Method and apparatus for managing federated machine learning performance in mobile communication system

Abstract

The present disclosure relates to a method of operating a central NWDAF, comprising the steps of: receiving an analysis request message from a network function; determining, based on the analysis request message, a local NWDAF included in a federated learning (FL) group; an FL local behavior instruction is sent to a local NWDAF; receiving a response message from the local NWDAF in response to the FL local behavior instruction; receiving a local model generated by means of the local NWDAF from the local NWDAF based on the FL local behavior instruction; generating a global model based on the local model; and sending the global model to the NWDAF.

Description

Technical Field

[0001] The present disclosure relates to a wireless communication system, and more particularly, to a method and apparatus for managing federated machine learning performance in a mobile communication system. Background Art

[0002] 5G mobile communication technology defines a wide frequency band so that high transmission rates and new services are possible, and can be realized not only in "below 6 GHz" frequency bands such as 3.5 GHz, but also in "above 6 GHz" frequency bands called mmWave including 28 GHz and 39 GHz. In addition, 6G mobile communication technology called a super 5G system has been considered to be implemented in a terahertz (THz) frequency band (e.g., 95 GHz to 3 THz frequency band) in order to achieve a transmission rate fifty times faster than 5G mobile communication technology and an ultra-low latency one-tenth of 5G mobile communication technology.

[0003] At the beginning of the development of 5G mobile communication technology, in order to support services and meet performance requirements related to enhanced mobile broadband (eMBB), ultra-reliable low latency communications (URLLC), and massive machine type communications (mMTC), there has been ongoing standardization on beamforming and massive multiple-input multiple-output (MIMO) for mitigating radio wave path loss and increasing radio wave transmission distance in mmWave, parameter sets to support dynamic operation for efficient utilization of mmWave resources and time slot formats (e.g., operating multiple subcarrier spacings), initial access technologies to support multi-beam transmission and wideband, definition and operation of bandwidth parts (BWPs), new channel coding methods such as low-density parity-check (LDPC) codes for large amounts of data transmission and polar codes for highly reliable transmission of control information, L2 pre-processing, and network slicing for providing dedicated networks dedicated to specific services.

[0004] Currently, in view of the services that 5G mobile communication technology is to support, there are ongoing discussions on improvements and performance enhancements of initial 5G mobile communication technology, and there is already physical layer standardization on the technology, such as Vehicle-to-Everything (V2X) for assisting autonomous vehicles in making driving decisions based on information about the location and status of the vehicle sent by the vehicle and for enhancing user convenience, New Radio Unlicensed (NR-U) for system operation intended to comply with various regulatory requirements in unlicensed bands, NR UE power saving, Non-Terrestrial Network (NTN) as UE-satellite direct communication for providing coverage in areas where communication with terrestrial networks is unavailable, and positioning.

[0005] In addition, there has been ongoing standardization on air interface architecture / protocols for technologies such as Industrial Internet of Things (IIoT) for supporting new services through interworking and convergence with other industries, Integrated Access and Backhaul (IAB) for providing nodes for network service area expansion by supporting wireless backhaul links and access links in an integrated manner, mobility enhancements including conditional handover and dual active protocol stack (DAPS) handover, and two-step random access (NR's two-step RACH) for simplifying the random access procedure. There is also ongoing standardization on system architecture / services for 5G baseline architecture (e.g., service-based architecture or service-based interface) for combining network function virtualization (NFV) and software-defined networking (SDN) technologies, and mobile edge computing (MEC) for receiving services based on UE location.

[0006] As 5G mobile communication systems are commercialized, the already exponentially growing number of connected devices will be connected to the communication network, and accordingly it is expected that enhanced functionality and performance of the 5G mobile communication systems and integrated operations of the connected devices will be necessary. To this end, new research related to extended reality (XR) is arranged for effectively supporting augmented reality (AR), virtual reality (VR), mixed reality (MR), etc., improving 5G performance and reducing complexity by utilizing artificial intelligence (AI) and machine learning (ML), AI service support, metaverse service support, and drone communication.

[0007] Furthermore, such development of 5G mobile communication systems will serve as a foundation for developing not only new waveforms for providing terahertz band coverage for 6G mobile communication technology, multi-antenna transmission technologies such as full-dimensional MIMO (FD-MIMO), array antennas and massive antennas, metamaterial-based lenses and antennas for improving terahertz band signal coverage, high-dimensional spatial multiplexing technology using orbital angular momentum (OAM), and reconfigurable smart surface (RIS) technology, but also full-duplex technology for improving frequency efficiency and improving system networks for 6G mobile communication technology, AI-based communication technology for achieving system optimization by utilizing satellites and artificial intelligence (AI) from the design stage and internalizing end-to-end AI support functions, and next-generation distributed computing technology for realizing services at a level of complexity that exceeds the limits of UE operating capabilities by utilizing ultra-high-performance communication and computing resources. Summary of the invention

[0008] Technical issues

[0009] When a network data analysis function (NWDAF) service consumer requests an analysis service in a mobile network, the analysis service can be provided within the required time. In an NWDAF learning configuration that does not support federated learning, the training process is performed within a single entity, and the judgment on whether to provide services and time management can be performed within a given time. In an NWDAF learning configuration that supports federated learning, training is performed within multiple local NWDAFs, and the operation of collecting them into a central NWDAF can occur repeatedly. In this case, time management in the training and collection operations of each local NWDAF must be performed so that the training can be completed and the analysis service results can be provided within the time requested by the NWDAF service consumer.

[0010] In addition, when NWDAF provides analysis services through federated learning, since multiple NWDAFs participate in learning, it is necessary to monitor abnormal events that may occur in each NWDAF. The accuracy and completion time of federated learning can change depending on the events that occur in each local NWDAF participating in federated learning. In order to perform stable federated learning between NWDAFs, a method for monitoring and responding to abnormal events is required.

[0011] Technical Solutions

[0012] An operating method of a central network data analysis function (NWDAF) according to the present disclosure includes: receiving an analysis request message from a network function; determining a local NWDAF included in a federated learning (FL) group based on the analysis request message; sending FL local behavior instructions to the local NWDAF; receiving a response message from the local NWDAF in response to the FL local behavior instructions; receiving a local model generated by means of the local NWDAF from the local NWDAF based on the FL local behavior instructions; generating a global model based on the local model; and sending the global model to the NWDAF.

