WO2024111996A1

WO2024111996A1 - Method by which service management and orchestration entity determines whether to update machine learning model, and service management and orchestration entity

Info

Publication number: WO2024111996A1
Application number: PCT/KR2023/018433
Authority: WO
Inventors: 이회주; 박희원; 설유환; 연훈제; 이민하; 조현성
Original assignee: 삼성전자 주식회사
Priority date: 2022-11-22
Filing date: 2023-11-16
Publication date: 2024-05-30
Also published as: KR20240076531A

Abstract

This method by which a service management and orchestration (SMO) entity determines whether to update a machine learning model may comprise the steps of: providing a first machine learning model to a target RIC that controls at least one cell corresponding to the first machine learning model; identifying first performance information that indicates the communication performance of the at least one cell of a first point in time; determining whether the identified first performance information satisfies a specific condition; and determining, on the basis of the determination that the identified first performance information satisfies the specific condition, whether to update the first machine learning model. The first point in time can be after a control operation performed, on the at least one cell, at a second point in time by means of the target RIC on the basis of the first machine learning model.

Description

How the service management and orchestration entity determines whether the machine learning model is updated, and the service management and orchestration entity

This disclosure relates to a service management and orchestration entity and a method for determining whether a machine learning model is updated.

5G mobile communication technology defines a wide frequency band to enable fast transmission speeds and new services, and includes sub-6 GHz ('Sub 6GHz') bands such as 3.5 gigahertz (3.5 GHz) as well as millimeter wave (mm) bands such as 28 GHz and 39 GHz. It is also possible to implement it in the ultra-high frequency band ('Above 6GHz') called Wave. In addition, in the case of 6G mobile communication technology, which is called the system of Beyond 5G, Terra is working to achieve a transmission speed that is 50 times faster than 5G mobile communication technology and an ultra-low delay time that is reduced to one-tenth. Implementation in Terahertz bands (e.g., 95 GHz to 3 THz) is being considered.

Currently, discussions are underway to improve and enhance the initial 5G mobile communication technology, considering the services that 5G mobile communication technology was intended to support, based on the vehicle's own location and status information. V2X (Vehicle-to-Everything) to help autonomous vehicles make driving decisions and increase user convenience, and NR-U (New Radio Unlicensed), which aims to operate a system that meets various regulatory requirements in unlicensed bands. ), NR terminal low power consumption technology (UE Power Saving), Non-Terrestrial Network (NTN), which is direct terminal-satellite communication to secure coverage in areas where communication with the terrestrial network is impossible, positioning, etc. Physical layer standardization for technology is in progress.

When this 5G mobile communication system is commercialized, an explosive increase in connected devices will be connected to the communication network. Accordingly, it is expected that strengthening the functions and performance of the 5G mobile communication system and integrated operation of connected devices will be necessary. To this end, eXtended Reality (XR) and Artificial Intelligence to efficiently support Augmented Reality (AR), Virtual Reality (VR), and Mixed Reality (MR). , AI) and machine learning (ML), new research will be conducted on 5G performance improvement and complexity reduction, AI service support, metaverse service support, and drone communication.

The present disclosure may be implemented in a variety of ways, including as a method, system, device, or computer program stored on a computer-readable storage medium.

A method for a service management and orchestration (SMO) entity to determine whether to update a machine learning model according to an embodiment of the present disclosure includes controlling a first machine learning model at least one cell corresponding to the first machine learning model. It may include the step of providing to the target RIC. In one embodiment of the present disclosure, the method may include identifying first performance information indicating communication performance of the at least one cell at a first time. In one embodiment of the present disclosure, the method may include determining whether the identified first performance information satisfies a specific condition. In one embodiment of the present disclosure, the method may include determining whether to update the first machine learning model, based on determining that the identified first performance information satisfies the specific condition. . In one embodiment of the present disclosure, the first time point may be after the control operation performed by the target RIC at the second time point based on the first machine learning model for the at least one cell.

According to an embodiment of the present disclosure, the SMO entity may include a memory that stores one or more instructions and at least one processor that executes the one or more instructions stored in the memory. In one embodiment of the present disclosure, the at least one processor executes the one or more instructions to apply a first machine learning model to a target RIC that controls at least one cell corresponding to the first machine learning model. can be provided. In one embodiment of the present disclosure, the at least one processor may identify first performance information indicating communication performance of the at least one cell at a first point in time. In one embodiment of the present disclosure, the at least one processor may determine whether the identified first performance information satisfies a specific condition. In one embodiment of the present disclosure, the at least one processor may determine whether to update the first machine learning model based on determining that the identified first performance information satisfies the specific condition. In an embodiment of the present disclosure, the at least one processor is configured to set the first time point to the date after the control operation performed by the target RIC at the second time point based on the first machine learning model for the at least one cell. You can.

A program for performing on a computer a method of determining whether to update a machine learning model according to an embodiment of the present disclosure may be recorded on a computer-readable recording medium.

Figure 1 is a diagram showing an O-RAN structure according to an embodiment.

Figure 2 is a flowchart showing how the SMO entity determines whether to update the machine learning model, according to one embodiment.

Figure 3 is a block diagram showing an example of the structure of a SMO entity and RIC according to an embodiment.

Figure 4 is a diagram illustrating an example of an SMO entity classifying a plurality of cells, according to an embodiment.

Figure 5 is a diagram illustrating an example of an SMO entity distributing a service to a target RIC, according to one embodiment.

FIG. 6 is a diagram illustrating an example of RIC controlling a cell and collecting data, according to an embodiment.

Figure 7 is a diagram illustrating an example in which the SMO entity detects characteristic drift and model drift for each cell, according to an embodiment.

Figure 8 is a diagram illustrating an example of a SMO entity reclassifying a cell according to an embodiment.

Figure 9 is a diagram illustrating an example of updating an ML model according to an embodiment.

Figure 10 is an example of a block diagram of the SMO entity 100 according to an embodiment.

In this specification, singular or plural expressions are selected to suit the presented situation for convenience of explanation, and even components expressed in the plural may be composed of the singular, or components expressed in the singular may be composed of the plural. That is, singular expressions include plural expressions, unless the context clearly specifies that they are singular. Additionally, plural expressions include singular expressions, unless the context clearly specifies plural expressions. When it is said that a certain part includes a certain element throughout the specification, it does not mean that other elements are excluded, but that other elements may be further included, unless stated to the contrary.

Prior to a detailed description of the invention, the terms used hereinafter may be defined or understood as follows.

In the present disclosure, the expression “at least one of a, b, or c” refers to “a”, “b”, “c”, “a and b”, “a and c”, “b and c”, “a, b and c", or variations thereof.

In this specification, the description of 'first', 'second', or 'third' is only used to distinguish each element and may not indicate order. Additionally, the first configuration and the second configuration may not necessarily mean different configurations. For example, the first model and the second model may each refer to models corresponding to different cell groups, but the first model and the second model may be substantially the same model. For example, the first cell group and the second cell group may refer to different groups, but the cells in the first cell group and the cells in the second group may be different, partially identical, or identical.

In this specification, 'ML (machine learning) model' may mean any ML model that can be used to control a cell or base station in a RAN communication environment. An example of an ML model may be a model for resource allocation or scheduling of a cell or base station, but is not limited thereto. Examples of ML models may include models learned to calculate specific scheduling parameters based on specific data about cells, data usage prediction models, etc., but are not limited to these and may be various models that can be applied in a RAN environment. .

In this specification, 'RIC (radio access network intelligence controller)' may refer to an entity for controlling at least one base station in an O-RAN (open-radio access network) environment, but is not limited thereto. For example, in another communication system environment, it may be an entity for controlling at least one base station (or cell) using a machine learning model.

In this specification, 'SMO (service management and orchestration)' may refer to an entity that manages RIC or network function entities in an O-RAN environment, but is not limited thereto. For example, in other communication system environments, it may be an entity that manages network function entities, etc., and monitors the performance of machine learning models.

In this specification, 'service' may refer to an operation, method, function, business, technical support, product, etc. performed, used, or provided in a communication network environment. For example, 'service' may be for improving the performance of the communication network of a radio access network (RAN), a base station, or a cell. For example, 'service' may mean performing operations based on the inference results of a machine learning model.

In this specification, ‘distributing a service’ may mean distributing information to provide a service. For example, distributing a service by an SMO entity may mean installing an application to provide the service in the RIC. For example, when an SMO entity distributes a service that calculates a specific parameter, it may mean distributing information about an ML model for calculating a specific parameter, an ML model-based control operation method, and commands.

In this specification, 'point in time' may refer to a certain moment and/or a certain section in the flow of time.

Figure 1 is a diagram showing an O-RAN structure according to an embodiment.

Figure 1 is an example of an O-RAN structure and shows O-RAN components, but is not limited thereto. That is, in the O-RAN structure shown in FIG. 1, unnecessary components and connection relationships may be omitted when explaining embodiments of the present disclosure. Additionally, at least some of the components and connection relationships shown in FIG. 1 may not correspond to essential components of the O-RAN structure. Accordingly, some components may be modified, omitted, replaced, or added in the structure of the O-RAN shown in FIG. 1. For example, the O-RAN structure may be modified according to at least one standard specification. Additionally, operations in O-RAN may be performed according to at least one standard specification.

