WO2020211722A1 - 模型的推送、模型的请求方法、装置、存储介质及电子装置 - Google Patents

模型的推送、模型的请求方法、装置、存储介质及电子装置 Download PDF

Info

Publication number
WO2020211722A1
WO2020211722A1 PCT/CN2020/084470 CN2020084470W WO2020211722A1 WO 2020211722 A1 WO2020211722 A1 WO 2020211722A1 CN 2020084470 W CN2020084470 W CN 2020084470W WO 2020211722 A1 WO2020211722 A1 WO 2020211722A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
platform
matching instruction
searched
destination node
Prior art date
Application number
PCT/CN2020/084470
Other languages
English (en)
French (fr)
Inventor
袁丽雅
孟伟
Original Assignee
中兴通讯股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from CN201910300243.XA external-priority patent/CN111083722B/zh
Application filed by 中兴通讯股份有限公司 filed Critical 中兴通讯股份有限公司
Priority to EP20790513.4A priority Critical patent/EP3958606A4/en
Priority to US17/603,622 priority patent/US20220197953A1/en
Publication of WO2020211722A1 publication Critical patent/WO2020211722A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/06—Testing, supervising or monitoring using simulated traffic
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F8/00—Arrangements for software engineering
    • G06F8/60—Software deployment
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/903—Querying
    • G06F16/90335—Query processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F8/00—Arrangements for software engineering
    • G06F8/70—Software maintenance or management
    • G06F8/71—Version control; Configuration management
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/55—Push-based network services
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F8/00—Arrangements for software engineering
    • G06F8/30—Creation or generation of source code
    • G06F8/35—Creation or generation of source code model driven
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/02—Arrangements for optimising operational condition

