Method and apparatus 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 embodiment disclosed in the present specification may comprise the steps of: acquiring a configuration for an artificial intelligence / machine learning (AI / ML) model on the basis of high layer signaling; monitoring one or more AI / ML models; and performing AI / ML model management for handover or updating based on monitoring the one or more AI / ML models. Monitoring the one or more AI / ML models includes monitoring at least one of a first AI / ML model in an inactive state or a second AI / ML model in an active state. The first AI / ML model monitoring in the inactive state may occupy fewer terminal processing units or occupy the terminal processing units for a shorter period of time than the second AI / ML model monitoring in the active state.

Description

TECHNICAL FIELD

[0001] The disclosure relates to a wireless communication system, and more particularly, to a method and apparatus for transmitting / receiving uplink / downlink wireless signals in a wireless communication system. BACKGROUND

[0002] Generally, wireless communication systems are developing to variously cover a wide range to provide communication services such as audio communication services, data communication services, etc. A wireless communication is a multiple access system that is capable of supporting communication with multiple users by sharing available system resources (e.g., bandwidth, transmit power, etc.). For example, the multiple access system can be any one of a code division multiple access (CDMA) system, a frequency division multiple access (FDMA) system, a time division multiple access (TDMA) system, an orthogonal frequency division multiple access (OFDMA) system, and a single carrier frequency division multiple access (SC-FDMA) system. SUMMARY

[0003] TECHNICAL PROBLEM

[0004] An object of the disclosure is to provide a method and apparatus that more accurately and efficiently perform a process of transmitting and receiving wireless signals.

[0005] The object to be achieved is not limited to this, and other objects can be inferred from the disclosure of the specification.

[0006] TECHNICAL SOLUTION

[0007] In an aspect of the disclosure, a method performed by a user equipment (UE) in a wireless communication system is provided herein. The method can include obtaining a configuration for artificial intelligence / machine learning (AI / ML) models based on higher layer signaling, monitoring at least one AI / ML model, and performing AI / ML model management for handover or update based on monitoring the at least one AI / ML model. Monitoring the at least one AI / ML model can include at least one of monitoring a first AI / ML model in an inactive state and monitoring a second AI / ML model in an active state. Monitoring the first AI / ML model in the inactive state can occupy a UE processing unit for a shorter duration than monitoring the second AI / ML model in the active state, or can occupy a less UE processing unit than monitoring the second AI / ML model in the active state.

[0008] The first AI / ML model to be monitored among the AI / ML models in the inactive state can be determined based on network signaling.

[0009] A start of a duration in which a UE processing unit is occupied due to monitoring the first AI / ML model can be determined based on a time point at which network signaling is received.

[0010] The end of the duration of occupation of the UE processing unit by monitoring the first AI / ML model can be determined based on at least one of: i) a time point of reporting a result of monitoring the first AI / ML model; ii) a time point of receiving a network response to the reporting; or iii) a time point of expiration of a predetermined timer.

[0011] The first AI / ML model to be monitored among the AI / ML models in the inactive state can be determined by the UE.

[0012] Based on triggering the operation for the second AI / ML model in the active state through network signaling, monitoring the AI / ML models in the inactive state can be stopped or suspended.

[0013] The UE can report information about the occupied state or unoccupied state of the UE processing unit to the network.

[0014] Monitoring the first AI / ML model can be performed based on the remaining UE processing unit not occupied by the second AI / ML model in the active state.

[0015] The second AI / ML model in the active state can occupy the UE processing unit with a higher priority than the first AI / ML model in the inactive state.

[0016] Monitoring the second AI / ML model in the active state can occupy one UE processing unit, and monitoring the first AI / ML model in the inactive state can occupy n UE processing units, where n can be a real number greater than 0 and less than or equal to 1.

[0017] In this case, n can be determined based on the UE capability, or can be provided through network signaling.

[0018] In another aspect of the disclosure, a computer-readable recording medium having recorded thereon a program for executing the above-described method is provided.

[0019] In another aspect of the disclosure, provided herein is a user equipment (UE) in a wireless communication system. The UE can include a transceiver; and a processor configured to control the transceiver. The processor can be configured to obtain a configuration for artificial intelligence / machine learning (AI / ML) models based on higher layer signaling, monitor at least one AI / ML model, and perform AI / ML model management for switching or updating based on monitoring the at least one AI / ML model. Monitoring the at least one AI / ML model can include at least one of monitoring a first AI / ML model in an inactive state, and monitoring a second AI / ML model in an active state. Monitoring the first AI / ML model in the inactive state can occupy a UE processing unit for a shorter duration than monitoring the second AI / ML model in the active state, or can occupy a less UE processing unit than monitoring the second AI / ML model in the active state.

[0020] In another aspect of the disclosure, provided herein is a device configured to control a user equipment (UE) in a wireless communication system. The device can include a memory configured to store instructions, and a processor configured to perform operations for controlling the UE by executing the instructions. The operations for controlling the UE can include obtaining a configuration for artificial intelligence / machine learning (AI / ML) models based on higher layer signaling, monitoring at least one AI / ML model, and performing AI / ML model management for switching or updating based on monitoring the at least one AI / ML model. Monitoring the at least one AI / ML model can include at least one of monitoring a first AI / ML model in an inactive state, and monitoring a second AI / ML model in an active state. Monitoring the first AI / ML model in the inactive state can occupy a device processing unit for a shorter duration than monitoring the second AI / ML model in the active state, or can occupy a less device processing unit than monitoring the second AI / ML model in the active state.

[0021] Advantageous Effects

[0022] According to the embodiments, wireless signal transmission and reception can be efficiently performed in a wireless communication system.

[0023] Other effects not mentioned above can be inferred from the description. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 Physical channels used in a 3rd generation partnership project (3GPP) system, which is an exemplary wireless communication system, and a general signal transmission method using the same are illustrated.

[0025] Figure 2 A radio frame structure is illustrated.

[0026] Figure 3 A resource grid of a time slot is shown.

[0027] Figure 4 An exemplary mapping of physical channels in a time slot is shown.

[0028] Figure 5 An exemplary physical downlink shared channel (PDSCH) and ACK / NACK transmission and reception process is shown.

[0029] Figure 6 An exemplary physical uplink shared channel (PUSCH) transmission process is shown.

[0030] Figure 7 An example of a channel state information (CSI) related procedure is shown.

[0031] Figure 8 is a diagram for explaining concepts of artificial intelligence / machine learning (AI / ML) / deep learning.

[0032] Figures 9 to 12 Various AI / ML models of deep learning are shown.

[0033] Figure 13 is a diagram showing split AI inference.

[0034] Figure 14 is a diagram showing a framework of 3GPP radio access network (RAN) intelligence.

[0035] Figures 15 to 17 An AI model training and inference environment is shown.

[0036] Figure 18 is a diagram for explaining AI-based CSI feedback.

[0037] Figure 19 is a diagram for explaining operations of a network and a user equipment (UE) in a wireless communication system according to an embodiment.

[0038] Figure 20 A flow of a method performed by a UE according to an embodiment is shown.

[0039] Figures 21 to 24 Examples of a communication system 1 and a wireless device to which the present disclosure is applied are shown. DETAILED DESCRIPTION

[0040] Embodiments of the present disclosure are applicable to various wireless access technologies such as 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). CDMA can be implemented as a radio technology such as universal terrestrial radio access (UTRA) or CDMA2000. TDMA can be implemented as a radio technology such as global system for mobile communications (GSM) / general packet radio service (GPRS) / enhanced data rates for GSM evolution (EDGE). OFDMA can be implemented as a radio technology such as institute of electrical and electronics engineers (IEEE) 802.11 (wireless fidelity (Wi-Fi)), IEEE 802.16 (worldwide interoperability for microwave access (WiMAX)), IEEE 802.20, and evolved UTRA (E-UTRA). UTRA is a part of universal mobile telecommunications system (UMTS). 3rd generation partnership project (3GPP) long term evolution (LTE) is a part of evolved UMTS (E-UMTS) using E-UTRA, and LTE-Advanced (A) 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.

[0041] As more and more communication devices require greater communication capacity, enhanced mobile broadband communication relative to legacy radio access technology (RAT) is required. In addition, massive machine type communications (MTC), which is capable of providing various services anytime anywhere by connecting multiple devices and objects, is another important issue to be considered in the next-generation communication. Also, communication system design considering services / UEs sensitive to reliability and latency is being discussed. Accordingly, introduction of a new radio access technology considering enhanced mobile broadband communication (eMBB), massive MTC, and ultra-reliable low-latency communication (URLLC) is being discussed. In the present disclosure, for simplicity, this technology will be referred to as NR (New Radio or new RAT).

[0042] For the sake of simplicity, mainly 3GPP NR is described, but the technical idea of the present disclosure is not limited thereto.

[0043] In the disclosure, the term "set" can be replaced with "configure", both of which are used interchangeably. Also, conditional expressions (for example, "if", "in the case of", or "when") can be replaced with "based on" or "in the state of". In addition, the operation or software / hardware (SW / HW) configuration of the user equipment (UE) / base station (BS) can be derived / understood based on the satisfaction of the corresponding conditions. When the processing of the receiving (or transmitting) side can be derived / understood from the processing of the transmitting (or receiving) side in signal transmission / reception between wireless communication devices (for example, BS and UE), its description can be omitted. For example, signal determination / generation / coding / transmission of the transmitting side can be understood as signal monitoring reception / decoding / determination of the receiving side. Also, when it is said that the UE performs (or does not perform) a specific operation, it can also be interpreted that the BS expects / hypothesizes (or does not expect / hypothesizes) the UE to perform the specific operation. When it is said that the BS performs (or does not perform) a specific operation, it can also be interpreted that the UE expects / hypothesizes (or does not expect / hypothesizes) the BS to perform the specific operation. In the following description, for the convenience of description, sections, embodiments, examples, options, methods, schemes, etc. are distinguished from and indexed to each other, which does not mean that each of them necessarily constitutes an independent invention or that each of them should be implemented only individually. Unless explicitly contradictory to each other, at least a part of the sections, embodiments, examples, options, methods, schemes, etc. can be implemented in combination or can be omitted.

[0044] In a wireless communication system, a user equipment (UE) receives information from a base station (BS) through a downlink (DL) and transmits information to the BS through an uplink (UL). The information transmitted and received by the BS and the UE includes data and various control information, and includes various physical channels according to the type / use of the information transmitted and received by the UE and the BS.

[0045] Figure 1 Physical channels used in the 3GPP NR system and a general signal transmission method using the same are shown.

[0046] When the UE is powered on again from a power-off state or enters a new cell, the UE performs an initial cell search procedure (for example, synchronization establishment with the BS) in step S101. To this end, the UE receives a synchronization signal block (SSB) from the BS. The SSB includes a primary synchronization signal (PSS), a secondary synchronization signal (SSS), and a physical broadcast channel (PBCH). The UE establishes synchronization with the BS based on the PSS / SSS and acquires information such as a cell identification (ID). The UE can acquire broadcast information in the cell based on the PBCH. The UE can receive a DL reference signal (RS) in the initial cell search procedure to monitor the DL channel state.

[0047] After the initial cell search, the UE can acquire more specific system information by receiving a physical downlink control channel (PDCCH) and receiving a physical downlink shared channel (PDSCH) based on information of the PDCCH in step S102.

[0048] The UE can perform a random access procedure to access the BS in steps S103 to S106. For random access, the UE can transmit a preamble to the BS on a physical random access channel (PRACH) (S103) and receive a response message to the preamble on a PDCCH and a PDSCH corresponding to the PDCCH (S104). In the case of contention-based random access, the UE can perform a contention resolution procedure by further transmitting a PRACH (S105) and receiving a PDCCH and a PDSCH corresponding to the PDCCH (S106).

[0049] After the foregoing procedure, the UE can receive a PDCCH / PDSCH (S107) and transmit a physical uplink shared channel (PUSCH) / physical uplink control channel (PUCCH) (S108) as a general downlink / uplink signal transmission procedure. Control information transmitted from the UE to the BS is referred to as uplink control information (UCI). The UCI includes a hybrid automatic repeat and request acknowledgement / negative determination (HARQ-ACK / NACK), a scheduling request (SR), channel state information (CSI), etc. The CSI includes a channel quality indicator (CQI), a precoding matrix indicator (PMI), a rank indicator (RI), etc. Although the UCI is generally transmitted on the PUCCH, when control information and traffic data need to be simultaneously transmitted, the UCI can be transmitted on the PUSCH. In addition, the UCI can be transmitted aperiodically through the PUSCH according to a request / command of the network.

[0050] Figure 2 A radio frame structure is shown. In NR, uplink transmission and downlink transmission are configured with a frame. Each radio frame has a length of 10 ms and is divided into two 5 ms half frames (HF). Each half frame is divided into five 1 ms subframes (SF). A subframe is divided into one or more slots, and the number of slots in a subframe depends on the subcarrier spacing (SCS). According to a cyclic prefix (CP), each slot includes 12 or 14 orthogonal frequency-division multiplexing (OFDM) symbols. When a normal CP is used, each slot includes 14 OFDM symbols. When an extended CP is used, each slot includes 12 OFDM symbols.

[0051] Table 1 exemplarily shows that the number of symbols per slot, the number of slots per frame, and the number of slots per subframe vary according to the SCS when a normal CP is used.

[0052] [Table 1] N slot symb : number of symbols in a slot N frame,u slot : number of slots in a frame N subframe,u slot : number of slots in a subframe Table 2 shows that the number of symbols per slot, the number of slots per frame, and the number of slots per subframe vary according to the SCS when the extended CP is used.

[0053] [Table 2]

[0054] The structure of the frame is merely an example. The number of subframes, the number of slots, and the number of symbols in the frame can vary.

