Method and device for signal transmission and reception in wireless communication system

Abstract

A method performed by a terminal in a wireless communication system according to at least one of the embodiments disclosed in the present specification may comprise the steps of: acquiring configurations for artificial intelligence / machine learning (AI / ML) models on the basis of higher layer signaling; monitoring one or more of the AI / ML models; and, on the basis of the monitoring of the one or more AI / ML models, performing AI / ML model management for switching or updating. The monitoring of the one or more AI / ML models includes at least one of the monitoring of a first AI / ML model in an inactive state or the monitoring of a second AI / ML model in an active state. The monitoring of the first AI / ML model in the inactive state may occupy fewer terminal processing units or occupy the terminal processing units for a shorter period of time than the monitoring of the second AI / ML model in the active state.

Description

Method and device for transmitting and receiving signals in a wireless communication system

[0001] The present disclosure relates to a wireless communication system, and more particularly, to a method and device for transmitting or receiving an uplink / downlink wireless signal in a wireless communication system.

[0002] Wireless communication systems are widely deployed to provide various types of communication services, such as voice and data. Typically, wireless communication systems are multiple access systems that support communication with multiple users by sharing available system resources (e.g., bandwidth, transmission power). Examples of multiple access systems include code division multiple access (CDMA), frequency division multiple access (FDMA), time division multiple access (TDMA), orthogonal frequency division multiple access (OFDMA), and single-carrier frequency division multiple access (SC-FDMA).

[0003] The technical task to be achieved is to provide a method for efficiently performing a wireless signal transmission and reception process and a device therefor.

[0004] The technical tasks to be achieved are not limited to the technical tasks mentioned above, and other technical tasks not mentioned can be inferred from the description below.

[0005] A method performed by a terminal in a wireless communication system according to one aspect comprises: obtaining a configuration for AI / ML (artificial intelligence / machine learning) models based on upper layer signaling; monitoring at least one of the AI / ML models; and performing AI / ML model management for switching or updating based on the monitoring of the at least one AI / ML model, wherein the monitoring of the at least one AI / ML model comprises at least one of monitoring of a first AI / ML model that is inactive and monitoring of a second AI / ML model that is active, wherein the monitoring of the first AI / ML model that is inactive may occupy a terminal processing unit for a shorter time or in a smaller number of units than the monitoring of the second AI / ML model that is active.

[0006] Among the AI / ML models that are inactive, the first AI / ML model to be monitored can be determined based on network signaling.

[0007] The start of the time interval in which the terminal processing unit is occupied by monitoring of the first AI / ML model may be determined based on the time at which the network signaling is received.

[0008] The end of the time interval during which the terminal processing unit is occupied by monitoring of the first AI / ML model may be determined based on at least one of: i) a time point at which the result of monitoring of the first AI / ML model is reported, ii) a time point at which a network response to the report is received, or iii) a time point at which a preset timer expires.

[0009] Among the AI / ML models that are inactive, the first AI / ML model to be monitored can be determined by the terminal.

[0010]

[0011] Monitoring of the inactive AI / ML model may be stopped or suspended based on whether an action on the active second AI / ML model is triggered via network signaling.

[0012] The terminal can report information about the occupied or unoccupied state of the terminal processing unit to the network.

[0013] Monitoring of the first AI / ML model can be performed based on the remaining terminal processing units that are not occupied by the second AI / ML model that is in an active state.

[0014] The second AI / ML model in the active state can occupy the terminal processing unit with a higher priority than the first AI / ML model in the inactive state.

[0015] Monitoring of the second AI / ML model in the active state occupies 1 terminal processing unit, and monitoring of the first AI / ML model in the inactive state occupies n terminal processing units, where n may be a real number greater than 0 and less than or equal to 1.

[0016] The above n may be determined based on terminal capability or provided through network signaling.

[0017] According to another aspect, a computer-readable recording medium having recorded thereon a program for performing the above-described method may be provided.

[0018] In another aspect, a terminal in a wireless communication system includes a transceiver; and a processor controlling the transceiver, wherein the processor obtains settings for AI / ML (artificial intelligence / machine learning) models based on upper layer signaling, monitors at least one of the AI / ML models, and performs AI / ML model management for switching or updating based on the monitoring of the at least one AI / ML model, wherein the monitoring of the at least one AI / ML model includes at least one of monitoring of an inactive first AI / ML model and monitoring of an active second AI / ML model, and the monitoring of the inactive first AI / ML model may occupy a terminal processing unit for a shorter time or in a smaller number of units than the monitoring of the active second AI / ML model.

[0019] In another aspect, a device for controlling a terminal in a wireless communication system includes a memory for storing commands; and a processor for performing operations for controlling the terminal by executing the commands, wherein the operations for controlling the terminal include obtaining settings for AI / ML (artificial intelligence / machine learning) models based on upper layer signaling; monitoring at least one of the AI / ML models; and performing AI / ML model management for switching or updating based on the monitoring of the at least one AI / ML model, wherein the monitoring of the at least one AI / ML model includes at least one of monitoring of a first AI / ML model that is inactive and monitoring of a second AI / ML model that is active, and wherein the monitoring of the first AI / ML model that is inactive may occupy a processing unit of the device for a shorter period of time or a smaller number of times than the monitoring of the second AI / ML model that is active.

[0020] According to one embodiment, wireless signal transmission and reception can be efficiently performed in a wireless communication system.

[0021] Other effects not mentioned can be inferred from the description below.

[0022] Figure 1 illustrates physical channels used in a 3GPP system, which is an example of a wireless communication system, and a general signal transmission method using the channels.

[0023] Figure 2 illustrates the structure of a radio frame.

[0024] Figure 3 illustrates a resource grid of slots.

[0025] Figure 4 illustrates an example of physical channels being mapped within a slot.

[0026] Figure 5 illustrates the PDSCH and ACK / NACK transmission process.

[0027] Figure 6 illustrates a PUSCH transmission process.

[0028] Figure 7 shows an example of a CSI-related procedure.

[0029] Figure 8 is a diagram to explain the concept of AI / ML / Deep learning.

[0030] Figures 9 to 12 illustrate various AI / ML models of deep learning.

[0031] Figure 13 is a diagram illustrating segmentation AI inference.

[0032] Figure 14 is a diagram illustrating a framework for 3GPP RAN Intelligence.

[0033] Figures 15 to 17 illustrate AI Model Training and Inference environments.

[0034] Figure 18 is a diagram for explaining AI-based CSI feedback.

[0035] FIG. 19 is a diagram for explaining network and terminal operations in a wireless communication system according to one embodiment.

[0036] FIG. 20 illustrates a flow of a method performed by a terminal according to one embodiment.

[0037] Figures 21 to 24 illustrate a communication system (1) and a wireless device applicable to the embodiments.

[0038] The following technologies can be used in various wireless access systems, such as CDMA (code division multiple access), FDMA (frequency division multiple access), TDMA (time division multiple access), OFDMA (orthogonal frequency division multiple access), and SC-FDMA (single carrier frequency division multiple access). CDMA can be implemented with radio technologies such as UTRA (Universal Terrestrial Radio Access) or CDMA2000. TDMA can be implemented with radio technologies such as GSM (Global System for Mobile communications) / GPRS (General Packet Radio Service) / EDGE (Enhanced Data Rates for GSM Evolution). OFDMA can be implemented with radio technologies such as IEEE 802.11 (Wi-Fi), IEEE 802.16 (WiMAX), IEEE 802-20, and E-UTRA (Evolved UTRA). UTRA is a part of UMTS (Universal Mobile Telecommunications System). 3GPP (3rd Generation Partnership Project) LTE (long term evolution) is part of E-UMTS (Evolved UMTS) that uses E-UTRA, and LTE-A (Advanced) is an evolved version of 3GPP LTE. 3GPP NR (New Radio or New Radio Access Technology) is an evolved version of 3GPP LTE / LTE-A.

[0039] As more and more communication devices demand greater communication capacity, the need for improved mobile broadband communications compared to existing Radio Access Technology (RAT) is emerging. Furthermore, massive Machine Type Communications (MTC), which connects multiple devices and objects to provide diverse services anytime, anywhere, is also a key issue to be considered in next-generation communications. Furthermore, communication system design that considers reliability and latency-sensitive services / terminals is being discussed. Accordingly, the introduction of next-generation RATs that consider enhanced Mobile BroadBand Communication (eMBB), massive MTC, and Ultra-Reliable and Low Latency Communication (URLLC) is being discussed. For convenience, this technology is referred to as NR (New Radio or New RAT) in the present invention.

[0040] For clarity of explanation, the description will focus on 3GPP NR, but the technical idea of ​​the present invention is not limited thereto.

[0041] In this specification, the expression "setting" can be replaced with the expression "configure / configuration", and the two can be used interchangeably. In addition, conditional expressions (e.g., "if", "in a case", or "when", etc.) can be replaced with the expression "based on that ~~" or "in a state / status". In addition, the operation of the terminal / base station or the SW / HW configuration according to the satisfaction of the condition can be inferred / understood. In addition, if the process of the receiving (or transmitting) side can be inferred / understood from the process of the transmitting (or receiving) side in signal transmission / reception between wireless communication devices (e.g., base stations, terminals), the description thereof can be omitted. For example, signal determination / generation / encoding / transmission, etc. of the transmitting side can be understood as signal monitoring reception / decoding / determination, etc. of the receiving side. In addition, the expression that the terminal performs (or does not perform) a specific operation can also be interpreted as meaning that the base station operates while expecting / assuming (or expecting / assuming that the terminal does not perform) the specific operation. In addition, the expression that the base station performs (or does not perform) a specific operation can also be interpreted as meaning that the terminal operates while expecting / assuming (or expecting / assuming that the base station does not perform) the specific operation. In addition, the division and index of each section, embodiment, example, option, method, plan, etc. in the following description are for the convenience of explanation and should not be interpreted as meaning that each constitutes an independent invention or that each must be implemented only individually. In addition, in describing each section, embodiment, example, option, method, plan, etc., if there is no explicitly conflicting / opposing description, it can be inferred / interpreted that at least some of them can be combined and implemented together, or at least some can be implemented with the omission of each.

[0042] In a wireless communication system, a terminal receives information from a base station via the downlink (DL) and transmits it to the base station via the uplink (UL). The information transmitted and received between the base station and the terminal includes data and various control information, and various physical channels exist depending on the type and purpose of the information being transmitted and received.

[0043] Figure 1 is a drawing for explaining physical channels used in a 3GPP NR system and a general signal transmission method using them.

[0044] When a terminal is powered on again from a powered-off state or enters a new cell, it performs an initial cell search operation, such as synchronizing with the base station, in step S101. To this end, the terminal receives a Synchronization Signal Block (SSB) from the base station. The SSB includes a Primary Synchronization Signal (PSS), a Secondary Synchronization Signal (SSS), and a Physical Broadcast Channel (PBCH). Based on the PSS / SSS, the terminal synchronizes with the base station and obtains information such as a cell ID (cell identity). In addition, the terminal can obtain broadcast information within the cell based on the PBCH. Meanwhile, the terminal can check the downlink channel status by receiving a Downlink Reference Signal (DL RS) during the initial cell search phase.

[0045] After completing the initial cell search, the terminal can obtain more specific system information by receiving a physical downlink control channel (PDCCH) and a physical downlink shared channel (PDSCH) based on the physical downlink control channel information in step S102.

[0046] Thereafter, the terminal may perform a random access procedure such as steps S103 to S106 to complete connection to the base station. To this end, the terminal may transmit a preamble through a physical random access channel (PRACH) (S103) and receive a response message to the preamble through a physical downlink control channel and a corresponding physical downlink shared channel (S104). In the case of contention-based random access, a contention resolution procedure such as transmission of an additional physical random access channel (S105) and reception of a physical downlink control channel and a corresponding physical downlink shared channel (S106) may be performed.

[0047] The terminal that has performed the procedure as described above can then perform the general uplink / downlink signal transmission procedure, such as receiving a physical downlink control channel / physical downlink shared channel (S107) and transmitting a physical uplink shared channel (PUSCH) / physical uplink control channel (PUCCH) (S108). The control information that the terminal transmits to the base station is collectively referred to as uplink control information (UCI). UCI includes HARQ ACK / NACK (Hybrid Automatic Repeat and request Acknowledgement / Negative-ACK), SR (Scheduling Request), CSI (Channel State Information), etc. CSI includes CQI (Channel Quality Indicator), PMI (Precoding Matrix Indicator), RI (Rank Indication), etc. UCI is generally transmitted through PUCCH, but can be transmitted through PUSCH when control information and traffic data must be transmitted simultaneously. Additionally, UCI can be transmitted aperiodically via PUSCH upon request / instruction from the network.

[0048] Figure 2 illustrates the structure of a radio frame. In NR, uplink and downlink transmissions are organized into frames. Each radio frame is 10 ms long and is divided into two 5 ms half-frames (HF). Each half-frame is divided into five 1 ms sub-frames (SF). A sub-frame is divided into one or more slots, and the number of slots within a sub-frame depends on the subcarrier spacing (SCS). Each slot contains 12 or 14 OFDM (Orthogonal Frequency Division Multiplexing) symbols, depending on the cyclic prefix (CP). When a normal CP is used, each slot contains 14 OFDM symbols. When an extended CP is used, each slot contains 12 OFDM symbols.

[0049] Table 1 illustrates that when CP is normally used, the number of symbols per slot, the number of slots per frame, and the number of slots per subframe vary depending on the SCS.

[0050] SCS (15*2^u)N slot symb N frame,u slot N subframe,u slot 15KHz (u=0)1410130KHz (u=1)1420260KHz (u=2)14404120KHz (u=3)14808240KHz (u=4)1416016

[0051] * N slot symb : Number of symbols in the slot

[0052] * N frame,u slot : Number of slots in the frame

[0053] * N subframe,u slot : Number of slots in a subframe

[0054] Table 2 illustrates that when extended CP is used, the number of symbols per slot, the number of slots per frame, and the number of slots per subframe vary depending on the SCS.

[0055] SCS (15*2^u)N slot symb N frame,u slot N subframe,u slot 60KHz (u=2)12404

[0056] The structure of the frame is only an example, and the number of subframes, number of slots, and number of symbols in the frame can be varied.

[0057] In an NR system, OFDM numerology (e.g., SCS) may be set differently between multiple cells that are merged into a single terminal. Accordingly, the (absolute time) interval of a time resource (e.g., SF, slot, or TTI) (conveniently referred to as TU (Time Unit)) consisting of the same number of symbols may be set differently between the merged cells. Here, the symbol may include an OFDM symbol (or CP-OFDM symbol), an SC-FDMA symbol (or Discrete Fourier Transform-spread-OFDM, DFT-s-OFDM symbol).