Definitions

  • This application relates to the field of communications, and in particular to a method, device, storage medium, and electronic device for pushing and requesting models.
  • 5G networks Compared with traditional networks, 5G networks have introduced new technologies such as Software Defined Network (SDN) and Network Function Virtualization (NFV), which increase network flexibility while increasing network flexibility. , It also brings complexity in management and operation and maintenance. Automated and intelligent operation and maintenance capabilities will become a rigid demand for telecom networks in the 5G era. Artificial intelligence technology has natural advantages in solving high-calculation data analysis, cross-domain feature mining, and dynamic strategy generation. It will give new models and capabilities for network operation and maintenance in the 5G era.
  • SDN Software Defined Network
  • NFV Network Function Virtualization
  • modeling refers to selecting an appropriate algorithm for a specific analysis requirement, and training the historical data to enable the model to obtain a higher confidence analysis result based on the input data.
  • Application refers to selecting a model that meets the application scenario from the existing models, deploying it and running it in a designated location.
  • the analysis data generated during operation can be used to continuously optimize the model parameters, so that the model can be adjusted in time to maintain the accuracy of model inference.
  • the embodiments of the present application provide a method, device, storage medium, and electronic device for pushing models and requesting models, so as to at least solve the problem of lack of a unified technical solution on how to select a required model from a large number of models in related technologies.
  • a method for pushing a model including: receiving a model matching instruction from an orchestrator, wherein the model matching instruction is generated based on analysis requirements; searching for a model corresponding to the model matching instruction; When the model is searched, the searched model is pushed to the destination node that needs the model.
  • a model request method including: generating a model matching instruction based on analysis requirements; sending the generated model matching instruction to a model platform to instruct the model platform to search for and The model corresponding to the model matching instruction, and when the model is searched, the searched model is pushed to the destination node that needs the model.
  • a model pushing device including: a receiving module configured to receive a model matching instruction from an orchestrator, wherein the model matching instruction is generated based on analysis requirements; a search module uses To search for the model corresponding to the model matching instruction; the push module is used to push the searched model to the destination node that needs the model when the model is found.
  • a device for requesting a model which includes: a generating module for generating model matching instructions based on analysis requirements; a sending module for sending the generated model matching instructions to the model platform, To instruct the model platform to search for a model corresponding to the model matching instruction, and when the model is found, push the searched model to the destination node that needs the model.
  • a storage medium in which a computer program is stored, wherein the computer program is configured to execute the steps in any one of the foregoing method embodiments when running.
  • an electronic device including a memory and a processor, the memory is stored with a computer program, and the processor is configured to run the computer program to execute any of the above Steps in the method embodiment.
  • the model matching instruction of the orchestrator is received, the model matching instruction is generated based on analysis requirements; the model corresponding to the model matching instruction is searched; when the model is searched, the searched model is pushed To the destination node where the model is required, the problem of how to select the required model from a large number of models and the lack of a unified technical solution can be solved, so as to provide a technical solution for how to select the model.
  • Fig. 1 is a flowchart of a method for pushing a model according to an embodiment of the present application
  • Fig. 2 is a flowchart of a request method of a model according to an embodiment of the present application
  • Figure 3 is a schematic structural diagram of a model pushing system according to an example of the present application.
  • Fig. 4 is a structural block diagram of a push device of a model according to an embodiment of the present application.
  • Figure 5 is a structural block diagram of a requesting device for a model according to an embodiment of the present application.
  • FIG. 6 is a schematic diagram of a flow of resource allocation according to an optional embodiment of the present application.
  • Fig. 7 is a schematic diagram of a model selection process according to an optional embodiment of the present application.
  • a method for pushing a model is provided.
  • Fig. 1 is a flowchart of the method for pushing a model according to an embodiment of the present application. As shown in Fig. 1, the process includes the following steps:
  • Step S102 Receive a model matching instruction from the orchestrator, where the model matching instruction is generated based on analysis requirements.
  • Step S104 searching for a model corresponding to the model matching instruction.
  • Step S106 When the model is searched, the searched model is pushed to the destination node that needs the model.
  • the model matching instruction of the orchestrator is received, the model matching instruction is generated based on analysis requirements; the model corresponding to the model matching instruction is searched; when the model is searched, the searched model is pushed To the destination node where the model is required, the problem of how to select the required model from a large number of models and the lack of a unified technical solution can be solved, so as to provide a technical solution for how to select the model.
  • searching for a model corresponding to the model matching instruction includes at least one of the following: searching for a model corresponding to the model matching instruction on the main model platform; searching on the joint model platform of the main model platform The model corresponding to the model matching instruction; the model corresponding to the model matching instruction can be searched on the main model platform and the joint model platform, that is, the main model platform or the joint model platform can be searched, and the main model The platform and the joint model platform serve as a search database for searching.
  • step S104 can be implemented in an optional embodiment in the following manner: when a model corresponding to the model matching instruction is searched, a search success message is fed back to the arranger; When a model corresponding to the model matching instruction is searched, a search failure message is fed back to the arranger.
  • pushing the searched model to the destination node that needs the model includes: at least packaging the model and the metadata of the model into a file; sending the packaged file to the destination For a node, the packaged file may be a docker file or other executable file, which is not limited in the embodiment of the application.
  • the searched model when the searched model is a single model, the single model is packaged into a file and sent to the destination node; when the searched model is multiple, at least one of the following will be performed on the multiple models Operation: arrange and combine, verify, package multiple models that have performed at least one of the above operations into files and send them to the destination node.
  • the model engine determines the model search domain according to the joint type provided by the orchestrator to find whether there is a model that meets the model matching instruction. If it does not exist, send a model matching failure message to the orchestrator. If the matching result is a single model, directly package the model into a deployable manual application. If the matching is multiple models, first complete the orchestration combination and verification of the models, and then package the orchestration information file and the model into a deployable manual application.
  • the method further includes: receiving a model update request sent by the orchestrator when it detects that the model reaches the model update condition, where the model update condition includes: pushing to The model accuracy rate of the destination node is lower than the preset threshold; the model corresponding to the model update request is sent to a sandbox (SandBox), where the model corresponding to the model update request is trained and optimized in the sandbox ; Receive the model after the sandbox training is successful.
  • SandBox sandbox
  • the model matching instruction includes at least one of the following parameters: model or algorithm attributes, model input and output requirements, indication information used to indicate whether to combine model platform search, and destination node information.
  • the technical solutions of the above embodiments of the present application relate to the application of artificial intelligence to network communications, for example, a next-generation network based on 5G, SDN, NFV and other technologies to realize the automatic management of artificial intelligence models and apply the technology and systems to the network.
  • FIG. 2 is a flowchart of the model request method according to an embodiment of the application. As shown in FIG. 2, the process includes the following steps:
  • Step S202 Generate a model matching instruction based on the analysis requirement.
  • Step S204 Send the generated model matching instruction to the model platform to instruct the model platform to search for a model corresponding to the model matching instruction, and when the model is found, push the searched model to the The destination node of the model.
  • the model matching instruction generated based on the analysis request is sent to the model platform to instruct the model platform to search for a model corresponding to the model matching instruction, and when the model is searched, the searched model It is pushed to the destination node that needs the model, so it can solve the problem of how to select the required model from a large number of models and lack a unified technical solution, so as to provide a technical solution for how to select the model.
  • the method further includes: receiving a search success message fed back by the model platform when a model corresponding to the model matching instruction is searched; receiving; The model platform feeds back a search failure message when the model corresponding to the model matching instruction is not found.
  • the method further includes: when it is detected that the model reaches the model update condition, sending a model update request to the model platform to instruct the model platform to communicate with all models.
  • the model corresponding to the model update request is sent to the sandbox, and the model after the successful training of the sandbox is received, wherein the model corresponding to the model update request is trained and optimized in the sandbox, and the model update condition Including: the accuracy of the model pushed to the destination node is lower than the preset threshold.
  • the example of this application provides an artificial intelligence platform system capable of real-time interaction with a telecommunications network, and a method for automatically selecting, deploying and optimizing artificial intelligence models in the telecommunications network through this interaction.
