborative model reasoning based on model segmentation.

[0041] Figure 2 A schematic diagram of an AI device-edge collaborative reasoning solution based on model segmentation in the device-edge collaborative intelligent task processing method in a mobile scenario provided by an embodiment of the present invention;

[0042] Figure 3 A diagram showing training results for a related example of the method for edge-to-edge collaborative intelligent task processing in a mobile scenario provided by an embodiment of the present invention;

[0043] Figure 4 This is an architecture diagram of a device-edge collaborative intelligent task processing apparatus in a mobile scenario provided by an embodiment of the present invention;

[0044] Figure 5 Schematic diagram of the structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.

[0046] It should be noted that, unless otherwise defined, technical or scientific terms used in the embodiments of the present invention should have the same general meaning as those understood by persons of ordinary skill in the art to which the present invention pertains. The words "include" or "comprise" and similar expressions used in the embodiments of the present invention mean that the elements or objects preceding the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0047] Currently, to address the issue of excessive centralized processing load, existing technologies consider shifting computing tasks from the cloud to the end and edge, proposing an edge computing architecture and combining edge computing with AI to create the concept of edge intelligence. To further improve the computing performance of AI tasks and fully utilize the available computing power and storage resources of end-side devices, a model-based end-edge collaborative intelligence paradigm has been proposed. For example, CN111445026B discloses multi-path inference acceleration for deep neural networks for edge intelligence applications (prior art 1). This approach, combined with the structural characteristics of AI models, monitors and analyzes network loads and the computing power of terminals and edge devices online in an edge computing environment. The multi-path model is segmented with the goal of minimizing inference latency, and the model segments are deployed separately at the terminal and edge layers, ultimately forming a multi-path inference acceleration framework. Furthermore, given the inherent mobility of mobile smart devices, their environments and available resources are constantly changing. When mobile smart devices switch between different edge nodes, task processing overhead increases due to service migration and the lack of required cache on edge nodes, posing a challenge to real-time dynamic offloading decisions.

[0048] In the related technology, CN109862592B discloses a resource management and scheduling method in a mobile edge computing environment based on multi-base station collaboration, which considers dynamically migrating terminal computing tasks between the intelligent base station to which it belongs, the adjacent base stations of the intelligent base station to which it belongs, and the cloud, and optimizes transmission and computing delays, cache allocation, system benefits, etc. (prior technology 2).

[0049] However, prior art 1 only considers edge-device collaborative AI inference solutions based on model partitioning deployment, without considering the challenges and challenges posed by terminal device mobility to model partitioning decisions and resource scheduling. Furthermore, prior art 2 only considers the migration and deployment of computing tasks in mobile scenarios, ignoring the computing power of terminal devices and failing to consider the performance improvements that edge-device collaborative intelligent solutions based on model partitioning can bring to AI intelligent services. Based on the existing technologies, there is a lack of edge-device collaborative intelligent task processing solutions based on model partitioning in mobile scenarios. Due to the dynamic changes in the communication environment and resources such as available computing nodes in mobile scenarios, and the coupled coupling between model partitioning and task processing decision-making, dynamic joint optimization of model partitioning and task processing solutions is an urgent challenge. Furthermore, due to the different cache probabilities of edge nodes, when a mobile terminal switches between different edge nodes, the migration of computing tasks between edge nodes may require the deployment of corresponding cache content, resulting in a cold start phenomenon, which in turn incurs latency overhead due to cache deployment. Therefore, it is crucial to develop a dynamic and efficient intelligent task processing solution for edge-device collaborative architectures based on model partitioning in mobile scenarios, which is also the technical problem addressed by this technical solution.

[0050] In view of this, the embodiments of the present invention propose a method, apparatus, device and storage medium for end-edge collaborative intelligent task processing in mobile scenarios, which is based on model segmentation when a mobile intelligent terminal switches between the coverage areas of multiple edge nodes. The task processing scheme and model segmentation decision are jointly optimized to solve the dynamic decision-making challenges brought about by the dynamic changes in the device environment and resources in mobile scenarios, as well as the cold start problem caused by task migration between edge nodes, and reduce the processing delay of end-edge collaborative intelligent tasks based on model segmentation during the node switching period of the end-side device. The following is the specific content of the embodiments of the present invention.

