Method and apparatus for performing artificial intelligence and machine learning model-based channel state information prediction performance monitoring

Abstract

The present embodiments provide a method by which a user equipment performs AI / ML model-based channel state information prediction performance monitoring, the method comprising the steps of: receiving, from a base station, configuration information for prediction performance monitoring; determining, on the basis of the configuration information, a performance metric indicating an error between reference channel state information and prediction channel state information for a plurality of time resources derived through an AI / ML model; and if the performance metric satisfies a preset reporting condition, transmitting, to the base station, a prediction performance monitoring result including the performance metric for a time resource indicated by the configuration information.

Description

Method and apparatus for performing monitoring of channel state information prediction performance based on artificial intelligence and machine learning models

[0001] The embodiments propose a method and apparatus for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in a next-generation wireless access network (in this disclosure, "5G", "NR [New Radio]", "5G-Advanced", "6G" or subsequent 3GPP wireless access networks).

[0002] As next-generation wireless communication technology evolves beyond 5G to 6G, it aims to achieve faster data transmission speeds and ultra-low latency compared to 5G in ultra-high frequency bands such as the terahertz (THz) band. Accordingly, technology is advancing in the direction of incorporating artificial intelligence (AI) and machine learning (ML) technologies from the design stage of communication systems. As a result, wireless communication systems are establishing a technical foundation to support new services and applications in an ultra-high performance, ultra-low latency, and ultra-connected environment.

[0003] In particular, AI / ML technologies are being introduced in wireless communication networks to optimize network operations and ensure real-time quality. By performing various roles such as situational awareness through big data analysis, adaptive utilization of network resources and data, and intelligent, data-driven system optimization, AI / ML can enable efficient resource management and service quality improvement in complex wireless environments, over which existing methods had limitations.

[0004] As part of this aspect, a specific design is required to enable wireless communication using AI / ML models.

[0005] Embodiments of the present disclosure may provide a method and apparatus for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in a next-generation wireless access network.

[0006] In one aspect, the present embodiments may provide a method for a terminal (user equipment; UE) to perform channel state information (CSI) prediction performance monitoring based on an artificial intelligence and machine learning (AI / ML) model, comprising the steps of: receiving configuration information for prediction performance monitoring from a base station; determining, based on the configuration information, a performance metric representing an error between a prediction channel state information for a plurality of time resources derived through an AI / ML model and a reference channel state information; and, when the performance metric satisfies a preset reporting condition, transmitting a prediction performance monitoring result including a performance metric for a time resource indicated by the configuration information to the base station.

[0007] In another aspect, the present embodiments provide a method for a base station to control channel state information (CSI) prediction performance monitoring based on an artificial intelligence and machine learning (AI / ML) model, comprising the steps of transmitting configuration information for prediction performance monitoring to a terminal and receiving a prediction performance monitoring result from the terminal, the result including a performance metric for a time resource indicated by the configuration information, wherein the performance metric represents an error between the prediction channel state information and the reference channel state information for a plurality of time resources derived by the terminal through an AI / ML model based on the configuration information, and the prediction performance monitoring result is received from the terminal when the performance metric satisfies a preset reporting condition.

[0008] In another aspect, the embodiments may provide a terminal (user equipment; UE) that performs channel state information (CSI) prediction performance monitoring based on an artificial intelligence and machine learning (AI / ML) model, comprising a transmitter, a receiver, and a control unit that controls the operation of the transmitter and the receiver, wherein the control unit receives configuration information for prediction performance monitoring from a base station, determines a performance metric representing an error between the prediction channel state information for a plurality of time resources derived through an AI / ML model and the reference channel state information based on the configuration information, and when the performance metric satisfies a preset reporting condition, transmits a prediction performance monitoring result including a performance metric for a time resource indicated by the configuration information to the base station.

[0009] In another aspect, the embodiments provide a base station for controlling channel state information (CSI) prediction performance monitoring based on artificial intelligence and machine learning (AI / ML) models, comprising a transmitter, a receiver, and a control unit for controlling the operation of the transmitter and the receiver, wherein the control unit transmits configuration information for prediction performance monitoring to a terminal and receives a prediction performance monitoring result from the terminal, the result including a performance metric for a time resource indicated by the configuration information, wherein the performance metric represents an error between the prediction channel state information for a plurality of time resources derived through an AI / ML model based on the configuration information by the terminal and the reference channel state information, and the prediction performance monitoring result can be provided when the performance metric satisfies a preset reporting condition.

[0010] According to the embodiments, a method and apparatus for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in a next-generation wireless access network can be provided.

[0011] In addition, according to these embodiments, the reliability of an AI / ML-based channel prediction model can be verified in real time, while unnecessary uplink feedback signaling overhead can be reduced.

[0012] FIG. 1 is a diagram briefly illustrating the structure of an NR wireless communication system to which the present embodiment can be applied.

[0013] FIG. 2 is a drawing illustrating the frame structure in an NR system to which the present embodiment can be applied.

[0014] FIG. 3 is a diagram illustrating a resource grid supported by wireless access technology to which the present embodiment can be applied.

[0015] FIG. 4 is a diagram illustrating the bandwidth part supported by the wireless access technology to which the present embodiment can be applied.

[0016] FIG. 5 is a diagram illustrating an exemplary synchronization signal block in a wireless access technology to which the present embodiment can be applied.

[0017] FIG. 6 is a diagram illustrating a random access procedure in a wireless access technology to which the present embodiment can be applied.

[0018] Figure 7 is a diagram for explaining CORESET.

[0019] FIG. 8 is a diagram illustrating channel state information being outdated to which the present embodiment can be applied.

[0020] FIG. 9 is a diagram illustrating the structure of a time resource area for predicting channel state information according to one embodiment.

[0021] FIG. 10 is a diagram illustrating a procedure in which a terminal according to one embodiment performs monitoring of channel state information prediction performance using artificial intelligence and machine learning.

[0022] FIG. 11 is a diagram illustrating a procedure in which a base station according to one embodiment controls channel state information prediction performance monitoring using artificial intelligence and machine learning.

[0023] FIG. 12 is a diagram illustrating performance indicators calculated for each slot and subband according to one embodiment.

[0024] FIG. 13 is a diagram illustrating a threshold-based reporting event based on performance indicators calculated for each slot and subband according to one embodiment.

[0025] FIG. 14 is a diagram illustrating a threshold-based reporting event based on performance indicators calculated for each slot and subband according to another embodiment.

[0026] FIG. 15 is a diagram showing the configuration of a terminal according to another embodiment.

[0027] FIG. 16 is a diagram showing the configuration of a base station according to another embodiment.

[0028] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to the exemplary drawings. In assigning reference numerals to the components of each drawing, the same components may have the same reference numeral as much as possible, even if they are shown in different drawings. Furthermore, in describing the embodiments, if it is determined that a detailed description of related known components or functions may obscure the essence of the technical concept, such detailed description may be omitted. Where terms such as "comprising," "having," or "consisting of" are used in this specification, other parts may be added unless "only" is used. Where a component is expressed in the singular, it may include a plural unless otherwise specified.

[0029] Additionally, terms such as first, second, A, B, (a), (b), etc., may be used to describe the components of the present disclosure. These terms are used merely to distinguish the components from other components, and the nature, order, sequence, or number of the components are not limited by such terms.

[0030] In describing the positional relationship of components, where it is stated that two or more components are "connected," "combined," or "joined," it should be understood that while the two or more components may be directly "connected," "combined," or "joined," they may also be "connected," "combined," or "joined" with other components "intervened." Here, the other components may be included in one or more of the two or more components that are "connected," "combined," or "joined" with one another.

[0031] In describing the temporal flow relationship regarding components, methods of operation, or methods of production, for example, when the temporal or sequential relationship is described using "after," "following," "next," or "before," it may include cases where the relationship is not continuous unless "immediately" or "directly" is used.

[0032] Meanwhile, where numerical values ​​or corresponding information regarding a component (e.g., levels, etc.) are mentioned, even without separate explicit notation, the numerical values ​​or corresponding information may be interpreted as including a range of error that may occur due to various factors (e.g., process factors, internal or external shocks, noise, etc.).

[0033] A wireless communication system in this specification refers to a system for providing various communication services, such as voice and data packets, using wireless resources, and may include a terminal, a base station, or a core network.

[0034] The embodiments disclosed below may be applied to wireless communication systems using various wireless access technologies. For example, the embodiments may be applied to various wireless access technologies such as CDMA (code division multiple access), FDMA (frequency division multiple access), TDMA (time division multiple access), OFDMA (orthogonal frequency division multiple access), SC-FDMA (single carrier frequency division multiple access), or NOMA (non-orthogonal multiple access). Furthermore, wireless access technology may refer not only to specific access technologies but also to communication technologies for each generation established by various telecommunication organizations such as 3GPP, 3GPP2, WiFi, Bluetooth, IEEE, and ITU. For example, CDMA may be implemented as a wireless technology such as UTRA (universal terrestrial radio access) or CDMA2000. TDMA may be implemented as a wireless technology such as GSM (global system for mobile communications), GPRS (general packet radio service), or EDGE (enhanced datarates for GSM evolution). OFDMA can be implemented using wireless technologies such as IEEE (Institute of Electrical and Electronic Engineers) 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802-20, and E-UTRA (evolved UTRA). IEEE 802.16m is an evolution of IEEE 802.16e and provides backward compatibility with systems based on IEEE 802.16e.UTRA is part of UMTS (universal mobile telecommunications system). 3GPP (3rd generation partnership project) LTE (long term evolution) is part of E-UMTS (evolved UMTS) that uses E-UTRA (evolved-UMTS terrestrial radio access), employing OFDMA in the downlink and SC-FDMA in the uplink. As such, these embodiments can be applied to currently disclosed or commercialized radio access technologies, and can also be applied to radio access technologies currently under development or to be developed in the future.

[0035] Meanwhile, the term "terminal" in this specification is a comprehensive concept meaning a device including a wireless communication module that communicates with a base station in a wireless communication system. It should be interpreted as a concept that includes not only User Equipment (UE) in WCDMA, LTE, NR, HSPA, and IMT-2020 (5G or New Radio), but also Mobile Station (MS), User Terminal (UT), Subscriber Station (SS), and wireless device in GSM. Furthermore, depending on the usage type, the terminal may be a user portable device such as a smartphone, or in a V2X communication system, it may refer to a vehicle or a device including a wireless communication module inside a vehicle. Additionally, in the case of a Machine Type Communication (MMC) system, it may refer to an MTC terminal, M2M terminal, URLLC terminal, etc., equipped with a communication module to perform machine type communication.

[0036] In this specification, "base station" or "cell" refers to an end that communicates with a terminal in terms of a network, and encompasses various coverage areas such as Node-B, eNB (evolved Node-B), gNB (gNode-B), LPN (Low Power Node), Sector, Site, various types of antennas, BTS (Base Transceiver System), Access Point, Point (e.g., Transmitter Point, Receiver Point, Transceiver Point), Relay Node, Mega Cell, Macro Cell, Micro Cell, Pico Cell, Femto Cell, RRH (Remote Radio Head), RU (Radio Unit), and Small Cell. Additionally, "cell" may include a Bandwidth Part (BWP) in the frequency domain. For example, a serving cell may refer to the Activation BWP of a terminal.

