A method and system for understanding and scheduling visual Internet of Things tasks based on continuous learning

Through the continuous learning-based visual network task understanding and scheduling method, predicting the resource requirements of video tasks and optimizing resource allocation, the problem that cloud servers cannot distinguish task performance requirements is solved, and the real-time processing capabilities and resource utilization efficiency of visual network applications are improved.

CN118521946BActive Publication Date: 2024-11-08BEIJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410991751.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-08
Estimated Expiration
2044-07-23

AI Technical Summary

Technical Problem

Cloud servers cannot distinguish the performance requirements of tasks when processing a large number of video tasks, resulting in non-real-time tasks blocking real-time tasks, affecting the performance of visual network applications.

Method used

Using the visual network task understanding and scheduling method based on continuous learning, by obtaining the visual network state of the historical time period, building resource state vectors and feature vectors, inputting them into the pre-trained resource demand characterization model and serialized resource demand model, and predicting the task's Resource requirements, and calculate the dependency value of the task combination, determine the final task combination and allocate to the node, ensuring the targeted resource allocation.

Benefits of technology

Effectively distinguish the performance requirements of tasks, ensure targeted resource allocation, avoid task blockage, and improve the real-time processing capabilities and resource utilization efficiency of visual network applications.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118521946B_ABST
    Figure CN118521946B_ABST
Patent Text Reader

Abstract

The present invention provides a method and system for understanding and scheduling visual network tasks based on continuous learning, the method comprising the following steps: obtaining the visual network status of a historical time period, the visual network status including the status of multiple resources; constructing the status of each resource at each historical time point in the historical time period as a resource state vector, and inputting it into a pre-trained resource demand characterization model to obtain a resource feature vector; constructing the resource feature vectors corresponding to multiple historical time points as a feature group, and inputting it into a pre-trained serialized resource demand model to obtain predicted resource demand vectors of multiple tasks; combining the tasks corresponding to the predicted resource demand vectors to obtain multiple task combinations, and calculating the dependency values ​​of various task combinations based on the resource demand vectors of multiple tasks; determining the task combination to be finally used based on the dependency value, and allocating the task combination to be finally used to a node.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of Internet of Things, and in particular to a method and system for understanding and scheduling visual Internet of Things tasks based on continuous learning. Background Art

[0002] The rise of technologies such as the Internet of Things, mobile Internet, 5G and artificial intelligence has stimulated the emergence of a large number of new video Internet of Things (visual Internet for short) applications with high bandwidth requirements and high latency requirements, such as ultra-high-definition video on demand / live broadcast, remote video security, cloud gaming, augmented reality / virtual reality / mixed reality.

[0003] Although the application prospects of visual networking are broad, cloud computing is still needed to achieve real-time and highly reliable video task processing. In short, after the video frame is generated from the terminal (such as smart camera, smart phone), it needs to be uploaded to the cloud server with the help of network transmission technology such as optical fiber and 5G to complete the processing of related tasks.

[0004] Efficient cloud network resource matching and scheduling with video tasks is the key to ensuring the performance of visual networking applications. However, when faced with a large number of video tasks, cloud servers cannot distinguish the performance requirements of tasks, and there is a disadvantage that non-real-time tasks block the execution of real-time tasks. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a method for understanding and scheduling visual Internet of Things tasks based on continuous learning to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides a method for understanding and scheduling visual networking tasks based on continuous learning, the method comprising the following steps:

[0007] Obtaining a visual network status in a historical time period, wherein the visual network status includes statuses of multiple resources;

[0008] The state of each resource at each historical time point in the historical time period is constructed as a resource state vector, and input into a pre-trained resource demand representation model, wherein the resource demand representation model outputs a resource feature vector;

[0009] The resource feature vectors corresponding to multiple historical time points are constructed into a feature group, and are input into a pre-trained serialized resource demand model, wherein the serialized resource demand model outputs predicted resource demand vectors of multiple tasks, and the value of each dimension in the resource demand vector is the demand of the corresponding task for a resource;

[0010] The tasks corresponding to the predicted resource demand vectors are combined to obtain multiple task combinations, and dependency values ​​of the various task combinations are calculated based on the resource demand vectors of the multiple tasks;

[0011] The task combination finally used is determined based on the dependency value, and the task combination finally used is allocated to the node.