  • Fig. 3 is a schematic structural diagram of the model pushing system according to an example of the present application.
  • the model platform is composed of a model engine, a model library, a sandbox and other modules.
  • the model library is used to store the artificial intelligence model and its corresponding metadata.
  • the metadata of the model is a series of descriptive data about the model, including model name, version, programming language and algorithm used, runtime dependency, and deployment Condition requirements, as well as requirements for the type of input data (such as images, values), content and format requirements, and a description of the type of output data.
  • the model engine is used to manage the models in the model library, including adding, deleting, modifying, and querying models.
  • the model engine selects the artificial intelligence model according to the model matching instructions issued by the orchestrator, and pushes the model to a designated location in the network.
  • the orchestrator is used to manage the artificial intelligence application instances in the network, including the issuance of model matching instructions, model optimization instructions, life cycle management of artificial intelligence application instances, and monitoring of operating status.
  • the orchestrator creates management entries for each artificial intelligence application instance in the management domain.
  • the management entries include the identity document (ID), state machine, location, model information, etc. of the artificial intelligence application instance.
  • the working mode of the model engine is divided into independent and combined. Stand-alone means that when the model engine receives a model matching request, it can only search for models in the model library in the model platform where it is located.
  • the model platform where the model engine that directly interacts with the orchestrator is located serves as the main platform, and the model engine of the main platform can obtain models from the model library of the joint model platform.
  • the joint model platform refers to the model platform that provides the management interface of the model library for the model engine of the main platform and does not directly interact with the orchestrator.
  • the sandbox is used to provide an operating environment for the retraining, verification and optimization of the model.
  • the orchestrator converts the demand for intelligent analysis into model requirements, and sends the model matching request to the model engine of the model platform.
  • the request message includes the orchestrator identity information, The model requirements and the node type and location information of the model deployment target node.
  • the identity information of the orchestrator includes information such as the type, address, and association type of the orchestrator.
  • Model requirements include functional requirements and performance requirements of the model.
  • the federation types in the orchestrator identity information are divided into three types: no federation, primary priority federation, and peer-to-peer federation.
  • the union type in the orchestrator's identity information can only be no union, that is, only one model of the model library can be requested.
  • the federation type in the orchestrator identity information is divided into no federation, primary priority federation and peer-to-peer federation.
  • the main priority combination refers to matching models from the model library (main model library) of the main platform first, and matching from the model library (joint model library) of the joint platform after the matching fails.
  • Peer-to-peer federation means that when matching, the main model library and the joint model library serve as search domains at the same time.
  • the model engine After receiving the model matching instruction, the model engine determines the model search domain according to the joint type provided by the orchestrator, and searches for whether there is a model that meets the model matching instruction. If there is no model that meets the model matching instruction, a model matching failure message is sent to the orchestrator. If the matching result is a single model, the model is directly packaged into a deployable manual application. If the matching is multiple models, first complete the orchestration combination and verification of the models, and then package the orchestration information file and the model into a deployable manual application.
  • the model engine pushes the artificial intelligence application and model metadata together with the orchestrator identity information to the model application target node.
  • the orchestrator receives a model matching failure message, it sends an alarm message. If the model matching success message is received, the artificial intelligence application instance management entry is pre-created, and the information in the entry is in the initialization state, and the artificial intelligence application instance creation request is sent to the model deployment target node.
  • the model deployment target node After the model deployment target node receives the application instance creation request, it verifies the identity information of the orchestrator. If the verification is successful, it creates and starts the instance, and sends an instance creation success message to the orchestrator.
  • the message includes the instance ID, instance operation information, etc. Among them, the instance ID should be able to and only be able to determine the only instance in the target node. If the verification fails, an illegal access alarm is issued.
  • the orchestrator After the orchestrator receives the instance creation success message, it activates and updates the management entry of the instance accordingly, including changing the state machine of the instance to the running state. Then start the monitoring process of the instance status, and send the model update strategy to the target node.
  • the inference result is obtained according to the input of the model, and the network orchestration or control plane determines the network configuration information according to the inference, configures it in the network, and optimizes the network.
  • the target node When the instance is in the running state, the target node sends the output of the model to the big data platform according to certain rules according to the model update strategy, as the training and test data when the model is updated.
  • the orchestrator monitors that the model reaches the model update condition (for example, the model accuracy is lower than the threshold), it sends a model update request to the model engine.
  • the model update condition for example, the model accuracy is lower than the threshold
  • the model engine selects the model to be updated according to the model update request and pushes it to the sandbox.
  • the sandbox extracts the latest data from the big data platform, and retrains and optimizes the model.
  • the sandbox will feedback the model training success message to the model engine.
  • the model engine After the model engine receives the message, it saves the updated model to the model library, packs the new artificial intelligence application and pushes it to the target node, and the orchestrator implements the deployment of the new instance in the target node. If the deployment is successful, the original instance is sent for destruction Instruction to stop the target node and delete the original instance. If the deployment fails, an alarm is issued.
  • the embodiments of this application propose architectures and methods for the automatic selection, deployment, and optimization of artificial intelligence models for the application of artificial intelligence to the network, which realizes the unified management of artificial intelligence models in the network and improves network intelligence. Autonomy.
  • the method according to the foregoing embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware.
  • the technical solution of the present application can be embodied in the form of a software product.
  • the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, and optical disk), and includes several instructions to make a terminal
  • a device which may be a mobile phone, a computer, a server, or a network device, etc.) executes the method described in each embodiment of the present application.
  • a model pushing device is also provided, which is used to implement the above-mentioned embodiments and optional implementation manners, and those that have been explained will not be repeated.
  • the term "module” can implement a combination of software and/or hardware with predetermined functions.
  • the devices described in the following embodiments can be implemented by software, implementation by hardware or a combination of software and hardware is also possible and conceived.
  • Fig. 4 is a structural block diagram of a model pushing device according to an embodiment of the present application.
  • the device includes: a receiving module 40 for receiving a model matching instruction from an orchestrator, wherein the model matching instruction is based on analysis Requirement generation; search module 42 for searching for a model corresponding to the model matching instruction; push module 44 for pushing the searched model to the destination node that needs the model when the model is searched.
  • the search module 42 is also used for at least one of the following: search for a model corresponding to the model matching instruction on the main model platform; search for a model matching the model on the joint model platform of the main model platform A model corresponding to the instruction; searching for a model corresponding to the model matching instruction on the main model platform and the joint model platform.
  • the search module 42 is further configured to: when a model corresponding to the model matching instruction is searched, feed back a search success message to the arranger; if no search is found corresponding to the model matching instruction When the model is selected, the search failure message is fed back to the arranger.
  • the pushing module 44 is further configured to: at least package the model and the metadata of the model into a file; and send the packaged file to the destination node.
  • the push module 44 is also used to: when the searched model is a single model, package the single model into a file and send it to the destination node;
  • the multiple models perform at least one of the following operations: arrange and combine, verify, and package multiple models that have performed at least one of the above operations into files and send them to the destination node.
  • the pushing module 44 is further configured to: receive a model update request sent by the orchestrator when it detects that the model meets the model update condition, where the model update condition includes: the accuracy of the model pushed to the destination node Lower than the preset threshold; send the model corresponding to the model update request to the sandbox, where the model corresponding to the model update request is trained and optimized in the sandbox; after receiving the sandbox training successfully Model.
  • the model matching instruction includes at least one of the following parameters: model or algorithm attributes, model input and output requirements, indication information used to indicate whether to combine model platform search, and destination node information.
  • a device for requesting a model is also provided, which is used to implement the above-mentioned embodiments and optional implementation manners, and those that have been described will not be repeated.
  • the term "module” can implement a combination of software and/or hardware with predetermined functions.
  • the devices described in the following embodiments can be implemented by software, implementation by hardware or a combination of software and hardware is also possible and conceived.
