Method and device for controlling performance monitoring operation of ai / ML model in wireless communication system

Abstract

The present disclosure relates to a method and a device for controlling a performance monitoring operation for an AI / ML model in a wireless communication system, and may provide a method and a device comprising the steps of: receiving, from a base station, configuration information for performance monitoring of the AI / ML model, the configuration information including mapping information between a prediction reference signal set and a monitoring reference signal set; identifying the monitoring reference signal set indicated in association with the prediction reference signal set outputted using the AI / ML model on the basis of the configuration information; and transmitting performance monitoring result information for the AI / ML model using inference result information for the prediction reference signal set and quality measurement result information for the monitoring reference signal set.

Description

Method and apparatus for controlling performance monitoring operation of AI / ML models in a wireless communication system

[0001] The present disclosure relates to a method and apparatus for controlling performance monitoring operations for an AI / ML model in a wireless communication system.

[0002] The International Telecommunication Union (ITU) is conducting standardization of next-generation 5th generation (5G) and 6th generation (6G) mobile communications through the International Mobile Telecommunication (IMT)-2020 and IMT-2030 programs.

[0003] The next-generation 6th generation (6G) mobile communication system is defined as an evolutionary system that further enhances the performance of 5th generation (5G) technology, and at the same time, as a next-generation communication platform that encompasses new services and industrial applications. Based on the ultra-high transmission rates, ultra-low latency, and massive terminal connectivity achieved by 5G, 6G aims for ultra-broadband, ultra-precision, ultra-intelligence, and ultra-convergence.

[0004] To achieve this, various technologies are being discussed in 6G, including the utilization of FR3 and terahertz (THz) bands, large-scale antennas and new numerology designs, AI / ML native network optimization, global coverage based on non-terrestrial networks (NTN), dynamic channel control by intelligent reflective surfaces (RIS), and integrated sensing and communication (ISAC). In addition, technologies to enhance network energy efficiency are also being treated as an important research axis in 6G. Therefore, to achieve the goals of ultra-high speed, ultra-low latency, intelligence, global coverage, and convergence services that the next-generation 6th generation (6G) mobile communication system aims for, technical challenges related to AI-Native RAN supporting AI / ML-based network optimization, non-terrestrial networks (NTN) ensuring global connectivity, ISAC converging communication, location, and sensing, intelligent reflective surfaces (RIS) enabling dynamic channel control, and securing energy efficiency for sustainable network operation need to be resolved.

[0005] When using AI / ML models to predict channel estimation results or estimate optimal beams, the accuracy of the AI / ML models configured in the terminal and base station is critical. However, the performance of AI / ML models can change depending on various factors, such as changes in the wireless environment. To address this, the terminal and base station must perform performance monitoring operations on the AI / ML models.

[0006] Therefore, there is an urgent need for the proposal and research of various technologies capable of meeting the technical requirements for this purpose.

[0007] The present embodiments aim to provide a method and apparatus for performing performance monitoring of an AI / ML model in a wireless communication system.

[0008] The present embodiments aim to provide a method and apparatus for configuring a monitoring reference signal set in a wireless communication system.

[0009] In addition, the present embodiments aim to provide a method and apparatus for efficiently monitoring the performance of an AI / ML model while reducing the measurement and reporting overhead of a terminal in a wireless communication system.

[0010] In addition, the present embodiments aim to provide a method and apparatus for configuring a monitoring reference signal set into a bitmap in a wireless communication system.

[0011] In addition, the present embodiments aim to provide a method and apparatus that support switching between a general monitoring mode and an advanced monitoring mode in a wireless communication system.

[0012] In addition, the present embodiments aim to provide a method and apparatus for variably setting a beam subset size in a wireless communication system.

[0013] In addition, the present embodiments aim to provide a method and apparatus that support switching between a general monitoring mode and an advanced monitoring mode in a wireless communication system.

[0014] The present embodiments aim to provide a method and apparatus for configuring a reference signal at a terminal to monitor the performance of an AI / ML model in a wireless communication system, and for the terminal to perform a monitoring operation using the same.

[0015] In order to solve the aforementioned problem, in one aspect, the present disclosure may provide a method for a terminal to perform a performance monitoring operation of an AI / ML model, comprising the steps of: receiving configuration information for performance monitoring of an AI / ML model including mapping information of a prediction reference signal set and a monitoring reference signal set from a base station; identifying a monitoring reference signal set indicated in conjunction with a prediction reference signal set output using the AI / ML model based on the configuration information; and transmitting performance monitoring result information for the AI / ML model using inference result information for the prediction reference signal set and quality measurement result information for the monitoring reference signal set.

[0016] In another aspect, the present disclosure may provide a method for a base station to control a performance monitoring operation of an AI / ML model of a terminal, comprising the steps of: transmitting configuration information for performance monitoring of an AI / ML model, including mapping information of a prediction reference signal set and a monitoring reference signal set, to the terminal; transmitting a monitoring reference signal indicated according to the configuration information to the terminal; and receiving performance monitoring result information for an AI / ML model generated using inference result information for a prediction reference signal set and quality measurement result information for a monitoring reference signal set from the terminal.

[0017] In another aspect, the present disclosure may provide a terminal device for performing performance monitoring of an AI / ML model, comprising: a receiving unit that receives configuration information for performance monitoring of an AI / ML model including mapping information of a prediction reference signal set and a monitoring reference signal set from a base station; a control unit that identifies a monitoring reference signal set indicated in conjunction with a prediction reference signal set output using an AI / ML model based on the configuration information; and a transmitting unit that transmits performance monitoring result information for an AI / ML model using inference result information for a prediction reference signal set and quality measurement result information for a monitoring reference signal set.

[0018] According to the present disclosure, the advantages of improving prediction accuracy and communication stability are provided by efficiently applying an AI / ML model as part of a communication procedure in a wireless communication system and continuously monitoring the performance of said AI / ML model at the network level.

[0019] Furthermore, according to the present disclosure, an AI / ML native communication structure is supported in which network control and resource management are performed based on the operation results of an AI / ML model in a wireless communication system, thereby enabling the AI / ML model to be utilized as an essential network component rather than a mere auxiliary function. Based on this structure, the present disclosure effectively detects performance degradation of the AI / ML model that may occur due to changes in the wireless environment, traffic characteristics, and user mobility, and enables adaptive correction and optimization thereof, thereby stably securing the high reliability and prediction accuracy required for next-generation mobile communication systems. This provides a technical effect of improving the reliability and operability of AI / ML-based communication functions.

[0020] This disclosure satisfies the performance requirements of next-generation 6th generation (6G) mobile communication systems aiming for ultra-high speed, ultra-low latency, and large-scale connectivity, while enabling the evolution into an intelligent and autonomous network in which the network perceives its environment and optimizes its operation. This reduces unnecessary measurement, reporting, and signal exchange, and allows for more efficient management of wireless and network resources, thereby providing the effect of improving network operational efficiency and energy efficiency.

[0021] FIG. 1 is a drawing illustrating the structure of a wireless communication system to which the present embodiment can be applied.

[0022] FIG. 2 is a diagram illustrating the logical layer structure between a terminal (UE) and a base station (gNB) in a wireless communication system to which the present embodiment can be applied.

[0023] FIG. 3 is a diagram illustrating an exemplary NG-RAN structure to which the present embodiment can be applied.