[0055] In the NR system, OFDM numerology (e.g., SCS) can be differently configured for a plurality of cells aggregated for one UE. Accordingly, the (absolute time) duration of a time resource (e.g., SF, slot, or TTI) (for simplicity, referred to as a time unit (TU)) consisting of the same number of symbols can be differently configured among the aggregated cells. Here, the symbol can include an OFDM symbol (or CP-OFDM symbol) and an SC-FDMA symbol (or discrete Fourier transform-spread-OFDM (DFT-s-OFDM) symbol).

[0056] Figure 3 A resource grid of a slot is shown. The slot includes a plurality of symbols in the time domain. For example, when the normal CP is used, the slot includes 14 symbols. However, when the extended CP is used, the slot includes 12 symbols. A carrier includes a plurality of subcarriers in the frequency domain. A resource block (RB) is defined as a plurality of consecutive subcarriers (e.g., 12 consecutive subcarriers) in the frequency domain. A bandwidth part (BWP) can be defined as a plurality of consecutive physical RBs (PRBs) in the frequency domain and corresponds to a single numerology (e.g., SCS, CP length, etc.). The carrier can include up to N (e.g., five) BWPs. Data communication can be performed through an enabled BWP, and only one BWP can be enabled for one UE. In the resource grid, each element is referred to as a resource element (RE), and one complex symbol can be mapped to each RE.

[0057] Figure 4An exemplary mapping of physical channels in a time slot is shown. PDCCH can be transmitted in a DL control region, PDSCH can be transmitted in a DL data region. PUCCH can be transmitted in a UL control region, PUSCH can be transmitted in a UL data region. A guard period (GP) provides a time gap for transmission mode to reception mode switching or reception mode to transmission mode switching at the BS and the UE. Some symbols in a subframe at the DL to UL switching can be configured as a GP.

[0058] The various physical channels will be described in more detail below.

[0059] PDCCH conveys DCI. For example, PDCCH (i.e., DCI) can carry information on transport format and resource allocation of DL shared channel (DL-SCH), resource allocation information of uplink shared channel (UL-SCH), paging information on paging channel (PCH), system information on DL-SCH, information on resource allocation of high speed shared control channel (HS-SCCH), transmit power control commands, information on activation / deactivation of configured scheduling, etc. DCI includes a cyclic redundancy check (CRC). The CRC is masked with various identifiers (IDs) (e.g., radio network temporary identifier (RNTI)) according to the owner or usage of the PDCCH. For example, if the PDCCH is for a specific UE, the CRC is masked with a UE ID (e.g., cell-RNTI (C-RNTI)). If the PDCCH is for a paging message, the CRC is masked with a paging-RNTI (P-RNTI). If the PDCCH is used for system information (e.g., system information block (SIB)), the CRC is masked with a system information RNTI (SI-RNTI). When the PDCCH is used for an RAR, the CRC is masked with a random access-RNTI (RA-RNTI).

[0060] A PDCCH includes 1, 2, 4, 8, or 16 control channel elements (CCEs) according to its aggregation level (AL). A CCE is a logical allocation unit for providing a certain code rate to a PDCCH according to a radio channel state. A CCE includes 6 resource element groups (REGs), each REG being defined by one OFDM symbol × one (P) RB. A PDCCH is transmitted in 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 one UE can overlap with each other in the time / frequency domain. A CORESET can be configured by system information (e.g., master information block (MIB)) or UE-specific higher layer signaling (e.g., radio resource control (RRC) signaling). Specifically, the number of RBs and the number of symbols (up to 3) in a CORESET can be configured by higher layer signaling.

[0061] For PDCCH reception / detection, a UE monitors PDCCH candidates. A PDCCH candidate is a CCE that the UE should monitor to detect a PDCCH. Each PDCCH candidate is defined as 1, 2, 4, 8, or 16 CCEs according to an AL. Monitoring includes (blind) decoding of the PDCCH candidate. A set of PDCCH candidates decoded by a UE is defined as a PDCCH search space (SS). An SS can be a common search space (CSS) or a UE-specific search space (USS). A UE can obtain DCI by monitoring PDCCH candidates in one or more SSs configured by MIB or higher layer signaling. Each CORESET is associated with one or more SSs, and each SS is associated with one CORESET. An SS can be defined based on the following parameters.

[0062] - controlResourceSetId: CORESET related to the SS.

[0063] - monitoringSlotPeriodicityAndOffset: PDCCH monitoring periodicity (slots) and PDCCH monitoring offset (slots).

[0064] - monitoringSymbolsWithinSlot: PDCCH monitoring symbols in a slot (e.g., the first symbol of the CORESET).

[0065] - nrofCandidates: The number of PDCCH candidates for each AL = {1, 2, 4, 8, 16} (one of 0, 1, 2, 3, 4, 5, 6, and 8).

[0066] The occasions (e.g., time / frequency resources) in which a UE is to monitor PDCCH candidates are defined as PDCCH (monitoring) occasions. One or more PDCCH (monitoring) occasions can be configured in a slot.

[0067] Table 3 shows the characteristics of each SS.

[0068] [Table 3]

[0069] Table 4 shows the DCI formats transmitted on PDCCH.

[0070] [Table 4]

[0071] DCI format 0_0 can be used to schedule TB (or TB-level) based PUSCH, DCI format 0_1 can be used to schedule TB (or TB-level) based PUSCH or code block group (CBG) (or CBG-level) based PUSCH. DCI format 1_0 can be used to schedule TB (or TB-level) based PDSCH, DCI format 1_1 can be used to schedule TB (or TB-level) based PDSCH or CBG (or CBG-level) based PDSCH (or DL grant DCI). DCI format 0_0 / 0_1 can be referred to as UL grant DCI or UL scheduling information, DCI format 1_0 / 1_1 can 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 slot format indicator (SFI)) to a UE, DCI format 2_1 is used to convey DL pre-emption information to a UE. DCI format 2_0 and / or DCI format 2_1 can be conveyed to a group of UEs on a group common PDCCH (PDCCH pointing to a group of UEs).

[0072] DCI format 0_0 and DCI format 1_0 can be referred to as fallback DCI formats, while DCI format 0_1 and DCI format 1_1 can be referred to as non-fallback DCI formats. Under fallback DCI formats, the DCI size / field configuration is maintained to be the same regardless of UE configuration. In contrast, under non-fallback DCI formats, the DCI size / field configuration varies depending on UE configuration.

[0073] The PDSCH transmits DL data (e.g., a DL shared channel transport block (DL-SCH TB)) and uses a modulation scheme such as quadrature phase shift keying (QPSK), 16-quadrature amplitude modulation (16QAM), 64QAM, or 256QAM. A TB is coded into a codeword. The PDSCH can transmit up to two codewords. Scrambling and modulation mapping can be performed on a codeword basis, and modulation symbols generated from each codeword can be mapped to one or more layers. Each layer is mapped to a resource together with a demodulation reference signal (DMRS), and an OFDM symbol signal is generated from the layer mapped with the DMRS and transmitted through a corresponding antenna port.

[0074] The PUCCH transmits uplink control information (UCI). The UCI includes the following information.

[0075] - SR (scheduling request): information for requesting a UL-SCH resource.

[0076] - HARQ (hybrid automatic repeat request)-ACK (acknowledgement): a response to a DL data packet (e.g., a codeword) on a PDSCH. The HARQ-ACK indicates whether the DL data packet is successfully received. In response to a single codeword, a 1-bit HARQ-ACK can be transmitted. In response to two codewords, a 2-bit HARQ-ACK can be transmitted. The HARQ-ACK response includes a positive ACK (simply ACK), a negative ACK (NACK), discontinuous transmission (DTX), or NACK / DTX. The term HARQ-ACK can be used interchangeably with HARQ ACK / NACK and ACK / NACK.

[0077] - CSI (channel state information): feedback information for a DL channel. Multiple-input multiple-output (MIMO)-related feedback information includes an RI and a PMI.

[0078] Table 5 shows exemplary PUCCH formats. Based on a PUCCH transmission duration, the PUCCH formats can be divided into short PUCCHs (formats 0 and 2) and long PUCCHs (formats 1, 3, and 4).

[0079] [Table 5]

[0080] The PUCCH format 0 transmits up to 2 bits of UCI and is mapped in a sequence-based manner for transmission. Specifically, the UE transmits specific UCI to the BS by transmitting one of a plurality of sequences on a PUCCH of the PUCCH format 0. The UE transmits the PUCCH of the PUCCH format 0 in a PUCCH resource configured for the corresponding SR only when the UE transmits a positive SR.

[0081] PUCCH format 1 transmits up to 2 bits of UCI, and the modulated symbols of UCI are spread in time domain with an orthogonal cover code (OCC) (whether frequency hopping is performed is configured differently). DMRS is transmitted in symbols where modulated symbols are not transmitted (i.e., transmitted in time-division multiplexing (TDM)).

[0082] PUCCH format 2 transmits more than 2 bits of UCI, and the modulated symbols of UCI are transmitted with DMRS in frequency-division multiplexing (FDM). DMRS is located in symbols #1, #4, #7, and #10 of a given RB with a density of 1 / 3. A pseudo-noise (PN) sequence is used for the DMRS sequence. For 2-symbol PUCCH format 2, frequency hopping can be enabled.

[0083] PUCCH format 3 does not support UE multiplexing in the same PRBS and transmits more than 2 bits of UCI. In other words, the PUCCH resource of PUCCH format 3 does not include OCC. The modulated symbols are transmitted with DMRS in TDM.

[0084] PUCCH format 4 supports multiplexing of up to 4 UEs in the same PRBS and transmits more than 2 bits of UCI. In other words, the PUCCH resource of PUCCH format 3 includes OCC. The modulated symbols are transmitted with DMRS in TDM.

[0085] At least one of one or two or more cells configured in the UE can be configured for PUCCH transmission. At least a primary cell can be configured as a cell for PUCCH transmission. Based on at least one cell configured for PUCCH transmission, at least one PUCCH cell group can be configured in the UE, and each PUCCH cell group includes one or more cells. The PUCCH cell group can be simply referred to as a PUCCH group. PUCCH transmission can be configured for an SCell as well as a primary cell. The primary cell belongs to a primary PUCCH group, and a PUCCH-SCell configured for PUCCH transmission belongs to a secondary PUCCH group. PUCCH on the primary cell can be used for cells belonging to the primary PUCCH group, and PUCCH on the PUCCH-SCell can be used for cells belonging to the secondary PUCCH group.

[0086] PUSCH transmits UL data (e.g., UL shared channel transport block (UL-SCH TB)) and / or UCI based on a CP-OFDM waveform or a DFT-s-OFDM waveform. When PUSCH is transmitted with a DFT-s-OFDM waveform, the UE transmits the PUSCH with transform precoding. For example, the UE can transmit the PUSCH with a CP-OFDM waveform when transform precoding is not possible (e.g., disabled), and the UE can transmit the PUSCH with a CP-OFDM or DFT-s-OFDM waveform when transform precoding is possible (e.g., enabled). The PUSCH transmission can be dynamically scheduled by an UL grant in DCI or semi-statically scheduled by higher layer (e.g., RRC) signaling (and / or layer 1 (L1) signaling such as PDCCH) (configured scheduling or configured grant). The PUSCH transmission can be performed in a codebook-based or non-codebook-based manner.

[0087] Figure 5 An example ACK / NACK transmission process is shown. Referring to Figure 5 , a UE can detect a PDCCH in slot #n. The PDCCH includes DL scheduling information (e.g., DCI format 1_0 or DCI format 1_1). The PDCCH indicates a DL assignment with a PDSCH offset K0 and a PDSCH to HARQ-ACK reporting offset K1. For example, DCI format 1_0 and DCI format 1_1 can include the following information.

[0088] - Frequency domain resource assignment: indicates a set of RBs assigned to the PDSCH.

[0089] - Time domain resource assignment: indicates K0 and a starting position (e.g., OFDM symbol index) and length (e.g., number of OFDM symbols) of the PDSCH in a slot.

[0090] - PDSCH-to-HARQ_feedback timing indicator: indicates K1.

[0091] - HARQ process number (4 bits): indicates a HARQ process ID of data (e.g., PDSCH or TB).

[0092] - PUCCH resource indicator (PRI): indicates a PUCCH resource out of a plurality of PUCCH resources in a PUCCH resource set to be used for UCI transmission.

[0093] After receiving the PDSCH in slot #(n+K0) according to the scheduling information of slot #n, the UE can transmit UCI on a PUCCH in slot #(n+K1). The UCI can include a HARQ-ACK response to the PDSCH. For convenience, Figure 5Based on the assumption that the SCS of PDSCH is equal to the SCS of PUCCH and slot #n1 = slot #(n+K0), this should not be interpreted as limiting the present disclosure. When the SCS is not the same, K1 can be indicated / interpreted based on the SCS of PUCCH.

[0094] In case PDSCH is configured to carry at most one TB, HARQ-ACK response can be configured in one bit. In case PDSCH is configured to carry at most two TBs, if spatial bundling is not configured, HARQ-ACK response can be configured in 2 bits, if spatial bundling is configured, HARQ-ACK response can be configured in 1 bit. When slot #(n+K1) is specified as the HARQ-ACK transmission timing for multiple PDSCHs, the UCI transmitted in slot #(n+K1) includes HARQ-ACK responses for multiple PDSCHs.

[0095] UE whether it should perform spatial bundling for HARQ-ACK responses can be configured (e.g., by RRC / high layer signaling) for each cell group. For example, spatial bundling can be configured for each individual HARQ-ACK response transmitted on PUCCH and / or HARQ-ACK responses transmitted on PUSCH.

[0096] When at most two (or two or more) TBs (or codewords) can be received at a time (may or can not be scheduled by one DCI) in the corresponding serving cells (e.g., when the higher layer parameter maxNrofCodeWordsScheduledByDCI indicates 2 TBs), spatial bundling can be supported. More than four layers can be used for 2 TB transmission, and at most four layers can be used for 1 TB transmission. As a result, when spatial bundling is configured for the corresponding cell group, spatial bundling can be performed for the serving cells among the cell group's serving cells for which more than four layers can be scheduled. A UE that wants to transmit HARQ-ACK responses through spatial bundling can generate the HARQ-ACK responses by performing a (bit-wise) logical AND operation on the A / N bits for multiple TBs.