[0012] Adopting the above scheme, this scheme determines the resource consumption characteristics of historical video tasks based on the historical visual network status, and predicts the resource requirements of video tasks in the next period of time. Among them, each deep learning network layer records the demand characteristics of a type of video tasks, and classifies the task requirements by comparing the representation distance between the demand characteristics of the new tasks and the characteristics of the old tasks that have been "learned". Finally, the requirements for various resources of the video tasks in the next period of time are obtained. Even when facing a large number of tasks, the actual resources are provided as a reference through the predicted situation, which can distinguish the performance requirements of the tasks and ensure the targeted resource allocation.

[0013] In some embodiments of the present invention, in the step of constructing the state of each resource at each historical time point in the historical time period as a resource state vector and inputting it into a pre-trained resource demand characterization model, wherein the resource demand characterization model outputs a resource feature vector, the resource demand characterization model includes an embedding layer encoder and a fully connected layer encoder.

[0014] In some embodiments of the present invention, in the step of constructing resource feature vectors corresponding to multiple historical time points into a feature group and inputting them into a pre-trained serialized resource requirement model, the serialized resource requirement model outputs predicted resource requirement vectors of multiple tasks, the serialized resource requirement model includes a position encoding layer, a multi-head attention layer, a feedforward response layer and a normalization layer.

[0015] In some embodiments of the present invention, in the step of combining tasks corresponding to the predicted resource demand vectors to obtain multiple task combinations, and calculating dependency values ​​of the various task combinations based on the resource demand vectors of the multiple tasks;

[0016] Combining all tasks to obtain multiple task combinations, each task combination includes multiple subtask groups, and each subtask group includes at least one task;

[0017] The dependency value of each task combination is calculated based on the resource requirement vectors of the tasks.

[0018] In some embodiments of the present invention, in the step of calculating the dependency value of each task combination based on the resource requirement vector of the task, the dependency value of each subtask group in each task combination is calculated, and the dependency value of the task combination is calculated based on the dependency values ​​of each subtask group in the task combination.

[0019] In some embodiments of the present invention, in the step of calculating the dependency value of each subtask group in each task combination, the dependency of the subtask group is calculated based on the following formula:

[0020]

[0021] in, Indicates the demand of the jth task for the i-th resource; It represents the maximum value of the demand for various types of resources among all tasks in the task group; k represents the number of resource types; m represents the number of tasks in the task group; Indicates the dependency value of the task group.

[0022] In some embodiments of the present invention, after determining the final used task combination based on the dependency value, the step of allocating the final used task combination to the node includes:

[0023] Get the current remaining amount of various resources of each node in the visual network;

[0024] Based on the various resource amounts required by the tasks in the subtask groups of the task combination and the various resource amounts currently remaining on each node, the matching degree between the subtask groups and the nodes is calculated, and the nodes allocated to each subtask group in the task combination are determined based on the matching degree.

[0025] In some embodiments of the present invention, in the step of calculating the matching degree of the subtask group and the node based on the amounts of various resources required for the tasks in the subtask group of the task combination and the current remaining amounts of various resources of each node, the resource difference value between each task in the subtask group and the node is calculated based on the amounts of various resources required for the tasks in the subtask group and the current remaining amounts of various resources of the node, and the matching degree of the subtask group and the node is calculated based on the resource difference value between each task in the subtask group and the node.

[0026] In some embodiments of the present invention, in the step of calculating the resource difference value between each task in the subtask group and the node based on the amount of each resource required by the task in the subtask group and the amount of each resource currently remaining at the node, the resource difference value between each task in the subtask group and the node is calculated based on the following formula:

[0027]

[0028] in, represents the demand for the i-th resource in the first task of the subtask group; represents the remaining amount of the i-th resource in the s-th node; represents the demand for the i-th resource in the j-th task of the subtask group; represents the demand for the i-th resource in the m-th task of the subtask group; k represents the number of resource types; β represents an array consisting of the resource difference values ​​of each task and node in the subtask group.

[0029] The second aspect of the present invention also provides a visual Internet of Things task requirement understanding and scheduling system based on continuous learning, the system includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0030] The third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps implemented by the aforementioned method for understanding and scheduling visual Internet tasks based on continuous learning.

[0031] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention can be specifically pointed out and obtained in the specification and the accompanying drawings.