[0024] FIG. 4 is a diagram illustrating the configuration of a terminal to which the present embodiment can be applied.

[0025] FIG. 5 is a diagram illustrating an AI / ML operation workflow according to one embodiment of the present disclosure.

[0026] FIG. 6 is a diagram illustrating the data collection and reporting procedure of an AI / ML model to which embodiments of the present disclosure can be applied.

[0027] FIG. 7 is a diagram illustrating the classification according to the form in which an AI / ML model to which embodiments of the present disclosure can be applied is configured.

[0028] FIG. 8 is a diagram illustrating a terminal-network collaboration level using an AI / ML model to which embodiments of the present disclosure can be applied.

[0029] FIG. 9 is a diagram illustrating terminal operation according to one embodiment.

[0030] FIG. 10 is a diagram illustrating the operation of a base station according to one embodiment.

[0031] FIG. 11 is a diagram illustrating the performance monitoring operation of a terminal according to one embodiment.

[0032] FIG. 12 is a diagram illustrating a performance monitoring operation according to a mode of a terminal according to another embodiment.

[0033] FIG. 13 is a diagram illustrating an operation for indicating a set of monitoring reference signals according to configuration information according to one embodiment.

[0034] FIG. 14 is a diagram illustrating an example of configuring a monitoring reference signal set according to one embodiment.

[0035] FIG. 15 is a diagram illustrating examples of a prediction reference signal set and a monitoring reference signal set according to one embodiment.

[0036] FIG. 16 is a drawing for explaining a terminal configuration according to one embodiment.

[0037] FIG. 17 is a drawing for explaining the configuration of a base station according to one embodiment.

[0038] Hereinafter, some embodiments of the present disclosure will be described in detail with reference to the exemplary drawings. In assigning reference numerals to the components of each drawing, the same components may have the same reference numeral as much as possible, even if they are shown in different drawings. Furthermore, in describing the embodiments, if it is determined that a detailed description of related known components or functions may obscure the essence of the technical concept, such detailed description may be omitted. Where terms such as "comprising," "having," or "consisting of" are used in this specification, other parts may be added unless "only" is used. Where a component is expressed in the singular, it may include a plural unless otherwise specified.

[0039] Additionally, terms such as first, second, A, B, (a), (b), etc., may be used to describe the components of the present disclosure. These terms are used merely to distinguish the components from other components, and the nature, order, sequence, or number of the components are not limited by such terms.

[0040] In describing the positional relationship of components, where it is stated that two or more components are "connected," "combined," or "joined," it should be understood that while the two or more components may be directly "connected," "combined," or "joined," they may also be "connected," "combined," or "joined" with other components "intervened." Here, the other components may be included in one or more of the two or more components that are "connected," "combined," or "joined" with one another.

[0041] In describing the temporal flow relationship regarding components, methods of operation, or methods of production, for example, when the temporal or sequential relationship is described using "after," "following," "next," or "before," it may include cases where the relationship is not continuous unless "immediately" or "directly" is used.

[0042] Meanwhile, where numerical values ​​or corresponding information regarding a component (e.g., levels, etc.) are mentioned, even without separate explicit notation, the numerical values ​​or corresponding information may be interpreted as including a range of error that may occur due to various factors (e.g., process factors, internal or external shocks, noise, etc.).

[0043] A wireless communication system in this specification refers to a system for providing various communication services, such as voice and data packets, using wireless resources, and may include a terminal, a base station, or a core network.

[0044] Meanwhile, the term "terminal" in this specification is a comprehensive concept meaning a device including a wireless communication module that communicates with a base station in a wireless communication system, and should be interpreted as a concept that includes not only User Equipment (UE) in WCDMA, LTE, NR, HSPA and IMT-2020 / IMT-2030 (5G, New Radio, and 6G), and wired / wireless communication sensing systems that interact with specific communication systems, but also Mobile Station (MS), User Terminal (UT), Subscriber Station (SS), wireless device, and sensing module / device in GSM. In this specification, the term "terminal" is a comprehensive concept meaning a device including a wireless communication module that communicates with a base station in a wireless communication system. It should be interpreted as a concept that includes not only User Equipment (UE) in WCDMA, LTE, NR, HSPA, and IMT-2020 / IMT-2030 (5G, New Radio, and 6G), and wired / wireless communication sensing systems that interact with specific communication systems, but also Mobile Station (MS), User Terminal (UT), Subscriber Station (SS), wireless device, sensing module / device, etc. in GSM. Furthermore, depending on the usage type, the terminal may be a user portable device such as a smartphone, and in a V2X communication system, it may refer to a device including the vehicle itself or a wireless communication module within the vehicle. Furthermore, in the case of a Machine Type Communication (MTC) system, it can be interpreted as a concept that includes not only MTC terminals, M2M terminals, URLLC terminals, and V2X terminals equipped with communication modules to perform machine type communication, but also sensor network terminals, IoT devices, wearable devices, etc.As such, a “terminal” may refer to another node that transmits and receives data / signals with a specific communication node.

[0045] In this specification, the term "base station" is a comprehensive concept referring to a network device that performs wireless communication with a terminal in a wireless communication system. It should be interpreted to include eNBs (evolved Node Bs) in 3GPP LTE systems, gNBs (gNode Bs) in NR (New Radio) systems, and next-generation base stations in IMT-2020 / IMT-2030 (5G and 6G) ​​systems, as well as Node Bs in WCDMA systems, BTSs (Base Transceiver Stations) in GSM systems, APs (Access Points) in Wi-Fi systems, and RSUs (Road Side Units) in V2X communication systems. Furthermore, base stations can be implemented in various forms, including not only a centralized architecture but also distributed Small Cell units, Repeaters, Relay Nodes, Satellite Gateways, and Intelligent Reflective Surface (RIS) controllers, and all such variations should be interpreted as being included in the term "base station" in this specification. As such, "base station" may refer to another communication node that transmits and receives data / signals with another node in a specific communication system. In this specification, the term "Cell" should be interpreted as a comprehensive concept referring to a logical / physical area in a wireless communication system where a terminal receives wireless services from a base station. In LTE systems, the coverage provided by the base station (eNB) may include not only physical cells, but also multiple cell concepts (e.g., SCell, PCell, BWP-based virtual cell, etc.) based on frequency resources, beams, slot structures, etc., in 5G NR systems. Furthermore, as beam-based communication structures are emphasized along with FR2 / FR3 and terahertz (THz) band expansion, beam-based coverage may be included as the basic unit of a cell.Furthermore, in a Non-Terrestrial Network (NTN) environment, as the beam footprint of an orbiting satellite changes over time, there is a characteristic where the boundaries of cells are not fixed but change dynamically, and this can be defined and included as a cell. Additionally, it may include cell expansion based on Intelligent Reflective Surfaces (RIS), dynamic cell structures considering Integrated Sensing and Communication (ISAC), and cells based on virtualized network slices. This may include a structure in which a terminal connects to multiple cells simultaneously or selectively according to physical / logical resource units. Therefore, the present specification should be interpreted as a concept that includes not only base station-based coverage but also beam-unit logical cells, NTN-based dynamic cells, and new types of cells formed based on RIS or virtualization.