[0051] In this embodiment, the solution proposed by the present invention is described by taking AI model reasoning as an example. Figure 1 As shown in the figure, it shows the schematic diagram of AI end-edge collaborative reasoning based on model segmentation. The AI ​​model is regarded as a layered structure, and the total number of layers of the AI ​​model is defined as N. Two virtual layer structures l0 and l1 are added at the input and output ends respectively. N+1 , it is stipulated that l0 is executed on the end-side device, l N+1Executed on the base station side. Each layer of the AI ​​model can be used as a split point. If the selected split point is layer m, then l0~l m Executed on the end device, m+1 ~l N+1 Executed on the base station side. Define the amount of intermediate feature data after the i-th layer split point as I i , the computational load of layer i is L i .

[0052] See also Figure 2 , an embodiment of the present invention provides a method for device-edge collaborative intelligent task processing in a mobile scenario, comprising the following steps:

[0053] Step 1: The central server receives a task processing request, which is initiated to the central server by a mobile smart terminal when generating an intelligent task. The intelligent task includes an AI workflow task of model reasoning and model training.

[0054] Step 2: The central server senses the available resources and environmental information of the devices in the current edge network to obtain perception data; the devices in the edge network include mobile smart terminals and edge nodes; wherein the perception data includes the available computing power resources of the mobile smart terminal and the edge node, the location information of the mobile smart terminal, the communication environment between the mobile smart terminal and the edge node, and the cache status of the edge node.

[0055] Step 3: Based on the perception data, the central server uses an intelligent task processing decision algorithm to minimize the latency of the intelligent task end-edge collaborative processing based on model segmentation during node switching, makes a joint optimization decision on the processing scheme and model segmentation, and sends the joint optimization decision instruction to the mobile intelligent terminal and the edge node.

[0056] Specifically, such as Figure 2 As shown, any layer of the AI ​​model can be selected for segmentation. For edge-to-edge collaborative AI reasoning tasks based on model segmentation when a mobile smart terminal switches between the coverage areas of multiple edge nodes, there are three different processing solutions: 1) Execute locally on the mobile terminal; 2) Offload the intelligent task segmentation to the current edge node after the switch for edge-to-edge collaborative reasoning; 3) Offload the intelligent task segmentation to the original service edge node before the switch for edge-to-edge collaborative reasoning, selecting the current edge node and its adjacent nodes as relays, maintaining a connection with the original service node to collaboratively execute the reasoning task.

[0057] The mobile intelligent terminal deploys the AI ​​model required for the intelligent task in advance;

[0058] The edge node randomly deploys the AI ​​model required for the intelligent task of the end-side device with a preset probability. If the edge node currently selected as the collaborative node has not deployed the AI ​​model in advance, the edge node currently selected as the collaborative node requests the data center equipped with the required model to download the model and complete the collaborative intelligent task processing;

[0059] The data center is a macro base station and a centralized network node of a cloud computing center to provide the AI ​​model cache required by the mobile smart terminal.

[0060] The available resources and environment information of the devices in the current edge network include:

[0061] The available computing resources of the mobile intelligent terminal and the edge node, the location information of the mobile intelligent terminal, the communication environment between the mobile intelligent terminal and the edge node, and the cache status of the edge node;

[0062] The model segmentation regards the AI ​​model corresponding to the intelligent task initiated by the mobile intelligent terminal as a hierarchical structure, and each layer of the structure is regarded as a segmentation point; the segmentation point of the AI ​​model is obtained according to the model segmentation decision, and the AI ​​model is divided into two parts. The first half is processed on the terminal side device, and the intermediate feature data is transmitted to the selected collaborative node, and the selected node completes the model processing of the second half.

[0063] In this embodiment, the model segmentation decision is defined as:

[0064] l∈[0,1,...,N]

[0065] If the number of selectable edge intelligent base stations is M, the service execution decision is defined as:

[0066] x=[x 1,s ,x 1,r ,...,x M,s ,x M,s ]

[0067] where x i,s =1(i∈[1,M]) means that the i-th base station is selected as the service node and the task is split and offloaded to the base station for execution; i,r =1(i∈[1,M]) means that the i-th base station is selected as the relay node, and a connection is established with the original serving base station through the relay node to perform the task; Indicates that the task is executed locally on the device. It means that at the same time, the mobile terminal can select at most one edge base station for task offloading.