[0097] For example, assuming that a UE receives a DCI that schedules two TBs and receives the two TBs on PDSCH based on the DCI, a UE that performs spatial bundling can generate a single A / N bit by a logical AND operation between a first A / N bit of a first TB and a second A / N bit of a second TB. As a result, when both the first TB and the second TB are ACK, the UE reports an ACK bit value to the BS, and when at least one TB is NACK, the UE reports a NACK bit value to the BS.

[0098] For example, when only one TB is actually scheduled in a serving cell configured for receiving two TBs, the UE can generate a single A / N bit by performing a logical AND operation on the A / N bits of the one TB and a bit value of 1. As a result, the UE reports the A / N bits of the one TB to the BS.

[0099] There are multiple parallel DL HARQ processes for DL transmission at the BS / UE. The multiple parallel HARQ processes allow for consecutive DL transmissions while the BS waits for HARQ feedback indicating successful or failed reception of a previous DL transmission. Each HARQ process is associated with a HARQ buffer in the medium access control (MAC) layer. Each DL HARQ process manages a number of MAC physical data unit (PDU) transmissions, HARQ feedback for MAC PDUs in the buffer, and state variables for the current redundancy version. Each HARQ process is identified by a HARQ process ID.

[0100] Figure 6 An example PUSCH transmission procedure is shown. Referring to Figure 6 , the UE can detect a PDCCH in slot #n. The PDCCH includes DL scheduling information (e.g., DCI format 1_0 or 1_1). The DCI format 1_0 or 1_1 can include the following information.

[0101] - Frequency domain resource assignment: indicates a set of RBs assigned to the PUSCH.

[0102] - Time domain resource assignment: indicates a slot offset K2 and a starting position (e.g., OFDM symbol index) and duration (e.g., number of OFDM symbols) of the PUSCH in a slot. The starting symbol and length of the PUSCH can be indicated by a start and length indicator value (SLIV) or separately indicated.

[0103] The UE can then transmit a PUSCH in slot #(n+K2) according to the scheduling information in slot #n. The PUSCH includes a UL-SCH TB.

[0104] CSI-related operations

[0105] Figure 7 An example of a CSI-related procedure is shown.

[0106] The UE receives, from the BS via RRC signaling, configuration information related to CSI (710). The CSI-related configuration information can include at least one of channel state information-interference measurement (CSI-IM)-related information, CSI measurement-related information, CSI resource configuration-related information, CSI-RS resource-related information, or CSI reporting configuration-related information.

[0107] - A CSI-IM resource can be configured for interference measurement (IM) of a UE. In time domain, a set of CSI-IM resources can be configured as periodic, semi-persistent, or aperiodic. A CSI-IM resource can be configured as a zero-power (ZP)-CSI-RS for a UE. A ZP-CSI-RS can be configured to be distinguished from a non-zero-power (NZP)-CSI-RS.

[0108] - A UE can assume that a CSI-RS resource configured for channel measurement and a CSI-IM / NZP CSI-RS resource configured for interference measurement for one CSI report have a QCL relationship with respect to “QCL-TypeD” of the respective resources (when a NZP CSI-RS resource is used for interference measurement).

[0109] - A CSI resource configuration can include at least one of a CSI-IM resource for interference measurement, a NZP CSI-RS resource for interference measurement, and a NZP CSI-RS resource for channel measurement. A channel measurement resource (CMR) can be a NZP CSI-RS for CSI acquisition, and an interference measurement resource (IMR) can be a NZP CSI-RS for CSI-IM and IM.

[0110] - A CSI-RS can be configured for one or more UEs. Different CSI-RS configurations can be provided for respective UEs, or the same CSI-RS configuration can be provided to multiple UEs. A CSI-RS can support up to 32 antenna ports. A 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 with CDM, FDM, and / or TDM methods. A CSI-RS can be mapped to remaining REs except for REs on which CORESET, DMRS, and SSB are mapped. In frequency domain, a CSI-RS can be configured for an entire bandwidth, a part of a bandwidth part (BWP), or a part of a bandwidth. A CSI-RS can be transmitted in each RB (i.e., density = 1) within a bandwidth in which the CSI-RS is configured, or a CSI-RS can be transmitted in every two RBs (e.g., even or odd RBs) (i.e., density = 1 / 2). When a CSI-RS is used as a tracking reference signal (TRS), a single-port CSI-RS can be mapped on three subcarriers in each resource block (i.e., density = 3). One or more CSI-RS resource sets can be configured for a UE in time domain. Each CSI-RS resource set can include one or more CSI-RS configurations. Each CSI-RS resource set can be configured as periodic, semi-persistent, or aperiodic.

[0111] - CSI reporting configuration can include configuration of feedback type, measurement resource, reporting type, etc. NZP-CSI-RS resource set can be used for CSI reporting configuration for corresponding UE. NZP-CSI-RS resource set can be associated with CSI-RS or SSB. Multiple periodic NZP-CSI-RS resource sets can be configured as TRS resource sets. (i) Feedback type includes channel quality indicator (CQI), precoding matrix indicator (PMI), CSI-RS resource indicator (CRI), SSB resource block indicator (SSBRI), layer indicator (LI), rank indicator (RI), first layer (L1)-reference signal received strength (RSRP), etc. (ii) Measurement resource can include configuration of downlink signal and / or downlink resource for which UE performs measurement to determine feedback information. Measurement resource can be configured as ZP and / or NZP CSI-RS resource set associated with CSI reporting configuration. NZP CSI-RS resource set can include CSI-RS set or SSB set. For example, L1-RSRP can be measured for CSI-RS set or SSB set. (iii) Reporting type can include configuration of time and uplink channel for which UE performs reporting. Reporting time can be configured as periodic, semi-persistent, or aperiodic. Periodic CSI reporting can be transmitted on PUCCH. Semi-persistent CSI reporting can be transmitted on PUCCH or PUSCH based on MAC CE indicating enabling / disabling. Aperiodic CSI reporting can be indicated by DCI signaling. For example, CSI request field of uplink grant can indicate one of various reporting trigger sizes. Aperiodic CSI reporting can be transmitted on PUSCH.

[0112] UE measures CSI based on configuration information related to CSI. CSI measurement can include receiving CSI-RS (720) and obtaining CSI by calculating received CSI-RS (730).

[0113] UE can transmit CSI report to BS (740). For CSI reporting, time resource and frequency resource available for UE are controlled by BS. Channel state information (CSI) includes at least one of channel quality indicator (CQI), precoding matrix indicator (PMI), CSI-RS resource indicator (CRI), SS / PBCH block resource indicator (SSBRI), layer indicator (LI), rank indicator (RI), L1-RSRP, and / or L-SINR.

[0114] The time domain behavior of CSI reporting supports periodic, semi-persistent, aperiodic. i) Periodic CSI reporting is performed in short PUCCH, long PUCCH. The periodicity and slot offset of periodic CSI reporting can be configured by RRC and refer to CSI-ReportConfig IE. ii) SP (semi-periodic) CSI reporting is performed in short PUCCH, long PUCCH or PUSCH. For SP CSI in short / long PUCCH, the periodicity and slot offset are configured by RRC and the CSI reporting is enabled / disabled by separate MAC CE / DCI. For SP CSI in PUSCH, the periodicity of SP CSI reporting is configured by RRC but the slot offset is not configured by RRC and the SP CSI reporting is enabled / disabled by DCI (format 0_1). For SP CSI reporting in PUSCH, a separate RNTI (SP-CSI C-RNTI) is used. The initial CSI reporting timing follows the PUSCH time domain allocation value indicated by DCI and the subsequent CSI reporting timing follows the periodicity configured by RRC. DCI format 0_1 can include CSI request field and enable / disable a specific configured SP-CSI trigger state. SP CSI reporting has the same or similar enable / disable mechanism as having data transmission in SPS PUSCH. iii) Aperiodic CSI reporting is performed in PUSCH and triggered by DCI. In this case, the information related to the trigger of aperiodic CSI reporting can be conveyed / indicated / configured by MAC-CE. For AP CSI with AP CSI-RS, the AP CSI-RS timing is configured by RRC and the timing of AP CSI reporting is dynamically controlled by DCI.

[0115] CSI codebooks (e.g., PMI codebooks) defined in NR specifications can be roughly classified into Type I codebooks and Type II codebooks. Type I codebooks are mainly for single-user MIMO (SU-MIMO) that supports both high order and low order. Type II codebooks are mainly for multi-user MIMO (MU-MIMO) that can handle up to two layers. Although Type II codebooks can provide more accurate CSI than Type I, Type II codebooks can also increase signaling overhead. On the other hand, an enhanced Type II codebook is introduced to address the CSI overhead problem associated with the existing Type II codebook. The enhanced Type II codebook can reduce the payload of the codebook by considering the correlation in the frequency domain.

[0116] CSI reporting on PUSCH can be configured as part 1 and part 2. Part 1 has a fixed payload size, which is used to identify the number of information bits in part 2. Part 1 is completely transmitted before part 2.

[0117] - For Type I CSI feedback, Part 1 includes RI (if reported), CRI (if reported) and CQI for the first codeword. Part 2 includes PMI and, when RI > 4, CQI for the second codeword.

[0118] - For Type II CSI feedback, Part 1 includes RI (if reported), CQI and an indication of the number of non-zero wideband amplitude coefficients for each layer of Type II CSI. Part 2 includes PMI for Type II CSI.

[0119] - For Enhanced Type II CSI feedback, Part 1 includes RI (if reported), CQI and an indication of the total number of non-zero wideband amplitude coefficients for all layers of Enhanced Type II CSI. Part 2 includes PMI for Enhanced Type II CSI.

[0120] For PUSCH, a CSI report includes two parts. If the CSI payload to be reported is smaller than the payload size provided by the PUSCH resources allocated for the CSI report, the UE can discard part of Part 2 CSI.

[0121] Semi-persistent CSI reporting performed in PUCCH format 3 or 4 supports Type II CSI feedback, but only supports Part 1 of Type II CSI feedback.

[0122] Quasi co-location (QCL)

[0123] Two antenna ports are quasi co-located when the channel properties of one antenna port are to be inferred from the channel of another antenna port. The channel properties can include one or more of delay spread, Doppler spread, frequency / Doppler shift, average received power, received timing / average delay and spatial RX parameters.

[0124] A list of multiple TCI state configurations can be configured in a UE by a higher layer parameter PDSCH-Config, each TCI state links to one or two QCL configuration parameters between one or two DL reference signals and the DM-RS ports of PDSCH. The QCL can include qcl-Type1 of the first DL RS and qcl-Type2 of the second DL RS. The QCL type can correspond to one of the following.

[0125] - “QCL-TypeA”: {Doppler shift, Doppler spread, average delay, delay spread}

[0126] - “QCL-TypeB”: {Doppler shift, Doppler spread}

[0127] - “QCL-TypeC”: {Doppler shift, average delay}

[0128] - "QCL-TypeD": {spatial Rx parameter}

[0129] Beam management (BM)

[0130] BM refers to a series of processes for acquiring and maintaining a set of BS beams (transmit receive point (TRP) beams) and / or a set of UE beams that can be used for DL and UL transmission / reception. The BM can include the following processes and terms.

[0131] - Beam measurement: an operation in which a BS or a UE measures a characteristic of a received beamformed signal.

[0132] - Beam determination: an operation in which a BS or a UE selects its Tx / Rx beam.

[0133] - Beam sweeping: an operation in which a Tx and / or Rx beam covers a spatial domain within a prescribed time interval according to a predetermined method.

[0134] - Beam reporting: an operation in which a UE reports information about a beamformed signal based on beam measurement.

[0135] The BM procedure can be divided into (1) a DL BM procedure using an SSB or a CSI-RS and (2) an UL BM procedure using an SRS. In addition, each BM procedure can include a Tx beam sweeping for determining a Tx beam and an Rx beam sweeping for determining an Rx beam.

[0136] The DL BM procedure can include (1) transmission of a beamformed DL RS (e.g., a CSI-RS or an SSB) from a BS and (2) beam reporting from a UE.

[0137] The beam reporting can include a preferred DL RS ID and a reference signal received power (RSRP) corresponding to the preferred DL RS ID. The DL RS ID can be an SSB resource indicator (SSBRI) or a CSI-RS resource indicator (CRI).

[0138] Artificial intelligence / machine learning (AI / ML)

[0139] With the technology development of AI / ML, the nodes and UEs constituting a wireless communication network are becoming more and more intelligent / advanced, and it is expected to quickly optimize and derive / apply various network / BS-determined parameter values (e.g., transmission and reception powers of respective BSs, transmission powers of respective UEs, precoders / beams of BSs / UEs, time / frequency resource allocations of respective UEs, or duplexing methods of BSs) according to various environmental parameters (e.g., distribution / location of BSs, distribution / location / materials of buildings / furniture, location / moving direction / speed of UEs, and climate information) due to the intelligence of the network / BS, in particular. In line with this trend, many standardization organizations (e.g., 3GPP or O-RAN) are considering the introduction of network / BS-determined parameter values, and research thereon is also actively ongoing.

[0140] In a narrow sense, AI / ML can be easily referred to as artificial intelligence based on deep learning, but is conceptually shown in Figure 8

[0141] - Artificial intelligence: This can correspond to all automation in which machines replace human work.

[0142] - Machine learning: A machine can learn a decision pattern from data by itself without explicit programming rules.

[0143] - Deep learning: This is an AI / ML model based on artificial neural networks in which a machine performs all at once from unstructured data to feature extraction and determination, and an algorithm relies on a biological nervous system, i.e., a multi-layer interconnected node network inspired by a neural network for feature extraction and transformation. Common deep learning network architectures can include a deep neural network (DNN), a recurrent neural network (RNN), and a convolutional neural network (CNN).