The O-RAN system can follow the artificial intelligence (AI)/machine learning (ML) workflow standard specifications of the O-RAN standardization organization (O-RAN alliance), and accordingly, AI/ML technology in the RAN environment. This can be applied. Through the O-RAN system, an environment in which equipment from different vendors can be integrated and operated can be provided by opening the existing vendor-specific RAN interface and defining an integrated interface.

Referring to FIG. 1, a service management and orchestration (SMO) entity 100 may be an entity that provides various management services and network management functions. The RAN Intelligent Controller (RIC) entity is a platform that can run third party applications that can acquire data from the RAN and deliver control messages using the open interface of O-RAN. It can be. Depending on the closed loop latency requirements, Non-Real Time (Non-RT) RIC (110) (e.g. latency greater than 1 second) and Near-Real Time (Non-RT) RIC (110) (e.g. latency greater than 1 second) Near-RT) RIC (120) (e.g., latency less than 1 second).

For example, the Non-RT RIC 110 may perform functions such as micro-service and policy management, wireless network analysis, and artificial intelligence model training. As shown, Non-RT RIC 110 may be embedded in the SMO entity 100. For example, the Near-RT RIC 120 may perform a radio resource management (RRM) function.

The O-RAN system includes O-RAN Central Unit (O-CU) (140, 150), O-RAN Distributed Unit (O-DU) (160), O- It includes entities such as a RAN radio unit (O-RAN Radio Unit, O-RU) 170, and may include an O-RAN network function introducing Near-RT RIC (120). The network function portion of the O-RAN may be a gNB supporting the 5th-Generation (5G) protocol or an eNB 130 supporting the 4th-Generation (4G) Long Term Evolution (LTE) protocol.

O-CU (140, 150) can be divided into O-CU-CP (140) of the control plane that transmits control information and O-CU-UP (150) of the user plane that transmits traffic, Near-RT Control commands received from the RIC 120 can be executed. O-CU-CP 140 and O-CU-UP 150 may additionally support E2 interfaces (e.g., E2-cp and E2-up) compared to CUs in non-O-RAN systems. The O-DU 160 may additionally support an E2 interface (e.g., E2-du) compared to the DU of a non-O-RAN system. In the O-RAN system, an open fronthaul interface may be used to connect the O-DU 160 and the O-RU 170.

The O1 interface may be used to connect the SMO entity 100 and O-RAN network functional entities. The A1 interface can be used to connect the non-RT RIC (110) and Near-RT RIC (120). The E1 interface may be an interface between O-CU-CP (140) and O-CU-UP (150). The F1-c interface may be an interface between the O-CU-CP (140) and the O-DU (160), and F1-u may be an interface between the O-CU-UP (150) and the O-DU (160).

The E2 interface may represent an interface between the Near-RT RIC 120 and RAN equipment. For example, the E2 interface can be an interface between Near-RT RIC (120) and O-eNB (130), O-CU-CP (140), O-CU-UP (150), and O-DU (160). there is. For example, the Near-RT RIC 120 collects measurement information for each functional entity of the wireless network through the E2 interface, transmits control commands to at least one base station, and controls the operation of at least one base station. You can. In O-RAN, other interfaces in addition to the interfaces shown may be used.

In a network environment, problems that are difficult to solve based on existing rules can be solved by introducing artificial intelligence technology. For example, in an unpredictable and diverse network environment, artificial intelligence technology can be applied for appropriate control according to the changing cell environment.

Logical modules necessary to apply AI/ML technology in the O-RAN system environment can be defined. For example, ML training module, ML inference module, and ML assisted solution module may be defined. The ML inference module can operate as an application (ML xApp) that performs ML model inference. The ML Assisted Solution module can operate as an application (Assist xApp) that uses inference results.

An ML model to be applied to a specific network environment can be created using data collected from that network environment. In other words, when learning an ML model, you can use data collected in the environment where you want to actually apply it. Accordingly, even if it is an ML model to solve the same problem, the ML model to be used in the first environment is created based on data collected in the first environment (or an environment similar to the first environment), and the ML model to be used in the second environment is It may be generated based on data collected in a second environment (or an environment similar to the second environment).

For example, a first environment with a relatively small number of terminals and a small amount of data communication and a second environment requiring large amounts of data communication from a large number of terminals are different environments, and the data collected in the first environment is different from the other. The ML model trained as a baseline may not be suitable for the second environment. Likewise, an ML model trained based on data collected in a second environment may not be suitable for the first environment. In other words, when each ML model is trained with data collected in different environments, the parameters of each ML model may not be learned identically.

Considering this, collecting data for each cell and training the ML model to create and use a customized ML model for each cell may be ideal for ML model performance. However, creating and using individual ML models for every cell may not be efficient due to computing resource issues and management complexity issues.

In order to solve the above-described problem, a method of grouping cells with similar characteristics into one cell group and generating and using an ML model suitable for each cell group may be considered. That is, an ML model for each cell group can be created by learning an ML model based on data collected from cells included in the cell group. By creating and using ML models for each cell group, some computing resource and management issues can be solved.

In grouping cells with similar characteristics into one cell group, that is, classifying or grouping multiple cells, the cell's communication environment (e.g. number of terminals, traffic volume, etc.), surrounding environment (e.g. : Various factors that can affect communication networks, such as weather characteristics, topographical characteristics, etc., can be considered. Additionally, multiple cells may be grouped based on different criteria for each service ML model. For example, for an ML model for a specific service, the first cell and the second cell may be grouped into the same cell group, while for an ML model for another service, they may be grouped into different cell groups.

Meanwhile, the network environment can change from time to time depending on various factors such as the number and type of terminals managed by each cell, and the types and characteristics of services used by the terminals. Accordingly, control by an ML model learned to suit the existing (i.e., previous) cell environment may no longer be optimal control for the cell. Therefore, when using a trained ML model, it is necessary to monitor whether the ML model is working well and whether updates are needed to maintain and improve the performance of the ML model.

For example, ML model updates may be periodically performed based on the latest data collected in an environment where the ML model is applied. For example, in a RAN environment, if one super ML model that applies to all cells is deployed to the RIC, the super ML model can be updated based on the latest data from all cells managed by the RIC. For example, when an ML model for each cell group is distributed, the ML model for each cell group may be updated based on the latest data of cells belonging to each cell group.

Methods of periodically updating ML models may have difficulty coping with rapidly changing environments. In addition, the method of periodically updating the ML model can be inefficient, such as wasting computer resources, because it also updates models that do not need to be updated. Therefore, it may be necessary to change the ML model for the cell or update the ML model according to changes in the cell environment or performance of the ML model.

The performance of an ML model can respond to the amount of change in KPI (key performance indicator) in the domain to which the ML model is applied. Applying ML models in communication systems aims to improve specific KPIs or overall performance. For example, in the RAN domain, KPIs for monitoring the performance of ML models may include average throughput and whether service level agreement (SLA) is guaranteed.

For example, when one super ML model is used for all cells, the SMO entity 100 monitors the performance of the super ML model by monitoring data (i.e. KPI information) of all cells managed by the RIC. can do. For example, when an ML model for each cell is used, the SMO entity 100 can monitor the performance of the ML model for each cell managed by the RIC based on the data of each cell. Alternatively, when an ML model for each cell group is used, the SMO entity 100 manages each cell group (or cells included in each cell group) managed by the RIC based on the data of the cells included in each cell group. You can monitor the performance of ML models.

A low KPI value of the target cell may indicate that the ML model already applied to the target cell no longer provides optimized inference results for the target cell. In this case, the ML model applied to the target cell needs to be updated even if the update cycle has not been reached. Alternatively, if there is another ML model more suitable for the changed characteristics of the target cell, performance for the target cell can be improved by allocating the other more suitable ML model to the target cell. In this case, the existing ML model for the target cell (i.e., the ML model before reallocation) may not be updated separately.

As described above, the SMO entity 100 monitors the performance of the ML model for the target cell (i.e., the performance of the target cell) and decides to update the ML model already applied to the target cell or selects another ML model that is more suitable. By allocating, the RIC can perform optimal control on the target cell. In this case, more reliable data can be used for periodic model updates. In other words, the performance and reliability of the ML model can be improved by updating the ML model based on data from cells that are still suitable for the ML model, rather than data from cells that are no longer suitable for the ML model due to environmental changes.

Embodiments according to the present disclosure may be implemented according to the standard specifications of a standardization organization, but are not limited thereto.

Referring to FIG. 2, a method 200 of determining whether an SMO entity (eg, 100 in FIG. 1) updates a machine learning model according to an embodiment may include steps 210 to 240. In one embodiment of the present disclosure, steps 210 to 240 may be executed by at least one processor included in the SMO entity. In one or more embodiments, the method 200 of determining whether to update a machine learning model is not limited to that shown in FIG. 2 and may further include steps not shown in FIG. 2 . For example, method 200 may further include performing an operation for updating the first machine learning model as the SMO entity determines to update the first machine learning model. Additionally, the method 200 may further include the SMO entity performing the operation of providing the updated model to the RIC.