[0046] The present disclosure describes an NR system as an example, but it will be obvious that the present disclosure can be applied to and used in the relevant functions of a 6G mobile communication network, and that the limitations of the present disclosure are not restricted by the said NR system. This specifies that while NR (New Radio) is used as a specific embodiment, the technical concept of the present disclosure can be modified or expanded and applied to future systems such as 6G and the newly defined Beyond_G, and is characterized by not being limited to a specific system (NR). Accordingly, the technical scope of the invention is not limited to an NR system and is characterized by being implementable in various wireless systems including 6G.

[0047] The 6G communication to which the present disclosure applies is a next-generation wireless communication technology following 5G, and standardization is currently underway with the goal of commercialization around 2030. Beyond simple speed improvements, 6G aims for artificial intelligence, hyper-spatial scalability, and reality-digital convergence, and is being prepared as a core infrastructure for the future digital society. This 6G communication aims to support the next-generation digital society by organically combining core technologies such as THz communication capable of ultra-broadband transmission, AI-Native networks integrating AI / ML across all network layers, RIS (Reconfigurable Intelligent Surface) technology that actively reflects / manipulates radio waves, NTN (Non-Terrestrial Network) providing 3D coverage such as satellites, Joint Communication & Sensing (JCAS) integrating communication and sensing functions, ultra-precise location estimation, and Quantum Key Distribution (QKD) for security.

[0048] The structure of these 6G communication networks is being discussed based on the following directions and characteristics. First, as an End-to-End Intelligent Autonomous Network (AI-Native Architecture) structure, AI / ML is fundamentally embedded across all network layers to perform data-driven policy optimization, autonomous network operation, anomaly detection, and Quality of Service (QoS / QoE) assurance; it is expected that AI functions will operate in a distributed manner across the RAN, Core, and service management. Second, by adopting a 3D scalability architecture, it supports a Non-Terrestrial Network (NTN) that integrates the terrestrial network (RAN), satellite, and High Altitude Platform (HAPS); to this end, an integrated cell structure and IAB-based backhaul / fronthaul design are planned to be supported. Third, the Service-Based Architecture (SBA) introduced in 5G will be advanced to strengthen the modularization of Network Functions (NF), and flexible API-based service creation will be enabled through NEF, NWDAF, PCF, etc. Fourth, through a distributed intelligence structure, we plan to move away from a centralized architecture and perform real-time AI-based decisions at edge nodes such as RAN, MEC, and UE, thereby supporting local decision-making in smart transportation, factories, and medical sites. Fifth, by introducing a structure that integrates communication and sensing (Joint Communication & Sensing), we will enable the RAN to perform environmental sensing functions beyond simple communication capabilities; to this end, structures for processing sensing features and asynchronous data collection will also be designed. Finally, through a digital twin-based virtualization structure, we will replicate the physical network state in real time and enable predictive autonomous operation and network simulation, thereby supporting the intelligence and efficiency of the entire network.The next-generation wireless system to which this disclosure applies includes support for both frequency bands below 6 GHz (FR1, Frequency Range 1) and frequency bands above 6 GHz (FR2, Frequency Range 2), and specifically defines support for FR3 (Frequency Range 3) between FR1 and FR2, and proposes various support technologies related to end-to-end interoperability and data connectivity for the FR3 band (7.125-24.25 GHz).

[0049] For example, NR has defined various operation scenarios by adding considerations for satellites, automobiles, and new verticals, and in terms of service, it supports eMBB (Enhanced Mobile Broadband) scenarios, mMTC (Massive Machine Communication) scenarios that require low data rates and asynchronous connections while having high terminal density and being deployed over a wide range, and URLLC (Ultra Reliability and Low Latency) scenarios that require high responsiveness and reliability and can support high-speed mobility.

[0050] To satisfy these scenarios, NR introduces a wireless communication system equipped with new waveform and frame structure technologies, low latency technologies, mmWave support technologies, and forward compatibility technologies. In particular, NR systems present various technical changes in terms of flexibility to provide forward compatibility.

[0051] Although the following description focuses on NR, it can also be applied to next-generation wireless communication systems (e.g., 6G, etc.) as described above. Additionally, while the aforementioned terms may be used as different terms in next-generation wireless communication systems, the device providing the same function may be the device described in this specification.

[0052] FIG. 1 is a drawing illustrating the structure of a wireless communication system to which the present embodiment can be applied.

[0053] Referring to FIG. 1, a wireless communication system may be composed of a core network unit (120), a base station (110), and terminals (100, 101). In addition, the wireless communication system may include additional nodes such as a Radio Unit (RU). From the perspective of terminal expansion, the wireless communication system may further include vehicles (102, 103), aerial mobile bodies (104), etc. From the perspective of base station expansion (110), the wireless communication system may further include satellites (111), etc.

[0054] The base station (110) can provide user plane and control plane protocol termination to the terminal (100). For example, the base station (110) may be named gNB (next generation-Node B) and / or eNB (evolved-Node B), and may be referred to by other terms in next-generation wireless communication systems. For example, the terminal (100) may be fixed or mobile and may be referred to by other terms such as MS (Mobile Station), UT (User Terminal), SS (Subscriber Station), MT (Mobile Terminal), or Wireless Device. For example, the base station (110) may be a fixed station communicating with the terminal (100) and may be referred to by other terms such as BTS (Base Transceiver System) or Access Point.

[0055] Base stations (110) can be connected to each other via protocol interfaces. For example, base stations (110) can be connected to each other via Xn interfaces. Additionally, base stations (110) can be connected to core network entities (120) via protocol interfaces. For example, base stations (110) can be connected to core network entities (120) via NG interfaces. More specifically, base stations (110) can be connected to an access and mobility management function (AMF) (120) via an NG-C interface and to a user plane function (UPF) (120) via an NG-U interface. The AMF is responsible for control planes such as terminal access and mobility control functions, and the UPF is responsible for control functions for user data.

[0056] The satellite (111) can provide communication services to a terminal through a ground antenna or gateway as a core network entity. The satellite (111) may provide communication services while orbiting the Earth or may provide communication services while fixed in orbit on the Earth. The satellite (111) may be configured to transparently transmit messages provided by a ground gateway or base station (110). Alternatively, the satellite (111) may be equipped with some or all of the functions of a base station (110) to perform the function of a base station for the core network entity (120) and the terminal.

[0057] Meanwhile, the terminal (100) can perform direct communication with other terminals (101) without passing through the base station (110) through terminal-to-terminal communication technology. For example, direct communication between terminals can be described as D2D or V2X technology, but is not limited to such terms. For example, vehicles (102, 103) can use terminal-to-terminal communication technology as vehicle-to-vehicle communication technology, and communication can also be performed between the vehicle (103) and the terminal (100). Of course, communication can also be performed between the vehicle (102) and the base station (110). Similarly, a communication link can be established between the aerial mobile body (104) and the base station (110) to perform communication, and communication can also be performed between the aerial mobile body (104) and the vehicle (102, 103) or the terminal (100, 101).

[0058] In NR, CP-OFDM waveforms using a cyclic prefix are used for downlink transmission, while CP-OFDM or DFT-s-OFDM are used for uplink transmission. OFDM technology is easy to combine with MIMO (Multiple Input Multiple Output) and has the advantage of enabling the use of low-complexity receivers along with high frequency efficiency. For next-generation wireless communication systems, various waveforms and modulation technologies may be applied.

[0059] FIG. 2 is a diagram illustrating the logical layer structure between a terminal (UE) and a base station (gNB) in a wireless communication system to which the present embodiment can be applied.