[0144] Classification of AI / ML types according to various references

[0145] 1. Offline vs. online

[0146] (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 and used for prediction. In most cases, this offline learning method is used. In offline learning, the system does not learn incrementally. Instead, learning is performed using all available collected data, and the results are applied without further learning. If new data needs to be learned, learning can be started again using new complete data.

[0147] (2) Online learning: Online learning is a method of improving performance little by little by incrementally learning with additionally generated data based on data to be used for learning that is continuously generated through the Internet. Real-time learning is performed on specific data units (batches) collected online, thereby allowing the system to quickly adapt to changing data.​

[0148] To build an AI system, learning can be done only with real-time generated data only through online learning. Alternatively, after offline learning using a specific data set, additional learning can be done using subsequently generated real-time data (online + offline learning).

[0149] 2. Classification according to AI / ML framework concept

[0150] (1) Centralized learning: When training data collected from multiple different nodes is reported to a centralized node, all data resources / storage / learning (e.g., supervised, unsupervised, and reinforcement learning) are performed by one central node.

[0151] (2) Federated learning: A collective AI / ML model is configured based on data across dispersed data owners. Instead of using data in an AI / ML model, a local node / each device collects data and trains a copy of its AI / ML model, so there is no need to report source data to a central node. In federated learning, the parameters / weights of the AI / ML model can be sent back to the centralized node to support general AI / ML model training. Advantages of federated learning include an increase in computing speed and advantages in information security. That is, processing that does not require uploading personal data to a central server can prevent leakage and misuse of personal information.

[0152] (3) Distributed learning: Machine learning processing is a concept that represents scaling and deployment across a cluster of nodes. AI / ML model training is split and shared across multiple nodes operating simultaneously to accelerate AI / ML model training.

[0153] 3. Classification according to learning method

[0154] (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. Input data is referred to as training data and has known labels or results. Examples of supervised learning can include: (i) Regression: linear regression, logistic regression; (ii) Instance-based algorithms: k-nearest neighbors (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 due to regression and classification problems, and classification predicts a label, and regression predicts a number.

[0155] (2) Unsupervised learning: This is a machine learning task that aims to learn a function that describes hidden structures in unlabeled data. Input data is unlabeled and has no known results. Some examples of unsupervised learning include K-means clustering, principal component analysis (PCA), nonlinear independent component analysis (ICA), and LSTM.

[0156] (3) Reinforcement Learning: In reinforcement learning (RL), an agent aims to optimize a long-term goal by processing interactions with an environment based on trial and error, which is goal-oriented learning based on interactions with an environment. Examples of RL algorithms can 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. RL can be further grouped into AI / ML model-based RL and AI / ML model-free RL. Model-based RL is a RL algorithm that uses a predictive AI / ML model of the environment using various dynamic states of the environment and an AI / ML model of the states leading to rewards to obtain probabilities of transitions between states. Model-free RL is a value- or policy-based RL algorithm that implements maximum future rewards, which is less computationally complex in terms of multi-agent environment / state and does not require an accurate representation of the environment. RL algorithms can also be classified as value-based RL vs. policy-based RL, policy-based RL vs. non-policy RL, etc.

[0157] AI / ML model

[0158] Figure 9 An example of a feedforward neural network (FFNN) AI / ML model is shown. Referring to Figure 9 , the FFNN AI / ML model includes an input layer, a hidden layer, and an output layer.

[0159] Figure 10 An example of a recurrent neural network (RNN) AI / ML model is shown. Referring to Figure 10 , the RNN AI / ML model is an artificial neural network in which hidden nodes are connected to directed edges to form directed loops, and is an AI / ML model suitable for processing sequentially occurring data (e.g., speech or text). One type of RNN is a long short-term memory (LSTM), which is a structure that adds a cell state to the hidden state of the RNN. In detail, in the 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 denotes a neural network, x t denotes an input value, h t denotes an output value. Here, h t may denote a current state value based on time, and h t-1 may denote a previous state value.

[0160] Figure 11A convolutional neural network (CNN) AI / ML model is illustrated. The CNN is used for two purposes, including reducing AI / ML model complexity and extracting good features by applying convolution calculation commonly used in the field of video processing or image processing. Referring to Figure 11 , a kernel or filter means a unit / structure that applies weights to an input within a certain range / unit. The kernel (or filter) can be modified through learning. A stride is a moving range in which the kernel moves within the input. A feature map is a result of applying a kernel to an input. Padding refers to adding values to adjust the size of a feature map. In order to enhance robustness to distortion and change, a plurality of feature maps can be extracted. Pooling refers to a calculation that reduces the size of a feature map by down-sampling the feature map (e.g., max-pooling or average-pooling).

[0161] Figure 12 An autoencoder AI / ML model is illustrated. Referring to Figure 12 , an autoencoder is a kind of unsupervised learning that receives a feature vector x and outputs the same or similar vector x', and an input node and an output node have the same features. The autoencoder reconstructs the input, and thus, the output can be referred to as reconstruction. A loss function can be expressed according to Equation 1 below.

[0162] [Equation 1] wherein

[0163] In Equation 12, a loss function of the autoencoder is calculated based on a difference between an input and an output. Based on the loss function of the autoencoder, the degree of loss in the input is evaluated, and an optimization process is applied to the autoencoder to minimize the loss.

[0164] Figure 13 is a diagram illustrating split AI inference.

[0165] Figure 13 A case in which a model inference function is cooperatively performed between a terminal device such as a UE and a network AI / ML endpoint during split AI operation is illustrated.

[0166] In addition to the model inference function, a model training function, an executor, and a data collection function can each be split into a plurality of parts based on a current task and an environment. These functions can be performed through cooperation of a plurality of entities.

[0167] For example, a computation-intensive part and an energy-intensive part can be performed at a network endpoint, and a privacy-sensitive part and a latency-sensitive part can be performed on a terminal device. In this case, the terminal device performs a task / model based on input data up to a certain part / layer, and then transmits intermediate data to the network endpoint. The network endpoint performs the remaining part / layer, and provides an inference output to one or more devices that perform an operation / task.

[0168] The following describes a functional framework for AI operations.

[0169] Herein, to explain AI (or AI / ML) in more detail, the following terms can be defined.

[0170] - Data collection: data collected from network nodes, management entities, or UEs, which serves as a basis for AI model training, data analysis, and inference.

[0171] - AI model: a data-driven algorithm in which AI techniques are applied to generate an output set based on an input set, including prediction information and / or decision parameters.

[0172] - AI / ML training: an online or offline process of training an AI model by learning features and patterns to best represent data and obtain an AI / ML model trained for inference.

[0173] - AI / ML inference: a process of making predictions or deriving decisions based on collected data and AI models using a trained AI model.

[0174] Referring to Figure 14 , the data collection function 10 collects input data and provides processed input data to the model training function 20 and the model inference function 30.

[0175] For example, the input data can include measurements from UEs or other network entities, feedback from actuators, and outputs from AI models.

[0176] The data collection function 10 performs data preparation based on the input data and provides input data processed through the data preparation. Here, the data collection function 10 does not perform specific data preparation (e.g., data preprocessing and cleaning, formation and transformation) for each AI algorithm, but the data collection function 10 can perform data preparation common to AI algorithms.

[0177] After completing the data preparation process, the data collection function 10 provides training data 11 to the model training function 20 and provides inference data 12 to the model inference function 30. Here, the training data 11 is input data required for the AI model training function 20. The inference data 12 is input data required for the AI model inference function 30.

[0178] The data collection function 10 can be performed by a single entity (e.g., a UE, a RAN node, a network node, etc.) or by multiple entities. In this case, the training data 11 and the inference data 12 can be provided from the multiple entities to the model training function 20 and the model inference function 30, respectively.

[0179] As part of the AI model testing process, the model training function 20 is responsible for performing the AI model training, validation, and testing required to generate model performance metrics. If needed, the model training function 20 can handle data preparation (e.g., data pre-processing and cleaning, formation, and transformation) based on the training data 11 provided by the data collection function 10.

[0180] Here, the model deployment / update 13 is used to initially deploy the trained, validated, and tested AI model to the model inference function 30 or to provide an updated model to the model inference function 30.

[0181] The model inference function 30 is responsible for providing the AI model inference output 16 (e.g., a prediction or a decision). If applicable, the model inference function 30 can also provide model performance feedback 14 to the model training function 20. Additionally, if needed, the model inference function 30 can handle data preparation (e.g., data pre-processing and cleaning, formation, and transformation) based on the inference data 12 provided by the data collection function 10.

[0182] Here, the 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 can vary depending on the use case.

[0183] The model performance feedback 14, when available, can be used to monitor the performance of the AI model. However, the feedback can be omitted.

[0184] The executor function 40 receives the output 16 from the model inference function 30 and triggers or performs related tasks / operations. The executor function 40 can trigger tasks / operations of other entities (e.g., one or more UEs, one or more RAN nodes, one or more network nodes, etc.) or its own.

[0185] The feedback 15 can be used to derive the training data 11 and the inference data 12 or to monitor the performance of the AI model, the impact on the network, etc.

[0186] The definitions of training, validation, and testing in the dataset used in AI / ML can be distinguished as follows: - Training data: Training data refers to a dataset used to train a model.

[0187] - Validation data: Validation data refers to a dataset used to validate a trained model. In other words, validation data means a dataset used to prevent overfitting of a typical training dataset.

[0188] In addition, validation data refers to a dataset used to select the best model among a plurality of trained models during the training process. Thus, the validation data can also be considered as a kind of learning.

[0189] - Test data: Test data refers to the dataset used for the final evaluation. Test data is independent of training data.

[0190] In the case of the datasets described above, it is common practice to split the training set so that the training and validation data are divided in an 8:2 or 7:3 ratio. When test data is included, the training, validation, and test data can be divided in a 6:2:2 ratio (training:validation:test).

[0191] Based on the AI / ML capabilities between the BS and UE, the collaboration level can be defined as follows. Multiple levels can also be combined, or a level can be modified by separating any one of them.

[0192] Cat 0a) No collaborative framework: AI / ML algorithms are purely based on the implementation and do not require changes to the wireless interface.

[0193] Cat 0b) This level relates to wireless interfaces suitable for modifications based on efficient implementations of AI / ML algorithms, but this level corresponds to a framework without cooperation.

[0194] Cat 1) provides support between nodes to enhance the AI / ML algorithms of each node. This level is applied 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.

[0195] Cat 2) enables collaborative ML tasks to be performed between the UE and gNB. This level requires AI / ML model command exchange between network nodes.

[0196] Figure 14 The functions shown can be implemented in RAN nodes (e.g., BS, TRP, BS central unit (CU)), network nodes, network operator operations management and maintenance (OAM) or UE.

[0197] Alternatively, two or more entities among the RAN node, network node, network operator's OAM, or UE can collaborate to achieve this. Figure 14 The functions shown. For example, an entity can perform... Figure 14 One entity can perform some of the functions, while another entity can perform the remaining functions. Thus, if Figure 14 If some of the functions shown are performed by a single entity (e.g., UE, RAN node, or network node), then the transmission / providing of data / information between the functions can be omitted. For example, if model training function 20 and model inference function 30 are performed by the same entity, then the transmission / providing of model deployment / update 13 and model performance feedback 14 can be omitted.

[0198] Alternatively, Figure 14Any of the functions shown can be performed by a RAN node, a network node, an OAM of a network operator, or a cooperation among two or more entities among the UE. This can be referred to as split AI operation.

[0199] Figure 15 A scenario is shown in which 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 BS, a TRP, or a CU of a BS).

[0200] Step 1: RAN node 1 and RAN node 2 send input data (i.e., training data) for AI model training to the network node. Here, RAN node 1 and RAN node 2 can send data collected from the UE (e.g., UE measurements related to RSRP, RSRQ, or SINR of a serving cell and neighboring cells, UE location, speed, etc.) to the network node.

[0201] Step 2: The network node trains an AI model using the received training data.

[0202] Step 3: The network node deploys / updates the AI model to RAN node 1 and / or RAN node 2. RAN node 1 (and / or RAN node 2) can continue to perform model training based on the received AI model.

[0203] For convenience of explanation, it is assumed that the AI model is deployed / updated to only RAN node 1.

[0204] Step 4: RAN node 1 receives input data (i.e., inference data) for AI model inference from the UE and RAN node 2.

[0205] Step 5: RAN node 1 performs AI model inference using the received inference data to generate output data (e.g., a prediction or a decision).

[0206] Step 6: If applicable, RAN node 1 can send model performance feedback to the network node.

[0207] Step 7: RAN node 1, RAN node 2, and the UE (or RAN node 1 and the UE or RAN node 1 and RAN node 2) perform an action based on the output data. For example, in the case of load balancing operation, the UE can move from RAN node 1 to RAN node 2.

[0208] Step 8: RAN node 1 and RAN node 2 send feedback information to the network node.

[0209] Figure 16 A scenario is shown in which both the AI model training function and the AI model inference function are performed by a RAN node (e.g., a BS, a TRP, or a CU of a BS).

[0210] Step 1: The UE and RAN node 2 send input data for AI model training (i.e., training data) to RAN node 1.

[0211] Step 2: RAN node 1 trains an AI model using the received training data.

[0212] Step 3: RAN node 1 receives input data for AI model inference (i.e., inference data) from the UE and RAN node 2.

[0213] Step 4: RAN node 1 performs AI model inference using the received inference data to generate output data (e.g., a prediction or a decision).

[0214] Step 5: RAN node 1, RAN node 2, and the UE (or RAN node 1 and the UE or RAN node 1 and RAN node 2) perform an action based on the output data. For example, in the case of a load balancing operation, the UE can move from RAN node 1 to RAN node 2.

