Method and device for measuring and transmitting channel state information in wireless communication system

Abstract

This method of a wireless communication system may include the steps of: performing, by a wireless user device, training and testing on an AI / ML model for CSI prediction, and reporting training and test information to a base station; receiving, from the base station, an inference triggering of the AI / ML model for CSI prediction, and generating predicted CSI by performing inference of the AI / ML model on the basis of same; receiving, from the base station, a monitoring indication for the AI / ML model for CSI prediction; and reporting, to the base station, fallback-related information on the basis of the monitoring indication for the AI / ML model.

Description

Method and device for measuring and transmitting channel state information in a wireless communication system

[0001] The present disclosure relates to a method and device for measuring and transmitting channel state information in a wireless communication system. Specifically, the present disclosure relates to a method and device for measuring and transmitting channel state information (CSI) based on artificial intelligence / machine learning (AI / ML) in a wireless communication system.

[0002]

[0003] The International Telecommunication Union (ITU) is developing the International Mobile Telecommunication (IMT) framework and standards, and is currently discussing fifth-generation (5G) communications through a program called "IMT for 2020 and beyond."

[0004] To meet the requirements presented in "IMT for 2020 and beyond," the 3rd Generation Partnership Project (3GPP) NR (New Radio) system is being discussed to support various numerologies based on time-frequency resource units, taking into account various scenarios, service requirements, and potential system compatibility.

[0005] Additionally, 5G communications can support the transmission of physical signals or physical channels through multiple beams to overcome adverse channel conditions such as high path loss, phase noise, and frequency offset that occur at high carrier frequencies. Through this, 5G communications can support applications such as enhanced Mobile Broadband (eMBB), massive Machine Type Communications (mMTC), and Ultra Reliable and Low Latency Communication (URLLC).

[0006]

[0007] The technical problem of the present disclosure is a method and device for measuring and transmitting CSI in a wireless communication system.

[0008] The technical problem of the present disclosure is a method and device for measuring and reporting CSI using an AI / ML model for CSI prediction.

[0009] The technical problem of the present disclosure is a method and device for evaluating model performance of an AI / ML model for CSI prediction.

[0010] The technical problem of the present disclosure is a method and device for falling back to an existing CSI operation based on performance evaluation of an AI / ML model for CSI prediction.

[0011] The technical problems to be achieved in the present disclosure are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by a person having ordinary skill in the technical field to which the present disclosure belongs from the description below.

[0012]

[0013] According to one aspect of the present disclosure, a method may include a step of a wireless user device performing learning and testing on an artificial intelligence (AI) / machine learning (ML) model for predicting channel state information (CSI) and reporting learning and testing information to a base station, a step of receiving an inference triggering of an AI / ML model for predicting CSI from the base station and performing inference of the AI / ML model for predicting CSI based thereon to generate predicted CSI, a step of receiving a monitoring instruction of the AI / ML model for predicting CSI from the base station, and a step of reporting fallback-related information to the base station based on the monitoring instruction of the AI / ML model.

[0014] In addition, according to one aspect of the present disclosure, a wireless user device includes at least one processor and a memory storing instructions for causing the wireless user device to perform a specific operation by the at least one processor, wherein the specific operation may include: performing learning and testing on an artificial intelligence (AI) / machine learning (ML) model for predicting channel state information (CSI), reporting learning and testing information to a base station, receiving inference triggering of the AI / ML model for predicting CSI from the base station, performing inference of the AI / ML model for predicting CSI based thereon to generate predicted CSI, receiving a monitoring instruction of the AI / ML model for predicting CSI from the base station, and reporting fallback-related information to the base station based on the monitoring instruction of the AI / ML model.

[0015] Additionally, the following may be commonly applied:

[0016] According to one aspect of the present disclosure, the fallback-related information is a performance monitoring result, and the performance monitoring result is determined based on a performance evaluation index derived based on a comparison of predicted CSI and ground-true CSI, and whether to perform fallback at a base station based on the performance monitoring result can be determined based on a threshold value.

[0017] In addition, according to one aspect of the present disclosure, the fallback-related information is a fallback recommendation report, and the fallback recommendation report is information derived based on a performance evaluation index and a threshold value derived based on a comparison of predicted CSI and ground-true CSI, and whether to perform fallback can be determined at the base station based on the fallback recommendation report.

[0018] Additionally, according to one aspect of the present disclosure, the fallback-related information is measurement information, and the measurement information includes predicted CSI and ground-true CSI, and a performance evaluation index is derived from the base station based on the measurement information to determine whether to perform fallback.

[0019] Additionally, according to one aspect of the present disclosure, the fallback-related information is a performance evaluation index, and the performance evaluation index is derived based on a comparison of the predicted CSI and the ground-true CSI, and the performance evaluation index and a threshold value are compared at the base station to determine whether to perform fallback.

[0020]

[0021] According to the present disclosure, a method for measuring and transmitting CSI in a wireless communication system can be provided.

[0022] According to the present disclosure, a method for measuring and reporting CSI using an AI / ML model for CSI prediction can be provided.

[0023] According to the present disclosure, a method for evaluating model performance of an AI / ML model for CSI prediction can be provided.

[0024] According to the present disclosure, a method for falling back to an existing CSI operation based on performance evaluation of an AI / ML model for CSI prediction can be provided.

[0025] The effects that can be obtained from the present disclosure are not limited to the effects mentioned above, and other effects that are not mentioned will be clearly understood by a person having ordinary skill in the art to which the present disclosure pertains from the description below.

[0026]

[0027] FIG. 1 is a drawing for explaining the frame structure of a wireless communication system to which the present disclosure can be applied.

[0028] FIG. 2 is a diagram showing the resource structure of a wireless communication system to which the present disclosure can be applied.

[0029] FIG. 3 is a diagram illustrating CSI-RS resources to which the present disclosure can be applied.

[0030] FIG. 4 is a diagram illustrating a type 1 CSI applicable to the present disclosure.

[0031] FIG. 5 is a diagram illustrating a method of using multiple SRSs that can be applied to the present disclosure.

[0032] FIG. 6 is a diagram illustrating a method for performing non-codebook based precoding that can be applied to the present disclosure.

[0033] Figure 7 is a diagram illustrating AI / ML model life cycle management applicable to the present disclosure.

[0034] FIG. 8 is a diagram illustrating a method for performing AI / ML model training based on Type 1 applicable to the present disclosure.

[0035] FIG. 9 is a diagram illustrating a method for performing AI / ML model training based on Type 2 applicable to the present disclosure.

[0036] FIG. 10 is a diagram illustrating a method for performing AI / ML model training based on Type 3 applicable to the present disclosure.

[0037] FIG. 11 is a diagram illustrating various types of functions applicable to the present disclosure.

[0038] FIG. 12 is a diagram showing the relationship between features, functions, and models applicable to the present disclosure.

[0039] FIG. 13 is a diagram illustrating an inference operation for CSI prediction applicable to the present disclosure.

[0040] FIG. 14 is a diagram illustrating a fallback operation for an AI / ML-based CSI prediction model applicable to the present disclosure.

[0041] FIG. 15 is a diagram illustrating a fallback operation for an AI / ML-based CSI prediction model applicable to the present disclosure.

[0042] FIG. 16 is a diagram illustrating a fallback operation for an AI / ML model for CSI prediction applicable to the present disclosure.

[0043] FIG. 17 is a diagram illustrating a fallback operation for an AI / ML-based CSI prediction model applicable to the present disclosure.

[0044] FIG. 18 is a flowchart illustrating a fallback operation for AI / ML-based CSI reporting applicable to the present disclosure.

[0045] Figure 19 is a drawing showing a base station device and a terminal device to which the present disclosure can be applied.

[0046]

[0047] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings so that those skilled in the art can easily implement the present disclosure. However, the present disclosure may be implemented in various different forms and is not limited to the embodiments described herein.

[0048] In describing embodiments of the present disclosure, detailed descriptions of known configurations or functions will be omitted if they are deemed to obscure the gist of the present disclosure. Furthermore, portions of the drawings that are irrelevant to the description of the present disclosure have been omitted, and similar portions are designated with similar reference numerals.

[0049] In the present disclosure, when a component is said to be "connected," "coupled," or "connected" to another component, this may include not only a direct connection, but also an indirect connection in which another component exists in between. Furthermore, when a component is said to "include" or "have" another component, unless otherwise specifically stated, this does not exclude the other component, but rather implies that the other component may be included.

[0050] In this disclosure, terms such as first, second, etc. are used solely to distinguish one component from another, and do not limit the order or importance of components unless specifically stated otherwise. Accordingly, within the scope of this disclosure, a first component in one embodiment may be referred to as a second component in another embodiment, and similarly, a second component in one embodiment may be referred to as a first component in another embodiment.

[0051] In this disclosure, distinct components are used to clearly illustrate their respective characteristics, and do not necessarily imply that the components are separated. That is, multiple components may be integrated into a single hardware or software unit, or a single component may be distributed into multiple hardware or software units. Therefore, even if not specifically mentioned, such integrated or distributed embodiments are also included within the scope of this disclosure.

[0052] In the present disclosure, the components described in various embodiments are not necessarily essential components, and some may be optional components. Therefore, embodiments comprising a subset of the components described in one embodiment are also within the scope of the present disclosure. Furthermore, embodiments including other components in addition to the components described in various embodiments are also within the scope of the present disclosure.

[0053] The present disclosure describes a wireless communication network, and operations performed in the wireless communication network may be performed in a process of controlling the network and transmitting or receiving a signal in a system (e.g., a base station) that manages the wireless communication network, or in a process of transmitting or receiving a signal in a terminal connected to the wireless network.

[0054] It is self-evident that various operations performed for communication with terminals in a network consisting of multiple network nodes including a base station can be performed by the base station or other network nodes other than the base station. 'Base station (BS)' can be replaced by terms such as fixed station, Node B, eNodeB (eNB), ng-eNB, gNodeB (gNB), and access point (AP). In addition, 'terminal' can be replaced by terms such as UE (User Equipment), MS (Mobile Station), MSS (Mobile Subscriber Station), SS (Subscriber Station), and non-AP station (non-AP STA).

[0055] In the present disclosure, transmitting or receiving a channel means transmitting or receiving information or a signal through the channel. For example, transmitting a control channel means transmitting control information or a signal through the control channel. Similarly, transmitting a data channel means transmitting data information or a signal through the data channel.

[0056] In the following description, the term NR (New Radio) system is used for the purpose of distinguishing the system to which various examples of the present disclosure are applied from existing systems; however, the scope of the present disclosure is not limited by this term.

[0057] NR systems support a variety of subcarrier spacings (SCS) to accommodate diverse scenarios, service requirements, and potential system compatibility. Furthermore, NR systems can support the transmission of physical signals / channels across multiple beams to overcome challenging channel conditions, such as high path loss, phase noise, and frequency offsets that occur at high carrier frequencies. This enables NR systems to support applications such as enhanced Mobile Broadband (eMBB), massive Machine Type Communications (mMTC) / ultra Machine Type Communications (uMTC), and Ultra Reliable and Low Latency Communications (URLLC).

[0058] Hereinafter, 5G mobile communication technology can be defined to include not only the NR system, but also the existing LTE-A (Long Term Evolution-Advanced) system and LTE (Long Term Evolution) system. 5G mobile communication may include technology that operates in consideration of backward compatibility with previous systems as well as the newly defined NR system. Therefore, the 5G mobile communication below may include technology that operates based on the NR system and technology that operates based on previous systems (e.g., LTE-A, LTE), and is not limited to a specific system.

[0059] First, we would like to briefly explain the physical resource structure of the wireless communication system to which the present invention is applied.

[0060] FIG. 1 is a drawing for explaining an NR frame structure to which the present disclosure can be applied.

[0061] In NR, the basic unit of time domain is It can be, , and N can be 4096. Meanwhile, in LTE, the basic unit of the time domain is It can be, And, =2048. The constant for the multiplication relationship between the NR time base unit and the LTE time base unit is k= can be defined as

[0062] Referring to Figure 1, the time structure of a frame for downlink / uplink (DL / UL) transmission is can have. Here, one frame is It consists of 10 subframes corresponding to time. The number of consecutive OFDM symbols in each subframe is = It can be. In addition, each frame is divided into two half frames of the same size, half frame 1 can be composed of sub frames 0-4, and half frame 2 can be composed of sub frames 5-9.

[0063] represents the timing advance (TA) between the downlink (DL) and uplink (UL). Here, the transmission timing of the uplink transmission frame i is determined based on the downlink reception timing at the terminal, based on the following mathematical expression 1.

[0064] [Mathematical Formula 1]

[0065]

[0066]

[0067] Here, It may be a TA offset value that occurs due to differences in duplex mode, etc. In FDD (Frequency Division Duplex), has a value of 0, but in TDD (Time Division Duplex), it takes into account the margin for DL-UL switching time. It can be defined as a fixed value. For example, in TDD (Time Division Duplex) of FR1 (Frequency Range 1), which is a frequency below 6 GHz, is 39936 or 25600 It could be 39936 is 20.327μs, and 25600 is 13.030μs. Also, at FR2 (Frequency Range 2), which is a millimeter wave (mmWave) frequency, is 13792 It could be. At this time, 13792 is 7.020 μs.

[0068] FIG. 2 is a diagram showing an NR resource structure to which the present disclosure can be applied.

[0069] Resource elements (REs) within a resource grid can be indexed according to each subcarrier spacing. Here, one resource grid can be created for each antenna port and each subcarrier spacing. Uplink and downlink transmission and reception can be performed based on the corresponding resource grid.

[0070] In the frequency domain, one Resource Block (RB) consists of 12 REs, and each of the 12 REs can be configured with an index (nPRB) for one RB. The index for an RB can be utilized within a specific frequency band or system bandwidth. The index for an RB can be defined as in the following mathematical expression 2. Here, represents the number of subcarriers per RB, and k represents the subcarrier index.

[0071] [Equation 2]

[0072]

[0073]

[0074] Different numerologies can be configured to meet the diverse services and requirements of NR systems. For example, while LTE / LTE-A systems can support a single subcarrier spacing (SCS), NR systems can support multiple SCSs.

[0075] A new numerology for NR systems supporting multiple SCSs can operate in frequency ranges or carriers such as below 3GHz, 3GHz-6GHz, 6GHZ-52.6GHz or above 52.6GHz to address the issue of not being able to use wide bandwidth in frequency ranges or carriers such as 700MHz or 2GHz.

[0076] Table 1 below shows examples of numerologies supported by the NR system.

[0077] [Table 1]

[0078]

[0079]

[0080] Referring to Table 1 above, the numeral can be defined based on the subcarrier spacing (SCS), cyclic prefix (CP) length, and number of OFDM symbols per slot used in the Orthogonal Frequency Division Multiplexing (OFDM) system. The above values ​​can be provided to the terminal through the upper layer parameters DL-BWP-mu and DL-BWP-cp for the downlink, and through the upper layer parameters UL-BWP-mu and UL-BWP-cp for the uplink.

[0081] In Table 1 above, when the subcarrier spacing setting index (u) is 2, the subcarrier spacing (Δf) is 60 kHz, and normal CP and extended CP can be applied. For other numerology indices, only normal CP can be applied.

[0082] A normal slot can be defined as the basic time unit used to transmit a single piece of data and control information in an NR system. The length of a normal slot can be set to the number of OFDM symbols, which is 14 by default. Furthermore, unlike slots, a subframe has an absolute time length equivalent to 1 ms in an NR system and can be used as a reference time for the length of other time intervals. Here, for coexistence or backward compatibility between LTE and NR systems, a time interval similar to an LTE subframe may be required in the NR standard.

[0083] For example, in LTE, data can be transmitted based on a unit of time called a Transmission Time Interval (TTI), which can be set to one or more subframes. Here, one subframe can be set to 1ms and can contain 14 OFDM symbols (or 12 OFDM symbols).

[0084] In addition, a non-slot can be defined in NR. A non-slot can mean a slot having a number that is at least one symbol smaller than a normal slot. For example, in the case of providing low latency such as URLLC service, latency can be reduced through a non-slot having a number of symbols smaller than a normal slot. Here, the number of OFDM symbols included in a non-slot can be determined by considering the frequency range. For example, a non-slot with a length of 1 OFDM symbol can be considered in a frequency range of 6 GHz or higher. As an additional example, the number of OFDM symbols defining a non-slot can include at least 2 OFDM symbols. Here, the range of the number of OFDM symbols included in a non-slot can be set as the length of a mini-slot up to a predetermined length (e.g., the normal slot length - 1). However, as a specification of a non-slot, the number of OFDM symbols may be limited to 2, 4, or 7 symbols, but is not limited thereto.