[0037] Since there is a base station controlling one or more of the various cells listed above, the term "base station" can be interpreted in two senses. 1) It may refer to the device itself that provides a mega cell, macro cell, micro cell, pico cell, femto cell, or small cell in relation to a wireless area, or 2) it may refer to the wireless area itself. In 1), all devices that provide a specific wireless area are controlled by the same entity or interact to configure the wireless area collaboratively are referred to as base stations. Depending on the configuration method of the wireless area, a point, a transmitting / receiving point, a transmitting point, a receiving point, etc., are examples of a base station. In 2), the wireless area itself that receives or transmits a signal from the perspective of a user terminal or from the perspective of a neighboring base station may also be referred to as a base station.

[0038] In this specification, "Cell" may refer to a component carrier having coverage of a signal transmitted from a transmitting / receiving point or coverage of a signal transmitted from a transmitting / receiving point (transmission point or transmission / reception point), or the transmitting / receiving point itself.

[0039] Uplink (UL, or Uplink) refers to the method of transmitting and receiving data from a terminal to a base station, and Downlink (DL, or Downlink) refers to the method of transmitting and receiving data from a base station to a terminal. Downlink may refer to communication or a communication path from multiple transmission and reception points to a terminal, and uplink may refer to communication or a communication path from a terminal to multiple transmission and reception points. In this case, in the downlink, the transmitter may be part of the multiple transmission and reception points, and the receiver may be part of the terminal. Additionally, in the uplink, the transmitter may be part of the terminal, and the receiver may be part of the multiple transmission and reception points.

[0040] The uplink and downlink transmit and receive control information through control channels such as PDCCH (Physical Downlink Control Channel) and PUCCH (Physical Uplink Control Channel), and transmit and receive data by configuring data channels such as PDSCH (Physical Downlink Shared Channel) and PUSCH (Physical Uplink Shared Channel). In the following description, the situation in which signals are transmitted and received through channels such as PUCCH, PUSCH, PDCCH, and PDSCH is also referred to as "transmitting and receiving PUCCH, PUSCH, PDCCH, and PDSCH."

[0041] To clarify the explanation, the technical concept described below is primarily based on 3GPP LTE / LTE-A / NR (New RAT) communication systems, but the technical features are not limited to said communication systems.

[0042] Following research on 4G (4th-Generation) communication technology, 3GPP develops 5G (5th-Generation) communication technology to meet the requirements of the ITU-R for next-generation radio access technology. Specifically, 3GPP develops LTE-A pro, which enhances LTE-Advanced technology to meet ITU-R requirements, and NR, a new communication technology distinct from 4G communication technology, as 5G communication technologies. Since both LTE-A pro and NR refer to 5G communication technology, the following description of 5G communication technology will focus on NR unless a specific technology is being identified.

[0043] The operational scenarios in NR define various operation scenarios by adding considerations for satellites, automobiles, and new verticals to the existing 4G LTE scenarios, and in terms of service, they support eMBB (Enhanced Mobile Broadband) scenarios, mMTC (Massive Machine Communication) scenarios which require low data rates and asynchronous access while having high terminal density and being deployed over a wide range, and URLLC (Ultra Reliability and Low Latency) scenarios which require high responsiveness and reliability and can support high-speed mobility.

[0044] To satisfy these scenarios, NR introduces a wireless communication system equipped with new waveform and frame structure technologies, low latency technology, mmWave support technology, and forward compatibility technology. In particular, the NR system presents various technical changes in terms of flexibility to provide forward compatibility. The main technical features of NR are explained below with reference to the drawings.

[0045]

[0046] <NR 시스템 일반>

[0047] FIG. 1 is a simplified diagram illustrating the structure of an NR system to which the present embodiment can be applied.

[0048] Referring to FIG. 1, the NR system is divided into a 5G Core Network (5GC) and an NR-RAN part. The NG-RAN consists of gNBs and ng-eNBs that provide control plane (RRC) protocol endpoints for the user plane (SDAP / PDCP / RLC / MAC / PHY) and User Equipment (UE). gNBs are interconnected with each other, or gNBs and ng-eNBs are interconnected via Xn interfaces. Each gNB and ng-eNB is connected to the 5GC via an NG interface. The 5GC may be configured to include an Access and Mobility Management Function (AMF), which is responsible for control plane functions such as terminal access and mobility control, and a User Plane Function (UPF), which is responsible for control functions for user data. The NR system includes support for both frequency bands below 6 GHz (FR1, Frequency Range 1) and frequency bands above 6 GHz (FR2, Frequency Range 2).

[0049] gNB refers to a base station that provides NR user plane and control plane protocol terminations to a terminal, and ng-eNB refers to a base station that provides E-UTRA user plane and control plane protocol terminations to a terminal. The base station described in this specification should be understood as encompassing both gNB and ng-eNB, and may also be used to refer to gNB or ng-eNB separately as necessary.

[0050] <NR 웨이브 폼, 뉴머롤러지 및 프레임 구조>

[0051] In NR, CP-OFDM waveforms using a cyclic prefix are used for downlink transmission, and CP-OFDM or DFT-s-OFDM are used for uplink transmission. OFDM technology is easy to combine with MIMO (Multiple Input Multiple Output) and has the advantage of allowing the use of low-complexity receivers along with high frequency efficiency.

[0052] Meanwhile, in NR, since the requirements for data rate, latency, coverage, etc. differ for each of the three scenarios mentioned above, it is necessary to efficiently satisfy the requirements for each scenario through the frequency bands that constitute an arbitrary NR system. To this end, a technology has been proposed to efficiently multiplex wireless resources based on multiple different numerologies.

[0053] Specifically, the NR transmission numerator is determined based on sub-carrier spacing and CP (Cyclic prefix), and as shown in Table 1 below, the μ value is used as an exponential value of 2 based on 15 kHz and changes exponentially.

[0054] μ서브캐리어 간격Cyclic prefixSupported for dataSupported for synch015NormalYesYes130NormalYesYes260Normal, ExtendedYesNo3120NormalYesYes4240NormalNoYes

[0055] As shown in Table 1 above, the numerators of NR can be classified into five types based on the subcarrier spacing. This differs from LTE, one of the 4G communication technologies, where the subcarrier spacing is fixed at 15 kHz. Specifically, the subcarrier spacings used for data transmission in NR are 15, 30, 60, and 120 kHz, while the subcarrier spacings used for synchronization signal transmission are 15, 30, 12, and 240 kHz. Additionally, extended CP applies only to the 60 kHz subcarrier spacing. Meanwhile, the frame structure in NR defines a frame with a length of 10ms, composed of 10 subframes of equal length of 1ms. A single frame can be divided into 5ms half-frames, and each half-frame contains 5 subframes. In the case of a 15 kHz subcarrier interval, one subframe consists of one slot, and each slot consists of 14 OFDM symbols. FIG. 2 is a diagram illustrating the frame structure in an NR system to which the present embodiment can be applied. Referring to FIG. 2, in the case of a normal CP, the slot is fixedly composed of 14 OFDM symbols, but the length of the slot in the time domain may vary depending on the subcarrier interval. For example, in the case of a numeral with a 15 kHz subcarrier interval, the slot is composed of a length of 1 ms, which is the same length as the subframe. In contrast, in the case of a numeral with a 30 kHz subcarrier interval, the slot is composed of 14 OFDM symbols, but two slots may be included in one subframe with a length of 0.5 ms. That is, the subframe and the frame are defined with a fixed time length, while the slot is defined by the number of symbols, and the time length may vary depending on the subcarrier interval.Meanwhile, NR defines the basic unit of scheduling as a slot and introduced mini-slots (or sub-slots or non-slot based schedules) to reduce transmission delay in the wireless section. Using a wide subcarrier spacing reduces transmission delay in the wireless section because the length of a single slot becomes inversely shorter. Mini-slots (or sub-slots) are designed for efficient support of URLLC scenarios and allow scheduling in units of 2, 4, or 7 symbols.

[0056] Furthermore, unlike LTE, NR defines uplink and downlink resource allocation at the symbol level within a single slot. To reduce HARQ latency, a slot structure was defined that allows HARQ ACK / NACK to be transmitted directly within the transmission slot; this slot structure is described as a self-contained structure.

[0057] NR is designed to support a total of 256 slot formats, of which 62 are used in 3GPP Rel-15. Additionally, it supports common frame structures that form FDD or TDD frames through various slot combinations. For example, it supports slot structures where all slot symbols are set to downlink, slot structures where all symbols are set to uplink, and slot structures where downlink and uplink symbols are combined. Furthermore, NR supports data transmission being distributed and scheduled across one or more slots. Therefore, base stations can use a Slot Format Indicator (SFI) to inform a terminal whether a slot is a downlink slot, an uplink slot, or a flexible slot. Base stations can indicate the slot format by using the SFI to indicate an index of a table configured via UE-specific RRC signaling, or they can indicate it dynamically via Downlink Control Information (DCI) or statically or semi-statically via RRC.

[0058] <NR 물리 자원 >

[0059] Regarding physical resources in NR, antenna ports, resource grids, resource elements, resource blocks, and bandwidth parts are considered.

[0060] An antenna port is defined such that the channel carrying a symbol on the antenna port can be inferred from the channel carrying another symbol on the same antenna port. If the large-scale property of the channel carrying a symbol on one antenna port can be inferred from the channel carrying a symbol on another antenna port, the two antenna ports can be said to be in a QC / QCL (quasi co-located or quasi co-location) relationship. Here, the large-scale property includes one or more of delay spread, Doppler spread, frequency shift, average received power, and received timing.

[0061] FIG. 3 is a diagram illustrating a resource grid supported by wireless access technology to which the present embodiment can be applied.

[0062] Referring to FIG. 3, a resource grid may exist for each numerator because NR supports multiple numerators on the same carrier. Additionally, a resource grid may exist depending on the antenna port, subcarrier spacing, and transmission direction.

[0063] A resource block consists of 12 subcarriers and is defined only in the frequency domain. Additionally, a resource element consists of one OFDM symbol and one subcarrier. Therefore, as shown in Fig. 3, the size of a single resource block can vary depending on the subcarrier spacing. Furthermore, NR defines "Point A," which serves as a common reference point for the resource block grid, as well as common resource blocks, virtual resource blocks, etc.

[0064] FIG. 4 is a diagram illustrating the bandwidth part supported by the wireless access technology to which the present embodiment can be applied.

[0065] In NR, unlike LTE where the carrier bandwidth is fixed at 20 MHz, the maximum carrier bandwidth is set from 50 MHz to 400 MHz depending on the subcarrier interval. Therefore, it is not assumed that all terminals use this entire carrier bandwidth. Accordingly, in NR, as shown in Fig. 4, a Bandwidth Part (BWP) can be designated within the carrier bandwidth for the terminal to use. Additionally, a Bandwidth Part is associated with a single numerator and consists of a subset of a continuous common resource block, and can be dynamically activated over time. Up to four Bandwidth Parts are configured for the uplink and downlink respectively, and data is transmitted and received using the Bandwidth Part activated at a given time.

[0066] In the case of paired spectrum, the uplink and downlink bandwidth parts are set independently, whereas in the case of unpaired spectrum, the downlink and uplink bandwidth parts are paired to share a center frequency in order to prevent unnecessary frequency re-tuning between downlink and uplink operations.

[0067] <NR 초기 접속>

[0068] In NR, the terminal performs cell search and random access procedures to connect to the base station and perform communication.

[0069] Cell search is a procedure in which a terminal uses a Synchronization Signal Block (SSB) transmitted by a base station to synchronize with the corresponding base station's cell, obtain a physical layer cell ID, and acquire system information.