[0058] Figure 3 illustrates a resource grid of a slot. A slot contains multiple symbols in the time domain. For example, in the case of a regular CP, one slot contains 14 symbols, but in the case of an extended CP, one slot contains 12 symbols. A carrier contains multiple subcarriers in the frequency domain. A Resource Block (RB) is defined as multiple (e.g., 12) consecutive subcarriers in the frequency domain. A Bandwidth Part (BWP) is defined as multiple consecutive Physical RBs (PRBs) in the frequency domain and can correspond to a single numerology (e.g., SCS, CP length, etc.). A carrier can contain up to N (e.g., 5) BWPs. Data communication is performed through activated BWPs, and only one BWP can be activated for a single terminal. Each element in the resource grid is referred to as a Resource Element (RE), to which one complex symbol can be mapped.

[0059] Figure 4 illustrates an example of how physical channels are mapped within a slot. A PDCCH can be transmitted in the DL control region, and a PDSCH can be transmitted in the DL data region. A PUCCH can be transmitted in the UL control region, and a PUSCH can be transmitted in the UL data region. GP provides a time gap between the base station and the terminal when switching from transmission mode to reception mode or from reception mode to transmission mode. Some symbols within a subframe at the time of transition from DL to UL can be set as GP.

[0060] Below, each physical channel is described in more detail.

[0061] The PDCCH carries Downlink Control Information (DCI). For example, the PCCCH (i.e., DCI) carries the transmission format and resource allocation of the downlink shared channel (DL-SCH), resource allocation information for the uplink shared channel (UL-SCH), paging information for the paging channel (PCH), system information on the DL-SCH, resource allocation information for upper layer control messages such as random access responses transmitted on the PDSCH, transmission power control commands, activation / deactivation of Configured Scheduling (CS), etc. The DCI includes a cyclic redundancy check (CRC), which is masked / scrambled with various identifiers (e.g., Radio Network Temporary Identifier, RNTI) depending on the owner or usage of the PDCCH. For example, if the PDCCH is for a specific terminal, the CRC is masked with a terminal identifier (e.g., Cell-RNTI, C-RNTI). If the PDCCH is for paging, the CRC is masked with the Paging-RNTI (P-RNTI). If the PDCCH is for system information (e.g., a System Information Block, SIB), the CRC is masked with the System Information RNTI (SI-RNTI). If the PDCCH is for a random access response, the CRC is masked with the Random Access-RNTI (RA-RNTI).

[0062] The PDCCH consists of 1, 2, 4, 8, or 16 Control Channel Elements (CCEs) depending on the Aggregation Level (AL). A CCE is a logical allocation unit used to provide a PDCCH with a predetermined code rate depending on the radio channel status. A CCE consists of six Resource Element Groups (REGs). A REG is defined as one OFDM symbol and one (P)RB. The PDCCH is transmitted through a Control Resource Set (CORESET). A CORESET is defined as a set of REGs with a given numerology (e.g., SCS, CP length, etc.). Multiple CORESETs for a single UE can overlap in the time / frequency domain. A CORESET can be configured through system information (e.g., Master Information Block, MIB) or UE-specific upper layer (e.g., Radio Resource Control, RRC, layer) signaling. Specifically, the number of RBs and the number of OFDM symbols (up to 3) that constitute the CORESET can be set by upper layer signaling.

[0063] To receive / detect PDCCH, the UE monitors PDCCH candidates. PDCCH candidates represent the CCE(s) that the UE should monitor for PDCCH detection. Each PDCCH candidate is defined as 1, 2, 4, 8, or 16 CCEs depending on the AL. Monitoring involves (blind) decoding the PDCCH candidates. The set of PDCCH candidates that the UE monitors is defined as a PDCCH Search Space (SS). The search space includes a Common Search Space (CSS) or a UE-specific search space (USS). The UE can acquire DCI by monitoring PDCCH candidates in one or more search spaces configured by the MIB or higher-layer signaling. Each CORESET is associated with one or more search spaces, and each search space is associated with one COREST. The search space can be defined based on the following parameters.

[0064] - controlResourceSetId: Indicates the CORESET associated with the search space.

[0065] - monitoringSlotPeriodicityAndOffset: Indicates the PDCCH monitoring period (in slots) and the PDCCH monitoring interval offset (in slots).

[0066] - monitoringSymbolsWithinSlot: Indicates the PDCCH monitoring symbols within the slot (e.g., the first symbol(s) of the CORESET).

[0067] - nrofCandidates: AL={1, 2, 4, 8, 16} indicates the number of PDCCH candidates (one of 0, 1, 2, 3, 4, 5, 6, 8)

[0068] * An opportunity (e.g., time / frequency resource) for monitoring PDCCH candidates is defined as a PDCCH (monitoring) opportunity. One or more PDCCH (monitoring) opportunities can be configured within a slot.

[0069] Table 3 illustrates the characteristics of each search space type.

[0070] TypeSearch SpaceRNTIUse CaseType0-PDCCHCommonSI-RNTI on a primary cellSIB DecodingType0A-PDCCHCommonSI-RNTI on a primary cellSIB DecodingType1-PDCCHCommonRA-RNTI or TC-RNTI on a primary cellMsg2, Msg4 decoding in RACHType2-PDCCHCommonP-RNTI on a primary cellPaging DecodingType3-PDCCHCommonINT-RNTI, SFI-RNTI, TPC-PUSCH-RNTI, TPC-PUCCH-RNTI, TPC-SRS-RNTI, C-RNTI, MCS-C-RNTI, or CS-RNTI(s)UE SpecificUE SpecificC-RNTI, or MCS-C-RNTI, or CS-RNTI(s)User specific PDSCH decoding

[0071] Table 4 illustrates DCI formats transmitted via PDCCH.

[0072] DCI formatUsage0_0Scheduling of PUSCH in one cell0_1Scheduling of PUSCH in one cell1_0Scheduling of PDSCH in one cell1_1Scheduling of PDSCH in one cell2_0Notifying a group of UEs of the slot format2_1Notifying a group of UEs of the PRB(s) and OFDM symbol(s) where UE may assume no transmission is intended for the UE2_2Transmission of TPC commands for PUCCH and PUSCH2_3Transmission of a group of TPC commands for SRS transmissions by one or more UEs

[0073] DCI format 0_0 is used to schedule a TB-based (or TB-level) PUSCH, and DCI format 0_1 ​​can be used to schedule a TB-based (or TB-level) PUSCH or a CBG (Code Block Group)-based (or CBG-level) PUSCH. DCI format 1_0 is used to schedule a TB-based (or TB-level) PDSCH, and DCI format 1_1 can be used to schedule a TB-based (or TB-level) PDSCH or a CBG-based (or CBG-level) PDSCH (DL grant DCI). DCI format 0_0 / 0_1 may be referred to as UL grant DCI or UL scheduling information, and DCI format 1_0 / 1_1 may be referred to as DL grant DCI or DL ​​scheduling information. DCI format 2_0 is used to convey dynamic slot format information (e.g., dynamic SFI) to the terminal, and DCI format 2_1 is used to convey downlink pre-emption information to the terminal. DCI format 2_0 and / or DCI format 2_1 can be conveyed to the terminals within a group through the group common PDCCH, which is a PDCCH conveyed to the terminals defined as a group.

[0074] DCI format 0_0 and DCI format 1_0 may be referred to as fallback DCI formats, while DCI format 0_1 ​​and DCI format 1_1 may be referred to as non-fallback DCI formats. In the fallback DCI format, the DCI size / field configuration remains the same regardless of the terminal configuration. On the other hand, in the non-fallback DCI format, the DCI size / field configuration varies depending on the terminal configuration.

[0075] PDSCH carries downlink data (e.g., DL-SCH transport block, DL-SCH TB) and applies modulation methods such as Quadrature Phase Shift Keying (QPSK), 16 Quadrature Amplitude Modulation (QAM), 64 QAM, and 256 QAM. TB is encoded to generate a codeword. PDSCH can carry up to two codewords. Scrambling and modulation mapping are performed for each codeword, and modulation symbols generated from each codeword can be mapped to one or more layers. Each layer is mapped to resources along with a Demodulation Reference Signal (DMRS), generated as an OFDM symbol signal, and transmitted through the corresponding antenna port.

[0076] PUCCH carries Uplink Control Information (UCI). UCI includes:

[0077] - SR (Scheduling Request): Information used to request UL-SCH resources.

[0078] - HARQ(Hybrid Automatic Repeat reQuest)-ACK(Acknowledgement): This is a response to a downlink data packet (e.g., codeword) on the PDSCH. It indicates whether the downlink data packet was successfully received. One HARQ-ACK bit can be transmitted in response to a single codeword, and two HARQ-ACK bits can be transmitted in response to two codewords. The HARQ-ACK response includes a positive ACK (simply, ACK), a negative ACK (NACK), a DTX, or a NACK / DTX. Here, HARQ-ACK is used interchangeably with HARQ ACK / NACK and ACK / NACK.

[0079] - CSI (Channel State Information): Feedback information for the downlink channel. MIMO (Multiple Input Multiple Output)-related feedback information includes the Rank Indicator (RI) and Precoding Matrix Indicator (PMI).

[0080] Table 5 illustrates PUCCH formats. Depending on the PUCCH transmission length, they can be classified into Short PUCCH (formats 0 and 2) and Long PUCCH (formats 1, 3, and 4).

[0081] PUCCH formatLength in OFDM symbols N PUCCH symb Number of bitsUsageEtc01 - 2≤2HARQ, SRSequence selection14 - 14≤2HARQ, [SR]Sequence modulation21 - 2>2HARQ, CSI, [SR]CP-OFDM34 - 14>2HARQ, CSI, [SR]DFT-s-OFDM(no UE multiplexing)44 - 14>2HARQ, CSI, [SR]DFT-s-OFDM(Pre DFT OCC)

[0082] PUCCH format 0 carries UCI of up to 2 bits in size and is mapped and transmitted based on sequence. Specifically, the terminal transmits a specific UCI to the base station by transmitting one of multiple sequences through the PUCCH of PUCCH format 0. The terminal transmits the PUCCH of PUCCH format 0 within the PUCCH resources for the corresponding SR configuration only when transmitting a positive SR.

[0083] PUCCH format 1 carries UCI of up to 2 bits in size, and modulation symbols are spread in the time domain using an orthogonal cover code (OCC) (which is set differently depending on whether frequency hopping is used). DMRS are transmitted in symbols where modulation symbols are not transmitted (i.e., transmitted using Time Division Multiplexing (TDM).

[0084] PUCCH format 2 carries UCI with a bit size greater than 2 bits, and modulation symbols are transmitted by frequency division multiplexing (FDM) with DMRS. DM-RSs are located at symbol indices #1, #4, #7, and #10 within a given resource block with a density of 1 / 3. Pseudo Noise (PN) sequences are used for DM_RS sequences. Frequency hopping can be enabled for 2-symbol PUCCH format 2.

[0085] PUCCH format 3 does not multiplex terminals within the same physical resource blocks and carries UCI with a bit size greater than 2 bits. In other words, PUCCH resources in PUCCH format 3 do not include orthogonal cover codes. Modulation symbols are transmitted through time division multiplexing (TDM) with DMRS.

[0086] PUCCH format 4 supports multiplexing of up to four terminals within the same physical resource blocks and carries UCI with a bit size greater than 2 bits. In other words, PUCCH resources in PUCCH format 3 include orthogonal cover codes. Modulation symbols are transmitted through time division multiplexing (TDM) with DMRS.

[0087] At least one of one or more configured cells in a terminal may be configured for PUCCH transmission. At least the primary cell may be configured as a cell for PUCCH transmission. At least one PUCCH cell group may be configured in the terminal based on at least one cell configured for PUCCH transmission, and each PUCCH cell group includes one or more cells. The PUCCH cell group may be simply referred to as a PUCCH group. PUCCH transmission may be configured not only for the primary cell but also for the SCell, and the primary cell belongs to the primary PUCCH group, and the PUCCH-SCell configured for PUCCH transmission belongs to the secondary PUCCH group. For cells belonging to the primary PUCCH group, the PUCCH on the primary cell may be used, and for cells belonging to the secondary PUCCH group, the PUCCH on the PUCCH-SCell may be used.

[0088] PUSCH carries uplink data (e.g., UL-SCH transport block, UL-SCH TB) and / or uplink control information (UCI), and is transmitted based on a CP-OFDM (Cyclic Prefix - Orthogonal Frequency Division Multiplexing) waveform or a DFT-s-OFDM (Discrete Fourier Transform - spread - Orthogonal Frequency Division Multiplexing) waveform. When the PUSCH is transmitted based on a DFT-s-OFDM waveform, the UE transmits the PUSCH by applying transform precoding. For example, when transform precoding is disabled (e.g., transform precoding is disabled), the UE transmits the PUSCH based on the CP-OFDM waveform, and when transform precoding is enabled (e.g., transform precoding is enabled), the UE can transmit the PUSCH based on the CP-OFDM waveform or the DFT-s-OFDM waveform. PUSCH transmissions can be dynamically scheduled by UL grants in DCI, or semi-statically scheduled (configured grant) based on higher layer (e.g., RRC) signaling (and / or Layer 1 (L1) signaling (e.g., PDCCH)). PUSCH transmissions can be performed in a codebook-based or non-codebook-based manner.

[0089] Figure 5 illustrates an ACK / NACK transmission process. Referring to Figure 5, a terminal can detect a PDCCH in slot #n. Here, the PDCCH includes downlink scheduling information (e.g., DCI formats 1_0, 1_1), and the PDCCH indicates a DL assignment-to-PDSCH offset (K0) and a PDSCH-HARQ-ACK reporting offset (K1). For example, DCI formats 1_0, 1_1 can include the following information:

[0090] - Frequency domain resource assignment: Indicates the set of RBs allocated to the PDSCH.

[0091] - Time domain resource assignment: K0 (e.g., slot offset), indicates the starting position of the PDSCH within slot #n+K0 (e.g., OFDM symbol index), and the length of the PDSCH (e.g., number of OFDM symbols).

[0092] - PDSCH-to-HARQ_feedback timing indicator: Indicates K1

[0093] - HARQ process number (4 bits): Indicates the HARQ process ID (Identity) for data (e.g., PDSCH, TB)

[0094] - PUCCH resource indicator (PRI): Indicates the PUCCH resource to be used for UCI transmission among multiple PUCCH resources within the PUCCH resource set.

