JP7636080B2 - AI model building platform, AI model building method and program - Google Patents

Description

The present invention relates to an AI model construction platform and an AI model construction method and program, and in particular to an AI model construction platform and an AI model construction method and program for constructing an AI model to be applied to a commercial network on a relatively small-scale verification network.

Non-patent document 1 discloses the concept of digital twins as a technology for collecting information in real space and simulating it in digital space, and consideration of its use in network areas such as 5G is also accelerating.

By using a digital twin to reproduce a commercial network in a verification network, it becomes easier to conduct tests that are difficult to carry out on a commercial network and it becomes possible to collect a large amount of data, which makes it possible to build AI and machine learning learning models that are essential for autonomous network operation.

IETF Digital Twin Network: Concepts and Reference Architecture https://datatracker.ietf.org/doc/draft-zhou-nmrg-digitaltwin-network-concepts/

However, when reproducing a commercial network, which is a global communications network, on a verification network, in principle, server and network resources (e.g., CPU, memory, storage capacity, network bandwidth) equivalent to or greater than those of a commercial network would be required, resulting in enormous equipment costs. As a result, the problem of being unable to build a verification network for building AI models arises.

In addition, since there is no integrated platform where commercial networks can request, train, and manage AI models for network operations, training must be carried out separately for each network, resulting in inefficient management of AI models.

Network operations using AI have been attracting a great deal of attention in recent years, from research to practical use. However, thanks to the benefits of virtualization and containerization, network services can now be deployed at high speed, resulting in a wide variety of network conditions. Therefore, it is necessary to build AI models that are trained specifically for each of these diverse network environments.

There are two elements to building an AI model on a network: the "network configuration" and the "traffic applied to the network." The network configuration includes the CPU/Memory scale of each node that makes up the network, the amount of network bandwidth connected, etc., and represents the underlying network itself.

Traffic includes the total number of IP packets generated when users of the network make requests to the network (e.g., web browsing, connection to 5GC, handover). In order to build an AI model specialized for a commercial network, the above two elements must be the same as those of a commercial network.

However, in order to build a network configuration and traffic volume that is exactly the same as a commercial network, it is necessary to have equipment of exactly the same scale as the commercial network, which doubles the equipment costs. In addition, in recent years, a concept known as chaos engineering has emerged in which random failures are intentionally generated on a commercial network to check the reliability of the system and build AI models. However, this concept makes it difficult to build AI models before the start of commercial services, and there is also the possibility that telecommunications carriers may intentionally drop emergency calls that users are currently making, making it difficult to implement from a legal perspective.

The object of the present invention is to solve the above technical problems, and to provide an AI model construction platform and an AI model construction method and program for constructing a verification network that is smaller than a commercial network in a verification environment, and constructing an AI model to be applied to the commercial network on the small-scale verification network.

In order to achieve the above object, the present invention is characterized in that an AI model construction platform that constructs an AI model to be applied to a commercial network on a verification network includes a means for acquiring network information of the commercial network, a means for acquiring computational resource information available in the verification environment, a means for constructing a verification network in the verification environment based on the network information and the computational resource information, a means for determining test scenario information that needs to be implemented to construct an AI model based on the network information, and a means for constructing an AI model based on the results of testing on the verification network using the test scenario information.

The present invention achieves the following effects:

(1) It will be possible to build a platform that dynamically manages AI models related to network operations, enabling autonomous network operations in commercial networks.

(2) Building a platform that dynamically manages AI models related to network operations makes it possible to manage AI models, and enables AI models that take a long time to train to be applied to other similar networks. As a result, the time required to build an AI model is reduced, and the validation network can be used more efficiently.

(3) By dynamically constructing a verification network for building AI models, it becomes unnecessary to simulate failures on a commercial network, which is essential for building AI models. This makes it possible to build AI models without affecting users of the commercial network.

1 is a functional block diagram showing a configuration of a network verification system according to an embodiment of the present invention. 13 is a sequence flow showing a procedure for constructing a verification network NW2. FIG. 1 is a functional block diagram of a network operation AI model manager. 13 is a sequence flow showing the operation of a network operation AI model manager. FIG. 2 is a functional block diagram showing the configuration of a verification network construction node. 13 is a sequence flow showing the operation of a verification network construction node. FIG. 1 is a functional block diagram showing the configuration of the main parts of an AI model training node. 1 is a sequence flow illustrating the operation of an AI model training node. FIG. 1 is a diagram illustrating an example of a relationship between a commercial network and a verification network.