[0070] FIG. 5 is a diagram illustrating an exemplary synchronization signal block in a wireless access technology to which the present embodiment can be applied.

[0071] Referring to FIG. 5, the SSB consists of a primary synchronization signal (PSS) and a secondary synchronization signal (SSS) each occupying 1 symbol and 127 subcarriers, and a PBCH spanning 3 OFDM symbols and 240 subcarriers.

[0072] The terminal monitors the SSB in the time and frequency domains and receives the SSB.

[0073] SSBs can be transmitted up to 64 times within 5ms. Multiple SSBs are transmitted via different transmission beams within the 5ms timeframe, and the terminal performs detection by assuming that an SSB is transmitted every 20ms based on a specific beam used for transmission. The number of beams available for SSB transmission within the 5ms timeframe can increase as the frequency band increases. For example, up to 4 SSB beams can be transmitted at 3GHz or lower, up to 8 beams in the frequency band from 3GHz to 6GHz, and up to 64 different beams can be used to transmit SSBs in the frequency band above 6GHz.

[0074] Two SSBs are included in a single slot, and the starting symbol and number of repetitions within the slot are determined according to the subcarrier interval as follows.

[0075] Meanwhile, unlike the SS of conventional LTE, the SSB is not transmitted at the center frequency of the carrier bandwidth. That is, the SSB can be transmitted even at locations other than the center of the system band, and multiple SSBs can be transmitted across the frequency domain when broadband operation is supported. Accordingly, the terminal monitors the SSB using a synchronization raster, which is a candidate frequency location for monitoring the SSB. The carrier raster, which is information on the center frequency location of the channel for initial connection, and the synchronization raster were newly defined in NR, and the synchronization raster is set with a wider frequency interval compared to the carrier raster, thereby supporting fast SSB search by the terminal.

[0076] The terminal can obtain the MIB through the PBCH of the SSB. The Master Information Block (MIB) contains minimum information for the terminal to receive the Remaining Minimum System Information (RMSI) broadcast by the network. Additionally, the PBCH may include information regarding the location of the first DM-RS symbol in the time domain, information for the terminal to monitor SIB1 (e.g., SIB1 numeral information, information related to SIB1 CORESET, search space information, PDCCH related parameter information, etc.), and offset information between the Common Resource Block and the SSB (the absolute location of the SSB within the carrier is transmitted via SIB1). Here, the SIB1 numeral information is applied identically to some messages used in the random access procedure for the terminal to connect to the base station after completing the cell search procedure. For example, the SIB1 numeral information may be applied to at least one of messages 1 to 4 for the random access procedure.

[0077] The aforementioned RMSI may refer to SIB1 (System Information Block 1), and SIB1 is broadcast periodically (e.g., 160ms) from the cell. SIB1 contains information necessary for the terminal to perform the initial random access procedure and is transmitted periodically via PDSCH. To receive SIB1, the terminal must receive the numerology information used for transmitting SIB1 and the CORESET (Control Resource Set) information used for scheduling SIB1 via PBCH. The terminal checks the scheduling information for SIB1 using SI-RNTI within the CORESET and obtains SIB1 on the PDSCH according to the scheduling information. The remaining SIBs, excluding SIB1, may be transmitted periodically or upon the terminal's request.

[0078] FIG. 6 is a diagram illustrating a random access procedure in a wireless access technology to which the present embodiment can be applied.

[0079] Referring to FIG. 6, when cell search is completed, the terminal transmits a random access preamble for random access to the base station. The random access preamble is transmitted via PRACH. Specifically, the random access preamble is transmitted to the base station via PRACH, which consists of a series of radio resources in specific slots that are repeated periodically. Generally, when the terminal initially connects to a cell, a contention-based random access procedure is performed, and when performing random access for Beam Failure Recovery (BFR), a non-contention-based random access procedure is performed.

[0080] The terminal receives a random access response for the transmitted random access preamble. The random access response may include a random access preamble identifier (ID), an UL Grant (uplink radio resource), a temporary C-RNTI (Temporary Cell - Radio Network Temporary Identifier), and a TAC (Time Alignment Command). Since a single random access response may contain random access response information for one or more terminals, the random access preamble identifier may be included to indicate which terminal the included UL Grant, temporary C-RNTI, and TAC are valid for. The random access preamble identifier may be an identifier for the random access preamble received by the base station. The TAC may be included as information for the terminal to coordinate uplink synchronization. The random access response may be indicated by the random access identifier on the PDCCH, namely the RA-RNTI (Random Access - Radio Network Temporary Identifier).

[0081] A terminal that receives a valid random access response processes the information contained in the random access response and performs a transmission scheduled to the base station. For example, the terminal applies a TAC and stores a temporary C-RNTI. Additionally, using a UL Grant, it transmits data stored in the terminal's buffer or newly generated data to the base station. In this case, information that can identify the terminal must be included.

[0082] Finally, the terminal receives a downlink message to resolve competition.

[0083] <NR CORESET>

[0084] The downlink control channel in NR is transmitted in a CORESET (Control Resource Set) with a length of 1 to 3 symbols, and transmits uplink / downlink scheduling information, SFI (Slot format Index), TPC (Transmit Power Control) information, etc.

[0085] In this way, NR introduced the concept of CORESET to ensure system flexibility. CORESET (Control Resource Set) refers to time-frequency resources for downlink control signals. A terminal can decode control channel candidates by using one or more search spaces from the CORESET time-frequency resources. Quasi CoLocation (QCL) assumptions were established for each CORESET, and these are used to indicate characteristics regarding the analog beam direction in addition to the characteristics assumed by conventional QCL, such as delay spread, Doppler spread, Doppler shift, and mean delay.

[0086] Figure 7 is a diagram for explaining CORESET.

[0087] Referring to FIG. 7, CORESET can exist in various forms within a single slot and within the carrier bandwidth, and in the time domain, CORESET can be composed of up to 3 OFDM symbols. Additionally, CORESET is defined as a multiple of 6 resource blocks up to the carrier bandwidth in the frequency domain.

[0088] The first CORESET is specified via the MIB as part of the initial bandwidth part configuration to enable the reception of additional configuration and system information from the network. After establishing a connection with the base station, the terminal can be configured by receiving one or more CORESET information via RRC signaling.

[0089] In this specification, frequencies, frames, subframes, resources, resource blocks, regions, bands, subbands, control channels, data channels, synchronization signals, various reference signals, various signals, or various messages related to NR (New Radio) may be interpreted in the sense used in the past or present, or in various senses used in the future.

[0090] Wider bandwidth operations

[0091] In the case of existing LTE systems, scalable bandwidth operation was supported for any LTC Component Carrier (CC). That is, depending on the frequency deployment scenario, any LTE operator could configure a bandwidth ranging from a minimum of 1.4 MHz to a maximum of 20 MHz when configuring a single LTE CC, and normal LTE terminals supported a transmit / receive capability of 20 MHz bandwidth for a single LTE CC.

[0092] However, in the case of NR, the design is made to enable support for NR terminals with different transmit / receive bandwidth capabilities through a single wideband NR CC. Accordingly, it is required to configure one or more bandwidth parts (BWPs) consisting of subdivided bandwidths for any NR CC, and to support flexible wider bandwidth operation through different bandwidth part configurations and activations for each terminal.

[0093] Specifically, in NR, one or more bandwidth parts can be configured through a single serving cell configured from the perspective of a terminal, and the terminal is defined to activate one downlink bandwidth part (DL bandwidth part) and one uplink bandwidth part (UL bandwidth part) in the serving cell to use for uplink / downlink data transmission and reception. In addition, for terminals where multiple serving cells are configured, i.e., terminals to which CA is applied, it is defined to activate one downlink bandwidth part and / or uplink bandwidth part for each serving cell to use the wireless resources of the serving cell for uplink / downlink data transmission and reception.

[0094] Specifically, an initial bandwidth part for the initial access procedure of a terminal is defined in any serving cell, and one or more terminal-specific (UE-specific) bandwidth parts are configured for each terminal through dedicated RRC signaling, and a default bandwidth part for a fallback operation can also be defined for each terminal.

[0095] However, depending on the capability and bandwidth part(s) configuration of the terminal in any serving cell, it may be defined to simultaneously activate and use multiple downlink and / or uplink bandwidth parts, but in NR rel-15, it is defined to activate and use only one downlink bandwidth part (DL bandwidth part) and one uplink bandwidth part (UL bandwidth part) at any time in any terminal.

[0096] In addition, terms regarding AI / ML-applied wireless communication can be defined as follows.

[0097] Data collection refers to the process by which network nodes, managed entities, or UEs collect data for the purpose of training AI / ML models, data analysis, and inference.

[0098] An AI / ML model (hereinafter also referred to as a 'model') refers to a data-driven algorithm that applies AI / ML technology to generate a series of outputs based on a series of inputs. AI / ML model training refers to the process of training an AI / ML model in a data-driven manner by learning input / output relationships, and obtaining the trained AI / ML model for inference. AI / ML model inference refers to the process of generating a series of outputs based on a series of inputs using the trained AI / ML model. AI / ML model validation refers to a sub-process of training that evaluates the quality of an AI / ML model using a dataset different from the one used for training. AI / ML model testing refers to a sub-process of training that evaluates the performance of the final AI / ML model using a dataset different from those used for training and validation. Unlike AI / ML model validation, testing does not assume subsequent tuning of the model.

[0099] A UE-side (AI / ML) model refers to an AI / ML model where inference is performed entirely within the UE. A Network-side (AI / ML) model refers to an AI / ML model where inference is performed entirely within the network. A One-sided (AI / ML) model refers to either a UE-side (AI / ML) model or a Network-side (AI / ML) model. A Two-sided (AI / ML) model refers to a pair of AI / ML models where joint inference is performed. Here, joint inference consists of AI / ML inference performed jointly across the UE and the network. That is, the first part of the inference is performed by the UE first, and the remaining part by the gNB, or vice versa.

[0100] AI / ML model transfer refers to the transmission of an AI / ML model via a wireless interface, using parameters of a model structure known to the receiving side or a new model with such parameters. The transmission may include a full model or a partial model. Model download refers to the transmission of a model from the network to the UE. Model upload refers to the transmission of a model from the UE to the network.

[0101] Federated learning / federated training refers to a machine learning technique that trains AI / ML models on multiple distributed edge nodes (e.g., UEs, gNBs), each performing local model training using local data samples. While this requires various interactions between the models, the exchange of local data samples is not necessary. Offline field data is data collected in the field and used for the offline training of AI / ML models. Online field data is data collected in the field and used for the online training of AI / ML models.

[0102] Model monitoring refers to the procedure of monitoring the inference performance of AI / ML models.

[0103] Supervised learning refers to the process of training a model using inputs and their corresponding labels. Unsupervised learning refers to the process of training a model without labeled data. Semi-supervised learning refers to the process of training a model by mixing labeled and unlabeled data. Reinforcement Learning (RL) refers to the process of training an AI / ML model from inputs (i.e., states) and feedback signals (i.e., rewards) resulting from the model's outputs (i.e., actions) in an environment where the model interacts.

[0104] Model activation refers to activating an AI / ML model for a specific function. Model deactivation refers to deactivating an AI / ML model for a specific function. Model switching refers to deactivating the currently activated AI / ML model and activating a different AI / ML model for a specific function.

[0105] When applying AI / ML models, the following levels of network-UE collaboration are considered.

[0106] 1. Level x: No collaboration.

