JP7803896B2 - Method for training learning models of analysis function servers located at each of multiple sites, analysis function server, and program - Google Patents

Description

The present invention relates to technology for analysis control servers located in core systems of mobile communication networks. In particular, it relates to technology for NWDAF (Network Data Analytics Function).

Common to both 5G (5th Generation) and 4G (4th Generation) standards, mobile communication networks consist of user equipment (UE), radio access networks (RAN), and core network systems.

Figure 1 is a functional configuration diagram of a core system in conventional technology.

The core system has a group of physical user plane network devices arranged on a virtualization platform of a group of logical control plane network devices. The control plane network devices accommodate a large number of user terminals via the RAN and transmit and receive control signals such as for establishing communications. The user plane network devices transmit and receive user data from user terminals via the RAN.
Furthermore, various NF (Network Function) servers are arranged in the core system to control the control plane network devices.

The core system includes an NWDAF (analysis control server) that acquires and analyzes a wide variety of log data from various data sources, which are communication equipment. The NWDAF receives analysis requests from NF consumers and returns the analysis results. This is standardized by the 3GPP (3rd Generation Partnership Project) (registered trademark) (see, for example, Non-Patent Document 1).
NWDAF introduces federated learning, which allows NWDAF to use learning (AI) models for inference while training them with network data (see, for example, Non-Patent Document 2).

According to FIG. 1, the system is configured with the following NF servers.
Application Function Server (NF Consumer) 10
Repository Function Server (NRF) 11
Analysis function server (NWDAF) 12
The NWDAF 12 holds a learning model corresponding to the analysis item, executes inference processing in response to an analysis request, and returns the analysis results. For this purpose, multiple NWDAFs 12 are provided for each type of analysis item.
The NF consumer 10 is an NF server that is the source of an analysis request.
In response to a discovery request from an NF consumer, the NRF 11 selects one NWDAF from among a plurality of NWDAFs 12 and responds.
The control plane network devices and the user plane network devices serve as data sources, and serve as input data for the NWDAF to execute inference processing using a learning model.

Figure 2 is a sequence diagram for conventional technology.

(s01) Assume that the NF consumer 10 wishes to obtain an analysis result for analysis item α. At this time, the NF consumer 10 transmits a discover request to the NRF 11 to inquire about an NWDAF that can analyze analysis item α.
It is also preferable to set a Preferred Supported Analytics Delay, which is a parameter for the time required to obtain the analysis result for analysis item α, in the discovery request (see, for example, Non-Patent Document 2).

(s02) When the NRF 11 receives a discovery request from the NF consumer 10, it selects an NWDAF 12 that can perform the analysis. It returns the address of the selected NWDAF 12 to the NF consumer 10. The NWDAF 12 selected here possesses a "learning model" for analyzing the analysis item α requested by the NF consumer 10.

(s03) The NF consumer 10 sends an "analysis request" to the address of the NWDAF returned from the NRF 11. The analysis request includes the following information:
Analysis item (Analytics ID): Specification of analysis processing Analysis target period: Period requested as analysis result Desired waiting time: Waiting time for response of analysis processing

If the NWDAF 12 receives an analysis request and is unable to return the analysis results within the desired waiting time, it returns an error response to the NF consumer 10. At this time, the NWDAF 12 may set the analysis processing waiting time (revised waiting time) in the error response.

(s04) The NWDAF 12 transmits a network data request to the data source to acquire input data for analysis on the analysis item α.
In response to this, the data source transmits the network data to the NWDAF 12. This "data collection time" is delayed as the processing time of the NWDAF 12.

(s05) NWDAF121 inputs network data into the learning model, performs inference processing, and obtains the analysis results. This "inference time" is delayed as processing time for NWDAF12.

(s06) Then, the NWDAF 121 returns the analysis results to the NF consumer 10.

3GPP TS 23.288 “Architecture enhancements for 5G System (5GS) to support network data analytics services,” [online], [Retrieved March 7, 2023], Internet <URL: https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3579> Hexa-X deliverables D4.2 “AI-driven communication & computation co-design: initial solutions,” [online], [Retrieved March 7, 2023], Internet <URL: https://hexa-x.eu/wp-content/uploads/2022/07/Hexa-X_D4.2_v1.0.pdf>

As mentioned above, conventionally, various NF servers in a core system are centrally managed at one site.
In contrast, in recent years, 5G and 6G have adopted architectures based on distributed computing consisting of a central site and an edge site. For this reason, multiple NWDAFs have been deployed not only at the central site but also at edge sites. Deploying an NWDAF at an edge site where an NF consumer is located has the advantage of reducing the transmission delay between sending an analysis request and receiving the analysis result.