[0013] A central NWDAF according to the present disclosure includes: a network interface; and a processor connected to the network interface. The processor receives an analysis request message from a network function, determines a local NWDAF included in a federated learning (FL) group based on the analysis request message, sends a FL local behavior instruction to the local NWDAF, receives a response message from the local NWDAF in response to the FL local behavior instruction, receives a local model generated by the local NWDAF from the local NWDAF based on the FL local behavior instruction, generates a global model based on the local model, and sends the global model to the NWDAF.

[0014] Beneficial Effects

[0015] When providing analysis services through federated learning, NWDAF can complete learning and provide analysis service results within a given time. In addition, when an abnormal event occurs during the operation of performing federated learning, NWDAF can respond to the abnormal event and ultimately provide analysis service results stably. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a conceptual diagram illustrating a connection structure between NWDAFs for performing federated learning according to the present disclosure.

[0017] Figure 2 is a conceptual diagram showing a 5G system supporting NWDAF according to the present disclosure.

[0018] Figure 3 is a flowchart illustrating a process for performing time control at each operation of federated learning in a central NWDAF according to the present disclosure.

[0019] Figure 4 is a flow chart illustrating a process for generating a federated learning group to provide analysis services to a central NWDAF according to the present disclosure.

[0020] Figure 5 is a flow chart illustrating a process for monitoring a delay situation occurrence event in a local NWDAF and notifying the central NWDAF of the event according to the present disclosure.

[0021] Figure 6 is a block diagram illustrating a central NWDAF according to the present disclosure.

[0022] Figure 7 is a block diagram illustrating a local NWDAF 1 according to the present disclosure.

[0023] Figure 8 is a block diagram illustrating a local NWDAF 2 according to the present disclosure. DETAILED DESCRIPTION

[0024] Hereinafter, the operating principle of the present disclosure will be described in detail with reference to the accompanying drawings. The terms described below are defined in consideration of the functions in the present disclosure and may differ according to the intention or habits of the user or operator. Therefore, the definition of the terms should be based on the content in the entire specification.

[0025] As used herein, for the sake of convenience of description, terms about network entities and edge computing system entities, terms about messages, terms about identification information, etc. are illustratively used. Therefore, the present disclosure is not limited to the terms used below, and other terms about the subject matter with equivalent technical meanings may be used.

[0026] In the following description, for the convenience of description, the present disclosure will be described using the terms and names defined in the 5G system standard. However, the present disclosure is not limited to these terms and names, and can be applied to systems conforming to other standards in the same manner.

[0027] The present disclosure proposes a method for supporting federated learning in a mobile communication system. Specifically, in a structure for performing federated learning between a central NWDAF and multiple local NWDAFs, the present disclosure proposes a time coordination method for providing analysis within a given time and a method for monitoring and responding to the occurrence of abnormal events during learning.

[0028] Figure 1 is a conceptual diagram illustrating a connection structure between NWDAFs for performing federated learning according to the present disclosure.

[0029] about Figure 1 , multiple local NWDAFs 1110, 1120, 1130 can be connected to each other based on the central NWDAF 1000 to perform federated learning. The local NWDAFs 1110, 1120, 1130 may include a model training logic function (MTLF), and may perform machine learning (ML) model training by using local data source network functions (NFs) 1210, 1220, 1230 (e.g., NFs connected to the local NWDAFs and that can provide data collected by them) or an operation, administration, and maintenance (OAM) system connected thereto, etc.

[0030] The central NWDAF 1000 may collect and calculate the learned ML model from the locally distributed local NWDAFs 1110, 1120, 1130, and thus generate a global machine learning (ML) model (or an aggregated machine learning model). The central NWDAF 1000 may distribute the generated global ML model to the local NWDAFs 1110, 1120, 1130 again. The local NWDAFs 1110, 1120, 1130 may train a local ML model based on the global ML model, and send the local ML model updated as above to the central NWDAF 1000. The central NWDAF 1000 and the local NWDAFs 1110, 1120, 1130 may repeatedly perform the above operations until a predetermined model training condition is met.

[0031] Figure 2 is a conceptual diagram showing a 5G system supporting NWDAF according to the present disclosure.

[0032] refer to Figure 2, the architecture of the 5G system supporting the NWDAF 1000 may include various network functions (NFs). For example, the architecture of the 5G system may include an access and mobility management function (AMF) 500, a session management function (SMF) 600, a policy control function (PCF) 700, an application function (AF) 800, a unified data management (UDM) 900, a data network (DN) 400, a user plane function (UPF) 300, a (radio) access network ((R)AN) 200, and a user equipment (UE) 100.

[0033] For example, each NF can support the following functions.

[0034] -AMF 500 may provide functionality for UE-based access and mobility management, and each UE may be connected to one AMF by default.

[0035] -DN 400 may mean, for example, a network outside 5GS, including operator services, Internet access, third-party services, etc. DN 400 may send a downlink protocol data unit (PDU) to UPF 300, or receive a PDU sent from UE 100 from UPF 300.

[0036] -PCF 700 may receive packet flow information from an application server and provide functionality for determining policies such as mobility management and session management. Specifically, PCF 700 may support functionality such as supporting a unified policy framework for controlling network operations, providing policy rules to allow control plane functions (e.g., AMF 500, SMF 600, etc.) to implement policy rules, and implementing a front end for accessing relevant subscription information to determine policies within a user data repository (UDR).

[0037] -SMF 600 can provide session management functions, and in the case where UE 100 has multiple sessions, each session can be managed by a different SMF.

[0038] -UDM 900 can store user subscription data, policy data, etc.

[0039] The UPF 300 may forward the downlink PDU received from the DN 400 to the UE 100 via the (R)AN 200 , and may forward the uplink PDU received from the UE 100 to the DN 400 via the (R)AN 200 .

[0040] - AF 800 can interoperate with the 3GPP core network to provide services (e.g., to support functions such as influencing applications on service routing, accessing network capability exposure, and interacting with the policy framework for policy control).

[0041] -NWDAF 1000 may include at least one of an analysis logic function (AnLF) and a model training logic function (MTLF), and may collect data in the 5G system and provide analysis services based on the collected data. For example, it may train an ML model based on the data and provide analysis data required for network management and operation based on the trained model.