[0060] Referring to FIG. 2, data transmission between a terminal and a base station in a wireless communication system is performed through a multilayer protocol structure. Generally, NR and NG-RAN structures are configured to include at least one of a Service Data Adaptation Protocol (SDAP), a Packet Data Collection Protocol (PDCP), Radio Link Control (RLC), Media Access Control (MAC), and a Physical Layer (PHY).

[0061] In the User Plane, the SDAP layer maps QoS flows defined in the 5G core (5GC) to the Data Radio Bearer (DRB), and the PDCP layer performs encryption, integrity protection, header compression, packet reordering, and deduplication functions. The RLC layer provides splitting and combining, retransmission, and reordering functions, while the MAC layer handles radio resource scheduling, multiplexing, and HARQ. The PHY layer performs modulation, encoding, channel estimation, and beamforming.

[0062] In the control plane, the RRC layer is responsible for wireless bearer setup, measurement and handover procedures, and security parameter exchange, and works in conjunction with the upper layer, NAS, to perform terminal authentication and session management.

[0063] The functions and operations of each layer described above are exemplary, and the functions of a specific layer may be separated and performed by other layers. For instance, the splitting and combining functions of the RLC layer may be performed by PDCP, and the retransmission and reordering functions may be absorbed into similar functions in the MAC or PHY layers, thereby eliminating the RLC layer.

[0064] The embodiments in this specification may be applied to the aforementioned logical layer structure and may also be applied to various logical layer structures that are modified in the future.

[0065] FIG. 3 is a diagram illustrating an exemplary NG-RAN structure to which the present embodiment can be applied.

[0066] Referring to FIG. 3, in the NG-RAN structure, base station (110) functions can be separated into a central unit (CU) and a distributed unit (DU). The CU is responsible for the control plane (RRC / PDCP-C) and the user plane (SDAP / PDCP-U), while the DU is responsible for the RLC / MAC / PHY layer. The interface between the CU and the DU is defined as F1-C and F1-U, and within the CU, an E1 interface is defined between the control plane and the user plane. This allows for support of both distributed and centralized architectures.

[0067] Additionally, the connection between base stations is performed by the CU through the Xn interface, and the base station (110) is connected to the core network entity (120) through the NG interface. Meanwhile, in a 6th generation (6G) wireless communication system, the above logic layer structure can be extended for AI / ML-based network optimization, FR3 and THz band expansion, non-terrestrial network (NTN) support, RIS (Reconfigurable Intelligent Surfaces)-based dynamic channel control, and ISAC (Integrated Sensing and Communication)-based integrated service provision. For example, the RRC and MAC layers may include new control signals for low-power mode and energy efficiency, and the SDAP / PDCP layer may simultaneously perform QoS processing and sensing quality assurance. Additionally, in an NTN environment, RLC / MAC procedures may be added to assist time / frequency synchronization, and new RRC information elements for RIS control may be defined.

[0068] As described above, the communication system to which this disclosure applies inherits the 5G-Advanced infrastructure for interoperability with 6G or next-generation communication systems, but may evolve in a form where each layer of the newly defined next-generation communication system interacts to expand or converges. Alternatively, the communication system to which this disclosure applies may include a separate Radio Access Network (RAN) and Core Network structure distinct from the 4G and 5G wireless communication structures. This disclosure is not limited to the network architecture of the existing generation and may be applied based on a next-generation network structure in which AI / ML native-based functions are essentially integrated into communication procedures and network control. Accordingly, each embodiment described in this disclosure is not limited to a specific generation of mobile communication systems and is characterized by being adaptively applicable not only to currently commercialized 4G and 5G mobile communication systems but also to 6G mobile communication systems that are currently under discussion and standardization or are scheduled for the future.

[0069] For example, the present disclosure may apply to AI-Native RAN interconnection operations. The AI-Native RAN may be configured to extend beyond the control layer within the Radio Access Network (RAN) to the Radio Resource Control (RRC), the orchestration of the Central Unit (CU) and the Distributed Unit (DU), and further to the Policy Control Function (PCF) of the core network. In this case, AI / ML models and policies may be arranged hierarchically according to the control target and time sensitivity; for example, real-time control centered on the MAC / PHY layer may be performed in the DU, and near-real-time control related to RRM, handover, and measurement policies may be performed in the CU-CP. Additionally, policy control related to slice control and QoS governance on a minute to hourly basis may be performed in the core network and PCF.

[0070] Higher-level policies can be propagated throughout the network from the PCF via the AMF and SMF paths, and these policies can be converted into model-derived parameters that are reflected in procedures such as measurement, handover, DRX, and cell reselection through the extension of RRC Information Elements (IEs), and then transmitted to terminals and base stations. Through this, terminals and base stations can adaptively adjust their behavior based on network policies and the inference results of AI / ML models.

[0071] In addition, in the present disclosure, network metrics generated at various layers can be systematically collected and utilized for AI / ML learning and inference. For example, retransmission and packet drop delay information can be collected at the PDCP layer, and buffer occupancy status and retransmission-related metrics can be collected at the RLC layer. At the MAC / PHY layer, beam quality information based on CQI, CSI, BLER, HARQ, PHR, SRS, and CSI-RS can be utilized, and at the RRC, NGAP, and XnAP layers, metrics related to measurement events, handover failures, delays, and jitter can be collected.

[0072]

[0073]

[0074] As another example, the frequency band used in the communication system to which the present disclosure applies may be extended beyond the existing FR1 and FR2 to the FR3 band. The FR3 band may consist of an ultra-high frequency band approaching or including terahertz (THz), in which case the complexity of beam configuration and wireless resource management may increase significantly. In particular, in a THz environment, link stability may fluctuate very rapidly due to high path loss, increased sensitivity to shielding, and amplification of the Doppler effect.

[0075] In such an environment, beam tracking operations performed at the MAC / PHY layer and policy control performed at the RRC and SDAP layers need to be closely coupled. For example, the AI ​​according to the present disclosure can predictively preempt future valid beams by utilizing various features related to the terminal's operating state, mobility, and surrounding environment, thereby effectively mitigating control overhead by reducing the size of the beam sweeping period or the set of synchronized signal blocks (SSB).

[0076] In addition, radio resource decisions, such as the selection of slot or mini-slot structures, density settings for DMRS and CSI-RS, and whether to use auxiliary carriers, can be performed in conjunction with the QoS mapping policy of the SDAP layer. This allows for controlling latency-sensitive traffic and bandwidth-intensive traffic to be subject to different beam and carrier strategies, and consequently, enables the provision of stable communication performance that meets service requirements even in ultra-high frequency FR3 environments.