In step 210, the SMO entity may provide the first machine learning model to the target RIC that controls at least one cell corresponding to the first machine learning model. At least one cell corresponding to the first machine learning model may refer to a cell controlled by the target RIC based on the first machine learning model. For example, at least one cell may include at least some cells among a plurality of cells included in the first cell group corresponding to the first machine learning model. The first machine learning model may be a model used to provide a specific service to at least one cell. The target RIC may perform a control operation for at least one cell based on the provided first machine learning model.

In step 220, the SMO entity may identify first performance information indicating communication performance of at least one cell at a first time. The first time point may represent a time point after the control operation performed by the target RIC at the second time point based on the first machine learning model for at least one cell. Additionally, the control operation performed by the target RIC at the second time point may be the latest control operation for a specific service (service associated with the first machine learning model) based on the first time point. Performance information may include at least one of a KPI value, a KPI change value, or a KPI statistical value.

In one embodiment of the present disclosure, the SMO entity may identify the amount of change in performance of at least one cell based on second performance information and first performance information indicating communication performance of at least one cell before the first time point. The second performance information may indicate the communication performance of at least one cell at any time interval or moment before the first time point. For example, the second performance information may indicate the performance of at least one cell before the second time point (i.e., before the control operation).

Control operations of the target RIC may be performed according to standard specifications, but are not limited thereto. For example, the control operation of the target RIC may include transmitting a control message to the base station or cell, receiving a response message from the base station or cell, etc. The time point after the control operation may indicate the time point after the control operation starts, after the control operation is completed, and/or after the cell or base station is set by the control operation. The first time may be a time according to the performance information monitoring cycle of the cell after the control operation.

At step 230, the SMO entity may determine whether the identified first performance information satisfies a specific condition. In one embodiment of the present disclosure, the specific condition may include a threshold condition. For example, the SMO entity may determine whether the first performance information is below (or above) a predetermined threshold. In this case, the SMO entity may determine that a specific condition is satisfied as the first performance information is determined to be below (or above) a predetermined threshold. For example, the SMO entity may determine whether the first performance information is within (or outside) a predetermined reference range. In this case, the SMO entity may determine that a specific condition is satisfied as the first performance information is determined to be included (or outside) a predetermined reference range.

In one embodiment of the present disclosure, the specific condition may include a change amount condition. That is, the SMO entity can determine whether the amount of change in performance of at least one cell satisfies a specific condition. For example, the SMO entity may determine whether the amount of performance change is below (or above) a predetermined threshold. In this case, the SMO entity may determine that the first performance information satisfies a specific condition as it determines that the performance change amount is below (or above) a predetermined threshold. For example, the SMO entity may determine whether the amount of performance variation is within (or outside of) a predetermined reference range. In this case, the SMO entity may determine that the first performance information satisfies a specific condition as it determines that the performance change amount is included (or deviates) from a predetermined reference range.

At step 240, based on determining that the first performance information satisfies a specific condition, the SMO entity may determine whether to update the first machine learning model. In one embodiment of the present disclosure, the SMO entity may determine to update the first machine learning model based on determining that the first performance information satisfies a specific condition. In one embodiment of the present disclosure, the SMO entity changes the machine learning model corresponding to at least one cell from the first machine learning model to the second machine learning model, based on determining that the first performance information satisfies a specific condition. Can be changed (or reassigned). In this case, the SMO entity may decide not to update the first machine learning model (i.e., freeze the first machine learning model).

Figure 2 shows only the step of the SMO entity identifying performance information of at least one cell at a first time and determining whether a specific condition is satisfied, but the present invention is not limited thereto. For example, the SMO entity may periodically or aperiodically monitor performance information of a plurality of cells associated with the target RIC (i.e., a plurality of cells managed by the target RIC).

The method 200 for the SMO entity to determine whether to update the machine learning model may further include performing at least some of the operations of the SMO entity, which will be described later.

In one embodiment of the present disclosure, the SMO entity may determine a group of at least one cell as the first group. The first machine learning model may be a machine learning model corresponding to the first group including at least one cell. For example, a first machine learning model and a second machine learning model for providing a specific service are trained with different learning data for the same service and may have different parameter values. In this case, in order to provide the specific service, a first machine learning model may be used for the first group, and a second machine learning model may be used for the second group.

In one embodiment of the present disclosure, the SMO entity may identify inference information including at least one of input data or output data of a first machine learning model for the control operation. In one embodiment of the present disclosure, the SMO entity may determine whether the environment of at least one cell has changed based on inferred information. In one embodiment of the present disclosure, based on determining that the inference information is outside the reference data range for the first group, the SMO entity may determine that the environment of at least one cell has changed.

In one embodiment of the present disclosure, based on determining that the environment of at least one cell has changed, the SMO entity may re-determine the group of at least one cell. In one embodiment of the present disclosure, based on determining that the environment of at least one cell has changed, the SMO entity may determine whether to update the first machine learning model. In one embodiment of the present disclosure, based on re-determining the group of at least one cell to the first group, the SMO entity may determine to update the first machine learning model. In one embodiment of the present disclosure, based on re-determining a group of at least one cell to a second group, the SMO entity may determine to freeze the first machine learning model. The machine learning model corresponding to the second group may be a second machine learning model. In one embodiment of the present disclosure, based on determining that the environment of at least one cell has not changed, the SMO entity may determine to update the first machine learning model.

A program for performing the method according to at least some of the above-described embodiments on a computer may be recorded on a computer-readable recording medium.

SMO entity 100 may refer to an entity for managing at least one RIC. That is, the SMO entity 100 can manage the life cycle of applications deployed to RIC (eg, Near-RT RIC) with an O-RAN defined service management module. The SMO entity 100 may include one or more hardware modules and/or software modules to perform operations such as service management. Referring to FIG. 3, the SMO entity 100 includes a service management module 310, an ML monitoring and analysis module 320, an App (application) database (DB) 330, and a model DB (340). It may include, but is not limited to this.

The service management module 310 can manage information for each service registered with the SMO entity 100. The service management module 310 may include a cell classifier 312 and a model updater 314. In one embodiment of the present disclosure, the service management module 310 may receive information about the service model and information for application distribution (App description) at the time of service registration from the service provider's device (e.g., server, entity). there is. Additionally or alternatively, the service management module 310 may receive information for grouping cells managed by the RIC, information for updating the model, etc. from the service provider's device.

The cell classification module 312 classifies a plurality of cells managed by at least one RIC associated with the SMO entity 100. Can be clustered or grouped. For example, the cell classification module 312 can classify each cell into a plurality of cell groups according to its characteristics. Additionally or alternatively, the cell classification module 312 may perform reclassification of cells included in the cell or cell group according to a re-classification request for the cell or cell group.

The cell classification module 312 may include cell grouping information for each service. That is, cell classification criteria, classification algorithms, classification models, etc. may differ depending on the service to be provided. For example, for the first service, cells may be classified based on the number of terminals, while for the second service, cells may be classified based on traffic. When the cell classification module 312 uses a machine learning model (e.g., a machine learning model learned to classify a plurality of cells) to classify cells, the cell classification module 312 provides information for using the machine learning model. (e.g. weights, parameters, etc.) may be included.

When the ML model needs to be updated (eg, when a retraining triggering request is received), the model learning module 314 may perform an operation to update the ML model. Additionally or alternatively, as the ML model is updated, the model learning module 314 may perform an update of the ML xApp 354 of the target RIC 120 using the updated ML model.

In one embodiment of the present disclosure, the model learning module 314 may include information about an update flow (eg, algorithm, etc.). In this case, the model learning module 314 can directly perform learning of the ML model. In one embodiment of the present disclosure, the model learning module 314 may request retraining from a device (eg, server) for retraining (ie, update). For example, the model learning module 314 may include training host information, model update triggering application programming interface (API), and uniform resource locator (URL) information for retraining the model. there is. In this case, as the model learning module 314 decides to retrain the ML model, it may request retraining of the ML model using training host information, API, URL information, etc.

The ML monitoring and analysis module 320 can receive data from the Near-RT RIC and perform analysis to analyze the performance of the machine learning model used for each service. As shown, the ML monitoring and analysis module 320 includes a KPI analysis module 322, a feature drift detector 324, and a model drift detector 326. may include. For example, the KPI analysis module 322 performs control operations (e.g., sending a control message) by an ML model for the service based on the KPI information and RIC control history information stored in the Near-RT RIC 120. KPI changes before and after can be monitored for each cell.

The feature drift detection module 324 may monitor and analyze information about the inference request data (i.e., input data) and inference result data (i.e., output data) of the ML model (ML xApp). For example, feature drift detection module 324 may monitor statistical information of inferred data. Drift in the characteristics of a cell may mean a change in the environmental characteristics of the cell. For example, the occurrence of a cell's characteristic drift may indicate that the environment and characteristics that may affect the cell's communication performance have changed. For example, the occurrence of characteristic drift of a cell may mean that the changed environment and characteristics of the cell no longer match the characteristics of the cell group to which the cell belongs. For example, a cell for which characteristic drift has been determined (i.e., a cell in which characteristic drift has occurred) may be unsuitable for an existing ML model.