The following describes in detail an embodiment of the present invention with reference to the drawings. FIG. 1 is a functional block diagram showing the configuration of an AI model construction platform according to one embodiment of the present invention. In this embodiment, a verification network NW2 is automatically constructed from a commercial network NW1 based on the resource status available in the verification environment, an AI model is dynamically created from the results of various tests performed on the constructed verification network NW2, and the AI model is finally reflected in the commercial network NW1.

In this embodiment, we focus on cloud-native architecture, which has also been adopted in recent Web applications and mobile cores (5GC). In this architecture, instead of launching a small number of huge monolithic network instances to process requests from a large number of users, we adopt a scale-out configuration in which a large number of uniform network instances are arranged in parallel and requests from a large number of users are uniformly load-balanced across the parallel network instances.

For example, if the commercial network NW1 uses 100 instances with 16 CPU cores and 160 GB of memory to process 1 million users, then the verification network NW2 can be configured to run one instance with 16 CPU cores and 160 GB of memory and run simulated traffic of 1 million/100 = 10,000 users to recreate an environment equivalent to the commercial network NW1.

The network operation AI model manager 1 manages the configuration information of the commercial network NW1 and the verification network NW2 and applies the AI model. The verification network construction node 2 constructs the verification network. The AI model training node 3 trains and manages the AI model.

Such an AI model construction platform can be constructed by implementing applications (programs) that realize the functions of the network operation AI model manager 1, the verification network construction node 2, and the AI model training node 3 on at least one general-purpose computer or server equipped with a CPU, ROM, RAM, bus, interface, etc. Alternatively, it can be constructed as a dedicated machine or a single-function machine in which part of the application is implemented as hardware or software.

Figure 2 shows a sequence flow showing the procedure for constructing the verification network NW2. At time t1, the commercial network NW1 sends its own network information to the network operation AI model manager 1 to request the construction of an AI model.

The network information of the commercial network NW1 includes the information necessary for constructing an AI model, such as (1) network configuration information (single network instance information, total number of network instances, configuration software information (name of software used, version)), (2) accommodated user information (user type, type of service used by user, number of accommodated users), (3) network statistical information (traffic volume, number of requests, number of sessions, number of handovers), and (4) required AI model information (AI type, required accuracy).

The network operation AI model manager 1 checks whether an AI model that matches the requested AI model already exists in its database, and if not, at time t2 it requests the verification network construction node 2 to construct a verification network. This request includes network information about the commercial network NW1.

At time t3, the verification network construction node 2 requests the verification environment (NW2) that constructs the verification network NW2 for computational resource information that can be used to construct the verification network NW2. At time t4, in response to the request, the verification environment (NW2) transmits computational resource information that can be used in the verification environment (NW2) to the verification network construction node 2.

When the verification network construction node 2 acquires the computational resource information from the verification environment (NW2), at time t5, it determines network information of the verification network NW2 for reproducing the commercial network NW1 in constructing the requested AI model. At time t6, it requests the verification environment (NW2) to construct the verification network NW2 for constructing the AI model.

In this embodiment, it is required to construct a verification network NW2 that is relatively smaller than the commercial network NW1. For example, if the commercial network NW1 is configured by arranging multiple uniform network instances in parallel, it is required to construct a verification network NW2 that has a smaller number of network instances than the commercial network NW1.

The verification environment (NW2) constructs the requested relatively small verification network NW2, and sends a completion notification to the verification network construction node 2 at time t7. At time t8, the verification network construction node 2 notifies the network operation AI model manager 1 that construction of the verification network NW2 for building the AI model has been completed. At time t9, the network operation AI model manager 1 requests the AI model training node 3 to train the AI model on the verification network NW2.

The AI model training node 3 determines the test scenario information that needs to be implemented to build the requested AI model, and transmits the test scenario information for creating the AI model to the verification network NW2 at time t10. This test scenario information includes (a) test traffic information (e.g., the amount of transmitted traffic, the number of transmitted requests) and (b) simulated failure information (e.g., the type of simulated failure event, the time when the simulated failure occurred, and the frequency of occurrence of the simulated failure).