[0042] Figure 3 is a flowchart illustrating a process for performing time control at each operation of federated learning in a central NWDAF according to the present disclosure.

[0043] about Figure 3 In operation 1, the central NWDAF 1000 may generate a federated learning (FL) group. For example, the central NWDAF 1000 may receive an analysis request from an analysis service consumer (e.g., a network function or another NWDAF). The central NWDAF 1000 may determine to perform federated learning based on the analysis request. The central NWDAF 1000 may select local NWDAFs 1110, 1120 that may participate in the federated learning based on the determination. The central NWDAF 1000 may generate a federated learning group (FL group) based on the selected local NWDAFs 1110, 1120.

[0044] In operation 2, the central NWDAF 1000 may provide FL local behavior instructions to the local NWDAFs 1110, 1120 selected in operation 1 for performing federated learning. For example, the central NWDAF 1000 may send a message including the FL local behavior instructions to the local NWDAFs 1110, 1120. For example, the FL local behavior instructions may include the following information:

[0045] -ML model information: federated learning model parameters, model structure, possible model acquisition address (URL for downloading ML model)

[0046] - Parameters required to generate the global model: the number of data samples, gradient values, weight values, or pre-committed profile index values ​​used to generate the local model

[0047] - A list of events that need to be monitored and reported: event IDs, including an event ID value representing at least one of the events, for example, late transmission (a reporting event notifying that the local model training time in the local NWDAF 1110, 1120 has increased and the local model supply time to the central NWDAF 1000 has been delayed), low accuracy (a reporting event notifying that the accuracy of the local model trained in the local NWDAF 1110, 1120 is lower than the reference in the case of indicating monitoring and reporting of low accuracy events), the accuracy threshold may be provided together with the event identifier), data shortage occurrence (a reporting event that data samples cannot be obtained due to an abnormal state of a data source that provides data samples for local model training in the local NWDAF 1110, 1120), etc. Such events may be expressed as a straggling event (a discrete event), and a subscription may be formed in the form of an implicit subscription to notify the discrete event while providing FL local behavior instructions instead of a separate subscription service. Alternatively, the subscription for notifying the discrete event may be performed by a separate subscription request.

[0048] - Epoch time request information

[0049] a) Expected epoch time or expected local ML model training and transfer time: When performing federated learning, the expected time value from performing local model training in the local NWDAF 1110, 1120 to transferring the trained local model to the central NWDAF 1000

[0050] b) Local ML model training time requirement: the required value of the time taken from completing local model training in the local NWDAF 1110, 1120 to sending the trained local model to the central NWDAF 1000

[0051] The expected epoch time or the expected local ML model training and transmission time or the local ML model training time requirement value may be provided as different values ​​according to the local NWDAFs 1110, 1120 participating in the federated learning, or may be provided as a unified single value. Figure 4The detailed process involved is described. For example, based on the load status of each NWDAF 1110, 1120, the number of data samples, etc., the expected epoch time or the expected local ML model training and transmission time or the local ML model training time requirement value can be configured as a different value for each local NWDAF 1110, 1120, which may affect the time taken for local model training. For example, the central NWDAF 1000 may determine the size of the federated learning epoch time requirement proportional to the size of the data set included in the data information available for ML model training of the local NWDAF 1110, 1120 managed by the local NWDAF 1110, 1120, the number of network function instances or the size of the network function set that can provide data by connecting to the local NWDAF 1110, 1120, the number of PDU sessions from which the local NWDAF 1110, 1120 can collect data, the number of data network names managed by the local NWDAF 1110, 1120, and the number of data network access identifiers. The information used by the central NWDAF 1000 to determine the size of the federated learning epoch time requirement value may be obtained from the network repository function (NRF) 1300 in the local NWDAF search operation used to generate the federated learning group, or may be obtained directly from the local NWDAF 1110, 1120 searching as a member of the federated learning group.

[0052] The local NWDAFs 1110 and 1120 may respectively perform data collection based on the FL local behavior instructions from the central NWDAF 1000. For example, the local NWDAFs 1110 and 1120 may determine the load status or data sample information (the number of data samples or the expected number of obtainable data samples, the size of the data collection area (the number of tracking area IDs), and the number of network function instances that can be used as data sources) of the local NWDAFs 1110 and 1120 through data collection.

[0053] In operation 3, the local NWDAF 1110, 1120 may send a response message to the FL local behavior instruction to the central NWDAF 1000. For example, the local NWDAF 1110, 1120 may determine whether the expected epoch time, the expected local ML model training and transmission time, or the local ML model training time requirement included in the FL local behavior instruction received from the central NWDAF 1000 can be met. For example, the local NWDAF 1110, 1120 may determine whether the expected epoch time, the expected local ML model training and transmission time, or the local ML model training time requirement included in the FL local behavior instruction can be met based on the load status or data sample information (the number of data samples or the expected number of obtainable data samples, the size of the data collection area (the number of tracking area IDs), and the number of network function instances that can be used as data sources) of the local NWDAF 1110, 1120.

[0054] In case the requirements are fully met, the local NWDAF 1110 , 1120 may send a response message including an indicator indicating that the requirements are fully met to the central NWDAF 1000 .

[0055] If the local NWDAF 1110, 1120 determines that the requirement cannot be fully met, it may (1) generate a new corrected time value and send a response message including the corrected time value to the central NWDAF 1000, or (2) send a response message including an indicator indicating that the requirement cannot be fully met and data sample information to the central NWDAF 1000. The central NWDAF 1000 may receive the response message from the local NWDAF 1110, 1120. The central NWDAF 1000 may determine a new epoch time requirement based on the response message.

[0056] The response to the epoch time request received from the central NWDAF 1000 may vary depending on the local NWDAF 1110 , 1120 .

[0057] In operation 4, the central NWDAF 1000 may provide the new desired epoch time requirement to the local NWDAFs 1110, 1120. For example, the central NWDAF 1000 may send a message including the new desired epoch time requirement to the local NWDAFs 1110, 1120. For example, the central NWDAF 1000 may determine the new epoch time requirement based on a response message to the epoch time requirement received from the local NWDAFs 1110, 1120. The central NWDAF 1000 may send a message including a value corresponding to the determined new epoch time requirement and an indicator indicating the start of federated learning to the local NWDAFs 1110, 1120. The message may include a timestamp value (absolute time value) for performing synchronous federated learning.