[0095] Afterwards, the terminal receives PDSCH from slot #(n+K0) according to the scheduling information of slot #n, and when reception of PDSCH is finished in slot #n1 (where, n+K0≤n1), UCI can be transmitted through PUCCH in slot #(n1+K1). Here, UCI may include HARQ-ACK response for PDSCH. In Fig. 5, for convenience, it is assumed that SCS for PDSCH and SCS for PUCCH are the same and slot # n1 = slot #n+K0, but the present invention is not limited thereto. If the SCSs are different, K1 can be indicated / interpreted based on the SCS of PUCCH.

[0096] When the PDSCH is configured to transmit at most 1 TB, the HARQ-ACK response may consist of 1 bit. When the PDSCH is configured to transmit at most 2 TB, the HARQ-ACK response may consist of 2 bits if spatial bundling is not configured, and may consist of 1 bit if spatial bundling is configured. When the HARQ-ACK transmission timing for multiple PDSCHs is designated as slot #(n+K1), the UCI transmitted in slot #(n+K1) includes HARQ-ACK responses for multiple PDSCHs.

[0097] Whether a UE should perform spatial bundling for a HARQ-ACK response can be configured (e.g., via RRC / higher layer signaling) for each cell group. For example, spatial bundling can be individually configured for each HARQ-ACK response transmitted over the PUCCH and / or each HARQ-ACK response transmitted over the PUSCH.

[0098] Spatial bundling can be supported when the maximum number of TBs (or codewords) that can be received at a time (or scheduled via 1 DCI) in the serving cell is 2 (or more than 2) (e.g., when the upper layer parameter maxNrofCodeWordsScheduledByDCI corresponds to 2-TB). Meanwhile, more than 4 layers can be used for 2-TB transmission, and up to 4 layers can be used for 1-TB transmission. Consequently, when spatial bundling is configured for the cell group, spatial bundling can be performed for serving cells that can schedule more than 4 layers among the serving cells in the cell group. On the serving cell, a terminal that wishes to transmit a HARQ-ACK response via spatial bundling can generate the HARQ-ACK response by performing a (bit-wise) logical AND operation on the A / N bits for multiple TBs.

[0099] For example, assuming that a terminal receives a DCI scheduling 2 TB and receives 2 TB via PDSCH based on the DCI, the terminal performing spatial bundling can generate a single A / N bit by logically ANDing the first A / N bit for the first TB and the second A / N bit for the second TB. Consequently, if both the first TB and the second TB are ACK, the terminal reports the ACK bit value to the base station, and if either TB is NACK, the terminal reports the NACK bit value to the base station.

[0100] For example, if only 1-TB is actually scheduled on a serving cell configured to receive 2-TB, the terminal can generate a single A / N bit by logically ANDing the A / N bit for the 1-TB with bit value 1. Consequently, the terminal reports the A / N bit for the 1-TB to the base station as is.

[0101] A base station / terminal has multiple parallel DL HARQ processes for DL ​​transmission. These multiple parallel HARQ processes allow DL transmissions to be performed continuously while waiting for HARQ feedback regarding the success or failure of the previous DL transmission. Each HARQ process is associated with a HARQ buffer in the MAC (Medium Access Control) layer. Each DL HARQ process manages state variables such as the number of transmissions of MAC Physical Data Blocks (PDUs) in the buffer, HARQ feedback for MAC PDUs in the buffer, and the current redundancy version. Each HARQ process is identified by a HARQ process ID.

[0102] Figure 6 illustrates a PUSCH transmission process. Referring to Figure 6, a terminal can detect a PDCCH in slot #n. Here, the PDCCH includes uplink scheduling information (e.g., DCI formats 0_0 and 0_1). DCI formats 0_0 and 0_1 can include the following information.

[0103] - Frequency domain resource assignment: Indicates the set of RBs allocated to PUSCH.

[0104] - Time domain resource assignment: Slot offset K2 indicates the starting position (e.g., symbol index) and length (e.g., number of OFDM symbols) of the PUSCH within the slot. The starting symbol and length can be indicated through SLIV (Start and Length Indicator Value) or can be indicated separately.

[0105] Thereafter, the terminal can transmit a PUSCH in slot #(n+K2) according to the scheduling information of slot #n. Here, the PUSCH includes a UL-SCH TB.

[0106] CSI-related actions

[0107] Figure 7 shows an example of a CSI-related procedure.

[0108] The terminal receives configuration information related to CSI from the base station via RRC signaling (710). The configuration information related to CSI may include at least one of CSI-IM (interference management) resource-related information, CSI measurement configuration-related information, CSI resource configuration-related information, CSI-RS resource-related information, or CSI report configuration-related information.

[0109] - CSI-IM resources can be configured for interference measurement (IM) of the terminal. In the time domain, the CSI-IM resource set can be configured periodically, semi-persistently, or aperiodicly. The CSI-IM resources can be configured as Zero Power (ZP)-CSI-RS for the terminal. The ZP-CSI-RS can be configured separately from the Non-Zero Power (NZP)-CSI-RS.

[0110] - The UE may assume that the CSI-RS resource(s) for channel measurement configured for one CSI reporting and the CSI-IM / NZP CSI-RS resource(s) for interference measurement (when NZP CSI-RS resource(s) are used for interference measurement) are in a QCL relationship with respect to 'QCL-TypeD' per resource.

[0111] - The CSI resource configuration may include at least one of a CSI-IM resource for interference measurement, an NZP CSI-RS resource for interference measurement, and an NZP CSI-RS resource for channel measurement. The CMR (channel measurement resource) may be an NZP CSI-RS for CSI acquisition, and the IMR (Interference measurement resource) may be an NZP CSI-RS for CSI-IM and IM.

[0112] - CSI-RS can be configured for one or more terminals. Different CSI-RS configurations may be provided for each terminal, or the same CSI-RS configuration may be provided to multiple terminals. CSI-RS can support up to 32 antenna ports. CSI-RS corresponding to N (N is 1 or greater) antenna ports can be mapped to N RE locations within a time-frequency unit corresponding to one slot and one RB. When N is 2 or greater, N-port CSI-RS can be multiplexed using CDM, FDM, and / or TDM schemes. CSI-RS can be mapped to REs other than REs to which CORESET, DMRS, and SSB are mapped. In the frequency domain, CSI-RS can be configured for the entire bandwidth, a portion of the bandwidth (BWP), or a portion of the bandwidth. CSI-RS may be transmitted in each RB within the bandwidth for which CSI-RS is configured (i.e., density = 1), or in every second RB (e.g., even or odd RB) (i.e., density = 1 / 2). When CSI-RS is used as a Tracking Reference Signal (TRS), a single-port CSI-RS may be mapped on three subcarriers in each resource block (i.e., density = 3). One or more CSI-RS resource sets may be configured for a UE in the time domain. Each CSI-RS resource set may include one or more CSI-RS configurations. Each CSI-RS resource set may be configured periodically, semi-persistently, or aperiodicly.

[0113] - The CSI report configuration may include configurations for feedback type, measurement resources, report type, etc. The NZP-CSI-RS resource set may be used for the CSI report configuration of the corresponding terminal. The NZP-CSI-RS resource set may be associated with CSI-RS or SSB. In addition, multiple periodic NZP-CSI-RS resource sets may be configured as TRS resource sets. (i) The feedback type may include a Channel Quality Indicator (CQI), a Precoding Matrix Indicator (PMI), a CSI-RS Resource Indicator (CRI), an SSB Resource block Indicator (SSBRI), a Layer Indicator (LI), a Rank Indicator (RI), a Layer 1-Reference Signal Received Strength (RSRP), etc. (ii) Measurement resources may include configurations for downlink signals and / or downlink resources on which the terminal performs measurements to determine feedback information. The measurement resources may be configured as ZP and / or NZP CSI-RS resource sets associated with CSI reporting configurations. The NZP CSI-RS resource set may include a CSI-RS set or an SSB set. For example, L1-RSRP may be measured for a CSI-RS set or an SSB set. (iii) Reporting types may include configurations for a time point at which the terminal performs reporting and an uplink channel, etc. The reporting time point may be configured as periodic, semi-persistent, or aperiodic. Periodic CSI reporting may be transmitted on PUCCH. Semi-persistent CSI reporting may be transmitted on PUCCH or PUSCH based on a MAC CE indicating activation / deactivation. Aperiodic CSI reporting may be indicated by DCI signaling.For example, the CSI request field of an uplink grant may indicate one of several report trigger sizes. Aperiodic CSI reports may be transmitted on the PUSCH.

[0114] The terminal measures CSI based on configuration information related to CSI. CSI measurement may include a procedure of receiving a CSI-RS (720) and computing the received CSI-RS to acquire CSI (730).

[0115] The UE can transmit a CSI report to the base station (740). For the CSI report, the time and frequency resources that the UE can use are controlled by the base station. The CSI (channel state information) can include at least one of a channel quality indicator (CQI), a precoding matrix indicator (PMI), a CSI-RS resource indicator (CRI), an SS / PBCH block resource indicator (SSBRI), a layer indicator (LI), a rank indicator (RI), L1-RSRP, and / or L-SINR.

[0116] The time domain behavior of CSI reporting supports periodic, semi-persistent, and aperiodic. i) Periodic CSI reporting is performed on short PUCCH and long PUCCH. The periodicity and slot offset of periodic CSI reporting can be configured by RRC, and refer to the CSI-ReportConfig IE. ii) SP (semi-periodic) CSI reporting is performed on short PUCCH, long PUCCH, or PUSCH. In case of SP CSI on short / long PUCCH, the periodicity and slot offset are configured by RRC, and CSI reporting is activated / deactivated by separate MAC CE / DCI. In case of SP CSI on PUSCH, the periodicity of SP CSI reporting is configured by RRC, but the slot offset is not configured by RRC, and SP CSI reporting is activated / deactivated by DCI (format 0_1). For SP CSI reporting on PUSCH, a separate RNTI (SP-CSI C-RNTI) is used. The initial CSI reporting timing follows the PUSCH time domain allocation value indicated in the DCI, and subsequent CSI reporting timings follow the cycle set by RRC. DCI format 0_1 ​​includes a CSI request field and can activate / deactivate a specific configured SP-CSI trigger state. SP CSI reporting has the same or similar activation / deactivation mechanism as the data transmission mechanism on the SPS PUSCH.iii) Aperiodic CSI reporting is performed on PUSCH and is triggered by DCI. In this case, information related to the triggering of aperiodic CSI reporting can be transmitted / indicated / configured via MAC-CE. For AP CSI with AP CSI-RS, the AP CSI-RS timing is configured by RRC, and the timing for AP CSI reporting is dynamically controlled by DCI.

[0117] CSI codebooks defined in the NR standard (e.g., PMI codebooks) can be broadly divided into Type I and Type II codebooks. Type I codebooks are primarily targeted at SU (Single User)-MIMO, which supports both high-order and low-order signals. Type II codebooks can primarily support MI-MIMO, which supports up to two layers. Compared to Type I, Type II codebooks can provide more accurate CSI, but may increase signaling overhead. Meanwhile, Enhanced Type II codebooks were introduced to address the CSI overhead shortcomings of existing Type II codebooks. Enhanced Type II codebooks were introduced by reducing the codebook payload by considering frequency-axis correlation.

[0118] CSI reporting via PUSCH can be configured as Part 1 and Part 2. Part 1 has a fixed payload size and is used to identify the number of information bits in Part 2. Part 1 is transmitted in its entirety before Part 2.

[0119] - For Type I CSI feedback, Part 1 contains the RI (if reported), the CRI (if reported), and the CQI of the first code word. Part 2 contains the PMI, and when RI > 4, Part 2 contains the CQI.

[0120] - For Type II CSI feedback, Part 1 contains the RI (if reported), CQI, and an indication of the number of non-zero WB amplitude coefficients per layer of Type II CSI. Part 2 contains the PMI of Type II CSI.

[0121] - For Enhanced Type II CSI feedback, Part 1 contains the RI (if reported), CQI, and the total number of non-zero WB amplitude coefficients for all layers of Enhanced Type II CSI. Part 2 contains the PMI of Enhanced Type II CSI.

[0122] If CSI reporting on PUSCH includes two parts and the CSI payload to be reported is less than the payload size provided by the PUSCH resources allocated for CSI reporting, the UE may omit part of Part 2 CSI.

[0123] Meanwhile, semi-persistent CSI reporting performed in PUCCH format 3 or 4 supports Type II CSI feedback, but only Part 1 of Type II CSI feedback.

[0124] QCL (quasi-co location)

[0125] Two antenna ports are quasi-co-located if the channel properties of one antenna port can be inferred from the channel properties of the other antenna port. The channel properties may include one or more of Delay spread, Doppler spread, Frequency / Doppler shift, Average received power, Received Timing / average delay, and Spatial RX parameters.

[0126] A terminal can configure a list of multiple TCI-State configurations via the upper layer parameter PDSCH-Config. Each TCI-State is associated with one or two DL reference signals and a QCL configuration parameter between the DM-RS port of the PDSCH. The QCL can include qcl-Type1 for the first DL RS and qcl-Type2 for the second DL RS. The QCL type can correspond to one of the following:

[0127] - 'QCL-TypeA': {Doppler shift, Doppler spread, average delay, delay spread}

[0128] - 'QCL-TypeB': {Doppler shift, Doppler spread}

[0129] - 'QCL-TypeC': {Doppler shift, average delay}

[0130] - 'QCL-TypeD': {Spatial Rx parameter}

[0131] Beam Management (BM)

[0132] The BM process is a process for acquiring and maintaining a set of BS (or transmission and reception point (TRP)) and / or UE beams that can be used for downlink (DL) and uplink (UL) transmission / reception, and may include the following processes and terms.

[0133] - Beam measurement: An operation in which a BS or UE measures the characteristics of a received beamforming signal.

[0134] - Beam determination: An operation in which a BS or UE selects its own transmit beam (Tx beam) / receive beam (Rx beam).

[0135] - Beam sweeping: An operation of covering a spatial domain using transmit and / or receive beams over a predetermined time interval in a predetermined manner.

[0136] - Beam report: An operation in which a UE reports information about a beamformed signal based on beam measurement.

[0137] The BM process can be divided into (1) a DL BM process using SSB or CSI-RS, and (2) a UL BM process using SRS (sounding reference signal). In addition, each BM process can include Tx beam sweeping to determine a Tx beam and Rx beam sweeping to determine an Rx beam.