In one embodiment of the present disclosure, when a normal statistic range of input data and output data exists in the ML model for a specific cell group, the characteristic drift detection module 324 detects the cell's normal statistic range based on information about the normal statistic range. It is possible to detect whether there is characteristic drift. In one embodiment of the present disclosure, when there is no normal statistical range between input data and output data, it is possible to detect whether cell characteristics drift by tracking the amount of change in statistical values. For example, a criterion for detecting feature drift may be determined according to the history of the inference data.

The characteristic drift detection module 324 may request reclassification of cells in which characteristic drift is determined to have occurred. For example, the characteristic drift detection module 324 may request the cell classification module 312 to reclassify the cell. That is, the characteristic drift detection module 324 may request to re-determine the group of cells whose characteristics have drifted.

The model drift detection module 326 may determine the impact of the RIC control operation on any service based on KPI monitoring information. The occurrence of model drift may indicate that the ML model is no longer achieving expected performance. In other words, the fact that the model has drifted may mean that the model no longer achieves expected performance (or standard performance). For example, a cell in which model drift is determined (i.e., a cell in which model drift occurs) may be a cell in which communication performance has deteriorated due to ML model-based control.

For example, the model drift detection module 326 may request reclassification of cells in which it is determined that model drift has occurred. For example, as the model drift detection module 326 determines that model drift has occurred, it may request reclassification of cells in a group corresponding to the model. For example, the model drift detection module 326 may request an update (retraining) of the ML model in which it is determined that model drift has occurred.

The SMO entity 100 can manage information stored in the App DB 330 and the Model DB 340. App DB 330 and Model DB 340 provide information (e.g. description) required to distribute applications registered in SMO entity 100 (e.g. ML xApp (354) and Assist xApp (356)) to RIC. You can save it. App DB 330 can store information about ML xApp 354 and Assist xApp 356 that constitute the service. The model DB 340 may store ML model information about the service. For example, the model DB 340 may include information about ML models for each cell group.

The target RIC 120 in FIG. 3 is a RIC managed by the SMO entity 100 in FIG. 3 and may be a Near-RT RIC. The target RIC 120 may include, but is not limited to, a monitoring agent module 352, ML xApp 354, Assist xApp 356, and KPI monitoring xApp 358. Additionally or alternatively, the target RIC 120 may include an inference data DB 360, a RIC control DB 362, and a performance information DB (KPIs per Cells/UEs) 364. You can.

The ML xApp 354 may be a (Near-RT) RIC application that provides a response to the request using an ML model according to the inference request. ML xApp 354 may be an application that performs inference operations of an ML model. In one embodiment of the present disclosure, the target RIC 120 may include a plurality of ML xApps, each corresponding to a plurality of groups for one service. For example, in the service for parameter calculation, the target RIC 120 includes a first ML xApp that calculates parameters of cells included in the first group and a second ML that calculates parameters of cells included in the second group. Can include xApp.

Assist xApp 356 may be an application for RIC that performs operations required for RAN control excluding the inference operation of the machine learning model (i.e., operation of ML xApp). Assist xApp, like other applications, supports interfaces such as E2, A1, and O1 and can enable RIC internal messaging. For example, Assist xApp (356) supports RIC internal communication and O-RAN interface and transmits data, signals, information, requests, messages, etc. to ML xApp (354), or transmits data, signals, etc. from ML xApp (354). , information, requests, messages, etc. can be received. For example, Assist xApp 356 can transmit data to or receive data from RAN when providing a service using an ML model.

The monitoring agent module 352 may provide RIC internal data according to requests received from entities, modules, devices, etc. external to the RIC. For example, monitoring agent module 352 may receive a monitoring data request from SMO entity 100 and transmit data according to the received request. The monitoring agent module 352 may use the O1 interface for data transmission and reception, but is not limited to this.

The KPI monitoring xApp 358 may collect data (e.g., KPI values) from the RAN and store it within the RIC. The performance information DB 364 (e.g., KPI storage) may be a database that stores KPI values collected by the KPI monitoring xApp 358. For example, according to the implementation of O-RAN SC, the performance information DB 364 can be accessed using a shared data layer (SDL) library that wraps Redis or InfluxDB.

The inference data DB 360 (e.g., inference data storage) may be a database that stores the inference request data received by the ML xApp 354 and the resulting data. That is, when the ML xApp 354 performs an inference operation of an ML model, data input to the ML model and data output from the ML model may be stored in the inference data DB 360. For example, the inference data DB 360 may store input data and/or output data input to the ML xApp 354 for each RAN, each cell group, or each cell.

The control information DB 362 (e.g., RIC Control Storage) may be a database that stores information about RIC control (e.g., control messages, etc.) transmitted from the target RIC 120. In one embodiment of the present disclosure, the control information DB 362 may store RIC control information transmitted from the target RIC 120 through E2. For example, the control information DB 362 may include the time when an RIC application (eg, Assist xApp) transmits a RIC control message to the RAN or cell and the contents of the control message.

For example, each RIC application (xApp) may store RIC control information in the control information DB 362. Additionally or alternatively, as the RIC application transmits a control message, the target RIC 120 (RIC platform) may hook the control message and store it in the control information DB 362. Additionally or alternatively, the target RIC 120 may identify a message type corresponding to the control message and store the control message in the control information DB 362. The method and process for storing RIC control information in the control information DB 362 are not limited to the above.

Each entity and module shown in FIG. 3 may perform operations by at least one processor. Each entity and module in FIG. 3 may be composed of at least one software component, at least one hardware component, and a combination thereof to perform each function.

To support an arbitrary service, there may be multiple ML models that consider the characteristics of the cell. For example, for an ML model that calculates the value of a specific parameter, there may be separate ML models corresponding to cell groups with large traffic and ML models corresponding to cell groups with low traffic. In this case, the ML model corresponding to a cell group with large traffic may be learned based on data from cells with large traffic, and the ML model corresponding to a cell group with low traffic may be learned based on data from cells with low traffic. there is. Accordingly, the SMO entity 100 can classify a plurality of cells according to the amount of traffic so that each cell can be controlled based on the inference results of the ML model suitable for traffic characteristics.

Before distributing a service (i.e., ML model for the service, information, etc.), the SMO entity 100 may perform grouping on a plurality of cells managed by the target RIC 120 to which the service is to be distributed. In one embodiment of the present disclosure, the cell classification module 312 of the SMO entity 100 uses information on cells managed by the target RIC 120 (i.e., cells controlled by the target RIC 120) to classify cells. Classification can be performed. For example, the cell classification module 312 may perform cell classification based on performance information (eg, KPI information) of a plurality of cells included in the performance information DB 364 of the target RIC 120. The performance information of the plurality of cells may include information about the KPI for each cell and/or the KPI for each terminal stored by the KPI monitoring xApp (356 in FIG. 3). Additionally or alternatively, the cell classification module 312 may perform cell classification based on basic information (eg, information registered during RAN setup) of a plurality of cells.

In one embodiment of the present disclosure, the cell classification module 312 may classify a plurality of cells based on feature clustering, simulation, etc. For example, based on the number of terminals in each cell, packet size of the terminal, request interval, mobility pattern of the terminal, data usage pattern of the terminal, cell average throughput trend, etc., among a plurality of cells with similar characteristics Cells can be grouped into the same group. In one embodiment of the present disclosure, the cell classification module 312 may classify a plurality of cells based on rules and/or ML models (e.g., models learned to classify cells), but is not limited thereto.

In one embodiment of the present disclosure, the cell classification module 312 may classify a plurality of cells in different ways (i.e., different criteria, different ML models, different algorithms, etc.) for each service. For example, for the first service, the first cell and the second cell may be classified as the same cell, but for the second service, the first cell and the second cell may be classified as different cells.

In one embodiment of the present disclosure, when each service is registered with the SMO entity 100, the SMO entity 100 may also obtain a cell classification method (i.e., classification criteria, etc.). The SMO entity 100 may classify cells using the obtained method, but is not limited to this.

In one embodiment of the present disclosure, even if the SMO entity 100 acquires one ML model for any service, it may group a plurality of cells and manage (e.g., monitor and update) the ML model for each group. there is. That is, the SMO entity 100 may develop one ML model for an arbitrary service by dividing it into ML models suitable for each group.

Referring to FIG. 4, the cell classification module 312 of the service management module 310 provides performance information of the target RIC 120 according to service information included in at least one of the App DB 330 or the model DB 340. A plurality of cells can be classified based on performance information of the plurality of cells included in the DB 364. For example, the service information included in the App DB 330 and the Model DB 340 may include, but is not limited to, classification criteria, methods, or the number of ML models (i.e., maximum number of groups). It may contain service information that can be considered when classifying multiple cells.

The SMO entity 100 may distribute the service application to the target RIC 120. In one embodiment of the present disclosure, the SMO entity 100 may package and distribute an inference application for an arbitrary service as many as the number of cell groups of cells for the target RIC 120. For example, when the SMO entity 100 classifies cells managed by the target RIC 120 into three cell groups, the SMO entity 100 creates an ML corresponding to each cell group, as shown in FIG. 5. xApp (354_1, 354_2, 354_3) may be provided as the target RIC (120). Each ML xApp may include an ML model, and the ML models included in each ML xApp (354_1, 354_2, 354_3) may be different from each other, but are not limited to this.