[0085] Additionally, for example, in unlicensed bands below 6 GHz, subcarrier spacing where u equals 1 and 2 may be used, and in unlicensed bands above 6 GHz, subcarrier spacing where u equals 3 and 4 may be used. For example, when u equals 4, it may be used for SSB (Synchronization Signal Block).

[0086] [Table 2]

[0087]

[0088]

[0089] Table 2 shows the number of OFDM symbols per slot for normal CP, by subcarrier spacing setting (u). ), number of slots per frame ( ), number of slots per subframe ( ) is shown. Table 2 shows the above-described values ​​based on a normal slot having 14 OFDM symbols.

[0090] [Table 3]

[0091]

[0092]

[0093] Table 3 shows the number of slots per frame and the number of slots per subframe based on normal slots with 12 OFDM symbols per slot when extended CP is applied (i.e., when u is 2 and the subcarrier spacing is 60 kHz).

[0094] As mentioned above, one subframe may correspond to 1ms on the time axis. Additionally, one slot may correspond to 14 symbols on the time axis. For example, one slot may correspond to 7 symbols on the time axis. Accordingly, the number of slots and symbols that can be considered within 10ms corresponding to one radio frame may be set differently. Table 4 may show the number of slots and symbols according to each SCS. In Table 4, the 480kHz SCS may not be considered, but is not limited to these examples.

[0095] [Table 4]

[0096]

[0097]

[0098] A terminal can calculate downlink CSI parameters (e.g., CQI, PMI, RI, L1-RSRP, etc.) through a DL (downlink) CSI (channel state information) reporting procedure and report the calculated CSI information to a base station. The base station can utilize the received CSI information for downlink data transmission. The CSI parameters are measurement values ​​associated with a channel state, and the base station can perform downlink data transmission based on the measurement values. The base station can instruct the terminal about data scheduling information based on the CSI parameters. The terminal generally reports the CSI calculation and derived CSI parameters to the base station as feedback information by receiving a CSI-RS transmitted by the base station. Alternatively, the terminal can measure the channel state through another downlink signal (e.g., SSB (synchronization signal block)) in addition to the CSI-RS and report the measured value to the base station.

[0099]

[0100] Select RI (rank indicator):

[0101] RI can be a CSI parameter that indicates the number of layers available for downlink transmission in a specific channel environment. RI can also indicate the maximum number of uncorrelated paths available for downlink transmission. Here, other CSI parameters (e.g., PMI, CQI) can be calculated based on the rank provided by RI.

[0102]

[0103] Select precoding matrix index (PMI):

[0104] A PMI may include a set of indices corresponding to a precoding matrix. The base station may apply the PMI reported by the terminal to downlink transmission. However, the base station may have the freedom to apply a precoding matrix other than the PMI. For example, the standard may support multiple types of codebooks to report the precoding matrix for the PMI. Each codebook type is described below.

[0105] Codebook type

[0106] - Type 1 single-panel codebooks

[0107] - Type 1 multi-panel codebooks

[0108] - Type 2 codebooks

[0109] - Enhanced Type 2 codebooks

[0110]

[0111] For example, PMI selection may be performed based on the codebook type, the number of transmission layers represented by RI values, and other CSI reporting configuration parameters (e.g., antenna panel dimensions). Each codebook may include a set of precoding matrices. For example, in NR, a dual-stage precoding matrix codebook design may be applied, but is not limited thereto.

[0112] The dual-stage precoding matrix may be as shown in Equation 3 below. In Equation 3, W1 may be a matrix representing a DFT (discrete Fourier transform) beam or a beam group for both polarizations according to the codebook type, and W2 may be selected from one of the following according to the codebook type.

[0113] W2 codebook type

[0114] - Beam selection from W1

[0115] - Weighting coefficients for the beams in W1

[0116] - Cophasing values ​​between two polarizations

[0117]

[0118] [Equation 3]

[0119]

[0120]

[0121] The terminal calculates a PMI value that represents the highest SINR (signal interference noise ratio) value at the receiver using the codebook selected for "Type I single-panel codebook" and "Type I multi-panel codebook" as the codebook type and all possible precoding matrices in the given channel environment. For example, a Type 1 codebook may be considered for PMI selection in a single user multi-input multi-output (SU-MIMO) scenario, but is not limited thereto.

[0122] A Type II codebook can consider a set of orthogonal DFT beams. In Equation 3, the W1 matrix contains information associated with the DFT beams, and the PMI selection function calculates beam amplitude scaling and cophasing values ​​(W2) for all DFT beams within each orthogonal beam group for a given channel environment and number of DFT beams. The terminal can report the i1 and i2 values ​​to the base station based on the PMI selection function, and each index corresponds to a precoding matrix (W) that provides the maximum SINR value. A Type II codebook-based PMI can provide more accurate channel information than a Type I codebook-based PMI. A Type II codebook can provide more accurate channel information because multiple beams are applied to estimate the channel eigenvector. For example, a Type II codebook-based PMI can be used in a MU (multi)-MIMO scenario because it considers multiple beams, but is not limited thereto. For example, a Type 2 codebook can only support up to Rank 2, while an enhanced Type 2 codebook can support up to Rank 4 with overhead reduction methods, which will be described later.

[0123] Furthermore, based on the above, CSI reporting performed by a terminal may consider explicit CSI feedback and implicit CSI feedback. Explicit CSI feedback is a method in which the terminal directly feeds back channel information (e.g., channel eigenvector / eigenvalue, SVD) indicating the channel state measured by the terminal, and may not consider how the base station processes the reported CSI. On the other hand, implicit CSI feedback may be a method in which a selected precoder matrix (desired precoder matrix) W within a set of candidate precoder matrices from a possible precoder codebook is indicated.

[0124] For example, CSI parameters (or CSI contents) related to CSI feedback may be as shown in Table 5 below, but may not be limited thereto.

[0125] [Table 5]

[0126]

[0127]

[0128] The CSI content within a CSI report can be set by different combinations of the CSI parameters in Table 5. Here, the CSI content within a CSI report can be set by reportQuantity.

[0129] For example, if set to "none", the base station may trigger aperiodic CSI-RS transmission. Aperiodic CSI-RS transmission may be triggered by the terminal for channel measurement and receiver parameter adjustment, and related CSI reporting may not be performed by the terminal. Aperiodic CSI-RS transmission may be, but is not limited to, aperiodic transmission of CSI-RS or TRS (tracking reference signal) for tracking purposes for synchronization purposes, or spatial aperiodic receive beam sweeping for the terminal to adjust the receive filter. The following operations do not require the base station to report aperiodic CSI to the terminal, but may operate within the same framework as aperiodic CSI requests.

[0130] - Aperiodic transmission of the Tracking Reference Signal (TRS or CSI-RS for tracking)

[0131] - Aperiodic Rx beam sweeping (so called P3 beam sweep) - adjusts the Rx spatial received filter to the UE

[0132]

[0133] Additionally, CSI contents within a single CSI report may be set as shown in Table 6 below, but may not be limited thereto. For example, the combinations of CSI contents in Table 6 below are representative combinations of CSI contents that a base station will primarily need.

[0134] [Table 6]

[0135]

[0136]

[0137] For example, “cri-RI-i1” in Table 6 can operate based on “Hybrid beamformed / non-precoded CSI acquisition operation” as shown in FIG. 3.

[0138] The base station can mainly use UE-specific beamformed CSI-RS (320) with several ports per CSI-RS resource for the purpose of CSI acquisition. However, in order for the base station to recognize how to beamform the CSI-RS for each UE, non-precoded CSI-RS resources (310) can be transmitted in the middle together with multiple antenna ports (e.g., 32 ports). Here, the UE can report PMI corresponding to all antenna ports (e.g., 32 ports) based on the received CSI-RS, thereby indicating a preferred beam direction. For example, beamforming of the CSI-RS can be based on the wideband / long-term part of the reported precoder. That is, the reported W1 metrics can be used to perform UE-specific CSI-RS beamforming. Additionally, the short-term subband CS (short-term / subband CSI) (W2) can be determined from the beamforming CSI-RS.

[0139] For example, non-precoded CSI-RS (310) can be used to recognize how CSI-RS beamforming is performed, and based on this, the terminal may only need to report CSI corresponding to W1 and not report for W2. This can reduce overhead.

[0140] Additionally, the terminal can select a PMI based on a specific number of transmit antennas. Specifically, the terminal can select a PMI based on a specific number of transmit antennas given by the number of antenna ports of a configured CSI-RS associated with the reporting configuration.

[0141] Additionally, in the case of MU-MIMO scheduling, the base station may select a PMI other than the PMI reported by the target terminal, taking into account interference or channel conditions from other terminals simultaneously scheduled. Considering the above, two types of CSI can be defined in a wireless communication system, as shown in Table 7 below.

[0142] [Table 7]

[0143]

[0144]

[0145] Specifically, type 1 CSI may have different sub-codebook formats based on different antenna configurations at the transmitter. As an example, FIG. 4 is a diagram illustrating type 1 CSI applicable to the present disclosure. Specifically, FIG. 4(a) may be a case where 16 antenna ports are configured with (N1, N2) = (4, 2). The precoding matrix W for type 1 single panel CSI may be expressed as the product of two matrices (W1, W2), each of which may be independently reported as a different part of the overall PMI. For example, W1 may reflect long-term frequency independent channel characteristics, and thus may report channel information corresponding to the entire bandwidth (Wideband reporting). For example, W1 may be as shown in Equation 4 below. W1 may indicate a set of beams pointing in different directions, and matrix B may define one beam. In mathematical expression 4, based on the 2 X 2 block structure, there can be two polarizations, and the selection of W1 can be a limited set selection with the same beam direction.

[0146] [Equation 4]

[0147]

[0148]

[0149] As a concrete example, for rank 1 or 2, a single beam or four neighboring beams can be indicated by the W1 matrix. The correct beam among the four neighboring beams can be selected by W2 reported for each subband, thereby selecting an optimized beam for each subband. Furthermore, W2 can provide cophasing between two polarizations. If W1 defines a single beam, then W2 in B, which is a single-column matrix, can only provide cophasing between two polarizations. On the other hand, if multiple neighboring beams are defined by W1, W2 can select one correct beam and provide cophasing between two polarizations within that beam.

[0150] On the other hand, if the rank is greater than 2, W1 can define N neighboring orthogonal beams, which can be as shown in Equation 5. In Equation 5, N beams used for transmission of R transmission layers can be indicated, and W2 can only provide cophasing between two polarizations. Here, transmission to the same terminal can be possible for up to 8 transmission layers.

[0151] [Equation 5]

[0152]

[0153]

[0154] On the other hand, W2 is short-term and can potentially indicate different channel characteristics depending on the frequency. W2 may not be reported in certain cases. Here, the base station can randomly select W2 for each physical resource block group (PRG), and the terminal can select the channel quality indicator (CQI) based on this assumption.

[0155] Also, Fig. 4(b) may be a Type 1 multi-panel CSI. Multi-panel CSI may be applied considering a case where multiple antenna panels are used at a base station. Here, coherence between different panels may be difficult to guarantee. For example, Fig. 4(b) may be a case where (N1, N2) = (4, 1) and there are 32 antenna ports by 4 panels, but the present invention is not limited thereto. Here, the difference between the single panel of Fig. 4(a) and the multi-panel of Fig. 4(b) is that coherence may be difficult to guarantee between the antenna ports of different antenna panels. In the multi-panel, W1 may be the same set of beams for different polarizations and panels, just like in the single panel. On the other hand, in the case of the multi-panel, W2 may provide cophasing not only between polarizations but also between multi-panels for each subband.

[0156] Type 2 CSI can provide channel information with higher spatial granularity than Type 1 CSI and can be primarily applied to MU-MIMO scenarios. For example, Type 2 CSI may be limited to a maximum of two ranks, while Enhanced Type 2 CSI may be extended to four ranks, but may not be limited to a specific form. For convenience of explanation, the terms Type 2 CSI and Enhanced Type 2 CSI are referred to below, but may not be limited thereto.

[0157] Here, in the above-described mathematical expressions 3 and 4, W1 may be a wideband report. Type 2 CSI may report up to four beams, corresponding to four columns within B. W2 may provide amplitude values ​​(partially wideband and partially subband reporting) and phase values ​​(subband reporting) for each of the four reported beams and two polarizations within B.

[0158] Type 2 CSI can provide more detailed channel model information than Type 1, including main ray and amplitude and phase information. Based on the CSI reported from multiple terminals, the base station can determine which combinations of terminals are appropriate for simultaneous transmission on the same time / frequency resource.

[0159] Additionally, enhanced Type 2 CSI (e.g., Rel-16 enhanced Type 2 CSI) may be considered. Enhanced Type 2 CSI can report a set of beams on the wideband, similar to Type 2 CSI, along with a set of combining coefficients on an additional narrowband. Here, the reported beams can be linearly combined using the combining coefficients to provide a set of precoder vectors for each transmission layer. For example, in the existing Type 2 CSI, the combining coefficient value can be reported independently for each subband. Therefore, even considering the fact that the channels between adjacent subbands are highly correlated, the existing Type 2 CSI must report a relatively large combining coefficient value on a per-subband basis, which can increase the overhead. Enhanced Type 2 CSI can utilize frequency correlation to reduce this overhead. Furthermore, enhanced Type 2 CSI can improve the frequency-domain granularity of the PMI report. For example, compression operations can be applied to subbands or half-subbands based on frequency domain units. While conventional Type 2 CSI uses one precoder per subband, enhanced Type 2 CSI can use a recommended precoder for each frequency domain unit. However, actual reporting can be performed together for all frequency units. Specifically, given layer k, the reported precoders for all frequency domain units can be expressed as in Equation 6 below, where N can be the number of frequency domain units.

[0160] [Equation 6]

[0161]

[0162]

[0163] Here, Equation 6 may not be a mapping of layers and antennas, but may mean a set of precoder vectors for the entire set of frequency domain units for layer k. For the entire k layers, there may be a set of k precoder vectors, each set may consist of N precoder vectors, and the layer and antenna port mapping for a specific frequency domain unit n may be as shown in Equation 6 below.

[0164] Additionally, in the expression for vectors in Equation 7, W1 can be the same as Equation 4 described above, which is the same as the existing Type 2 CSI.

[0165] [Equation 7]

[0166]

[0167] Here, the B columns can correspond to the L selected beams, and W1 can be the same for all frequency domain units. (wideband reporting) Also, W1 can be the same for all layers.

[0168] Additionally, as an example, the enhanced Type 2 CSI has a compression matrix (size M x N) can be considered. can be constructed as a set of row vectors from the DFT basis, and can provide a transformation from N dimensions in the frequency domain (N frequency domain units) to M dimensions in the smaller delay domain. can be frequency independent, but can be reported independently for each layer. That is, can be a single common metric for all frequency domain units, but can be reported independently for each layer. The number of rows can be equal to M in Equation 8. In Equation 8, R can be the number of frequency domain units per subframe (R=1 or 2), and p can be a configurable parameter that controls the amount of compression.

[0169] [Equation 8]

[0170]

[0171] Finally (2L x M) can perform mapping from the delay domain to the beam domain. For example, W2 of the existing Type 2 CSI can generate a similar matrix for all subbands, and the matrix can be mapped from the frequency (subband) domain to the beam domain. On the other hand, can perform allocation from a smaller delay domain to a beam domain. can be adjusted by the reported rank and compression factor parameters p, which allows reporting based on fewer parameters. In addition, You can adjust the amount of CSI to be reported through the value, which allows you to get a smaller Fewer parameters can be reported based on size.

[0172] Additionally, as an example, the enhanced Type 2 CSI may be expanded to a maximum of 4 reported ranks and a maximum number of selectable beams to 6, with reduced overhead and improved frequency domain granularity. Table 8 below may illustrate possible configurations of the enhanced Type 2 CSI, but is not limited thereto.