[0138] At this time, the DL BM process may include (1) transmission of beamformed DL RSs (e.g., CSI-RS or SSB) by the BS and (2) beam reporting by the UE.

[0139] Here, the beam report may include preferred DL RS ID(s) and corresponding reference signal received power (RSRP). The DL RS ID may be an SSB Resource Indicator (SSBRI) or a CSI-RS Resource Indicator (CRI).

[0140] AI / ML (Artificial intelligence / machine learning)

[0141] Technological advancements in AI / ML are leading to the intelligence / advanced development of node(s) and terminal(s) that make up wireless communication networks. In particular, the intelligence of networks / base stations will enable the rapid optimization / derivation / application of various network / base station decision parameter values ​​(e.g., transmit / receive power of each base station, transmit power of each terminal, precoder / beam of base stations / terminals, time / frequency resource allocation for each terminal, duplexing method of each base station, etc.) based on various environmental parameters (e.g., distribution / location of base stations, distribution / location / material of buildings / furniture, etc., location / movement direction / speed of terminals, climate information, etc.). In line with this trend, many standardization organizations (e.g., 3GPP, O-RAN) are considering its introduction, and research on it is also actively underway.

[0142] AI / ML can be easily referred to as artificial intelligence based on deep learning in a narrow sense, but conceptually it is as shown in Figure 8.

[0143] - Artificial Intelligence: This can refer to all automation where machines can replace tasks that people would otherwise do.

[0144] - Machine Learning: Machines can learn patterns for decision-making from data without explicitly programming rules.

[0145] Deep Learning: An AI / ML model based on artificial neural networks. Machines simultaneously extract features from unstructured data and make judgments. The algorithms rely on multilayer networks of interconnected nodes for feature extraction and transformation, inspired by the biological nervous system, or neural networks. Common deep learning network architectures include deep neural networks (DNNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs).

[0146] Classification of AI / ML types based on various criteria

[0147] 1. Offline vs. Online

[0148] (1) Offline Learning: This follows a sequential process of database collection, learning, and prediction. In other words, collection and learning are performed offline, and the completed program can be installed on-site for use in prediction tasks. This offline learning method is used in most situations. In offline learning, the system does not learn incrementally; learning is performed using all available collected data and applied to the system without further training. If learning on new data is required, learning can be restarted using the entire new data set.

[0149] (2) Online Learning: Online learning leverages the continuous availability of data available for learning via the Internet. This method incrementally improves performance by learning from additional data. Learning is performed in real time on specific data (sets) collected online, enabling the system to quickly adapt to changing data.

[0150] To build an AI system, only online learning may be used, so that learning is performed using only real-time data, or offline learning may be performed using a predetermined data set, and then additional learning may be performed using additional real-time data (online + offline learning).

[0151] 2. Classification by AI / ML Framework Concept

[0152] (1) Centralized Learning: Training data collected from multiple different nodes are reported to a centralized node, and all data resources / storage / learning (e.g., supervised, unsupervised, reinforcement learning) are performed in one central node.

[0153] (2) Federated Learning: A collective AI / ML model is built based on data across distributed data owners. Instead of importing data into an AI / ML model, the AI / ML model is imported as a data source, allowing local nodes / individual devices to collect data and train their own copies of the AI / ML model, eliminating the need to report source data to a central node. In federated learning, the parameters / weights of the AI / ML model are simply sent back to the centralized node to support general AI / ML model training. The advantages of federated learning include increased computational speed and superior information security. This eliminates the need to upload personal data to a central server, preventing personal information leaks and misuse.

[0154] (3) Distributed Learning: This concept represents the concept of machine learning processes being scaled and distributed across a cluster of nodes. Training AI / ML models are split and shared across multiple nodes operating simultaneously to accelerate AI / ML model training.

[0155] 3. Classification by learning method

[0156] (1) Supervised Learning: Supervised learning is a machine learning task that aims to learn a mapping function from input to output given a labeled data set. The input data is called training data and has known labels or outcomes. Examples of supervised learning include (i) Regression: Linear Regression, Logistic Regression, (ii) Instance-based Algorithms: k-Nearest Neighbor (KNN), (iii) Decision Tree Algorithms: CART, (iv) Support Vector Machines: SVM, (v) Bayesian Algorithms: Naive Bayes, and (vi) Ensemble Algorithms: Extreme Gradient Boosting, Bagging: Random Forest. Supervised learning can be further grouped into regression and classification problems, where classification predicts labels and regression predicts quantities.

[0157] (2) Unsupervised Learning: A machine learning task that aims to learn features that explain hidden structures in unlabeled data. The input data is unlabeled and has no known outcome. Some examples of unsupervised learning include K-means clustering, principal component analysis (PCA), nonlinear independent component analysis (ICA), and long-term memory (LSTM).

[0158] (3) Reinforcement Learning: In reinforcement learning (RL), an agent interacts with the environment through a trial-and-error process, aiming to optimize a long-term goal. It is a goal-oriented learning method based on interaction with the environment. Examples of RL algorithms include (i) Q-learning, (ii) multi-armed bandit learning, (iii) deep Q network, state-action-reward-state-action (SARSA), (iv) temporal difference learning, (v) actor-critic reinforcement learning, (vi) deep deterministic policy gradient, and (vii) Monte-Carlo tree search. Reinforcement learning can be further grouped into AI / ML model-based reinforcement learning and AI / ML model-free reinforcement learning. Model-based reinforcement learning is an RL algorithm that uses a predictive AI / ML model to obtain transition probabilities between states by using various dynamic states of the environment and the AI / ML model that leads to these states as rewards. Model-free reinforcement learning is a value- or policy-based RL algorithm that maximizes future rewards. It is computationally less complex in multi-agent environments / states and does not require an accurate representation of the environment. RL algorithms can also be categorized into value-based RL versus policy-based RL, and policy-based RL versus non-policy RL.

[0159] AI / ML models

[0160] Figure 9 illustrates an FFNN (Feed-Forward Neural Network) AI / ML model. Referring to Figure 9, the FFNN AI / ML model includes an input layer, a hidden layer, and an output layer.

[0161] Figure 10 illustrates an RNN (Recurrent Neural Network) AI / ML model. Referring to Figure 10, the RNN AI / ML model is a type of artificial neural network in which hidden nodes are connected by directed edges to form a cyclic structure (directed cycle), and is an AI / ML model suitable for processing data that appears sequentially, such as voice and text. One type of RNN is LSTM (Long Short-Term Memory), and LSTM is a structure that adds a cell state to the hidden state of the RNN. Specifically, in LSTM, an input gate, a forget gate, and an output gate are added to the RNN cell, and a cell state is added. In Figure 10, A represents a neural network, x t is the input value, h t represents the output value. Here, h t can mean a status value that represents the present based on time, and h t-1 can represent the previous state value.

[0162] Figure 11 illustrates a CNN (Convolution Neural Network) AI / ML model. CNN uses convolution operations commonly used in image processing and video processing to achieve two goals: reducing AI / ML model complexity and extracting good features. Referring to Figure 11, a kernel or filter refers to a unit / structure that applies weights to inputs within a specific range / unit. The kernel (or filter) can be modified through learning. The stride refers to the range of movement of the kernel within the input. The feature map refers to the result of applying the kernel to the input. Padding refers to a value added to adjust the size of the feature map. Multiple feature maps can be extracted to induce robustness to distortion and changes. Pooling refers to an operation (e.g., max pooling, average pooling) to reduce the size of the feature map by downsampling it.

[0163] Figure 12 illustrates an auto-encoder AI / ML model. Referring to Figure 12, an auto-encoder is a neural network that receives a feature vector x as input and outputs the same or similar vector x'. The input and output nodes have the same features, and it is a type of unsupervised learning. Since an auto-encoder reconstructs the input, the output can be referred to as a reconstruction. The loss function can be expressed as in Mathematical Formula 1.

[0164]

[0165] The loss function of the auto encoder exemplified in Figure 12 is calculated based on the difference between the input and the output, and based on this, the degree of loss of the input is identified, and the auto encoder performs an optimization process to minimize the loss.

[0166] Figure 13 is a diagram illustrating segmentation AI inference.

[0167] Figure 13 illustrates a case where, among split AI operations, the Model Inference function is performed collaboratively by an end device such as a UE and a network AI / ML endpoint.

[0168] In addition to the Model Inference function, the Model Training function, Actor, and Data Collection functions can each be split into multiple parts depending on the current task and environment, and performed by multiple entities collaborating.

[0169] For example, computationally intensive and energy-intensive parts may be performed at the network endpoint, while privacy-sensitive and latency-sensitive parts may be performed at the end device. In this case, the end device may execute the task / model from input data up to a specific part / layer, and then transmit the intermediated data to the network endpoint. The network endpoint then executes the remaining parts / layers and provides the inference outputs to one or more devices that perform the actions / tasks.

[0170] The following describes a functional framework for AI operations.

[0171] Below, to explain AI (or AI / ML) more specifically, the terms can be defined as follows.

[0172] - Data collection: Data collected from network nodes, management entities, or UEs as a basis for AI model training, data analysis, and inference.

[0173] - AI Model: A data-driven algorithm that applies AI technology to generate a set of outputs containing predictive information and / or decision parameters based on a set of inputs.

[0174] - AI / ML Training: An online or offline process of training an AI model by learning features and patterns that best represent the data and obtain a trained AI / ML model for inference.

[0175] - AI / ML Inference: The process of making predictions or inducing decisions based on collected data and the AI ​​model using a trained AI model.

[0176] Referring to FIG. 14, the Data Collection function (10) is a function that collects input data and provides processed input data to the Model Training function (20) and the Model Inference function (30).

[0177] Examples of input data may include measurements from UEs or other network entities, feedback from actors, and output from AI models.

[0178] The Data Collection function (10) performs data preparation based on input data and provides input data processed through data preparation. Here, the Data Collection function (10) does not perform data preparation specific to each AI algorithm (e.g., data pre-processing and cleaning, formatting, and transformation), but can perform data preparation common to AI algorithms.

[0179] After the data preparation process is performed, the Model Training function (10) provides training data (11) to the Model Training function (20) and provides inference data (Inference Data) (12) to the Model Inference function (30). Here, the Training Data (11) is data required as input for the AI ​​Model Training function (20). The Inference Data (12) is data required as input for the AI ​​Model Inference function (30).

[0180] The Data Collection function (10) may be performed by a single entity (e.g., UE, RAN node, network node, etc.) or may be performed by multiple entities. In this case, Training Data (11) and Inference Data (12) may be provided to the Model Training function (20) and Model Inference function (30), respectively, from multiple entities.

[0181] The Model Training function (20) is a function that performs AI model training, validation, and testing, which can generate model performance metrics as part of the AI ​​model testing process. If necessary, the Model Training function (20) also handles data preparation (e.g., data pre-processing and cleaning, forming, and transformation) based on the Training Data (11) provided by the Data Collection function (10).

[0182] Here, Model Deployment / Update (13) is used to initially deploy the trained, verified, and tested AI model to the Model Inference function (30) or to provide the updated model to the Model Inference function (30).

[0183] The Model Inference function (30) is a function that provides AI model inference output (16) (e.g., prediction or decision). If applicable, the Model Inference function (30) may provide model performance feedback (14) to the Model Training function (20). In addition, the Model Inference function (30) is also responsible for data preparation (e.g., data pre-processing and cleaning, forming, and transformation) based on the Inference Data (12) provided by the Data Collection function (10), if necessary.

[0184] Here, Output (16) refers to the inference output of the AI ​​model generated by the Model Inference function (30), and the details of the inference output may vary depending on the use case.

[0185] Model Performance Feedback (14) can be used to monitor the performance of the AI ​​model if available, and this feedback may be omitted.

[0186] The actor function (40) is a function that receives the output (16) from the model inference function (30) and triggers or performs a corresponding task / action. The actor function (40) can trigger tasks / actions for other entities (e.g., one or more UEs, one or more RAN nodes, one or more network nodes, etc.) or for itself.

[0187] Feedback (15) can be used to derive training data (11), inference data (12), or to monitor the performance of the AI ​​model, its impact on the network, etc.

[0188] Meanwhile, the definitions of training / validation / test in the data set used in AI / ML can be distinguished as follows.

[0189] - Training data: This refers to the data set for learning the model.

[0190] - Validation data: This refers to a data set used to validate a model that has already completed training. In other words, it refers to a data set typically used to prevent overfitting of the training data set.

[0191] It also refers to a data set for selecting the best model among the various models learned during the learning process. Therefore, it can be viewed as a type of learning.

[0192] - Test data: This refers to the data set for final evaluation. This data is unrelated to learning.

[0193] In the case of the above data set, if the training set is generally divided, the training data and validation data can be divided and used in a ratio of 8:2 or 7:3 within the entire training set, and if the test is included, it can be divided and used in a ratio of 6:2:2 (training: validation: test).

[0194] Depending on the capability of the AI / ML function between the base station and the terminal, the level of cooperation can be defined as follows, and variations due to combination of multiple levels or separation of any one level are also possible.

[0195] Cat 0a) No collaboration framework: AI / ML algorithms are purely implementation-based and do not require any changes to the wireless interface.

[0196] Cat 0b) This level corresponds to a framework with a modified wireless interface tailored to efficient implementation-based AI / ML algorithms, but without collaboration.

[0197] Category 1) involves inter-node support to improve the AI / ML algorithms of each node. This applies when the UE receives support from the gNB (for training, adaptation, etc.) and vice versa. At this level, model exchange between network nodes is not required.

[0198] Category 2) Joint ML tasks can be performed between the UE and gNB. This level requires the exchange of AI / ML model commands or network nodes.

[0199] The functions exemplified in FIG. 14 above may be implemented in a RAN node (e.g., a base station, a TRP, a central unit (CU) of a base station, etc.), a network node, an operation administration maintenance (OAM) of a network operator, or a UE.

[0200] Alternatively, two or more entities, such as a RAN, a network node, a network operator's OAM, or a UE, may cooperate to implement the functions exemplified in FIG. 14. For example, one entity may perform some of the functions of FIG. 14, and another entity may perform the remaining functions. In this way, since some of the functions exemplified in FIG. 14 are performed by a single entity (e.g., a UE, a RAN node, a network node, etc.), the transmission / provision of data / information between each function may be omitted. For example, if the Model Training function (20) and the Model Inference function (30) are performed by the same entity, the transmission / provision of Model Deployment / Update (13) and Model Performance Feedback (14) may be omitted.

[0201] Alternatively, any one of the functions illustrated in FIG. 14 may be performed through collaboration between two or more entities, including a RAN, a network node, a network operator's OAM, or a UE. This may be referred to as a split AI operation.