[0215] Step 6: RAN node 2 sends feedback information to RAN node 1.

[0216] Figure 17 A scenario is shown in which the AI model training function is performed by a RAN node (e.g., a BS, a TRP, or a CU of a BS), and the AI model inference function is performed by a UE.

[0217] Step 1: The UE sends input data for AI model training (i.e., training data) to a RAN node. Here, the RAN node can collect data from various UEs and / or other RAN nodes (e.g., UE measurements related to RSRP, RSRQ, or SINR of a serving cell and neighboring cells, UE location, speed, etc.).

[0218] Step 2: The RAN node trains an AI model using the received training data.

[0219] Step 3: The RAN node deploys / updates the AI model to the UE. The UE can continue to perform model training based on the received AI model.

[0220] Step 4: The UE receives input data for AI model inference (i.e., inference data) from the RAN node (and / or from other UEs).

[0221] Step 5: The UE performs AI model inference using the received inference data to generate output data (e.g., a prediction or a decision).

[0222] Step 6: If applicable, the UE can send model performance feedback to the RAN node.

[0223] Step 7: The UE and the RAN node perform an action based on the output data.

[0224] Step 8: The UE sends feedback information to the RAN node.

[0225] CPU occupation rule for AI / ML-based CSI reporting

[0226] With the advancement of computing technology and AI / ML technology, nodes and UEs included in a wireless communication network are becoming more and more intelligent and complex. In particular, due to the intelligence of the network, it is expected that various network decision parameter values (e.g., transmission and reception power of each BS, transmission power of each UE, precoder / beam of each BS / UE, time / frequency resource allocation of each UE, duplex mode of each BS, etc.) can be quickly optimized and applied based on various network environment parameters (e.g., distribution / location of BSs, distribution / location / material of buildings / furniture, location / moving direction / speed of UEs, climate information, etc.).

[0227] Hereinafter, a proposal of a CSI processing unit (CPU) when CSI reporting is performed in / by an evolved network / BS / UE based on AI / ML will be described.

[0228] Figure 18 is a diagram for explaining AI-based CSI feedback.

[0229] In a method for reducing CSI feedback payload overhead, common AI models can be based on compression at the UE and / or BS. For example, a CSI encoder of the UE and a CSI decoder of the BS can be assumed as shown in Figure 18 The encoder and the decoder can be related to an autoencoder, and a convolutional neural network (CNN) can be used for efficiently reducing channel size / dimension.

[0230] In AI / ML-based payload reduction, what to compress is a key factor. There are two main types of channel feedback: explicit feedback and implicit feedback. LTE and NR use implicit feedback, in which RI / CQI / PMI is reported to the BS, not the original 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, information about the compressed original channel can be fed back to the gNB, which can result in explicit feedback.

[0231] As another example of AI / ML based CSI feedback, compression of existing PMI, e.g., Type I and Type II CSI, can be considered. In Rel-16 / 17 Type II CSI, a linear vector (e.g., DFT vector) is used for compression in the frequency domain to reduce the payload of Rel-15 Type II CSI. Thus, AI / ML can be used to further reduce the payload of existing CSI and improve its performance.

[0232] As described above, AI / ML based CSI reporting can be considered. Assuming both BS and UE have the capability of AI / ML (training and / or inference), CSI reporting like autoencoder can be performed. Alternatively, one-sided AI / ML (at the network or UE) can be considered. For one-sided AI / ML, specific parameter optimization (performance and / or payload) can be considered to optimize existing legacy codebook based on AI / ML.

[0233] For the generation / computation of CSI (based on AI / ML), to manage the CSI processor of the UE, CPU is introduced in NR Rel.17. Table 6 is an excerpt from TS 38.214 regarding CPU.

[0234] [Table 6]

[0235] Proposal 1

[0236] In the case of AI / ML based CSI reporting, when the UE performs CSI generation through the inference operation of the AI / ML model, the CPU occupancy is assumed to be X, where the value of X is a value configured to the UE by the BS or a pre-defined value.

[0237] AI / ML based CSI reporting can be roughly classified into two types: two-sided AI / ML and one-sided AI / ML. In both cases, it can be assumed that the BS (or network AI / ML server) performs training / inference and signals the trained AI / ML model to the UE, and the UE operates based on only AI / ML inference. As an entity in charge of training, the BS (or network AI / ML server) can configure / indicate the corresponding value based on the computational complexity of inference (floating point operations (FLOP) required based on AI / ML inference) and / or the report of UE capability related to CSI processing / memory. When indicating the value X, the UE assumes the CPU occupancy configured / indicated by the BS as X (e.g., O CPU= X) and compute AI / ML based CSI after CPU occupancy operation of the UE, which is specified in Table 6 extracted from TS 38.214.

[0238] In the case of AI / ML inference, without training burden, variables multiplication between nodes in a pre-trained model and activation functions of individual nodes (e.g., sigmoid, rectified linear unit (ReLU), hyperbolic tangent (tanh), etc.) are processed. Thus, compared to training, the complexity of inference operation is much lower. However, for performance improvement, deeper and larger models are considered. In these cases, it is obvious that the computational complexity increases due to the increase in the number of variables. Thus, in AI / ML models, CPU occupancy can vary according to the amount of computation associated with AI / ML model parameters (e.g., the number of variables, the number of nodes, the number of hidden layers, etc.) and / or the amount of computation involved in pre-processing / post-processing of input / output data. Furthermore, AI / ML models are scalable, and depending on the use case, different AI / ML models can be configured. To reflect the diversity of AI / ML models uniformly, the following proposals are made.

[0239] Proposal 1-1

[0240] The UE can report information on AI / ML models that can be processed with one CPU occupancy to the BS (as UE capability).

[0241] The information on AI / ML models that can be processed with one CPU occupancy in Proposal 1-1 can include some or all of the following: the number of variables, the number of nodes, and the number of hidden layers. Alternatively, a value based on a specific computational unit (e.g., FLOP) that can be processed with one CPU occupancy can be reported.

[0242] As another example, the BS can indicate / configure a threshold value to the UE that can be processed with one CPU occupancy. Since this value can vary depending on the performance of the UE, a relatively loose value can be indicated. Based on UE capability, the threshold value can be divided into multiple groups (e.g., high, medium, low, etc.), and a specific threshold value can be indicated for each group. Alternatively, the BS can configure or predefine the CPU required for CSI reporting based on a specific AI / ML model, and then indicate / configure the CPU of each group to the UE. For example, the value X in Proposal 1 can be configured or predefined as shown in Table 1. UE capability related to AI / ML can be reported by the UE.

[0243] Table 7 shows an example of configuring the value of X based on UE capability.

[0244] [Table 7]

[0245] Alternatively, if a specific AI / ML model or version of a model is specified, the UE can report its capability per model / version.

[0246] Proposal 1-2

[0247] In addition to the CPU, an AI / ML processing unit (APU) can be introduced to allow the UE to efficiently manage the processing of multiple AI / ML models.

[0248] In Proposal 1-2, the APU can allow reporting the number of AI / ML modules / functions (A) that the UE can compute per CC or across CCs simultaneously. If B APUs are occupied, the UE can assume that the UE has A-B unoccupied APUs. In this case, if the number of AI / ML models enabled or triggered exceeds A-B, the selected A-B AI / ML models can be enabled according to a predetermined priority rule, and the UE can report information about the AI / ML models that are not selected to the BS. It can be assumed that the APU is occupied from the time at which signaling (e.g., MAC-CE / DCI) to trigger / enable a specific AI / ML model is received until the time at which deactivation, termination, or switching signaling is received.

[0249] Alternatively, the APU can be managed / applied integrally via the AI / ML model, and the number of APUs required per AI / ML model can vary.

[0250] As another example, it can be considered that the APU is separately configured and indicated for AI / ML-based CSI reporting, and thus the UE can report both the maximum supported APU and CPU values as capability. In this case, the APU occupation rule is the same as that of the CPU, but the APU occupation rule is exclusively used to manage AI / ML-based CSI reporting. The declaration that the APU occupation rule is the same as that of the CPU means that, as in the conventional system, one CSI report associated with one CSI-RS resource occupies K APUs (e.g., K=1). Alternatively, since the APU can be used across multiple CSI resources as in the case of the UE's CSI prediction, the K value can have a value of 0.5. In this case, it can be assumed that one CSI report associated with two resources occupies one APU.

[0251] The main use case of the APU can be for AI / ML inference. However, if the UE has a training capability, a separate unit for managing training and inference can be introduced. For example, the APU can be managed separately as APU-training and APU-inference. In this case, if the UE’s APU-inference is all occupied and some or all of the APU-training is not occupied, the APU-training can be borrowed and used according to certain rules. Conversely, if the UE’s APU-training is all occupied and some or all of the APU-inference is not occupied, the APU-inference can be borrowed and used according to certain rules. The certain rules refer to rules that exchange the use of the APU-training and the APU-inference. For example, the individual APU occupation can be described by a function (e.g., a linear equation or a polynomial equation). The following linear equation can be used: w APU-inference + b = APU-training, where w is a coefficient for matching the capability of the APU-training and the APU-inference, and b is an offset. For example, if w = 2 and b = 0, it means that the APU required for the processing of the APU-training is twice the APU required for the APU-inference.

[0252] Further, when the APU-training and the APU-inference are integrated and utilize a single APU operation / application, if there is an APU shortage, the training can take priority over the inference and occupy the APU (or the inference can return the APU already occupied by the inference for use by the training) (a reverse operation is also possible). If the inference is not allowed due to the APU shortage, a default mode can be configured. For example, in the case of a CSI report, it can fall back to a Type I CSI report as a default mode. Alternatively, if the inference is not allowed due to the APU shortage, the AI / ML model inference can be discarded or delayed to apply the inference. Here, “delayed” means waiting until an unoccupied APU equal to or greater than the APU occupied by the AI / ML model is available. However, if the delay is too long, problems such as buffer size, memory size, etc. can occur. Therefore, a maximum window size or a timer can be introduced to limit the delay. The waiting can be performed until the window or the timer expires. If an unoccupied APU is secured before the expiration, the AI / ML inference can be applied. Otherwise, the AI / ML inference can be discarded. Information about the discard or delay of the AI / ML model can be reported to the BS or the AI / ML server.

[0253] Proposal 1-3

[0254] If the current APU occupation value of the UE is 0 (i.e., L = 0) and the BS triggers the training / update / inference of the AI / ML model, in order for faster training / update / inference of the AI / ML model, it is assumed that the UE occupies as many APU as the UE is capable of processing at the same time per CC or across CCs (N APU: UE indicates the number of simultaneous AI / ML model training / update / inference supported N APU . In other words, for fast training / update / inference of a particular AI / ML model, the APU occupancy is O APU = N APU .

[0255] In Proposal 1-3, for fast training / update / inference, the BS can configure / indicate the UE to occupy all necessary APUs. Alternatively, if there is a priority for a particular AI / ML model, it can be agreed that Proposal 1-3 is only applied when the training / update / inference of the AI / ML model with the highest priority is triggered.

[0256] Proposal 2

[0257] When AI / ML functions are modularized and multiple AI / ML models (e.g., Y) are involved in a single CSI report, the CPU occupancy is considered to be Y or Y+1.

[0258] Typically, CSI reporting is performed by measuring raw channel data through channel measurement (CMR) and / or interference measurement (IMR) based on a specific RS (e.g., CSI-RS or SSB) configured by the BS. The data is then pre-processed as needed (e.g., eigenvalue decomposition), and based on the pre-processing, the preferred CRI / RI / CQI / PMI or SSBRI / CRI / L1-RSRP / L1-SINR is calculated. The entire process can be handled with a single AI / ML model. Alternatively, the process can be divided into multiple modules, and modules / functions with higher computational complexity or that can be replaced with lower complexity modules / functions that provide moderate performance can be replaced by AI / ML models. Thus, the number of modules (Y) involved in the modularized AI / ML-based CSI reporting is considered as the CPU occupancy. If these modules are mixed with other legacy modules / functions, an additional count of 1 can be added, and this is considered as Y+1. Even if a particular module (e.g., channel estimation) is implemented in the UE, the UE can report information (e.g., the number of modules or model information) to the BS, which supports the scheduling flexibility of the BS.

[0259] Proposal 3

[0260] If the current CPU occupancy value of the UE is 0 (i.e., L=0) and the BS triggers AI / ML-based CSI reporting, for faster AI / ML-based CSI reporting, it is assumed that the UE occupies as many CPUs as the UE is capable of processing simultaneously per CC or across CCs (N CPU : UE indicates the number of simultaneous AI / ML model training / update / inference supported N simultaneousCSI-ReportsPerCC with the parameters simultaneousCSI-ReportsAllCC across all component carriersCPU ). In other words, for AI / ML based CSI reporting, CPU occupancy is O CPU = N CPU .

[0261] In case of AI / ML based CSI computation, if the size of AI / ML model performing inference is not significant, the amount of computation can be low, allowing faster CSI reporting. In particular, in a bilateral model where both BS and UE use AI / ML model to perform inference, the UE can occupy all available CPU to perform fast AI / ML based CSI reporting as described in Proposal 3.

[0262] As another example, in Proposal 3, N CPU values can be separately configured and managed for AI / ML based CSI reporting and legacy CSI reporting. For example, N CPU _ AI , and N CPU _ AI can be exclusively used for AI / ML based CSI reporting. Alternatively, if a certain number of AI / ML based CSI reporting is triggered (e.g., more than one report), N CPU can be replaced with N CPU _ AI . Here, N CPU _ AI can be a value determined based on UE capability report. Alternatively, BS can define multiple values for N CPU and indicate to UE which value to use. This is used to allow BS to configure / indicate a specific N CPU value according to CSI reporting scheduled by BS, increasing flexibility of both BS and UE. Alternatively, multiple values for N CPU _ AI can be configured to UE and applied exclusively to different AI / ML model groups (AI / ML model (group) specific). For example, if two values: N CPU_AI-1 and N CPU_AI-2 are applied, and there are multiple AI / ML models, the multiple AI / ML models can be mapped to each of N CPU_AI-1 and N CPU_AI-2 for use. In other words, for AI / ML models belonging to group 1, CPU occupancy is managed separately with N CPU_AI-1 as upper limit, while for AI / ML models in group 2, CPU occupancy is managed separately with N CPU_AI-2 as upper limit. Information about such grouping can be based on UE capability report or configured / indicated by BS.