[0173] [Table 8]

[0174]

[0175]

[0176] Additionally, multi-antenna precoding can be considered in uplink transmission. PUSCH (physical uplink shared channel) MIMO precoding can support up to four layers. For uplink transmission, DFT-based precoding can only support single-user (SU) MIMO. For example, a terminal can perform codebook-based transmission and non-codebook-based transmission with two different PUSCH precoding modes. In case of codebook-based precoding, the uplink grant transmitted by the base station can include precoder-related information (e.g., PMI). It is assumed that the terminal uses the precoder provided by the base station in the uplink differently from the downlink transmission. Furthermore, coherence can be assumed between the terminal antennas in uplink MIMO transmission, and the phase of the signals transmitted on the two antennas can be adjusted. Here, if there is no coherence between the antenna ports, each antenna port may result in a random relative phase, which may mean that the use of antenna port specific weight factors is limited.

[0177] For example, the coherence between antenna ports can correspond to any one of full coherence, partial coherence, and no coherence, and the corresponding information can be provided as terminal capability information. In addition, the use of multiple sounding reference signals (SRS) for codebook-based PUSCH can be considered. FIG. 5 is a diagram illustrating a method of using multiple SRSs applicable to the present disclosure. Referring to FIG. 5, a terminal can transmit relatively large beams on a multi-port SRS. For example, the beams can correspond to different terminal antenna panels (in different directions), and each panel can include a set of antenna elements. That is, each panel can correspond to antenna ports of a multi-port SRS. The base station can transmit an SRI (SRS Resource Indicator) to the terminal to determine whether to perform transmission through which beam. The terminal can determine which transmission to perform within the selected beam based on the number of layers and precoder using SRI and precoder information.

[0178] FIG. 6 is a diagram illustrating a method for performing non-codebook-based precoding applicable to the present disclosure. Referring to FIG. 6, when performing non-codebook-based precoding, a terminal can perform transmission based on a precoder indication and channel measurement received from a base station. In codebook-based precoding, the terminal performs transmission based on an uplink precoder selected based on channel measurement by the base station, but in non-codebook, uplink transmission can be performed through a precoder selected based on measurement based on channel reciprocity. Here, each column of the precoder matrix W can be a digital beam for a corresponding layer. When the terminal selects a precoder for N layers, it can be regarded as selecting N different beam directions, and each beam can correspond to one possible layer.

[0179] For example, referring to FIG. 6, a terminal may be capable of PUSCH transmission based on the selected precoding. However, the precoder selected by the terminal based on downlink channel measurements may not be the optimal precoder from the base station's perspective. Considering the above, the base station may modify the precoder selected by the terminal. For example, the base station may remove some beams or some columns selected by the terminal from the precoder and transmit information about this to the terminal. Specifically, the base station may indicate the result of beam selection based on measured SRS information back to the terminal through SRI. Based on the indicated information, the terminal may perform only some beam transmissions, which may implicitly indicate reduced layer transmission.

[0180]

[0181] A Channel Quality Indicator (CQI) may be a channel measurement value reported by a terminal to a base station for the purpose of reporting a desired Modulation and Coding Scheme (MCS) of the terminal to the base station. A precoding matrix may be indicated by a precoding matrix indicator (PMI) and a rank indicator (RI) reported based on the CQI, and may be applied to CSI-RS transmission. The CSI-RS may be transmitted from the base station to the terminal for channel measurement of the terminal. For example, multiple CSI-RS transmissions for a virtual PDSCH transmission may assume different PMIs and RIs, and multiple CSI-RSs may be transmitted to the terminal. The terminal may first determine (or derive) an RI value through the received CSI-RS, and then determine (or derive) a PMI value based on the determined RI value. The terminal may finally calculate the CQI based on the determined (or derived) RI and PMI values. A precoding matrix indicated based on specific RI and PMI values ​​can be used for subsequent PDSCH transmissions by the base station. Here, when the base station applies the precoding matrix reported by the terminal to PDSCH precoding, the base station can set the MCS based on the reported CQI. That is, the terminal can determine the CQI based on the assumptions about the precoding matrix and RI values ​​based on the PDSCH and report it to the base station. The base station can determine the MCS value for subsequent PDSCH scheduling based on the CQI value reported by the terminal. Here, the base station may not only use the CSI information reported by the terminal, but may utilize the information for PDSCH scheduling for the corresponding terminal by considering the CSI information reported by the terminal.That is, the CSI parameters (PMI, RI, CQI, LI (layer indicator), R1-RSRP (reference signal received power), CRI (CSI-RS resource indicator)) reported by the terminal may be channel state report values ​​recommended by the terminal to the base station. Among the CSI parameters, CQI may be a value corresponding to one SINR (signal interference noise ratio) value and may be used as an input value for link adaptation. When determining the CQI value, the terminal may consider a virtual PDSCH transmission that schedules and transmits one TB (transport block). Here, it may be assumed that the virtual PDSCH is allocated to a radio resource set by the CSI reference resource parameter. The CSI reference resource may mean information corresponding to assumptions about information on the used precoding, RS (reference signal) overhead, bandwidth, and other information. In addition, the CSI reference resource may include relationship information on time resources associated with the CSI. That is, the CSI reference resource may be used as a timing reference. A CSI reference resource may refer to a specific downlink slot to which it is allocated. Accordingly, slots later than the specific downlink slot may not be used as slot resources for CSI reporting.

[0182] The terminal can determine the highest CQI index that can provide virtual PDSCH scheduling with a TB error probability that satisfies the target BLER (Block Error Rate) as the CQI value. For example, the target BLER may be 10%, but may not be limited thereto. As another example, in case of supporting URLLC (ultra reliable low latency communications), the target BLER may be It could be a level.

[0183] In addition, the CQI table may be indicated in the form of a subset of the MCS table. The CQI table may consist of 4 or 5 bits, and four tables may be used to support up to 64QAM, 256QAM, URLLC, and 1024QAM, respectively, but may not be limited to the embodiment. In addition, a single wideband MCS value may be applied and used to the entire TB, but if the base station wishes to use frequency selective scheduling considering multiple sub-bands, the terminal may measure an independent CQI value for each subband and report it to the base station. For example, if a subband CQI is configured, the terminal may use a differential subband CQI in addition to the wideband CQI for each configured subband, as shown in Table 9 below. Additionally, in Table 9, the sub-band offset level can be as shown in Equation 9 below, and the terminal can report an independent CQI value for each subband through a 2-bit subband CQI difference value.

[0184]

[0185] [Table 9]

[0186]

[0187]

[0188] [Equation 9]

[0189]

[0190]

[0191] For example, AI / ML (artificial intelligence / machine learning)-based operations may be performed below. The terms in Table 10 below may be used within the AI / ML-based framework, but may not be limited thereto.

[0192] [Table 10]

[0193]

[0194]

[0195]

[0196] For example, deep learning, a subcategory of machine learning (ML), focuses on parameterizing multi-layer neural networks. Deep learning can be a method for learning data representation and has demonstrated high performance in image classification, speech recognition, and natural language processing. However, deep learning can require a large amount of data during the training process, and the computational complexity required to train deep learning-based models on large data sets can be high, necessitating high hardware performance.

[0197] Regarding AI / ML, AI / ML-based mobile communications can be implemented. For example, AI has been able to reduce CAPEX and OPEX costs by managing data from devices operating on existing wireless communication systems. Furthermore, the introduction of next-generation networks may require low operating costs and high return-to-cost compensation, and AI can help address the complexity and optimization challenges of network operations. Furthermore, network intelligence and automation may be critical issues in next-generation mobile communication systems, and these challenges can be addressed using AI / ML. For example, budget management, network management, equipment lifecycle management, service level agreement (SLA) management, network performance management, and network planning can be implemented using AI / ML. AI / ML can enhance the Quality of Experience (QoE) of network operation and usage, but is not limited to this. Furthermore, AI / ML can improve network quality and enhance the provision of more personalized services, but is not limited to a specific form.

[0198] For example, AI / ML algorithms could be implemented within network equipment for some purposes of equipment and network operation. Here, the issue of optimizing network operation between manufacturers and operators is a major technical issue in operating next-generation communication systems, as described above. Therefore, an AI / ML model that considers this issue may be necessary. In particular, a method of applying AI / ML algorithms within the wireless interface required to transmit and receive wireless signals between multiple network entities, such as between base stations and terminals or between terminals, may be necessary. For example, AI / ML technologies may be utilized in 5G NR (New Radio), a mobile communication technology, or in future 6G mobile communication systems to significantly improve wireless performance and reduce complexity / latency, but may not be limited thereto. Table 11 below may be a basic framework considered in a wireless interface based on AI / ML technologies, but may not be limited thereto.

[0199] [Table 11]

[0200]

[0201]

[0202]

[0203] For example, AI / ML-based operations may consider two types of operations: model generation and inference operations. Model generation may be performed based on model training, model validation, model testing, and application stages. For example, model training may be performed based on at least one of input / output, pre-processing / post-processing, and online / offline application, but may not be limited thereto.

[0204] In addition, the inference operation can be performed based on the input / output, pre / post-processing, and application stages. The life cycle management (LCM) procedure of a general AI / ML framework and an AI / ML model can be performed based on the following steps. As an example, FIG. 7 is a diagram illustrating the AI / ML model life cycle management applicable to the present disclosure. Referring to FIG. 7, data collection for an AI / ML model can be performed. Here, the collected data can include data for model training, monitoring data for model management and performance monitoring, and inference data for model inference. Model training can perform model learning based on training data, and the learned and updated model can be stored. Here, the model can be stored based on a function or model identifier or an identifier related to the model and additional information, but is not limited to a specific form. In addition, the model training can be performed based on model training control information based on model management and performance monitoring.

[0205] Model management and performance monitoring can control model inference based on monitoring data. Model management and performance monitoring can control the activation / deactivation / selection / switching / fallback of model inference, as well as other operations. Model management and performance monitoring can obtain information about the results of model inference. Model inference can obtain inference data, perform inference operations, and obtain result information. Model inference can receive models stored in storage and perform inference operations based on them.

[0206] In relation to the LCM (life cycle management) process, one AI / ML model may have one model ID and related information and / or model functions for AI / ML operations, as described above.

[0207] Actions within the LCM procedure

[0208] - Data Collection

[0209] - Model Training

[0210] - Functionality / model identification

[0211] - Model transfer

[0212] - Model Inference

[0213] - Functionality / model selection, activation, deactivation, switching and fallback operation

[0214] - Functionality / model monitoring

[0215] - Model update

[0216] - Model Storage

[0217] - UE capability

[0218]

[0219] For example, AI / ML-based operations in wireless communication systems can consider the following cases. However, this is only one example, and AIML-based operations can also be applied to other cases.

[0220] AI / ML-based operation of wireless communication systems

[0221] CSI (Channel State Information): compression (**), temporal prediction (*)

[0222] BM(Beam Management): spatial prediction (*), temporal prediction (*)

[0223] Positioning: direct (*), assisted (*)

[0224] (**) is a two-side model (*) is a single-side model

[0225]

[0226] As a concrete example, one can consider AI / ML model training collaborations for CSI compression using a two-sided model.

[0227] For example, with respect to CSI compression, AI / ML model inference can be performed on both the terminal and the base station. That is, AI / ML model inference can be performed on the two-side. To this end, a CSI generation part can be configured on the terminal, and a CSI reconstruction part corresponding to the CSI generation part can be configured on the base station. The CSI generation part is a part for CSI compression, and the CSI reconstruction part can be used to obtain (recover) more accurate CSI for better MU (multi-user) scheduling operation in a massive MIMO scenario, but is not limited thereto, and cooperation on the two-side AI / ML model can be performed in various forms.

[0228] As another example, considering the limitations of the terminal's capability or computing power, it may be considered that a third entity (e.g., a server for the terminal) performs AI / ML model training or inference operations. That is, the terminal-side operations (e.g., model training / inference / monitoring / switching / activation / deactivation / validation / testing) in the above-described AI / ML operations may include the operations of other entities and are not limited to a specific form.

[0229] For example, training collaboration can consider Types 1 through 3 below. In the following, all AI / ML model-related operations performed at the base station or network are assumed to be performed at the network level. However, this may not be limited. Since the "network" is an inclusive term that includes the base station, the two terms may be used interchangeably. What is performed at the network can encompass both the base station side and the broader network side. For convenience of explanation, the following description will be described as being performed at the network.

[0230]

[0231] For example, joint training can be performed based on Type 1. Type 1 can be a method in which training for an AI / ML model is performed on one side / entity. Other sides / entities can then obtain specific model parts through downloading / uploading, and CSI feedback operations can be performed based on these.

[0232] As a specific example, FIG. 8 is a diagram illustrating a method for performing AI / ML model training based on Type 1. Referring to FIG. 8, model training for the CSI generation part of the terminal and the CSI reconstruction part of the base station can be performed on the network (820) side in a joint training manner. Thereafter, the network side can provide the trained model information to the terminal (810) through downloading and can also apply it to the network (820). As another example, model training for the CSI generation part of the terminal and the CSI reconstruction part of the base station can be performed on the terminal (810) side in a joint training manner. Thereafter, the terminal side can apply the trained model information to the terminal (810) and provide it to the network (820) through an uplink. In other words, an AI / ML model can be trained on one side / entity, and the trained AI / ML model can be transferred to another side / entity.

[0233] As a specific example, network-side model 1 training can train the CSI generation part using a data set related to CSI generation for training purposes in the network, and train the CSI reconstruction part by transferring the result value into the training loop of the CSI reconstruction part. Here, the CSI reconstruction part located in the base station can be utilized for subsequent AI / ML model-based CSI feedback. On the other hand, the CSI generation part (i.e., CSI encoder) that performed model training by the base station can transfer the trained model information to the terminal via the wireless interface (UE download AI / ML model). The terminal can perform the AI / ML model-based CSI feedback operation using the downloaded CSI generation part model. That is, CSI compression can be performed through the CSI generation part to perform CSI feedback.

[0234] On the other hand, when the AI / ML model is trained on the terminal side, terminal-side Type 1 training can train the CSI generation part using a data set of CSI generation for training purposes at the terminal. Then, the output value can be provided to the training loop of the CSI reconstruction part, and the CSI reconstruction part can be trained based on the output value. For example, the CSI generation part located at the terminal can be used to generate CSI feedback information based on the AI / ML model. The CSI reconstruction part (i.e., the CSI decoder) that performed the model training by the base station can transmit the trained model information to the base station via the wireless interface (UE upload AI / ML model). Thereafter, the base station can perform the AI / ML model-based CSI feedback operation using the CSI reconstruction part model received through the terminal upload. That is, CSI feedback can be performed by recovering CSI feedback information through the CSI reconstruction part.

[0235] Additionally, AI / ML model joint training type 1 may require AI / ML model exchange. For example, AI / ML model inference may require a common reference for model inference and a common AI / ML model inference algorithm, which may require AL / ML model exchange. For example, terminal-specific joint training type 1 may perform AI / ML model transfer based on the open format of the AI / ML model structure recognized by the terminal. On the other hand, non-terminal-specific joint training type 1 may mean model transfer within an open format without the terminal being aware of the AI / ML model structure. For example, when training a terminal-side model on the network side, the terminal-specific model may be trained on the network side based on the terminal capabilities and the terminal model structure, but may not be limited to a specific form.

[0236]

[0237] As another example, FIG. 9 is a diagram illustrating a method for performing AI / ML model training based on Type 2 applicable to the present disclosure. Referring to FIG. 9, Type 2 model training may be identical to Type 1 in that both the CSI generation part and the CSI reconstruction part are performed within a single, identical loop. However, Type 2 model training differs from Type 1 in that training is not performed on a specific side / entity, but on both sides / entities. That is, both the network and the terminal may be involved in the training.

[0238] For example, referring to FIG. 9, the terminal side can perform a forward propagation (FP) operation based on a data sample. That is, the terminal (910) can generate a CSI feedback result and transmit the FP result as information thereon to the network side. Thereafter, the network (920) can perform a CSI reconstruction operation based on the FP result, thereby training the CSI reconstruction part. The network side can generate BP (backward propagation) information (e.g., gradients) based on the above-described method and transmit the generated BP information back to the terminal side. The terminal side can perform training for the CSI generation part based on the BP information received from the network side. Here, the data set can be aligned on both the terminal side and the network side, and training can be performed through the above-described method.