[0202] Figure 15 illustrates a case where the AI ​​Model Training function is performed by a network node (e.g., a core network node, an OAM of a network operator, etc.) and the AI ​​Model Inference function is performed by a RAN node (e.g., a base station, a TRP, a CU of a base station, etc.).

[0203] Step 1: RAN node 1 and RAN node 2 transmit input data (i.e., training data) for AI model training to the network node. Here, RAN node 1 and RAN node 2 can also transmit data collected from the UE (e.g., UE measurements related to RSRP, RSRQ, SINR of the serving cell and neighboring cells, UE location, speed, etc.) to the network node.

[0204] Step 2: Network nodes train the AI ​​model using the received training data.

[0205] Step 3: The network node distributes / updates the AI ​​Model to RAN Node 1 and / or RAN Node 2. RAN Node 1 (and / or RAN Node 2) may continue model training based on the received AI Model.

[0206] For convenience of explanation, we assume that the AI ​​Model is deployed / updated only to RAN node 1.

[0207] Step 4: RAN node 1 receives input data for AI Model Inference (i.e., Inference data) from UE and RAN node 2.

[0208] Step 5: RAN node 1 performs AI Model Inference using the received Inference data to generate output data (e.g., prediction or decision).

[0209] Step 6: If applicable, RAN node 1 may transmit model performance feedback to the network nodes.

[0210] Step 7: RAN node 1, RAN node 2, and the UE (or 'RAN node 1 and the UE', or 'RAN node 1 and the RAN node 2') perform actions based on the output data. For example, in the case of a load balancing operation, the UE may move from RAN node 1 to RAN node 2.

[0211] Step 8: RAN node 1 and RAN node 2 transmit feedback information to the network nodes.

[0212] Figure 16 illustrates a case where both the AI ​​Model Training function and the AI ​​Model Inference function are performed by RAN nodes (e.g., base stations, TRPs, CUs of base stations, etc.).

[0213] Step 1: UE and RAN node 2 transmit input data (i.e., training data) for AI model training to RAN node 1.

[0214] Step 2: RAN node 1 trains the AI ​​model using the received training data.

[0215] Step 3: RAN node 1 receives input data for AI Model Inference (i.e., Inference data) from the UE and RAN node 2.

[0216] Step 4: RAN node 1 performs AI Model Inference using the received Inference data to generate output data (e.g., prediction or decision).

[0217] Step 5: RAN node 1, RAN node 2, and the UE (or 'RAN node 1 and the UE', or 'RAN node 1 and the RAN node 2') perform actions based on the output data. For example, in the case of a load balancing operation, the UE may move from RAN node 1 to RAN node 2.

[0218] Step 6: RAN node 2 sends feedback information to RAN node 1.

[0219] Figure 17 illustrates a case where the AI ​​Model Training function is performed by a RAN node (e.g., a base station, a TRP, a CU of the base station, etc.) and the AI ​​Model Inference function is performed by a UE.

[0220] Step 1: The UE transmits input data (i.e., training data) for AI model training to the RAN node. Here, the RAN node can collect data (e.g., UE measurements related to RSRP, RSRQ, SINR of the serving cell and neighboring cells, UE location, speed, etc.) from various UEs and / or from other RAN nodes.

[0221] Step 2: The RAN node trains the AI ​​model using the received training data.

[0222] Step 3: The RAN node distributes / updates the AI ​​model to the UE. The UE may also continue model training based on the received AI model.

[0223] Step 4: Receive input data (i.e., Inference data) for AI Model Inference from the UE and RAN nodes (and / or from other UEs).

[0224] Step 5: The UE performs AI Model Inference using the received Inference data to generate output data (e.g., prediction or decision).

[0225] Step 6: If applicable, the UE may send model performance feedback to the RAN node.

[0226] Step 7: The UE and RAN nodes perform actions based on the output data.

[0227] Step 8: The UE transmits feedback information to the RAN node.

[0228] CPU occupation rule for AI / ML based CSI report

[0229] Due to the advancement of computational processing technology and AI (artificial intelligence) / ML (machine learning) technology, the nodes and terminals that make up the wireless communication network are becoming more intelligent / advanced. In particular, due to the intelligence of the network, it is expected that various network decision parameter values ​​(e.g., transmission / reception power of each base station, transmission power of each terminal, precoder / beam of base station / terminal, time / frequency resource allocation for each terminal, duplex method of each base station, etc.) can be quickly optimized and derived / applied according to various network environment parameters (e.g., distribution / location of base stations, distribution / location / material of buildings / furniture, etc., location / movement direction / speed of terminals, climate information, etc.).

[0230] Below, we describe proposals for a CPU (CSI processing unit) for performing AI / ML-based CSI reporting in evolved networks / base stations / terminals.

[0231] Figure 18 is a diagram for explaining AI-based CSI feedback.

[0232] To reduce CSI feedback payload overhead, a common AI model can be based on compression at the terminal and / or base station. For example, a terminal-side CSI encoder and a base station-side CSI decoder can be assumed, as shown in Figure 18. The encoder and decoder can be associated with an autoencoder, and a convolutional neural network (CNN) can be used to efficiently reduce the channel size / dimension.

[0233] When it comes to AI / ML-based payload reduction, determining what to compress is a critical factor. Broadly speaking, there are two types of channel feedback: explicit and implicit. In LTE and NR, implicit feedback is used, reporting RI / CQI / PMI to the base station rather than the raw channel matrix or channel covariance matrix. Implicit feedback has the advantage of lower feedback overhead compared to explicit feedback. When AI / ML is used for CSI payload reduction, compressed raw channel information can be fed back to the gNB, which can then be used as explicit feedback.

[0234] Another example of AI / ML-based CSI feedback is compressing existing PMI, such as Type I or Type II CSI. In Rel-16 / 17 Type II CSI, linear-based (i.e., DFT vectors) frequency-domain compression is used to reduce the payload of Rel-15 Type II CSI. Similarly, AI / ML can be used to further reduce the payload of existing CSI and improve performance.

[0235] As described above, AI / ML-based CSI reporting can be considered. Assuming both the base station and the UE have AI / ML capabilities (training and / or inference), auto-encoder-like CSI reporting can be performed. Alternatively, single-sided (at the network or UE) AI / ML can be considered. In single-sided cases, existing legacy codebooks can be optimized for specific parameters (performance and / or payload) based on AI / ML.

[0236] To manage the terminal's CSI processor and other components during (AI / ML-based) CSI generation / calculation, NR Rel. 17 introduced a CSI processing unit (CPU). Table 6 is an excerpt from the standard document TS 38.214 regarding the CSI processing unit.

[0237]

[0238]

[0239]

[0240] Proposal 1

[0241] In the case of AI / ML-based CSI reporting, when the terminal generates CSI through the inference operation of the AI / ML model, CPU occupation is assumed to be X, and the X value is a value that the base station configures for the terminal or a value agreed upon in advance.

[0242] In the case of AI / ML-based CSI reporting, it can be largely classified into two types - Two-sided AI / ML and one-sided AI / ML. In both cases, it can be assumed that the base station (or network AI / ML server) performs training / inference, and the trained AI / ML model is signaled to the terminal, so that the terminal operates only with AI / ML inference. In addition, the base station (or network AI / ML server), which is the subject of training, can set / indicate the corresponding value based on the computational complexity according to inference based on FLOP (floating operation) required for AI / ML inference and / or the capability report of the terminal's CSI processing / memory. If the corresponding value X is indicated, the terminal regards the CPU occupation set / indicated by the base station as X (e.g., O) when calculating the AI / ML-based CSI. CPU =X), follows the behavior for CPU occupancy of the terminal as described in TS38.214 in Table 6 above.

[0243] AI / ML inference processes multiply variables between nodes within a pre-trained model and apply activation functions (e.g., sigmoid, relu, tanh) for each node without the burden of training. Therefore, the complexity of the inference operation is significantly lower than that of training. However, to achieve desired performance improvements, deeper and larger models are considered. In such cases, computational complexity increases due to the increasing number of variables. Therefore, CPU occupation may vary depending on the amount of computation associated with AI / ML model parameters such as the number of variables, number of nodes, and number of hidden layers, and / or the amount of computation associated with pre-processing / post-processing of input / output data. Furthermore, AI / ML models are highly scalable and can be composed of different AI / ML models depending on their use cases. To evenly reflect this diversity of AI / ML models, we propose the following.

[0244] Proposal 1-1

[0245] The terminal can report (as capability) information about the AI / ML model that can be processed by 1 CPU occupation to the base station.

[0246] The information regarding the AI / ML model that can be processed by 1 CPU occupation in the above proposal 1-1 may include some or all of the number of variables, number of nodes, and number of hidden layers. Alternatively, a value based on a specific computational unit (e.g., FLOP) that can be processed by 1 CPU occupation may be reported.

[0247] In another embodiment, the base station can instruct / set a threshold that can be processed by 1 CPU occupation to the terminal. Since the above value may vary depending on the performance of the terminal, a relatively loose value can be instructed, or a method may be used to divide the terminal into multiple groups (e.g., high, middle, low) based on the capability of the terminal and instruct a specific threshold value for each group. Alternatively, the base station can set or pre-arrange the CPU required for CSI reporting based on a specific AI / ML model for each group and instruct / set it to the terminal. For example, in the above proposal 1, the value X can be set or pre-arranged as shown in Table 1 below. The terminal can report the capability for AI / ML.

[0248] Table 7 shows an example of setting X values ​​based on terminal capability.

[0249] Group idX1 - High capable UE12 - Middle capable UE23 - low capable UE3

[0250] Alternatively, if a specific AI / ML model or version of the model is specified, the terminal can report the capability for each model / version.

[0251] Proposal 1-2

[0252] By introducing an APU (AI / ML processing unit) separately from the CPU, the terminal can effectively manage the processing of multiple AI / ML models.

[0253] In the above proposal 1-2, the APU can report the number of AI / ML modules / functions (A) that the terminal can compute simultaneously per CC or across CCs. If B APUs are occupied, the terminal can be assumed to have AB unoccupied APUs. In this situation, if more than AB AI / ML models are activated or triggered, the AB selected AI / ML models are selected and operated according to the preset priority rule, and information about the unselected AI / ML models can be reported to the base station. In the case of the above APU, it can be assumed to be occupied from the time of receiving signaling (e.g., MAC-CE / DCI) for triggering / activating a specific AI / ML model to the time of receiving deactivation, termination, or switching signaling.

[0254] Alternatively, the above APU may be managed / applied in an integrated manner to the AI / ML model, and the number of APUs required for each AI / ML model may differ.

[0255] As another example, the APU can be considered as being specifically configured only for AI / ML-based CSI reporting, and thus the terminal can simultaneously report the maximum supportable APU and CPU values ​​as capabilities. In this case, the occupation rule for the APU is the same as that for the CPU, but is used to separately manage only AI / ML-based CSI reporting. In the above description, the APU occupation rule being the same as that for the CPU can be assumed to mean that a single CSI report associated with a single CSI-RS resource occupies K APUs (e.g., K=1), similar to legacy. Alternatively, since it can be utilized across multiple CSI resources, such as CSI prediction of the terminal, the K value can have a value such as 0.5, which can be assumed to mean that a single CSI report associated with two resources occupies one APU.

[0256] For the above APU, AI / ML inference can be used as the main use case. However, if the terminal has training capability, units that manage training and inference separately can be introduced, for example, APU-training and APU-inference can be managed respectively. In this case, if the terminal's APU-inference is fully occupied and some or all of the APU-training is unoccupied, the APU-training can be borrowed and used according to a specific rule. Conversely, if the terminal's APU-training is fully occupied and some or all of the APU-inference is unoccupied, the APU-inference can be borrowed and used according to a specific rule. The specific rule may be a rule for the interchangeable use of APU-training and APU-inference. For example, each APU occupancy can be given by the following function (e.g., a linear linear equation or a polynomial equation), and an example of a linear linear equation is as follows. w *APU-inference + b = APU-training Here, w is a coefficient to match the capabilities of APU-training and APU-inference, and b is an offset. For example, w=2, b=0 can be interpreted to mean that the processing required for APU-training is twice that of APU-inference.

[0257] Meanwhile, if APU-training / APU-inference are integrated and operated / applied as a single APU without distinction, when there is a shortage of APUs, training takes priority over inference and occupies the APU (or it is possible to return the APU already occupied by inference and use it for training) (or vice versa), and in cases where inference is not possible due to a shortage of APUs, a default mode can be set, for example, in the case of CSI reporting, it can fallback to Type I CSI reporting and default mode. Alternatively, if inference is not possible due to a shortage of APUs, the AI / ML model inference can be dropped or delayed to apply inference. Here, delay means waiting until an unoccupied APU equal to or larger than the APU occupied by the AI / ML model is secured. However, if the delay becomes too long, problems such as buffer size / memory size may occur, so a maximum window size or timer that can be delayed can be introduced, and the window or timer can be waited for to expire, and if an unoccupied APU is secured before that, the AI / ML inference can be applied, and if not, the model can be dropped. Information about the drop or delay of the AI / ML model can be reported to the base station or AI / ML server.

[0258] Proposals 1-3

[0259] If the current APU occupation value of the terminal is 0 (i.e., L=0) and the base station triggers training / update / inference of the AI / ML model, the number of APUs (N) that the terminal can process simultaneously per CC or across CCs for fast training / update / inference of the AI / ML model APU : The UE indicates the number of supported simultaneous AI / ML model training / update / inference N APU ) is considered to occupy as much APU occupation as O. That is, in the case of fast training / update / inference of a specific AI / ML model, APU occupation is O. APU =N APU .

[0260] For Proposals 1-3 above, the base station can configure / instruct all APU occupations for fast training / update / inference by occupying all APUs. Alternatively, if a priority exists for a specific AI / ML model, Proposals 1-3 can be guaranteed to be applied only when training / update / inference of the AI / ML model corresponding to the highest priority is triggered.

[0261] Proposal 2

[0262] When AI / ML functions are modularized and multiple (e.g., Y) AI / ML models participate in a CSI report and perform a single CSI report, CPU occupation is considered as Y or Y+1.