  • Fig. 5 is a structural block diagram of a device for requesting a model according to an embodiment of the present application.
  • the device includes: a generating module 50 for generating model matching instructions based on analysis requirements; and a sending module 52 for transferring the generated
  • the model matching instruction is sent to the model platform to instruct the model platform to search for a model corresponding to the model matching instruction, and when the model is searched, the searched model is pushed to the destination node that needs the model.
  • the model matching instruction generated based on the analysis request is sent to the model platform to instruct the model platform to search for a model corresponding to the model matching instruction, and when the model is searched, the searched model It is pushed to the destination node that needs the model, so it can solve the problem of how to select the required model from a large number of models and lack a unified technical solution, so as to provide a technical solution for how to select the model.
  • the sending module 52 is further configured to: receive the search success message fed back by the model platform when it searches for a model corresponding to the model matching instruction; The search failure message fed back when the model matches the model corresponding to the instruction.
  • the sending module 52 is further configured to: when it is detected that the model meets the model update condition, send a model update request to the model platform to indicate that the model platform will correspond to the model update request.
  • the model is sent to the sandbox, and the model after the sandbox training is successfully received, wherein the model corresponding to the model update request is trained and optimized in the sandbox, and the model update condition includes: pushing to the destination node The accuracy of the model is lower than the preset threshold.
  • each of the above modules can be implemented by software or hardware.
  • it can be implemented in the following manner, but not limited to this: the above modules are all located in the same processor; or, the above modules are combined in any combination The forms are located in different processors.
  • the following models refer to systems that are learned from existing data or experience to achieve specific analysis, prediction and other functions through machine learning, deep learning, and other methods.
  • each model has its own functions. For example, certain models can be used to predict when the number of new users and sessions reaches the required number. Another example is some models that can be used to predict the time for slice expansion. In addition, it is also possible to determine the location of the alarm in the device according to the quantity or type of the alarm information. At the same time, the various models are also related. For example, the time used to predict the number of new users and the number of sessions to reach the required number needs to be used as an output for predicting the expansion of the slice. Therefore, the model and the model can be connected in series, and can also be connected in parallel or in series. The function of the specific model needs to be determined according to the function of the application instance.
  • Figure 6 shows a process of using artificial intelligence to implement 5G core network slicing intelligent resource allocation, as shown below.
  • the orchestrator issues a model matching request to the model engine, requiring the model to automatically scale the slice based on the traffic in the next 10 minutes; the accuracy of the model reaches 90%; the deployment network element type of the model is the network data analysis function network element ( Network Data Analysis Function (NWDAF for short); the IP address of the deployed network element is 192.168.1.107; the federation type is no federation.
  • NWDAF Network Data Analysis Function
  • the model engine finds available models from the model library: the convolutional neural network model with ID 1001 is used for network traffic prediction; the reinforcement learning model with ID 1002 is used for traffic information and service quality (Quality of Service) based on each slice. , QoS) requirements, etc. to realize intelligent resource allocation of slices.
  • the convolutional neural network model with ID 1001 is used for network traffic prediction; the reinforcement learning model with ID 1002 is used for traffic information and service quality (Quality of Service) based on each slice. , QoS) requirements, etc. to realize intelligent resource allocation of slices.
  • the deployment preparation includes: completing the series connection of the two models according to the requirements; packaging the series connection model, the runtime environment that the model depends on, and the metadata into a docker image named Image1.
  • the model engine sends Image1 together with the identity information of the orchestrator to the NWDAF network element corresponding to 192.168.1.107, and NWDAF feeds back a response that the packaged file is received successfully.
  • the model engine notifies the orchestrator that the model matching has been completed, and the orchestrator pre-creates a management item whose state is the initial state.
  • S510 Send Image1 running request and model update strategy to 192.168.1.107 through the orchestrator (when the accuracy of model 1001 drops below 80%, perform model update, the accuracy here can be understood as the model searched by the model engine The ratio of the number of models required by the model matching instruction to the number of models requested by the model matching instruction).
  • NWDAF verifies that the orchestrator is consistent with the orchestrator identity information carried by Image1, creates and starts a service named Instance1 (instance 1) based on the image of Image1, and notifies the orchestrator.
  • the orchestrator activates the management entry of artificial intelligence Instance1, sets the state machine of instance 1 to the running state; and sends a keep-alive message every 60 seconds to confirm the status of instance 1; sends an accuracy rate statistics request every other day to determine whether to execute the model Update.
  • Instance1 receives the corresponding data in the network according to the data input requirements in the metadata, and outputs the analysis result.
  • the Packet Control Function determines the slice resource allocation strategy based on the result, and sends it to the network by NSMF
  • the subnet slice management function network element Network Subnet Slice Management Function, referred to as NSSMF
  • NSSMF Network Subnet Slice Management Function
  • NFVO Network Function Virtualization Orchestrator
  • S516 The arranger is notified that the accuracy of the convolutional neural network model 1001 is reduced to 75%.
  • the orchestrator sends an update request of the model 1001 to the model engine.
  • the model engine selects a training environment containing 1 GPU as a sandbox for the convolutional neural network model 1001.
  • the sandbox takes out the latest 100,000 pieces of analysis data of Instance1 stored in the big data platform, of which 80,000 pieces are used as training data and 20,000 pieces are used as test data, and the convolutional neural network model 1001 is retrained and verified.
  • S520 Receive a notification of completion of model training, and after verifying the convolutional neural network model 1001, it is found that the accuracy of the convolutional neural network model 1001 reaches 90%.
  • the model engine stores the convolutional neural network model 1001 in the model library.
  • Figure 7 shows a flowchart of the application of the network fault root cause analysis model to the network to achieve precise fault location and source tracing.
  • step S602 the orchestrator issues a model matching request to the model engine, requesting the model to locate the root cause alarm information according to the alarm information of the network management system at the same time, the deployment location of the model is the network management system, and the joint mode is the priority local model library.
  • step S604 the model engine fails to find the adapted model from the main model library, finds the adapted model in the joint model library, and selects this model.
  • step S606 the model engine successfully finds the adapted model from the joint model library, finds the adapted model in the joint model library, and returns the matched model.
  • step S608 the model engine packages the model, the runtime environment on which the model depends, and metadata into a docker image with the image name Image1, and sends it to the network management platform together with the identity information of the orchestrator, and receives the file reception success message fed back by the network management platform .
  • step S612 the model engine notifies the orchestrator that the model matching has been completed, and the orchestrator sends an Image1 running request and a model update strategy (not updated) to the network management platform.
  • step S614 the network management system verifies that this orchestrator is consistent with the orchestrator identity information carried by Image1, creates and starts a container named Instance1 based on the image of Image1, and notifies the orchestrator.
  • the orchestrator activates the management entry of Instance1 and sets the state machine of Instance1 to the running state.
  • Instance1 receives corresponding data in the network according to the data input requirements in the metadata, and outputs the alarm analysis result to the PCF, and the PCF formulates an alarm processing strategy accordingly.
  • the embodiment of the present application also provides a storage medium in which a computer program is stored, wherein the computer program is configured to execute the steps in any of the foregoing method embodiments when running.
  • the foregoing storage medium may be configured to store a computer program for executing the following steps:
  • the foregoing storage medium may be configured to store a computer program for executing the following steps:
  • the foregoing storage medium may include, but is not limited to: U disk, Read-Only Memory (Read-Only Memory, ROM for short), Random Access Memory (Random Access Memory, RAM for short), Various media that can store computer programs, such as mobile hard disks, magnetic disks, or optical disks.
  • the embodiment of the present application also provides an electronic device, including a memory and a processor, the memory is stored with a computer program, and the processor is configured to run the computer program to execute the steps in any of the foregoing method embodiments.
  • the aforementioned electronic device may further include a transmission device and an input-output device, wherein the transmission device is connected to the aforementioned processor, and the input-output device is connected to the aforementioned processor.
  • the foregoing processor may be configured to execute the following steps through a computer program:
  • the foregoing processor may be configured to execute the following steps through a computer program:
  • modules or steps of the present application can be implemented by a general computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented by program codes executable by the computing device, so that they can be stored in the storage device for execution by the computing device, and in some cases, they can be executed in a different order than here.
  • the steps shown or described can be implemented by making them into individual integrated circuit modules, or making multiple modules or steps of them into a single integrated circuit module. In this way, this application is not limited to any specific hardware and software combination.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Computer Security & Cryptography (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Debugging And Monitoring (AREA)
  • Information Transfer Between Computers (AREA)