[0068] Step 4: If the intelligent task is determined by the intelligent task processing decision algorithm to be executed locally on the mobile terminal, the AI model does not split offloading, and all are executed by the mobile terminal.

[0069] Step 5: If the intelligent task is determined by the intelligent task processing decision algorithm to be executed by the edge node after switching for end-edge collaborative inference, the central server splits the model according to the local computing power, the computing power of the edge node, the channel state between the terminal and the edge node, and other environmental information, and distributes the model splitting decision to the mobile terminal and the edge node for collaborative intelligent processing. If the edge node does not have the corresponding model cached in advance, it needs to request the download of the model from the data center equipped with the required model and complete the end-edge collaborative inference.

[0070] Step 6: If the intelligent task is determined by the intelligent task processing decision algorithm to be executed by the original service edge node before switching for end-edge collaborative inference, the current edge node and the neighboring nodes of the current edge node are used as relays to establish a connection with the original service edge node. The central server splits the model according to the local computing power, the computing power of the original service edge node, the channel state between the terminal and the edge node, and other environmental information, and distributes the model splitting decision to the mobile terminal and the original service edge node for end-edge collaborative inference. The intermediate feature data after model splitting is transmitted to the original service edge node by the edge node and the neighboring nodes of the edge node as relays. The overhead of data migration between nodes is H, and the maximum number of relay nodes is set to e max .

[0071] Step 7: The mobile intelligent terminal and the edge node execute the intelligent task according to the joint optimization decision of the processing scheme and the model splitting, and return the execution result to the mobile intelligent terminal, completing the task processing request initiated by the mobile intelligent terminal.

[0072] Specifically, if the task is executed locally, the intelligent task processing delay is:

[0073]

[0074] where f local is the user's local computing resources;

[0075] If the end-edge collaborative inference is selected on the edge node after switching, the intelligent task processing delay includes the local computing delay, the intermediate feature transmission delay, the base station side computing delay, and the cache download delay:

[0076]

[0077] where l jis the selected model split point, Calculate the latency locally. Calculate the delay on the edge node side, f i is the computing resource of the selected node, is the intermediate characteristic transmission delay, R i is the data transmission rate of the wireless channel between the mobile terminal and its edge node, is the cache download delay, Q is the required cache data volume, R c is the data transmission rate between the data center and the edge node. i Used to represent the cache status of edge node i, if S i =1, it means that node i has cached the required model. If S i =0, it means that node i has not cached the required model and needs to download the cache from the data center.

[0078] Specifically, if the original service edge node is selected for device-edge collaborative inference before switching, the intelligent task processing latency includes local computing latency, intermediate feature transmission latency, node-side computing latency, and inter-node switching latency, and is:

[0079]

[0080] Where H is the switching delay between the current node and the adjacent node, which is related to the data transmission between nodes and the node reconfiguration overhead. sum is the cumulative switching delay between the neighboring node of the node and the original service node. If the neighboring node of the node is the original service node, then H sum = 0. Γ is the penalty factor. If the number of selected relay nodes exceeds the threshold e max , then Γ is an infinite positive real number, otherwise Γ=0.

[0081] Among them, by combining different task execution plan decision variables with the corresponding task processing delay, the total task processing delay can be obtained as:

[0082]

[0083] Furthermore, minimizing the latency of device-edge collaborative inference based on model segmentation when a mobile smart terminal switches between the coverage areas of multiple edge nodes can be transformed into the following problem:

[0084]

[0085] This optimization problem is a discrete optimization problem, and a reinforcement learning algorithm can be used to jointly optimize and schedule task reasoning scheme decisions and model splitting decisions in real time. Specifically, a deep reinforcement learning algorithm framework based on the actor-critic structure is designed. The model input is the perception data, including the available computing resources of the mobile smart terminal and the edge node, the location information of the mobile smart terminal, the communication environment between the mobile smart terminal and the edge node, and the cache status of the edge node; the model output is the service execution decision and the model splitting decision. The inverse of the end-side collaborative reasoning delay is used as the reward function to guide the reinforcement learning model to train until the model converges, and a joint optimization decision is obtained to minimize the end-side collaborative intelligent task processing delay. It solves the dynamic decision-making challenges brought about by the dynamic changes in device environment and resources in mobile scenarios, as well as the cold start problem caused by task migration between edge nodes, and reduces the processing delay of end-side collaborative intelligent tasks based on model splitting during node switching on the end-side device.