[0107] 2. Level y: Signaling-based collaboration without model migration.

[0108] 3. Level z: Signal-based collaboration via model transmission.

[0109] In relation to life cycle management (LCM) procedures for AI / ML models, the AI / ML model may have a model ID containing relevant information and / or model functions regarding at least some AI / ML operations.

[0110] With regard to model selection, activation, deactivation, switching, and replacement for the UE-side model and both-side models, if determined by the network, the network may initiate the process, or the terminal may initiate and request the network. If determined by the UE, the UE's decision may be reported to the network in accordance with an event configured by the network.

[0111] For AI / ML-based feature groups (FGs), an additional condition refers to all aspects assumed for model training, but it is not part of the UE capability for the AI / ML-based feature group. This does not mean that additional conditions are necessarily explicitly specified. Additional conditions can be divided into two categories: network-side additional conditions and UE-side additional conditions.

[0112] For the inference of the UE-side model, to ensure consistency between training and inference with respect to additional NW-side conditions (if identified), the following options can be taken as possible approaches:

[0113] - Model identification to achieve alignment of additional NW-side conditions between the NW side and the UE side

[0114] - Train the model under additional conditions in the network and deliver it to the UE

[0115] - Provide the UE with information and / or instructions regarding additional conditions on the NW side

[0116] - Consistency is supported by monitoring (model / function selection by UE and / or NW based on the performance of candidate models / functions on the UE side).

[0117] - Other approaches are not excluded

[0118] - The possibility that different approaches can achieve the same function is not denied.

[0119] Regarding data collection, it can be defined as follows.

[0120] For the UE-side AI / ML model on the UE side, the UE reports to the NW a supported / preferred configuration for downlink reference signal (DL RS) transmission. For the data collection trigger / start, data collection may be started / triggered by the NW's configuration or by a request from the UE for data collection.

[0121] Signals / configurations / measurements / reporting for data collection, for example, signaling aspects relate to assistance information (if supported), reference signals, the content / type of collected data, configurations related to Set A and / or Set B, and information regarding the association / mapping between Set A and Set B.

[0122] Support information provided by the network to the UE for the collection of UE data to classify data for the purpose of differentiating data characteristics (where supported). Support information must protect privacy / proprietary information.

[0123] For NW-side AI / ML models, mechanisms related to reporting, additional information regarding report content, reduction of reporting overhead, and signals / configuration / measurement / reporting for data collection, for example, signal aspects are related to supporting information (if supported) and reference signals.

[0124] Regarding data collection for the NW-side AI / ML models of BM-Case1 and BM-Case2, the following approach for overhead reduction is identified:

[0125] - Omission / Selection of collected data

[0126] - Compression of collected data

[0127] - If the purpose of data collection differs, the overhead reduction mechanism and the resulting impact on specifications may differ.

[0128] - For each LCM objective, the support of any mechanisms (if necessary) and potential specification impacts (if any) are subject to separate discussion.

[0129] With regard to data collection for the NW-side AI / ML models of BM-Case1 and BM-Case2, the following reporting signals for beam-specific aspects may apply:

[0130] - L1 signal for reporting collected data

[0131] - Higher-level signals for reporting collected data

[0132] - Does not apply to AI / ML model inference, at least.

[0133] - Existing signaling principles (e.g., L1 RSRP reporting) can be reused.

[0134] RAN1 studies model identification type A, including more details related to use cases.

[0135] RAN1 studies the following options for model identification type B as a starting point, including more details related to all use cases.

[0136] - MI-Option 1: Identify models with data collection-related configurations and / or instructions

[0137] - MI-Option 2: Model identification with dataset transmission

[0138] - MI-Option 3: Model identification in model transfer from NW to UE

[0139] - The names (MI-Option 1, MI-Option 2, MI-Option 3) are used for discussion purposes only.

[0140] - The following other options are suggested for model identification type B:

[0141] - MI-Option 4: Model identification through standardization of reference models (for CSI compression)

[0142] - MI-Option 5: Model identification through model monitoring

[0143] With respect to MI-Option 1 of Model Identification Type B (Model identification with configurations and / or instructions related to data collection), RAN1 further investigates the following aspects:

[0144] - Relationship between Model ID and data collection-related configurations and / or instructions

[0145] - Information transmitted from NW to UE (if any)

[0146] - Information transmitted from UE to NW (if any)

[0147] - Related procedures

[0148] - Usable / Applicable Use Cases of MI-Option 1

[0149] For Model Identification Type B of MI-Option 1 (including data collection configurations and / or instructions related to model identification), RAN1 further investigates the following aspects:

[0150] - Relationship between Model ID and data collection-related configurations and / or instructions

[0151] - Information transmitted from the network (NW) to the UE (if any)

[0152] - Information transmitted from the UE to the network (NW) (if any)

[0153] - Use of MI-Option 1 or applicable use cases

[0154] From the perspective of RAN1, for a UE-side model developed (e.g., trained, updated) on the UE side, the following procedure is an example (AI-Example 1) for further research (including feasibility / necessity) of MI-Option 1.

[0155] - A: For data collection, the NW transmits the data collection-related configuration and its / their associated IDs as a signal.

[0156] Association ID for each sub-use case related to NW-side additional conditions

[0157] - B: The UE collects data corresponding to the associated ID.

[0158] - C: AI / ML models are developed (e.g., trained, updated) on the UE side based on collected data corresponding to the associated ID.

[0159] - D: The UE reports its AI / ML model information corresponding to the associated ID to the NW. A model ID is determined / assigned for each AI / ML model.

[0160] Relationship between Model ID and Associate ID

[0161] Regarding how the model ID is determined / assigned, for example, whether the NW assigns the model ID, the UE assigns / reports the model ID, or an associated ID is considered the model ID, the statement in D “a model ID is determined / assigned for each AI / ML model” is not necessary, and the model ID is determined according to predefined rules.

[0162] D is intended to facilitate AI / ML model inference.

[0163] Additional interaction of steps A / B / C and association IDs between the UE and NW can be considered as an alternative solution for consistency resolution without model identification.

[0164] Regarding association IDs, the UE assumes that NW-side additional conditions with the same association ID are consistent at least within the cell. Whether and how the UE's assumption is applicable to multiple cells (including feasibility studies) is further investigated.

[0165] To ensure consistency of additional NW-side conditions throughout the training and inference of the UE-side model for BM-Case 1 and BM-Case 2, an association ID-based or performance monitoring-based method may be defined.

[0166] The UE may assume that DL Tx beams or beam sets / lists associated with the same Associated ID have similar properties. With respect to the Associated ID, the UE assumes that Network (NW)-side addition conditions with the same Associated ID are consistent within at least one cell.

[0167]

[0168] The present disclosure relates to a technology for measuring and reporting performance indicators of an AI model (hereinafter referred to as the CSI Prediction AI model) used by a terminal to predict Channel State Information (CSI). Traditionally, the terminal has estimated the CSI based on a pilot signal received and reported it to the Network. However, as shown in FIG. 8, the channel changes between the time when the terminal estimates the channel and provides CSI feedback, and the time when the Network utilizes the CSI feedback information for scheduling and precoding, and then transmits actual data, which may result in an outdated issue regarding the CSI feedback information.

[0169] To address this, standardization is underway to utilize AI / ML to predict future channel information where the terminal will receive actual data transmission and to provide feedback.

[0170] CSI feedback enhancement may include the following. CSI prediction (UE-sided model) using a terminal-side model may enable Functionality-based Lifecycle Management (LCM) utilizing other use cases, where necessary and applicable.

[0171] It was agreed to support the standardized 'typeII-Doppler-r18' codebook as follows. For CSI prediction using terminal-side models, the Rel-18 CSI framework is reused at least during the inference stage. Periodic, semi-persistent, and aperioditic CSI-RS are supported as CSI-RS resource types for Channel Measurement Resources (CMR). For inference reports, typeII-Doppler-r18 is supported when N4 >= 1.

[0172] Figure 9 and Table 2 illustrate the parameters associated with the type II-Doppler-r18 codebook.

[0173]

[0174] The point is that when the terminal provides PMI feedback via the type II-Doppler-r18 codebook, it can transmit channel information for slot N4 in the time domain to the network. For example, if the terminal provides CSI feedback at the n-th slot, the corresponding CSI is transmitted to the network using a CSI Prediction AI model from the l-th slot to the l+N4-1-th slot in the time domain.

[0175] The network transmits data to the terminal by scheduling and precoding based on the CSI feedback information transmitted by the terminal. Therefore, for the downlink performance of the terminal, the channel prediction performance of the CSI prediction AI model must be continuously monitored by the network.

[0176] The following points were agreed upon for performance monitoring.

[0177] In CSI prediction using a terminal-side model, if performance monitoring type 1 or type 3 is supported, the following is supported for the calculation of monitoring metrics. Calculation based on intermediate KPIs is based on at least one of Squared Generalized Cosine Similarity (SGCS) and Normalized Mean Square Error (NMSE), and the selection of specific metrics may be determined later. The definitions of SGCS and NMSE and the specific method for calculating monitoring metrics may be determined through further discussion.

[0178] In a CSI prediction use case using a terminal-side model, the following aspects were proposed regarding performance monitoring for functionality-based lifecycle management (LCM). Regarding performance monitoring types, Type 1 involves the terminal generating a performance metric. The terminal reports a performance monitoring output that assists in the network's functionality fallback decision. The network can set a threshold criterion to assist in the terminal-side performance monitoring. The network makes a decision regarding the functionality fallback behavior (a fallback mechanism to the existing CSI reporting method).

[0179] Type 2 involves the terminal reporting the predicted CSI and / or the corresponding ground truth. The network directly calculates the performance metrics. The network makes a decision regarding the function fallback behavior.

[0180] Type 3 involves the terminal calculating performance metrics. The terminal reports the calculated performance metrics to the network. The network makes decisions regarding function fallback behavior. Regarding configuration and procedures, selection / activation / deactivation / switching procedures defined for other terminal-side use cases may be reused where applicable. Configuration information and procedures for performance monitoring are included. CSI-RS configuration for performance monitoring is included. Performance metrics include at least intermediate KPIs (e.g., NMSE or SGCS). Terminal reporting includes periodic, semi-persistent, aperioditic, and event-driven reporting.

[0181] Down-selection of some of the options is not excluded. The terminal can independently determine model selection, activation, deactivation, and switching behaviors within the same function in a manner that the network does not recognize.

[0182] To measure the CSI prediction performance metric, the terminal estimates the channel for the N4 slots in the time domain using a CSI prediction AI model. The channel vector estimated through the AI ​​model for the n4th (≤N4)th slot is Let's call it the ground-truth channel vector Let's assume there are two measurement methods.

[0183] 1) SGCS-based:

[0184] In the above equation, ||x|| is the Euclidean norm (aka, L2 norma) value of vector x, and x His the conjugate transpose (aka Hermitian) value of vector x.

[0185] 2) NMSE-based:

[0186]

[0187] In the following, a method for specifically monitoring the performance of Channel State Information (CSI) prediction based on artificial intelligence and machine learning (AI / ML) models will be explained with reference to the relevant drawings.

[0188] FIG. 10 is a diagram illustrating a procedure (1000) in which a terminal according to one embodiment performs monitoring of channel state information prediction performance using artificial intelligence and machine learning.

[0189] Referring to FIG. 10, the terminal can receive configuration information for predictive performance monitoring from a base station (S1010).