[0058] For example, the central NWDAF 1000 may determine the maximum epoch time value in the corrected epoch time requirement included in the response message received from the local NWDAFs 1110, 1120 as the new epoch time value. The central NWDAF 1000 may generate a new epoch time requirement based on the determined new epoch time value. The central NWDAF 1000 may provide the generated new epoch time requirement to the local NWDAFs 1110, 1120 participating in the federated learning.

[0059] Alternatively, the central NWDAF 1000 may determine to exclude from the federated learning a local NWDAF that has responded with an epoch time value exceeding a predetermined threshold. The central NWDAF 1000 may send a message including an indicator of not participating in the federated learning to the local NWDAF determined to be excluded from the federated learning.

[0060] The central NWDAF 1000 may perform the above operations on the corrected epoch time requirement provided from the local NWDAFs 1110 , 1120 based on operator policies or predefined rules.

[0061] The local NWDAF 1110, 1120 may receive a message including a new desired epoch time requirement from the central NWDAF 1000. The local NWDAF 1110, 1120 may perform local model training based on the new epoch time requirement and the timestamp value included in the message received from the central NWDAF 1000. The local NWDAF 1110, 1120, 1130 may start a timer set to the epoch time requirement value at the same time as the local model training performance begins.

[0062] In operation 5, the local NWDAFs 1110, 1120 may send the completed ML local model to the central NWDAF 1000. In addition, the local NWDAFs 1110, 1120, 1130 may send required parameters for generating a global model to the central NWDAF 1000 based on the FL local behavior instruction received in operation 2.

[0063] The central NWDAF 1000 may receive the completed ML local model from each of the local NWDAFs 1110 , 1120 .

[0064] In operation 6, the central NWDAF 1000 may send, to the local NWDAFs 1110 , 1120 , confirmation messages for the local models received from the local NWDAFs 1110 , 1120 .

[0065] The local NWDAFs 1110, 1120 may receive a confirmation message from the central NWDAF 1000. The central NWDAF 1000 may generate a global model based on the ML local models received from the local NWDAFs 1110, 1120 and parameter values ​​required to generate the global model.

[0066] In operation 7, the central NWDAF 1000 may transmit the generated global model, an epoch ID (an identifier indicating which epoch number is a corresponding epoch), and a timestamp value to the local NWDAFs 1110, 1120 participating in the federated learning.

[0067] Figure 4 is a flow chart illustrating a process for generating a federated learning group to provide analysis services to NFs at a central NWDAF according to the present disclosure.

[0068] refer to Figure 4 , the NF 1200 may send an analysis request message to the C-NWDAF 1000. The NF 1200 may be at least one of the data source NF 1210, the data source NF 1220, and the data source NF 1230. The NF 1200 may receive the analysis request message from the C-NWDAF 1000. The C-NWDAF 1000 may determine to perform FL.

[0069] In operation 1, the local NWDAF 1110, 1120, 1130 that may participate in federated learning and serve as a local NWDAF may send a NF registration request message to the NRF 1300. The NRF 1300 may receive the NF registration request message from the local NWDAF 1110, 1120, 1130. The NF registration request message may include the following information.

[0070] -NWDAF identifier and address information

[0071] - Supported model IDs and information: Supported ML model identifiers and meta information (e.g., may include supported model sizes, number of parameters, etc.)

[0072] -Supported analysis IDs: Analysis IDs provided by federated learning

[0073] -Data information available for ML model training: (information of data sets available at the local NWDAF): number of data samples or data set size, a list of identifiers of network functions (data source NF) connected to the local NWDAF 1110, 1120, 1130 and capable of providing data, type and number or size of instances of a network function set (cardinality of NF sets associated with the local NWDAF), which may be in the form of a number or list of identifiers of PDU sessions and include information about the number of PDU sessions for collecting data by the local NWDAF 1110, 1120, 1130), number of UEs managed by the local NWDAF 1110, 1120, 1130 (number of UEs for collecting data by the local NWDAF 1110, 1120, 1130), network slice instance information managed by the local NWDAF 1110, 1120, 1130 (number of network slice instances, S-NSSAI list), data distribution information (e.g., data distribution information). May be information about how independently and identically distributed the data held or available by the local NWDAF 1110, 1120, 1130 are as a data distribution property; may indicate the degree of independence and identical distribution (iid) of the data sets, or the degree of (non) IID in the form of "low", "medium", "high", or "highest". Data distribution information may be included as some parameter values ​​of the statistical properties of the data

[0074] - Service area (tracking area ID list, cell ID list, etc.)

[0075] -Data network name or data network access identifier: data network information for the local NWDAF 1110, 1120, 1130 to collect data and provide analysis services

[0076] In operation 2, the NRF 1300 may send a NF registration response message to the local NWDAF 1110, 1120, 1130. For example, the NRF 1300 may store information of the local NWDAF 1110, 1120, 1130. The NRF 1300 may send a response message including whether the successful registration of the local NWDAF 1110, 1120, 1130 is completed to the local NWDAF 1110, 1120, 1130. The local NWDAF 1110, 1120, 1130 may receive the response message from the NRF 1300.

[0077] In operation 3, the central NWDAF 1000 may send an NF search request message to the NRF 1300. For example, the central NWDAF 1000 may receive an analysis request message from an unspecified NF (any NF). For example, the NF may be Figure 2 At least one of the UE 100, (R) AN 200, UPF 300, DN 400, AMF 500, SMF 600, PCF 700, AF 800 and UDM 900 in the NRF 1300. For example, the analysis request message may include at least one of the analysis ID, the FL performance request indicator, the data sharing permission indicator provided by the NF, the data distribution information owned or provided by the NF, the analysis service provision time requirement and the analysis service accuracy requirement. For example, if the analysis service provision time requirement is included in the analysis request message received from the NF, the central NWDAF 1000 may send an NWDAF search request message including the corresponding requirement to the NRF 1300. In response to the NWDAF search request message, the central NWDAF 1000 may receive a response message including a list of NWDAFs that can meet the corresponding requirement from the NRF 1300. Alternatively, the central NWDAF 1000 may send an NWDAF search request message not including the corresponding requirement to the NRF 1300. The central NWDAF 1000 may determine the local NWDAFs 1110 , 1120 , 1130 that may satisfy the corresponding requirements based on a previously stored list of available local NWDAFs.