[0086] In order to verify the technical effect of the present invention, the following calculation examples are designed:

[0087] Consider a traffic scenario of 6300m*3500m, including a mobile vehicle user and multiple roadside base station units, where the maximum speed of the mobile vehicle user is 25m / s, and the roadside base station units are deployed at intervals of 250m. The mobile vehicle user and the roadside base station unit collaborate to perform image classification tasks based on Alexa Net. The computing power of the mobile vehicle is set to a CPU clock frequency of 20GHz, and the computing power of the roadside base station unit is randomly generated with a CPU clock frequency in the range of 80-100GHz. The mobile vehicle signal transmission power is 0.2W, the transmission bandwidth is 10MHz, the carrier frequency is 5.9GHz, and the path loss model refers to the formula 16.7log 10 d+38.77+18.2log 10 f c , where is the distance between the mobile vehicle and the roadside base station unit, is the carrier frequency, and the channel noise power is -174dBm / Hz.

[0088] See also Figure 3 To address the problem of minimizing the latency of device-edge collaborative inference based on model partitioning, this embodiment uses an actor-critic deep reinforcement learning algorithm, trained with the inverse of the device-edge collaborative inference latency as the reward function, to achieve a joint optimization decision for service execution and model partitioning. The training convergence curve in the figure below shows that the inference latency achieved by the algorithm in this embodiment is close to the optimal performance of the exhaustive approach, and the convergence results are stable, demonstrating that it is possible to achieve joint optimization of service execution and model partitioning decisions and achieve near-optimal performance.

[0089] To summarize, an embodiment of the present invention receives a task processing request through a central server, and the task processing request is initiated to the central server when a mobile smart terminal generates an intelligent task, and the intelligent task includes AI workflow tasks of model reasoning and model training; the central server perceives the available resources and environmental information of the devices in the current edge network to obtain perception data; the devices in the edge network include mobile smart terminals and edge nodes; the central server uses the intelligent task processing decision algorithm based on the perception data to minimize the end-edge collaborative processing delay of the intelligent task based on model segmentation during node switching, and makes a joint optimization decision on the processing scheme and model segmentation, and sends the joint optimization decision instruction to the mobile smart terminal and the edge node; the mobile smart terminal and the edge node collaboratively execute the intelligent task according to the joint optimization decision of the processing scheme and the model segmentation, and return the execution result to the mobile smart terminal to complete the task processing request initiated by the mobile smart terminal. Among them, if the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the original service edge node before the switch, the current edge node and the adjacent nodes of the edge node are used as relays to establish a connection with the original service edge node, and model segmentation and collaborative processing are performed according to the model segmentation decision; the original service edge node is deployed with the AI ​​model required for the intelligent task before the switch; if the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the edge node after the switch, it is determined whether the current edge node has a cache of the corresponding intelligent task model. If there is no cache of the corresponding intelligent task model, the model is requested from the data center; when the edge node has a cache of the corresponding intelligent task model, model segmentation and collaborative processing are performed according to the model segmentation decision; if the intelligent task is determined by the intelligent task processing decision algorithm to be processed locally, all intelligent task processing is completed on the end-side device. The present invention solves the dynamic decision-making challenges brought about by the dynamic changes of device environment and resources in mobile scenarios, as well as the cold start problem caused by task migration between edge nodes, and reduces the processing delay of end-side collaborative intelligent tasks based on model segmentation during node switching on the end-side device.

[0090] It should be noted that the method of the embodiment of the present invention can be performed by a single device, such as a computer or server. The method of this embodiment can also be applied in a distributed scenario, where multiple devices cooperate to perform the method. In such a distributed scenario, one of the multiple devices may only perform one or more steps of the method of the embodiment of the present invention, and the multiple devices will interact with each other to complete the method.

[0091] It is noted that some embodiments of the application have been described. In some cases, the recited acts or steps can be performed in different order than that described above and still achieve desirable results. Additionally, the process depicted in the figures does not necessarily require the particular order shown or sequential order to achieve desirable results. In some implementations, multitasking and parallel processing can be advantageous.