In one embodiment of the present disclosure, the SMO entity 100 uses text-based App information (e.g., App description) included in the App DB (330 in FIG. 3) to the target RIC 120 through the O2 interface. You can perform the operation of installing a service application. The SMO entity 100 may follow standard specifications in packaging and distributing service applications, but is not limited thereto.

Referring to FIG. 5, the SMO entity 100 includes a first ML xApp 354_1 including a first ML model corresponding to the first group, and a second ML xApp including a second ML model corresponding to the second group. (354_2) and a third ML xApp (354_3) including a third ML model corresponding to the third group may be provided to the target RIC (120). In one embodiment of the present disclosure, the SMO entity 100 may include and distribute information about the ML xApp corresponding to each group and group information of each cell in the Assist xApp 356. For example, the SMO entity 100 may include information indicating the correspondence between each group and each ML xApp (354_1, 354_2, 354_3) in the Assist xApp (356) and provide it to the target RIC (120). For example, as shown in FIG. 5, the SMO entity 100 has a first cell (C1) and a second cell (C2) belonging to the first group (group 1), and a third cell (C3). Information (i.e., group information) indicating that the cell belongs to the second group (group 2) and the fourth cell (C4) belongs to the third group (group 3) can be provided by including it in the Assist xApp 356.

Figure 5 shows an example in which the SMO entity 100 provides ML xApp for each group, but the present invention is not limited to this. For example, the SMO entity 100 may provide one ML xApp including all ML models for each group.

Assist xApp 356 can set up a subscription with the RAN and request a message (e.g., a message containing data and information for each cell). The Assist xApp 356 requests data (e.g. KPI data) for each cell from the RAN, and the operation of receiving data for each cell from the RAN may be performed according to the standard specifications of O-RAN E2, but It is not limited. Assist xApp (356) can transfer the data of each cell transmitted from the RAN to the ML xApp (354_1, 354_2, 354_2) corresponding to the group containing each cell based on the cell group information and request inference. . Assist xApp 356 can obtain inference result data for each cell in response to the request.

Referring to FIG. 6, Assist xApp (356) transmits data of the first cell (C1) and second cell (C2) included in the first group (group 1) to the ML xApp (354_1) for the first group. And, the inference result for the first cell and the inference result for the second cell of the first ML model may be received from the ML xApp (354_1). In addition, Assist xApp (356) transfers the data of the third cell (C3) included in the second group (group 2) to the ML xApp (354_2) for the second group, and transfers the data from the ML xApp (354_2) to the second ML Inference results for the third cell of the model can be received. In addition, Assist xApp (356) transfers the data of the fourth cell (C4) included in the third group (group 3) to the ML xApp (354_3) for the third group, and transfers the data from the ML xApp (354_3) to the third ML The inference result for the fourth cell of the model can be received.

Assist xApp 356 can transmit a control message for each cell based on the inference result for each cell. In one embodiment of the present disclosure, Assist xApp 356 may transmit a control message to the RAN. Each cell can be controlled according to a control message. Assist xApp 356 can perform control on each cell periodically or aperiodically by repeatedly performing the operations of receiving data, sending inference requests, receiving inference results, and sending control messages for each cell as described above. .

The target RIC 120 may periodically or aperiodically collect various data about each cell by monitoring them while performing control on each cell through a service application. In one embodiment of the present disclosure, the target RIC 120 may collect at least some of performance information indicating communication performance of each cell, control information performed for each cell, and inference information of the ML model for each cell. . For example, the target RIC 120 may store descriptive statistics such as the average, minimum value, maximum value, variance, and standard deviation of a certain time interval for at least one of performance information, control information, or inference information, depending on the resource environment. there is.

In one embodiment of the present disclosure, the KPI monitoring xApp 358 of the target RIC 120 may collect KPI information of the RAN. RAN KPI information may include at least one of a RAN-level KPI value or a cell-level KPI value. For example, the KPI monitoring xApp 358 can collect RAN KPI information through the E2 or O1 interface. For example, KPI monitoring xApp 358 can collect and store KPIs over time. In collecting KPI information, the KPI monitoring xApp 358 may follow the O-RAN E2SM-KPM standard, but is not limited thereto. Data monitored and collected by the KPI monitoring xApp 358 may be stored in the performance information DB 364.

In one embodiment of the present disclosure, Assist xApp 356 may store a history of when and which RAN control message was transmitted for which cell in the control information DB 362. For example, the target RIC 120 (i.e., RIC platform) may store RIC control messages. For example, due to security or latency issues, Assist xApp 356 may store control information. For example, because RIC control information can greatly affect the RAN environment, the target RIC 120 decodes the payload when the type of RMR (RIC message routing) message transmitted by the application is RIC Control. You can save it. For example, the E2 terminator of the target RIC 120 may store information about the control message. The E2 terminator can represent a module that supports SCTP to transfer data to the E2 interface.

In one embodiment of the present disclosure, ML xApp (354_1, 354_2, 354_3) may store at least some of the input data and output data for the inference operation of the ML model in the inference data DB 360. For example, the inference data DB 360 may include raw data for all inference operations. For example, the inference data DB 360 may include some sampled or processed data. Inference data may correspond to classification criteria, but is not limited to this. For example, if the input data of the ML model corresponds to packet size and request interval, a plurality of cells may be classified based on packet size and request interval.

Although the plurality of DBs shown in FIG. 6 are shown as separate components, at least some DBs may be included in the same hardware storage device. Additionally or alternatively, each DB is shown as a single component, but may be included in multiple storage devices in hardware.

The SMO entity 100 can monitor the performance of the ML model using data stored in the target RIC 120. For example, the SMO entity 100 displays statistics such as the average value, variance, standard deviation, maximum/minimum, and range of data for each cell (e.g., KPI, ML model input data, output data, etc.). Changes can be monitored. The performance of the ML model can correspond to the communication performance of the cell. Excellent communication performance of a cell indicates excellent control performance for the cell, and may indicate excellent performance of the ML model used to control the cell.

Meanwhile, even if it is the same ML model, the performance of the ML model may be different for each cell. For example, a certain ML model may achieve baseline performance for a first cell, but may not achieve baseline performance for a second cell. In this case, the corresponding ML model may be an appropriate model for the first cell but an inappropriate model for the second cell.

In one embodiment of the present disclosure, the SMO entity 100 may include reference data information for each ML model (e.g., reference ranges of input data and output data, etc.). The reference data information of the ML model may be reference data information of a group corresponding to the ML model. In one embodiment of the present disclosure, the reference data information of each ML model may include a preset normal range. For example, the reference data information of the first ML model is reference data information of the first cell group corresponding to the first ML model, and includes the normal range of data according to the characteristics of cells included in the first cell group, etc. can do. In one embodiment of the present disclosure, reference data information for each ML model may be set based on data history (eg, data statistics, etc.). For example, the reference data information of the first ML model may be set according to the statistical value (eg, average, maximum/minimum value, etc.) of the accumulated input data of the first ML model.

The SMO entity 100 can determine whether feature drift occurs in each cell based on reference data information for each ML model. In one embodiment of the present disclosure, the characteristic drift detection module 324 identifies input data for a first cell included in a first group, and determines that the identified input data is within a reference range for the first group (e.g., 1 As it is determined that the input data is outside the reference range of the ML model for the group, it may be determined that the environment of the first cell has changed. That is, it may be determined that the first cell no longer substantially belongs to the first group, and that the ML model corresponding to the first group is no longer suitable for the first cell. For example, the characteristic drift detection module 324 may determine that the group of the first cell should be redetermined based on determining that the environment of the first cell has changed. Additionally, the SMO entity 100 may further perform operations for cells in which characteristic drift is detected to determine whether the drift is a temporary phenomenon or corresponds to an error in data collection.

For example, the characteristic drift detection module 324 may determine whether there is characteristic drift for each cell, and may generate an alert indicating characteristic drift for cells determined to have drifted characteristics. For example, the characteristic drift detection module 324 may transmit information about cells in which characteristic drift has occurred to the service management module 310. For example, the characteristic drift detection module 324 may request the cell classification module 312 to reclassify cells that are determined to have drifted characteristics.

For example, according to the statistical values of the packet size and request interval of the cells belonging to the cell group with high traffic, the standard data for the cell group with high traffic is set to a packet size of 80~100mb and a request interval of 100us or less. It can be. In this case, when the packet size of any cell included in the cell group is lowered to 20~30mb or the request interval is identified as 1ms, the SMO entity 100 determines that characteristic drift has occurred for the cell, and re-processes. Classification can be performed.

The KPI analysis module 322 and the model drift detection module 326 may determine the performance of the ML model for each cell based on data stored in the control information DB 362 and the performance information DB 364. In one embodiment of the present disclosure, the KPI analysis module 322 and the model drift detection module 326 determine the ML model for each cell based on the data stored in the control information DB 362 and the performance information DB 364. It is possible to detect whether drift has occurred. For example, the KPI analysis module 322 sends information about cells whose KPI value has fallen below the threshold (or target value) after the Assist xApp 356 delivers a control message to the RAN to the model drift detection module 326. It can be delivered.