Abstract

本申请提供了一种模型的推送、模型的请求方法、装置、存储介质及电子装置，上述模型的推送方法包括：接收编排器发送的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；搜索与所述模型匹配指令对应的模型；在搜索到所述模型的情况下，将搜索到的模型推送至需要所述模型的目的节点。

Description

模型的推送、模型的请求方法、装置、存储介质及电子装置
本申请要求在2019年04月15日提交中国专利局、申请号为201910300243.X的中国专利申请的优先权，该申请的全部内容通过引用结合在本申请中。
技术领域
本申请涉及通信领域，具体而言，涉及一种模型的推送、模型的请求方法、装置、存储介质及电子装置。
背景技术
跟传统网络相比，5G网络中引入了软件定义网络(Software Defined Network，简称为SDN)、网络功能虚拟化(Network Function Virtualization，简称为NFV)等新技术，这些技术在增加网络灵活性的同时，也带来了管理和运维方面的复杂性，自动化、智能化的运营运维能力，将成为5G时代电信网络的刚需。人工智能技术在解决高计算量数据分析、跨领域特性挖掘、动态策略生成等方面，具备天然优势，将赋予5G时代网络运营运维新的模式和能力。
使用人工智能模型须考虑以下两方面：建模和应用。其中建模是指针对某种特定的分析需求，选择适当的算法，并通过历史数据的训练使得模型能够根据输入数据得到置信度较高的分析结果。应用是指在已有的模型中选择符合应用场景的模型，将其部署并运行在指定位置。利用运行中产生的分析数据可以实现模型参数的不断优化，使得模型能够随及时调整，维持模型推理的准确率。
目前在大数据及人工智能技术引入网络运维管理中，技术人员专注于人工智能建模部分，模型的部署和优化往往是针对单个模型的，对于大量现成模型的管理及其在网络中的应用(包括模型选择、部署及优化)则缺乏统一的方案。
针对相关技术中，如何从大量模型中选择需要的模型缺乏统一的技术方案等问题，尚未提出有效的解决方案。
发明内容
本申请实施例提供了一种模型的推送、模型的请求方法、装置、存储介质及电子装置，以至少解决相关技术中如何从大量模型中选择需要的模型缺乏统一的技术方案等问题。
根据本申请的一个实施例，提供了一种模型的推送方法，包括：接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；搜索与所述 模型匹配指令对应的模型；在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
根据本申请的另一个实施例，还提供了一种模型的请求方法，包括：基于分析需求生成模型匹配指令；将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
根据本申请的另一个实施例，还提供了一种模型的推送装置，包括：接收模块，用于接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；搜索模块，用于搜索与所述模型匹配指令对应的模型；推送模块，用于在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
根据本申请的另一个实施例，还提供了一种模型的请求装置，包括：生成模块，用于基于分析需求生成模型匹配指令；发送模块，用于将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
根据本申请的又一个实施例，还提供了一种存储介质，所述存储介质中存储有计算机程序，其中，所述计算机程序被设置为运行时执行上述任一项方法实施例中的步骤。
根据本申请的又一个实施例，还提供了一种电子装置，包括存储器和处理器，所述存储器中存储有计算机程序，所述处理器被设置为运行所述计算机程序以执行上述任一项方法实施例中的步骤。
通过本申请，由于接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；搜索与所述模型匹配指令对应的模型；在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点，因此，可以解决如何从大量模型中选择需要的模型缺乏统一的技术方案等问题，达到提供了如何选择模型的技术方案。
附图说明
此处所说明的附图用来提供对本申请的进一步理解，构成本申请的一部分，本申请的示意性实施例及其说明用于解释本申请，并不构成对本申请的不当限定。在附图中：
图1是根据本申请实施例的模型的推送方法的流程图；
图2是根据本申请实施例的模型的请求方法的流程图；
图3为根据本申请示例的模型推送系统的结构示意图；
图4是根据本申请实施例的模型的推送装置的结构框图；
图5是根据本申请实施例的模型的请求装置的结构框图；
图6为根据本申请可选实施例的资源分配的流程示意图；
图7为根据本申请可选实施例的模型选择流程的示意图。
具体实施方式
下文中将参考附图并结合实施例来说明本申请。需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
需要说明的是，本申请的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。
在本实施例中提供了一种模型的推送方法，图1是根据本申请实施例的模型的推送方法的流程图，如图1所示，该流程包括如下步骤：
步骤S102，接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成。
步骤S104，搜索与所述模型匹配指令对应的模型。
步骤S106，在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
通过本申请，由于接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；搜索与所述模型匹配指令对应的模型；在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点，因此，可以解决如何从大量模型中选择需要的模型缺乏统一的技术方案等问题，达到提供了如何选择模型的技术方案。
在本申请实施例中，搜索与所述模型匹配指令对应的模型，至少包括以下之一：在主模型平台搜索与所述模型匹配指令对应的模型；在所述主模型平台的联合模型平台搜索与所述模型匹配指令对应的模型；在所述主模型平台和所述联合模型平台搜索与所述模型匹配指令对应的模型，即可以在主模型平台或联合模型平台搜索，还可以将主模型平台和联合模型平台作为一个搜索库进行搜索。
在本申请实施例中，上述步骤S104在一个可选的实施例中可以通过以下方式实现：在搜索到与所述模型匹配指令对应的模型时，向所述编排器反馈搜索成功消息；在未搜索到与所述模型匹配指令对应的模型时，向所述编排器反馈 搜索失败消息。
在本申请实施例中，将搜索到的模型推送至需要所述模型的目的节点，包括：至少将所述模型，以及所述模型的元数据打包成文件；将打包的文件发送至所述目的节点，打包成的文件可以是docker文件，也可以是其他可执行文件，本申请实施例对此不作限定。
在本申请实施例中，在搜索到的模型为单个时，将单个的模型打包成文件发送至所述目的节点；在搜索到的模型为多个时，将对多个模型至少执行以下之一操作：编排组合，验证，将至少执行了以上之一操作的多个模型打包成文件发送至所述目的节点。
即模型引擎收到模型匹配指令后，根据编排器提供的联合类型确定模型搜索域，查找是否存在符合模型匹配指令的模型。若不存在，则向编排器发送模型匹配失败消息，若匹配结果为单个模型，则直接将模型打包为可部署的人工应用。若匹配为多个模型，则先完成模型的编排组合、验证，然后将编排信息文件与模型一起打包为可部署的人工应用。
在本申请实施例中，为了提高模型推送的准确率，所述方法还包括：接收编排器在检测到模型达到模型更新条件时发送的模型更新请求，其中，所述模型更新条件包括：推送至目的节点的模型准确率低于预设阈值；将与所述模型更新请求对应的模型发送至沙盒(SandBox)，其中，在所述沙盒中对所述模型更新请求对应的模型进行训练优化；接收所述沙盒训练成功后的模型。
在本申请实施例中，所述模型匹配指令至少包括以下参数之一：模型或算法属性，模型输入和输出要求，用于指示是否联合模型平台查找的指示信息，目的节点的信息。
本申请以上实施例的技术方案，涉及人工智能应用于网络通信，例如是下一代基于5G、SDN、NFV等技术的网络，以实现人工智能模型自动化管理并应用于网络的技术及其系统。
在本实施例中提供了一种模型的请求方法，图2是根据本申请实施例的模型的请求方法的流程图，如图2所示，该流程包括如下步骤：
步骤S202，基于分析需求生成模型匹配指令。
步骤S204，将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
通过本申请，由于将基于分析请求生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模 型时，将搜索到的模型推送至需要所述模型的目的节点，因此，可以解决如何从大量模型中选择需要的模型缺乏统一的技术方案等问题，达到提供了如何选择模型的技术方案。