The test traffic information is obtained based on the accommodated user information and network statistical information deflated at a ratio according to the scale of the verification network NW2 relative to the commercial network NW1 when the network operations AI model manager 1 requests training of the AI model at the time t9. The simulated failure information is determined when the AI model training node 3 requests execution of a test at the time t10. For example, if the requested AI model information includes a statement that "we want to detect IF failures," simulated failure information is determined to be to execute IF failures 100 times.

The verification network NW2 performs the requested test based on the test scenario information, and transmits the test results to the AI model training node 3 at time t11. The test results include (c) network statistical information (e.g., resource utilization rate and number of events for each device and network instance) and (d) test scenario information (e.g., amount of transmitted and received traffic and number of successful/failed requests) as information necessary for constructing the AI model.

At time t12, the AI model training node 3 creates an AI model using the test results received from the verification network NW2, and upon confirming that it meets the AI type and accuracy required by the commercial network NW1, it sends information about the created AI model to the network operations AI model manager 1 at time t13.

The network operation AI model manager 1 transmits the AI model received from the AI model training node 3 to the commercial network NW1 at time t14. The commercial network NW1 applies the AI model received from the network operation AI model manager 1 to each instance of the commercial network NW1.

At time t15, the commercial network NW1 starts network operation using an AI model for predicting future network metric values (e.g., CPU utilization) and detecting network anomalies. If improvements to the AI model are required after operation begins for reasons such as improving required accuracy, the AI model is updated using data actually obtained from the commercial network NW1. The commercial network NW1 sends the AI model updated at time t15 to the network operation AI model manager 1 at time t16.

Figure 3 is a functional block diagram of the network operation AI model manager 1, whose main components are a request receiving unit 101, an AI model management unit 102, and an AI model database (DB) 103, and it manages the configuration information of the commercial network NW1 and the verification network NW2, as well as AI model requests from the commercial network NW1.

The request receiving unit 101 accepts the network configuration information and request contents received from the commercial network NW1. The AI model DB 103 stores the AI model created by the verification network NW2. The AI model management unit 102 manages the overall processing of the network operation AI model manager 1, and requests the verification network construction node 2 and the AI model training node 3 to create an AI model.

Figure 4 is a sequence flow showing the operation of the network operation AI model manager 1 in the sequence of Figure 2. When the network operation AI model manager 1 receives a request to build an AI model including network information from the commercial network NW1 at time t1, the request receiving unit 101 accepts the build request at time t1_1 and notifies the AI model management unit 102 of the request contents.

The AI model management unit 102 checks the network information of the requested commercial network NW1, and at time t1_2, queries the AI model DB 103 as to whether an AI model for the requested network already exists. The AI model DB 103 checks its own database, and returns the query result to the AI model management unit 102 at time t1_3.

If an AI model already exists, the AI model management unit 102 transmits it to the commercial network NW1, but if the AI model does not exist, it requests the verification network construction node 2 to construct the verification network NW2 at time t1_4 (corresponding to the time t2) and further requests the AI model training node 3 to create an AI model in the verification network NW2 at the time t9.

The AI model training node 3 transmits the created AI model to the AI model management unit 102 of the network operation AI model manager 1 at time t1_5 (corresponding to t13 above). At time t1_6, the AI model management unit 102 stores the received AI model in the AI model DB 103 together with network information of the commercial network NW1 that uses the AI model. At time t1_7, the AI model management unit 102 transmits the created AI model to the commercial network NW1 for application.

When the AI model is updated in the commercial network NW1, the commercial network NW1 sends the updated AI model to the network operation AI model manager 1 at time t1_8. When the request receiving unit 101 receives the updated AI model from the commercial network NW1, it stores it in the AI model DB 103 at time t1_9.

Figure 5 is a functional block diagram showing the configuration of the verification network construction node 2, which mainly consists of a verification network configuration management unit 201, a verification network information acquisition unit 202, and a verification network construction unit 203, and constructs a verification network NW2 in the verification environment (NW2) to be used for constructing an AI model.

The verification network information acquisition unit 202 acquires configuration information of the verification environment (NW2). The verification network construction unit 203 constructs the verification network NW2 in the verification environment (NW2). The verification network configuration management unit 201 manages the overall processing within the verification network construction node 2 and determines the configuration information of the verification network NW2.