[0078] The central NWDAF 1000 may generate an NF search request message based on the analysis request message received from the NF. For example, the NF search request message may include the following information.

[0079] -FL Aggregator Capability: Information for requesting information about the central NWDAF 1000 serving as a FL aggregator from the NRF 1300

[0080] -Federated Learning Indicator: An indicator used to request information about NWDAFs that may be involved in a flight.

[0081] - Central NWDAF information (address and ID)

[0082] -UE identifier or UE group identifier (may be included in case the analysis requested by the NF corresponds to a specific UE or UE group)

[0083] -Supported ML model information: ML model information (identifier, etc.) supported in the central NWDAF 1000 and used for federated learning, and since training needs to be performed on the same model according to the characteristics of federated learning, it can be used to search for a local NWDAF (for local NWDAFs participating in the FL group) that supports the same model as the model supported by the central NWDAF 1000 (supported by the central NWDAF 1000 and performs local ML model training).

[0084] - Analysis ID: The analysis identifier requested by the NF from the central NWDAF 1000, the analysis information used to perform federated learning

[0085] -Requirements for data information available for ML model training: It may have data information values ​​available for ML model training specified in operation 1. It may be used to obtain data that meets the corresponding requirements or select an available NWDAF. For example, the following information may be included: the number of data samples or the size of the data set, a list of identifiers of network functions (data source NF) that are connected to the local NWDAF and can provide data, the type and number or size of instances of the network function set (the cardinality of the NF set associated with the local NWDAF), and PDU session information managed by the local NWDAF (e.g., PDU session information). For example, it may be in the form of a list of numbers and PDU session identifiers, and include information on the number of PDU sessions for which the local NWDAF collects data), the number of UEs managed by the local NWDAF (the number of UEs for which the local NWDAF collects data), and network slice instance information managed by the local NWDAF (the number of network slice instances, S-NSSAI list).

[0086] In operation 4, the NRF 1300 may send an NF search response message to the central NWDAF 1000. The central NWDAF 1000 may receive the NF search response message from the NRF 1300. For example, the NRF 1300 may send an NF search response message to the central NWDAF 1000, the NF search response message including local NWDAF information corresponding to the NF search request message received from the central NWDAF 1000. For example, the local NWDAF information (which may provide information on a plurality of local NWDAFs) may include the following.

[0087] - List of local NWDAF identifiers and addresses

[0088] -Data information available for ML model training for each local NWDAF: the number of data samples or data set size, a list of identifiers of network functions (data source NF) that can connect to the local NWDAF and provide data, the type of network function set and the number or size of instances (the cardinality of NF sets associated with the local NWDAF), PDU session information managed by the local NWDAF (e.g., the number of data samples or data set size. For example, it can be in the form of the number and list of PDU session identifiers and include information about the number of PDU sessions from which the local NWDAF can collect data), the number of UEs managed by the local NWDAF (the number of UEs from which the local NWDAF can collect data), network slice instance information managed by the local NWDAF (the number of network slice instances, S-NSSAI list), a list of data network names managed by the local NWDAF or a list of data network access identifiers.

[0089] (For example, being managed by NWDAF may mean that ML model training may be performed by collecting corresponding data and / or learning from the collected data)

[0090] -ML model information of each local NWDAF supported (Supported ML model information)

[0091] The central NWDAF 1000 may generate an FL group based on the NF search response message received from the NRF 1300. For example, the operation of generating the FL group by the central NWDAF 1000 may be similar to the operation of generating the FL group by the central NWDAF 1000. Figure 3 The operation is the same as or similar to the operation described in operation 1.

[0092] In operation 5, the central NWDAF 1000 may send a FL request message including a FL local behavior instruction to the local NWDAFs 1110, 1120, 1130. The local NWDAFs 1110, 1120, 1130 may receive a FL request message including a FL local behavior instruction from the central NWDAF 1000. For example, the operations of sending and receiving the FL request message may be similar to those described in connection with the embodiment of the present invention. Figure 3The operation of operation 2 in the above example is the same as or similar to that of operation 2 in the above example. For example, the central NWDAF 1000 may determine an epoch time requirement value (a requirement value of the time taken to complete local ML model training on the local NWDAFs 1110, 1120, 1130 in federated learning) based on the data information of the ML model training of the local NWDAFs available for search. For example, the central NWDAF 1000 may determine the size of the federated learning epoch time requirement value proportional to the size of the data set included in the data information of the ML model training of the local NWDAFs 1110, 1120, 1130 available for search, the number of network function instances or the size of the network function set that can be connected to the local NWDAFs 110, 1120, 1130 and provide data, the number of PDU sessions from which the local NWDAFs 110, 1120, 1130 can collect data, the number of data network names managed by the local NWDAFs 1110, 1120, 1130, and the number of data network access identifiers. The central NWDAF 1000 may send a FL request message including a federated learning epoch time requirement value to the local NWDAFs 1110 , 1120 , 1130 .

[0093] In operation 6, the local NWDAF 1110, 1120, 1130 may send a FL response message to the central NWDAF 1000 in response to the FL request message. The central NWDAF 1000 may receive the FL response message from the local NWDAF 1110, 1120, 1130. For example, the local NWDAF 1110, 1120, 1130 may perform an operation of sending a response to the epoch time to the central NWDAF 1000, such as regarding Figure 3 as described in Operation 3.

[0094] Figure 5 is a flow chart illustrating a process for monitoring a delay situation occurrence event in a local NWDAF and notifying the central NWDAF of the event according to the present disclosure.

[0095] about Figure 5 , in operation 1, the central NWDAF 1000 may generate a FL group by selecting local NWDAFs 1110, 1120 that may participate in federated learning (FL group creation).

[0096] In operation 2, the central NWDAF 1000 may send FL local behavior instructions to the local NWDAFs 1110, 1120 participating in the federated learning. The local NWDAFs 1110, 1120 may receive the FL local behavior instructions from the central NWDAF 1000. The FL local behavior instructions may include an indicator indicating that a report on an abnormal event or a discrete event occurring during the federated learning is performed. For example, the FL local behavior instructions may include the following information.