在本申请实施例中，将生成的模型匹配指令发送至模型平台之后，所述方法还包括：接收所述模型平台在搜索到与所述模型匹配指令对应的模型时反馈的搜索成功消息；接收所述模型平台在未搜索到与所述模型匹配指令对应的模型时反馈的搜索失败消息。
在本申请实施例中，为了提高模型推送的准确率，所述方法还包括：在检测到模型达到模型更新条件时，向所述模型平台发送模型更新请求，以指示所述模型平台将与所述模型更新请求对应的模型发送至沙盒，并接收所述沙盒训练成功后的模型，其中，在所述沙盒中对所述模型更新请求对应的模型进行训练优化，所述模型更新条件包括：推送至目的节点的模型准确率低于预设阈值。
以下结合一示例对上述模型的推送，模型的请求过程进行说明，但不用于限定本申请实施例的技术方案。
本申请示例提供了一个能够与电信网络实时交互的人工智能平台系统，及通过这种交互实现在电信网络中自动选择、部署和优化人工智能模型的方法。
图3为根据本申请示例的模型推送系统的结构示意图，如图3所示，模型平台由模型引擎、模型库、沙盒等模块组成。
其中，模型库用于存储人工智能模型及其对应的元数据，模型的元数据是对模型的一系列描述性数据，包括模型名称、版本、所使用的编程语言和算法、运行时依赖、部署条件要求，以及对于输入数据的类别(如图像、数值)、内容及格式要求，对输出数据的类型描述等。
模型引擎用于管理模型库中的模型，包括对模型的增加、删除、修改、查询。模型引擎根据编排器下发的模型匹配指令选择人工智能模型，将模型推送至网络中的指定位置。其中，编排器用于管理网络中的人工智能应用实例，包括模型匹配指令、模型优化指令的下发及人工智能应用实例的生命周期管理、运行状态的监控等。网络中可以存在多个编排器，各编排器负责其对应的人工智能实例管理域。编排器为管理域中每一个的人工智能应用实例创建管理条目，管理条目包括人工智能应用实例的身份标识号(Identity Document，ID)、状态机、位置、模型信息等。
模型引擎的工作方式分为独立式和联合式。独立式是指模型引擎接收到模型匹配请求时，只能在其所在模型平台内的模型库查找模型。联合式中，与编排器直接交互的模型引擎所在的模型平台作为主平台，该主平台的模型引擎可 从联合模型平台的模型库获取模型。联合模型平台是指为主平台的模型引擎提供模型库的管理接口、不与编排器直接交互的模型平台。
沙盒用于为模型的再训练、验证和优化提供运行环境。
如图3所示，当通信网络中产生智能分析的需求时，编排器将智能分析的需求转化为模型要求，将模型匹配请求发送至模型平台的模型引擎，请求消息中包括编排器身份信息、模型要求及模型部署目标节点的节点类型、位置信息等。其中，编排器的身份信息包括编排器的类型、地址、联合类型等信息。模型要求包括模型的功能要求、性能要求等。
编排器身份信息中的联合类型根据模型引擎的工作方式分为无联合、主优先联合和对等联合三种。模型引擎的工作方式为独立式时，编排器身份信息中的联合类型只能为无联合，即仅能请求一个模型库的模型。模型引擎的工作方式为联合式时，编排器身份信息中的联合类型分为无联合、主优先联合和对等联合。其中，主优先联合是指优先从主平台的模型库(主模型库)中匹配模型，匹配失败后再从联合平台的模型库(联合模型库)中匹配。对等联合是指匹配时，主模型库和联合模型库同时作为搜索域。
模型引擎收到模型匹配指令后，根据编排器提供的联合类型确定模型搜索域，查找是否存在符合模型匹配指令的模型。若不存在符合模型匹配指令的模型，则向编排器发送模型匹配失败消息。若匹配结果为单个模型，则直接将模型打包为可部署的人工应用。若匹配为多个模型，则先完成模型的编排组合、验证，然后将编排信息文件与模型一起打包为可部署的人工应用。
模型引擎将人工智能应用及模型元数据连同编排器身份信息推送至模型应用目标节点。
编排器若收到模型匹配失败消息，则发出告警信息。若收到模型匹配成功信息，则预创建人工智能应用实例管理条目，条目中的信息为初始化状态，向模型部署目标节点发送人工智能应用实例创建请求。
模型部署目标节点收到应用实例创建请求后，验证该编排器的身份信息，若验证成功，则创建并启动实例，并向编排器发送实例创建成功消息，消息包括实例ID、实例运行信息等。其中，实例ID应能且仅能确定目标节点中的唯一实例。若验证失败，则发出非法访问告警。
编排器收到实例创建成功消息后，则据此激活并更新该实例的管理条目，包括将该实例的状态机改为运行态等。随后启动对实例状态的监控流程，并向目标节点发送模型更新策略。
实例处于运行态时，根据模型的输入获得推理结果，网络编排或控制平面 根据推理确定网络配置信息，配置到网络中，对网络进行优化。
实例处于运行态时，目标节点根据模型更新策略，将模型的输出按一定规则发送至大数据平台，作为此模型更新时的训练和测试数据。
编排器若监控到模型达到模型更新条件时(如模型准确度低于阈值)，则向模型引擎发送模型更新请求。
模型引擎根据模型更新请求选择待更新模型，并推送至沙盒中。
沙盒从大数据平台中提取最新的数据，进行模型的再训练和优化。
训练完毕后，沙盒将模型训练成功消息反馈给模型引擎。
模型引擎收到消息后，将更新后的模型保存至模型库，打包新的人工智能应用推送至目标节点，编排器实现目标节点中的新实例的部署，若部署成功，则发送原实例的销毁指令，使目标节点停止并删除原实例。若部署失败，则发出告警。
与相关技术相比较，本申请实施例针对人工智能应用于网络，提出了人工智能模型的自动化选择、部署和优化的架构和方法，实现了网络中人工智能模型的统一管理、提高了网络智能化的自主性。
通过以上的实施方式的描述，本领域的技术人员可以了解到根据上述实施例的方法可借助软件加必需的通用硬件平台的方式来实现，当然也可以通过硬件来实现。基于这样的理解，本申请的技术方案可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中，包括若干指令用以使得一台终端设备(可以是手机，计算机，服务器，或者网络设备等)执行本申请各个实施例所述的方法。
在本实施例中还提供了一种模型的推送装置，该装置用于实现上述实施例及可选实施方式，已经进行过说明的不再赘述。如以下所使用的，术语“模块”可以实现预定功能的软件和/或硬件的组合。尽管以下实施例所描述的装置可以以软件来实现，但是硬件，或者软件和硬件的组合的实现也是可能并被构想的。
图4是根据本申请实施例的模型的推送装置的结构框图，如图4所示，该装置包括：接收模块40，用于接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；搜索模块42，用于搜索与所述模型匹配指令对应的模型；推送模块44，用于在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
在本申请实施例中，搜索模块42，还用于至少以下之一：在主模型平台搜索与所述模型匹配指令对应的模型；在所述主模型平台的联合模型平台搜索与 所述模型匹配指令对应的模型；在所述主模型平台和所述联合模型平台搜索与所述模型匹配指令对应的模型。
在本申请实施例中，搜索模块42，还用于：在搜索到与所述模型匹配指令对应的模型时，向所述编排器反馈搜索成功消息；在未搜索到与所述模型匹配指令对应的模型时，向所述编排器反馈搜索失败消息。
在本申请实施例中，推送模块44，还用于：至少将所述模型，以及所述模型的元数据打包成文件；将打包的文件发送至所述目的节点。
在本申请实施例中，推送模块44，还用于：在搜索到的模型为单个时，将单个的模型打包成文件发送至所述目的节点；在搜索到的模型为多个时，将对多个模型至少执行以下之一操作：编排组合，验证，将至少执行了以上之一操作的多个模型打包成文件发送至所述目的节点。
在本申请实施例中，推送模块44，还用于：接收编排器在检测到模型达到模型更新条件时发送的模型更新请求，其中，所述模型更新条件包括：推送至目的节点的模型准确率低于预设阈值；将与所述模型更新请求对应的模型发送至沙盒，其中，在所述沙盒中对所述模型更新请求对应的模型进行训练优化；接收所述沙盒训练成功后的模型。
在本申请实施例中，所述模型匹配指令至少包括以下参数之一：模型或算法属性，模型输入和输出要求，用于指示是否联合模型平台查找的指示信息，目的节点的信息。
在本实施例中还提供了一种模型的请求装置，该装置用于实现上述实施例及可选实施方式，已经进行过说明的不再赘述。如以下所使用的，术语“模块”可以实现预定功能的软件和/或硬件的组合。尽管以下实施例所描述的装置可以以软件来实现，但是硬件，或者软件和硬件的组合的实现也是可能并被构想的。
图5是根据本申请实施例的模型的请求装置的结构框图，如图5所示，该装置包括：生成模块50，用于基于分析需求生成模型匹配指令；发送模块52，用于将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
通过本申请，由于将基于分析请求生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点，因此，可以解决如何从大量模型中选择需要的模型缺乏统一的技术方案等问题，达到提供了如何选择模型的技术方案。
在本申请实施例中，发送模块52，还用于：接收所述模型平台在搜索到与所述模型匹配指令对应的模型时反馈的搜索成功消息；接收所述模型平台在未搜索到与所述模型匹配指令对应的模型时反馈的搜索失败消息。