[0077] Furthermore, in this disclosure, slot or mini-slot-based transmission structures may be applied differentially depending on the service type. In particular, since Ultra-Reliable and Ultra-Low Latency Communication (URLLC), Extended Reality (XR) services, and traffic involving sensing and communication have different time sensitivities and transmission characteristics, a policy distinction regarding wireless resource allocation units is required. First, in the case of URLLC traffic, since short latency requirements and high reliability are simultaneously required, a mini-slot-based transmission structure may be applied preferentially in this disclosure. Mini-slots enable immediate transmission not restricted by slot boundaries, thereby supporting fast scheduling and HARQ operations even in situations requiring rapid changes in link quality or retransmission. Accordingly, the AI ​​can dynamically select mini-slot-based transmission for URLLC traffic by considering beam quality changes, Doppler characteristics, packet arrival intervals, etc., and reinforce link reliability by increasing DMRS and CSI-RS densities as needed. Meanwhile, XR traffic requires high bandwidth and continuous data transmission, but it has characteristics that allow for a certain level of latency tolerance. Accordingly, the present disclosure applies a slot-based transmission structure as the basis for XR traffic, but may partially parallel mini-slot transmission in sections where frame loss or degradation of user perceived quality is predicted. For example, AI can simultaneously secure spectrum efficiency and user perceived quality by combining slot-based high-capacity transmission and mini-slot-based complementary transmission based on user viewpoint movement, display frame requirements, and beam stability prediction results. Furthermore, in the case of traffic involving the linkage of sensing and communication, resource management is required considering the periodicity and accuracy requirements of the sensing signal, as well as the possibility of interference with communication traffic.In this disclosure, regular resource allocation in slot units is applied to sensing-related traffic to ensure sensing accuracy, while rapid beam switching or retransmission can be performed through mini-slot-based transmission when abrupt changes in the communication link are predicted based on the sensing results. This allows for improved adaptability to dynamic environmental changes while minimizing mutual interference between the sensing signal and the communication signal.

[0078] In this way, the present disclosure does not apply a fixed slot or mini-slot structure according to the characteristics of URLLC, XR, and sensing-linked traffic, but rather adaptively selects it based on AI / ML-based prediction and policy control, thereby efficiently supporting various service requirements within a single wireless interface structure.

[0079] As another example, since NTN and terrestrial network duality entails orbit / visibility / Doppler / path delay variations, measurement / handover procedures on the RRC side and timing / synchronization on the DU side must be complemented. For instance, it may be required to predict and correct timing advances and frequency offsets based on satellite orbit / ephemeris / terminal movement models, and to utilize visibility predictions between the terrestrial network and NTN or between NTNs to perform preparatory handovers.

[0080] As another example, from an energy efficiency perspective, green RAN improves energy efficiency by maintaining traffic and SLAs while implementing DU sleep / wake, MIMO chain off, and DRX optimization. For instance, hierarchical coverage strategies can be employed, such as turning off small cells / auxiliary carriers and maintaining wide-area cell anchors during nighttime or low-load periods. To achieve this, RRC low-power profiles by service class, power state transition commands / telemetry (KPM) from CU to DU, and OAM collection of energy KPIs (cell power, RF chain uptime, cooling alarms) can be defined. To address increased handover failures or delays caused by excessive power saving, relevant policies and procedures regarding minimum resources and maximum wake-up delays per SLA can also be established.

[0081] As another example, the present disclosure may apply an Integrated Sensing and Communication (ISAC) coupled operation. In this case, physical layer (PHY) signals, such as a Position Reference Signal (PRS) or a Reference Signal (RS), may be utilized for sensing purposes as well as for communication purposes, and sensing performance may be treated as a key element of network quality management. For example, sensing-related performance indicators, such as sensing accuracy, sensing latency, and sensing information update cycle, may be considered as part of the Quality of Service (QoS) of communication.

[0082] Specifically, in the present disclosure, auxiliary indicators such as sensing-accuracy, sensing-latency, and sensing-update-rate may be included in the SDAP layer and 5QI scheme, and these sensing-related indicators may be reflected together in the objective function for policy optimization. Through this, the network can perform resource management and policy decisions to satisfy both sensing performance and communication performance simultaneously, rather than allocating resources based solely on communication performance.

[0083] Furthermore, at the RRC layer, the configuration of sensing-related PHY signals can be adaptively controlled according to service characteristics. For example, in cases requiring high sensing accuracy and reliability, such as industrial position estimation or sensing services, the RRC can adjust the pattern, density, and transmit power of the PRS or CSI-RS to meet the requirements of the service. This allows for improved sensing quality while minimizing unnecessary signal overhead.

[0084] Through such an ISAC linkage structure, the present disclosure goes beyond conventional structures where sensing and communication operate as separate functions, enabling more precise and intelligent network control by integrating sensing performance into network policy and QoS management.

[0085] FIG. 4 is a diagram illustrating the configuration of a terminal to which the present embodiment can be applied.

[0086] Referring to FIG. 4, the terminal (100) may be composed of various elements, components, units, and / or modules. For example, the terminal (100) may include a communication unit (300), a processor (310), and a memory (320). The communication unit (300) may include a communication circuit and transceiver(s). For example, the communication circuit may include circuit configurations such as a signal oscillator for transmitting or receiving a signal through an antenna. For example, the transceiver(s) may be composed of one or more and may include one or more antennas. The processor (310) is electrically connected to the communication unit (300) and the memory (320) and controls the overall operation of the terminal (100). For example, the processor (310) may control the electrical / mechanical operation of the terminal based on a program / code / command / information stored in the memory (320). Additionally, the processor (310) can transmit information stored in memory (320) to an external (e.g., another communication device or base station) via a wireless / wired interface through the communication unit (300), or store information received from an external (e.g., another communication device or base station) via a wireless / wired interface through the communication unit (300) in memory (320).

[0087] Additionally, the processor (310) can perform AI / ML-related data processing by loading program code into memory (320). Alternatively, the processor (310) may perform data analysis and processing operations using signals and / or data received through the communication unit (300).

[0088] As described above, the present disclosure enables the support of a standard performance monitoring structure required for an AI / ML-based wireless communication system at the terminal level. Accordingly, the terminal (100) can determine the performance status of an AI / ML model based on various measurement indicators and operation results of the PHY / MAC / RRC layer collected through the communication unit (300), and can store the determination result in memory (320) or report it to the network. Such performance monitoring results can be utilized for adjusting model parameters, changing inference policies, triggering retraining, or controlling the activation or deactivation of AI / ML functions. In particular, as the AI / ML model is involved in network core control loops such as wireless resource control, beam management, handover, and measurement settings, the inference accuracy of the model, stability degradation, and performance deterioration due to environmental changes can be monitored in real-time or near-real-time, and the relevant information can be transmitted and received.

[0089] In addition to this, the terminal (100) may include various additional elements. For example, the additional elements may be configured in various ways depending on the type of terminal. For example, the additional elements may include at least one of a power unit / battery, an input / output unit (I / O unit), a driving unit, and a computing unit. Although not limited thereto, the terminal (100) may be implemented in the form of a drone, robot, vehicle, XR device, portable device, home appliance, IoT device, digital broadcasting terminal, hologram device, public safety device, MTC device, medical device, fintech device (or financial device), security device, climate / environment device, AI server / device, sensor device, network node, etc. Depending on the use—e.g., service—the terminal (100) may be movable or used in a fixed location.

[0090] Meanwhile, various elements, components, units / parts, and / or modules within the terminal (100) may be entirely interconnected via a wired interface, or at least some of them may be wirelessly connected via a communication unit (300). For example, within the terminal (100), the processor (310) and the communication unit (300) may be wired, and the processor (310) and the memory (320) and / or additional elements may be wirelessly connected via the communication unit (300). Additionally, each element, component, unit / part, and / or module within the terminal (100) may include one or more additional elements. For example, the processor (310) may be composed of a set of a communication control processor, an application processor, an Electronic Control Unit (ECU), a graphics processing processor, a memory control processor, etc. As another example, memory (320) may be composed of RAM (Random Access Memory), DRAM (Dynamic RAM), ROM (Read Only Memory), flash memory, volatile memory, non-volatile memory and / or a combination thereof.