However, an NF consumer at an edge site transmits an "analysis request" to an NWDAF located within the edge site. In this case, there is a problem that analysis requests generated within the edge site are concentrated in the NWDAF at the edge site.
Many NF servers at edge sites generally have limited computing resources compared to those at central sites. As a result, NWDAFs at edge sites are also prone to running out of computing resources. When an NF consumer sends an analysis request to an NWDAF at its own edge site, the expected waiting time cannot be met.

In addition, various NF servers in the core system will be distributed across central and edge sites, which will increase the amount of network resources used between sites. In particular, the more geographically dispersed the central site and multiple edge sites are, the greater the problem of network resource capacity.

The present invention therefore aims to provide a learning model training method, analysis function server, and program for an analysis function server that can reduce the processing load on the analysis control server at the edge site and reduce the network resource usage capacity between the edge site and the central site.

According to the present invention, there is provided a method for training a learning model of a plurality of analysis function servers that analyzes network data using a learning model for a predetermined analysis item, the method comprising:
A first analytics function server is located at the edge site and a second analytics function server is located at the central site;
The first analysis function server at the edge site
a first step of determining whether or not the capacity of the network data in a data source at the edge site is equal to or less than the capacity of the learning model held by the edge site when training the learning model from the network data;
a 21st step of transmitting the network data to a second analysis function server at a central site when the first step determines that the answer is true, and causing the second analysis function server to train a learning model using the network data;
If the first step determines that the result is false, a 22nd step is included in which a learning model is trained using the network data and the trained learning model is sent to a second analysis function server at the central site.

According to another embodiment of the learning model training method of the analysis function server of the present invention,
It is also preferable that, as a preliminary step to the first step, the first analysis function server at the edge site proceeds to step 21 if its own computing resource utilization rate is higher than a predetermined threshold.

According to another embodiment of the learning model training method of the analysis function server of the present invention,
The first analysis function server at the edge site performs the following as a preliminary step of the first step:
receive an analysis request including an analysis item and an analysis request period from the consumer server, acquire network data corresponding to the analysis request period for the analysis item from a data source at the edge site, execute inference processing using a learning model, and return the analysis results to the consumer server ;
Or,
An analysis request including an analysis item and an analysis request period is received from the consumer server, network data corresponding to the analysis request period for the analysis item is obtained from the data source at the edge site, inference processing is performed using the learning model, and the analysis results are returned to the consumer server. If the utilization rate of its own computing resources is higher than a predetermined threshold, the process proceeds to step 21.
It is also preferable.

According to another embodiment of the learning model training method of the analysis function server of the present invention,
It is also preferable that the second analysis function server at the central site accumulates and shares learning models of the multiple first analysis function servers at the multiple edge sites.

According to another embodiment of the learning model training method of the analysis function server of the present invention,
The analysis function server is NWDAF (Network Data Analytics Function),
The learning model of the analysis function server is trained by MTLF (Model Training logical function),
It is also preferable that the inference process of the analysis function server is executed by AnLF (Analytics logical function).

According to another embodiment of the learning model training method of the analysis function server of the present invention,
Further, an analysis control function server is provided at the edge site,
The consumer server sends an analysis request to the analysis control function server at the edge site where it is located,
It is also preferable that the analysis control function server selects an analysis control server at the edge site and transfers the analysis request to that analysis control server.

According to the present invention, an analysis function server at an edge site is capable of analyzing network data using a learning model for a predetermined analysis item and communicating with other analysis function servers at a central site,
a determination means for determining whether or not the capacity of the network data in the data source of the edge site is equal to or less than the capacity of the learning model held by the edge site when training the learning model from the network data;
a network data transmission means for transmitting the network data to another analysis function server at the central site when the first step determines that the answer is true, and causing the second analysis function server to train a learning model using the network data;
If the first step determines that the result is false, the method is characterized by having a learning model transmission means that trains a learning model using the network data and transmits the trained learning model to other analysis function servers at the central site.