[0097] - Notify the destination address

[0098] - A list of events that need to be monitored and reported: event ID (one or more), which may include, for example, an event ID value representing at least one of a late transmission event (a reporting event notifying that the local model training time is delayed in the local NWDAF 1110, 1120 and thus the time to send the local model to the central NWDAF 1000 is delayed), a low accuracy event (a reporting event notifying that the accuracy of the local model trained on the local NWDAF 1110, 1120 is lower than a reference value), in the case of indicating monitoring and reporting of low accuracy events, an accuracy threshold may be sent together with the event ID, a data shortage occurrence event (a reporting event notifying that it is difficult to obtain data samples due to an abnormal state of a data source that provides data samples for local model training in the local NWDAF 1110, 1120), etc. The above events may be expressed as discrete events, and in the case of sending FL local behavior instructions instead of a separate subscription service, a subscription service may be formed in the form of an implicit subscription to notify discrete events. Alternatively, the subscription for notifying a discrete event service may be performed through a separate subscription request message.

[0099] In the case where a report on an event related to late transmission (local model transmission delay due to a local model training delay event) is indicated in the above-mentioned list of things to be monitored and reported, epoch time requirement information for transmitting a reference time value may be provided together. For example, the epoch time requirement information may be expressed in the following form.

[0100] - Epoch time request information

[0101] a) Expected epoch time or expected local ML model training and transfer time: When performing federated learning, the expected time value from performing local model training in the local NWDAF 1110, 1120 to transferring the trained local model to the central NWDAF 1000

[0102] b) Local ML model training time requirement: the required value of the time taken from completing local model training in the local NWDAF 1110, 1120 to sending the trained local model to the central NWDAF 1000

[0103] In operation 3, the local NWDAF 1110, 1120 may send each response message for the FL local behavior instruction to the central NWDAF 1000. The central NWDAF 1000 may receive each response message for the FL local behavior instruction from the local NWDAF 1110, 1120. Each response message may include whether the local NWDAF 1110, 1120 has successfully completed its subscription, and a subscription ID.

[0104] After each response message of the FL local behavior instruction is sent, a timer for the start of the epoch time may be operated in the central NWDAF 1000 and the local NWDAFs 1110 , 1120 .

[0105] The local NWDAF 1110, 1120 may perform each local model training operation.

[0106] As another example, the local NWDAF 1110, 1120 may send a response to the FL local behavior instruction to the central NWDAF 1000 before the timer operation for starting the epoch time, such as in Figure 3 In this case, the local NWDAF 1110, 1120 may operate a timer for starting the epoch time after receiving a new desired epoch time requirement from the central NWDAF 1000, as in operation 3 of FIG. Figure 3 Just like in operation 4.

[0107] In operation 4-1, the local NWDAF 1 1110 may send the ML local model to the central NWDAF 1000. The central NWDAF 1000 may receive the ML local model from the local NWDAF 1 1110. For example, the local NWDAF 1 1110 may successfully complete the local training within the time indicated by the epoch time requirement. The local NWDAF 1 1110 may generate the ML local model based on the result of completing the local training. The local NWDAF 1 1110 may send the generated ML local model to the central NWDAF 1000.

[0108] In operation 4-2, the central NWDAF 1000 may send an acknowledgement (ACK) message to the local NWDAF 1 1110 to notify that the ML local model has been successfully received. The local NWDAF 1 1110 may receive the acknowledgement message from the central NWDAF 1000.

[0109] In operation 5, the local NWDAF 2 1120 may send a delay situation occurrence event report message to the central NWDAF 1000. The central NWDAF 1000 may receive the delay situation occurrence event report message from the local NWDAF 2 1120. For example, the local NWDAF 2 1120 may determine that the time indicated by the epoch time requirement is approaching and cannot meet the requirement. For example, the local NWDAF 2 1120 may determine that the local training has been completed and the local model cannot be sent to the central NWDAF 1000 within the time indicated by the epoch time requirement. The local NWDAF 2 1120 may generate a delay situation occurrence event report message based on the above determination. For example, the delay situation occurrence event report message may include an event ID, a subscription ID, and a reason code to determine the cause of the event occurrence.

[0110] For example, the local NWDAF 2 1120 may determine a given situation to determine an event ID and a cause code. For example, the local NWDAF 2 1120 may determine that the local model training is being performed normally and the accuracy is improving during the training process. In this case, the local NWDAF 2 1120 may generate a delayed situation occurrence event report message including a delayed transmission event identifier and an additional time requirement cause code. As another example, the local NWDAF 2 1120 may detect that during the local model training, an instance of a network function that is a data source connected to the local NWDAF 2 1120 is terminated or can no longer transmit data samples due to an overload. In this case, the local NWDAF 2 1120 may generate a delayed situation occurrence event report message including a data shortage event identifier and a cause code indicating that the data source is in an abnormal state. Alternatively, in the case where the accuracy value is lower than a predetermined threshold accuracy value during the local model training, the local NWDAF 2 1120 may generate a delayed situation occurrence event report message including a low accuracy event identifier.

[0111] The central NWDAF 1000 may send a confirmation message to the local NWDAF 2 1120 in response to the delay situation occurrence event report message received from the local NWDAF 2 1120. The local NWDAF 2 1120 may receive the confirmation message from the central NWDAF 1000.

[0112] In operation 6, the central NWDAF 1000 may generate a global model. For example, the central NWDAF 1000 may generate a global model based on the ML local model received from the local NWDAF 1 1110 and the delay situation occurrence event report message received from the local NWDAF 2 1120. For example, the central NWDAF 1000 may determine that the local model of the local NWDAF 2 1120 cannot be received within the epoch time based on the delay situation occurrence event report message received from the local NWDAF 2 1120. For example, the central NWDAF 1000 may determine, based on the reason code received from the local NWDAF 2 1120, whether to generate a global model using only some of the ML local models received from the local NWDAF 2 1120 or to exclude the local NWDAF 2 1120 from federated learning. For example, in the case where the central NWDAF 1000 receives a delay situation occurrence event report message including an event ID of late transmission and a reason code indicating that additional time is required, the central NWDAF 1000 may generate a global model using only local models that have been received from the local NWDAF 2 1120, excluding local models whose transmission is delayed from the local NWDAF 2 1120. As another example, in the case where the central NWDAF 1000 receives a delay situation occurrence event report message in which the event ID indicates data shortage or low accuracy and the reason code indicates an abnormal state of a data source, the central NWDAF 1000 may determine to exclude the local NWDAF 2 1120 from federated learning.