[0190] According to one example, configuration information may be a higher-layer signal (e.g., RRC message) or a physical layer signal (e.g., DCI, MAC CE) containing overall control parameters required for a terminal to monitor the performance of an AI / ML prediction model and report the results for function-based lifecycle management (LCM). For example, configuration information may comprehensively include periodic, semi-static, or non-periodic CSI-RS configuration information to be utilized as a channel measurement resource (CMR), the method of calculating performance metrics (e.g., SGCS or NMSE), and threshold and timer configuration parameters for event-based reporting.

[0191] In particular, the configuration information may include identification information indicating the reporting target among multiple time resources. Here, the multiple time resources may refer to multiple slots that are the targets for CSI prediction by an AI / ML model. The identification information may be provided, for example, in the form of a bit string, and the base station may use the identification information to explicitly instruct the terminal on a specific slot or a specific resource group (e.g., totalslots, sub-slots, etc.) among the total time resources to calculate and report performance indicators. Through this, the terminal can prevent unnecessary computation on the total resources and efficiently limit the monitoring targets.

[0192] Referring again to FIG. 10, the terminal can determine (S1020) a performance metric representing the error between the predicted channel state information and the reference channel state information for multiple time resources derived through an AI / ML model based on configuration information.

[0193] For example, the predicted channel state information may refer to the CSI predicted by the terminal-side AI / ML model, and the reference channel state information may refer to the actual ground-truth channel measured by the terminal. The performance metric may be calculated as at least one of Squared Generalized Cosine Similarity (SGCS) or Normalized Mean Square Error (NMSE), which represent the accuracy of the prediction.

[0194] The terminal can divide multiple time resources into at least one resource group and determine performance indicators only for the resource group indicated by identification information. For example, if the total time resources to be predicted consist of multiple slots, the terminal can divide them into multiple sub-slot groups having a specific length. If the identification information included in the configuration information received from the base station is in the form of a bitmap (e.g., '1010') indicating a specific sub-slot group, the terminal can optimize the computational load (complexity) of the terminal by performing SGCS or NMSE operations only on the first and third resource groups where the bit is activated as '1'.

[0195] Referring again to FIG. 10, if the performance indicator satisfies a preset reporting condition, the terminal can transmit a predicted performance monitoring result including a performance indicator for a time resource indicated by the configuration information to the base station (S1030).

[0196] According to one example, the terminal may report the results of predictive performance monitoring to the base station based on events. The reporting conditions may include events in which the ratio or number of times a performance indicator falls below a first threshold value exceeds a second threshold value. As a specific example, the terminal may count the number of defective slots in which the calculated SGCS value for each slot falls below '0.5 (first threshold value)', and determine that a performance degradation event has occurred when the ratio of these defective slots to the total number of target slots exceeds '30% (second threshold value)'.

[0197] In addition, according to one example, the terminal may monitor the cumulative number of times an event occurs within a predetermined time interval and transmit a prediction performance monitoring result when the cumulative number exceeds a threshold. For example, the terminal may trigger to transmit a monitoring result for fallback to a base station only when an event accumulates and occurs three or more times (threshold) within an observation window consisting of 100 slots or a specific timer interval.

[0198] In addition, a data compression method may be applied to prevent large uplink (UL) overhead from occurring when the terminal reports performance indicators for multiple time resources to the base station. The terminal can approximate the patterns of performance indicators determined for each time resource using a codebook pre-configured between the terminal and the base station, and transmit the identifiers and coefficients of the codewords within the codebook used for approximation as the predicted performance monitoring results. As a specific example, the terminal can construct performance indicator values ​​for multiple slots into a single vector and model it as a linear combination of various codeword patterns within the codebook previously shared with the base station. Instead of individually transmitting dozens of performance indicators, the terminal can drastically reduce the size of the feedback payload by transmitting only the specific pattern number (codeword index) selected for approximation and its combination ratio (coefficient).

[0199] Additionally, when the terminal transmits prediction performance monitoring results for multiple resource groups, a situation may arise where uplink transmission resources (e.g., PUCCH or PUSCH payload) are insufficient. In this case, the terminal may decide whether to transmit prediction performance monitoring results for at least some resource groups based on a priority determined by at least one of the range of time resources included in the resource group and the index assigned to the resource group. For example, the terminal may assign the highest priority to groups (totalslots) that encompass all slots to be predicted and transmit them preferentially, and among the sub-slot groups, it may assign higher priority to groups with lower index numbers (e.g., priority of earlier time resources) or groups corresponding to specific even / odd index rules. Within the limits of allocated resources, the terminal can prevent transmission errors and reliably provide essential monitoring information to the base station by transmitting the results of high-priority groups first and dropping the results of lower-priority groups.

[0200] According to this, a method and device for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in next-generation wireless access networks can be provided. In addition, the reliability of AI / ML-based channel prediction models can be verified in real time, while simultaneously reducing unnecessary uplink feedback signaling overhead.

[0201] FIG. 11 is a diagram illustrating a procedure for a base station according to one embodiment to control channel state information prediction performance monitoring using artificial intelligence and machine learning. The description in FIG. 10 above may be omitted to avoid redundant descriptions, and in this case, the omitted content may be applied substantially the same to the base station as long as it does not contradict the technical concept of the invention.

[0202] Referring to FIG. 11, the base station can transmit configuration information for monitoring the predicted performance to the terminal (S1110).

[0203] According to one example, configuration information may be a higher-layer signal (e.g., RRC message) or a physical layer signal (e.g., DCI, MAC CE) containing overall parameters to control the terminal to monitor the performance of an AI / ML prediction model and report the results for function-based lifecycle management (LCM). For example, configuration information may comprehensively include periodic, semi-static, or non-periodic CSI-RS configuration information to be utilized as a channel measurement resource (CMR), the method of calculating performance metrics (e.g., SGCS or NMSE), and threshold and timer configuration parameters for event-based reporting.

[0204] In particular, the configuration information may include identification information indicating the reporting target among multiple time resources. Here, the multiple time resources may refer to multiple slots that are the targets for CSI prediction by the terminal's AI / ML model. The identification information may be provided, for example, in the form of a bit string, and the base station may use the identification information to explicitly instruct the terminal on a specific slot or a specific resource group (e.g., totalslots, sub-slots, etc.) among the total time resources to calculate and report performance indicators. Through this, the base station can prevent the terminal from performing unnecessary calculations on the total resources and control it to efficiently limit the monitoring targets.

[0205] Referring again to FIG. 11, the base station can receive a predicted performance monitoring result from the terminal, which includes a performance metric for a time resource indicated by configuration information (S1120).

[0206] Here, the performance metric may be a performance metric representing the error between the predicted channel state information and the reference channel state information for multiple time resources derived by the terminal through an AI / ML model based on configuration information. For example, the predicted channel state information may refer to the CSI predicted by the terminal-side AI / ML model, and the reference channel state information may refer to the actual ground-truth channel measured by the terminal. The performance metric may be calculated as at least one of Squared Generalized Cosine Similarity (SGCS) or Normalized Mean Square Error (NMSE), which represent the accuracy of the prediction.

[0207] At this time, the performance indicator may be determined only for the resource group indicated by the identification information after multiple time resources are divided into at least one resource group by the terminal. For example, if the total time resources to be predicted consist of multiple slots, they may be divided into multiple sub-slot groups having a specific length. If the identification information transmitted by the base station included in the configuration information is in the form of a bitmap (e.g., '1010') indicating a specific sub-slot group, the terminal optimizes the computational load (complexity) by performing SGCS or NMSE operations only on the first and third resource groups where the bit is activated as '1', and the base station can receive the results calculated for the corresponding resource groups.

[0208] According to one example, a base station may receive predicted performance monitoring results from a terminal based on events. The reporting conditions may include events in which the ratio or number of times a performance indicator calculated by the terminal falls below a first threshold value exceeds a second threshold value. As a specific example, when the number of defective slots in which the SGCS value calculated by the terminal for each slot falls below '0.5 (first threshold value)' is counted, and the ratio of these defective slots to the total number of target slots exceeds '30% (second threshold value)', it is determined that a performance degradation event has occurred and a report may be made to the base station.

[0209] In addition, according to one example, a base station may receive a prediction performance monitoring result that is triggered by a terminal when the cumulative number of events occurring within a predetermined time interval exceeds a threshold. For example, the base station may receive a monitoring result for fallback only when an event accumulates and occurs three or more times (threshold) within an observation window consisting of 100 slots or a specific timer interval at the terminal.

[0210] In addition, when a terminal reports performance indicators for multiple time resources to a base station, a data compression method may be applied to prevent large uplink (UL) overhead from occurring. When the patterns of performance indicators determined for each time resource are approximated using a codebook pre-configured between the terminal and the base station, the base station may receive the identifier and coefficient of the codeword within the codebook used for approximation from the terminal as a result of predictive performance monitoring. As a specific example, the terminal may construct performance indicator values ​​for multiple slots into a single vector and model it as a linear combination of various codeword patterns within a codebook previously shared with the base station. Instead of receiving dozens of performance indicators individually, the base station can achieve the effect of significantly reducing the size of the uplink feedback payload by receiving only the specific pattern number (codeword index) selected by the terminal's approximation and its combination ratio (coefficient).

[0211] In addition, when a terminal transmits prediction performance monitoring results for multiple resource groups, a situation may arise where the terminal's uplink transmission resources (e.g., PUCCH or PUSCH payload) are insufficient. In this case, the base station may receive prediction performance monitoring results for at least some resource groups whose transmission status has been determined by the terminal, based on a priority determined by at least one of the range of time resources included in the resource group and the index assigned to the resource group. For example, the terminal may assign the highest priority to groups (totalslots) that encompass all slots to be predicted and transmit them preferentially, and among the sub-slot groups, it may assign higher priority to groups with lower index numbers (e.g., priority of earlier time resources) or groups corresponding to specific even / odd index rules. Within the limits of the resources allocated to the terminal, the base station receives the results of high-priority groups first from the terminal, and receives the results of lower-priority groups with transmission omitted (dropped) by the terminal, thereby preventing transmission errors and reliably obtaining essential monitoring information.

[0212] According to this, a method and device for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in next-generation wireless access networks can be provided. In addition, the reliability of AI / ML-based channel prediction models can be verified in real time, while simultaneously reducing unnecessary uplink feedback signaling overhead.

[0213]

[0214] Hereinafter, with reference to the relevant drawings, each embodiment related to the method for monitoring the performance of Channel State Information (CSI) prediction based on artificial intelligence and machine learning (AI / ML) models will be described in detail.

[0215] A method for setting slots for measuring and reporting performance monitoring results of a CSI prediction AI model is proposed. In this disclosure, the performance monitoring results can be applied to a type 3 performance metric and can also be applied to a type 1 performance output.

[0216] 1. Report Slot Configuration: Set whether to report performance metrics per slot or to aggregate multiple slots for reporting by slot set.

[0217]

[0218] Slot Set Definition: The terminal is configured to report the CSI prediction performance metric for any of the N4 slots.

[0219] Total slots

[0220] The terminal is configured with a single slot set that includes all N4 slots.

[0221] The terminal calculates and reports the CSI prediction performance metric for all slots.

[0222] 'sub-slots(sub-slots)n':

[0223] The terminal is configured with a bit string of size n.

[0224] The terminal classifies N4 slots into slot sets divided into n portions.

[0225] The terminal assigns an index to distinguish each set of slots.

[0226] The terminal calculates and reports the CSI prediction performance metric for the slots belonging to the slot set corresponding to '1' in the BIT STRING.