According to the present invention, a program for causing a computer installed in an analysis function server at an edge site to function is provided, which analyzes network data using a learning model for a predetermined analysis item and is capable of communicating with other analysis function servers at a central site, and
a determination means for determining whether or not the capacity of the network data in the data source of the edge site is equal to or less than the capacity of the learning model held by the edge site when training the learning model from the network data;
a network data transmission means for transmitting the network data to another analysis function server at the central site when the first step determines that the answer is true, and causing the second analysis function server to train a learning model using the network data;
If the first step determines that the result is false, the method is characterized in that the computer functions as a learning model transmission means that trains a learning model using the network data and transmits the trained learning model to other analysis function servers at the central site.

The present invention's analysis function server learning model training method, analysis function server, and program can reduce the processing load on the edge site analysis control server and reduce the network resource usage capacity between the edge site and the central site.

FIG. 1 is a diagram illustrating a system configuration according to a conventional technique. FIG. 1 is a sequence diagram according to the prior art. 1 is a system configuration diagram according to the present invention. 10 is a flowchart of an analysis function server at an edge site in the present invention. FIG. 10 is a system configuration diagram further including an analysis control function server.

The following describes in detail an embodiment of the present invention using the drawings.

Figure 3 is a system configuration diagram for the present invention.

The core system in Figure 3 is composed of multiple edge sites and one central site. The multiple edge sites and the central site may each be geographically distributed, or may be logically distributed in consideration of computing resources.

The NRF 11 can discover various NF servers required by the NF consumer 10, as in the prior art.
The NF consumer 10 can inquire of the NRF 11 arranged in the edge site to which the NF consumer 10 belongs about the address of the NWDAF 12 that can perform analysis of the analysis item in question.

At each of the edge sites and the central site, a plurality of NWDAFs (analysis function servers) 12 are arranged, which analyze predetermined analysis items using a learning model from network data.
The NWDAF 12 is an existing device and has the following two functions.
AnLF(Analytics Logical Function)
MTLF (Model Training Logical Function)
The learning model of the NWDAF 12 is trained by the MTLF, which retrains the learning model using log data acquired from a data source and allows the AnLF to use the retrained learning model.
The inference process of the NWDAF 12 is executed by the AnLF, which receives an analysis request and returns the analysis result.

3, the NWDAF 12 at the edge site X can communicate with the NWDAF 22 at the central site. The NWDAF 12 at the edge site X may query the NRF 21 at the central site to discover the NWDAF 22 at the central site.
The NWDAF 22 at the central site may also store the learning models of multiple NWDAFs 12 at multiple edge sites, allowing the NWDAF 22 at the central site to build a shared learning model that can be used by multiple NWDAFs 12 at edge sites.

Figure 4 is a flowchart of the edge site analysis function server in the present invention.

(s03) This is the same as in Fig. 2. The NWDAF 12 receives an "analysis request" from the NF consumer 10. The analysis request includes at least an analysis item and an analysis request period.
It is assumed that the NWDAF 12 has a learning model A corresponding to the analysis item α included in the analysis request. Here, the learning model A has metadata such as the last update date and time and the period during which the learned network data can be analyzed.

(s04) This is the same as in Figure 2 described above. In order to obtain input data for analysis of analysis item α, NWDAF12 sends a "network data request" to the data source requesting network data corresponding to the analysis request period. In response, NWDAF12 receives the network data from the data source.

(s05) Same as in Figure 2 described above. The NWDAF 12 inputs network data into the learning model for analysis item α, performs inference processing, and returns the analysis results to the NF consumer 10.

<Retraining the learning model>
After returning the analysis results, the NWDAF 12 preferably retrains using the network data. This allows the learning model to be updated. The learning model is also shared between the NWDAF 12 at the edge site and the NWDAF 22 at the central site. The learning model held by the NWDAF 22 at the central site can also be used at other edge sites.

(S0) First, the NWDAF 12 determines whether its own computational resource utilization rate is equal to or lower than a predetermined threshold. If it determines that this is true (computational resource utilization rate≦predetermined threshold), it proceeds to S1. On the other hand, if it determines that this is false (computational resource utilization rate>predetermined threshold), it is determined that the NWDAF 12 is in an overload state, and it proceeds to S21. This allows the processing load of the NWDAF 12 at the edge site to be constantly reduced.
The NWDAF 12 may use the processor usage rate or the reception rate of analysis requests per unit time as the computational resource usage rate that it can measure itself.