[0263] Proposal 4

[0264] To enhance scheduling flexibility, the BS can adjust the CSI priority of the AI / ML based CSI report by additionally signaling a CSI priority indicator.

[0265] 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 certain CSI report can be dropped based on the following priority rules. Regarding the CSI priority rules, according to the rules in Table 8 extracted from the standard document, Pri iCSI(y,k,c,s) The CSI report with a lower value of k has a higher priority. To adjust the priority of the AI / ML based CSI report based on this structure, the BS can add a 1-bit field to the DCI scheduling the CSI report and indicate whether the priority is high or low. For example, the priority adjustment can be performed as follows: by setting k to 0 or 1 in Equation 2 below, a higher or equal priority can be indicated compared to the existing normal CSI report (e.g., Type 1 CSI, Type 2 CSI, etc.). As another example, the range of the k value in Equation 2 can be expanded to 0, 1, and 2, allowing higher priority flexibility. In this case, the field size of the priority can be increased to 2 bits. The indication in Proposal 4 can be specified differently for each AI / ML model.

[0266] [Equation 2]

[0267] Pri iCSI (y,k,c,s) = 3 N cells M s y + N cells M s k + M s c + s

[0268] Alternatively, by setting k = 0 for the AI / ML based CSI report, the AI / ML based CSI report can be configured to have the same priority as the beam report.

[0269] [Table 8]

[0270] Proposal 5

[0271] To support fast CSI reporting in AI / ML model inference based CSI reporting of a UE, a value (e.g., Z, Z', or CSI reference resource timing) smaller than previously defined CSI computation delay can be defined / configured.

[0272] Proposal 5 aims to support the case where model training is done and the UE needs to perform fast CSI reporting based on AI / ML model inference. As proposed below, a value smaller than those in Table 5.4-1 or Table 5.4-2 of TS 38.214 can be defined and used for AI / ML based CSI reporting, as shown in Table 9 below. Alternatively, Z / Z' values from Table 5.4-1 can be used for AI / ML based CSI reporting. Proposal 5 can be further extended by defining and using specific Z / Z' values depending on AI / ML model or model group. Model group refers to a group of AI / ML models that can perform CSI computation within a specific range of Z / Z' values. If the UE reports information related to corresponding Z / Z' value as capability, the BS can configure which Z / Z' value to use based on the information.

[0273] [Table 9]

[0274] As described above, when a network or BS to which AI technology is applied or a network / BS with advanced performance (achieved by other means) is implemented, integrated control between BSs and / or real-time cooperation is expected to allow dynamic optimization of interference and traffic. The present disclosure proposes a resource control and CSI related signaling method suitable for the above purpose. Although the present disclosure is described based on a 3GPP NR system within the current system architecture, it is obvious that the present disclosure is also applicable to other communication systems.

[0275] [APU occupation for inactive AI / ML model]

[0276] When an AI / ML model processing unit (APU) is introduced to efficiently manage AI / ML models, the present disclosure proposes an occupancy rule for efficient management of the APU. In particular, the present disclosure proposes an APU management method for inactive AI / ML models / functions during model monitoring, aiming at fast activation, switching, or updating of inactive models. Inactive AI / ML models can refer to models that are not currently in an active state among one or more AI / ML models registered, deployed, or implemented in a UE (or network) (e.g., models that have received an activation command or are currently performing AI / ML inference / monitoring / training operations).

[0277] Proposal 6

[0278] Information on whether monitoring is performed per model, per model group, or per AI / ML function in the inactive mode can be signaled based on at least some of the following implementations.

[0279] 1) Information on whether monitoring can be performed in the inactive AI / ML model can be reported through the capability report of the UE. For example, the information can be reported per AI / ML model, per AI / ML model group, or per AI / ML function, and the following is an example of reporting capability per AI / ML model.

[0280] - AI / ML Model #1 - capable of monitoring in the inactive model state

[0281] - AI / ML Model #2 - capable of monitoring in the inactive model state

[0282] - AI / ML Model #3 - not capable of monitoring in the inactive model state

[0283] 2) As another implementation, based on the capability report of the UE, the BS can indicate, through higher layer signaling (e.g., RRC, MAC-CE, DCI, or a combination of signaling), the AI / ML model, AI / ML model group, or AI / ML function for which monitoring is to be performed even when the model state is inactive. Alternatively, the active / inactive / idle states can be defined for the AI / ML model as shown below, and monitoring can be indicated based on these states.

[0284] i) AI / ML Model / Function Active State: a state in which an AI / ML model / function enable command is received or an operation such as AI / ML inference, monitoring, or training is currently being performed

[0285] ii) AI / ML Model / Function Inactive State: a state in which an AI / ML model / function enable command is not received or an operation such as AI / ML model inference, monitoring, or training is not being performed, but reporting of monitoring information is allowed (based on BS indication or configuration)

[0286] iii) AI / ML Model / Function Idle State: a state in which an AI / ML model / function enable command is not received and an operation such as AI / ML inference, monitoring, or training is not being performed, and reporting of monitoring information is not allowed (based on BS indication or configuration)

[0287] As mentioned above, AI / ML model monitoring can be performed even in the inactive state of AI / ML model. In this case, the inactive AI / ML model needs to occupy the APU. The APU occupancy priority of AI / ML model / model group / function can correspond to the following order: i) AI / ML model / function active state > ii) AI / ML model / function inactive state > iii) AI / ML model / function idle state. In each state, the priority can be determined per model according to certain rules (e.g., trigger order, order of lowest / highest model ID, or order of recent inference / monitoring).

[0288] In this case, the duration of APU occupancy can not be explicit, the following Proposal 7 addresses the APU occupancy duration.

[0289] Proposal 7

[0290] The APU occupancy duration of inactive AI / ML model, model group, or function can be controlled according to the following two methods.

[0291] i) Method 1: When the BS orders / control the monitoring or evaluation of (a particular) inactive AI / ML model, model group, or function, the APU can be occupied from the time point when the BS orders the UE to start model monitoring / evaluation to the time point when the UE reports the corresponding result.

[0292] ii) Method 2: When the UE determines to perform monitoring or evaluation on (a particular) inactive AI / ML model, model group, or function, the UE can use the idle APU to perform monitoring or evaluation on the inactive model with the best effort.

[0293] In the case of Method 1, the BS orders or controls the model monitoring or evaluation of the inactive AI / ML model, model group, or function. In this case, for the inactive model, model group, or AI / ML function, the UE performs monitoring or evaluation on the corresponding model from the time point n (or n+X) when the UE receives a command from the BS indicating that the UE performs monitoring or evaluation (e.g., a command via MAC-CE or DCI), and the UE can occupy the APU until the time point m when the reported value is reported to the BS (or the time point when the response to the report from the BS is received). The above-mentioned time points can be slot-level or symbol-level, the values X and Y can be based on UE capability (e.g., UE processing time), can be configured by the BS, or can be predetermined.

[0294] Alternatively, the UE can perform monitoring or evaluation on the model only without reporting the corresponding result, or can collect multiple results and report the results to the BS at once. In this case, a designated time such as a specific timer or monitoring duration can be defined. For example, APU occupation from the point in time n (or n+X) at which a command related to AI / ML monitoring or evaluation is received until the point in time at which the timer expires, or for a predetermined / configured duration Z, can be defined / configured.

[0295] In the case of Method 2, the UE determines to perform model monitoring or evaluation on an inactive AI / ML model, model group, or function. In this case, for model monitoring or evaluation of an inactive model, model group, or AI / ML function, the UE utilizes an idle APU among all APUs based on UE capability and performs AI / ML model monitoring or evaluation with the greatest effort. If the BS triggers (new or additional) AI / ML-based operation to the UE during model monitoring or evaluation of an inactive model, model group, or AI / ML function, the UE can need to stop, hold, or suspend ongoing model monitoring or evaluation of an inactive AI / ML model, model group, or function, and give priority to handling the AI / ML-based operation triggered by the BS. In this case, the idle APU of the UE is occupied by the (active) AI / ML model, model group, or function triggered by the BS (new or additional). Also in this case, the active AI / ML model occupies the APU with higher priority than the inactive AI / ML model. As another implementation, when the BS triggers an AI / ML-based operation while the UE performs monitoring or evaluation of an inactive AI / ML model, model group, or function based on an idle APU, the UE can determine to ignore the BS trigger and report that the trigger is ignored. Alternatively, the UE can delay application of the trigger rather than immediately applying the trigger, and can report information about the delay to the BS. For example, when the BS triggers (new or additional) AI / ML-based operation at a time close to when APU occupation for monitoring or evaluation of an inactive AI / ML model, model group, or function is completed, such an operation can be useful. In this case, discarding processing already performed for monitoring or evaluation of an inactive model can be inefficient, and thus the UE can ignore the trigger and complete the ongoing operation.

[0296] In Method 2, when the UE periodically, semi-periodically, or non-periodically reports monitoring or evaluation results of inactive AI / ML models, model groups, or functions to the BS, insufficient idle APU may prevent monitoring or evaluation of some or all models. To address this issue, when reporting monitoring or evaluation results of inactive AI / ML models, model groups, or functions, the UE can separately report the inactive AI / ML models, model groups, or functions that were monitored or evaluated and those that were not. When monitoring or evaluation was not performed, the UE can report a separate reason (e.g., insufficient APU). Alternatively, after the monitoring or evaluation of inactive AI / ML models, model groups, or functions is completed, the UE can report the results all at once via MAC-CE.

[0297] For smooth operation of Method 2, the UE may report information to the BS related to currently occupied APUs (e.g., the number of occupied APUs, the time points when the APUs were occupied). Reports may include periodic, semi-periodic, non-periodic, and event-based reports. For example, for event-based reports, specific thresholds for monitoring results may be predefined or configurable by the BS. Alternatively, instead of reporting information about occupied APUs, the UE may report information about the number of currently available APUs. Furthermore, when reporting occupied or unoccupied APUs, the UE may also report information about the (expected) occupied or unoccupied APUs at various time points. For example, the UE may report {the number of APUs remaining at t1, the number of APUs remaining at t2, the number of APUs remaining at t3, ...} as a set. Time points t1, t2, t3, ... may be predefined or configurable by the BS. Alternatively, the UE may additionally report information about time points, such as start time or duration (offset) values.

[0298] Proposal 8

[0299] When monitoring or evaluation is performed on inactive AI / ML models, model groups, or functions as described in Proposals 6 and 7, it is necessary to determine the priority among the candidate AI / ML models, model groups, or functions if multiple candidate AI / ML models, model groups, or functions are being monitored or evaluated.

[0300] The monitoring or evaluation priority of multiple inactive AI / ML models, model groups, or AI / ML functions can be determined based on at least one of the following three methods.

[0301] i) Method 1: Priority determined by the UE

[0302] ii) Method 2: Priority determined by network (NW) indication

[0303] iii) Method 3: Priority determined by priority rules

[0304] In case of Method 1, the UE determines the priority and occupies the APU to perform monitoring. In order to inform the BS which models the UE has decided to monitor, the UE can explicitly report the model, model group, or function ID to the BS, and such reporting can be sent separately or together with additional information (e.g., monitoring results or the number of occupied / unoccupied APUs).

[0305] In case of Method 2, the NW can specify which AI / ML models, model groups, or functions to monitor. For example, the BS can semi-statically configure the AI / ML models, model groups, or functions through RRC signaling, or can dynamically specify the AI / ML models, model groups, or functions through MAC-CE or DCI. This information can be sent to the UE together with the monitoring trigger information.

[0306] In case of Method 3, the priority of inactive AI / ML models, model groups, or functions can be determined based on certain priority rules. For example, each AI / ML use case can be assigned a priority. Mobility-related use cases and / or beam management (BM) use cases can have higher priority to occupy the APU than other use cases (e.g., CSI feedback or positioning). In addition, for the same use case, the priority rules can be defined according to the model-specific requirements or performance of certain AI / ML models, model groups, or functions. For example, in the CSI use case, if model #A provides higher accuracy (or higher compression ratio) than model #B, the priority rules can be defined such that model #A is assigned a higher priority than model #B.

[0307] For example, the priority of inactive AI / ML models, model groups, or functions can be determined according to Method 3. The UE can perform model monitoring or evaluation by applying the APU occupation rules according to the priority. In addition, the predefined or configured priority can also change in real-time or dynamically based on whether the UE has recently performed monitoring or evaluation. For example, the UE can assign a higher priority to inactive AI / ML models, model groups, or functions that have not been recently monitored or evaluated, allowing for monitoring or evaluation of inactive models in a fair manner. Although the above priority determination methods are described for inactive models or functions, the same methods can also be applied to active models or functions.

[0308] Additionally, when the priority is determined according to Method 3, there can be ambiguity on how the UE should operate if the first inactive AI / ML model, model group, or function with higher priority is triggered later than the second inactive AI / ML model, model group, or function with lower priority. If the second inactive AI / ML model, model group, or function has occupied the APU, the UE will not discard the monitoring or evaluation of the second inactive AI / ML model, model group, or function even when the first inactive AI / ML model, model group, or function with higher priority is triggered. In this case, the UE can delay the monitoring or evaluation of the newly triggered first inactive AI / ML model, model group, or function and perform the monitoring or evaluation once the APU becomes available. Alternatively, the UE can discard the second inactive AI / ML model, model group, or function in order of priority when the first inactive AI / ML model, model group, or function with higher priority is triggered. Then, after ensuring the APU, the UE can perform the monitoring or evaluation of the first inactive AI / ML model, model group, or function with higher priority.