In one embodiment of the present disclosure, the KPI analysis module 322 determines whether or not RAN control using the inference result of the ML model was successful based on the KPI change before and after the point when the Assist xApp 356 delivers the control message to the RAN. It can be determined. Accordingly, the model drift detection module 326 can determine whether drift of the ML model for the cell occurs. For example, the KPI analysis module 322 may include a KPI value after the control operation performed by the target RIC 120 at a second time point based on the first ML model for an arbitrary cell and a KPI value before the control operation (e.g. : previous value or average value, etc.) can be compared and analyzed, and the analysis results can be provided to the model drift detection module 326.

The model drift detection module 326 can determine whether the ML model drifts for each cell. For example, the model drift detection module 326 may determine that drift has occurred for the first cell and that no drift has occurred for the second cell in the same ML model. Specifically, based on the analysis result of the KPI analysis module 322, the model drift detection module 326 detects that the KPI value of the first cell falls below the threshold value, decreases sharply, or falls within the reference range after the control operation. As it deviates, it can be determined that drift of the ML model has occurred with respect to the first cell. In contrast, the model drift detection module 326 detects the drift of the ML model for the second cell as the KPI value of the second cell is within the reference range after the control operation, based on the analysis result of the KPI analysis module 322. It can be determined that has not occurred.

For example, the model drift detection module 326 may generate an alert indicating model drift for an ML model that is determined to have drifted. For example, the model drift detection module 326 may output a notification indicating that drift of the ML model has occurred for an arbitrary cell. For example, the model drift detection module 326 may transmit information about cells in which drift of the ML model has occurred to the service management module 310. For example, the model drift detection module 326 may request the model learning module 314 to update the ML model that is determined to have drifted.

For example, in a service that predicts the data usage of a cell in advance and performs load balancing in advance when the usage is predicted to exceed the limit of the cell to prevent the cell from being overloaded, the RIC 120 continuously sends RAN control messages. Despite being delivered, if the data usage of the cell exceeds a certain level or the data usage of the cell soars sharply, the inference performance of the prediction model may be judged to not meet the expected performance. That is, control based on the inference result of the corresponding prediction model may be judged as abnormal control, and accordingly, the model drift detection module 326 may generate a notification indicating model drift.

In one embodiment of the present disclosure, the ML monitoring and analysis module 320 updates the ML model in the service management module 310 when a drift in the characteristics of the cell is detected even if the ML model for any cell is detected. You may not request.

In one embodiment of the present disclosure, the SMO entity 100 detects only characteristic drift for a certain cell, and when model drift is not detected, the SMO entity 100 generates reference data information (e.g., reference data information) for the corresponding ML model. The standard data information of the group corresponding to the ML model) can be modified.

The above-described characteristic drift detection and model drift detection are performed on a per-cell basis, but are not limited to this. For example, feature drift detection and/or model drift detection may be performed on a cell group basis, RAN unit, RIC unit, ML model unit, and service unit. In one embodiment of the present disclosure, the model drift detection module 326 may determine whether there is drift for each ML model. That is, the model drift detection module 326 may determine whether there is drift on an ML model basis rather than on a cell basis. For example, the model drift detection module 326 may determine whether the ML model is drifting based on the KPI change before and after the control operation of a plurality of cells corresponding to the ML model. Additionally or alternatively, the model drift detection module 326 may determine whether the ML model is drifting based on the degree of change in the KPI of cells and the number of cells that satisfy specific conditions.

The service management module 310 of the SMO entity 100 may determine whether to reclassify cells and update the model based on at least one of the characteristic drift detection result or the model drift detection result of the ML monitoring and analysis module 320. . In one embodiment of the present disclosure, service management module 310 may identify or receive notifications (or requests) from characteristic drift detection module 324. In one embodiment of the present disclosure, service management module 310 may identify or receive notifications (or requests) from model drift detection module 326. The service management module 310 may determine whether to reclassify cells and update the model based on the notification (or request) from the ML monitoring and analysis module 320.

In one embodiment of the present disclosure, the cell classification module 312 may re-determine the group of cells in which characteristic drift has occurred. That is, the cell classification module 312 may perform a reclassification operation on cells in which characteristic drift has occurred. In one embodiment of the present disclosure, the cell classification module 312 may re-determine the group of cells in which both characteristic drift and model drift occur.

Referring to FIG. 8, the cell classification module 312 may perform a reclassification operation on the second cell in which characteristic drift has occurred. For example, the cell classification module 312 may re-determine the group of the second cell from the first group to the second group by performing regrouping based on the latest data of the second cell.

In this case, the cell classification module 312 may provide new group information of the second cell to Assist xApp 356 of the target RIC 120. That is, the cell classification module 312 may provide information indicating that the second cell belongs to the second group to the Assist xApp 356. Assist xApp (356), which has received new group information, then inputs the data of the second cell to the ML xApp (354_2) corresponding to the second group to control the second cell, and based on the inference result, the second cell A control message can be transmitted. That is, the target RIC 120 may use the second ML model corresponding to the second group according to the changed group information to control the second cell.

In one embodiment of the present disclosure, when it is determined that not only characteristic drift but also model drift has occurred with respect to the second cell, as the group of the second cell is redetermined from the first group to the second group, the SMO entity 100 It may be decided not to update (i.e., freeze) the first ML model corresponding to the first group. Additionally, if model drift occurs in other cells with respect to the first ML model, the SMO entity 100 may determine an update of the first ML model based on the latest data of cells included in the first group after the reclassification operation. .

In one embodiment of the present disclosure, when characteristic drift detection is performed on a cell group basis, the SMO entity 100 may perform regrouping on all cells belonging to the group for which characteristic drift has been detected. In one embodiment of the present disclosure, when a cell in which model drift occurs exists among cells in a group in which characteristic drift is detected, the SMO entity 100 may perform regrouping of the cell in which model drift occurred.

When the model learning module 314 determines whether to update an arbitrary ML model, the decision may be made based on at least some of the characteristic drift monitoring results and model drift monitoring results of cells included in the cell group corresponding to the ML model. . In one embodiment of the present disclosure, the model learning module 314 may determine an update of the ML model corresponding to the cell for which model drift is detected. In one embodiment of the present disclosure, model learning module 314 may determine to update the ML model for which model drift is detected.

In one embodiment of the present disclosure, the cell classification module 312 may perform a reclassification operation on cells in which characteristic drift and model drift are detected. If the corresponding cell is determined to still belong to the same group despite the reclassification operation, the model learning module 314 may determine to update the ML model corresponding to the group of the corresponding cell. For example, referring to FIG. 9, despite the reclassification operation of the cell classification module 312 for the second cell in which characteristic drift and model drift are detected, the group of the second cell may still be determined to be the first group. . In this case, the model learning module 314 may determine an update of the first ML model corresponding to the first group.

In one embodiment of the present disclosure, when model drift occurs in other cells with respect to the first ML model, even if the group of the second cell is redetermined to another group, the model learning module 314 determines the first group after the reclassification operation. An update of the first ML model may be determined based on the latest data of the included cell. In one embodiment of the present disclosure, the model learning module 314 may determine an update of the ML model based on the number of cells in which model drift is detected for the ML model, the degree of drift (e.g., degree of KPI change, etc.), etc. . For example, the model learning module 314 may determine the update of the ML model corresponding to the cell group by considering the number of model drift alarms of cells included in the cell group.

As the SMO entity 100 determines to update the ML model, it may perform operations for updating the ML model. For example, ML model updates can occur offline or along a learning pipeline. For example, the model learning module 314 may perform an operation to update the ML model based on model update-related information registered at the time of service registration. For example, updates to the ML model may be performed on a separate learning server or may be performed within the SMO entity 100.

The model learning module 314 of the SMO entity 100 may provide the updated ML model to the ML xApp of the corresponding cell group. Therefore, referring to FIG. 9, the target RIC 120 may use the updated first ML model to control cells included in the first group.

In FIGS. 4 to 9 described above, for convenience of explanation, four cells managed by the target RIC 120 are shown, but the number of cells is not limited thereto.

In addition, in FIGS. 4 to 9 described above, embodiments in which a plurality of cells are grouped are described, but the present invention is not limited thereto. For example, even when using a super ML model applied to all cells or an ML model for each cell without grouping a plurality of cells, at least some of the above-described embodiments can be applied. For example, even in an embodiment that does not provide a plurality of cell groups and a model corresponding to each group, operations such as the above-described KPI analysis module, model drift detection module, and model learning module can be performed.

For example, the SMO entity 100 may determine whether the ML model is updated by monitoring changes in performance information (e.g., KPI value) according to the control operation of the RIC for the cell. In this case, some of the components related to grouping, such as the cell classification module 312 and the characteristic drift detection module 324, may be omitted from the SMO entity 100.

SMO entity 100 of FIG. 10 may include at least one processor 1010 and memory 1020. The processor 1010 is electrically connected to components included in the SMO entity 100 and may perform operations or data processing related to control and/or communication of components included in the SMO entity 100. According to one embodiment, the processor 1010 may load a request, command, or data received from at least one of the other components into the memory 1020 to process it, and store the resulting data in the memory 1020. According to one or more embodiments, the processor 1010 includes a general-purpose processor such as a central processing unit (CPU), an application processor (AP), a digital signal processor (DSP), a graphic processing unit (GPU), and a vision processing unit (VPU). It may include at least one of a processor dedicated to graphics, or a processor dedicated to artificial intelligence, such as a neural processing unit (NPU).