[0032] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

[0034] Figure 1 A schematic diagram of an implementation of a method for understanding and scheduling visual networking tasks based on continuous learning according to the present invention;

[0035] Figure 2 A schematic diagram of another implementation of the method for understanding and scheduling visual networking tasks based on continuous learning of the present invention;

[0036] Figure 3 A schematic diagram of the visual networking architecture;

[0037] Figure 4 A schematic diagram of the processing architecture of a resource requirement representation model based on continuous learning and a serialized resource requirement model;

[0038] Figure 5 Schematic diagram of the processing architecture for model training. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0040] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0041] like Figure 1 and 3 As shown, the present invention proposes a method for understanding and scheduling visual networking tasks based on continuous learning, which includes the following steps:

[0042] Step S100, obtaining the visual network status of a historical time period, wherein the visual network status includes the status of multiple resources;

[0043] During the specific implementation process, the visual network includes multiple nodes, and the status of each resource in the visual network is the sum of the remaining resources of the resource of all nodes in the visual network or the sum of the consumed resources of each node. Specifically, the resources include storage resources, transmission resources and computing resources.

[0044] In a specific implementation process, the visual network status of the historical time period includes the visual network status of multiple historical time points.

[0045] Step S200, constructing the state of each resource at each historical time point in the historical time period into a resource state vector, and inputting it into a pre-trained resource demand characterization model, wherein the resource demand characterization model outputs a resource feature vector;

[0046] In the specific implementation process, in the step of constructing the state of each resource at each historical time point in the historical time period into a resource state vector, the state of each resource at the historical time point is used as the value of a dimension in the resource state vector to construct the resource state vector corresponding to the historical time point.

[0047] In a specific implementation process, the resource feature vector is used to represent the resource situation at a historical time point.

[0048] Step S300, constructing resource feature vectors corresponding to multiple historical time points into a feature group, and inputting them into a pre-trained serialized resource demand model, wherein the serialized resource demand model outputs predicted resource demand vectors of multiple tasks, wherein the value of each dimension in the resource demand vector is the demand of the corresponding task for a resource;

[0049] In the specific implementation process, this solution estimates the resource consumption of a period of time based on the visual network status of the historical time period, and then predicts the various resources required for tasks in the subsequent period of time based on the resource consumption.

[0050] Step S400, combining tasks corresponding to the predicted resource demand vectors to obtain multiple task combinations, and calculating dependency values ​​of various task combinations based on the resource demand vectors of the multiple tasks;

[0051] In a specific implementation process, tasks corresponding to the predicted resource demand vectors are combined to obtain different subtask groups, and different combinations are all a task combination.

[0052] Step S500: determining a task combination to be finally used based on the dependency value, and allocating the task combination to be finally used to a node.

[0053] In the specific implementation process, this scheme determines the node assigned to each subtask group in the task combination through the dependency value. When the resources are planned in advance, when the task arrives, the corresponding subtask group can be directly sent to the corresponding node for processing. This scheme ensures the efficiency of task processing by combining tasks.

[0054] Adopting the above scheme, this scheme determines the resource consumption characteristics of historical video tasks based on the historical visual network status, and predicts the resource requirements of video tasks in the next period of time. Among them, each deep learning network layer records the demand characteristics of a type of video tasks, and classifies the task requirements by comparing the representation distance between the demand characteristics of the new tasks and the characteristics of the old tasks that have been "learned". Finally, the requirements for various resources of the video tasks in the next period of time are obtained. Even when facing a large number of tasks, the actual resources are provided as a reference through the predicted situation, which can distinguish the performance requirements of the tasks and ensure the targeted resource allocation.

[0055] In some embodiments of the present invention, in the step of constructing the state of each resource at each historical time point in the historical time period as a resource state vector and inputting it into a pre-trained resource demand characterization model, wherein the resource demand characterization model outputs a resource feature vector, the resource demand characterization model includes an embedding layer encoder and a fully connected layer encoder.

[0056] like Figure 4As shown in the figure, in the specific implementation process, the resource demand characterization model extracts the storage, transmission and computing resource requirements of the video task based on the fully connected network. First, the cloud network storage, transmission and computing resource status, namely the memory size, network link capacity and CPU and GPU remaining capacity, is used as input, and the high-dimensional raw data is mapped to a low-dimensional manifold vector through the embedding layer encoder, and then the feature vector is output to the fully connected layer encoder through the splicing operation.