[0263] Typically, CSI reporting is performed by performing channel measurement (CMR) and / or interference measurement (IMR) based on specific RS (e.g., CSI-RS or SSB) set by the base station, acquiring raw channel data, performing pre-processing (e.g., eigen value decomposition) on this data as needed, and calculating preferred CRI / RI / CQI / PMI or SSBRI / CRI / L1-RSRP / L1-SINR, etc. based on this. This series of processes can be processed by a single AI / ML, or divided into multiple modules, and modules / functions with high computational complexity or modules / functions that can be replaced with moderate performance with low complexity can be configured by replacing them with AI / ML models. Therefore, the number of modules (Y) participating in this modularized AI / ML-based CSI reporting is regarded as CPU occupation, or if it is mixed with other legacy modules / functions, 1 is added to count this and regarded as Y+1. Even if a specific module (e.g., channel estimation) is implemented as a terminal implementation, the terminal can report information (e.g., number of modules or model information) to the base station, which is for the scheduling flexibility of the base station.

[0264] Proposal 3

[0265] If the current CPU occupation value of the terminal is 0 (i.e., L=0) and the base station triggers AI / ML-based CSI reporting, the number of CPUs (N) that the terminal can process simultaneously per CC or across CCs for fast AI / ML-based CSI reporting CPU: The UE indicates the number of supported simultaneous CSI calculations N CPU It is considered that the CPU occupation is occupied by parameters simultaneousCSI-ReportsPerCCin a component carrier, and simultaneousCSI-ReportsAllCCacross all component carriers. In other words, CPU occupation is O when reporting AI / ML-based CSI. CPU =N CPU .

[0266] For AI / ML-based CSI calculations, if the AI / ML model performing inference is not very large, the computational effort may be small, enabling faster CSI reporting. In such cases, particularly for two-sided models where inference is performed on both the base station and the terminal, all available CPUs can be occupied for AI / ML-based CSI reporting on the terminal, as in Proposal 3, to achieve faster CSI reporting.

[0267] As another example, in the above proposal 3, N CPU You can also manage CPU by setting AI / ML based CSI reporting and legacy CSI reporting separately. For example, N CPU_AI By introducing separately, the above N CPU_AI It is used only for AI / ML based CSI reporting, or when a certain number of AI / ML based CSI reports are triggered (e.g., 1). CPU Instead N CPU_AI It can be understood as a substitute for N. Here, N CPU_AImay be a value determined based on UE capability reporting. Or, the base station may have multiple N CPU can define and instruct the terminal which value to use. This is done based on which CSI report the base station schedules. CPU It is used for the purpose of setting / instructing the value to be written, and as a result, it can increase the flexibility of the base station and terminal. Or, the N of the terminal CPU_AI It can be applied to AI / ML model (group)-specifically by setting multiple values. For example, N CPU_AI-1 , N CPU_AI-2 Apply two, and if there are multiple AI / ML models, each N CPU_AI-1 , N CPU_AI-2 It can be used by mapping to. That is, in the case of AI / ML models belonging to group 1, CPU occupation is N CPU_AI-1 is managed separately as an upper limit value, and for AI / ML models belonging to group 2, CPU occupation is N CPU_AI-2 can be managed separately as an upper limit value. Information on the above grouping can be based on the capability report of the terminal or set / instructed by the base station.

[0268] Proposal 4

[0269] Base stations can improve base station scheduling flexibility by additionally signaling a CSI priority indicator to adjust the CSI priority of AI / ML-based CSI reporting.

[0270] As described in Table 8 below, when multiple CSI reports collide in time and / or frequency resources, or when multiple PUCCHs and / or PUSCHs carrying CSI are multiplexed, a specific CSI report can be dropped according to the priority rule below. In the case of the CSI priority rule, the CSI report with a lower value of PriiCSI (y, k, c, s) is given priority according to the rules of the standard document extracted in Table 8 below. By utilizing this structure, in order for the base station to adjust the priority of AI / ML-based CSI reports, a 1-bit field can be added to the DCI that schedules the CSI report to indicate high or low priority. For example, priority adjustment can be indicated as k=0 or 1 in Equation 2 below, indicating a higher or equal priority than the existing normal CSI report (e.g., Type 1 CSI, Type 2 CSI). As another example, the range of values ​​for k in Equation 2 below can be increased to 0, 1, and 2, allowing for more flexible instructions. In this case, the field size based on priority would increase to 2 bits. The instructions in Proposal 4 above can be given differently for each AI / ML model.

[0271]

[0272] Alternatively, AI / ML-based CSI reporting can be instructed to set k=0 so that AI / ML-based CSI reporting has the same priority as beam reporting.

[0273]

[0274] Proposal 5

[0275] To support fast CSI reporting in AI / ML model inference-based CSI reporting of the terminal, a value smaller than the previously defined CSI calculation latency (e.g. Z, Z', CSI reference resource timing) is defined / set.

[0276] In the case of the above proposal 5, the purpose is to support cases where model training is completed and the terminal performs AI / ML model inference to require fast CSI reporting. As in the above proposal, a table with values ​​smaller than table 5.4-1 or table 5.4-2 of TS 38.214, as extracted in Table 9 below, can be defined and used for AI / ML-based CSI reporting. Alternatively, the Z / Z' value of table 5.4-1 can be used for AI / ML-based CSI reporting. Extending the above proposal 5, there is also a method of defining and using Z / Z' according to a specific AI / ML model or model group. The model group may mean a group of AI / ML models for which CSI calculation can be performed within a specific Z / Z' value, and when the terminal reports information related to the Z / Z' to the capability, the base station can configure which Z / Z' value to use based on the value.

[0277]

[0278]

[0279]

[0280] As described above, when a network / base station utilizing AI technology or a network / base station with advanced performance (via other means) is implemented, it is expected that dynamic optimization of interference / traffic, etc. will be possible through real-time cooperation and / or integrated control between base stations. Accordingly, a suitable resource control and CSI-related signaling method are proposed. While the present invention is based on the 3GPP NR system in terms of the current system structure, this is not a limitation and the invention can be applied to other communication systems.

[0281] [APU occupation for inactive AI / ML model(s)]

[0282] When an APU (AI / ML model processing unit) is introduced to effectively manage AI / ML models, occupation rules are proposed to effectively manage it. In particular, we propose an APU management method for inactive AI / ML model(s) / functionality when monitoring models for rapid activation / switching / update of inactive models. Inactive AI / ML models can refer to single or multiple AI / ML models registered / deployed or implemented on a terminal (or network) that are not currently active (e.g., have received an activation command or are currently performing operations such as AI / ML inference / monitoring / training).

[0283] Proposal 6

[0284] Signaling of information about whether monitoring is performed on a per model(s) / per model group / per AI / ML functionality basis in an inactive model may be based on at least some of the following embodiments:

[0285] 1) Information regarding the availability of monitoring for inactive AI / ML model(s) can be reported through terminal capability reports, etc. For example, this information can be reported by AI / ML model, AI / ML model group, or AI / ML functionality. Below is an example of capability reporting by AI / ML model.

[0286] - AI / ML Model #1 - capable on monitoring at inactive model state

[0287] - AI / ML Model #2 - capable on monitoring at inactive model state

[0288] - AI / ML Model #3 - incapable on monitoring at inactive model state

[0289] 2) In another embodiment, based on the capability report of the terminal, the base station can instruct a higher layer (e.g., RRC / MAC-CE / DCI or a combination of the above signaling) to perform monitoring on the AI / ML model, AI / ML model group, or AI / ML functionality even when the base station is in an inactive model state. Alternatively, active / inactive / idle states can be defined for the AI / ML model as follows, and whether or not to monitor can be instructed based on these states.

[0290] i) AI / ML Model / functionality active state: The AI / ML Model / functionality activation command has been received, or the current state is performing operations such as AI / ML inference / monitoring / training.

[0291] ii) AI / ML Model / functionality inactive state: The AI / ML Model / functionality activation command has not been received, or the AI / ML model inference / monitoring / training operation is not currently being performed, but monitoring information can be reported (based on base station instructions / settings).

[0292] iii) AI / ML Model / functionality idle state: A state in which the AI / ML Model / functionality activation command has not been received, or the AI / ML inference / monitoring / training operation is not currently being performed, and monitoring information cannot be reported (based on base station instructions / settings).

[0293] Likewise, AI / ML model monitoring can be performed even when the AI / ML model is inactive, in which case the inactive AI / ML model must occupy the APU. The priority for APU occupancy of AI / ML model(s) / model group / functionality can be in the following order: i) AI / ML Model / functionality active state > ii) AI / ML Model / functionality inactive state > iii) AI / ML Model / functionality idle state, and within each state, the priority can be determined for each model by specific rules (e.g., trigger order, lowest / highest model ID order, most recent inference / monitoring order).

[0294] In this case, the time of APU occupancy may be ambiguous, and the following proposal 7 is proposed for APU occupancy duration.

[0295] Proposal 7

[0296] The APU occupancy duration of inactive AI / ML model(s) / model group / AI / ML functionality can be controlled in the following two ways.

[0297] i) Method 1: When the base station commands / controls monitor / assess for (specific) Inactive model(s) / model group / AI / ML functionality, the APU can be occupied from the time the base station commands to start model monitoring / assessment until the time the terminal reports the result.

[0298] ii) Method 2: If the terminal determines and performs monitor / assess on (specific) inactive model(s) / model group / AI / ML functionality, the terminal can perform monitor / access on the inactive model with best effort by utilizing the surplus APU.

[0299] Method 1 is when the base station commands / controls model monitoring / assessment for inactive model(s) / model group / AI / ML functionality. In this case, from time n (or n+X) when the base station receives a command (e.g., command via MAC-CE / DCI) instructing monitoring / assessment for the inactive model / model group / AI / ML functionality, the terminal may occupy the APU until time m (or time n+Y) when the terminal performs monitoring / assessment for the related model and reports the report value thereto to the base station. The above time points may be slot level or symbol level, and the X and Y values ​​may be based on the terminal capability (considering terminal processing time, etc.) or may be values ​​set by the base station or agreed upon in advance.

[0300] Alternatively, the terminal may only perform monitoring / assessment on the model(s) and not report the results, or may collect multiple results and report them to the base station at once. In this case, a specific timer or a specified time such as a monitoring duration may be defined. For example, it may be defined / configured to occupy the APU for a timer expiration or a predefined / configured time Z from the time n (or n+X) of receiving a command related to AI / ML monitoring / assessment.

[0301] Method 2 is when the terminal determines to perform model monitoring / assessment for inactive model(s) / model group / AI / ML functionality. In this case, the terminal performs AI / ML model monitoring / assessment with best effort by utilizing the remaining APU (surplus APU) among the total APUs based on the terminal's capability for model monitoring / assessment for inactive model / model group / AI / ML functionality. If the base station triggers a (new / additional) AI / ML-based operation to the terminal during the model monitoring / assessment operation for inactive model / model group / AI / ML functionality, the terminal may have to stop / hold / suspend the model monitoring / assessment operation for the currently in progress inactive model(s) / model group / AI / ML functionality and give priority to the AI / ML-based operation triggered by the base station. In this case, the spare APU of the terminal is occupied by the (active) AI / ML model / group / functionality (newly / additionally) triggered by the base station. As described above, in this case too, the active AI / ML model takes precedence over the inactive AI / ML model and occupies the APU.In another embodiment, if the base station triggers an AI / ML-based operation while the terminal is monitoring / assessing inactive AI / ML model(s) / model group / functionality based on the surplus APU, the terminal may determine to ignore the base station trigger and report that it has ignored it, or may not apply the trigger immediately but may apply it with a delay and report information about the delay to the base station. For example, if the base station triggers a (new / additional) AI / ML-based operation when the APU occupy time for monitoring / assessment of inactive AI / ML model(s) / model group / functionality is almost over, it may be inefficient to drop the processing done for monitoring / assessment of the existing inactive model(s), so the trigger may be ignored and the existing operation may be completed.

[0302] In Method 2, when reporting monitoring / assessment results for inactive AI / ML model(s) / model group / functionality to the base station periodically / semi-periodically / aperiodically, monitoring / assessment may not be possible for some or all models due to a lack of surplus APUs. To this end, when reporting the monitoring / assessment results for inactive AI / ML model(s) / model group / functionality, the inactive model(s) / model group / functionality for which monitoring / assessment was performed and the inactive model(s) / model group / functionality for which monitoring / assessment was not performed are reported separately, and if monitoring / assessment is not possible, the reason (e.g., APU shortage) can be additionally reported. Alternatively, monitoring / assessment for the inactive model(s) / model group / functionality can be reported all at once to MAC-CE after completion.

[0303] For smooth operation of Method 2, the terminal can report information on the currently occupied APUs (e.g., number of occupied APUs, time points) to the base station. The above reporting may include periodic / semi-periodic / aperiodic reporting and event-based reporting. For example, for event-based reporting, a specific threshold value for the monitoring value may be predefined or set by the base station. Alternatively, the terminal may report information on the number of currently available APUs in addition to information on occupied APUs. In addition, when reporting on the occupied / unoccupied APUs, information on (expected) occupied / unoccupied APUs for each time point may also be reported. For example, {number of remaining APUs at t1, number of remaining APUs at t2, number of remaining APUs at t3, ...} may be reported as a set. Information on the corresponding time points t1, t2, t3, ... may be predefined or set by the base station. Alternatively, the terminal may additionally report information about those points in time to the base station, such as start time / duration (offset) values.

[0304] Proposal 8

[0305] When performing monitoring / assessment on inactive AI / ML model(s) / model group / functionality as described in Proposal 6 / 7, if multiple candidate AI / ML model(s) / model group / functionality are subject to monitoring / assessment, it is necessary to prioritize them.

[0306] The monitoring / assessment priorities of multiple inactive AI / ML model(s) / model groups / AI / ML functionality can be determined based on at least one of the following three methods.

[0307] i) Method 1: Priority based on UE decision

[0308] ii) Method 2: Priority based on network (NW) instructions

[0309] iii) Method 3: Priority according to priority rules

[0310] In Method 1, the terminal sets priorities, occupies the APU, and performs monitoring. To indicate which models it has decided to monitor, the terminal can explicitly report model / model group / functionality IDs to the base station. Such reports can be transmitted alone or with additional information (e.g., monitoring results, number of occupied / unoccupied APUs).

[0311] For Method 2, the base station can specify which model(s) / model group / functionality the NW will monitor. For example, the base station can semi-statically configure this using RRC signaling, or dynamically configure it using MAC-CE / DCI, etc., and this can be transmitted to the terminal along with monitoring triggering information.

[0312] For method 3, the priorities of inactive AI / ML model(s) / model groups / AI / ML functionality can be determined based on specific priority rules. For example, priorities can be assigned to specific AI / ML use cases. For example, mobility-related use cases and / or beam management (BM) use cases can occupy the APU with higher priority than other use cases (e.g., CSI feedback, positioning). Furthermore, for the same use case, priority rules can be defined based on the model-specific requirements / performance of specific AI / ML model(s) / model groups / functionality. For example, in a CSI use case, if model #A provides higher accuracy (or higher compression ratio, higher accuracy) than model #B, a priority rule can be defined to assign model #A a higher priority than model #B.