[0091] Wireless communication systems are in the process of adopting artificial intelligence technology to provide more effective and faster communication services. For example, advancements in AI / ML (Artificial intelligence / machine learning) technology are leading to the intelligentization and sophistication of node(s) and terminal(s) constituting wireless communication networks.

[0092] In particular, due to the intelligence of networks and base stations, it is expected that various network / base station decision parameter values ​​can be rapidly optimized, derived, and applied according to various environmental parameters. Environmental parameters may include at least one of the distribution / location of base stations, the distribution / location / material of buildings / furniture, the location / movement direction / speed of terminals, and climate information. However, the parameters described above are merely examples, and the environmental parameters may include other environmental parameters related to the network / base station decision parameters in addition to the parameters listed. Network / base station decision parameter values ​​may include at least one of the transmit / receive power of each base station, transmit power of each terminal, precoder / beam of the base station / terminal, time / frequency resource allocation for each terminal, and duplex method of each base station. However, the parameters described above are merely examples, and the network / base station decision parameter values ​​may include other parameters determined by the network / base station in addition to the parameters listed.

[0093] In a narrow sense, AI / ML can easily be referred to as artificial intelligence based on deep learning, but conceptually, it can be distinguished as follows.

[0094] - Artificial Intelligence: This refers to all automation where machines can take over tasks that humans would otherwise have to perform.

[0095] - Machine Learning: Machines learn patterns for decision-making from data on their own, without explicitly programming rules.

[0096] - Deep Learning: A model based on artificial neural networks that enables machines to perform everything from feature extraction to judgment from unstructured data in a single step. The algorithm relies on multi-layered networks composed of interconnected nodes for feature extraction and transformation, inspired by biological nervous systems, or neural networks. Common deep learning network architectures include Deep Neural Networks (DNN), Recurrent Neural Networks (RNN), and Convolutional Neural Networks (CNN).

[0097] As mentioned above, artificial intelligence (AI) is the broadest concept of AI / ML, and deep learning is the narrowest concept of AI / ML. Machine learning (ML) can be interpreted as a concept that is narrower than artificial intelligence but broader than deep learning.

[0098] In this specification, AI / ML or artificial intelligence is a comprehensive concept that includes the aforementioned deep learning and does not limit specific models or specific learning / inference methods.

[0099] FIG. 5 is a diagram illustrating an AI / ML operation workflow according to one embodiment of the present disclosure.

[0100] Referring to Fig. 5, the AI / ML operations workflow highlights the key stages of the ML model lifecycle. The ML model lifecycle includes Training (510), Testing (520), Emulation (530), Deployment (540), and Inference (550).

[0101] ML Model Training (510): Includes initial training and retraining of an ML model (or family of models). It also includes validation to evaluate the performance of the model when performed on training and validation data. If the validation results fall short of expectations (e.g., variance exceeds the acceptable range), the model may be retrained. During the training phase, validation of the model's generalization performance, stability, and operability can be performed, and the results of such performance validation can be used as defined threshold values ​​in the subsequent operation phase.

[0102] ML Model Testing (Testing, 520): A validated ML model is tested to evaluate the performance of the ML model trained on test data. If the test results meet expectations, the next step can be performed. If the test results do not meet expectations, the model can be retrained. This can be defined as the concept of pre-deployment validation of AI / ML capabilities and can also function as a standard safeguard to prevent the deployment of models that may compromise network stability.

[0103] AI / ML Inference Emulation: Before applying to a target network or system, the model is run in an emulation environment to evaluate inference performance. If the emulation results fall short of expectations (e.g., failure to meet target inference performance, negative impact on the performance of existing features), the model must be retrained. AI / ML inference emulation is optional and can be omitted from the ML model lifecycle.

[0104] ML Model Deployment: Includes a model loading procedure that makes the trained model available for use in the target AI / ML inference function. Deployment may not be necessary in cases where the training function and the inference function are co-located in the same location, etc. The model deployment may include a model identifier, version information, scope, activation conditions, etc., and the present disclosure may manage such deployment metadata in conjunction with the performance monitoring and inference steps.

[0105] AI / ML Inference: The AI / ML inference function performs inference using a trained model. It may also trigger retraining or model updates based on performance monitoring and evaluation. Here, run-time performance monitoring or closed-loop monitoring of the AI / ML model is considered an essential function for detecting model degradation over time or due to environmental changes and for taking appropriate countermeasures; control over the aforementioned time and environment can be variably configured considering the service, stability, and reliability.

[0106] Depending on the system implementation method and the configuration of AI / ML functions, both the AI / ML inference emulation step and the ML model deployment step may be omitted. Meanwhile, since model performance may degrade over time or due to changes in the wireless environment or traffic characteristics, retraining and model update procedures based on model performance feedback may be performed periodically or on an event basis. As such, when AI / ML models are applied in a communication system, training, storage, deployment, inference, performance monitoring, and retraining can be adaptively applied and operated, taking into account the reliability, stability, and operability of the AI / ML-based network.

[0107] FIG. 6 is a diagram illustrating the data collection and reporting procedure of an AI / ML model to which embodiments of the present disclosure can be applied.

[0108] Referring to FIG. 6, 610 and 620 can each be nodes in a communication system. For example, 610 can be a base station and 620 can be a terminal. Conversely, 610 can be a terminal and 620 can be a base station. In addition, 610 can be a first base station and 620 can be a second base station. That is, a data request procedure between base stations can be performed. Similarly, 610 can be a first terminal and 620 can be a second terminal.

[0109] The exchange of Data Collection Request, Data Collection Failure, and Data Collection Update messages between the first node (610) and the second node (620) is described as an example. The first node (610) can send a message requesting data collection to the second node (620) for training, inference, management, and updating of an artificial intelligence model (S601). The message requesting data collection may vary depending on the subject of each node. For example, the message requesting data collection may be transmitted through the Xn interface, NG interface, or Uu interface. The message requesting data collection may be an upper-layer message or a lower-layer message. S601 is a procedure in which the first node (610) requests specific operational data, performance indicators, or measurement data from the second node (620). At this time, the data may be used for integration with the Self-Organizing Network (SON), Network Data Analytics Function (NWDAF), or OAM. Through this, it is possible to link with OAM / MDT (Minimization of Drive Test) data collection or E2E performance analysis.

[0110] The second node (620) determines whether the requested data can be collected. For example, if the 620 does not possess the requested data or if collection is difficult, the second node (620) sends a data collection failure message to the first node (610) (S602). For example, the data collection failure (S502) procedure may return a failure response when the requested data does not exist or when the node cannot provide it due to a lack of authority or resources. This corresponds to an error handling procedure and can be transmitted as a failure cause IE (Information Element) specified in the standardization. For example, it may include unsupported parameter, resource not available, etc. As described above, the data collection failure message may also be configured with various interfaces and message formats depending on the subject of each node. The data collection failure message may include information on the cause of the data collection failure. Alternatively, information reported through updates may include predicted resource status information, terminal performance feedback information, measured terminal path, energy cost, etc.