[0057] In the specific implementation process, in order to improve the execution efficiency of the algorithm, the resource requirement representation model of this scheme sets a memo layer for the fully connected layer encoder to record the characteristic value of the experience quality requirement of each task. The task experience quality requirements are classified by comparing the distance between the requirement characteristics of the new task and the old task requirement characteristics learned by the fully connected layer. If the feature distance is large, a new connection layer is added and a service level agreement requirement matching level is assigned to it; if the feature requirement distance is within the established threshold, the service level agreement level corresponding to the previous fully connected layer is referred to and set as the experience quality requirement of the task. The above operation maintains a certain memory capacity during the training process-by adaptively adjusting the fully connected network structure to adapt to the differences between the experience quality requirements of different tasks, it can also save the processed task requirement characteristics and timely perceive the experience quality requirements of subsequent similar tasks.

[0058] In some embodiments of the present invention, in the step of constructing resource feature vectors corresponding to multiple historical time points into a feature group and inputting them into a pre-trained serialized resource requirement model, the serialized resource requirement model outputs predicted resource requirement vectors of multiple tasks, the serialized resource requirement model includes a position encoding layer, a multi-head attention layer, a feedforward response layer and a normalization layer.

[0059] In the specific implementation process, in order to further improve the accuracy of understanding the task resource requirements, the serialized resource requirement model fine-tunes the resource requirements of the current task based on the changes in the storage, transmission, and computing resource utilization of the historical task flow. Figure 4 As shown in the figure, the algorithm combines the historical task requirement features into feature groups in a serialized order and inputs them into the Transformer encoder, which includes position encoding (Position Embedding), multi-head attention response mechanism (Multi-head Attention), feed forward response (Feed Forward) and normalization (Add & Normalization) operations to accurately adjust the task's requirements for storage, transmission and computing resources.

[0060] like Figure 5As shown, in the specific implementation process, this solution places the resource demand characterization model and the serialized resource demand model in an online reinforcement learning framework for training. The online reinforcement learning framework enables the model to adapt to the changes in cloud network resources and the suddenness of massive tasks in a timely manner during the training process. The training of the algorithm is iterated in steps. Each step requires the cloud network storage, transmission and computing resource status as input. The task demand perception agent based on continuous learning makes scheduling action decisions, and then the computing power resource encapsulation and scheduling environment executes the scheduling action and returns the new state as the input for the next iteration. In order to guide the agent to make correct decisions, the present invention gives the agent a reward value according to the decision action and the optimization goal after each action is executed, so as to adjust the resource allocation decision action online so that it can find the best decision that matches the action space. This solution uses the policy gradient algorithm to improve the training efficiency of the algorithm, and uses the trained reward value to perform gradient descent on the neural network parameters, and finally converges to a stable state.

[0061] In some embodiments of the present invention, in the step of combining tasks corresponding to the predicted resource demand vectors to obtain multiple task combinations, and calculating dependency values ​​of the various task combinations based on the resource demand vectors of the multiple tasks;

[0062] Combining all tasks to obtain multiple task combinations, each task combination includes multiple subtask groups, and each subtask group includes at least one task;

[0063] The dependency value of each task combination is calculated based on the resource requirement vectors of the tasks.

[0064] In some embodiments of the present invention, in the step of calculating the dependency value of each task combination based on the resource requirement vector of the task, the dependency value of each subtask group in each task combination is calculated, and the dependency value of the task combination is calculated based on the dependency values ​​of each subtask group in the task combination.

[0065] In some embodiments of the present invention, in the step of calculating the dependency value of the task combination based on the dependency values ​​of each subtask group in the task combination, the sum of the dependency values ​​of each subtask group in each task combination is calculated, and the task combination corresponding to the smallest value is selected as the final selected task combination.

[0066] In some embodiments of the present invention, in the step of calculating the dependency value of each subtask group in each task combination, the dependency value of the subtask group is calculated based on the following formula:

[0067]

[0068] in, Indicates the demand of the jth task for the i-th resource; It represents the maximum value of the demand for various types of resources among all tasks in the task group; k represents the number of resource types; m represents the number of tasks in the task group; Indicates the dependency value of the task group.

[0069] By adopting the above scheme, this scheme can understand the resource balance degree of each task in the subtask group through the calculation of the above formula. If the dependency value is smaller, it means that the resource balance degree of each task in the subtask group is higher, and the consumption of each resource of the subtask group is more balanced. If this scheme is continued, it can ensure that the resource consumption of the subtask group received by each node is relatively balanced, and the consumption of each resource of the node will also be more balanced. The node will not have a large difference in the consumption of various resources, and there will not be a situation where a certain resource is left in large quantities but other resources are consumed. This scheme can improve the overall utilization efficiency of node resources through the above scheme.