[0239] Also, as an example, FIG. 10 is a diagram illustrating a method for performing AI / ML model training based on Type 3 applicable to the present disclosure. Referring to FIG. 10, Type 3 model training is a sequential training method, and unlike the joint training of Types 1 and 2, model training can be performed sequentially / independently for each node. That is, Type 3 model training can be performed based on the exchange of aligned data set information between a branch station and a terminal within a separate loop, without training the CSI generation part and the CSI reconstruction part within the same loop. As an example, Type 3 model training can be distinguished into network-first training or terminal-first training, similar to Type 1 model training.

[0240] Referring to Figure 10, in the case of network-first training, the network may use the CSI generation part for training the CSI reconstruction model set on the network side. However, the CSI generation part is for training purposes and may not be used in subsequent model inference.

[0241] The network (1020) side inputs the CSI generation part's input values ​​( ) based on the corresponding CSI generation part and its output value ( ) and the final output value ( through model training of the CSI reconstruction part based on this ) can be generated. The CSI reconstruction part for which training is completed can be used for the AI / ML-based CSI feedback operation of the network. Thereafter, the network (1020) side can share the data set information (e.g., dataset including labels and intermediate results) utilized in the model training of the network side with the terminal after the model training is completed. The terminal (1010) side can apply the data set information to the CSI generation part training. Thereafter, a joint inference operation can be performed through the CSI generation part performed on the terminal (1010) side and the CSI reconstruction part performed on the network (1020). In addition, information on the CSI generation part trained on the terminal (1020) side (e.g., model / functionality identification process - information on available CSI generation parts) can be reported or transmitted to the network (1020). The network (1020) can select or optimize a CSI reconstruction part corresponding to the CSI generation part model of the terminal based on information about the CSI generation part obtained from the terminal (1010).

[0242] As an example, options for training a terminal-side CSI generation model for network-first type 3 joint training can be considered as shown in Table 12 below.

[0243] [Table 12]

[0244]

[0245]

[0246] The terminal-side CSI output format or the base station-side CSI input format may be as shown in Table 13 below. That is, the CSI generated by the terminal by considering AI / ML and transmitted to the base station may be in the form of Table 13 below. In Table 13, Option 1 may be a case where the terminal-side CSI output format or the base station-side CSI input format is in the form of a precoding matrix similar to the existing CSI feedback. The terminal may report CSI feedback to the base station in the form of a precoding matrix together with the confirmed layer number information. For example, the terminal may report CSI feedback in the angular-delay domain as the domain of Option 1-b of Table 13, but may not be limited to the embodiment. (e.g., via spatial-frequency DFT domain transformation to angular / delay domain)

[0247] On the other hand, option 2 in Table 13 may be a case where the terminal-side CSI output format or the base station-side CSI input format is a channel matrix with explicit CSI feedback. That is, the explicit CSI feedback of option 2 may not include precoding vectors assigned to each layer as a channel matrix. The explicit channel matrix may be based on channels in the space-frequency domain (option 2-a) or based on channels in the angle-delay domain (option 2-b), but is not limited to a specific form. For example, in the angle-delay domain of option 2-b, loss may be caused by sub-selection of angle-delay domain-based indices, but is not limited thereto.

[0248] [Table 13]

[0249]

[0250]

[0251] Additionally, as an example, applicable new content and transmission methods between AI CSI reports for AI / ML-based CSI compression may be as shown in Tables 14 and 15 below. The priority applied by reflecting AI CSI reports may be determined based on Table 14 below.

[0252] [Table 14]

[0253]

[0254]

[0255] [Table 15]

[0256]

[0257]

[0258]

[0259] For example, the input values ​​for AI / ML-based encoders can be the eigenvectors of the raw channel matrix (H) or the ray channel matrix (V). Here, if the raw channel matrix is ​​the input, the eigenvector calculation may be performed at the base station, which may increase the complexity of the base station. For example, the raw channel matrix may contain unnecessary information for precoder selection, which may increase unnecessary feedback overhead.

[0260] Therefore, considering the aforementioned characteristics, the AI ​​reporting procedure can be performed by considering which CSI content information will be used as input values ​​for the AI ​​encoder. Furthermore, by determining which CSI content information may impact the AI ​​decoder and how its performance may change accordingly, the AI ​​CSI reporting procedure can be performed by combining appropriate input values ​​from the input values ​​in Tables 14 and 15 described above, but is not limited to a specific format.

[0261] For example, when comparing AI / ML-based quantized CSI compression information for an AI / ML-based CSI feedback mechanism with existing CSI report / contents, the AI / ML-based CSI report fields may have differences as shown in Table 16 below.

[0262] [Table 16]

[0263]

[0264]

[0265] Below, we describe procedures and methods for verifying AI / ML functions or models to ensure that the base station and terminal have the same understanding (or assumptions) of AI / ML model functions. For example, the procedures and methods for verifying AI / ML functions or models may include, but are not limited to, function-based verification methods or model-based verification methods.

[0266] Functionality identification based on a feature-based verification method may mean that the base station and the UE identify the corresponding AI / ML functionality based on UE capability reports. For example, one or more AI / ML functionality may exist within an AI / ML-enabled feature. An AI / ML-enabled feature may be a sub-use case in which AI / ML is used. For example, an AI / ML-enabled feature may be CSI compression, CSI prediction, beam prediction, positioning prediction, or other parts. In other words, an AI / ML-enabled feature may mean a part in which AI / ML is applied and used.

[0267] Additionally, functionality may be associated with one or more AI / ML models. A functionality may refer to an AI / ML-based operation corresponding to a specific configuration or set of configurations. For example, a functionality may have switching, activation, and deactivation as basic units, but is not limited thereto and may further include other operations. For example, Table 17 may be information indicated by terminal capability signaling for verifying AI / ML functionality or for functionality-based life cycle management (LCM) (terminal-side model). However, this is only one example and may not be limited thereto. Furthermore, functionality may have various different definitions depending on the purpose of use or use case. In the supported functionality, functionality may refer to UE-capability information / parameters. For example, as shown in Table 17 below, functionality may correspond to Rel-19 AI / ML-enabled Features / FGs. However, this is merely an example and may not be limited thereto. In the applicable functionality, the functionality may correspond to a higher-layer configuration for an inference configuration or a set of parameters related to inference (e.g., CSI-ReportConfig for beam management or CSI prediction / compression). Finally, in the activated functionality, the functionality may refer to a specific function that performs the inference procedure (e.g., CSI framework).Therefore, depending on the use case in which the AI / ML model is installed and utilized to perform the LCM procedure including the inference operation, the required functions can be defined in the specification.

[0268] In Table 17, FG (feature group) may be functions configured for a UE by a base station based on a report on the UE's capabilities. The UE capability reporting procedure may be initiated in advance by signaling (e.g., UECapabilityEnquiry) directed by the base station. Subsequently, the base station and UE may exchange signaling for RRC reconfiguration and applicable functionality reporting. Additionally, if necessary, the RRC reconfiguration procedure may be performed by the base station. LCM procedures, such as model activation / deactivation, inference, and monitoring, may be performed based on these procedures. For example, the above-described procedure may be required when the AI / ML model is located on the UE side. However, when the AI / ML model is located at the base station, some of the above-described signaling may be omitted depending on the base station implementation. The base station may configure a feature / feature group for the UE that enables AI / ML functions through upper layer signaling. Additionally, the base station can instruct the terminal to activate the corresponding feature / feature group via upper-layer signaling. For example, from an LCM perspective, functionality-based LCM operation may be performed based on the configuration of the aforementioned AI / ML-enabled feature / feature group, but may not be limited thereto.

[0269] [Table 17]

[0270]

[0271]

[0272] FIG. 11 is a diagram illustrating various types of functions applicable to the present disclosure. Referring to FIG. 11 and Table 18 below, identified functionalities may refer to functions that are determined based on terminal capability information among functions that the terminal can support. In other words, they may refer to functions that the terminal can support and may be similar to the supported functionality. Here, configured functionalities may refer to functions that the network configures for actual use in the terminal among functions identified from the terminal. In addition, applicable functionalities may refer to functions that the terminal is in a state of being applicable to among the identified functions.

[0273] Additionally, an activated functionality may refer to a function that is applicable to a terminal and is actually activated and can currently be used in the terminal among the functions configured in the terminal. For example, referring to FIG. 11, functions may be configured in the terminal by the base station among functions identified based on the terminal's capabilities. Here, applicable functions may exist among the functions configured in the terminal, and actual functions may be activated and used based on whether the applicable functions are activated. (Alt 1, 1110) Alternatively, functions applicable to the terminal may be considered among the functions identified based on the terminal's capabilities. Here, functions applicable to the terminal may be configured in the terminal by the base station among the functions applicable to the terminal. The terminal may activate and use actual functions based on whether the configured functions are activated. (Alt 2, 1120)

[0274] Alternatively, functions applicable to the terminal may be considered among the functions identified based on terminal capabilities. Here, the terminal can activate and use the actual functions based on whether the applicable functions are activated. (Alt 3, 1130)

[0275] Alternatively, among the functions identified based on terminal capabilities, functions applicable to the terminal and functions configured within the terminal can be considered. Here, the terminal can activate and use the actual functions based on whether the functions applicable to the terminal and the configurations are activated. (Alt 4, 1140)

[0276] [Table 18]

[0277]

[0278]

[0279] Next, model identification can be considered. The model can be related to the terminal capabilities of a single AI / ML-enabled feature / feature group (FG). Furthermore, the model may consider specific settings / conditions related to additional conditions mutually confirmed by the base station and the terminal (e.g., scenarios, sites, and datasets), but may not be limited thereto.

[0280] For example, feature-based LCM may differ from model-based LCM. The model-based LCM may be a model ID-based LCM, which uses a model ID to enable the base station and the terminal to have the same understanding (or assumption) about the mutual use of AI / ML models. Here, the base station and the terminal can distinguish the AI / ML model based on the model ID. The model ID of the AI / ML model may be a logical ID, and multiple physical AI / ML models (physical model IDs) may be associated with and utilized under a single logical model ID, and are not limited to a specific form. For example, the model verification operation may be defined as the three types shown in Table 19 below. Specifically, type A may be a type in which the model is verified without any over-the-air signaling between the base station and the terminal. The model ID may be assigned through offline work and training. Once development or training is completed, a model corresponding to type A may be independently or separately verified between the base station and the terminal through offline work.

[0281] Type B1 may be a type in which the model is verified based on signaling between the base station and the terminal, and the model verification operation may be initiated by the terminal. For example, the terminal-side model may perform at least one of updating, switching, and other operations.

[0282] Type B2 is a model in which the model is verified based on signaling between the base station and the terminal, and the model verification operation can be initiated by the base station. For example, the base station can apply a new model or initialize the model through signaling to the terminal, but this may not be limited to this.

[0283] [Table 19]

[0284]

[0285]

[0286] That is, both feature-based LCM and model-based LCM can share the concept of features in terms of functionality / model identification. For example, terminal features provided by terminal capability information signaling can be applied to both feature-based LCM and model-based LCM. Here, model-based LCM can be viewed as a superset of the two LCMs, and functionality-based LCM may be a special case of model ID-based LCM that uses a fixed model ID or a dummy model ID in model ID-based LCM, but is not limited thereto.

[0287] For example, feature-based LCM can manage AI / ML operations at the feature level, while model-based LCM can manage AI / ML models at the model level with greater granularity than feature-based LCM. Furthermore, model-based LCM can dynamically and adaptively utilize models based on additional conditions (e.g., scenario / configuration / site / dataset).

[0288] FIG. 12 is a diagram showing the relationship between features, functions, and models applicable to the present disclosure.

[0289] Referring to Fig. 12, when a function / model is confirmed, at least one of activation, deactivation, switching, fallback, and monitoring of the function / model can be performed within the LCM procedure, as described above. This operation can be function-based confirmation or model-based confirmation, as described above. Fig. 12 can show the interrelationships between features (1211, 1212), functions (1221, 1222, 1223, 1224), and models (1231, 1232). However, Fig. 12 is only an example and may not be limited thereto. Specifically, features (feature 1, feature 2, 1211, 1212) applied between a base station and a terminal can be determined. For example, a feature may mean that an AI / ML function is applied, such as at least one of CSI compression, CSI prediction, beam prediction, and positioning prediction. However, this is just one example and may not be limited thereto. In other words, a feature may refer to a part to which an AI / ML function can be applied. Functionality (1221, 1222, 1223, 1224) may exist within the feature according to various settings and terminal capabilities. For example, a function may be a unit in which an AI / ML function is performed, but may not be limited thereto. In addition, models (1231, 1232) may be implemented by considering more detailed management depending on at least one of the channel environment, scenario, and performance capability of the base station terminal in which the function is used. The relationship between the above-described functions and models may be shared through signaling or offline learning between the base station and the terminal, but is not limited to a specific form.

[0290] For example, the term "base station" described below refers to a network node in general and may not be limited to actual base station equipment. Therefore, the term "base station" generally refers to network equipment and may include OAM, AI / ML-related servers, and other general terms used by network operators or related equipment companies for network operation. Similarly, the term "terminal" is not limited to only terminal equipment but can be expanded to include equipment related to the terminal side (e.g., terminal equipment company servers or representative terminal equipment, etc.) and is not limited to a specific form. In particular, with respect to AI / ML operations, the terms "base station" and "terminal" may each refer to various types of equipment or entities that take the corresponding operation into account and are not limited to a specific form. However, for the convenience of explanation, the terms "base station" and "terminal" are referred to below, but may not be applied in a limited manner.

[0291]

[0292] Below, a CSI prediction method based on AI / ML algorithms is described. As an example, the AI / ML algorithm-based CSI prediction method may be a method for predicting CSI values ​​in the time domain using a terminal-side model. However, this is not limited to this method, and it may be a method for predicting CSI values ​​in at least one of the spatial domain, frequency domain, delay domain, and other domains.

[0293] An AI / ML CSI prediction method may be applicable to mobile communication wireless systems (e.g., LTE / NR / 6G) and may be a solution therefor. Here, CSI prediction may operate based on a one-side model, unlike CSI compression. That is, an AI / ML model for CSI prediction may be located in one transmission node (NW or UE) and may generate an output value through an inference operation in one transmission node. The generated output value may be utilized as a CSI prediction value. The following description is based on a case where an AI / ML model for CSI prediction is located in one transmission node and operates using a one-side model, but the present invention may not be limited thereto. In addition, a transmission node mentioned herein includes a server or specific nodes located on the NW or UE side.

[0294] For example, the data required for AI / ML model training can be generated by the terminal. Specifically, the terminal can measure channel information (CSI) based on a reference signal transmitted from the base station, and the measured channel information can be utilized by the terminal for AI / ML model training for CSI prediction. In addition, the terminal can define a specific associated ID value to ensure that the contents of auxiliary information (e.g., base station antenna configuration, cell environment, scenario, etc.) to support model training / inference from another side (e.g., base station side) are provided. The associated ID can represent consistency between training and inference for additional auxiliary information on the NW side. From the terminal's perspective, all downlink beams in the set or list of beams associated with the specific associated ID described above can be understood (or assumed) to have similar transmission characteristics, and based on this terminal understanding (or assumption), the terminal can effectively operate the AI / ML model. The terminal can be provided with at least one of the upper layer configurations for the CSI prediction model and can perform training on the AI / ML model using the same. The terminal may utilize the above-described information for at least one of the following purposes: AI / ML model learning, inference, monitoring, and other operations.

[0295] That is, the AI / ML model for CSI prediction can be located and trained on the terminal side, and the inference operation can be performed on the terminal side. Here, the input data value for the AI / ML model for CSI prediction can be utilized internally in the terminal. In addition, the terminal can transmit the input data value for the AI / ML model for CSI prediction to the base station, and the input data value transmitted to the base station can be utilized by the base station as data necessary for calculating evaluation indices related to performance evaluation. As another example, the terminal can report the performance evaluation indices calculated by the terminal to the base station, and the form may not be limited to a specific form.

[0296] FIG. 13 is a diagram illustrating an inference operation for CSI prediction applicable to the present disclosure.

[0297] Referring to FIG. 13, the terminal can train an AI / ML model (1310) for CSI prediction by utilizing historical CSI measurement information acquired over a specific time window to generate input values ​​for CSI prediction. That is, the terminal can train an AI / ML model (1310) for CSI prediction by utilizing previous CSI measurement information based on a time window. In addition, other data may be utilized for training the AI / ML model (1310) for CSI prediction, and may not be limited to a specific format.