在本申请实施例中，发送模块52，还用于：在检测到模型达到模型更新条件时，向所述模型平台发送模型更新请求，以指示所述模型平台将与所述模型更新请求对应的模型发送至沙盒，并接收所述沙盒训练成功后的模型，其中，在所述沙盒中对所述模型更新请求对应的模型进行训练优化，所述模型更新条件包括：推送至目的节点的模型准确率低于预设阈值。
需要说明的是，上述各个模块是可以通过软件或硬件来实现的，对于后者，可以通过以下方式实现，但不限于此：上述模块均位于同一处理器中；或者，上述各个模块以任意组合的形式分别位于不同的处理器中。
以下结合有可选实施例对上述实施例以及示例的技术方案进行说明，但不用于限定本申请实施例的技术方案。
以下所述模型是指通过机器学习、深度学习等方法，从已有的数据或经验中学习得到的实现特定的分析、预测等功能的系统。
具体而言，每个模型都有其制定的功能。例如，某些模型可以用于预测新建用户以及会话的数目达到所需数量的时间。又如某些模型可以用于预测进行切片扩容的时间。此外，还可以是根据告警信息的数量或者类型确定设备中告警的位置。同时各个模型之间也是存在关联的。例如，用于预测新建用户以及会话的数目达到所需数量的时间，需要作为预测进行切片扩容的输出。因此，模型与模型之间可以通过串联连接，同时还可以并联连接，或者混联连接。具体模型的功能需根据应用实例的功能来决定。
可选实施例一：图6示出了一个运用人工智能实现5G核心网切片智能资源分配的流程，如下所示。
S502，编排器向模型引擎下发模型匹配请求，要求模型基于未来10分钟后的流量实现切片的自动缩放；模型的准确率达到90％；模型的部署网元类型是网络数据分析功能网元(Network Data Analysis Function，简称为NWDAF)；部署网元的IP地址是192.168.1.107；联合类型是无联合。
S504，模型引擎从模型库中查找到可用模型：ID为1001的卷积神经网络模型用于网路流量预测；ID为1002的强化学习模型用于根据各切片流量信息、服务质量(Quality of Service，QoS)要求等实现切片的智能资源分配。
部署准备包括：根据需求完成两个模型的串联；将串联后的模型、模型所依赖的运行时环境、元数据打包成镜像名称为Image1的docker镜像。
S506，模型引擎将Image1连同编排器的身份信息，发送至192.168.1.107对应的NWDAF网元，NWDAF反馈打包文件接收成功响应。
S508，模型引擎通知编排器模型匹配已完成，编排器预创建状态为初始态的管理条目。
S510，通过编排器向192.168.1.107发送Image1运行请求及模型更新策略(当模型1001的准确率降至80％以下时，执行模型更新，此处的准确率可以理解为模型引擎所搜索到的模型符合模型匹配指令所需要的模型的次数与模型匹配指令共请求的模型的次数的比值)。
S512，NWDAF验证此编排器与Image1携带的编排器身份信息一致，基于Image1镜像新建并启动名为Instance1(实例1)的服务，并通知编排器。编排器激活人工智能Instance1的管理条目，设置实例1的状态机为运行态；并且，每60秒发送保活报文以确认实例1的状态；每隔一天发送准确率统计请求以判断是否执行模型更新。
S514，Instance1根据元数据中的数据输入要求，接收网络中的相应数据，输出分析结果，分组控制功能(Packet Control Function，简称为PCF)基于此结果确定切片资源分配策略，由NSMF下发至网络子网切片管理功能网元(Network Subnet Slice Management Function，简称NSSMF)，通过网络功能虚拟化编排器(Network Function Virtualisation Orchestrator，简称为NFVO)执行该策略，实现切片资源分配。输出的结果在固定时间(可以是每天午夜)打包发送至大数据平台。
S516，编排器得到通知，当卷积神经网络模型1001的准确率降低至75％。编排器向模型引擎发送模型1001的更新请求。
S518，模型引擎为卷积神经网络模型1001选择包含1个GPU的训练环境作为沙盒。该沙盒取出大数据平台中存储的最新的Instance1的10万条分析数据，其中8万条作为训练数据，2万条作为测试数据，对卷积神经网络模型1001进行再训练和验证。
S520，接收模型训练完毕通知，对卷积神经网络模型1001进行验证后，发现卷积神经网络模型1001的准确率达到90％。
S522，模型引擎将卷积神经网络模型1001存储至模型库。
接下来的流程见步骤S502-S520。其中，编排器接收到Model 1001的新实例Instance2进入运行态的消息后，即向192.168.1.107发送Instance1的销毁指令，该节点停止并清除Instance1。
可选实施例二：图7所示出了网络故障根因分析模型应用于网络中，实现 故障的精准定位和溯源的流程图。
步骤S602，编排器向模型引擎下发模型匹配请求，要求模型根据网管系统同一时刻的告警信息定位根源告警信息，模型的部署位置是网管系统，联合模式为优先本地模型库。
步骤S604，模型引擎从主模型库中查找适配模型失败，在联合模型库中查找到适配模型，选择此模型。
步骤S606，模型引擎从联合模型库中查找适配模型成功，在联合模型库中查找到适配模型，返回匹配模型。
步骤S608，模型引擎将模型、模型所依赖的运行时环境、元数据打包成镜像名称为Image1的docker镜像，连同编排器的身份信息，发送至网管平台，并接收网管平台反馈的文件接收成功消息。
步骤S612，模型引擎通知编排器模型匹配已完成，编排器向网管平台发送Image1运行请求及模型更新策略(不更新)。
步骤S614，网管系统验证此编排器与Image1携带的编排器身份信息一致，基于Image1镜像新建并启动名为Instance1的容器，并通知编排器。编排器激活Instance1的管理条目，设置实例1的状态机为运行态。
步骤S616，Instance1根据元数据中的数据输入要求，接收网络中的相应数据，向PCF输出告警分析结果，PCF据此制定告警处理策略。
本申请的实施例还提供了一种存储介质，该存储介质中存储有计算机程序，其中，该计算机程序被设置为运行时执行上述任一项方法实施例中的步骤。
可选地，在本实施例中，上述存储介质可以被设置为存储用于执行以下步骤的计算机程序：
S1，接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成。
S2，搜索与所述模型匹配指令对应的模型。
S3，在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
可选地，在本实施例中，上述存储介质可以被设置为存储用于执行以下步骤的计算机程序：
S1，基于分析需求生成模型匹配指令。
S2，将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与 所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
可选地，在本实施例中，上述存储介质可以包括但不限于：U盘、只读存储器(Read-Only Memory，简称为ROM)、随机存取存储器(Random Access Memory，简称为RAM)、移动硬盘、磁碟或者光盘等各种可以存储计算机程序的介质。
本申请的实施例还提供了一种电子装置，包括存储器和处理器，该存储器中存储有计算机程序，该处理器被设置为运行计算机程序以执行上述任一项方法实施例中的步骤。
可选地，上述电子装置还可以包括传输设备以及输入输出设备，其中，该传输设备和上述处理器连接，该输入输出设备和上述处理器连接。
可选地，在本实施例中，上述处理器可以被设置为通过计算机程序执行以下步骤：
S1，接收编排器的模型匹配指令，其中，所述模型匹配指令基于分析需求生成。
S2，搜索与所述模型匹配指令对应的模型。
S3，在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
可选地，本实施例中的具体示例可以参考上述实施例及可选实施方式中所描述的示例，本实施例在此不再赘述。
可选地，在本实施例中，上述处理器可以被设置为通过计算机程序执行以下步骤：
S1，基于分析需求生成模型匹配指令。
S2，将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型时，将搜索到的模型推送至需要所述模型的目的节点。
本领域的技术人员应该明白，上述的本申请的各模块或各步骤可以用通用的计算装置来实现，它们可以集中在单个的计算装置上，或者分布在多个计算装置所组成的网络上，可选地，它们可以用计算装置可执行的程序代码来实现，从而，可以将它们存储在存储装置中由计算装置来执行，并且在某些情况下， 可以以不同于此处的顺序执行所示出或描述的步骤，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。这样，本申请不限制于任何特定的硬件和软件结合。
以上所述仅为本申请的实施例而已，并不用于限制本申请，对于本领域的技术人员来说，本申请可以有各种更改和变化。凡在本申请的原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。