[0227] The terminal expects that the size of the bit string is less than or equal to N4.

[0228] For example, if a bit string (0,1,0,0) of size 4 is received with N4=8, the terminal classifies the 8 slots into sets of 4 slots:

[0229]

[0230] Reports the CSI prediction performance metric measurements for the slots corresponding to slot set index 2, namely the 3rd and 4th slots, according to BIT STRING (0,1,0,0).

[0231] If parameters determining the size of the reporting slot (e.g., totalslots, sub-slots2, sub-slots4, sub-slots8, ...) are not separately set, the terminal expects to be set a bit string with a length equal to N4, which is the total number of slots for which performance monitoring was performed. In this case, the slots are not aggregated and report the performance monitoring results for the slot corresponding to 1 in the bit string.

[0232] If a bit string indicating which slot or set of slots to report is not separately configured, the terminal assumes that it has received a bit string in which all elements are 1. In this case, instead of selectively reporting some slots or sets of slots, it reports performance monitoring results for all slots or sets of slots.

[0233]

[0234] 2. Priority rule for omitting terminal CSI prediction performance metric reporting

[0235] If multiple performance monitoring requests are received from the NW within the same CSI report configuration but the payload of the uplink (UL) resource is insufficient, the terminal determines the priority as follows.

[0236] Totalslots has the highest priority.

[0237] When the number of slot sets is multiple (e.g., sub-slots2, sub-slots4, sub-slots8)

[0238] Slot sets with even (or odd) indices have higher priority.

[0239] The lower the index of the slot set (or the higher it is), the higher the priority.

[0240] For example, if N4 = 8 and sub-slots4 = (1,1,1,1), and it is configured that even indices have higher priority, or that lower indices have higher priority, the priorities are as follows:

[0241] Slot Set 2 > Slot Set 4 > Slot Set 1 > Slot Set 3

[0242] The terminal can omit reporting for low-priority slots.

[0243] The granularity in the time domain for reporting CSI prediction performance monitoring results is not defined in 3GPP. Through the report slot configuration proposed in this disclosure, the terminal can report in a granular manner for N4 slots. For example, it can be decided whether to process the performance monitoring values ​​for N4 slots as a single average or to process them by dividing them into multiple intervals. This allows for the control of the precision and overhead of the performance monitoring result reporting.

[0244] Priorities were assigned to reporting slots to prepare for situations where resources for CSI reports are insufficient. This allows the terminal to skip reporting on low-priority slots.

[0245]

[0246] According to another embodiment, a method is proposed to report CSI prediction performance monitoring results to NW for performance monitoring of a CSI Prediction AI model.

[0247] In the following, it is assumed that the terminal has received the slots and subbands targeted for CSI prediction from the network in advance. For the sake of convenience and for illustrative purposes, the target slots are assumed to be slots 1 through 4 and subbands 1 through 5, as shown in FIG. 12. It is assumed that the terminal has calculated performance metrics for each slot and subband using SGCS and NMSE. Since SGCS calculates cosine similarity, the value is closer to 1 as the channel is predicted more accurately. On the other hand, since NMSE calculates error, the value is closer to 0 as the channel is predicted more accurately. For the sake of convenience of explanation, it is assumed in this disclosure that prediction accuracy is higher the closer the value is to 1, regardless of which method, SGCS or NMSE, is used.

[0248] 1. Threshold-based

[0249] A. Single threshold based

[0250] a. The terminal receives a threshold setting.

[0251] b. The terminal calculates the ratio of performance metric values ​​that are less than (or greater than) threshold-1 and reports this as a performance monitoring output.

[0252] i. The terminal calculates the Performance metric value using the SGCS or NMSE method.

[0253] c. For example, as shown in FIG. 13, when the threshold is set to 0.5, the proportion of all 20 performance metric values ​​that are less than 0.5 is 25%, and the terminal reports this to the NW.

[0254] B. Multiple threshold-based

[0255] a. The terminal is set N thresholds.

[0256] i. The terminal expects that the N thresholds are sorted in ascending order.

[0257] b. The terminal sets intervals based on threshold intervals, calculates the proportion of total performance metric values ​​belonging to each interval, and reports this as a performance monitoring output.

[0258] c. For example, assume that the terminal is set to three thresholds (0.3, 0.5, 0.8).

[0259]

[0260] C. Where the terminal supports non-AI-based (legacy) CSI prediction capability

[0261] a. The terminal receives a threshold setting.

[0262] b. The terminal calculates the ratio of performance metric values ​​calculated based on AI that are below (or exceed) the threshold compared to performance metric values ​​calculated based on non-AI, and reports this.

[0263]

[0264] 2. Transmission of codebook-based channel prediction accuracy information

[0265] A. Procedure

[0266] a. The terminal calculates a representative value of the performance metric for each slot and vectorizes it.

[0267] i. For example, let's assume that the number of slots is N4 = 4. For each n4 ∈ {1, ..., N4} slot, the performance metric value for the entire subband is averaged and converted into a column vector.

[0268]

[0269] b. The terminal is set to integer B.

[0270] c. The terminal is a codebook of size S, C=[c1, c2, ..., c s ] Determine the coefficients to approximate x as a linear combination of my B codewords. The codebook is shared between the terminal and the base station.

[0271] i. For example, assume that S=8, B=2 and c is given as follows. Each codeword models the change in the performance metric per slot. For example, as in Equation 2, the example c1 models the performance metric not changing for N4 slots, and c2 models it gradually increasing.

[0272]

[0273]

[0274] ii. The terminal calculates codewords and coefficients that can describe x.

[0275]

[0276] b. The terminal reports B codewords and coefficients.

[0277] i. For example, the terminal reports in the following format.

[0278]

[0279] 3. Transmission of Channel Prediction Error Covariance Matrix

[0280] A. Notation:

[0281] a. and These are the channel matrix predicted by AI and the ground-truth channel matrix for the i-th slot and the j-th subband, respectively.

[0282] bL is the number of layers, N T is the number of transmitting antennas.

[0283] B. Error Matrix Transmission Method

[0284] a. and Calculate the error covariance matrix.

[0285] i. For example,

[0286] b. Report the diagonal elements of the covariance matrix.

[0287]

[0288] Through the method proposed in the present disclosure, the terminal can report performance monitoring results.

[0289] (single / multiple-) Threshold-based: The terminal can quantify and report the ratio of performance metric values ​​below a certain level.

[0290] Comparison with non-AI-based prediction methods: Base stations can intuitively report information to determine whether to fallback to legacy methods.

[0291] Codebook-based transmission method: Can concisely report performance monitoring metric trends per slot.

[0292] Error Covariance Matrix: It can report not only information on CSI prediction accuracy but also information advantageous for actual data transmission. For example, a base station can calculate the terminal's expected SINR performance through the error covariance matrix.

[0293]

[0294] According to another embodiment, it is assumed that the terminal has received in advance the slots and subbands to be subject to CSI prediction from the NW. In this embodiment, for the sake of example and convenience, it is assumed that the target slots are slots 1 through 4 and the subbands are subbands 1 through 5. It is assumed that the terminal has calculated a performance metric for each slot and subband using SGCS and NMSE. Since SGCS calculates cosine similarity, the value is closer to 1 as the channel is predicted more accurately. On the other hand, since NMSE calculates error, the value is closer to 0 as the channel is predicted more accurately. For the convenience of explanation in this disclosure, it is assumed that prediction accuracy is higher the closer the value is to 1, regardless of which method, SGCS or NMSE, is used.

[0295] 1. Event Definition

[0296] A. If the number or proportion of CSI prediction accuracy (CSI prediction performance metric) values ​​less than threshold-1 exceeds threshold-2

[0297] a. For example, in Fig. 14, the proportion of values ​​less than or equal to threshold-1 (0.5) among the total 20 CSI prediction accuracy values ​​is 25%, which is less than threshold-2 (30%); therefore, the event is not satisfied and the event does not occur.

[0298] A. When the average of the K worst CSI prediction accuracy values ​​is below the threshold

[0299] B. When the CSI prediction accuracy value for the wideband is below the threshold

[0300] C. If the average value of the AI-based CSI prediction accuracy (CSI prediction performance metric) is lower than the average value of the non-AI-based CSI prediction accuracy by a threshold amount

[0301] D. If the number or proportion of AI-based CSI prediction accuracy values ​​that are threshold-1 lower than non-AI-based values ​​exceeds threshold-2

[0302] E. If the average of the K worst AI-based CSI prediction accuracy values ​​is less than the threshold compared to the average of the K worst non-AI-based values

[0303] F. When the CSI prediction accuracy value for the AI-based wideband is below the threshold compared to the non-AI-based wideband CSI prediction accuracy value

[0304]

[0305] 2. Report immediately

[0306] A. The terminal is set the event defined above and parameters associated therewith, such as threshold(s), for example.

[0307] B. The terminal reports when a configured event occurs.

[0308]

[0309] 3. Timer-based reporting

[0310] A. The terminal receives the event defined above and parameters associated therewith, such as threshold(s).

[0311] B. The terminal receives the timer's start time and duration.

[0312] C. The terminal receives the Max-Event-Count value.

[0313] D. The terminal counts the number of event occurrences while the timer is running.

[0314] E. If the count value exceeds Max-Event-Count within the Timer operation time, the terminal reports.

[0315] a. Initialize count to 0 after reporting.

[0316] F. If the timer ends without the count value exceeding Max-Event-Count within the timer operation time, the count value is initialized to 0 when the timer is updated.

[0317]

[0318] 4. Window-based reporting

[0319] A. The terminal receives the event defined above and parameters associated therewith, such as threshold(s).

[0320] B. The terminal receives the window duration setting.

[0321] C. The terminal receives the Max-Event-Count value.

[0322] D. The terminal counts the number of times an event is satisfied from the present to the past, up to the window duration.

[0323] E. If the Count value exceeds Max-Event-Count, the terminal reports.

[0324] a. Initialize count to 0 after reporting.

[0325] By determining whether the event presented in the present invention has occurred, the terminal can determine when to report the performance monitoring results.

[0326] The aforementioned embodiments may be applied independently of each other, or two or more embodiments may be combined and implemented.

[0327]

[0328] According to one example, the terminal may quantize the calculated performance indicator (e.g., SGCS) into a 4-bit field to report it to the base station. Specifically, the terminal may map the entire range of the SGCS (0, 0.3) below a specific threshold value to a single first code point (e.g., '0000'). Additionally, the terminal may map the remaining range of values ​​(0.3, 1.0) to the remaining 15 code points by quantizing them evenly with a linear scale step size of 0.05. Through this, the terminal can reduce payload overhead by compressing the range where the AI / ML model’s prediction performance is significantly poor and fallback is obvious (0.3 or less) into a single code point, and report the prediction accuracy to the base station with a fine resolution of 0.05 for the key observation range (greater than 0.3) where performance is good.

[0329] Additionally, when a terminal transmits predictive performance monitoring results (e.g., CSI-PAI, CSI Predictive Accuracy Indicator) containing performance indicators via Uplink Control Information (UCI), the payload size of the UCI may be fixed to reduce the decoding complexity of the base station. Specifically, the payload size may be determined to have a fixed size (e.g., 2 × max_rank × 4 bits) based on the maximum rank (max_rank) determined by the Rank Indicator (RI) limit setting, regardless of the rank (v) value, which is the actual prediction result of the terminal. In this case, if the rank (v), which is the actual prediction result, is smaller than the maximum rank, the terminal may transmit to the base station by padding the remaining performance indicator fields corresponding to the rank difference (max_rank - v) with a value of '1' in the least significant bit (LSB) of the Uplink Control Information sequence.