(S1) NWDAF12 determines whether the capacity of the network data is less than or equal to the capacity of the learning model it holds. If it determines that this is true (the capacity of the network data is less than or equal to the capacity of the learning model), it proceeds to S21. On the other hand, if it determines that this is false (the capacity of the network data is greater than the capacity of the learning model), it proceeds to S22.

(S21) Here, the NWDAF 12 at the edge site determines that, from the viewpoint of computational resources and network resources, it would be better to have the NWDAF 22 at the central site retrain the learning model A rather than retraining it itself. To this end, the NWDAF 12 transmits network data to the NWDAF 22 at the central site. This causes the NWDAF 22 to train the learning model A using the network data.
However, if the network between the edge site and the central site is congested and the transmission delay exceeds a predetermined threshold, it is also preferable to proceed to S22.

In addition, the NWDAF12 at the edge site trains its learning model using only the network data acquired from the data source that exceeds the last update date and time and analysis period held by the learning model.

(S22) Here, the NWDAF12 at the edge site trains the learning model A using network data from the viewpoint of computational resources and network resources, and transmits the trained learning model A to the NWDAF22 at the central site.
By transmitting and receiving learning models between the edge site and the central site rather than transmitting and receiving network data, the amount of traffic is reduced and the network resource capacity is reduced.

Furthermore, the NWDAF12 at the central site trains its learning model using only the network data received from the NWDAF12 at the edge site that exceeds the last update date and time and analysis period held by the learning model.

As shown in Figure 4, it is possible to achieve both inference that takes into account the computational resources of the NWDAF12 at the edge site and retraining of the learning model that takes into account the network resources between the edge site and the central site.

Figure 5 is a system configuration diagram that further includes an analysis control function server.

5, in comparison with FIG. 3, analysis control function servers (NWDAF controllers) 13 and 23 are arranged at the edge site and the central site, respectively.
The NWDAF controller 13 at the edge site functions as a producer server for the NF consumer 10 .

The NF consumer 10 inquires of the NRF 11 at the edge site where the NF consumer 10 is located about the address of the NWDAF controller 13 .
Next, the NF consumer 10 sends an “analysis request” to the NWDAF controller 13 .
In this way, by disposing the NWDAF controller 13, the NRF 11 does not need to discover the NWDAF 12 according to the analysis items of the analysis request, and it also does not need to manage the Supported Analytics Delay for each NWDAF 12. The NRF 11 can reduce the processing load in managing the NWDAF 12, and can allocate that processing load to discovering various other NF servers.

<Management of current computing resource utilization rate for each NWDAF 12 in the NWDAF controller>
The NWDAF controller 13 may manage the current computational resource utilization rate for each NWDAF 12. The NWDAF controller 13 may periodically receive computational resource utilization rates from multiple NWDAFs 12 under its management.
The NWDAF controller 13 selects, as search candidates, multiple NWDAFs 12 that satisfy a predetermined condition, in descending order of current computational resource utilization rate (in descending order of computational resource efficiency). Here, the "predetermined condition" refers to a predetermined rate or less. In other words, if there are only NWDAFs 12 with computational resource utilization rates higher than the predetermined rate, no search candidates are selected. On the other hand, the NWDAF with the highest computational resource utilization rate and a rate less than the predetermined rate may be selected as a search candidate.

<Selection of NWDAF 12 in NWDAF Controller>
The NWDAF controller 13 selects an NWDAF 12 that satisfies the expected waiting time of the analysis request from among the NWDAFs 12 that are search candidates.
The NWDAF controller 13 calculates the desired waiting time as an inference time according to the time length of the analysis request period (Analytics target period), and calculates the network data acquisition time according to the time length of the analysis request period.
Then, the NWDAF 12 for which the sum of the inference time and the data collection time is equal to or less than the desired waiting time is selected as a search candidate.
Inference time + data collection time ≦ desired waiting time If the NWDAF 12 cannot be found, it returns an error response to the NF consumer 10. At this time, the NWDAF controller 13 sets the inference time in the error response as the revised waiting time. This allows the NF consumer 10 to send the analysis request again, taking the revised waiting time into consideration.
The NWDAF controller 13 may schedule and manage the time slots for inference of analysis items to be executed for each NWDAF 12 .

Then, the NWDAF controller 13 forwards the "analysis request" received from the NF consumer 10 to the selected NWDAF 12.

As explained in detail above, the learning model training method for an analysis function server, analysis function server, and program of the present invention can reduce the processing load on the analysis control server at the edge site and reduce the network resource usage capacity between the edge site and the central site.