[0070] like Figure 2 As shown, in some embodiments of the present invention, after determining the final used task combination based on the dependency value, the step of allocating the final used task combination to the node includes:

[0071] Step S510, obtaining the current remaining amount of various resources of each node in the visual network;

[0072] In the specific implementation process, this scheme adopts a resource encapsulation mechanism to obtain the current remaining amount of various resources of each node in the visual network. This mechanism is used to measure the remaining amount of cloud network storage, transmission and computing resources, and logically encapsulates various resources for subsequent scheduling. Its operations include the encapsulation of wireless radio frequency resources between the terminal and the wireless access network, and the measurement of storage, transmission and computing resources between the wireless access network and the edge / center cloud. For the former, the wireless channel quality is obtained by capturing the wireless access network signaling. In the wireless access network, the terminal is identified by the International Mobile Subscriber Identity (IMSI), which can be obtained in the signaling of the terminal random access process. The wireless channel quality is obtained by parsing the corresponding control frame through the link layer signaling (MAC CE). For the latter, it mainly provides a virtual resource view for the upper layer application and controls resource allocation. Among them, the virtual resource view is related to the frame structure configuration of the wireless frame, and adopts a combination of semi-static radio resource control (RRC) configuration and dynamic downlink control information (DCI) configuration. The number of available resources in each scheduling unit is determined by capturing all user control signaling configuration information.

[0073] Step S520, based on the amounts of resources required by the tasks in the subtask groups of the task combination and the amounts of resources currently remaining on each node, calculate the matching degree between the subtask groups and the nodes, and determine the nodes allocated to each subtask group in the task combination based on the matching degree.

[0074] In some embodiments of the present invention, in the step of calculating the matching degree of the subtask group and the node based on the amounts of various resources required for the tasks in the subtask group of the task combination and the current remaining amounts of various resources of each node, the resource difference value between each task in the subtask group and the node is calculated based on the amounts of various resources required for the tasks in the subtask group and the current remaining amounts of various resources of the node, and the matching degree of the subtask group and the node is calculated based on the resource difference value between each task in the subtask group and the node.

[0075] In some embodiments of the present invention, in the step of calculating the resource difference value between each task in the subtask group and the node based on the amount of each resource required by the task in the subtask group and the amount of each resource currently remaining at the node, the resource difference value between each task in the subtask group and the node is calculated based on the following formula:

[0076]

[0077] in, represents the demand for the i-th resource in the first task of the subtask group; represents the remaining amount of the i-th resource in the s-th node; represents the demand for the i-th resource in the j-th task of the subtask group; represents the demand for the i-th resource in the m-th task of the subtask group; k represents the number of resource types; β represents an array consisting of the resource difference values ​​of each task and node in the subtask group.

[0078] During the specific implementation process, in the step of calculating the matching degree between the subtask group and the node based on the resource difference value between each task and the node of the subtask group, the standard deviation of each value in the array composed of the resource difference values ​​of each task and the node of the subtask group is calculated, and the node corresponding to the smallest standard deviation among the standard deviations calculated between the subtask group and each node is used as the task group assigned to the subtask group.

[0079] Adopting the above scheme, this scheme matches the logical resources of the node with the corresponding task resource request. Assuming that a task has a large and urgent demand for CPU resources and a small demand for memory resources, while other tasks have the opposite demand for the above two resources, there is a dependency relationship between the two tasks - the two can be packaged and sent to the same node for parallel processing. This can not only avoid the decrease in overall resource utilization caused by the unbalanced use of CPU resources and memory resources when the two tasks are executed sequentially, but also improve the overall processing efficiency of multiple tasks; in addition, this scheme traverses all tasks to find multiple tasks with the highest resource demand matching degree and packages them. In order to assign each task group to a suitable server node for processing. The present invention selects the matching pair with the smallest standard deviation between the demand for each type of resource in each task group and the remaining amount of each type of resource in each server as the basis for scheduling decisions to ensure the accuracy of task scheduling.