[0330] In addition, if the terminal is configured to calculate wideband performance indicators for multiple frequency subbands, the terminal may not use a method of simply deriving an average precoder for the wideband first to calculate the performance indicators. Instead, the terminal may first calculate individual performance indicator values ​​(e.g., SGCS per subband) for each of the configured multiple subbands, and then determine the final wideband performance indicators by averaging the calculated performance indicator values ​​per subband across all subbands. This is intended to accurately reflect the overall prediction accuracy of the AI / ML model without distortion, even if channel variability varies by frequency band in a frequency-selective fading environment.

[0331] For example, a terminal can determine the applicability of a corresponding AI / ML model to the current wireless environment based on configuration information (e.g., CSI-ReportConfig) for monitoring prediction performance received from a base station. The terminal can report the initial applicability to the base station via an RRCReconfigurationComplete message, and if the applicability status changes due to the terminal's mobility or changes in the channel environment, it can report the updated status to the base station via a UE Assistance Information (UAI) message. If the terminal reports an 'Inapplicable' status for a specific configuration, the terminal may include a specific cause value for the inapplicability (e.g., model unavailability) in the message. This allows the base station to clearly identify the cause of the terminal's prediction failure and take appropriate follow-up actions.

[0332] Furthermore, even if the terminal reports an 'Inapplicable' status to the base station for a specific periodic CSI-ReportConfig, the terminal must maintain the configuration without autonomously releasing it. That is, until the terminal receives an explicit release instruction for the configuration from the base station, it continues to perform AI / ML model-based inference computations and report performance monitoring results in accordance with the existing settings. How the reported prediction data is processed and when the configuration is released are entirely up to the base station's network implementation.

[0333] In addition, if the base station determines that poor performance has occurred because the prediction performance monitoring results (e.g., SGCS, etc.) received from the terminal consistently fall below a threshold, or if it receives a report of 'inapplicability' from the terminal, it may release the AI / ML model-based CSI reporting configuration. At the same time or immediately thereafter, in order to ensure the stability of communication performance, the base station may newly provide (fallback) to the terminal a reference signal measurement and CSI reporting configuration of an existing method that does not use AI / ML (non-AI / ML configuration).

[0334] For example, if the AI / ML model is a model for Channel State Information Prediction (CSI Prediction) rather than Beam Management, the configuration information for controlling the terminal may be provided only in the form of a full inference configuration (e.g., Option A) that fully includes all parameter sets required for inference. That is, the configuration information for CSI prediction may be transmitted to the terminal in a form that is not separated into a separate auxiliary configuration (e.g., Option B method via OtherConfig) but is included entirely within a single upper-level parameter (e.g., CSI-ReportConfig).

[0335]

[0336] Hereinafter, the configuration of a terminal and a base station capable of performing some or all of the embodiments described with reference to FIGS. 1 to 14 will be described with reference to the drawings. The foregoing description may be omitted to avoid redundant descriptions, and in such cases, the omitted content may be applied substantially identically to the following description, provided that it does not contradict the technical concept of the invention.

[0337] FIG. 15 is a diagram showing the configuration of a terminal (1500) according to another embodiment.

[0338] Referring to FIG. 15, a terminal (1500) according to another embodiment includes a transmitter (1520), a receiver (1530), and a control unit (1510) that controls the transmitter and the receiver.

[0339] The control unit (1510) controls the overall operation of the terminal (1500) according to the method of performing monitoring of channel state information (CSI) prediction performance based on artificial intelligence and machine learning (AI / ML) models necessary to perform the above-described embodiments.

[0340] The transmitting unit (1520) and the receiving unit (1530) are used to transmit and receive signals, messages, and data to and from a base station that are necessary to perform the above-described embodiments.

[0341] The control unit (1510) can receive configuration information for predictive performance monitoring from the base station through the receiving unit (1530).

[0342] According to one example, the configuration information may be a higher-level signal (e.g., RRC message) or a physical-level signal (e.g., DCI, MAC CE) containing overall control parameters required for the control unit (1510) to monitor the performance of an AI / ML prediction model for function-based lifecycle management (LCM) and report the results. For example, the configuration information may comprehensively include periodic, semi-static, or non-periodic CSI-RS configuration information to be utilized as a channel measurement resource (CMR), a method for calculating performance metrics (e.g., SGCS or NMSE), and threshold and timer configuration parameters for event-based reporting.

[0343] In particular, the configuration information may include identification information indicating the reporting target among multiple time resources. Here, the multiple time resources may refer to multiple slots that are the targets for CSI prediction by an AI / ML model. The identification information may be provided in the form of a bit string, for example, and the base station may explicitly instruct the control unit (1510) through the identification information to calculate and report performance indicators among the total time resources, or to specify a specific slot or a specific resource group (e.g., totalslots, sub-slots, etc.). Through this, the control unit (1510) can prevent unnecessary computation on the total resources and efficiently limit the monitoring targets.

[0344] Additionally, the control unit (1510) can determine a performance metric representing the error between the predicted channel state information and the reference channel state information for a plurality of time resources derived through an AI / ML model, based on the received configuration information. According to one example, the predicted channel state information may refer to the CSI predicted by the AI / ML model on the terminal (1500) side, and the reference channel state information may refer to the actual ground-truth channel measured by the control unit (1510). The performance metric may be calculated as at least one of SGCS (Squared Generalized Cosine Similarity) or NMSE (Normalized Mean Square Error), which represent the accuracy of the prediction.

[0345] The control unit (1510) can divide multiple time resources into at least one resource group and determine performance indicators only for the resource group indicated by identification information. For example, if the total time resources to be predicted consist of multiple slots, the control unit (1510) can divide them into multiple sub-slot groups having a specific length. If the identification information included in the configuration information received from the base station is in the form of a bitmap (e.g., '1010') indicating a specific sub-slot group, the control unit (1510) can optimize the computation load (complexity) by performing SGCS or NMSE operations only on the first and third resource groups where the bit is activated as '1'.

[0346] Subsequently, if the performance indicator satisfies a preset reporting condition, the control unit (1510) can transmit a predicted performance monitoring result, including a performance indicator for a time resource indicated by the configuration information, to a base station through the transmission unit (1520). According to one example, the control unit (1510) can report the predicted performance monitoring result to the base station based on events. The reporting condition may include an event in which the ratio or number of times the performance indicator falls below a first reference value exceeds a second reference value. As a specific example, the control unit (1510) can count the number of defective slots in which the SGCS value calculated for each slot falls below '0.5 (first reference value)', and determine that a performance degradation event has occurred when the ratio of these defective slots to the total target slots exceeds '30% (second reference value)'.

[0347] Additionally, according to one example, the control unit (1510) monitors the cumulative number of times an event occurs within a predetermined time interval, and when the cumulative number exceeds a threshold, it can control the transmission unit (1520) to transmit the prediction performance monitoring result. For example, the control unit (1510) can trigger to transmit the monitoring result for fallback to the base station only when the event accumulates and occurs three or more times (threshold) within an observation window composed of 100 slots or a specific timer interval.

[0348] Additionally, when the control unit (1510) reports performance indicators for multiple time resources to the base station, a data compression method may be applied to prevent large uplink (UL) overhead from occurring. The control unit (1510) may approximate the patterns of performance indicators determined for each time resource using a codebook pre-configured between the terminal (1500) and the base station, and transmit the identifier and coefficient of the codeword within the codebook used for approximation as a result of predictive performance monitoring. As a specific example, the control unit (1510) may form performance indicator values ​​for multiple slots into a single vector and model it as a linear combination of multiple codeword patterns within the codebook shared with the base station in advance. Instead of individually transmitting all dozens of performance indicators, the control unit (1510) can drastically reduce the size of the feedback payload by transmitting only the specific pattern number (codeword index) selected for approximation and its combination ratio (coefficient) through the transmitter (1520).

[0349] Additionally, when the control unit (1510) transmits prediction performance monitoring results for multiple resource groups, a situation may occur where uplink transmission resources (e.g., PUCCH or PUSCH payload) are insufficient. In this case, the control unit (1510) may determine whether to transmit prediction performance monitoring results for at least some resource groups based on a priority determined by at least one of the range of time resources included in the resource group and the index assigned to the resource group. For example, the control unit (1510) may give the highest priority to the group (totalslots) that encompasses all slots to be predicted and transmit them preferentially, and among the sub-slot groups, it may give higher priority to the group with a lower index number (e.g., priority of earlier time resources) or the group corresponding to a specific even / odd index rule. The control unit (1510) can prevent transmission errors and reliably provide essential monitoring information to the base station by first transmitting the results of the high-priority group through the transmission unit (1520) within the limits of the allocated resources and omitting (dropping) the transmission of the results of the low-priority group.

[0350] According to this, a method and device for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in next-generation wireless access networks can be provided. In addition, the reliability of AI / ML-based channel prediction models can be verified in real time, while simultaneously reducing unnecessary uplink feedback signaling overhead.

[0351] FIG. 16 is a drawing showing the configuration of a base station (1600) according to another embodiment.

[0352] Referring to FIG. 16, a base station (1600) according to another embodiment includes a transmitter (1620), a receiver (1630), and a control unit (1610) that controls the transmitter and the receiver.

[0353] The control unit (1610) controls the overall operation of the base station (1600) and the operation of the repeater according to the method of performing communication using artificial intelligence and machine learning necessary to perform the above-described embodiments.

[0354] The transmitting unit (1620) and the receiving unit (1630) are used to transmit and receive signals, messages, and data to and from the terminal that are necessary to perform the above-described embodiments.

[0355] The control unit (1610) can transmit configuration information for monitoring the prediction performance to the terminal via the transmission unit (1620). According to one example, the configuration information may be a higher-layer signal (e.g., RRC message) or a physical layer signal (e.g., DCI, MAC CE) containing overall parameters to control the terminal to monitor the performance of an AI / ML prediction model for function-based lifecycle management (LCM) and report the results. For example, the configuration information may comprehensively include periodic, semi-static, or non-periodic CSI-RS configuration information to be utilized as a channel measurement resource (CMR), a method for calculating performance metrics (e.g., SGCS or NMSE), and threshold and timer setting parameters for event-based reporting.

[0356] In particular, the configuration information may include identification information indicating the reporting target among multiple time resources. Here, the multiple time resources may refer to multiple slots that are the targets for CSI prediction by the terminal's AI / ML model. The identification information may be provided in the form of a bit string, for example, and the control unit (1610) may explicitly instruct the terminal, through the identification information, to calculate and report performance indicators among the total time resources, and to specify a specific slot or a specific resource group (e.g., totalslots, sub-slots, etc.). Through this, the control unit (1610) can prevent the terminal from performing unnecessary operations on the total resources and control it to efficiently limit the monitoring targets.

[0357] Additionally, the control unit (1610) may receive a prediction performance monitoring result from the terminal via the receiver (1630), which includes a performance metric for a time resource indicated by configuration information. Here, the performance metric may be a performance metric representing the error between the prediction channel state information and the reference channel state information for a plurality of time resources derived by the terminal through an AI / ML model based on the configuration information. According to one example, the prediction channel state information may refer to the CSI predicted by the terminal-side AI / ML model, and the reference channel state information may refer to the actual ground-truth channel measured by the terminal. The performance metric may be calculated as at least one of SGCS (Squared Generalized Cosine Similarity) or NMSE (Normalized Mean Square Error), which indicates the accuracy of the prediction.