[0298] Historical CSI information may be provided as an input value to an AI / ML model (1310) for CSI prediction, for which learning has been completed. The AI / ML model (1310) for CSI prediction may utilize at least one of the correlation and the degree of correlation between previous CSI measurement information, and may generate CSI prediction information as an output value through an inference operation based on the correlation. Here, the AI / ML model (1310) for CSI prediction may be located on the terminal side. That is, the inference operation for CSI prediction may be performed on the terminal side. However, the present invention is not limited thereto, and the CSI prediction may be located and performed on the base station side depending on the application and installation environment and method. For the convenience of explanation, the following description will be based on a case where the AI / ML model (1310) for CSI prediction is located on the terminal side and the inference operation is performed.

[0299] For example, CSI compression can be inferred based on a two-side AI / ML model by an AI / ML-based model located at both the base station and the terminal. On the other hand, an AI / ML model for CSI prediction can be a one-side AI / ML model located at only one transmission node (NW or UE). When an AI / ML model for CSI prediction is located at the terminal, the terminal can generate CSI prediction information as an output value through an inference operation of the AI / ML model for CSI prediction located at the terminal. In addition, the terminal can report the generated predicted CSI information to the base station. The base station can perform PDSCH (physical downlink shared channel) scheduling for data transmission based on the same understanding (or assumption) of the predicted CSI information received from the terminal, and the base station and the terminal can communicate based on this. As another example, if the AI / ML model for CSI prediction is located at the base station, the terminal can measure target CSI information (e.g., channel response or eigenvector with high accuracy) with high accuracy as ground-truth information (or label information) and report it to the base station. That is, the terminal can report actual measurement information to the base station on resources set to have high reliability. The base station can perform learning, inference, and monitoring for the AI / ML model for CSI prediction using the received target CSI information. The base station can perform an inference operation of the CSI prediction model based on previous CSI information received from the terminal to generate an output value, which can be used for future data transmission scheduling.

[0300]

[0301] Below, we describe a CSI fallback technique for using AI / ML-based CSI prediction techniques.

[0302] When training an AI / ML model for CSI prediction, data information for AI / ML training can be generated by the terminal by default. When the AI / ML model for CSI prediction is located on the terminal side as a one-side model, the terminal can utilize CSI information acquired within the terminal, configuration information provided by the base station, and other model input information held by the terminal itself as input data values. For example, when the base station monitors the related performance of an AI / ML model for CSI prediction, performance indicators and related data information for evaluating the AI / ML-based CSI prediction model can be generated and utilized by both the base station and the terminal.

[0303] Specifically, CSI feedback may be an operation of performing CSI measurement based on a reference signal for CSI measurement and then reporting CSI information. The base station may transmit data (DL transmission) to the terminal based on the CSI feedback from the terminal. Here, considering the timeline of the CSI feedback, the times (slots) of the CSI measurement, CSI feedback, and CSI feedback-based data transmission may each be different. For example, when the base station performs data transmission to the terminal, the case where the validity period of the CSI feedback information used by the base station for data transmission has expired may be considered. In other words, the case where the channel environment changes over time and the measured CSI information differs from the actual data transmission environment may be considered. In the above case, when the base station transmits data to the terminal, the DL scheduling / precoding may be determined differently from the actual channel environment and may be applied improperly, which may cause the terminal to not properly receive the data, resulting in degradation of system performance.

[0304] Considering the above, a CSI prediction technique based on AI / ML can be applied. Here, the CSI compression method or the CSI prediction method can be applied based on the same characteristics. Specifically, both CSI compression and CSI prediction can use channel correlation in the time domain. The purpose of CSI compression may be to reduce feedback overhead in the time domain, while the purpose of CSI prediction may be to predict future CSI information with high accuracy. In other words, CSI compression and CSI prediction may have different purposes in the time domain, and the procedures and methods required for AI / ML-based CSI prediction are described below. In particular, when using an AI / ML model for CSI prediction, a fallback operation may be required based on the AI / ML model monitoring operation for CSI prediction, and this is described below.

[0305] For example, an AI / ML model for CSI prediction may be located on the terminal side, and a performance monitoring operation for a function-based LCM operation may be performed, and a method for performing a fallback operation based on this may be required. Here, if the AI / ML model for CSI prediction located on the terminal side performs an inference operation for CSI prediction, it may be necessary to continuously evaluate the reliability of the result value of the AI / ML model for CSI prediction. That is, if the reliability is low based on the performance evaluation of the AI / ML model for CSI prediction, using the existing CSI measurement operation rather than the operation of reporting the CSI prediction result value based on the AI / ML model can improve system stability, and a fallback operation that takes this into consideration may be required. For example, the following methods may be applied to the monitoring operation and fallback operation for the CSI prediction model, but may not be limited thereto.

[0306] Furthermore, while the above and below examples illustrate the case where AI / ML-based CSI reporting falls back to the existing CSI reporting method, the terminology may not be limited to this context. In other words, the same applies to operations where AI / ML-based CSI reporting falls back to the existing CSI reporting method. For convenience of explanation, the term "fallback" is used here as a context, but "fallback" can be used interchangeably with "transition" or "switching" and is not limited to a specific form.

[0307]

[0308] FIG. 14 is a diagram illustrating a fallback operation for an AI / ML-based CSI prediction model applicable to the present disclosure. Referring to FIG. 14, a terminal (1420) may transmit terminal capability information to a base station (1410), and the base station (1410) may transmit CSI-RS (channel state information) to the terminal (1420) as a reference signal for CSI measurement. In addition, the base station (1410) may transmit assistance information to the terminal (1420). Here, an AI / ML model for CSI prediction may be located on the terminal side as a one-side model. The terminal (1420) may perform measurements based on information acquired from the base station (1410) and, based on the measurements, may perform training on the AI / ML model for CSI prediction. Thereafter, the terminal (1420) may report the training and testing results to the base station (1410). Thereafter, the base station (1410) can transmit triggering / signaling to activate model inference to the terminal (1420). The terminal (1420) can perform an inference operation on the AI / ML model for CSI prediction based on the triggering / signaling for the received model inference, generate a CSI prediction value as an output value, and report the result to the base station (1410).

[0309] In addition, the base station (1410) and the terminal (1420) can monitor the AI / ML model for CSI prediction. Here, the terminal (1420) can calculate a performance metric for the AI / ML model for CSI prediction. The performance metric can be derived by utilizing at least one of a CSI prediction value and a corresponding ground truth CSI (i.e., actual data) value. The terminal (1420) can report a performance monitoring result derived based on the performance metric to the base station (or network, 1410), and the base station can determine whether to perform functionality fallback based on the performance monitoring result.

[0310] For example, the performance monitoring result may correspond to a value obtained by comparing the ground-truth CSI value with the output value of the CSI prediction model, or a performance metric value calculated using the values. That is, the CSI value actually derived by the terminal (1420) and the predicted CSI value are compared to calculate a performance evaluation index, and the performance monitoring result may be determined using the performance evaluation index. Alternatively, the performance monitoring result may be derived by utilizing other information in addition to the performance evaluation index derived from the actually derived CSI value and the predicted CSI value, and is not limited to a specific form.

[0311] For example, the terminal (1420) can calculate the monitoring accuracy (e.g., percentage(%)) / confidence level (e.g., level 0~3 (2bits))) and confidence information on the monitoring accuracy based on the loss function result value described above. The terminal (1420) can report the calculated monitoring accuracy information and the confidence information on the monitoring accuracy to the base station (1410) as a performance monitoring result. Here, the monitoring accuracy can be calculated as a specific value, such as a percentage value, and reported to the base station (1410). That is, the percentage value for accuracy can be derived based on the predicted CSI and the actually measured CSI (e.g., Target CSI). In addition, the confidence level of the monitoring accuracy can be information derived based on a certain section. The terminal (1420) can report information on the intervals containing the confidence level of the accuracy based on the calculated monitoring accuracy to the base station (1410), and the base station (1410) can thereby check the monitoring accuracy and the confidence level information for the monitoring accuracy. For example, considering the overhead of reporting the performance monitoring results, the confidence intervals / levels of the monitoring accuracy can be divided into a preset or predetermined number. If the number of confidence intervals of the monitoring accuracy is large, the base station (1410) can specifically recognize the confidence level of the model currently being monitored, which can be helpful in determining whether to perform a fallback. However, this may increase the overhead. On the other hand, if the number of confidence intervals of the monitoring accuracy is small, the overhead is reduced, but the base station (1410) may not specifically recognize the confidence level, and may perform an unintended fallback or not perform a fallback in a situation where a fallback is required. The monitoring accuracy and the confidence intervals of the monitoring accuracy can be determined considering the above-described points, but are not limited to a specific embodiment.

[0312] In addition, the terminal (1420) may report the performance monitoring results to the base station (1410) based on a preset cycle. Considering the characteristic that performance monitoring is performed less frequently, the terminal (1420) may periodically report the performance monitoring results to the base station (1410) based on a relatively long cycle. Such performance monitoring result reporting with a long cycle may be performed by the terminal based on upper layer settings provided by the base station in advance. As another example, the terminal (1420) may determine a situation requiring fallback on its own and, if necessary, may report the performance monitoring results to the base station (1410) aperiodically. As an example, the terminal (1420) may report the performance monitoring results to the base station (1410) when a specific condition is met in an event-triggered manner, but may not be limited thereto. Here, the specific condition may be a case where the model monitoring confidence interval / level or the amount of change in the model monitoring performance indicator changes significantly beyond a preset or predetermined threshold value. The terminal may report the performance monitoring results to the base station in the above-described cases. For example, the above-described operation can be determined to trigger an event not only when the model performance monitoring results change in a negative direction, but also when the monitoring results change in a positive direction. Of course, an event can also be triggered based solely on the monitoring performance indicator / confidence level itself. The performance monitoring result reporting method using the event triggering method proposed above can also be used as a criterion for determining whether to switch to the existing CSI reporting method or the AI / ML-based CSI reporting method, and can be applied to both within the method proposed in the present invention.

[0313] When the terminal (1420) periodically reports model performance monitoring results to the base station (1410), the terminal (1420) may receive CSI resources for monitoring result reporting and upper layer settings related to the reporting from the base station (1410). For example, the terminal (1420) may receive upper layer settings for performance monitoring results from the base station (1410), but is not limited thereto.

[0314] Additionally, a case may be considered in which the terminal (1420) reports information on whether to fallback to the base station (1410) at a time when necessary. The base station (1410) may provide the terminal (1420) with upper layer settings (e.g., CSI resources and reports) for the report in advance. If the terminal (1420) decides to report the performance monitoring results to the base station (1410) on its own, the terminal (1420) may report the performance monitoring results to the base station (1410) based on the upper layer settings previously provided by the base station (1410).

[0315] That is, the terminal (1420) can provide performance monitoring result information to the base station (1410), and the base station (1410) can determine whether to execute fallback based on the performance monitoring result information. Here, the base station (1410) can consider a threshold value to determine whether to execute fallback. For example, the threshold value can be a threshold value related to a performance evaluation index. The threshold value can be determined in advance based on an AI / ML model for CSI prediction or can be determined by the base station (1410), and is not limited to the embodiment. As a specific example, the performance evaluation index can be derived as a value compared by utilizing a loss function between a ground-true CSI value and a CSI prediction model output value, and the base station (1410) can recognize the performance evaluation index through the performance monitoring result. The base station (1410) can check the performance evaluation index based on the performance monitoring result, and compare the performance evaluation index with the threshold value to determine whether to perform a functionality fallback operation.

[0316] As another example, the threshold value may be set by further considering at least one of the information included in Table 20 below. Referring to Table 20, NMSE (normalized mean squared error) may be an indicator that measures the difference between data as a normalized mean squared error, and may quantitatively represent the error between original data and predicted data. SGCS (squared generalized cosine similarity) may be a squared generalized cosine similarity, and may be an indicator that measures the similarity between two vectors. Here, NMSE and SGCS may be further considered when determining the threshold value, but are not limited thereto. As another example, the threshold value may be determined by further utilizing RSRP (reference signal received power) or information on the best K beam IDs. As another example, the threshold value may be determined by further utilizing at least one of the transmission rate, the assumed BLER (block error rate), BLER, and NACK / ACK ratio information. As another example, the threshold may further consider at least one of the function ID, the model ID, and the associated ID, but may not be limited to that embodiment.

[0317] [Table 20]

[0318]

[0319]

[0320] Information regarding at least one of the functionality ID, model ID, and associated ID may not directly indicate a performance indicator, but may be utilized in combination with other performance indicators. That is, the base station (1410) can determine whether to perform a fallback by further considering the performance evaluation indicators confirmed based on the performance monitoring results and the aforementioned information, thereby improving the reliability of the fallback operation. As a specific example, if the ID information corresponds to at least one of the system settings, network scenario environment, and cell structure, the reliability of the CSI prediction result value may vary depending on the environment or settings. Therefore, if information regarding the environment or settings is provided in advance, the threshold setting can be configured in relation to the environment or settings, thereby improving the reliability of the final fallback decision.

[0321] Based on the above, if the base station (1410) determines to perform a function fallback, the base station (1410) needs to indicate that the terminal (1420) is switching to a different method of CSI reporting than the method performed by the base station (1410). For example, signaling may be required when falling back from an AI / ML-based CSI reporting method to a conventional CSI reporting method. As another example, the terminal (1420) may consider a case where it operates in the conventional CSI reporting method due to a temporary performance degradation of the AI / ML model even though the function corresponding to AI / ML-based CSI prediction has been set by the base station (1410).

[0322] As a specific example, when the terminal (1420) falls back from the AI / ML-based CSI reporting method to the existing CSI reporting method, the terminal (1420) can measure CSI based on the existing CSI reporting method and report the measured CSI to the base station (1410). Thereafter, the base station (1410) can switch back to the AI / ML-based CSI reporting method by utilizing at least one of the model performance evaluation methods proposed in the present invention, and at least one or more threshold values ​​associated with the AI / ML model performance evaluation index / performance monitoring result or the amount of change from the previous result. Here, the one or more threshold values ​​can be transmitted to the terminal (1420) through upper layer signaling. Specifically, the base station (1410) can perform continuous AI / ML model performance evaluation as described above and compare the values ​​and amounts of change corresponding to the evaluation index with the set threshold values. Here, if the AI / ML model evaluation performance result / change value has a reliability higher than a threshold value, if the configured timer / time interval expires, or if a signal for model switching by the base station is transmitted to the terminal, the existing CSI-based reporting may be converted (or switched) back to AI / ML-based CSI reporting in at least one of the following cases. Of course, the method is not limited to the case of switching to AI / ML-based CSI reporting, and may also be applied when switching from the AI / ML-based CSI reporting method to the existing CSI reporting method as described above. The methods for fallback / switching described here can be viewed as implicit signaling or condition-based triggering rather than explicit signaling, and in the latter case, explicit signaling for model switching by the base station or terminal may not be provided.The evaluation of whether to fallback / switch can also be performed based on one of the AI / ML model evaluation methods performed on the base station or terminal side, similar to the methods considered in the proposed methods below. The final decision on whether to fallback / switch can be made primarily on the network side, but the terminal can also make the final decision on the switch based on the switching triggering conditions set in advance by the base station to the terminal. Of course, according to the proposed method, the base station and the terminal can exchange target CSI information, performance evaluation indices, and thresholds required for the model evaluation indices through signaling. The calculation of the performance evaluation indices can be performed according to the method applied on the base station or terminal side. When switching from AI / ML-based CSI reporting to the existing CSI reporting method or from the existing CSI reporting method back to AI / ML-based CSI reporting, the corresponding action can be determined based on one or more evaluation results. For example, to prevent unnecessary frequent fallbacks or switches from occurring continuously, an additional threshold can be set when based on multiple evaluation results, and the fallback action can be performed when an evaluation result exceeds the threshold. For example, if the number of times a model has low confidence due to its performance evaluation metrics / outcome values ​​exceeds a threshold, a fallback / switching operation may be performed for that model. Conversely, no fallback / switching may be performed for existing AI / ML model-based CSI reporting. Or, conversely, if the number of times a model has high confidence due to its performance evaluation metrics / outcome values ​​exceeds a threshold, the model may be maintained (if currently in use) or switched to an AI / ML-based model (e.g., if the existing CSI reporting method is in use).