[0092] Referring to Figure 4 , based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the embodiments of the application also provide an edge-side collaborative intelligent task processing device in a mobile scenario, comprising:

[0093] A task processing request receiving module 100 is configured to receive a task processing request by a central server, wherein the task processing request is initiated by the central server when a mobile intelligent terminal generates an intelligent task, and the intelligent task includes an AI workflow task of model inference and model training;

[0094] A perception data acquisition module 200 is configured to perceive available resources and environmental information of devices in a current edge network by a central server to obtain perception data, wherein the devices in the edge network include a mobile intelligent terminal and an edge node;

[0095] A joint optimization decision module 300 is configured to make a joint optimization decision on a processing scheme and model splitting by a central server according to the perception data and using an intelligent task processing decision algorithm, with the goal of minimizing the intelligent task end-edge collaborative processing delay based on model splitting during node switching, and to issue a joint optimization decision instruction to the mobile intelligent terminal and the edge node;

[0096] A processing result returning module 400 is configured to execute an intelligent task by the mobile intelligent terminal and the edge node according to the joint optimization decision of the processing scheme and the model splitting, and to return an execution result to the mobile intelligent terminal, thereby completing the task processing request initiated by the mobile intelligent terminal.

[0097] In the present embodiment, the joint optimization decision module 300 comprises:

[0098] If the intelligent task is determined by the intelligent task processing decision algorithm to be processed collaboratively by the original service edge node before switching, the current belonging edge node and the neighboring node of the belonging edge node are used as relays to establish a connection with the original service edge node, and model splitting and collaborative processing are performed according to the model splitting decision; the original service edge node before switching is deployed with an AI model required by the intelligent task;

[0099] If the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the edge node after the handover, it is determined whether the current edge node has a cache of the corresponding intelligent task model. If not, the model is requested from the data center. If the edge node has a cache of the corresponding intelligent task model, model segmentation and collaborative processing are performed according to the model segmentation decision.

[0100] If the intelligent task is determined by the intelligent task processing decision algorithm to be processed locally, all intelligent task processing is completed on the terminal side device.

[0101] In this embodiment, the mobile intelligent terminal deploys the AI ​​model required for the intelligent task in advance;

[0102] The edge node randomly deploys the AI ​​model required for the intelligent task of the end-side device with a preset probability. If the edge node currently selected as the collaborative node has not deployed the AI ​​model in advance, the edge node currently selected as the collaborative node requests the data center equipped with the required model to download the model and complete the collaborative intelligent task processing;

[0103] The data center is a macro base station and a centralized network node of a cloud computing center to provide the AI ​​model cache required by the mobile smart terminal.

[0104] In this embodiment, the available resources and environment information of the devices in the current edge network include:

[0105] The available computing resources of the mobile intelligent terminal and the edge node, the location information of the mobile intelligent terminal, the communication environment between the mobile intelligent terminal and the edge node, and the cache status of the edge node;

[0106] In the joint optimization decision module, the model segmentation regards the AI ​​model corresponding to the intelligent task initiated by the mobile intelligent terminal as a hierarchical structure, and each layer of the structure is regarded as a segmentation point; the segmentation point of the AI ​​model is obtained according to the model segmentation decision, and the AI ​​model is divided into two parts. The first half is processed on the terminal side device, and the intermediate feature data is transmitted to the selected collaborative node, and the selected node completes the model processing of the second half.

[0107] The device of the above embodiment is used to implement the corresponding end-edge collaborative intelligent task processing method in a mobile scenario in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0108] Based on the same inventive concept, corresponding to any of the above-mentioned embodiments and methods, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements the end-edge collaborative intelligent task processing method in a mobile scenario described in any of the above embodiments.

[0109] Figure 5 A more specific hardware structure diagram of an electronic device provided in this embodiment is shown. The device may include: a processor 510, a memory 520, an input / output interface 530, a communication interface 540, and a bus 550. The processor 510, the memory 520, the input / output interface 530, and the communication interface 540 are connected to each other within the device via the bus 540.

[0110] The processor 510 can be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0111] The memory 520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage devices, dynamic storage devices, etc. The memory 520 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program codes are stored in the memory 520 and are called and executed by the processor 510.

[0112] The input / output interface 530 is used to connect input / output modules to implement information input and output. The input / output modules can be configured as components in the device (not shown in the figure) or can be externally connected to the device to provide corresponding functions. Input devices may include a keyboard, mouse, touch screen, microphone, various sensors, etc., and output devices may include a display, speaker, vibrator, indicator light, etc.