Furthermore, this will enable, for example, "reducing the processing load on core systems of mobile communications networks," which will contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "build resilient infrastructure, promote sustainable industrialization and foster innovation."

With regard to the various embodiments of the present invention described above, those skilled in the art will be able to easily make various changes, modifications, and omissions that fall within the scope of the technical spirit and principles of the present invention. The foregoing description is merely illustrative and is not intended to be limiting in any way. The present invention is limited only by the claims and their equivalents.

10 NF consumer, application function server 11 NRF, repository function server 12 NWDAF, analysis function server 13 NWDAF controller, analysis control function server 21 NRF, repository function server 22 NWDAF, analysis function server 23 NWDAF controller, analysis control function server

Claims (8)

A method for training a learning model of a plurality of analysis function servers that analyzes network data using a learning model for a predetermined analysis item, comprising:
A first analytics function server is located at the edge site and a second analytics function server is located at the central site;
The first analysis function server at the edge site
a first step of determining whether or not the capacity of the network data in a data source at the edge site is equal to or less than the capacity of the learning model held by the edge site when training the learning model from the network data;
a 21st step of transmitting the network data to a second analysis function server at a central site when the first step determines that the answer is true, and causing the second analysis function server to train a learning model using the network data;
A method for training a learning model for an analysis function server, characterized in that if the first step determines that the result is false, a 22nd step trains a learning model using the network data and sends the trained learning model to a second analysis function server at a central site. A method for training a learning model for an analytical function server as described in claim 1, characterized in that, as a preliminary step to the first step, the first analytical function server at the edge site proceeds to step 21 if its own computing resource utilization rate is higher than a predetermined threshold. The first analysis function server at the edge site performs the following as a preliminary step of the first step:
receive an analysis request including an analysis item and an analysis request period from the consumer server, acquire network data corresponding to the analysis request period for the analysis item from a data source at the edge site, execute inference processing using a learning model, and return the analysis results to the consumer server ;
Or,
An analysis request including an analysis item and an analysis request period is received from the consumer server, network data corresponding to the analysis request period for the analysis item is obtained from the data source at the edge site, inference processing is performed using the learning model, and the analysis results are returned to the consumer server. If the utilization rate of its own computing resources is higher than a predetermined threshold, the process proceeds to step 21.
2. The method for training a learning model of an analysis function server according to claim 1 . A method for training a learning model of an analytical function server as described in claim 1 or 2, characterized in that a second analytical function server at a central site accumulates and shares learning models of multiple first analytical function servers at multiple edge sites. The analysis function server is NWDAF (Network Data Analytics Function),
The learning model of the analysis function server is trained by MTLF (Model Training logical function),
3. The method for training a learning model for an analytical function server according to claim 1, wherein the inference process of the analytical function server is executed by an AnLF (Analytics logical function). Further, an analysis control function server is provided at the edge site,
The consumer server sends an analysis request to the analysis control function server at the edge site where it is located,
3. The method for training a learning model of an analysis function server according to claim 1, wherein the analysis control function server selects an analysis control server at an edge site and transfers the analysis request to the analysis control server. An analysis function server at an edge site that analyzes network data using a learning model for a predetermined analysis item and can communicate with other analysis function servers at a central site,
a determination means for determining whether or not the capacity of the network data in the data source of the edge site is equal to or less than the capacity of the learning model held by the edge site when training the learning model from the network data;
a network data transmission means for transmitting the network data to another analysis function server at the central site when the first step determines that the answer is true, and causing the second analysis function server to train a learning model using the network data;
An analysis function server at an edge site, characterized in that if the first step determines that the result is false, it has a learning model transmission means for training a learning model using the network data and transmitting the trained learning model to other analysis function servers at the central site. A program that causes a computer installed in an analysis function server at an edge site to function, the computer analyzing network data using a learning model and capable of communicating with other analysis function servers at a central site,
a determination means for determining whether or not the capacity of the network data in the data source of the edge site is equal to or less than the capacity of the learning model held by the edge site when training the learning model from the network data;
a network data transmission means for transmitting the network data to another analysis function server at the central site when the first step determines that the answer is true, and causing the second analysis function server to train a learning model using the network data;
A program for an analysis function server, characterized in that if the first step determines that the result is false, the program causes the computer to function as a learning model transmission means that trains a learning model using the network data and transmits the trained learning model to other analysis function servers at the central site.