The processor 1010 can control input data to be processed according to predefined operation rules or artificial intelligence models stored in memory. Alternatively, when the processor 1010 is an artificial intelligence-specific processor, the artificial intelligence-specific processor may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

Predefined operation rules or artificial intelligence models may be characterized as being created through learning. Here, being created through learning means that the basic artificial intelligence model is learned using a large number of learning data by a learning algorithm, thereby creating a predefined operation rule or artificial intelligence model set to perform the desired characteristics (or purpose). It can mean a burden. This learning may be performed on the device itself that performs the artificial intelligence according to the present disclosure, or may be performed through a separate server and/or system. Examples of learning algorithms include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but are not limited to the examples described above.

The memory 1020 is electrically connected to the processor 1010 and may store one or more modules, at least one learning model, program, command, or data related to the operation of components included in the SMO entity 100. For example, the memory 1020 may store one or more modules, learning models, programs, instructions, or data for processing and control of the processor 1010. The memory 1020 is a flash memory type, hard disk type, multimedia card micro type, card type memory (for example, SD or XD memory, etc.), RAM (Random Access Memory), SRAM (Static Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic It may include at least one type of storage medium among disks and optical disks.

The module or model included in the memory 1020 is executed according to the control or command of the processor 1010 and may include a program, model, or algorithm configured to perform operations that derive output data for input data. The memory 1020 may include a software module and a software configuration among the modules of the SMO entity 100 shown in FIG. 3 . In one embodiment of the present disclosure, each module included in the memory 1020 may perform an operation by being executed by the processor 1010. For example, at least some of the modules of the SMO entity 100 shown in FIG. 3 may be implemented as separate hardware modules. For example, at least some of the modules of the SMO entity 100 shown in FIG. 3 may be implemented as a combination of software configuration and hardware configuration. Memory 1020 may include DBs of the SMO entity 100 shown in FIG. 3.

Memory 1020 of SMO entity 100 may include an ML model for any service. For example, the memory 1020 may include a plurality of parameter values (weight values) constituting an ML model. An ML model may be a learning model trained to provide an arbitrary service. For example, the ML model may be, but is not limited to, a learning model trained using supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning. An ML model may be composed of multiple neural network layers. Each of the plurality of neural network layers has a plurality of weight values, and a neural network calculation can be performed through calculation between the calculation result of the previous layer and the plurality of weights. Multiple weights of multiple neural network layers can be optimized based on the learning results of the ML model. For example, a plurality of weights may be updated so that loss or cost values obtained from the ML model are reduced or minimized during the learning process. ML models may include deep neural networks (DNN), such as Convolutional Neural Network (CNN), Deep Neural Network (DNN), Recurrent Neural Network (RNN), Restricted Boltzmann Machine (RBM), These include, but are not limited to, Deep Belief Network (DBN), Bidirectional Recurrent Deep Neural Network (BRDNN), or Deep Q-Networks.

SMO entity 100 may include more components than those shown in FIG. 8 . For example, the SMO entity 100 may further include a communication interface (or communication module) for communication with other entities and external devices. For example, the SMO entity 100 may further include a transmitter and receiver for transmitting and receiving signals with other entities and external devices. Signals that the SMO entity 100 transmits and receives through the transmitter and receiver may include control information and data. The transceiver may include a wired or wireless transceiver, and may include various components for transmitting and receiving signals. For example, the transceiver may receive a signal, output it to the processor 1010, and transmit the signal output from the processor 1010.

At least one processor 1010 may control a series of processes so that the SMO entity 100 can operate according to at least one embodiment disclosed in FIGS. 1 to 10. In one embodiment of the present disclosure, at least one processor 1010 may execute one or more instructions stored in memory. In one embodiment of the present disclosure, at least one processor 1010 may perform an operation by executing one or more instructions.

In one embodiment of the present disclosure, at least one processor 1010 may provide a first machine learning model to a target RIC that controls at least one cell corresponding to the first machine learning model. In one embodiment of the present disclosure, at least one processor 1010 may identify first performance information indicating communication performance of the at least one cell at a first time. In one embodiment of the present disclosure, at least one processor 1010 may determine whether the identified first performance information satisfies a specific condition. In one embodiment of the present disclosure, at least one processor 1010 may determine whether to update the first machine learning model based on determining that the identified first performance information satisfies the specific condition. For example, the at least one processor 1010 may determine whether to update the first machine learning model in response to determining that the identified first performance information satisfies the specific condition. For example, when the identified first performance information satisfies the specific condition, at least one processor 1010 may determine whether to update the first machine learning model. In one embodiment of the present disclosure, the first time point may be after the control operation performed by the target RIC at the second time point based on the first machine learning model for the at least one cell.

In one embodiment of the present disclosure, at least one processor 1010 is configured to configure the at least one cell based on the first performance information and second performance information indicating communication performance of the at least one cell before the first time point. The amount of change in performance can be identified. For example, at least one processor 1010 may identify the amount of change in performance of the at least one cell using the second performance information and the first performance information. In one embodiment of the present disclosure, at least one processor 1010 may determine whether the performance change satisfies the specific condition.

In one embodiment of the present disclosure, at least one processor 1010 may determine the group of the at least one cell as the first group. In one embodiment of the present disclosure, the first machine learning model may be a machine learning model corresponding to the first group including the at least one cell.

In one embodiment of the present disclosure, at least one processor 1010 may identify inference information including at least one of input data or output data of the first machine learning model for the control operation. In one embodiment of the present disclosure, at least one processor 1010 may determine whether the environment of the at least one cell changes based on the inference information. For example, at least one processor 1010 may determine whether the environment of the at least one cell has changed using the inference information.

In one embodiment of the present disclosure, the at least one processor 1010 may determine that the environment of the at least one cell has changed, based on determining that the inference information is outside the reference data range for the first group. there is. For example, in response to determining that the inference information is outside the reference data range for the first group, the at least one processor 1010 may determine that the environment of the at least one cell has changed. For example, when the inference information is outside the reference data range for the first group, the at least one processor 1010 may determine that the environment of the at least one cell has changed.

In one embodiment of the present disclosure, at least one processor 1010 may re-determine the group of the at least one cell based on determining that the environment of the at least one cell has changed. For example, the at least one processor 1010 may re-determine the group of the at least one cell in response to determining that the environment of the at least one cell has changed. For example, when the environment of the at least one cell changes, the at least one processor 1010 may re-determine the group of the at least one cell.

In one embodiment of the present disclosure, at least one processor 1010 may determine whether to update the first machine learning model based on determining that the environment of the at least one cell has changed. For example, the at least one processor 1010 may determine whether to update the first machine learning model in response to determining that the environment of the at least one cell has changed. For example, when the environment of the at least one cell changes, at least one processor 1010 may determine whether to update the first machine learning model.

In one embodiment of the present disclosure, at least one processor 1010 may determine an update of the first machine learning model based on re-determining the group of the at least one cell as the first group. For example, the at least one processor 1010 may determine an update of the first machine learning model in response to re-determining the group of the at least one cell as the first group. For example, when re-determining the group of the at least one cell as the first group, the at least one processor 1010 may determine an update of the first machine learning model.

In one embodiment of the present disclosure, at least one processor 1010 may determine to freeze the first machine learning model based on re-determining the group of the at least one cell into the second group. For example, the at least one processor 1010 may determine to freeze the first machine learning model in response to re-determining the group of the at least one cell into the second group. For example, when re-determining the group of the at least one cell as the second group, the at least one processor 1010 may determine to freeze the first machine learning model.

In one embodiment of the present disclosure, the machine learning model corresponding to the second group may be a second machine learning model. For example, the second machine learning model and the first machine learning model may be models for the same service. For example, the second machine learning model and the first machine learning model may be models learned with different training data.

In one embodiment of the present disclosure, at least one processor 1010 may determine an update of the first machine learning model based on determining that the environment of the at least one cell has not changed. For example, the at least one processor 1010 may determine to update the first machine learning model in response to determining that the environment of the at least one cell has not changed. For example, when the environment of the at least one cell does not change, the at least one processor 1010 may determine to update the first machine learning model.

In one embodiment of the present disclosure, the method for the SMO entity to determine whether to update the machine learning model includes providing a first machine learning model to a target RIC that controls at least one cell corresponding to the first machine learning model. May include steps. In one embodiment of the present disclosure, a method for an SMO entity to determine whether to update a machine learning model may include identifying first performance information indicating communication performance of the at least one cell at a first time. . In one embodiment of the present disclosure, a method for an SMO entity to determine whether to update a machine learning model may include determining whether the identified first performance information satisfies a specific condition. In one embodiment of the present disclosure, the method for the SMO entity to determine whether to update the machine learning model is based on determining that the identified first performance information satisfies the specific condition. It may include a step of determining whether to update. In one embodiment of the present disclosure, the first time point may be after the control operation performed by the target RIC 120 at the second time point based on the first machine learning model for the at least one cell.