Figure 6 is a sequence flow showing the operation of the verification network construction node 2 in the sequence of Figure 2. At time t2, the network operation AI model manager 1 requests the verification network configuration management unit 201 of the verification network construction node 2 to construct a verification network. At time t2_1, the verification network configuration management unit 201 requests the verification network information acquisition unit 202 for information on the computational resources available in the verification environment (NW2).

At time t2_2 (corresponding to the time t3 mentioned above), the verification network information acquisition unit 202 requests the verification environment (NW2) for information on computational resources available in the verification network NW2. At time t2_3 (corresponding to the time t4 mentioned above), the verification environment (NW2) transmits information on computational resources available in the verification environment (NW2) to the verification network information acquisition unit 202 of the verification network construction node 2.

At time t2_4, the verification network information acquisition unit 202 transmits information on the computational resources available in the verification environment (NW2) to the verification network configuration management unit 201. At time t2_5, the verification network configuration management unit 201 creates network configuration information for the verification network NW2 based on the requested network information on the commercial network NW1 and the computational resource information available in the verification environment (NW2).

In this embodiment, as shown in an example in Figure 9, the number of instances in the commercial network NW1 is reduced from 100 to 2 in the verification network NW2, making the verification network NW2 1/50 the size of the commercial network NW1.

At time t2_6, the verification network configuration management unit 201 sends the network configuration information of the created verification network NW2 to the verification network construction unit 203, requesting construction of the verification network NW2. At time t2_7 (corresponding to the time t6), the verification network construction unit 203 constructs the verification network NW2 in the verification environment (NW2).

At time t2_8, the verification environment (NW2) constructs a verification network NW2 at 1/50 scale, and at time t2_9 (t7 above) notifies the verification network construction unit 203 of the verification network construction node 2.

At time t2_10, the verification network construction unit 203 notifies the verification network configuration management unit 201 of the completion of construction of the verification network NW2. At time t2_11 (corresponding to the time t8), the verification network construction management unit 201 notifies the network operation AI model manager 1 of the completion of construction of the verification network NW2.

Figure 7 is a functional block diagram showing the configuration of the main parts of the AI model training node 3, which mainly consists of an AI model training management unit 301, an AI model training unit 302, and a measurement data DB 303, and which trains and evaluates the AI model in the verification network NW2.

The AI model training unit 302 creates test scenarios and trains and evaluates the AI model. The measurement data DB 303 stores the test results of the AI model and the verification network. The AI model training management unit 301 manages the entire AI model training node 3, accepts requests from the network operation AI model manager 1, and transmits the created AI model.

Figure 8 is a sequence flow showing the operation of the AI model training node 3 in the sequence of Figure 2. At time t9, the network operation AI model manager 1 requests the AI model training node 3 to train the AI model. At time t9_1, the AI model training management unit 301 requests the AI model training unit 302 to train the AI model on the verification network NW2.

At time t9_2, the AI model training unit 302 determines and constructs test scenario information that needs to be implemented to construct the requested AI model, for example, as shown in an example in FIG. 9. In this embodiment, in accordance with the network being scaled down to 1/50, the traffic volume of the test scenario is reduced from 100 Gbps to 2 Gbps, the number of requests from 10,000/sec to 200/sec, the number of handovers from 50,000/sec to 100/sec, and the number of simultaneous sessions from 500,000 to 10,000.

At time t9_3 (corresponding to time t10), the AI model training unit 302 transmits test scenario information to the verification network NW2 to request the start of a test. The verification network NW2 performs the requested test and transmits the test results to the AI model training unit 302 at time t9_4 (corresponding to time t11).

At time t9_5, the AI model training unit 302 stores the test results in the measurement data DB 303. At time t9_6, the AI model training unit 302 uses the test results received from the verification network NW2 to train the AI model. At time t9_7, the AI model training unit 302 evaluates the accuracy of the trained AI model.

If the evaluation result does not satisfy the required accuracy of the AI model included in the network information of the commercial network NW1, the processes at times t9_2 to t9_6 are repeated until the required accuracy is satisfied. If the required accuracy is satisfied, the created AI model is sent from the AI model training unit 302 to the AI model training management unit 301 at time t9_8. At time t9_9 (corresponding to the time t13), the AI model training management unit 301 sends information about the created AI model to the network operation AI model manager 1.

According to the above embodiment, a platform that dynamically manages AI models related to network operation can be constructed, enabling autonomous network operation of commercial networks, and thus providing a low-cost and good network environment. As a result, it will be possible to contribute to Goal 9 "Develop resilient infrastructure and promote inclusive and sustainable industrialization" and Goal 11 "Make cities inclusive, safe, resilient and sustainable" of the Sustainable Development Goals (SDGs) led by the United Nations.