[0111] If the second node (620) succeeds in collecting the requested data, it may send a data collection update message to the first node (610) (S603). For example, the data collection update (S503) procedure may subsequently send an update message to the first node (610) when the situation improves or when some data becomes available. This includes sending via asynchronous reporting or periodic / conditional update methods. Examples may include RAN performance counters, QoE measurements, QoS Flow utilization, location / traffic patterns, etc. The data collection update message may include data information collected regarding the requested data. In the case of previously transmitted data, it may be included in the data collection update message by indicating an update.

[0112] Through this process, data collection and transmission between each node can be performed.

[0113] As explained above, the artificial intelligence model can be configured in the terminal and / or base station.

[0114] In particular, a data collection framework centered on linkage with NWDAF (Network Data Analytics Function), SON (Self-Organizing Network), OAM (Operation, Administration and Maintenance), and MDT (Minimization of Drive Test) is being discussed as a function to systematically collect operational data, performance indicators, and measurement results generated from RAN, Core, and terminals for use in AI / ML learning and inference. Accordingly, the data collection request, failure reporting, and data collection update procedures between the first node (610) and the second node (620) described in FIG. 6 can directly correspond to the analysis and management functions defined in 3GPP standards. For example, the procedure in which the first node requests data collection from the second node as the entity operating the network analysis or AI / ML model aligns with the standard mechanism for NWDAF or RAN intelligence functions to collect input data from terminals or base stations. At this time, the data subject to collection can be expanded to include RAN performance counters, QoS Flow statistics, QoE measurement results, location and mobility information, traffic patterns, and energy consumption indicators, which can be utilized for training, inference, and performance monitoring of AI / ML models. Additionally, the failure cause information included in data collection failure messages can be used for evaluating the reliability of input data, analyzing the causes of model performance degradation, or selecting alternative data sources in terms of AI / ML model operation.In addition, the data collection update procedure can be performed using asynchronous reporting, periodic reporting, or event-triggered reporting methods. For example, data collection updates can be triggered when a specific threshold is exceeded, performance degradation is detected, or environmental changes occur, allowing AI / ML models to update inference results or policy decisions by reflecting the latest network status.

[0115] FIG. 7 is a diagram illustrating a classification according to the configuration of an AI / ML model to which embodiments of the present disclosure can be applied. FIG. 7 performs wireless resource management and beam management optimization using an artificial intelligence / machine learning (AI / ML) model.

[0116] First, (a) in the case of a network-sided AI / ML model, an AI / ML model located at a base station (710) performs resource optimization operations such as scheduling, beam selection, and interference management based on channel state information, measurements, performance indicators, etc. collected from a terminal (700).

[0117] Next, (b) in the case of a terminal-side AI / ML model (UE-sided AI / ML Model), the AI / ML model located at the terminal (700) performs beam prediction, performance estimation, and movement pattern analysis on its own and reports the results to the base station (710), thereby improving the accuracy of network control.

[0118] Finally, (c) in the case of a two-sided AI / ML model, the base station (710) and the terminal (700) each perform AI / ML inference, and the utilization of network resources is maximized by mutually exchanging inference results, performance indicators, and confidence level information.

[0119] Accordingly, the present embodiments perform wireless resource optimization operations in a network-only, terminal-only, or network-terminal collaborative structure depending on the deployment location of the AI / ML model, thereby providing the effect of enabling the implementation of ultra-high-speed, ultra-low-latency, and intelligent services for next-generation wireless communication systems.

[0120] Referring again to FIG. 6, the artificial intelligence model may be configured in only one of the terminal (700) or the base station (710). This can be described as a one-side AI / ML model. Alternatively, the artificial intelligence model may be configured in both the terminal (700) and the base station (710). This can be described as a two-side AI / ML model.

[0121] In the form of a one-side AI / ML model, an artificial intelligence model is configured at a base station (710), and a terminal (700) transmits information necessary for model input to the base station (710) via a wireless interface. The base station (710) can generate inference results by providing the received information as input to the model. Through this, the base station (710) can directly utilize the model to perform beam management, terminal mobility management, etc. This is called a NW-side AI / ML model.

[0122] In another form of the one-side AI / ML model, an artificial intelligence model is configured in the terminal (700), and the base station (710) transmits information necessary for model input to the terminal (700) through a wireless interface. For example, a reference signal can be transmitted through the base station (710) beam, or a reference signal for channel measurement can be transmitted. In this case, the base station (710) can transmit the signal that was previously transmitted. The terminal (700) can generate an inference result by providing the received information as input to the model. The terminal (700) transmits the generated inference result to the base station (710). Through this, the terminal (700) performs inference using the model and transmits the inference result to the base station (710) to assist with beam management of the base station (710), mobility management of the terminal, etc. This is called a UE-side AI / ML model.

[0123] In the case of a two-sided AI / ML model, a model may be configured at the terminal (700) and the base station (710), respectively. In this case, the terminal (700) and the base station (710) can input mutually received data into the model to derive an inference result. The terminal (700) can transmit the inference result to the base station (710). The base station (710) can control the operation of the terminal using the inference result received from the terminal (700) and the inference result of the base station (710). The models configured at the terminal (700) and the base station (710) may each be configured with different purposes or functions.

[0124] As such, AI / ML capabilities vary depending on specific use cases and sub-use cases, and may focus on the interaction between the network and the terminal.

[0125] FIG. 8 is a diagram illustrating a terminal-network collaboration level using an AI / ML model to which embodiments of the present disclosure can be applied.

[0126] Referring to FIG. 8, the collaboration levels between the base station (810) and the terminal (800) can be defined as follows.

[0127] 1. Level X (or Collaboration Level 0) is characterized by a method in which the terminal (800) and the base station (810) each independently operate AI / ML models, and the network-side (NW-side) AI model and the terminal-side (UE-side) AI model perform only self-optimization without standardized signaling. For example, the base station (810) uses its own beam management model, and the terminal (800) uses its own power optimization model. That is, the base station (810) and the terminal (800) each independently operate their own AI / ML models, and both sides operate through AI / ML logic implemented within their own or the device without standardized signaling or model sharing. For example, beam management, scheduling, and power control of the base station can be performed using their own AI models, and operations such as power consumption optimization based on the terminal's movement pattern, antenna selection, and local beam prediction can be performed independently. In this Level X, unpredictable operations may occur between them due to the opacity of the AI ​​operations.

[0128] 2. Level Y (Collaboration Level 1) allows the terminal (UE, 800) and the base station (810) to perform the step of exchanging signals with each other to assist in the inference and operation of an AI model. Specifically, the terminal (800) reports the inference results or performance indicators (e.g., Top-K beam prediction accuracy, confidence level, etc.) of the AI / ML model it has performed itself to the base station (810). After receiving the reported information, the base station (810) performs network optimization decisions, such as resource management, beam management, and scheduling, based on this information. In particular, as part of the AI / ML model performance monitoring and auxiliary signaling procedures, signaling definitions such as RRC information elements (RRC IE) or MAC control elements (MAC CE) are required. Therefore, compared to the case where the terminal (800) and the base station (810) operate the AI ​​model independently according to the present disclosure, this provides a technical advantage of being able to efficiently manage signaling overhead while improving mutual collaboration performance. The present disclosure includes the ability for a terminal to report “inference output” or “auxiliary information (e.g., Top-K candidate beams, confidence / probability, etc.)” to be utilized for network beam management / resource determination, and may also include the ability to apply UE-side inference in a collaborative form that combines it with network procedures. Additionally, for the operation of AI / ML functions, the network may instruct the UE regarding the activation / deactivation, fallback, switching, etc., of AI / ML functions through RRC, MAC-CE, and DCI signaling, and in this case, the concept of lifecycle management (LCM) may be considered.