In one embodiment of the present disclosure, the step of identifying the first performance information includes the at least one cell based on the first performance information and second performance information indicating communication performance of the at least one cell before the first time point. It may include identifying the amount of change in performance of one cell. In one embodiment of the present disclosure, determining whether the first performance information satisfies a specific condition may include determining whether the performance change amount satisfies the specific condition.

In one embodiment of the present disclosure, the method may include determining the group of the at least one cell as the first group. In one embodiment of the present disclosure, the first machine learning model may be a machine learning model corresponding to the first group including the at least one cell.

In one embodiment of the present disclosure, the method may include identifying inference information including at least one of input data or output data of the first machine learning model for the control operation. In one embodiment of the present disclosure, the method may include determining whether the environment of the at least one cell has changed based on the inference information.

In one embodiment of the present disclosure, the step of determining whether the environment of the at least one cell changes includes determining that the inference information is outside the reference data range for the first group, It may include determining that the environment has changed.

In one embodiment of the present disclosure, the method may include re-determining the group of the at least one cell based on determining that the environment of the at least one cell has changed.

In one embodiment of the present disclosure, the step of determining whether to update the first machine learning model based on determining that the identified first performance information satisfies the specific condition includes the environment of the at least one cell. Based on determining that this has changed, it may include determining whether to update the first machine learning model.

In one embodiment of the present disclosure, the step of determining whether to update the first machine learning model based on determining that the environment of the at least one cell has changed includes dividing the group of the at least one cell into the first Based on the group re-determination, determining an update to the first machine learning model may be included.

In one embodiment of the present disclosure, the step of determining whether to update the first machine learning model based on determining that the environment of the at least one cell has changed includes dividing the group of the at least one cell into a second group. It may include determining to freeze the first machine learning model based on the re-determination. In one embodiment of the present disclosure, the machine learning model corresponding to the second group may be a second machine learning model.

In one embodiment of the present disclosure, the step of determining whether to update the first machine learning model based on determining that the identified first performance information satisfies the specific condition includes the environment of the at least one cell. Based on determining that this has not changed, determining an update to the first machine learning model may be included.

Methods according to embodiments described in the claims or specification of the present disclosure may be implemented in the form of hardware, software, or a combination of hardware and software. When implemented as software, a computer-readable storage medium that stores one or more programs (software modules) may be provided. One or more programs stored in a computer-readable storage medium are configured to be executable by one or more processors in an electronic device (configured for execution). One or more programs include instructions that cause the electronic device to execute methods according to embodiments described in the claims or specification of the present disclosure.

These programs (software modules, software) may include random access memory, non-volatile memory, including flash memory, read only memory (ROM), and electrically erasable programmable ROM. (electrically erasable programmable read only memory, EEPROM), magnetic disc storage device, compact disc-ROM (CD-ROM), digital versatile discs (DVDs), or other types of disk storage. It can be stored in an optical storage device or magnetic cassette. Alternatively, it may be stored in a memory consisting of a combination of some or all of these. Additionally, multiple configuration memories may be included.

In addition, the program may be distributed through a communication network such as the Internet, an intranet, a local area network (LAN), a wide area network (WAN), or a storage area network (SAN), or a combination thereof. It may be stored on an attachable storage device that is accessible. This storage device can be connected to a device performing an embodiment of the present disclosure through an external port. Additionally, a separate storage device on a communication network may be connected to the device performing an embodiment of the present disclosure.

Securing and maintaining a super model that can cover the environment of all cells can increase the model development period and model size. Additionally, optimal control of all cells cannot be guaranteed. Learning and managing models for each cell separately can increase resource use and management/operation costs of RIC and SMO. According to one embodiment, cells can be grouped by service-specific features and multiple ML xApps can be used to achieve more optimized control and save resources. In other words, resources and management/operation costs can be reduced by managing multiple cells as a cell group.

In one embodiment of the present disclosure, group information can be updated through ML model performance monitoring. In other words, by monitoring model performance and updating the ML model when an update is needed, the time of abnormal operation (i.e., operation with poor performance) can be minimized. Additionally, resource use can be reduced by not performing updates when updates are not needed.

In one embodiment of the present disclosure, before retraining and updating the ML model, the group information is updated using the environment information of each cell, thereby reassigning the model to suit changes in the cell, thereby reducing management and operation costs. You can.

A storage medium that can be read by a device may be provided in the form of a non-transitory storage medium. Here, 'non-transitory storage medium' only means that it is a tangible device and does not contain signals (e.g. electromagnetic waves). This term refers to cases where data is semi-permanently stored in a storage medium and temporary storage media. It does not distinguish between cases where it is stored as . For example, a 'non-transitory storage medium' may include a buffer where data is temporarily stored.

According to one embodiment, methods according to various embodiments disclosed in this document may be provided and included in a computer program product. Computer program products are commodities and can be traded between sellers and buyers. A computer program product may be distributed in the form of a machine-readable storage medium (e.g. compact disc read only memory (CD-ROM)) or through an application store or between two user devices (e.g. smartphones). It may be distributed in person or online (e.g., downloaded or uploaded). In the case of online distribution, at least a portion of the computer program product (e.g., a downloadable app) is stored on a machine-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server. It can be temporarily stored or created temporarily.

Claims

In the method of the SMO (service management and orchestration) entity 100 determining whether to update the machine learning model,

Providing a first machine learning model to a target radio access network intelligence controller (RIC) 120 that controls at least one cell corresponding to the first machine learning model;

Identifying first performance information indicating communication performance of the at least one cell at a first time;

determining whether the identified first performance information satisfies a specific condition; and

Based on determining that the identified first performance information satisfies the specific condition, determining whether to update the first machine learning model;

The first time point is after the control operation performed by the target RIC 120 at the second time point based on the first machine learning model for the at least one cell.
According to paragraph 1,

The step of identifying the first performance information includes:

Comprising: identifying a performance change amount of the at least one cell based on the first performance information and second performance information indicating communication performance of the at least one cell before the first time point,

The step of determining whether the first performance information satisfies a specific condition is,

Method comprising determining whether the amount of performance change satisfies the specific condition.
According to claim 1 or 2,

Further comprising determining a group of the at least one cell as a first group,

The first machine learning model is a machine learning model corresponding to the first group including the at least one cell.
According to any one of claims 1 to 3,

identifying inference information including at least one of input data or output data of the first machine learning model for the control operation; and

The method further includes determining whether the environment of the at least one cell has changed based on the inference information.
According to clause 4,

The step of determining whether the environment of the at least one cell has changed,

Based on determining that the inference information is outside the reference data range for the first group, determining that the environment of the at least one cell has changed.
According to any one of claims 1 to 5,

The method further includes re-determining the group of the at least one cell based on determining that the environment of the at least one cell has changed.
According to any one of claims 1 to 6,

The step of determining whether to update the first machine learning model based on determining that the identified first performance information satisfies the specific condition includes,

A method comprising determining whether to update the first machine learning model based on determining that the environment of the at least one cell has changed.
In clause 7,

The step of determining whether to update the first machine learning model based on determining that the environment of the at least one cell has changed includes:

Based on re-determining the group of the at least one cell to the first group, determining an update to the first machine learning model.
According to paragraph 7 or 8,

The step of determining whether to update the first machine learning model based on determining that the environment of the at least one cell has changed includes:

Determining to freeze the first machine learning model based on re-determining the group of at least one cell into a second group,

The method wherein the machine learning model corresponding to the second group is a second machine learning model.
According to any one of claims 1 to 9,

The step of determining whether to update the first machine learning model based on determining that the identified first performance information satisfies the specific condition includes,

Based on determining that the environment of the at least one cell has not changed, determining an update to the first machine learning model.
A computer-readable recording medium on which a program for performing the method of any one of claims 1 to 10 is recorded on a computer.
In the SMO entity (service management and orchestration entity) 100,

A memory 1020 that stores one or more instructions; and

At least one processor 1010 that executes the one or more instructions stored in the memory 1020,

The at least one processor 1010 executes the one or more instructions,

Providing a first machine learning model to a target RIC (radio access network intelligence controller) 120 that controls at least one cell corresponding to the first machine learning model,

Identify first performance information indicating communication performance of the at least one cell at a first point in time,

Determine whether the identified first performance information satisfies a specific condition,

Based on determining that the identified first performance information satisfies the specific condition, determine whether to update the first machine learning model,

The first time point is after the control operation performed by the target RIC 120 at the second time point based on the first machine learning model for the at least one cell.
According to clause 12,

The at least one processor 1010,

Identifying a performance change amount of the at least one cell based on the first performance information and second performance information indicating communication performance of the at least one cell before the first time point,

An entity that determines whether the performance change satisfies the specific condition.
According to claim 12 or 13,

The at least one processor 1010,

Determine a group of the at least one cell as a first group,

The first machine learning model is a machine learning model corresponding to the first group including the at least one cell.
According to any one of claims 12 to 14,

The at least one processor 1010,

Identifying inference information including at least one of input data or output data of the first machine learning model for the control operation,

An entity that determines whether the environment of the at least one cell changes based on the inference information.