[0313] For example, based on method 3, priorities are determined for inactive model(s) / model group / AI / ML functionality, and the terminal can perform model monitoring / assessment by applying APU occupancy rules according to the priorities. Meanwhile, the predefined / set priorities may be changed in real time / dynamically based on whether the terminal has recently performed monitoring / assessment. For example, the terminal can fairly monitor / assess inactive models by giving a higher priority to inactive model(s) / model group / AI / ML functionality that has not been monitored / assessed recently. Although the above priority determination method is described for inactive model(s) / functionality, the same can be applied to active model(s) / functionality.

[0314] In addition, if the first inactive model(s) / model group / AI / ML functionality(s) with a higher priority based on the priority determination in Method 3 is triggered later than the second inactive model(s) / model group / AI / ML functionality(s) with a lower priority, ambiguity may arise in how the terminal should operate. If there is an APU occupied by the second inactive model(s) / model group / AI / ML functionality(s), even if the first inactive model(s) / model group / AI / ML functionality with a higher priority is triggered, the terminal may not drop the monitoring / assessment of the second inactive model(s) / model group / AI / ML functionality(s). In this case, the terminal delays the newly triggered first inactive model(s) / model group / AI / ML functionality, etc., and performs the monitoring / assessment as soon as the APU is secured. Alternatively, when the first inactive model(s) / model group / AI / ML functionality(s) with a high priority is triggered, the terminal may drop the second inactive model(s) / model group / AI / ML functionality(s) in order of lower priority to secure APU, and then perform monitoring / assessment of the first inactive model(s) / model group / AI / ML functionality with a high priority.

[0315] Proposal 9

[0316] When inactive AI / ML model(s) / model group / AI / ML functionality are managed with a common APU with active AI / ML model(s) / model group / AI / ML functionality, the APU occupancy for monitoring / assessment of the inactive model(s) / model group / AI / ML functionality may occupy fewer APUs than the active case. For example, if one active AI / ML model(s) / functionality occupies 1 APU, the inactive AI / ML model(s) / functionality may occupy α APUs (α <= 1). For example, the value of α may be determined based on terminal capability or base station configuration / pre-arrangement.

[0317] This is because the processing capabilities required for monitoring inactive model(s) / model group / AI / ML functionality can be implemented / operated more simply than for active models. Alternatively, rather than a common APU for active / inactive AI / ML models / functionality, separate APUs can be managed for each active or inactive AI / ML model(s) / functionality. For example, a terminal can be managed by an APU. active , APU inactive You can also manage APUs by distinguishing them.

[0318] The APU occupancy methods proposed above can be extended to active models as well as inactive models. For example, the APU occupancy timing can be as follows. From time n (or n+X) when the base station receives an activation command (via MAC-CE / DCI) indicating training / inference / monitoring of an AI / ML model / model group / AI / ML functionality, the terminal performs training / inference / monitoring for the same, and the AI / ML model / model group / AI / ML functionality can occupy the APU until time m (or n+Y) when the base station receives a deactivation / switching / update command for the same model. In the case of model monitoring, the APU can be occupied until time m when the terminal transmits a report value thereto to the base station (or time n+Y) when the base station receives a response from the report. In the case of the above point in time, it can be a slot level or a symbol level, and the X and Y values ​​can be based on the terminal capability (considering terminal processing time, etc.) or can be values ​​set by the base station or agreed upon in advance.

[0319] As described above, if a network / base station utilizing AI / ML technology or a network / base station with advanced performance (via other means) is implemented, dynamic optimization of interference / traffic, etc. can be achieved through real-time cooperation and / or integrated control between base stations. This specification proposes suitable control and AI / ML-related signaling methods. For convenience, the description is based on the 3GPP system, but is not limited thereto, and the embodiments can be applied to various other communication systems. Furthermore, each of Proposals 1 to 9 can be applied singly or in combination.

[0320] FIG. 19 is a diagram for explaining network and terminal operations in a wireless communication system according to one embodiment.

[0321] Referring to Figure 19, the network can perform at least one upper-layer signaling operation on the terminal (A05). The upper-layer signaling operation may include configuration of AI / ML (artificial intelligence / machine learning) models. The terminal can configure the AI / ML models based on the upper-layer signaling operation.

[0322] The terminal can activate at least one of the configured AI / ML models (A10). For example, activation / deactivation of the AI / ML model may be performed based on network instructions such as MAC or DCI. For convenience of explanation, the deactivated AI / ML model will be referred to as the first AI / ML model, and the activated AI / ML model will be referred to as the second AI / ML model.

[0323] The terminal can monitor at least one of the configured AI / ML models (A15).

[0324] The AI / ML model monitored by the terminal may include a first AI / ML model. The first AI / ML model may be determined by network instructions or by the terminal itself.

[0325] The AI / ML model monitored by the terminal may include a second AI / ML model.

[0326] Monitoring of an inactive first AI / ML model may occupy a smaller number of APU(s) or for a shorter period of time than monitoring of an active second AI / ML model.

[0327] Monitoring of the first AI / ML model can be performed based on the remaining APUs not occupied by the second AI / ML model that is active.

[0328] The second AI / ML model in the above active state can occupy the APU with a higher priority than the first AI / ML model in the above inactive state.

[0329] The monitoring of the second AI / ML model in the active state occupies 1 APU, and the monitoring of the first AI / ML model in the inactive state occupies n APU, where n can be a real number greater than 0 and less than or equal to 1.

[0330] The above n may be determined based on terminal capability or provided through network signaling.

[0331] The terminal can transmit the monitoring results of the AI / ML model to the network (A20).

[0332] Based on the monitoring results of the AI / ML model, the terminal and network can perform the LCM procedure for the AI / ML model (A25).

[0333] The start of the time interval in which the APU is occupied by monitoring of the first AI / ML model can be determined based on the point in time at which monitoring of the first AI / ML model is instructed.

[0334] The end of the time interval in which the APU is occupied by monitoring of the first AI / ML model may be determined based on at least one of: i) the time at which the result of monitoring of the first AI / ML model is reported (A20), ii) the time at which a network response to the report (A20) is received, or iii) the time at which a preset timer expires.

[0335] Meanwhile, the terminal can report information about the APU occupied or unoccupied state to the network.

[0336] FIG. 20 illustrates a flow of a method performed by a terminal according to one embodiment.

[0337] Referring to FIG. 20, the terminal can obtain settings for AI / ML (artificial intelligence / machine learning) models based on upper layer signaling (B05).

[0338] The terminal can monitor at least one of the above AI / ML models (B10).

[0339] The terminal can perform AI / ML model management for switching or updating based on monitoring of at least one AI / ML model (B15).

[0340] The monitoring of the at least one AI / ML model may include at least one of monitoring a first AI / ML model that is inactive and monitoring a second AI / ML model that is active.

[0341] Monitoring of the first AI / ML model in the inactive state may occupy a smaller number of terminal processing units or for a shorter period of time than monitoring of the second AI / ML model in the active state.

[0342] Among the AI / ML models that are inactive, the first AI / ML model to be monitored can be determined based on network signaling.

[0343] The start of the time interval in which the terminal processing unit is occupied by monitoring of the first AI / ML model may be determined based on the time at which the network signaling is received.

[0344] The end of the time interval during which the terminal processing unit is occupied by monitoring of the first AI / ML model may be determined based on at least one of: i) a time point at which the result of monitoring of the first AI / ML model is reported, ii) a time point at which a network response to the report is received, or iii) a time point at which a preset timer expires.

[0345] Among the AI / ML models that are inactive, the first AI / ML model to be monitored can be determined by the terminal.

[0346] Monitoring of the inactive AI / ML model may be stopped or suspended based on whether an action on the active second AI / ML model is triggered via network signaling.

[0347] The terminal can report information about the occupied or unoccupied state of the terminal processing unit to the network.

[0348] Monitoring of the first AI / ML model can be performed based on the remaining terminal processing units that are not occupied by the second AI / ML model that is in an active state.

[0349] The second AI / ML model in the active state can occupy the terminal processing unit with a higher priority than the first AI / ML model in the inactive state.

[0350] Monitoring of the second AI / ML model in the active state occupies 1 terminal processing unit, and monitoring of the first AI / ML model in the inactive state occupies n terminal processing units, where n may be a real number greater than 0 and less than or equal to 1.

[0351] The above n may be determined based on terminal capability or provided through network signaling.

[0352] Figure 21 illustrates a communication system (1) to which the present invention can be applied.

[0353] Referring to FIG. 21, a communication system (1) includes a wireless device, a base station, and a network. Here, the wireless device refers to a device that performs communication using a wireless access technology (e.g., 5G NR (New RAT), LTE (Long Term Evolution)) and may be referred to as a communication / wireless / 5G device. Although not limited thereto, the wireless device may include a robot (100a), a vehicle (100b-1, 100b-2), an XR (eXtended Reality) device (100c), a hand-held device (100d), a home appliance (100e), an IoT (Internet of Things) device (100f), and an AI device / server (400). For example, the vehicle may include a vehicle equipped with a wireless communication function, an autonomous vehicle, a vehicle capable of performing vehicle-to-vehicle communication, etc. Here, the vehicle may include an Unmanned Aerial Vehicle (UAV) (e.g., a drone). XR devices include AR (Augmented Reality) / VR (Virtual Reality) / MR (Mixed Reality) devices, and can be implemented in the form of HMD (Head-Mounted Device), HUD (Head-Up Display) installed in a vehicle, television, smartphone, computer, wearable device, home appliance, digital signage, vehicle, robot, etc. Mobile devices can include smartphone, smart pad, wearable device (e.g., smart watch, smart glass), computer (e.g., laptop, etc.), etc. Home appliances can include TV, refrigerator, washing machine, etc. IoT devices can include sensors, smart meters, etc. For example, base stations and networks can also be implemented as wireless devices, and a specific wireless device (200a) can act as a base station / network node to other wireless devices.

[0354] Wireless devices (100a to 100f) can be connected to a network (300) via a base station (200). Artificial Intelligence (AI) technology can be applied to the wireless devices (100a to 100f), and the wireless devices (100a to 100f) can be connected to an AI server (400) via the network (300). The network (300) can be configured using a 3G network, a 4G (e.g., LTE) network, a 5G (e.g., NR) network, etc. The wireless devices (100a to 100f) can communicate with each other via the base station (200) / network (300), but can also communicate directly (e.g., sidelink communication) without going through the base station / network. For example, vehicles (100b-1, 100b-2) can communicate directly (e.g., V2V (Vehicle to Vehicle) / V2X (Vehicle to Everything) communication). In addition, IoT devices (e.g., sensors) can communicate directly with other IoT devices (e.g., sensors) or other wireless devices (100a to 100f).

[0355] Wireless communication / connection (150a, 150b, 150c) can be established between wireless devices (100a~100f) / base stations (200), and base stations (200) / base stations (200). Here, wireless communication / connection can be achieved through various wireless access technologies (e.g., 5G NR) such as uplink / downlink communication (150a), sidelink communication (150b) (or D2D communication), and communication between base stations (150c) (e.g., relay, IAB (Integrated Access Backhaul). Through wireless communication / connection (150a, 150b, 150c), wireless devices and base stations / wireless devices, and base stations and base stations can transmit / receive wireless signals to each other. For example, wireless communication / connection (150a, 150b, 150c) can transmit / receive signals through various physical channels. To this end, at least some of various configuration information setting processes for transmitting / receiving wireless signals, various signal processing processes (e.g., channel encoding / decoding, modulation / demodulation, resource mapping / demapping, etc.), and resource allocation processes can be performed based on various proposals of the present invention.

[0356] Figure 22 illustrates a wireless device applicable to the present invention.

[0357] Referring to FIG. 22, the first wireless device (100) and the second wireless device (200) can transmit and receive wireless signals through various wireless access technologies (e.g., LTE, NR). Here, {the first wireless device (100), the second wireless device (200)} can correspond to {the wireless device (100x), the base station (200)} and / or {the wireless device (100x), the wireless device (100x)} of FIG. 21.

[0358] A first wireless device (100) includes one or more processors (102) and one or more memories (104), and may further include one or more transceivers (106) and / or one or more antennas (108). The processor (102) controls the memories (104) and / or the transceivers (106), and may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in this document. For example, the processor (102) may process information in the memory (104) to generate first information / signal, and then transmit a wireless signal including the first information / signal via the transceiver (106). In addition, the processor (102) may receive a wireless signal including second information / signal via the transceiver (106), and then store information obtained from signal processing of the second information / signal in the memory (104). The memory (104) may be connected to the processor (102) and may store various information related to the operation of the processor (102). For example, the memory (104) may perform some or all of the processes controlled by the processor (102), or may store software code including commands for performing the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this document. Here, the processor (102) and the memory (104) may be part of a communication modem / circuit / chip designed to implement a wireless communication technology (e.g., LTE, NR). The transceiver (106) may be connected to the processor (102) and may transmit and / or receive wireless signals via one or more antennas (108). The transceiver (106) may include a transmitter and / or a receiver. The transceiver (106) may be used interchangeably with an RF (Radio Frequency) unit. In the present invention, a wireless device may also mean a communication modem / circuit / chip.

[0359] The second wireless device (200) includes one or more processors (202), one or more memories (204), and may further include one or more transceivers (206) and / or one or more antennas (208). The processor (202) controls the memories (204) and / or the transceivers (206), and may be configured to implement the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in this document. For example, the processor (202) may process information in the memory (204) to generate third information / signals, and then transmit a wireless signal including the third information / signals via the transceivers (206). Furthermore, the processor (202) may receive a wireless signal including fourth information / signals via the transceivers (206), and then store information obtained from signal processing of the fourth information / signals in the memory (204). The memory (204) may be connected to the processor (202) and may store various information related to the operation of the processor (202). For example, the memory (204) may perform some or all of the processes controlled by the processor (202), or may store software code including commands for performing the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this document. Here, the processor (202) and the memory (204) may be part of a communication modem / circuit / chip designed to implement wireless communication technology (e.g., LTE, NR). The transceiver (206) may be connected to the processor (202) and may transmit and / or receive wireless signals via one or more antennas (208). The transceiver (206) may include a transmitter and / or a receiver. The transceiver (206) may be used interchangeably with an RF unit. In the present invention, a wireless device may also mean a communication modem / circuit / chip.