[0080] An embodiment of the present invention also provides a visual Internet of Things task understanding and scheduling system based on continuous learning, the system comprising a computer device, the computer device comprising a processor and a memory, the memory storing computer instructions, the processor being used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0081] The embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps implemented by the aforementioned visual networking task understanding and scheduling method based on continuous learning are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0082] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0083] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0084] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for understanding and scheduling visual Internet of Things tasks based on continuous learning, characterized in that: The method comprises the following steps: Obtaining a visual network status in a historical time period, wherein the visual network status includes statuses of multiple resources; The state of each resource at each historical time point in the historical time period is constructed as a resource state vector, and input into a pre-trained resource demand representation model, wherein the resource demand representation model outputs a resource feature vector; The resource feature vectors corresponding to multiple historical time points are constructed into a feature group and input into a pre-trained serialized resource demand model, wherein the serialized resource demand model outputs predicted resource demand vectors of multiple tasks, wherein the value of each dimension in the resource demand vector is the demand of the corresponding task for a resource; The tasks corresponding to the predicted resource demand vectors are combined to obtain multiple task combinations, and the dependency values ​​of various task combinations are calculated based on the resource demand vectors of the multiple tasks. All tasks are combined to obtain multiple task combinations, each task combination includes multiple subtask groups, and each subtask group includes at least one task; the dependency value of each task combination is calculated based on the resource demand vectors of the tasks, the dependency value of each subtask group in each task combination is calculated, the dependency value of the task combination is calculated based on the dependency values ​​of each subtask group in the task combination, and the dependency value of the subtask group is calculated based on the following formula: in, Indicates the demand of the jth task for the i-th resource; It represents the maximum value of the demand for various types of resources among all tasks in the task group; k represents the number of resource types; m represents the number of tasks in the task group; Indicates the dependency value of the task group; The task combination finally used is determined based on the dependency value, and the task combination finally used is allocated to the node.

2. The method for understanding and scheduling visual networking tasks based on continuous learning according to claim 1, characterized in that: In the step of constructing the state of each resource at each historical time point in the historical time period as a resource state vector and inputting it into a pre-trained resource demand characterization model, wherein the resource demand characterization model outputs a resource feature vector, the resource demand characterization model includes an embedding layer encoder and a fully connected layer encoder.

3. The method for understanding and scheduling visual networking tasks based on continuous learning according to claim 1, characterized in that: In the step of constructing resource feature vectors corresponding to multiple historical time points into a feature group and inputting them into a pre-trained serialized resource requirement model, the serialized resource requirement model outputs predicted resource requirement vectors of multiple tasks, the serialized resource requirement model includes a position encoding layer, a multi-head attention layer, a feedforward response layer and a normalization layer.

4. The method for understanding and scheduling visual networking tasks based on continuous learning according to claim 1, characterized in that: After determining the final used task combination based on the dependency value, the step of allocating the final used task combination to the node includes: Get the current remaining amount of various resources of each node in the visual network; Based on the various resource amounts required by the tasks in the subtask groups of the task combination and the various resource amounts currently remaining on each node, the matching degree between the subtask groups and the nodes is calculated, and the nodes allocated to each subtask group in the task combination are determined based on the matching degree.

5. The method for understanding and scheduling visual networking tasks based on continuous learning according to claim 4 is characterized in that: In the step of calculating the degree of matching between the subtask group and the node based on the amounts of various resources required for the tasks in the subtask group of the task combination and the amounts of various resources currently remaining for each node, the resource difference value between each task in the subtask group and the node is calculated based on the amounts of various resources required for the tasks in the subtask group and the amounts of various resources currently remaining for the node, and the degree of matching between the subtask group and the node is calculated based on the resource difference value between each task in the subtask group and the node.

6. The method for understanding and scheduling visual networking tasks based on continuous learning according to claim 5, characterized in that: In the step of calculating the resource difference value between each task in the subtask group and the node based on the various resource amounts required by the tasks in the subtask group and the various resource amounts currently remaining at the node, the resource difference value between each task in the subtask group and the node is calculated based on the following formula: in, represents the demand for the i-th resource in the first task of the subtask group; represents the remaining amount of the i-th resource in the s-th node; represents the demand for the i-th resource in the j-th task of the subtask group; represents the demand for the i-th resource in the m-th task of the subtask group; k represents the number of resource types; β represents an array consisting of the resource difference values ​​of each task and node in the subtask group.

7. A visual Internet of Things task understanding and scheduling system based on continuous learning, characterized in that: The system includes a computer device, which includes a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method as described in any one of claims 1 to 6.