1...Network operation AI model manager, 2...Verification network construction node, 3...AI model training node, 101...Request receiving unit, 102...AI model management unit, 103...AI model database, 201...Verification network configuration management unit, 202...Verification network information acquisition unit, 203...Verification network construction unit, 301...AI model training management unit, 302...AI model training unit, 303...Measurement data database, NW1...Commercial network, NW2...Verification network

Claims (8)

In an AI model construction platform that builds AI models to be applied to commercial networks on a verification network,
A means for obtaining network information of a commercial network;
A means for acquiring information about available computing resources in a network verification environment;
A means for constructing a verification network in the verification environment based on the network information and the computational resource information;
A means for determining test scenario information that needs to be implemented to build an AI model based on the network information;
and a means for constructing an AI model based on a result of a test using the test scenario information in a verification network ;
The network information includes commercial network configuration information, accommodated user information, and network statistics information;
The means for constructing the verification network constructs a verification network smaller than a commercial network based on the configuration information,
The AI model building platform is characterized in that the means for determining the test scenario information determines the test scenario information using accommodated user information and network statistical information that have been shrunk in a ratio according to the size of the verification network relative to the commercial network . the commercial network is configured by arranging multiple uniform network instances in parallel;
The means for constructing the verification network constructs a verification network having a smaller number of network instances than the commercial network;
The AI model building platform according to claim 1, characterized in that the means for determining the test scenario information determines the test scenario information using accommodated user information and network statistical information deflated at a ratio corresponding to the number of network instances of the verification network relative to the number of network instances of the commercial network. 3. The AI model building platform according to claim 1, wherein the means for building the AI model determines whether the built AI model conforms to the type and accuracy required for the AI model, and if not, rebuilds the AI model. 4. The AI model construction platform according to any one of claims 1 to 3 , further comprising: a means for storing the constructed AI model; and applying the existing AI model in response to an AI model request from a commercial network to which an existing AI model can be applied. In an AI model construction method in which a computer constructs an AI model to be applied to a commercial network on a verification network,
Obtain network information of commercial networks,
Obtaining information about available computing resources in the network verification environment;
constructing a verification network in the verification environment based on the network information and the computational resource information;
Determine test scenario information that needs to be implemented to build an AI model based on the network information;
Constructing an AI model based on a result of a test using the test scenario information in a verification network;
The network information includes commercial network configuration information, accommodated user information, and network statistics information;
constructing a verification network that is smaller than the commercial network based on the configuration information;
An AI model construction method characterized by determining test scenario information using accommodated user information and network statistical information that have been shrunk in proportion to the scale of the verification network relative to the commercial network . the commercial network is configured by arranging multiple uniform network instances in parallel;
Construct a verification network with a smaller number of network instances than the commercial network;
The AI model construction method according to claim 5 , characterized in that test scenario information is determined using accommodated user information and network statistical information that have been deflated in a ratio corresponding to the number of network instances in the verification network relative to the number of network instances in the commercial network. In an AI model construction program that builds AI models to be applied to commercial networks on a verification network,
A procedure for obtaining network information of a commercial network;
obtaining information about available computing resources in a network verification environment;
constructing a verification network in the verification environment based on the network information and the computational resource information;
A step of determining test scenario information that needs to be implemented to build an AI model based on the network information;
A procedure for constructing an AI model based on a result of a test using the test scenario information in a verification network ;
The network information includes commercial network configuration information, accommodated user information, and network statistics information;
The step of constructing the verification network includes constructing a verification network smaller than a commercial network based on the configuration information;
The AI model construction program is characterized in that, in the step of determining the test scenario information, the test scenario information is determined using accommodated user information and network statistical information that have been shrunk in proportion to the scale of the verification network relative to the commercial network. the commercial network is configured by arranging multiple uniform network instances in parallel;
The step of constructing the verification network includes constructing a verification network having a smaller number of network instances than the commercial network;
The AI model construction program according to claim 7, characterized in that in the step of determining the test scenario information, the test scenario information is determined using accommodated user information and network statistical information that have been deflated at a ratio corresponding to the number of network instances in the verification network relative to the number of network instances in the commercial network.

Images