[0113] The communication interface 540 is used to connect to a communication module (not shown) to enable communication between the device and other devices. The communication module can communicate via a wired method (such as USB, network cable, etc.) or a wireless method (such as mobile network, Wi-Fi, Bluetooth, etc.).

[0114] The bus 540 comprises a pathway for transmitting information between the various components of the device (eg, the processor 510 , the memory 520 , the input / output interface 530 , and the communication interface 540 ).

[0115] It should be noted that although the above device only shows the processor 510, the memory 520, the input / output interface 530, the communication interface 540, and the bus 540, in a specific implementation, the device may also include other components necessary for normal operation. In addition, it will be understood by those skilled in the art that the above device may only include the components necessary to implement the embodiments of this specification, and does not necessarily include all the components shown in the figure.

[0116] The electronic device of the above embodiment is used to implement the corresponding end-edge collaborative intelligent task processing method in a mobile scenario in any of the above embodiments, and has the beneficial effects of the corresponding method embodiment, which will not be repeated here.

[0117] Based on the same inventive concept, corresponding to any of the above-mentioned embodiment methods, the present invention also provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute a terminal-edge collaborative intelligent task processing method in a mobile scenario as described in any of the above embodiments.

[0118] The computer-readable media of this embodiment include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device.

[0119] The computer instructions stored in the storage medium of the above embodiment are used to enable the computer to execute a terminal-edge collaborative intelligent task processing method in a mobile scenario as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0120] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Within the scope of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the embodiments of the present invention as described above, which are not provided in detail for the sake of simplicity.

[0121] In addition, to simplify the description and discussion, and in order not to obscure the embodiments of the present invention, known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided figures. In addition, devices may be shown in the form of block diagrams to avoid obscuring the embodiments of the present invention, and this also takes into account the fact that the details of the implementation of these block diagram devices are highly dependent on the platform on which the embodiments of the present invention will be implemented (i.e., these details should be fully within the scope of understanding of those skilled in the art). Where specific details (e.g., circuits) are set forth to describe exemplary embodiments of the present invention, it will be apparent to those skilled in the art that embodiments of the present invention may be implemented without these specific details or with variations in these specific details. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0122] Although the present invention has been described in conjunction with specific embodiments thereof, many alternatives, modifications, and variations of these embodiments will be apparent to those skilled in the art in light of the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may utilize the embodiments discussed.

[0123] The embodiments of the present invention are intended to cover all such substitutions, modifications, and variations that fall within the scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for device-edge collaborative intelligent task processing in a mobile scenario, wherein: include: The central server receives a task processing request, which is initiated by a mobile smart terminal when generating an intelligent task. The intelligent task includes an AI workflow task such as model reasoning and model training. The central server senses the available resources and environmental information of devices in the current edge network and obtains perception data; The devices in the edge network include mobile intelligent terminals and edge nodes; The central server uses an intelligent task processing decision algorithm based on the perception data to minimize the latency of the intelligent task end-edge collaborative processing based on model segmentation during node switching, makes a joint optimization decision on the processing solution and model segmentation, and sends the joint optimization decision instruction to the mobile intelligent terminal and the edge node; The mobile intelligent terminal and the edge node collaboratively execute the intelligent task according to the processing solution and the joint optimization decision of the model segmentation, and return the execution result to the mobile intelligent terminal, completing the task processing request initiated by the mobile intelligent terminal; The latency of intelligent task processing includes local computing latency, intermediate feature transmission latency, base station computing latency, cache download latency, node computing latency, and inter-node switching latency. The latency minimization of device-edge collaborative inference based on model segmentation when a mobile smart terminal switches between the coverage areas of multiple edge nodes can be transformed into: st0≤l≤N, H sum ≤H·e max Where x is the service execution decision, l is the model segmentation decision, t is the total task processing delay, t1 is the intelligent task processing delay when executing the task locally, t2 is the intelligent task processing delay when the edge node to which it belongs performs device-edge collaborative reasoning after the switch, and t3 is the intelligent task processing delay when the original service edge node performs device-edge collaborative reasoning before the switch. i,s =1, i∈[1,M] means selecting the i-th base station as the service node and offloading the task to the base station for execution; i,r =1, i∈[1,M] means that the i-th base station is selected as the relay node, and a connection is established with the original serving base station through the relay node to perform the task; M is the number of edge intelligent base stations; H is the switching delay between the current node and the adjacent node, which is related to the data transmission between nodes and the node reconfiguration overhead. sum is the cumulative switching delay between the neighboring node of the node and the original service node; e max is the maximum number of relay nodes; The optimization problem is a discrete optimization problem, and the task reasoning scheme decision and the model splitting decision are jointly optimized and scheduled in real time through the reinforcement learning algorithm; the model input is the perception data, including the available computing power resources of the mobile smart terminal and the edge node, the location information of the mobile smart terminal, the communication environment between the mobile smart terminal and the edge node, and the cache status of the edge node; the model output is the service execution decision and the model splitting decision, and the opposite of the end-edge collaborative reasoning delay is used as the reward function to guide the reinforcement learning model for training until the model converges, and a joint optimization decision that minimizes the end-edge collaborative intelligent task processing delay is obtained.