[0323] As another example, the base station (1410) may transmit AI / ML model monitoring instruction information to the terminal (1420) while the existing CSI reporting is being performed. The terminal (1420) may obtain the model monitoring instruction information from the base station (1410), perform an inference operation on the AI / ML model for CSI prediction, and transmit the performance monitoring result derived based on the performance evaluation index based on the model monitoring instruction information to the base station (1410). The base station (1410) may compare the performance monitoring result with a threshold value to determine whether to switch back to AI / ML-based CSI reporting. For example, the threshold value may be a value different from the threshold value described above as a threshold value for determining whether the terminal (1420) reverts to AI / ML-based CSI reporting. In addition, a case where the threshold value is the same as the threshold value described above may also be considered, and is not limited to a specific form. The base station (1410) may switch back from the existing CSI reporting to AI / ML-based CSI reporting based on the above-described information. Here, the terminal (1420) may also need to be instructed to switch to AI / ML-based CSI reporting operation to perform AI / ML-based CSI reporting again. Additionally, if such fallback or transition is required on the terminal side, the terminal can report this information to the base station, and the base station can then signal whether to initiate fallback or transition using the proposed method.

[0324] The model switching indication method described below can be applied both to switching from AI / ML-based CSI reporting to existing CSI reporting (fallback) and to switching from existing CSI reporting to AI / ML-based CSI reporting (fallback).

[0325] As a specific example, the switching instruction can be implicitly indicated by providing the terminal with existing CSI reporting and resource configurations via upper layer configurations. That is, if the terminal (1420) receives existing CSI reporting and resource configurations via upper layer configurations, it can recognize that it is falling back to existing CSI reporting. Conversely, if the terminal (1420) receives AI / ML-based CSI reporting and resource configurations via upper layer configurations while performing existing CSI reporting, it can recognize that it is falling back to AI / ML-based CSI reporting.

[0326] As another example, if the terminal (1420) has already been provided with existing CSI reporting and resource settings from the base station (1410), the terminal (1420) may not be implicitly instructed to fallback. In the above-described case, the terminal (1420) may receive explicit signaling from the base station (1410) via at least one of radio resource control (RRC) and medium access control element (MAC CE) that instructs the terminal (1420) to perform CSI reporting using the existing CSI reporting method.

[0327] Additionally, if the terminal (1420) has already been provided with AI / ML-based CSI reporting and resource settings from the base station (1410), the terminal (1420) may not be implicitly instructed to fallback. In the above-described case, the terminal (1420) may receive explicit signaling from the base station (1410) via at least one of RRC and MAC CE instructing the terminal (1420) to perform CSI reporting using the AI / ML-based CSI reporting method.

[0328] As another example, the terminal may be instructed to fallback through L1 signaling corresponding to dynamic signaling (e.g., recycling an existing field (e.g., CSI request) in the DCI (downlink control information) format or an additional DCI field). The DCI format is a format distinct from the existing DCI format and may scramble the CRC (cyclic redundancy check) using a new RNTI (radio network temporary identifier) ​​value (e.g., AI / ML-CSI-RNTI) for AI / ML-based CSI reporting. For example, the DCI format defined by the above-described RNTI value may indicate whether to fallback by recycling an existing field or defining a new field, and is not limited to a specific form.

[0329] FIG. 15 is a diagram illustrating a fallback operation for an AI / ML-based CSI prediction model applicable to the present disclosure. A terminal (1520) transmits terminal capability information to a base station (1510), and the base station (1510) may transmit CSI-RS (channel state information) to the terminal (1520) as a reference signal for CSI measurement. In addition, the base station (1510) may transmit assistance information to the terminal (1520). Here, an AI / ML model for CSI prediction may be located on the terminal side as a one-side model. The terminal (1520) may perform measurements based on information acquired from the base station (1510) and, based on the measurements, may train the AI / ML model for CSI prediction. Thereafter, the terminal (1520) may report the training and testing results to the base station (1510). Thereafter, the base station (1510) can transmit triggering / signaling to activate model inference to the terminal (1520). The terminal (1520) can perform an inference operation on the AI / ML model for CSI prediction based on the triggering / signaling for the received model inference, generate a CSI prediction value as an output value, and report the result to the base station (1510).

[0330] In addition, the terminal (1520) can receive a threshold value provided based on the performance monitoring result from the base station (1510). For example, the threshold value may be a threshold value related to a performance evaluation index. The threshold value may be transmitted from the base station (1510) to the terminal (1520) via upper layer signaling. In another example, the threshold value may be a value determined in advance based on an AI / ML model for CSI prediction, but is not limited to the embodiment. In addition, as proposed above, one or more threshold values ​​may be provided to the terminal for various purposes. The terminal (1520) can recognize the threshold value before performing the inference operation of the AI / ML model for CSI prediction, and the proposed one or more threshold values ​​may be utilized in the process of monitoring the AI / ML model for CSI prediction.

[0331] When the base station (1510) and the terminal (1520) monitor an AI / ML model for CSI prediction, the terminal (1520) can calculate a performance evaluation index for the AI / ML model for CSI prediction. The performance evaluation index can be derived by utilizing at least one of the ground truth CSI (i.e., actual data) values ​​corresponding to the CSI prediction value. Thereafter, the terminal (1520) can generate fallback identification information using the performance evaluation index and the threshold value and report a fallback recommendation reporting to the base station (1510). Here, the fallback recommendation reporting can be information indicating whether the base station (1510) will perform a fallback operation.

[0332] Here, the threshold associated with the model performance evaluation index may be set by further considering at least one of the information included in Table 20 described above. Referring to Table 20, NMSE may be an index that measures the difference between data with the normalized mean squared error, and may quantitatively represent the error between the original data and the predicted data. SGCS may be a squared generalized cosine similarity, and may be an index that measures the similarity between two vectors. Here, when determining the threshold, NMSE and SGCS may be further considered, but are not limited thereto. As another example, the threshold may be determined by further utilizing RSRP or the best K beam ID information. As another example, the threshold may be determined by further utilizing at least one of the transmission rate, the assumed BLER (block error rate), BLER, and NACK / ACK ratio information. As another example, the threshold may further consider at least one of the function ID, the model ID, and the related ID, but may not be limited to the present embodiment.

[0333] As a specific example, the performance evaluation index can be derived as a value that is compared by utilizing a loss function between the ground-true CSI value and the CSI prediction model output value. Here, whether to perform a functionality fallback operation can be determined based on comparison information about the performance evaluation index and a threshold value, and fallback identification information can be generated based on this. Specifically, if the performance evaluation index is greater than the threshold value, the terminal (1520) can transmit a fallback recommendation report to the base station (1510) based on the fallback identification information, and the fallback recommendation report can include information indicating a fallback operation. On the other hand, if the performance evaluation index is less than the threshold value, the terminal (1520) can transmit a fallback recommendation report to the base station (1510) based on the fallback identification information. Here, the fallback recommendation report can indicate that a fallback operation is not required. For example, it may be possible to instruct a fallback action when a performance evaluation metric is less than a threshold, and not to instruct a fallback action when a performance evaluation metric is greater than a threshold, and it is not limited to a specific form.

[0334] The fallback recommendation report may include functionality fallback indicator (FFI) information. That is, the terminal (1520) can instruct the base station (1510) whether to execute fallback using the FFI information. Here, the base station (1510) can decide whether to actually reflect the FFI information or perform another decision. That is, even if the terminal (1520) indicates fallback using the FFI information, the base station (1510) can determine whether to actually perform fallback by reflecting other information, and an operation without performing fallback may also be possible. For example, the FFI information is similar to the existing PMI (precoding matrix indicator) and is merely a value reported by the terminal. Whether or not the base station uses the reported information as is may depend on the base station implementation / decision, but may not be limited thereto.

[0335] For example, as in the method already proposed above, the terminal (1520) may report the fallback recommendation report to the base station (1510) based on a preset cycle. Considering the characteristic that the fallback recommendation report is performed less frequently, the terminal (1520) may periodically report the fallback recommendation report to the base station (1510) based on a relatively long cycle. Such a fallback recommendation report with a long cycle may be performed by the terminal based on a higher layer setting provided by the base station in advance. As another example, the terminal (1520) may determine a situation requiring fallback on its own and, if necessary, may report the fallback recommendation report to the base station (1510) aperiodically. For example, the terminal (1520) may compare a performance evaluation index with a threshold value in an event-triggered manner and transmit the fallback recommendation report to the base station (1510), but may not be limited thereto.

[0336] When a terminal (1520) periodically reports a fallback recommendation report to a base station (1510), the terminal (1520) may receive CSI resources for the fallback recommendation report and upper layer settings related to the report from the base station (1510). As an example, the terminal (1520) may receive upper layer settings for the fallback recommendation report from the base station (1510), but is not limited thereto.

[0337] In addition, a case may be considered in which the terminal (1520) reports information on whether to perform fallback to the base station (1510) at a time when it needs to do so. The base station (1510) may provide the terminal (1520) with upper layer settings (e.g., CSI resources and reports) for the report in advance. If the terminal (1520) decides to report a fallback recommendation report to the base station (1510), the terminal (1520) may perform the report to the base station (1510) based on the upper layer settings previously provided by the base station (1510). That is, the terminal (1520) may provide a fallback recommendation report to the base station (1510), and the base station (1510) may execute a fallback operation.

[0338] In addition, for example, information about at least one of the functionality ID, model ID, and associated ID may not directly indicate a performance indicator, but may be utilized in combination with the ID information and other performance indicators. That is, the base station (1510) may further consider the above-described information to determine whether to perform a fallback, thereby improving reliability. As a specific example, if the ID information corresponds to at least one of the system settings, network scenario environment, and cell structure, the reliability of the CSI prediction result value may vary depending on the environment or settings. Therefore, if information about the environment or settings is provided in advance, the threshold value setting may be configured as a value related to the environment or settings, thereby improving the reliability of the final fallback decision.

[0339] Based on the above, if the base station (1510) determines a function fallback, the base station (1510) needs to indicate that the terminal (1520) is switching to a different method of CSI reporting. For example, signaling may be required when falling back from an AI / ML-based CSI reporting method to a conventional CSI reporting method. Furthermore, for example, the terminal (1520) may operate with the conventional CSI reporting method due to a temporary performance degradation of the AI / ML model even if the function corresponding to AI / ML-based CSI prediction is configured by the base station (1510).

[0340] As a specific example, when the terminal (1520) falls back from the AI / ML-based CSI reporting method to the existing CSI reporting method, the terminal can measure CSI based on the existing CSI reporting method and report the measured CSI to the base station (1510).

[0341] Thereafter, the base station (1510) can switch back to the AI / ML-based CSI reporting method by utilizing at least one threshold value associated with the AI / ML model performance evaluation index / performance monitoring result or the amount of change from the previous result based on at least one of the model performance evaluation methods proposed in the present invention. Here, the one or more threshold values ​​can be transmitted to the terminal (1520) through upper layer signaling. Specifically, the base station (1510) can perform continuous AI / ML model performance evaluation as described above and compare the value and amount of change corresponding to the evaluation index with the set threshold value. Here, in at least one of the cases where the corresponding AI / ML model evaluation performance result / amount of change value has a reliability higher than the threshold value, when the set timer / time interval expires, or when a signal for model switching by the base station is transmitted to the terminal, the existing CSI-based reporting can be switched back to the AI / ML-based CSI reporting. Of course, this method is not limited to the application only when switching to AI / ML-based CSI reporting, and can also be applied when switching from AI / ML-based CSI reporting to the existing CSI reporting method as described above. The fallback / switching methods described above can be viewed as implicit signaling or condition-based triggering rather than explicit signaling, and in the latter case, explicit signaling for model switching by the base station or terminal may not be provided. The evaluation of whether to perform the fallback / switching can also be performed based on one of the AI / ML model evaluation methods performed on the base station or terminal side, similar to the methods considered in the methods proposed below. The final decision on whether to perform the fallback / switching can be primarily performed by the network side, but the terminal can make the final decision on the switch based on the triggering conditions for switching set by the base station to the terminal in advance.Of course, according to the proposed method, the base station and the terminal can exchange target CSI information, performance evaluation indices, and thresholds required for model evaluation indices through signaling. The calculation of performance evaluation indices can be performed according to the method applied on the base station or terminal side. When switching from AI / ML-based CSI reporting to the existing CSI reporting method or switching back from the existing CSI reporting method to the AI / ML-based CSI reporting method, the corresponding operation can be determined based on one or more evaluation results. For example, in order to prevent unnecessary frequent fallbacks or switching operations from continuously occurring, an additional threshold value can be set when based on multiple evaluation results, and the fallback operation can be performed when an evaluation result exceeds the threshold value. For example, if the number of times that the performance evaluation indices / results have low model reliability is greater than the threshold value, the fallback / switching operation for the corresponding model can be performed. On the other hand, fallback / switching is not performed for the existing AI / ML model-based CSI reporting. Or, conversely, if the number of times the performance evaluation metrics / outcome values ​​have high model confidence is greater than a threshold, the model can be maintained (if currently in use) or switched to an AI / ML-based model (e.g., if the existing CSI reporting method is in use).

[0342] As another example, the base station (1510) may transmit AI / ML model monitoring instruction information to the terminal (1520) while the existing CSI reporting is being performed. The terminal (1520) may obtain the model monitoring instruction information from the base station (1510), perform an inference operation on the AI / ML model for CSI prediction, and calculate a performance evaluation index based on the obtained information. The terminal (1520) may compare the performance evaluation index with a threshold value to determine whether to switch back to AI / ML-based CSI reporting, and may transmit a fallback recommendation report to the base station (1510) based on the comparison. Here, the fallback recommendation report may indicate switching from the existing CSI reporting to the AI / ML-based CSI reporting. As an example, the threshold value may be a different value from the threshold value described above as a threshold value for determining whether the terminal (1520) reverts to the AI / ML-based CSI reporting. In addition, a case where the threshold value is the same as the threshold value described above may also be considered, and is not limited to a specific form. The base station (1510) may perform an operation to switch back from conventional CSI reporting to AI / ML-based CSI reporting based on the above-described information. Here, the terminal (1520) may also need to be instructed to switch back to AI / ML-based CSI reporting in order to perform AI / ML-based CSI reporting again. Additionally, if such fallback or transition is required on the terminal side, the terminal may report the relevant information to the base station, and the base station may then signal whether to instruct fallback or transition using the proposed method.

[0343] The fallback indication or model switching indication method described below can be applied both when switching from AI / ML-based CSI reporting to conventional CSI reporting and when switching from conventional CSI reporting to AI / ML-based CSI reporting, and is not limited to a specific form.

[0344] As a specific example, the fallback instruction can be implicitly indicated by providing the terminal with existing CSI reporting and resource configurations via higher layer configurations. That is, if the terminal (1520) receives existing CSI reporting and resource configurations via higher layer configurations, it can recognize that it is falling back to the existing CSI reporting. Conversely, if the terminal (1520) receives AI / ML-based CSI reporting and resource configurations via higher layer configurations while performing existing CSI reporting, it can recognize that it is falling back to AI / ML-based CSI reporting.

[0345] As another example, if the terminal (1520) has already been provided with configurations for existing CSI reporting and resources from the base station (1510), the terminal (1520) may not be implicitly instructed to fallback. In the above-described case, the terminal (1520) may be provided with explicit signaling from the base station (1510) via at least one of an RRC and a MAC CE that instructs the terminal (1520) to perform CSI reporting using the existing CSI reporting method. In addition, if the terminal (1520) has already been provided with configurations for AI / ML-based CSI reporting and resources from the base station (1510), the terminal (1520) may not be implicitly instructed to fallback. In the above-described case, the terminal (1520) may be provided with explicit signaling from the base station (1510) via at least one of an RRC and a MAC CE that instructs the terminal (1520) to perform CSI reporting using the AI / ML-based CSI reporting method.