[0113] In operation 7-1, the local NWDAF 2 1120 may transmit the transmission-delayed ML local model to the central NWDAF 1000. The central NWDAF 1000 may receive the transmission-delayed ML local model from the local NWDAF 2 1120. For example, the local NWDAF 2 1120 may complete delayed local training after receiving the delay situation occurrence event report message including the delayed transmission event in operation 5. The local NWDAF 2 1120 may generate the transmission-delayed local model based on the delayed completed local training. The local NWDAF 2 1120 may transmit the transmission-delayed local model to the central NWDAF 1000. The local NWDAF 2 1120 may transmit the epoch time ID indicating at which epoch the transmission-delayed local model was generated together with the transmission-delayed local model to the central NWDAF 1000. The central NWDAF 1000 may receive the transmission-delayed local model and the epoch time ID from the local NWDAF 2 1120.

[0114] In operation 7-2, the central NWDAF 1000 may send a confirmation message to the local NWDAF 2 1120. The local NWDAF 2 1120 may receive the confirmation message from the central NWDAF 1000. For example, the confirmation message may include an indicator indicating that the global generation model has been completed.

[0115] In operation 8, the central NWDAF 1000 may transmit the generated global model to the local NWDAFs 1110, 1120. The local NWDAFs 1110, 1120 may receive the global model from the central NWDAF 1000. The global model may not be transmitted to the local NWDAFs excluded from the federated learning. The central NWDAF 1000 may transmit the epoch time ID together with the generated global model. The local NWDAFs 1110, 1120 may receive the global model and the epoch time ID from the central NWDAF 1000. The operation of operation 8 may be performed before the operation of operation 7 based on the local NWDAF 2 1120.

[0116] If the local NWDAF 2 1120 transmits a delay situation occurrence event report message to the central NWDAF 1000, and then transmits the local model whose transmission is delayed to the central NWDAF 1000, the central NWDAF 1000 may buffer the local model whose transmission is delayed and reflect the local model whose transmission is delayed in the global model generated in the next epoch. When buffering the local model whose transmission is delayed, the central NWDAF 1000 may store the local NWDAF identifier and the epoch time ID (which may be used to record information at which epoch step / time the corresponding local model should be provided) that have provided the corresponding local model by mapping them.

[0117] The local NWDAF 1110, 1120 may update the local model based on the received global model. The local NWDAF 1110, 1120 may perform additional local training on the updated local model.

[0118] Figure 6 is a block diagram illustrating a central NWDAF 1000 according to the present disclosure.

[0119] about Figure 6 , the central NWDAF 1000 according to the present disclosure may include a controller 1002 that controls the overall operation of the central NWDAF 1000, a network interface 1001 including a transmitter and a receiver, and a memory 1003. Obviously, it is not limited to the above example, and the central NWDAF 1000 may include more than Figure 6 The compositions shown may be more or less than those shown.

[0120] According to the present disclosure, the network interface 1001 may transmit and receive signals with at least one of the other network entities 100, 200, 300, 400, 500, 600, 700, 800, 900, 1210, 1220, 1230, 1300 and the local NWDAF 1110, 1120, 1130. The signals transmitted and received with at least one of the other network entities 100, 200, 300, 400, 500, 600, 700, 800, 900, 1210, 1220, 1230, 1300 and the local NWDAF 1110, 1120, 1130 may include control information and data.

[0121] According to the present disclosure, the controller 1002 can control the central NWDAF 1000 to perform the above Figures 3 to 5 At the same time, the controller 1002, the memory 1003 and the network interface 1001 do not necessarily have to be implemented as separate modules, and can obviously be implemented as a single component in the form of a single chip. In addition, the controller 1002 and the network interface 1001 can be electrically connected. In addition, the controller 1002 can be an application processor (AP), a communication processor (CP), a circuit, a dedicated circuit, or at least one processor.

[0122] According to the present disclosure, the memory 1003 may store data such as basic programs, application programs, and configuration information for the operation of the central NWDAF 1000. In particular, the memory 1003 provides the stored data according to the request of the controller 1002. The memory 1003 may be composed of a storage medium such as a ROM, a RAM, a hard disk, a CD-ROM, and a DVD, or a combination thereof. In addition, the memory 1003 may be plural. In addition, the controller 1002 may execute the above-mentioned embodiments based on the program for executing the above-mentioned embodiments of the present disclosure stored in the memory 1003.

[0123] Figure 7 is a block diagram illustrating a local NWDAF 1 1110 according to the present disclosure.

[0124] about Figure 7 , the local NWDAF 1 1110 according to the present disclosure may include a controller 1112 that controls the overall operation of the local NWDAF 1 1110, a network interface 1111 including a transmitter and a receiver, and a memory 1113. Obviously, it is not limited to the above example, and the local NWDAF 1 1110 may include more than Figure 7 The compositions shown in the figure may be more or less than those shown in the figure.

[0125] According to the present disclosure, the network interface 1111 may transmit and receive signals with at least one of the other network entities 100, 200, 300, 400, 500, 600, 700, 800, 900, 1210, 1220, 1230, 1300 and the central NWDAF 1000. The signals transmitted and received with at least one of the other network entities 100, 200, 300, 400, 500, 600, 700, 800, 900, 1210, 1220, 1230, 1300 and the central NWDAF 1000 may include control information and data.

[0126] According to the present disclosure, the controller 1112 may control the local NWDAF 1110 to perform the above Figures 3 to 5 At the same time, the controller 1112, the memory 1113, and the network interface 1111 do not necessarily have to be implemented as separate modules, and can obviously be implemented as a single component in the form of a single chip. In addition, the controller 1112 and the network interface 1111 can be electrically connected. In addition, the controller 1112 can be an application processor (AP), a communication processor (CP), a circuit, a dedicated circuit, or at least one processor.

[0127] According to the present disclosure, the memory 1113 may store data such as basic programs, application programs, and configuration information for the operation of the local NWDAF 11110. In particular, the memory 1113 provides the stored data according to the request of the controller 1112. The memory 1113 may be composed of a storage medium such as a ROM, a RAM, a hard disk, a CD-ROM, and a DVD, or a combination thereof. In addition, the memory 1113 may be multiple. In addition, the controller 1112 may execute the above-mentioned embodiments based on the program for executing the above-mentioned embodiments of the present disclosure stored in the memory 1113.