[0358] At this time, the performance indicator may be determined only for the resource group indicated by the identification information after the multiple time resources are divided into at least one resource group by the terminal. For example, if the total time resources to be predicted consist of multiple slots, they may be divided into multiple sub-slot groups having a specific length. If the identification information transmitted by the control unit (1610) including in the configuration information is in the form of a bitmap (e.g., '1010') indicating a specific sub-slot group, the terminal optimizes the computation load (Complexity) by performing SGCS or NMSE operations only on the first and third resource groups where the bit is activated as '1', and the control unit (1610) can receive the calculated result for the corresponding resource group through the receiver (1630).

[0359] According to one example, the control unit (1610) may receive predicted performance monitoring results from the terminal on an event basis through the receiving unit (1630). The reporting conditions may include an event in which the ratio or number of times a performance indicator calculated by the terminal falls below a first reference value exceeds a second reference value. As a specific example, when the number of defective slots in which the SGCS value calculated by the terminal for each slot falls below '0.5 (first reference value)' is counted, and the ratio of these defective slots to the total number of target slots exceeds '30% (second reference value)', it is determined that a performance degradation event has occurred and a report may be made to the base station (1600).

[0360] Additionally, according to one example, the control unit (1610) may receive a prediction performance monitoring result that is triggered by the terminal when the cumulative number of events occurring within a predetermined time interval exceeds a threshold through the receiver (1630). For example, the control unit (1610) may receive a monitoring result for fallback only when an event accumulates and occurs three or more times (threshold) within an observation window consisting of 100 slots or a specific timer interval at the terminal.

[0361] Additionally, when the terminal reports performance indicators for multiple time resources to the base station (1600), a data compression method may be applied to prevent large uplink (UL) overhead from occurring. When the pattern of performance indicators determined for each time resource through the receiver (1630) is approximated using a codebook pre-configured between the terminal and the base station (1600), the control unit (1610) may receive the identifier and coefficient of the codeword within the codebook used for approximation from the terminal as a result of predictive performance monitoring. As a specific example, the terminal may compose performance indicator values ​​for multiple slots into a single vector and model this as a linear combination of multiple codeword patterns within the codebook shared with the base station (1600) in advance. Instead of receiving all dozens of performance indicators individually, the control unit (1610) can obtain the effect of drastically reducing the size of the uplink feedback payload by receiving only the specific pattern number (codeword index) selected by the terminal's approximation and the combination ratio (coefficient).

[0362] Additionally, when the terminal transmits the prediction performance monitoring results for multiple resource groups, a situation may occur where the terminal's uplink transmission resources (e.g., PUCCH or PUSCH payload) are insufficient. In this case, the control unit (1610) may receive the prediction performance monitoring results for at least some resource groups, the transmission status determined by the terminal, through the receiver (1630), according to a priority determined based on at least one of the range of time resources included in the resource group and the index assigned to the resource group. For example, the terminal may assign the highest priority to the group (totalslots) that encompasses all slots to be predicted and transmit them preferentially, and among the sub-slot groups, it may assign a higher priority to the group with a lower index number (e.g., priority of earlier time resources) or the group corresponding to a specific even / odd index rule. The control unit (1610) receives the results of a high-priority group first from the terminal within the limits of the resources allocated to the terminal, and receives the results of a low-priority group with transmission omitted (dropped) by the terminal, thereby preventing transmission errors and reliably obtaining essential monitoring information.

[0363] According to this, a method and device for monitoring channel state information prediction performance based on artificial intelligence and machine learning models in next-generation wireless access networks can be provided. In addition, the reliability of AI / ML-based channel prediction models can be verified in real time, while simultaneously reducing unnecessary uplink feedback signaling overhead.

[0364] The aforementioned embodiments may be supported by standard documents disclosed in at least one of the wireless access systems IEEE 802, 3GPP, and 3GPP2. That is, steps, configurations, and parts in the embodiments that are not described to clearly reveal the technical concept may be supported by the aforementioned standard documents. Furthermore, all terms disclosed in this specification may be explained by the standard documents disclosed above.

[0365] The embodiments described above may be implemented through various means. For example, the embodiments may be implemented by hardware, firmware, software, or a combination thereof.

[0366] In the case of implementation by hardware, the method according to the embodiments may be implemented by one or more ASICs (Application Specific Integrated Circuits), DSPs (Digital Signal Processors), DSPDs (Digital Signal Processing Devices), PLDs (Programmable Logic Devices), FPGAs (Field Programmable Gate Arrays), processors, controllers, microcontrollers, or microprocessors.

[0367] In the case of implementation by firmware or software, the method according to the embodiments may be implemented in the form of a device, procedure, or function that performs the functions or operations described above. Software code may be stored in a memory unit and executed by a processor. The memory unit may be located inside or outside the processor and may exchange data with the processor by various means already known.

[0368] Additionally, terms such as "system," "processor," "controller," "component," "module," "interface," "model," or "unit" described above may generally refer to computer-related entities, hardware, combinations of hardware and software, software, or running software. For example, the aforementioned components may be, but are not limited to, processes driven by a processor, processors, controllers, control processors, objects, execution threads, programs, and / or computers. For example, both the application running on the controller or processor and the controller or processor may be components. One or more components may reside within a process and / or execution thread, and the components may be located on a single device (e.g., a system, a computing device, etc.) or distributed across two or more devices.

[0369] The foregoing description is merely an illustrative explanation of the technical concept of the present disclosure, and those skilled in the art to which the present disclosure pertains may make various modifications and variations within the scope of the essential characteristics of the technical concept. Furthermore, since these embodiments are intended to explain, not limit, the scope of the technical concept is not limited by these embodiments. The scope of protection of the present disclosure shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present disclosure.

[0370]

[0371] CROSS-REFERENCE TO RELATED APPLICATION

[0372] This patent application claims priority under Section 119(a) of the U.S. Patent Act (35 USC § 119(a)) to Patent Application No. 10-2025-0036909, filed in Korea on March 21, 2025, Patent Application No. 10-2025-0036907, filed in Korea on March 21, 2025, Patent Application No. 10-2025-0036908, filed in Korea on March 20, 2026, Patent Application No. 10-2026-0050982, filed in Korea on March 20, 2026, all of which are incorporated into this patent application by reference. Furthermore, if this patent application claims priority in countries other than the United States for the same reasons as above, all such contents shall be incorporated into this patent application by reference.

Claims

1. A method for a terminal (user equipment; UE) to perform monitoring of channel state information (CSI) prediction performance based on artificial intelligence and machine learning (AI / ML) models, A step of receiving configuration information for monitoring the predicted performance from a base station; Based on the above configuration information, a step of determining a performance metric representing the error between the predicted channel state information and the reference channel state information for a plurality of time resources derived through the AI / ML model; and A method comprising the step of transmitting a predicted performance monitoring result, including the performance indicator for a time resource indicated by the configuration information, to the base station when the performance indicator satisfies a preset reporting condition.

2. In Paragraph 1, The above configuration information includes identification information indicating a reporting target among the plurality of time resources, and A method comprising the step of determining the above performance indicator, dividing the plurality of time resources into at least one resource group and determining the above performance indicator only for the resource group indicated by the identification information.

3. In Paragraph 2, The above-mentioned transmitting step comprises a step of determining whether to transmit the predicted performance monitoring results for at least some resource groups according to a priority determined based on at least one of the range of time resources included in the resource group and an index assigned to the resource group.

4. In Paragraph 1, The above-mentioned transmitting step comprises the step of approximating the pattern of performance indicators determined by the above-mentioned time resources using a codebook pre-configured between the terminal and the base station, and transmitting the identifier and coefficient of the codeword in the codebook used for the approximation as the predicted performance monitoring result.

5. In Paragraph 1, The above reporting condition includes an event in which the ratio or number of times the above performance indicator falls below a first threshold value exceeds a second threshold value, and A method comprising the step of transmitting, wherein the above-mentioned transmitting step monitors the cumulative number of times the above-mentioned event occurs within a predetermined time interval, and transmits the above-mentioned prediction performance monitoring result when the cumulative number exceeds a threshold.

6. A method for a base station to control the monitoring of channel state information (CSI) prediction performance based on artificial intelligence and machine learning (AI / ML) models, A step of transmitting configuration information for monitoring the predicted performance to a terminal; and The method includes the step of receiving a predictive performance monitoring result from the terminal, the result of which includes a performance metric for a time resource indicated by the configuration information. The above performance indicator represents the error between the predicted channel state information and the reference channel state information for a plurality of time resources derived through the AI / ML model based on the configuration information by the terminal, and The method of receiving the above-mentioned predictive performance monitoring result from the terminal when the above-mentioned performance indicator satisfies a preset reporting condition.

7. In Paragraph 6, The above configuration information includes identification information indicating a reporting target among the plurality of time resources, and A method in which the above performance indicator is determined by the terminal only for the resource group indicated by the identification information after the plurality of time resources are divided into at least one resource group.

8. In Paragraph 7, The receiving step comprises receiving the predicted performance monitoring result for at least some resource groups, the transmission status determined by the terminal according to a priority determined based on at least one of the range of time resources included in the resource group and an index assigned to the resource group.

9. In Paragraph 6, The receiving step comprises, when the pattern of performance indicators determined by the time resource is approximated using a codebook pre-configured between the terminal and the base station, receiving from the terminal the identifier and coefficient of a codeword in the codebook used for the approximation as the predicted performance monitoring result.

10. In Paragraph 6, The above reporting condition includes an event in which the ratio or number of times the above performance indicator falls below a first threshold value exceeds a second threshold value, and The receiving step comprises a method for receiving the prediction performance monitoring result when the cumulative number of times the event occurs within a predetermined time interval exceeds a threshold.

11. In a terminal (user equipment; UE) that performs monitoring of channel state information (CSI) prediction performance based on artificial intelligence and machine learning (AI / ML) models, Transmitter; Receiver; and It includes a control unit that controls the operation of the transmitting unit and the receiving unit, wherein The above control unit is, A terminal that receives configuration information for predictive performance monitoring from a base station, determines a performance metric representing an error between predictive channel state information and reference channel state information for a plurality of time resources derived through the AI / ML model based on the configuration information, and transmits a predictive performance monitoring result including the performance metric for a time resource indicated by the configuration information to the base station when the performance metric satisfies a preset reporting condition.

12. In Paragraph 11, The above configuration information includes identification information indicating a reporting target among the plurality of time resources, and The above control unit divides the plurality of time resources into at least one resource group and determines the performance indicator only for the resource group indicated by the identification information.

13. In Paragraph 12, The control unit determines whether to transmit the predicted performance monitoring result for at least some resource groups according to a priority determined based on at least one of the range of time resources included in the resource group and an index assigned to the resource group.

14. In Paragraph 11, The control unit approximates the pattern of performance indicators determined by the time resource using a codebook pre-configured between the terminal and the base station, and transmits the identifier and coefficient of the codeword in the codebook used for the approximation as the predicted performance monitoring result.

15. In Paragraph 11, The above reporting condition includes an event in which the ratio or number of times the above performance indicator falls below a first threshold value exceeds a second threshold value, and The above control unit monitors the cumulative number of times the event occurs within a predetermined time interval, and when the cumulative number exceeds a threshold, the terminal transmits the prediction performance monitoring result.

Images