Claims (14)

  1. 一种模型的推送方法，包括：
    接收编排器发送的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；
    搜索与所述模型匹配指令对应的模型；
    在搜索到所述模型的情况下，将搜索到的模型推送至需要所述模型的目的节点。
  2. 根据权利要求1所述的方法，其中，所述搜索与所述模型匹配指令对应的模型，至少包括以下之一：
    在主模型平台搜索与所述模型匹配指令对应的模型；
    在所述主模型平台的联合模型平台搜索与所述模型匹配指令对应的模型；
    在所述主模型平台和所述联合模型平台搜索与所述模型匹配指令对应的模型。
  3. 根据权利要求1所述的方法，还包括：
    在搜索到与所述模型匹配指令对应的模型的情况下，向所述编排器反馈搜索成功消息；
    在未搜索到与所述模型匹配指令对应的模型的情况下，向所述编排器反馈搜索失败消息。
  4. 根据权利要求1所述的方法，其中，所述将搜索到的模型推送至需要所述模型的目的节点，包括：
    至少将所述模型，以及所述模型的元数据打包成文件；
    将打包的文件发送至所述目的节点。
  5. 根据权利要求4所述的方法，还包括：
    在搜索到的模型为单个的情况下，将单个的模型打包成文件并将所述文件发送至所述目的节点；
    在搜索到的模型为多个的情况下，对多个模型执行目标操作，将执行所述目标操作后的多个模型打包成文件并将所述文件发送至所述目的节点，其中，所述目标操作包括以下至少之一：编排和组合；验证。
  6. 根据权利要求1所述的方法，还包括：
    接收编排器在检测到模型达到模型更新条件的情况下发送的模型更新请求，其中，所述模型更新条件包括：推送至所述目的节点的模型准确率低于预 设阈值；
    将与所述模型更新请求对应的模型发送至沙盒，其中，所述模型更新请求对应的模型在所述沙盒中进行训练；
    接收所述沙盒发送的训练成功后的模型。
  7. 根据权利要求1-6任一项所述的方法，其中，所述模型匹配指令至少包括以下参数之一：模型或算法属性，模型输入和输出要求，用于指示是否在联合模型平台查找的指示信息，目的节点的信息。
  8. 一种模型的请求方法，包括：
    基于分析需求生成模型匹配指令；
    将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型的情况下，将搜索到的模型推送至需要所述模型的目的节点。
  9. 根据权利要求8所述的方法，将生成的模型匹配指令发送至模型平台之后，还包括：
    接收所述模型平台在搜索到与所述模型匹配指令对应的模型的情况下反馈的搜索成功消息；
    接收所述模型平台在未搜索到与所述模型匹配指令对应的模型的情况下反馈的搜索失败消息。
  10. 根据权利要求8所述的方法，还包括：
    在检测到模型达到模型更新条件的情况下，向所述模型平台发送模型更新请求，以指示所述模型平台将与所述模型更新请求对应的模型发送至沙盒，并接收所述沙盒发送的训练成功后的模型，其中，所述模型更新请求对应的模型在所述沙盒中进行训练，所述模型更新条件包括：推送至所述目的节点的模型准确率低于预设阈值。
  11. 一种模型的推送装置，包括：
    接收模块，设置为接收编排器发送的模型匹配指令，其中，所述模型匹配指令基于分析需求生成；
    搜索模块，设置为搜索与所述模型匹配指令对应的模型；
    推送模块，设置为在搜索到所述模型的情况下，将搜索到的模型推送至需要所述模型的目的节点。
  12. 一种模型的请求装置，包括：
    生成模块，设置为基于分析需求生成模型匹配指令；
    发送模块，设置为将生成的模型匹配指令发送至模型平台，以指示所述模型平台搜索与所述模型匹配指令对应的模型，并在搜索到所述模型的情况下，将搜索到的模型推送至需要所述模型的目的节点。
  13. 一种存储介质，存储有计算机程序，所述计算机程序被设置为运行时执行所述权利要求1至10任一项所述的方法。
  14. 一种电子装置，包括存储器和处理器，所述存储器中存储有计算机程序，所述处理器被设置为运行所述计算机程序以执行所述权利要求1至10任一项所述的方法。
PCT/CN2020/084470 2019-04-15 2020-04-13 模型的推送、模型的请求方法、装置、存储介质及电子装置 WO2020211722A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP20790513.4A EP3958606A4 (en) 2019-04-15 2020-04-13 METHODS AND DEVICES FOR PUSHING AND REQUEST MODEL, STORAGE MEDIA AND ELECTRONIC DEVICE
US17/603,622 US20220197953A1 (en) 2019-04-15 2020-04-13 Model pushing method and device, model requesting method and device, storage medium and electronic device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910300243.XA CN111083722B (zh) 2019-04-15 模型的推送、模型的请求方法及装置、存储介质
CN201910300243.X 2019-04-15

Publications (1)

Publication Number Publication Date
WO2020211722A1 true WO2020211722A1 (zh) 2020-10-22

Family

ID=70310304

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/084470 WO2020211722A1 (zh) 2019-04-15 2020-04-13 模型的推送、模型的请求方法、装置、存储介质及电子装置

Country Status (3)

Country Link
US (1) US20220197953A1 (zh)
EP (1) EP3958606A4 (zh)
WO (1) WO2020211722A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106533714A (zh) * 2015-09-09 2017-03-22 中兴通讯股份有限公司 重新实例化虚拟网络功能的方法和装置
CN107707381A (zh) * 2017-08-04 2018-02-16 北京天元创新科技有限公司 虚拟网元智能切片管理系统及方法
US20180131621A1 (en) * 2016-11-07 2018-05-10 Hewlett Packard Enterprise Development Lp Middlebox modeling
CN109218046A (zh) * 2017-06-30 2019-01-15 中国移动通信有限公司研究院 网络切片的管理方法及系统和存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106156127B (zh) * 2015-04-08 2020-06-16 深圳市腾讯计算机系统有限公司 选择数据内容向终端推送的方法及装置
US10685332B2 (en) * 2016-06-24 2020-06-16 Intel Corporation Contextual model-based event scheduling
US10310821B2 (en) * 2017-06-03 2019-06-04 Apple Inc. Integration of learning models into a software development system

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106533714A (zh) * 2015-09-09 2017-03-22 中兴通讯股份有限公司 重新实例化虚拟网络功能的方法和装置
US20180131621A1 (en) * 2016-11-07 2018-05-10 Hewlett Packard Enterprise Development Lp Middlebox modeling
CN109218046A (zh) * 2017-06-30 2019-01-15 中国移动通信有限公司研究院 网络切片的管理方法及系统和存储介质
CN107707381A (zh) * 2017-08-04 2018-02-16 北京天元创新科技有限公司 虚拟网元智能切片管理系统及方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3958606A4 *

Also Published As

Publication number Publication date
CN111083722A (zh) 2020-04-28
EP3958606A1 (en) 2022-02-23
US20220197953A1 (en) 2022-06-23
EP3958606A4 (en) 2023-01-25

Similar Documents

Publication Publication Date Title
US9201644B2 (en) Distributed update service
CN109816116A (zh) 机器学习模型中超参数的优化方法及装置
US20180234308A1 (en) Service orchestration
US10453165B1 (en) Computer vision machine learning model execution service
Mennes et al. GRECO: A distributed genetic algorithm for reliable application placement in hybrid clouds
CN114039858B (zh) 一种算网资源融合方法、装置、设备及存储介质
US10970127B1 (en) Systems and methods for virtual machine resource optimization using machine learning techniques
CN109815025A (zh) 业务模型调用方法、装置及存储介质
US11983189B2 (en) Data pipeline controller
CN113934763A (zh) 分布式数据库的sql查询方法及装置
US20230071201A1 (en) Data blending for multiple data pipelines
CN110297713A (zh) 一种云主机的配置管理系统和方法
CN114296933A (zh) 端边云架构下轻量化容器的实现方法、数据处理系统
CN115168031A (zh) 一种雾计算系统、方法、电子设备及存储介质
CN112995303B (zh) 跨集群调度方法及装置
WO2020211722A1 (zh) 模型的推送、模型的请求方法、装置、存储介质及电子装置
CN114693220B (zh) 基于数字孪生DaaS平台的算法仓库管理方法及系统
WO2022166715A1 (zh) 一种智能流水线处理方法、装置、存储介质及电子装置
CN111083722B (zh) 模型的推送、模型的请求方法及装置、存储介质
US20220413896A1 (en) Selecting a node group of a work group for executing a target transaction of another work group to optimize parallel execution of steps of the target transaction
WO2022016969A1 (zh) 一种数据处理方法及装置
US11775819B2 (en) Automated configuration determinations for data center devices using artificial intelligence techniques
US20220019460A1 (en) Dynamically adjusting reconciliation time-delays
Tomczak et al. Development of service composition by applying ICT service mapping
CN113852479B (zh) 一种安全网络构建方法、装置、设备和计算机存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20790513

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2020790513

Country of ref document: EP

Effective date: 20211115