[0309] Proposal 9

[0310] When the inactive AI / ML model, model group, or function is managed with the active AI / ML model, model group, or function in a common APU, the APU occupation for the monitoring or evaluation of the inactive AI / ML model, model group, or function can occupy less APU than in the active case. For example, if one active AI / ML model or function occupies one APU, the inactive AI / ML model or function can occupy a APU (a < 1), the value of a can be determined based on the UE capability, the configuration of the BS, or a predefined agreement.

[0311] This is because the processing capability required for monitoring can be more simply implemented or operated for the inactive AI / ML model, model group, or function than for the active model. Alternatively, instead of using a common APU for the active and inactive AI / ML models or functions, the APU can be managed separately for each active or inactive AI / ML model or function. For example, the UE can manage the APU by distinguishing the APU into APU active and APU inactive.

[0312] In addition to the inactive model, the APU occupation method proposed above can also be extended and applied to the active model. For example, the APU occupation time can be determined as follows. The UE can perform training, inference, or monitoring from the time point n (or n+X) at which the UE receives an enabling command indicating the training, inference, or monitoring of the AI / ML model, model group, or AI / ML function from the BS (via MAC-CE or DCI). The UE can occupy the APU for the AI / ML model, model group, or AI / ML function until the time point m (or n+Y) at which a disabling, switching, or updating command for the model is subsequently received from the BS. In the case of model monitoring, the APU can be occupied until the time point m (or the time point at which the UE receives a response to the report from the BS or n+Y) at which the UE sends the corresponding report value to the BS. The above time points can be slot-level or symbol-level, and the values of X and Y can be based on the UE capability (e.g., UE processing time), can be configured by the BS, or can be predetermined.

[0313] As described above, when a network or BS that applies AI / ML technology or a network or BS that has enhanced performance (through other means) is implemented, through real-time cooperation and / or integrated control between BSs, interference or traffic can be dynamically optimized. The present disclosure proposes a control and AI / ML-related signaling method suitable for the above operation. Although described based on the 3GPP system for convenience, the embodiments are not limited thereto and can be applied to various other communication systems. In addition, Proposals 1 to 9 can be independently applied or applied in any combination.

[0314] Figure 19 is a diagram for explaining the operation of a network and a UE in a wireless communication system according to an embodiment.

[0315] Referring to Figure 19 , the network can perform high-layer signaling to the UE at least once (A05). The high-layer signaling can include a configuration for the AI / ML model. The UE can configure the AI / ML model based on the high-layer signaling.

[0316] The UE can enable at least one configured AI / ML model (A10). For example, the enabling or disabling of the AI / ML model can be performed based on a network indication such as MAC or DCI. For convenience of explanation, the AI / ML model in the inactive state is referred to as a first AI / ML model, and the AI / ML model in the active state is referred to as a second AI / ML model.

[0317] The UE can monitor at least one configured AI / ML model (A15).

[0318] The AI / ML model monitored by the UE can include the first AI / ML model. The first AI / ML model can be determined according to a network indication or can be autonomously determined by the UE.

[0319] The AI / ML model monitored by the UE can include a second AI / ML model.

[0320] The first AI / ML model in the inactive state can occupy the APU for a shorter duration than the second AI / ML model in the active state, or can occupy less APU than the second AI / ML model in the active state.

[0321] The monitoring of the first AI / ML model can be performed based on remaining APU not occupied by the second AI / ML model in the active state.

[0322] The second AI / ML model in the active state can occupy the APU with a higher priority than the first AI / ML model in the inactive state.

[0323] The monitoring of the second AI / ML model in the active state can occupy one APU, and the monitoring of the first AI / ML model in the inactive state can occupy n APUs, where n can be a real number greater than 0 and less than or equal to 1.

[0324] Here, n can be determined based on a UE capability, or can be provided through network signaling.

[0325] The UE can transmit, to the network, a result of monitoring the AI / ML model (A20).

[0326] Based on the result of monitoring the AI / ML model, the UE and the network can perform a lifecycle management (LCM) procedure on the AI / ML model (A25).

[0327] A start of a duration of occupation of the APU due to the monitoring of the first AI / ML model can be determined based on a time point at which the first AI / ML model is instructed to be monitored.

[0328] An end of the duration of occupation of the APU due to the monitoring of the first AI / ML model can be determined based on at least one of: i) a time point at which a result of monitoring the first AI / ML model is reported (A20); ii) a time point at which a network response to the reporting is received (A20); or iii) a time point at which a predetermined timer expires.

[0329] The UE can report, to the network, information on an occupied state or an unoccupied state of the APU.

[0330] Figure 20 A flow illustrating a method performed by a UE according to an embodiment is shown.

[0331] Referring to Figure 20 , the UE can obtain a configuration for the AI / ML model based on higher layer signaling (B05).

[0332] The UE can monitor the at least one AI / ML model (B10).

[0333] Based on monitoring the at least one AI / ML model, the UE can perform AI / ML model management for handover or update (B15).

[0334] Monitoring the at least one AI / ML model can include at least one of: monitoring a first AI / ML model in an inactive state; and monitoring a second AI / ML model in an active state.

[0335] Monitoring the first AI / ML model in the inactive state can occupy the UE processing unit for a shorter duration or can occupy less of the UE processing unit than monitoring the second AI / ML model in the active state.

[0336] The first AI / ML model among the AI / ML models in the inactive state to be monitored can be determined based on network signaling.

[0337] A start of a duration of occupation of the UE processing unit due to monitoring the first AI / ML model can be determined based on a time point of receiving network signaling.

[0338] An end of the duration of occupation of the UE processing unit due to monitoring the first AI / ML model can be determined based on at least one of: i) a time point of reporting a result of monitoring the first AI / ML model; ii) a time point of receiving a network response to the reporting; or iii) a time point of expiration of a predetermined timer.

[0339] The first AI / ML model among the AI / ML models in the inactive state to be monitored can be determined by the UE.

[0340] Based on triggering, by network signaling, operation for the second AI / ML model in the active state, monitoring the AI / ML models in the inactive state can be stopped or paused.

[0341] The UE can report information about an occupied state or an unoccupied state of the UE processing unit to the network.

[0342] Monitoring the first AI / ML model can be performed based on a remaining UE processing unit that is not occupied by the second AI / ML model in the active state.

[0343] The second AI / ML model in the active state can occupy the UE processing unit with a higher priority than the first AI / ML model in the inactive state.

[0344] The monitoring of the second AI / ML model in the active state can occupy one UE processing unit, and the monitoring of the first AI / ML model in the inactive state can occupy n UE processing units, where n can be a real number greater than 0 and less than or equal to 1.

[0345] Here, n can be determined based on a UE capability, or can be provided through network signaling.

[0346] Figure 21 A communication system 1 to which the present disclosure is applied is illustrated.

[0347] Referring to Figure 21 The communication system 1 to which the present disclosure is applied includes wireless devices, base stations (BSs), and a network. Herein, a wireless device denotes a device that performs communication using a radio access technology (RAT) (e.g., 5G new RAT (NR) or long term evolution (LTE)) and can be referred to as a communication / radio / 5G device. The wireless device can include, but is not limited to, a robot 100a, vehicles 100b-1 and 100b-2, an extended reality (XR) device 100c, a handheld device 100d, a home appliance 100e, an Internet of Things (IoT) device 100f, and an artificial intelligence (AI) device / server 400. For example, the vehicles can include vehicles with wireless communication functionality, self-driving vehicles, and vehicles capable of performing communication between vehicles. Herein, the vehicles can include unmanned aerial vehicles (UAVs) (e.g., drones). The XR device can include an augmented reality (AR) / virtual reality (VR) / mixed reality (MR) device and can be implemented in the form of a head-mounted device (HMD), a head-up display (HUD) mounted in a vehicle, a television, a smartphone, a computer, a wearable device, a home appliance device, a digital signage, a vehicle, a robot, etc. The handheld device can include a smartphone, a smartpad, a wearable device (e.g., a smartwatch or smartglasses), and a computer (e.g., a notebook). The home appliance can include a TV, a refrigerator, and a washing machine. The IoT device can include a sensor and a smartmeter. For example, the BSs and the network can be implemented as wireless devices, and a specific wireless device 200a can operate as a BS / network node with respect to other wireless devices.

[0348] The wireless devices 100a-100f can be connected to the network 300 via the BS 200. The AI technology can be applied to the wireless devices 100a-100f, and the wireless devices 100a-100f can be connected to the AI server 400 via the network 300. The network 300 can be configured using a 3G network, a 4G (e.g., LTE) network, or a 5G (e.g., NR) network. Although the wireless devices 100a-100f can communicate with each other through the BS 200 / network 300, the wireless devices 100a-100f can perform direct communication (e.g., sidelink communication) with each other without going through the BS / network. For example, the vehicles 100b-1 and 100b-2 can perform direct communication (e.g., vehicle-to-vehicle (V2V) / vehicle-to-everything (V2X) communication). The IoT devices (e.g., sensors) can perform direct communication with other IoT devices (e.g., sensors) or other wireless devices 100a-100f.

[0349] Wireless communication / connection 150a, 150b, or 150c can be established between the wireless devices 100a-100f / BS 200 or the BS 200 / BS 200. Herein, the wireless communication / connection can be established by various RATs (e.g., 5G NR) such as uplink / downlink communication 150a, sidelink communication 150b (or D2D communication), or inter-BS communication (e.g., relay, integrated access backhaul (IAB)). The wireless devices and the BS / wireless devices can transmit / receive radio signals to / from each other through the wireless communication / connection 150a and 150b. For example, the wireless communication / connection 150a and 150b can transmit / receive signals through various physical channels. To this end, at least a part of various configuration information of procedures for transmitting / receiving radio signals, various signal processing procedures (e.g., channel coding / decoding, modulation / demodulation, and resource mapping / demapping), and resource allocation procedures can be performed based on various proposals of the disclosure.

[0350] Figure 22 A wireless device suitable for the disclosure is illustrated.

[0351] Referring to Figure 22 , the first wireless device 100 and the second wireless device 200 can transmit radio signals through various RATs (e.g., LTE and NR). Herein, {the first wireless device 100 and the second wireless device 200} can correspond to Figure 21 {the wireless device 100x and the BS 200} and / or {the wireless device 100x and the wireless device 100x}.

[0352] The first wireless device 100 can include one or more processors 102 and one or more memories 104, and additionally further include one or more transceivers 106 and / or one or more antennas 108. The processor(s) 102 can control the memory(s) 104 and / or the transceiver(s) 106, and can be configured to implement the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed in the present document. For example, the processor(s) 102 can process information within the memory(s) 104 to generate first information / signals, and then transmit radio signals including the first information / signals through the transceiver(s) 106. The processor(s) 102 can receive radio signals including second information / signals through the transceiver(s) 106, and then store information obtained by processing the second information / signals in the memory(s) 104. The memory(s) 104 can be connected to the processor(s) 102, and can store various information related to operations of the processor(s) 102. For example, the memory(s) 104 can store software code including commands for executing some or all of the procedures controlled by the processor(s) 102 or for executing the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed in the present document. Herein, the processor(s) 102 and the memory(s) 104 can be a part of a communication modem / circuitry / chip designed to implement a RAT (e.g., LTE or NR). The transceiver(s) 106 can be connected to the processor(s) 102 and transmit and / or receive radio signals through the one or more antennas 108. Each of the transceiver(s) 106 can include a transmitter and / or a receiver. The transceiver(s) 106 can be used interchangeably with radio frequency (RF) unit(s). In the present disclosure, a wireless device can represent a communication modem / circuitry / chip.

[0353] The second wireless device 200 can include one or more processors 202 and one or more memories 204, and additionally further include one or more transceivers 206 and / or one or more antennas 208. The processor(s) 202 can control the memory(s) 204 and / or the transceiver(s) 206, and can be configured to implement the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed in the present document. For example, the processor(s) 202 can process information within the memory(s) 204 to generate third information / signal, and then transmit a radio signal including the third information / signal through the transceiver(s) 206. The processor(s) 202 can receive a radio signal including fourth information / signal through the transceiver(s) 206, and then store information obtained by processing the fourth information / signal in the memory(s) 204. The memory(s) 204 can be connected to the processor(s) 202, and can store a variety of information related to the operation of the processor(s) 202. For example, the memory(s) 204 can store software code including commands for executing some or all of the procedures controlled by the processor(s) 202 or for executing the descriptions, functions, procedures, proposals, methods and / or operational flowcharts disclosed in the present document. Herein, the processor(s) 202 and the memory(s) 204 can be a part of a communication modem / circuitry / chip designed to implement a RAT (e.g., LTE or NR). The transceiver(s) 206 can be connected to the processor(s) 202 and transmit and / or receive a radio signal through the one or more antennas 208. Each of the transceiver(s) 206 can include a transmitter and / or a receiver. The transceiver(s) 206 can be used interchangeably with RF unit(s). In the present disclosure, a wireless device can represent a communication modem / circuitry / chip.

[0354] Hereinafter, the hardware elements of the wireless devices 100 and 200 will be described more specifically. One or more protocol layers can be implemented by, but not limited to, the one or more processors 102 and 202. For example, the one or more processors 102 and 202 can implement one or more layers (e.g., functional layers such as PHY, MAC, RLC, PDCP, RRC, and SDAP). The one or more processors 102 and 202 can 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 operational flowcharts disclosed in the present document. The one or more processors 102 and 202 can generate messages, control information, data, or information according to the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document. The one or more processors 102 and 202 can generate signals (e.g., baseband signals) including the PDUs, SDUs, messages, control information, data, or information according to the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document, and provide the generated signals to the one or more transceivers 106 and 206. The one or more processors 102 and 202 can receive signals (e.g., baseband signals) from the one or more transceivers 106 and 206 and acquire the PDUs, SDUs, messages, control information, data, or information according to the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document.