[0346] As another example, the terminal may be instructed to fallback through L1 signaling corresponding to dynamic signaling (e.g., recycling an existing field (e.g., CSI request) in the DCI (downlink control information) format or an additional DCI field). The DCI format is a format distinct from the existing DCI format and may scramble the CRC (cyclic redundancy check) using a new RNTI (radio network temporary identifier) ​​value (e.g., AI / ML-CSI-RNTI) for AI / ML-based CSI reporting. For example, the DCI format defined by the above-described RNTI value may indicate whether to fallback by recycling an existing field or defining a new field, and is not limited to a specific form.

[0347]

[0348] FIG. 16 is a diagram illustrating a fallback operation for an AI / ML model for CSI prediction applicable to the present disclosure.

[0349] Referring to FIG. 16, a terminal (1620) may transmit terminal capability information to a base station (1610), and the base station (1610) may transmit a CSI-RS to the terminal (1620) as a reference signal for CSI measurement. In addition, the base station (1610) may transmit assistance information to the terminal (1620). Here, an AI / ML model for CSI prediction may be located on the terminal side as a one-side model. The terminal (1620) may perform measurements based on information acquired from the base station (1610), and may perform learning on the AI / ML model for CSI prediction based on the measurements. Thereafter, the terminal (1620) may report learning and testing results to the base station (1610). Thereafter, the base station (1610) may transmit triggering / signaling to activate model inference to the terminal (1620). The terminal (1620) can perform an inference operation on an AI / ML model for CSI prediction based on triggering / signaling for the received model inference, generate a CSI prediction value as an output value, and report it to the base station (1610).

[0350] In addition, the base station (1610) and the terminal (1620) can monitor the AI / ML model for CSI prediction. Here, the terminal (1620) can report the predicted CSI information and the corresponding ground-true CSI information to the base station (1610). That is, unlike FIGS. 14 and 15 described above, the terminal (1620) can generate a CSI prediction result (model inference output) and corresponding actual CSI measurement information, and report the same to the base station (1610) as a measurement result (measurement report). Here, the base station can calculate a performance index for evaluating the CSI prediction model. The base station (1610) can determine whether to perform a function fallback operation based on the performance index. The performance evaluation index for the AI / ML model for CSI prediction can be derived by utilizing at least one of a CSI prediction value and a corresponding ground-true CSI (ground truth CSI, i.e., actual data) value. The base station (1610) can determine whether to perform functionality fallback by calculating performance evaluation indicators.

[0351] Here, the base station may consider a threshold value for determining a fallback. For example, the threshold value may be a threshold value related to a performance evaluation index. The base station (1610) may compare the performance evaluation index with the threshold value to determine whether to perform a functionality fallback operation. For example, the threshold value may be set by considering not only the performance evaluation index but also at least one of the information included in Table 20 described above. Referring to Table 20, NMSE may be an index that measures the difference between data as a normalized mean squared error and may quantitatively represent the error between original data and predicted data. SGCS may be a squared generalized cosine similarity and may be an index that measures the similarity between two vectors. Here, NMSE and SGCS may be further considered when determining the threshold value, but are not limited thereto. As another example, the threshold value may be determined by further utilizing RSRP or the best K beam ID information. As another example, the threshold value may be determined by further utilizing at least one of the following information: transmission rate, assumed BLER, BLER, and NACK / ACK ratio. As another example, the threshold value may further consider at least one of the function ID, model ID, and related ID, but may not be limited to the present embodiment.

[0352] In addition, information about at least one of the functionality ID, model ID, and associated ID may not directly indicate a performance indicator, but may be utilized in combination with the ID information and other performance indicators. That is, the base station (1610) can further consider the above-described information to determine whether to perform a fallback, thereby improving reliability. As a specific example, if the ID information corresponds to at least one of the system settings, network scenario environment, and cell structure, the reliability of the CSI prediction result value may vary depending on the environment or settings. Therefore, if information about the environment or settings is provided in advance, the threshold value setting can be configured as an independent value related to the environment or settings, thereby improving the reliability of the final fallback decision.

[0353] Based on the above, if the base station (1610) determines a function fallback, the base station (1610) needs to indicate that the terminal (1620) is switching to a different method of CSI reporting than the one performed by the base station (1610). For example, signaling may be required when falling back from an AI / ML-based CSI reporting method to a conventional CSI reporting method. Furthermore, for example, the terminal (1620) may operate with the conventional CSI reporting method due to a temporary performance degradation of the AI / ML model even if the function corresponding to AI / ML-based CSI prediction is configured by the base station (1610).

[0354] As a specific example, when the terminal (1620) falls back from the AI / ML-based CSI reporting method to the existing CSI reporting method, the terminal can measure CSI based on the existing CSI reporting method and report the measured CSI to the base station (1610).

[0355] Thereafter, the base station (1610) can switch back to the AI / ML-based CSI reporting mode by utilizing at least one threshold value associated with the AI / ML model performance evaluation index / performance monitoring result or the amount of change from the previous result based on at least one of the model performance evaluation methods proposed in the present invention. Here, the one or more threshold values ​​can be transmitted to the terminal (1620) through upper layer signaling. Specifically, the base station (1610) can perform continuous AI / ML model performance evaluation as described above and compare the value and amount of change corresponding to the evaluation index with the set threshold value. Here, in at least one of the cases where the corresponding AI / ML model evaluation performance result / amount of change value has a reliability higher than the threshold value, when the set timer / time interval expires, or when a signal for model switching by the base station is transmitted to the terminal, the existing CSI-based reporting can be switched back to the AI / ML-based CSI reporting. Of course, this method is not limited to the application only when switching to AI / ML-based CSI reporting, and can also be applied when switching from AI / ML-based CSI reporting to the existing CSI reporting method as described above. The fallback / switching methods described above can be viewed as implicit signaling or condition-based triggering rather than explicit signaling, and in the latter case, explicit signaling for model switching by the base station or terminal may not be provided. The evaluation of whether to perform the fallback / switching can also be performed based on one of the AI / ML model evaluation methods performed on the base station or terminal side, similar to the methods considered in the methods proposed below. The final decision on whether to perform the fallback / switching can be primarily performed by the network side, but the terminal can make the final decision on the switch based on the triggering conditions for switching set by the base station to the terminal in advance.Of course, according to the proposed method, the base station and the terminal can exchange target CSI information, performance evaluation indices, and thresholds required for model evaluation indices through signaling. The calculation of performance evaluation indices can be performed according to the method applied on the base station or terminal side. When switching from AI / ML-based CSI reporting to the existing CSI reporting method or switching back from the existing CSI reporting method to the AI / ML-based CSI reporting method, the corresponding operation can be determined based on one or more evaluation results. For example, in order to prevent unnecessary frequent fallbacks or switching operations from continuously occurring, an additional threshold value can be set when based on multiple evaluation results, and the fallback operation can be performed when an evaluation result exceeds the threshold value. For example, if the number of times that the performance evaluation indices / results have low model reliability is greater than the threshold value, the fallback / switching operation for the corresponding model can be performed. On the other hand, fallback / switching is not performed for the existing AI / ML model-based CSI reporting. Or, conversely, if the number of times the performance evaluation metrics / outcome values ​​have high model confidence is greater than a threshold, the model can be maintained (if currently in use) or switched to an AI / ML-based model (e.g., if the existing CSI reporting method is in use).

[0356] As another example, the base station (1610) may transmit AI / ML model monitoring instruction information to the terminal (1620) while the existing CSI reporting is being performed. The terminal (1620) may obtain the model monitoring instruction information from the base station (1610) and perform an inference operation on the AI / ML model for CSI prediction. The terminal (1620) may transmit the CSI prediction result and the measurement result including the corresponding ground-true CSI to the base station (1610). The base station (1610) may calculate a performance evaluation index using the CSI prediction result and the ground-true CSI and compare it with a threshold value to determine whether to switch back to AI / ML-based CSI reporting. As an example, the threshold value may be a value different from the threshold value described above as a threshold value for determining whether the terminal (1620) reverts to AI / ML-based CSI reporting. In addition, a case where the threshold value is the same as the above-described threshold value may also be considered, and is not limited to a specific form. The base station (1610) may switch back from conventional CSI reporting to AI / ML-based CSI reporting based on the information described above. Here, the terminal (1620) may also need an instruction to switch back to AI / ML-based CSI reporting operation in order to perform AI / ML-based CSI reporting again. The fallback indication method described below can be applied to both cases of switching from AI / ML-based CSI reporting to conventional CSI reporting and from conventional CSI reporting to AI / ML-based CSI reporting, and is not limited to a specific form.

[0357] As a specific example, the fallback instruction can be implicitly indicated by providing the terminal with existing CSI reporting and resource configurations via higher layer configurations. That is, if the terminal (1620) receives existing CSI reporting and resource configurations via higher layer configurations, it can recognize that it is falling back to the existing CSI reporting. Conversely, if the terminal (1620) receives AI / ML-based CSI reporting and resource configurations via higher layer configurations while performing existing CSI reporting, it can recognize that it is falling back to AI / ML-based CSI reporting.

[0358] As another example, if the terminal (1620) has already been provided with the configuration for existing CSI reporting and resources from the base station (1610), the terminal (1620) may not be implicitly instructed to fallback. In the above-described case, the terminal (1620) may receive explicit signaling from the base station (1610) via at least one of radio resource control (RRC) and medium access control element (MAC CE) that instructs to perform CSI reporting using the existing CSI reporting method. In addition, if the terminal (1620) has already been provided with the configuration for AI / ML-based CSI reporting and resources from the base station (1610), the terminal (1620) may not be implicitly instructed to fallback. In the above-described case, the terminal (1620) may be provided with explicit signaling from the base station (1610) via at least one of RRC and MAC CE that instructs to perform CSI reporting using the AI / ML-based CSI reporting method.

[0359] As another example, the terminal may be instructed to fallback through L1 signaling corresponding to dynamic signaling (e.g., recycling an existing field (e.g., CSI request) in the DCI (downlink control information) format or an additional DCI field). The DCI format is a format distinct from the existing DCI format and may scramble the CRC (cyclic redundancy check) using a new RNTI (radio network temporary identifier) ​​value (e.g., AI / ML-CSI-RNTI) for AI / ML-based CSI reporting. For example, the DCI format defined by the above-described RNTI value may indicate whether to fallback by recycling an existing field or defining a new field, and is not limited to a specific form.

[0360] FIG. 17 is a diagram illustrating a fallback operation for an AI / ML-based CSI prediction model applicable to the present disclosure. Referring to FIG. 17, a terminal (1720) may transmit terminal capability information to a base station (1710), and the base station (1710) may transmit CSI-RS (channel state information) to the terminal (1720) as a reference signal for CSI measurement. In addition, the base station (1710) may transmit assistance information to the terminal (1720). Here, an AI / ML model for CSI prediction may be located on the terminal side as a one-side model. The terminal (1720) may perform measurements based on information acquired from the base station (1710) and, based on the measurements, may perform training on the AI / ML model for CSI prediction. Thereafter, the terminal (1720) may report the training and testing results to the base station (1710). Thereafter, the base station (1710) can transmit triggering / signaling to activate model inference to the terminal (1720). The terminal (1720) can perform an inference operation on the AI / ML model for CSI prediction based on the triggering / signaling for the received model inference, generate a CSI prediction value as an output value, and report the result to the base station (1710).

[0361] In addition, the base station (1710) and the terminal (1720) can monitor the AI / ML model for CSI prediction. Here, the terminal (1720) can calculate a performance evaluation index for the AI / ML model for CSI prediction. The performance evaluation index can be derived by utilizing at least one of the CSI prediction value and the corresponding ground truth CSI (i.e., actual data) value. The terminal (1720) can report the performance evaluation index to the base station (or network, 1710), and the base station (1710) can determine whether to perform functionality fallback based on the performance evaluation index. That is, in FIG. 14, the terminal derives a performance monitoring result value using the performance evaluation index and reports it to the base station, but in FIG. 17, the terminal (1720) can report the performance evaluation index itself to the base station (1710).

[0362] Here, the evaluation performance indicator can be determined as an intermediate value by considering at least one of NMSE and SGCS. Additionally, metrics that identify performance issues when data with a sudden drop in correlation between the data distribution and relationship between the input and output of the AI / ML model can be considered as other performance indicators. Furthermore, final system performance indicators (e.g., throughput, hypothetical BLER, BLER, NACK / ACK ratio) can be considered as model performance indicators and are not limited to a specific form. Here, whether the model performance indicator can be utilized in the AI / ML model can be determined based on the above-described information. This can have the advantage of avoiding the drawbacks (CSI reporting or out-of-band CSI indication overhead) that may arise from the above-described methods using NMSE and SGCS. However, the accuracy of the AI / ML model's applicability to the above-described information may not be clear. For example, it may be unclear whether the information changes due to other network or channel environment variables. Additionally, the need to monitor the aforementioned information for a relatively long period of time to obtain stable results may result in delay issues. Furthermore, for example, model performance metrics may be scored as a single generalized value and used by the terminal (1720) or reported to the base station (1710). Furthermore, whether or not to report model performance metrics may be determined based on a threshold value set through upper layer signaling, but is not limited thereto.

[0363] In addition, the terminal (1720) can report the performance evaluation index to the base station (1710) based on a preset cycle. The terminal (1720) can report to the base station (1710) periodically based on a relatively long cycle. As another example, the terminal (1720) can determine a situation requiring fallback on its own and aperiodically report the performance evaluation index to the base station (1710). As an example, the terminal (1720) can report the performance evaluation index to the base station (1710) when a specific condition is met in an event-triggered manner, but the present invention is not limited thereto. When the terminal (1720) periodically reports the performance evaluation index to the base station (1710), the terminal (1720) can receive upper layer settings related to CSI resources and reporting from the base station (1710). As an example, the terminal (1720) can receive upper layer settings for the performance evaluation index from the base station (1710), but the present invention is not limited thereto.

[0364] Additionally, a case may be considered in which the terminal (1720) reports information on whether to fallback to the base station (1710) at a time when necessary. The base station (1710) may provide the terminal (1720) with upper layer settings (e.g., CSI resources and reports) for the report in advance. If the terminal (1720) decides to report performance evaluation indicators to the base station (1710) on its own, the terminal (1720) may report the information to the base station (1710) based on the upper layer settings previously provided by the base station (1710).

[0365] That is, the terminal (1720) can provide a performance evaluation index to the base station (1710), and the base station (1710) can determine whether to execute a fallback based on the information. In addition, the base station (1710) can consider a threshold value for the fallback decision. For example, the threshold value can be a threshold value related to a performance evaluation index. In addition, the threshold value can be set by further considering at least one of the information included in Table 20 described above. Referring to Table 20, NMSE can be an index that measures the difference between data with a normalized mean square error, and can quantitatively represent the error between original data and predicted data. SGCS can be a squared generalized cosine similarity, and can be an index that measures the similarity between two vectors. Here, NMSE and SGCS can be further considered when determining the threshold value, but are not limited thereto. As another example, the threshold value can be determined by further utilizing RSRP or the information of the best K beam IDs. As another example, the threshold value may be determined by further utilizing at least one of the transmission rate, the assumed BLER, the BLER, and the NACK / ACK ratio information. As another example, the threshold value may further consider at least one of the function ID, the model ID, and the associated ID, but may not be limited to the embodiment. Information about at least one of the function ID, the model ID, and the associated ID may not directly indicate a performance indicator, but may be utilized through a combination of the ID information and other performance indicators. That is, the base station (1710) may further consider the above-described information to determine whether to perform a fallback, thereby improving reliability.For example, if the ID information corresponds to at least one of the following: system settings, network scenario environment, or cell structure, the reliability of the CSI prediction result value may vary depending on the environment or setting. Therefore, if information about the environment or setting is provided in advance, the threshold setting can be configured as a value independent of the environment or setting, thereby increasing the reliability of the final fallback decision.

[0366] Based on the above, if the base station (1710) determines a function fallback, the base station (1710) needs to indicate that the terminal (1720) is switching to a different method of CSI reporting. For example, signaling may be required when falling back from an AI / ML-based CSI reporting method to a conventional CSI reporting method. Furthermore, for example, the terminal (1720) may operate with the conventional CSI reporting method due to a temporary performance degradation of the AI / ML model even if the function corresponding to AI / ML-based CSI prediction is configured by the base station (1710).

[0367] As a specific example, when the terminal (1720) falls back from the AI / ML-based CSI reporting method to the existing CSI reporting method, the terminal can measure CSI based on the existing CSI reporting method and report the measured CSI to the base station (1710).