[0360] Hereinafter, the hardware elements of the wireless device (100, 200) will be described in more detail. Although not limited thereto, one or more protocol layers may be implemented by one or more processors (102, 202). For example, one or more processors (102, 202) may implement one or more layers (e.g., functional layers such as PHY, MAC, RLC, PDCP, RRC, SDAP). One or more processors (102, 202) may generate one or more Protocol Data Units (PDUs) and / or one or more Service Data Units (SDUs) according to the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this document. One or more processors (102, 202) may generate messages, control information, data, or information according to the descriptions, functions, procedures, proposals, methods, and / or operation flowcharts disclosed in this document. One or more processors (102, 202) can generate signals (e.g., baseband signals) including PDUs, SDUs, messages, control information, data or information according to the functions, procedures, proposals and / or methods disclosed herein, and provide the signals to one or more transceivers (106, 206). One or more processors (102, 202) can receive signals (e.g., baseband signals) from one or more transceivers (106, 206) and obtain PDUs, SDUs, messages, control information, data or information according to the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed herein.

[0361] One or more processors (102, 202) may be referred to as a controller, a microcontroller, a microprocessor, or a microcomputer. One or more processors (102, 202) may be implemented by hardware, firmware, software, or a combination thereof. For example, one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), one or more Digital Signal Processing Devices (DSPDs), one or more Programmable Logic Devices (PLDs), or one or more Field Programmable Gate Arrays (FPGAs) may be included in one or more processors (102, 202). The descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in this document may be implemented using firmware or software, and the firmware or software may be implemented to include modules, procedures, functions, etc. The descriptions, functions, procedures, suggestions, methods and / or operation flowcharts disclosed in this document may be implemented using firmware or software configured to perform one or more processors (102, 202) or stored in one or more memories (104, 204) and executed by one or more processors (102, 202). The descriptions, functions, procedures, suggestions, methods and / or operation flowcharts disclosed in this document may be implemented using firmware or software in the form of codes, instructions and / or sets of instructions.

[0362] One or more memories (104, 204) may be coupled to one or more processors (102, 202) and may store various forms of data, signals, messages, information, programs, codes, instructions, and / or commands. The one or more memories (104, 204) may be configured as ROM, RAM, EPROM, flash memory, hard drives, registers, cache memory, computer-readable storage media, and / or combinations thereof. The one or more memories (104, 204) may be located internally and / or externally to the one or more processors (102, 202). Additionally, the one or more memories (104, 204) may be coupled to the one or more processors (102, 202) via various technologies, such as wired or wireless connections.

[0363] One or more transceivers (106, 206) can transmit user data, control information, wireless signals / channels, etc., as mentioned in the methods and / or flowcharts of this document, to one or more other devices. One or more transceivers (106, 206) can receive user data, control information, wireless signals / channels, etc., as mentioned in the descriptions, functions, procedures, proposals, methods and / or flowcharts of this document, from one or more other devices. For example, one or more transceivers (106, 206) can be connected to one or more processors (102, 202) and can transmit and receive wireless signals. For example, one or more processors (102, 202) can control one or more transceivers (106, 206) to transmit user data, control information, or wireless signals to one or more other devices. Additionally, one or more processors (102, 202) may control one or more transceivers (106, 206) to receive user data, control information, or wireless signals from one or more other devices. Additionally, one or more transceivers (106, 206) may be coupled to one or more antennas (108, 208), and one or more transceivers (106, 206) may be configured to transmit and receive user data, control information, wireless signals / channels, or the like, as referred to in the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed herein, via one or more antennas (108, 208). In this document, one or more antennas may be multiple physical antennas or multiple logical antennas (e.g., antenna ports). One or more transceivers (106, 206) can convert received user data, control information, wireless signals / channels, etc. from RF band signals to baseband signals in order to process the received user data, control information, wireless signals / channels, etc. using one or more processors (102, 202).One or more transceivers (106, 206) may convert user data, control information, wireless signals / channels, etc. processed by one or more processors (102, 202) from baseband signals to RF band signals. For this purpose, one or more transceivers (106, 206) may include an (analog) oscillator and / or filter.

[0364] Figure 23 illustrates another example of a wireless device applicable to the present invention. The wireless device may be implemented in various forms depending on the use case / service (see Figure 21).

[0365] Referring to FIG. 23, the wireless device (100, 200) corresponds to the wireless device (100, 200) of FIG. 22 and may be composed of various elements, components, units / units, and / or modules. For example, the wireless device (100, 200) may include a communication unit (110), a control unit (120), a memory unit (130), and additional elements (140). The communication unit may include a communication circuit (112) and a transceiver(s) (114). For example, the communication circuit (112) may include one or more processors (102, 202) and / or one or more memories (104, 204) of FIG. 22. For example, the transceiver(s) (114) may include one or more transceivers (106, 206) and / or one or more antennas (108, 208) of FIG. 22. The control unit (120) is electrically connected to the communication unit (110), the memory unit (130), and the additional elements (140) and controls the overall operation of the wireless device. For example, the control unit (120) may control the electrical / mechanical operation of the wireless device based on the program / code / command / information stored in the memory unit (130). In addition, the control unit (120) may transmit information stored in the memory unit (130) to an external device (e.g., another communication device) via a wireless / wired interface through the communication unit (110), or store information received from an external device (e.g., another communication device) via a wireless / wired interface in the memory unit (130).

[0366] The additional element (140) may be configured in various ways depending on the type of the wireless device. For example, the additional element (140) may include at least one of a power unit / battery, an input / output unit (I / O unit), a driving unit, and a computing unit. Although not limited thereto, the wireless device may be implemented in the form of a robot (Fig. 21, 100a), a vehicle (Fig. 21, 100b-1, 100b-2), an XR device (Fig. 21, 100c), a portable device (Fig. 21, 100d), a home appliance (Fig. 21, 100e), an IoT device (Fig. 21, 100f), a digital broadcasting terminal, a hologram device, a public safety device, an MTC device, a medical device, a fintech device (or a financial device), a security device, a climate / environmental device, an AI server / device (Fig. 21, 400), a base station (Fig. 21, 200), a network node, etc. Wireless devices may be mobile or stationary depending on the use / service.

[0367] In FIG. 23, various elements, components, units / parts, and / or modules within the wireless device (100, 200) may be entirely interconnected via a wired interface, or at least some may be wirelessly connected via a communication unit (110). For example, within the wireless device (100, 200), the control unit (120) and the communication unit (110) may be wired, and the control unit (120) and a first unit (e.g., 130, 140) may be wirelessly connected via the communication unit (110). In addition, each element, component, unit / part, and / or module within the wireless device (100, 200) may further include one or more elements. For example, the control unit (120) may be composed of one or more processor sets. For example, the control unit (120) may be composed of a set of a communication control processor, an application processor, an electronic control unit (ECU), a graphics processing processor, a memory control processor, etc. As another example, the memory unit (130) may be composed of RAM (Random Access Memory), DRAM (Dynamic RAM), ROM (Read Only Memory), flash memory, volatile memory, non-volatile memory, and / or a combination thereof.

[0368] Figure 24 illustrates a vehicle or autonomous vehicle applicable to the present invention. The vehicle or autonomous vehicle may be implemented as a mobile robot, car, train, manned / unmanned aerial vehicle (AV), ship, etc.

[0369] Referring to FIG. 24, a vehicle or autonomous vehicle (100) may include an antenna unit (108), a communication unit (110), a control unit (120), a driving unit (140a), a power supply unit (140b), a sensor unit (140c), and an autonomous driving unit (140d). The antenna unit (108) may be configured as a part of the communication unit (110). Blocks 110 / 130 / 140a to 140d correspond to blocks 110 / 130 / 140 of FIG. 23, respectively.

[0370] The communication unit (110) can transmit and receive signals (e.g., data, control signals, etc.) with external devices such as other vehicles, base stations (e.g., base stations, road side units, etc.), and servers. The control unit (120) can control elements of the vehicle or autonomous vehicle (100) to perform various operations. The control unit (120) can include an ECU (Electronic Control Unit). The drive unit (140a) can drive the vehicle or autonomous vehicle (100) on the ground. The drive unit (140a) can include an engine, a motor, a power train, wheels, brakes, a steering device, etc. The power supply unit (140b) supplies power to the vehicle or autonomous vehicle (100) and can include a wired / wireless charging circuit, a battery, etc. The sensor unit (140c) can obtain vehicle status, surrounding environment information, user information, etc. The sensor unit (140c) may include an IMU (inertial measurement unit) sensor, a collision sensor, a wheel sensor, a speed sensor, an incline sensor, a weight detection sensor, a heading sensor, a position module, a vehicle forward / backward sensor, a battery sensor, a fuel sensor, a tire sensor, a steering sensor, a temperature sensor, a humidity sensor, an ultrasonic sensor, an illuminance sensor, a pedal position sensor, etc. The autonomous driving unit (140d) may implement a technology for maintaining a driving lane, a technology for automatically controlling speed such as adaptive cruise control, a technology for automatically driving along a set path, a technology for automatically setting a path and driving when a destination is set, etc.

[0371] For example, the communication unit (110) can receive map data, traffic information data, etc. from an external server. The autonomous driving unit (140d) can generate an autonomous driving route and driving plan based on the acquired data. The control unit (120) can control the drive unit (140a) so that the vehicle or autonomous vehicle (100) moves along the autonomous driving route according to the driving plan (e.g., speed / direction control). During autonomous driving, the communication unit (110) can irregularly / periodically acquire the latest traffic information data from an external server and can acquire surrounding traffic information data from surrounding vehicles. In addition, during autonomous driving, the sensor unit (140c) can acquire vehicle status and surrounding environment information. The autonomous driving unit (140d) can update the autonomous driving route and driving plan based on newly acquired data / information. The communication unit (110) can transmit information regarding the vehicle location, autonomous driving route, driving plan, etc. to the external server. External servers can predict traffic information data in advance using AI technology or other technologies based on information collected from vehicles or autonomous vehicles, and provide the predicted traffic information data to the vehicles or autonomous vehicles.

[0372] The embodiments described above are combinations of components and features of the present invention in a predetermined form. Each component or feature should be considered optional unless explicitly stated otherwise. Each component or feature may be implemented without being combined with other components or features. Furthermore, it is also possible to form an embodiment of the present invention by combining some components and / or features. The order of operations described in the embodiments of the present invention may be changed. Some components or features of one embodiment may be included in another embodiment or may be replaced with corresponding components or features of another embodiment. It is self-evident that claims that do not have an explicit citation relationship in the patent claims may be combined to form an embodiment or may be incorporated as a new claim through a post-application amendment.

[0373] It will be apparent to those skilled in the art that the present invention can be embodied in other specific forms without departing from the scope of the invention. Therefore, the above detailed description should not be construed as limiting in any respect, but rather as illustrative. The scope of the present invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the scope of equivalents of the present invention are intended to be included within the scope of the present invention.

[0374] The present invention can be used in a terminal, base station, or other equipment of a wireless mobile communication system.

Claims

1. In a method performed by a terminal in a wireless communication system, Obtain settings for AI / ML (artificial intelligence / machine learning) models based on upper-layer signaling; Monitoring at least one of the above AI / ML models; and Based on monitoring of at least one AI / ML model, performing AI / ML model management for switching or updating, The monitoring of at least one AI / ML model includes at least one of monitoring of a first AI / ML model that is inactive and monitoring of a second AI / ML model that is active, A method in which monitoring of the first AI / ML model in the inactive state occupies a smaller number of terminal processing units or for a shorter period of time than monitoring of the second AI / ML model in the active state.

2. In paragraph 1, A method wherein the first AI / ML model to be monitored among the inactive AI / ML models is determined based on network signaling.

3. In paragraph 2, A method wherein the start of the time interval in which the terminal processing unit is occupied by monitoring of the first AI / ML model is determined based on the time at which the network signaling is received.

4. In paragraph 2, The end of the time interval in which the terminal processing unit is occupied by monitoring of the first AI / ML model is A method determined based on at least one of i) the time at which the results of monitoring of the first AI / ML model are reported, ii) the time at which a network response to the report is received, or iii) the time at which a preset timer expires.

5. In paragraph 1, A method wherein the first AI / ML model to be monitored among the inactive AI / ML models is determined by the terminal.

6. In paragraph 5, A method in which monitoring of an inactive AI / ML model is stopped or suspended based on an action for the active second AI / ML model being triggered via network signaling.

7. In paragraph 1, A method further comprising reporting information about the occupied or unoccupied state of the terminal processing unit to the network.

8. In paragraph 1, A method in which monitoring of the first AI / ML model is performed based on the remaining terminal processing units that are not occupied by the second AI / ML model that is in an active state.

9. In paragraph 1, A method in which the second AI / ML model in the active state occupies the terminal processing unit with a higher priority than the first AI / ML model in the inactive state.

10. In paragraph 1, Monitoring of the second AI / ML model in the above active state occupies 1 terminal processing unit, A method wherein monitoring of the first AI / ML model in the above inactive state occupies n terminal processing units, where n is a real number greater than 0 and less than or equal to 1.

11. In paragraph 10, The above n is determined based on terminal capability or provided through network signaling.

12. A computer-readable recording medium having recorded thereon a program for performing the method described in paragraph 1.

13. In a wireless communication system, at a terminal, Transmitter and receiver; and including a processor controlling the above transmitter and receiver, The processor obtains settings for AI / ML (artificial intelligence / machine learning) models based on upper layer signaling, monitors at least one of the AI / ML models, and performs AI / ML model management for switching or updating based on the monitoring of the at least one AI / ML model. The monitoring of at least one AI / ML model includes at least one of monitoring of a first AI / ML model that is inactive and monitoring of a second AI / ML model that is active, A terminal in which monitoring of the first AI / ML model in the inactive state occupies a smaller number of terminal processing units or for a shorter period of time than monitoring of the second AI / ML model in the active state.

14. In a device that controls a terminal in a wireless communication system, memory that stores commands; and A processor that performs operations for controlling the terminal by executing the above commands, The operations for controlling the terminal include obtaining settings for AI / ML (artificial intelligence / machine learning) models based on upper layer signaling; monitoring at least one of the AI / ML models; and performing AI / ML model management for switching or updating based on the monitoring of the at least one AI / ML model. The monitoring of at least one AI / ML model includes at least one of monitoring of a first AI / ML model that is inactive and monitoring of a second AI / ML model that is active, A device wherein monitoring of the first AI / ML model in the inactive state occupies a smaller number of processing units of the device or for a shorter period of time than monitoring of the second AI / ML model in the active state.

Images