[0355] The one or more processors 102 and 202 can be referred to as controllers, microcontrollers, microprocessors, or microcomputers. The one or more processors 102 and 202 can be implemented by hardware, firmware, software, or a combination thereof. As an 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) can be included in the one or more processors 102 and 202. The descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document can be implemented using firmware or software, and the firmware or software can be configured to include modules, procedures, or functions. Firmware or software configured to perform the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document can be included in the one or more processors 102 and 202 or stored in the one or more memories 104 and 204 to be driven by the one or more processors 102 and 202. The descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document can be implemented using firmware or software in the form of codes, commands, and / or command sets.

[0356] One or more memories 104 and 204 can be connected to the one or more processors 102 and 202 and store various types of data, signals, messages, information, programs, codes, instructions, and / or commands. The one or more memories 104 and 204 can be configured by read-only memory (ROM), random-access memory (RAM), electrically programmable read-only memory (EPROM), flash memory, hard disk drive, register, cache memory, computer readable storage media, and / or a combination thereof. The one or more memories 104 and 204 can be located internal and / or external to the one or more processors 102 and 202. The one or more memories 104 and 204 can be connected to the one or more processors 102 and 202 through various technologies, such as wired or wireless connections.

[0357] The one or more transceivers 106 and 206 can transmit user data, control information, and / or radio signals / channels mentioned in the methods and / or operational flowcharts of the present document, to one or more other apparatuses. The one or more transceivers 106 and 206 can receive user data, control information, and / or radio signals / channels mentioned in the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document, from one or more other apparatuses. For example, the one or more transceivers 106 and 206 can be connected to the one or more processors 102 and 202 and transmit and receive radio signals. For example, the one or more processors 102 and 202 can perform control so that the one or more transceivers 106 and 206 can transmit user data, control information, or radio signals to one or more other apparatuses. The one or more processors 102 and 202 can perform control so that the one or more transceivers 106 and 206 can receive user data, control information, or radio signals from one or more other apparatuses. The one or more transceivers 106 and 206 can be connected to the one or more antennas 108 and 208, and the one or more transceivers 106 and 206 can be configured to transmit and receive user data, control information, and / or radio signals / channels mentioned in the descriptions, functions, procedures, proposals, methods, and / or operational flowcharts disclosed in the present document, through the one or more antennas 108 and 208. In the present document, the one or more antennas can be a plurality of physical antennas or a plurality of logical antennas (for example, antenna ports). The one or more transceivers 106 and 206 can convert received radio signals / channels, etc. from RF band signals to baseband signals in order to process received user data, control information, radio signals / channels, etc. using the one or more processors 102 and 202. The one or more transceivers 106 and 206 can convert user data, control information, radio signals / channels, etc. processed using the one or more processors 102 and 202 from baseband signals to RF band signals. To this end, the one or more transceivers 106 and 206 can include (analog) oscillators and / or filters.

[0358] Figure 23 Another example applied to the wireless device of the present disclosure is illustrated. The wireless device can be implemented in various forms according to use cases / services (refer to Figure 21 ).

[0359] Referring to Figure 23 , the wireless devices 100 and 200 can correspond to Figure 22The wireless devices 100 and 200 are configured with various elements, components, units / parts, and / or modules. For example, each of the wireless devices 100 and 200 may include a communication unit 110, a control unit 120, a memory unit 130, and an additional component 140. The communication unit may include a communication circuit 112 and a transceiver 114. For example, the communication circuit 112 may include... Figure 22 One or more processors 102 and 202 and / or one or more memories 104 and 204. For example, transceiver 114 may include Figure 22 The device comprises one or more transceivers 106 and 206 and / or one or more antennas 108 and 208. Control unit 120 is electrically connected to communication unit 110, memory unit 130, and add-on components 140, and controls the overall operation of the wireless device. For example, control unit 120 may control the electrical / mechanical operation of the wireless device based on programs / code / commands / information stored in memory unit 130. Control unit 120 may transmit information stored in memory unit 130 to an external source (e.g., other communication devices) via communication unit 110 through a wireless / wired interface, or store information received from an external source (e.g., other communication devices) via communication unit 110 in memory unit 130 via a wireless / wired interface.

[0360] The additional component 140 can be configured differently depending on the type of wireless device. For example, the additional component 140 may include at least one of a power supply unit / battery, an input / output (I / O) unit, a drive unit, and a computing unit. The wireless device can be configured according to (but is not limited to) a robot ( Figure 21 100a), vehicles ( Figure 21 100b-1 and 100b-2), XR device ( Figure 21 100c), handheld device ( Figure 21 100d), home appliances ( Figure 21 100e), IoT devices ( Figure 21 100f), digital broadcasting terminals, holographic devices, public safety devices, MTC devices, medical devices, fintech devices (or financial devices), security devices, climate / environment devices, AI servers / devices ( Figure 21 400), BS ( Figure 21 This can be achieved through 200 network nodes, etc. Wireless devices can be used in mobile or fixed locations depending on the use case / service.

[0361] exist Figure 23In some embodiments, various elements, components, units / portions, and / or modules in the wireless devices 100 and 200 can all be connected to each other through a wired interface, or at least part of them can be wirelessly connected through the communication units 110. For example, in each of the wireless devices 100 and 200, the control unit 120 and the communication unit 110 can be connected through a wired interface, and the control unit 120 and the first units (e.g., 130 and 140) can be wirelessly connected through the communication unit 110. The various elements, components, units / portions, and / or modules within the wireless devices 100 and 200 can also include one or more elements. For example, the control unit 120 can be configured by a set of one or more processors. As an example, the control unit 120 can be configured by a set of communication control processor, application processor, electronic control unit (ECU), graphic processing unit, and memory control processor. As another example, the memory unit 130 can be configured by random access memory (RAM), dynamic RAM (DRAM), read only memory (ROM), flash memory, volatile memory, non-volatile memory, and / or a combination thereof.

[0362] Figure 24 A vehicle or an autonomous driving vehicle applied to the present disclosure is illustrated. The vehicle or the autonomous driving vehicle can be implemented by a mobile robot, a car, a train, a manned / unmanned aerial vehicle (AV), a ship, etc.

[0363] Referring to Figure 24 , the vehicle or the autonomous driving vehicle 100 can 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 can be configured as a part of the communication unit 110. The blocks 110 / 130 / 140a to 140d correspond to the blocks 110 / 130 / 140 of Figure 23 , respectively.

[0364] The communication unit 110 can transmit and receive signals (e.g., data and control signals) to and from external devices such as other vehicles, BSs (e.g., gNBs and roadside units), and servers. The control unit 120 can perform various operations by controlling elements of the vehicle or the autonomous driving vehicle 100. The control unit 120 can include an electronic control unit (ECU). The driving unit 140a can cause the vehicle or the autonomous driving vehicle 100 to travel on a road. The driving unit 140a can include an engine, a motor, a powertrain, a wheel, a brake, a steering device, etc. The power supply unit 140b can supply power to the vehicle or the autonomous driving vehicle 100, and includes a wired / wireless charging circuit, a battery, etc. The sensor unit 140c can acquire vehicle states, surrounding environment information, user information, etc. The sensor unit 140c can include an inertial measurement unit (IMU) sensor, a collision sensor, a wheel sensor, a speed sensor, a slope sensor, a weight 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 depth sensor, an ultrasonic sensor, an illumination sensor, a pedal position sensor, etc. The autonomous driving unit 140d can implement a technology for maintaining a lane in which the vehicle is traveling, a technology for automatically adjusting a speed (e.g., adaptive cruise control), a technology for autonomously traveling along a determined path, a technology for traveling by automatically setting a path if a destination is set, etc.

[0365] 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 path and a driving plan from the obtained data. The control unit 120 can control the driving unit 140a so that the vehicle or the autonomous driving vehicle 100 can move along the autonomous driving path according to the driving plan (e.g., speed / direction control). In the middle of autonomous driving, the communication unit 110 can acquire recent traffic information data from an external server aperiodically / periodically, and surrounding traffic information data from a neighboring vehicle. In the middle of autonomous driving, the sensor unit 140c can obtain vehicle states and / or surrounding environment information. The autonomous driving unit 140d can update the autonomous driving path and the driving plan based on newly obtained data / information. The communication unit 110 can transmit information about a vehicle position, an autonomous driving path, and / or a driving plan to an external server. The external server can predict traffic information data using an AI technology, etc. based on information collected from vehicles or autonomous driving vehicles, and provide the predicted traffic information data to the vehicles or the autonomous driving vehicles.

[0366] The above-described embodiments of the present disclosure can correspond to combinations of elements and features in the specified form of the present disclosure. Also, unless explicitly mentioned, each of the elements or features can be considered selective. Each of the elements or features can be implemented in a form that is not combined with the other elements or features. Also, embodiments of the present disclosure can be implemented by partially combining elements and / or features. The order of operations explained for each of the embodiments of the present disclosure can be modified. Some configurations or features of one embodiment can be included in another embodiment, or can replace corresponding configurations or features of another embodiment. Also, it can be obviously understood that the embodiments are configured by combining the claims that have no explicit relationship with each other in the attached claims, or can be included as new claims by amendment after the filing of the application.

[0367] Those skilled in the art will appreciate that the disclosure can be practiced in other specific forms without departing from the spirit and essential characteristics of the disclosure. Thus, the above-described embodiments are to be considered in all respects as illustrative and not restrictive. The scope of the disclosure should be determined from the appended claims and their legal equivalents rather than from the foregoing description, which is to be construed merely as illustrative of the embodiments. All changes that come within the meaning and range of equivalency of the claims are intended to be embraced therein.

[0368] Industrial applicability

[0369] The present disclosure is applicable to a UE, a BS, or other devices in a wireless mobile communication system.

Claims

1. A method performed by a user equipment (UE) in a wireless communication system, the method comprising the following steps: Configurations for artificial intelligence / machine learning (AI / ML) models are obtained based on higher-level signaling. Monitor at least one AI / ML model in the AI / ML models; and Based on monitoring of at least one AI / ML model, AI / ML model management is performed for switching or updating. The step of monitoring the at least one AI / ML model includes at least one of the following operations: Monitoring the first AI / ML model in an inactive state; and Monitor the second AI / ML model that is active, and Specifically, the operation of the first AI / ML model in the inactive state occupies the UE processing unit for a shorter duration than the operation of the second AI / ML model in the active state, or occupies fewer UE processing units than the operation of the second AI / ML model in the active state.

2. The method according to claim 1, wherein, The first AI / ML model to be monitored in the inactive AI / ML model is determined based on network signaling.

3. The method according to claim 2, wherein, The start of the duration occupied by the UE processing unit due to monitoring the first AI / ML model is determined based on the time point at which the network signaling is received.

4. The method according to claim 2, wherein, The end of the duration occupied by the UE processing unit due to monitoring the first AI / ML model is determined based on at least one of the following: i) Report the time point at which the results of the first AI / ML model are monitored; ii) The time point at which the network response to the report was received; or iii) Schedule the timer to expire at the designated time.

5. The method according to claim 1, wherein, The first AI / ML model to be monitored in the inactive AI / ML model is determined by the UE.

6. The method according to claim 5, wherein, Based on the operation triggered by network signaling for the second AI / ML model in the active state, monitoring of the AI / ML model in the inactive state is stopped or paused.

7. The method according to claim 1, further comprising the following step: Report information to the network regarding the occupied or unoccupied status of the UE processing unit.

8. The method according to claim 1, wherein, Monitoring the first AI / ML model is performed based on the remaining UE processing units that are not occupied by the second AI / ML model, which is in the active state.

9. The method according to claim 1, wherein, The second AI / ML model in the active state occupies the UE processing unit with a higher priority than the first AI / ML model in the inactive state.

10. The method according to claim 1, wherein, Monitoring the operation of the second AI / ML model in the active state occupies one UE processing unit, and The operation of monitoring the first AI / ML model in the inactive state occupies n UE processing units, where n is a real number greater than 0 and less than or equal to 1.

11. The method according to claim 10, wherein, n is determined based on UE capabilities or provided through network signaling.

12. A computer-readable recording medium having a program recorded thereon for performing the method according to claim 1.

13. A user equipment (UE) in a wireless communication system, the UE comprising: transceiver; as well as A processor configured to control the transceiver. The processor is configured as follows: Configurations for artificial intelligence / machine learning (AI / ML) models are obtained based on higher-level signaling. Monitor at least one AI / ML model among the AI / ML models; and Based on monitoring of at least one AI / ML model, AI / ML model management is performed for switching or updating. The operation of monitoring the at least one AI / ML model includes at least one of the following operations: Monitoring the first AI / ML model in an inactive state; and Monitor the second AI / ML model that is active, and Specifically, the operation of the first AI / ML model in the inactive state occupies the UE processing unit for a shorter duration than the operation of the second AI / ML model in the active state, or occupies fewer UE processing units than the operation of the second AI / ML model in the active state.

14. An apparatus configured to control a user equipment (UE) in a wireless communication system, the apparatus comprising: Memory configured to store instructions; as well as A processor configured to perform operations for controlling the UE by executing the instructions. The operations used to control the UE include the following: Configurations for artificial intelligence / machine learning (AI / ML) models are obtained based on higher-level signaling. Monitor at least one AI / ML model in the AI / ML models; and Based on monitoring of at least one AI / ML model, AI / ML model management is performed for switching or updating. The operation of monitoring the at least one AI / ML model includes at least one of the following operations: Monitoring the first AI / ML model in an inactive state; and Monitor the second AI / ML model that is active, and Specifically, monitoring the operation of the first AI / ML model in the inactive state occupies the device processing unit for a shorter duration than monitoring the operation of the second AI / ML model in the active state, or occupies fewer device processing units than monitoring the operation of the second AI / ML model in the active state.

Images