2. The method for edge-to-edge collaborative intelligent task processing in a mobile scenario according to claim 1, wherein: If the intelligent task is determined by the intelligent task processing decision algorithm to require device-edge collaborative processing by the original service edge node before switching, the current edge node and its neighboring nodes are used as relays to establish a connection with the original service edge node, and model segmentation and collaborative processing are performed according to the model segmentation decision. Before the switch, the original service edge node was deployed with the AI ​​model required for intelligent tasks; If the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the edge node after the handover, it is determined whether the current edge node has a cache of the corresponding intelligent task model. If not, the model is requested from the data center. If the edge node has a cache of the corresponding intelligent task model, model segmentation and collaborative processing are performed according to the model segmentation decision. If the intelligent task is determined by the intelligent task processing decision algorithm to be processed locally, all intelligent task processing is completed on the terminal side device.

3. The method for device-edge collaborative intelligent task processing in a mobile scenario according to claim 1, wherein: The mobile intelligent terminal deploys the AI ​​model required for the intelligent task in advance; The edge node randomly deploys the AI ​​model required for the intelligent task of the end-side device with a preset probability. If the edge node currently selected as the collaborative node has not deployed the AI ​​model in advance, the edge node currently selected as the collaborative node requests the data center equipped with the required model to download the model and complete the collaborative intelligent task processing; The data center is a macro base station and a centralized network node of a cloud computing center to provide the AI ​​model cache required by the mobile smart terminal.

4. The method for device-edge collaborative intelligent task processing in a mobile scenario according to claim 1, wherein: The available resources and environment information of devices in the current edge network include: The available computing resources of the mobile intelligent terminal and the edge node, the location information of the mobile intelligent terminal, the communication environment between the mobile intelligent terminal and the edge node, and the cache status of the edge node; The model segmentation regards the AI ​​model corresponding to the intelligent task initiated by the mobile intelligent terminal as a hierarchical structure, and each layer of the structure is regarded as a segmentation point; the segmentation point of the AI ​​model is obtained according to the model segmentation decision, and the AI ​​model is divided into two parts. The first half is processed on the terminal side device, and the intermediate feature data is transmitted to the selected collaborative node, and the selected node completes the model processing of the second half.