[0368] Thereafter, the base station (1710) can switch back to the AI / ML-based CSI reporting mode by utilizing at least one threshold value associated with the AI / ML model performance evaluation index / performance monitoring result or the amount of change from the previous result based on at least one of the model performance evaluation methods proposed in the present invention. Here, the one or more threshold values ​​can be transmitted to the terminal (1720) through upper layer signaling. Specifically, the base station (1710) can perform continuous AI / ML model performance evaluation as described above and compare the value and amount of change corresponding to the evaluation index with the set threshold value. Here, in at least one of the cases where the corresponding AI / ML model evaluation performance result / amount of change value has a reliability higher than the threshold value, when the set timer / time interval expires, or when a signal for model switching by the base station is transmitted to the terminal, the existing CSI-based reporting can be switched back to the AI / ML-based CSI reporting. Of course, this method is not limited to the application only when switching to AI / ML-based CSI reporting, and can also be applied when switching from AI / ML-based CSI reporting to the existing CSI reporting method as described above. The fallback / switching methods described above can be viewed as implicit signaling or condition-based triggering rather than explicit signaling, and in the latter case, explicit signaling for model switching by the base station or terminal may not be provided. The evaluation of whether to perform the fallback / switching can also be performed based on one of the AI / ML model evaluation methods performed on the base station or terminal side, similar to the methods considered in the methods proposed below. The final decision on whether to perform the fallback / switching can be primarily performed by the network side, but the terminal can make the final decision on the switch based on the triggering conditions for switching set by the base station to the terminal in advance.Of course, according to the proposed method, the base station and the terminal can exchange target CSI information, performance evaluation indices, and thresholds required for model evaluation indices through signaling. The calculation of performance evaluation indices can be performed according to the method applied on the base station or terminal side. When switching from AI / ML-based CSI reporting to the existing CSI reporting method or switching back from the existing CSI reporting method to the AI / ML-based CSI reporting method, the corresponding operation can be determined based on one or more evaluation results. For example, in order to prevent unnecessary frequent fallbacks or switching operations from continuously occurring, an additional threshold value can be set when based on multiple evaluation results, and the fallback operation can be performed when an evaluation result exceeds the threshold value. For example, if the number of times that the performance evaluation indices / results have low model reliability is greater than the threshold value, the fallback / switching operation for the corresponding model can be performed. On the other hand, fallback / switching is not performed for the existing AI / ML model-based CSI reporting. Or, conversely, if the number of times the performance evaluation metrics / outcome values ​​have high model confidence is greater than a threshold, the model can be maintained (if currently in use) or switched to an AI / ML-based model (e.g., if the existing CSI reporting method is in use).

[0369] As another example, the base station (1710) may transmit AI / ML model monitoring instruction information to the terminal (1720) while the existing CSI reporting is being performed. The terminal (1720) may obtain the model monitoring instruction information from the base station (1710), perform an inference operation on the AI / ML model for CSI prediction, and transmit a performance evaluation index based on the inference operation to the base station (1710). The base station (1710) may compare the performance evaluation index with a threshold value to determine whether to switch back to AI / ML-based CSI reporting. For example, the threshold value may be a value different from the threshold value described above as a threshold value for determining whether the terminal (1720) reverts to AI / ML-based CSI reporting. In addition, a case where the threshold value is the same as the threshold value described above may also be considered, and is not limited to a specific form. The base station (1710) may switch back from the existing CSI reporting to AI / ML-based CSI reporting based on the above-described information. Here, the terminal (1720) may also need to be instructed to switch to AI / ML-based CSI reporting operation in order to perform AI / ML-based CSI reporting again. The fallback indication method described below can be applied to both cases of switching from AI / ML-based CSI reporting to conventional CSI reporting and from conventional CSI reporting to AI / ML-based CSI reporting, and is not limited to a specific form. Additionally, if such fallback or transition is required on the terminal side, the terminal can report the relevant information to the base station, and the base station can then signal whether to instruct the fallback or transition using the proposed method.

[0370] As a specific example, the fallback instruction / model switching instruction can be implicitly indicated by providing the terminal with existing CSI reporting and resource configurations via upper layer configurations. That is, if the terminal (1720) receives existing CSI reporting and resource configurations via upper layer configurations, it can recognize that it is falling back to the existing CSI reporting. Conversely, if the terminal (1720) receives AI / ML-based CSI reporting and resource configurations via upper layer configurations while performing existing CSI reporting, it can recognize that it is falling back to AI / ML-based CSI reporting.

[0371] As another example, if the terminal (1720) has already been provided with the configuration for existing CSI reporting and resources from the base station (1710), the terminal (1720) may not be implicitly instructed to fallback. In the above-described case, the terminal (1720) may receive explicit signaling from the base station (1710) via at least one of radio resource control (RRC) and medium access control element (MAC CE) that instructs the terminal to perform CSI reporting using the existing CSI reporting method. In addition, if the terminal (1720) has already been provided with the configuration for AI / ML-based CSI reporting and resources from the base station (1710), the terminal (1720) may not be implicitly instructed to fallback. In the above-described case, the terminal (1720) may be provided with explicit signaling from the base station (1710) via at least one of RRC and MAC CE that instructs the terminal to perform CSI reporting using the AI / ML-based CSI reporting method.

[0372] As another example, the terminal may be instructed to fallback through L1 signaling corresponding to dynamic signaling (e.g., recycling an existing field (e.g., CSI request) in the DCI (downlink control information) format or an additional DCI field). The DCI format is a format distinct from the existing DCI format and may scramble the CRC (cyclic redundancy check) using a new RNTI (radio network temporary identifier) ​​value (e.g., AI / ML-CSI-RNTI) for AI / ML-based CSI reporting. For example, the DCI format defined by the above-described RNTI value may indicate whether to fallback by recycling an existing field or defining a new field, and is not limited to a specific form.

[0373] Additionally, for example, monitoring operations for a CSI prediction model may be essential operations and procedures for performing fallback operations. However, such monitoring operations may not always be performed at the terminal or base station. For example, when an AI / ML model is installed, it may be considered that the channel or network environment, which makes it difficult to apply the model, may change rapidly. Accordingly, monitoring may not always be performed, but at least one of periodic monitoring, event / condition-based monitoring, or network triggering-based monitoring operations may be performed, and is not limited to a specific form.

[0374] FIG. 18 is a flowchart illustrating a fallback operation for AI / ML-based CSI reporting applicable to the present disclosure. Referring to FIG. 18, a wireless user device may perform learning and testing on an AI / ML model for CSI prediction, and report learning and testing information to a base station (S1810). The wireless user device may receive an inference triggering of an AI / ML model for CSI prediction from the base station, and perform inference of the AI / ML model for CSI prediction based on the inference, thereby generating predicted CSI (S1820). Thereafter, the wireless user device may receive a monitoring instruction of the AI / ML model for CSI prediction from the base station (S1830), and report fallback-related information to the base station based on the monitoring instruction of the AI / ML model (S1840).

[0375] Here, the fallback-related information may be the performance monitoring results described above. The performance monitoring results may be determined based on a performance evaluation index derived from a comparison of the predicted CSI and the ground-true CSI. A base station that receives the performance monitoring results from a wireless user device may compare the performance evaluation index with a threshold value to determine whether to perform fallback, as illustrated in FIG. 14.

[0376] As another example, fallback-related information may be a fallback recommendation report. The fallback recommendation report may be information derived based on performance evaluation metrics and thresholds derived from a comparison of predicted CSI and ground-true CSI. That is, a wireless user device can directly compare the performance evaluation metrics and thresholds to generate an indicator indicating whether fallback is necessary and report this to the base station, where the base station can then determine whether fallback is necessary. This may be illustrated in Figure 15.

[0377] As another example, fallback-related information may be measurement information. The measurement information may include predicted CSI and ground-true CSI. The base station may derive performance evaluation metrics based on the measurement information received from the wireless user device and compare them with threshold values ​​to determine whether to initiate fallback, as illustrated in Figure 16.

[0378] As another example, fallback-related information may be a performance evaluation metric. This performance evaluation metric can be derived based on a comparison of predicted CSI and ground-true CSI. The base station can obtain the performance evaluation metric from the wireless user device and compare it to a threshold value to determine whether to perform fallback, as illustrated in Figure 17.

[0379] Figure 19 is a drawing showing a base station device and a terminal device to which the present disclosure can be applied.

[0380] The base station device (1900) may include a processor (1920), an antenna unit (1912), a transceiver (1926), and a memory (1916).

[0381] The processor (1920) performs baseband-related signal processing and may include a higher layer processing unit (1930) and a physical layer processing unit (1940). The higher layer processing unit (1930) may process operations of a MAC (Medium Access Control) layer, an RRC (Radio Resource Control) layer, or higher layers. The physical layer processing unit (1940) may process operations of a physical (PHY) layer (e.g., uplink reception signal processing, downlink transmission signal processing). In addition to performing baseband-related signal processing, the processor (1920) may also control the overall operation of the base station device (1900).

[0382] The antenna unit (1912) may include one or more physical antennas, and when it includes multiple antennas, it may support MIMO (Multiple Input Multiple Output) transmission and reception. In addition, it may support beamforming.

[0383] The memory (1916) can store information processed by the processor (1920), software related to the operation of the base station device (1900), an operating system, applications, etc., and may also include components such as a buffer.

[0384] The processor (1920) of the base station (1900) may be configured to implement the operations of the base station in the embodiments described in the present invention.

[0385] The terminal device (1950) may include a processor (1970), an antenna unit (1962), a transceiver (1964), and a memory (1966). For example, the terminal device (1950) in the present invention may communicate with a base station device (1900). As another example, the terminal device (1950) in the present invention may perform sidelink communication with another terminal device. That is, the terminal device (1950) in the present invention refers to a device that can communicate with at least one of the base station device (1900) and another terminal device, and is not limited to communication with a specific device.

[0386] The processor (1970) performs baseband-related signal processing and may include a higher layer processing unit (1980) and a physical layer processing unit (1990). The higher layer processing unit (1980) may process operations of the MAC layer, the RRC layer, or higher layers. The physical layer processing unit (1990) may process operations of the PHY layer (e.g., downlink reception signal processing, uplink transmission signal processing). In addition to performing baseband-related signal processing, the processor (1970) may also control the overall operation of the terminal device (1950).

[0387] The antenna unit (1962) may include one or more physical antennas, and when it includes multiple antennas, it may support MIMO transmission and reception. In addition, it may support beamforming.

[0388] The memory (1966) can store information processed by the processor (1970), software related to the operation of the terminal device (1950), an operating system, applications, etc., and may also include components such as a buffer.

[0389] A terminal device (1950) according to an example of the present invention may be associated with a vehicle. For example, the terminal device (1950) may be integrated into, positioned in, or on the vehicle. Furthermore, the terminal device (1950) according to the present invention may be the vehicle itself. Furthermore, the terminal device (1950) according to the present invention may be at least one of a wearable terminal, an AV / VR, an IoT terminal, a robot terminal, and a public safety terminal. The terminal device (1950) to which the present invention is applicable may include any type of communication device that supports interactive services utilizing sidelink for services such as Internet access, service execution, navigation, real-time information, autonomous driving, and safety and risk diagnosis. Furthermore, any type of communication device capable of sidelink operation or an AR / VR device or a sensor that performs relay operation may be included.

[0390] Here, the vehicles to which the present invention is applied may include autonomous vehicles, semi-autonomous vehicles, non-autonomous vehicles, etc. Meanwhile, while the terminal device (1950) according to one example of the present invention is described as being associated with a vehicle, one or more of the UEs may not be associated with a vehicle. This is merely an example, and should not be construed as limiting the application of the present invention to the described example.

[0391] In addition, a terminal device (1950) according to an example of the present invention can perform learning and testing on an AI / ML model for CSI prediction and report learning and testing information to a base station (1900). The terminal device (1950) can receive an inference triggering of an AI / ML model for CSI prediction from the base station (1900), and perform inference of the AI / ML model for CSI prediction based on the inference, thereby generating predicted CSI. Thereafter, the terminal device (1950) can receive a monitoring instruction of the AI / ML model for CSI prediction from the base station (1900), and report fallback-related information to the base station (1900) based on the monitoring instruction of the AI / ML model.

[0392] Here, the fallback-related information may be the performance monitoring results described above. The performance monitoring results may be determined based on a performance evaluation index derived from a comparison of the predicted CSI and the ground-true CSI. The base station (1900), which receives the performance monitoring results from the terminal device (1950), may compare the performance evaluation index with a threshold value to determine whether to perform fallback. As another example, the fallback-related information may be a fallback recommendation report. The fallback recommendation report may be information derived based on a performance evaluation index and a threshold value derived from a comparison of the predicted CSI and the ground-true CSI. That is, the terminal device (1950) may directly compare the performance evaluation index with the threshold value to generate an indicator for whether to perform fallback and report the result to the base station (1900), whereupon the base station (1900) may determine whether to perform fallback. As another example, the fallback-related information may be measurement information. The measurement information may include predicted CSI and ground-true CSI. The base station (1900) may derive a performance evaluation index based on the measurement information received from the terminal device (1950) and compare it with a threshold value to determine whether to perform a fallback. As another example, the fallback-related information may be a performance evaluation index. The performance evaluation index may be derived based on a comparison of the predicted CSI and ground-true CSI. The base station (1900) may obtain the performance evaluation index from the terminal device (1950) and compare it with a threshold value to determine whether to perform a fallback.

[0393] Additionally, various embodiments of the present disclosure may be implemented by hardware, firmware, software, or a combination thereof. In the case of hardware implementation, the embodiments may be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), general processors, controllers, microcontrollers, microprocessors, etc.

[0394] The scope of the present disclosure includes software or machine-executable instructions (e.g., operating systems, applications, firmware, programs, etc.) that cause operations according to the methods of various embodiments to be executed on a device or a computer, and a non-transitory computer-readable medium having such software or instructions stored thereon and executable on the device or computer.

[0395] The various embodiments of the present disclosure are not intended to list all possible combinations but rather to illustrate representative aspects of the present disclosure, and the matters described in the various embodiments may be applied independently or in combination of two or more.

[0396]

[0397] The above may also apply to other systems.

Claims

1. In terms of method, A step in which a wireless user device performs training and testing on an artificial intelligence (AI) / machine learning (ML) model for predicting channel state information (CSI), and reports the training and testing information to a base station; A step of receiving an inference triggering of an AI / ML model for CSI prediction from the base station, and performing inference of the AI / ML model for CSI prediction based thereon to generate a predicted CSI; A step of receiving a monitoring instruction of an AI / ML model for CSI prediction from the base station; and A method comprising the step of reporting fallback related information to the base station based on the monitoring instructions of the AI / ML model.

2. In paragraph 1, The above fallback related information is a performance monitoring result, and the performance monitoring result is determined based on a performance evaluation index derived based on a comparison between the predicted CSI and the ground-true CSI. A method in which whether to fallback at the base station is determined based on a threshold value based on the performance monitoring results.

3. In paragraph 1, The above fallback related information is a fallback recommendation report, and the fallback recommendation report is information determined based on a performance evaluation index and threshold value derived based on a comparison of the predicted CSI and the ground-true CSI. A method in which whether to fallback is determined at the base station based on the above fallback recommendation report.

4. In paragraph 1, The above fallback related information is measurement information, wherein the measurement information includes the predicted CSI and ground-true CSI, A method in which a performance evaluation index is derived from the base station based on the above measurement information and whether or not to fallback is determined.

5. In paragraph 1, In the first paragraph, The above fallback related information is a performance evaluation index, and the performance evaluation index is derived based on a comparison between the predicted CSI and the ground-true CSI. A method in which the performance evaluation index and threshold value are compared at the base station to determine whether to fallback.

6. For wireless devices, at least one processor; and A memory storing instructions that cause the wireless user device to perform a specific operation by the at least one processor, The above specific actions are: Perform training and testing on AI (artificial intelligence) / ML (machine learning) models for predicting CSI (channel state information), and report training and testing information to the base station. Receive inference triggering of the AI / ML model for CSI prediction from the base station, and perform inference of the AI / ML model for CSI prediction based on the inference to generate predicted CSI. Receive a monitoring instruction of the AI / ML model for CSI prediction from the base station, and A wireless user device that reports fallback related information to the base station based on the monitoring instructions of the AI / ML model.

Images