[0128] Figure 8 is a block diagram illustrating a local NWDAF 2 1120 according to the present disclosure.

[0129] about Figure 8 , the local NWDAF 2 1120 according to the present disclosure may include a controller 1122 that controls the overall operation of the local NWDAF 2 1120, a network interface 1121 including a transmitter and a receiver, and a memory 1123. Obviously, it is not limited to the above example, and the local NWDAF 2 1120 may include more Figure 8 The compositions shown in the figure may be more or less than those shown in the figure.

[0130] According to the present disclosure, the network interface 1121 may transmit and receive signals with at least one of the other network entities 100, 200, 300, 400, 500, 600, 700, 800, 900, 1210, 1220, 1230, 1300 and the central NWDAF 1000. The signals transmitted and received with at least one of the other network entities 100, 200, 300, 400, 500, 600, 700, 800, 900, 1210, 1220, 1230, 1300 and the central NWDAF 1000 may include control information and data.

[0131] According to the present disclosure, the controller 1122 may control the local NWDAF 2 1120 to perform the above Figures 3 to 5 At the same time, the controller 1122, the memory 1123, and the network interface 1121 do not necessarily have to be implemented as separate modules, and can obviously be implemented as a single component in the form of a single chip. In addition, the controller 1122 and the network interface 1121 can be electrically connected. In addition, the controller 1122 can be an application processor (AP), a communication processor (CP), a circuit, a dedicated circuit, or at least one processor.

[0132] According to the present disclosure, the memory 1123 may store data such as basic programs, applications, and configuration information for the operation of the local NWDAF 2 1120. In particular, the memory 1123 provides the stored data according to the request of the controller 1122. The memory 1123 may be composed of a storage medium such as a ROM, a RAM, a hard disk, a CD-ROM, and a DVD, or a combination thereof. In addition, the memory 1123 may be multiple. In addition, the controller 1122 may execute the above-mentioned embodiments based on the program for executing the above-mentioned embodiments of the present disclosure stored in the memory 1123.

[0133] For ease of explanation, Figures 6 to 8 Only the configuration of the central NWDAF 1000, the local NWDAF 1 1110, and the local NWDAF 2 1120 is described, but Figure 2 Data sources NF 1210, 1220, 1230 and Figure 3 The composition of each of the UE 100, (R)AN200, UPF 300, DN 400, AMF 500, SMF 600, PCF 700, AF 800, UDM 900 and local NWDAF 3 1130 may be the same as or similar to that of the central NWDAF 1000 or the local NWDAF 1 1110 or the local NWDAF 2 1120.

[0134] Although the detailed description of the present disclosure has described specific embodiments, it is apparent that various modifications are possible without departing from the scope of the present disclosure. Therefore, the scope of the present disclosure should not be limited to the described embodiments, but should be determined not only by the scope of the claims described below, but also by the equivalents of the claims.

[0135] Industrial Applicability

[0136] The present disclosure can be used in the electronics industry and the information and communication industry.

Claims

1. A method of operating a central network data analysis function (NWDAF), include: receiving an analysis request message from a network function; Based on the analysis request message, determining a local NWDAF included in a federated learning (FL) group; Sending a FL local behavior instruction to the local NWDAF; In response to the FL local behavior instruction, receiving a response message from the local NWDAF; Based on the FL local behavior instruction, receiving from the local NWDAF a local model generated by means of the local NWDAF; generating a global model based on the local models; and The global model is sent to the NWDAF.

2. The operating method according to claim 1, in, The FL local behavior instruction includes delay situation occurrence event notification request information, parameters required to generate the global model, and epoch time requirement information.

3. The operating method according to claim 2, in, The epoch time requirement information includes an expected epoch time, an expected local machine learning (ML) model training and transmission time, and a local ML model training time requirement.

4. The operating method according to claim 3, in, The response message includes an indicator indicating whether the expected epoch time or the expected local ML model training and transmission time or the local ML model training time requirement can be met.

5. The operating method according to claim 4, in, In the event that the expected epoch time or the expected local ML model training and transmission time or the local ML model training time requirement cannot be met, the response message includes a new expected epoch time, a new expected local ML model training and transmission time, or a new local ML model training time requirement. 6 . The operating method according to claim 5 , further comprising, in response to the response message, sending epoch time requirement information updated based on the response message to the local NWDAF. 7 . The operating method according to claim 1 , further comprising, in response to the received local model, sending a confirmation message to the local NWDAF. 8 . The operating method according to claim 7 , further comprising, in response to the confirmation message, receiving a delay situation occurrence event report message.

9. The operating method according to claim 8, in, The global model is generated based on the delay situation occurrence event report message and the local model. 10 . The operating method according to claim 8 , further comprising receiving a local model sent later from the local NWDAF based on the delay situation occurrence event report message.

11. A central NWDAF, include: Network interface; as well as A processor is connected to the network interface, wherein the processor receives an analysis request message from a network function, determines a local NWDAF included in a federated learning (FL) group based on the analysis request message, sends a FL local behavior instruction to the local NWDAF, receives a response message from the local NWDAF in response to the FL local behavior instruction, receives a local model generated by the local NWDAF from the local NWDAF based on the FL local behavior instruction, generates a global model based on the local model, and sends the global model to the NWDAF.

12. The central NWDAF according to claim 11, in, The FL local behavior instruction includes delay situation occurrence event notification request information, parameters required to generate the global model, and epoch time requirement information.

13. The central NWDAF according to claim 12, in, The epoch time requirement information includes an expected epoch time, an expected local machine learning (ML) model training and transmission time, and a local ML model training time requirement.

14. The central NWDAF according to claim 13, in, The response message includes an indicator indicating whether the expected epoch time or the expected local ML model training and transmission time or the local ML model training time requirement can be met.

15. The central NWDAF according to claim 14, in, In the event that the expected epoch time or the expected local ML model training and transmission time or the local ML model training time requirement cannot be met, the response message includes a new expected epoch time, a new expected local ML model training and transmission time, or a new local ML model training time requirement.

Images