5. A device-edge collaborative intelligent task processing device in a mobile scenario, wherein: include: A task processing request receiving module is used for the central server to receive task processing requests, wherein the task processing requests are initiated to the central server when the mobile intelligent terminal generates an intelligent task, and the intelligent task includes an AI workflow task of model reasoning and model training; A perception data acquisition module is used by the central server to perceive the available resources and environmental information of devices in the current edge network to obtain perception data; the devices in the edge network include mobile smart terminals and edge nodes; A joint optimization decision module is configured for the central server to make joint optimization decisions on the processing scheme and model segmentation based on the perception data and using an intelligent task processing decision algorithm, with the goal of minimizing the latency of the intelligent task end-edge collaborative processing based on model segmentation during node switching, and to issue joint optimization decision instructions to the mobile intelligent terminal and the edge node; a processing result returning module, configured for the mobile intelligent terminal and the edge node to collaboratively execute the intelligent task according to the processing solution and the joint optimization decision of the model segmentation, and to return the execution result to the mobile intelligent terminal, thereby completing the task processing request initiated by the mobile intelligent terminal; The latency of intelligent task processing includes local computing latency, intermediate feature transmission latency, base station computing latency, cache download latency, node computing latency, and inter-node switching latency. The latency minimization of device-edge collaborative inference based on model segmentation when a mobile smart terminal switches between the coverage areas of multiple edge nodes can be transformed into: st0≤l≤N, H sum ≤H·e max Where x is the service execution decision, l is the model segmentation decision, t is the total task processing delay, t1 is the intelligent task processing delay when executing the task locally, t2 is the intelligent task processing delay when the edge node to which it belongs performs device-edge collaborative reasoning after the switch, and t3 is the intelligent task processing delay when the original service edge node performs device-edge collaborative reasoning before the switch. i,s =1, i∈[1,M] means selecting the i-th base station as the service node and offloading the task to the base station for execution; i,r =1, i∈[1,M] means that the i-th base station is selected as the relay node, and a connection is established with the original serving base station through the relay node to perform the task; M is the number of edge intelligent base stations; H is the switching delay between the current node and the adjacent node, which is related to the data transmission between nodes and the node reconfiguration overhead. sum is the cumulative switching delay between the neighboring node of the node and the original service node; e max is the maximum number of relay nodes; The optimization problem is a discrete optimization problem, and the task reasoning scheme decision and the model splitting decision are jointly optimized and scheduled in real time through the reinforcement learning algorithm; the model input is the perception data, including the available computing power resources of the mobile smart terminal and the edge node, the location information of the mobile smart terminal, the communication environment between the mobile smart terminal and the edge node, and the cache status of the edge node; the model output is the service execution decision and the model splitting decision, and the opposite of the end-edge collaborative reasoning delay is used as the reward function to guide the reinforcement learning model for training until the model converges, and a joint optimization decision that minimizes the end-edge collaborative intelligent task processing delay is obtained.

6. The device-edge collaborative intelligent task processing apparatus in a mobile scenario according to claim 5, wherein: In the joint optimization decision module: If the intelligent task is determined by the intelligent task processing decision algorithm to require device-edge collaborative processing by the original service edge node before switching, the current edge node and its neighboring nodes are used as relays to establish a connection with the original service edge node, and model segmentation and collaborative processing are performed according to the model segmentation decision. Before the switch, the original service edge node was deployed with the AI ​​model required for intelligent tasks; If the intelligent task is determined by the intelligent task processing decision algorithm to be processed by the edge node after the handover, it is determined whether the current edge node has a cache of the corresponding intelligent task model. If not, the model is requested from the data center. If the edge node has a cache of the corresponding intelligent task model, model segmentation and collaborative processing are performed according to the model segmentation decision. If the intelligent task is determined by the intelligent task processing decision algorithm to be processed locally, all intelligent task processing is completed on the terminal side device.

7. The device-edge collaborative intelligent task processing apparatus in a mobile scenario according to claim 5, wherein: The mobile intelligent terminal deploys the AI ​​model required for the intelligent task in advance; The edge node randomly deploys the AI ​​model required for the intelligent task of the end-side device with a preset probability. If the edge node currently selected as the collaborative node has not deployed the AI ​​model in advance, the edge node currently selected as the collaborative node requests the data center equipped with the required model to download the model and complete the collaborative intelligent task processing; The data center is a macro base station and a centralized network node of a cloud computing center to provide the AI ​​model cache required by the mobile smart terminal.

8. The device-edge collaborative intelligent task processing apparatus in a mobile scenario according to claim 5, wherein: The available resources and environment information of devices in the current edge network include: The available computing resources of the mobile intelligent terminal and the edge node, the location information of the mobile intelligent terminal, the communication environment between the mobile intelligent terminal and the edge node, and the cache status of the edge node; In the joint optimization decision module, the model segmentation regards the AI ​​model corresponding to the intelligent task initiated by the mobile intelligent terminal as a hierarchical structure, and each layer of the structure is regarded as a segmentation point; the segmentation point of the AI ​​model is obtained according to the model segmentation decision, and the AI ​​model is divided into two parts. The first half is processed on the terminal side device, and the intermediate feature data is transmitted to the selected collaborative node, and the selected node completes the model processing of the second half.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the end-edge collaborative intelligent task processing method in a mobile scenario as described in any one of claims 1 to 4.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the end-edge collaborative intelligent task processing method in a mobile scenario as described in any one of claims 1 to 4.

Images