[0129] 3. Level Z (Collaboration Level 2) allows the base station (810) and the terminal (UE, 800) to directly exchange the AI ​​model itself or learned parameters. Specifically, the base station (810) can download the AI ​​model learned on the network side to the terminal (800) to enhance the inference function at the terminal, while conversely, the terminal (800) can upload parameters learned independently or model update information to the base station (810) to enhance the central management function of the network. To this end, procedures for distributing and updating the AI ​​model can be performed, and include defining dedicated signaling procedures for model transmission, security / integrity guarantees, and transmission efficiency optimization techniques. For example, RRC protocol extensions, new MAC CE definitions, or interoperability structures with NWDAF / AF can be considered. Accordingly, according to the present disclosure, the base station (810) and the terminal (800) raise the level of cooperation to the highest level (full collaboration), providing the advantage that the network and the terminal can perform resource management and service optimization based on the same AI model. This can be applied adaptively while simultaneously considering issues such as increased signaling overhead and standardization complexity. The base station (810) and the terminal (800) can deliver / transfer an AI / ML model (or model parameters) via RRC signaling, which may include the transmission of a model identifier (model ID) and metadata (version / coverage / validity period, etc.). Here, the exchange of models / parameters can be applied adaptively according to the service / situation by considering not only transmission overhead but also integrity and security (e.g., prevention of model tampering, secure download / update) and storage / cache policies (e.g., whether the UE can store the model). That is, Level Z can be applied adaptively by considering standardization complexity.

[0130] More specifically, Collaboration Level 0 involves the terminal operating through its own AI / ML model or performing only RRC / RRM procedures without AI functions, and includes operations where the base station does not separately configure or intervene with the AI ​​model for the terminal. In other words, it is characterized by the absence of AI / ML-related interaction between the base station and the terminal. Meanwhile, Collaboration Level 1 is characterized by the fact that while the model itself is not transmitted, configuration / signaling information to support AI / ML functions is exchanged between the base station and the terminal. For example, the terminal receives configuration information (e.g., reference signal configuration, evaluation conditions, reporting cycle, etc.) for using the AI / ML model from the base station. Considering this, the terminal reports only the results of AI inference or measurement results, while the model itself exists independently within the terminal. That is, the terminal can report the results of local AI model execution used for CSI enhancement, beam management, mobility enhancement, etc., to the base station. In addition, Collaboration Level 2 enables the base station to transmit the AI / ML model itself or some parameter / configuration information to the terminal, and the terminal to perform AI functions based on the model (e.g., inference module, weight, etc.) received from the base station, or to feed back the model results learned by the terminal to the base station, or to support bidirectional model updates. This can be referred to as collaborative inference and federated learning-like operations. To this end, the terminal and the base station can transmit and receive model-related signaling, such as the configuration, parameters, and training cycle of the AI ​​model. The technology described herein may be applied to at least one of the three collaboration levels mentioned above.

[0131] Meanwhile, when a model is configured on a terminal, various methods may be applied to deliver the model to the terminal.

[0132] For example, the model can be trained on a terminal, a base station, or a separate device. The trained model can be stored on a server via model transfer / delivery. The model stored on the server can be delivered to the terminal as needed and configured on the terminal.

[0133] As another example, the model may be trained on a terminal or a separate device. The trained model is transmitted to a base station through a model delivery procedure and stored at the base station. The base station can then transmit the model to the terminal as needed for configuration on the terminal.

[0134] As another example, a model can be trained at a base station. The trained model is stored at the base station. The base station storing the model may be the same as or different from the base station that performed the training. The base station storing the model can then transmit the model to a terminal as needed to configure the model on the terminal.

[0135] Meanwhile, AI / ML models can be applied to various usage scenarios.

[0136] For example, it can be used for the purpose of saving network energy. For instance, to balance performance and energy savings, it is necessary to optimize cell activation or deactivation decisions based on predicted traffic loads by utilizing AI / ML technology. To this end, internal base station information such as current and predicted resource status, UE path prediction, currently predicted UE traffic, and predicted resource status of adjacent RAN nodes can be used as inputs to the model. Additionally, terminal-related information such as UE location data, UE measurement reports, and cell-level and beam-level UE measurement reports can be used as inputs. Furthermore, neighbor network information such as currently predicted energy efficiency, currently predicted resource status, and current energy status can be used as inputs. The model's output can infer energy saving strategies (recommended cell activation / deactivation), handover strategies, recommended candidate cells to take over traffic, predicted energy efficiency, and predicted energy status information. Feedback considers the resource status of adjacent RAN nodes, energy efficiency, UE performance affected by energy saving measures, and system KPIs.

[0137] As another example, AI / ML technology may be used for load balancing purposes. For instance, this functionality can be applied to enhance the quality of user experience and improve system capacity by improving load balancing performance through AI / ML model-based solutions and predicted loads, based on various metrics, feedback collected from UEs and network nodes, and historical data. To this end, the model's inputs may include base station information such as UE mobility path predictions, currently predicted energy efficiency, and currently predicted resource status, as well as terminal information such as UE location data, UE movement history data, and cell-level and beam-level UE measurement reports. Additionally, neighbor network information may include terminal performance measurement data from adjacent cells to which traffic has been offloaded. The model's outputs may infer target cell selection for load balancing, predicted local resource status information, a resource status model of predicted adjacent RAN nodes, and predicted UE information selected for handover to the target RAN node. Feedback information may include UE performance data of the target NG RAN, updates to the target NG RAN's resource status information, and system KPIs.

[0138] As another example, AI / ML technology may be used from a mobility perspective. For instance, based on AI / ML, functions such as (1) reducing the probability of unintended mobility events (early / late handover, incorrect cell handover), (2) predicting UE location mobility performance, and (3) optimizing traffic steering can be performed. To this end, the input to the model may include base station information such as UE mobility path prediction, currently predicted energy efficiency, and currently predicted resource status, as well as terminal measurement information such as UE location information, UE movement history information, and wireless measurements of adjacent cells to the UE service cell. Additionally, neighbor base station information may include UE history information, the location of the handovered UE, QoS parameter performance information, and information on past UE handovers that were successful or failed in the currently predicted resource status. The output of the model may include UE path prediction, conditional handover-related prediction, UE traffic prediction, and the validity period of the model output.

[0139] In addition, AI / ML capabilities can be used to analyze metrics related to network and UE performance to perform optimal resource management and mobility decisions for network slicing, thereby ensuring the Quality of Service (QoS) of each slice and improving network efficiency. Alternatively, AI / ML capabilities can provide network performance improvements and maintain consistency in the UE experience by proactively detecting and responding to Coverage and Capacity Optimization issues.

[0140] In addition to these scenarios and usability, the operations described in this specification can be applied to various past, present, and future scenarios.

[0141] Below, we will explain several representative scenarios and their applications.

[0142] AI / ML models can be used to enhance CSI feedback. Based on channel state feedback information, CSI feedback enables base stations to accurately identify the user's wireless channel status, allowing for optimal wireless resource allocation and transmission. However, a disadvantage exists in that the feedback load required for accurate channel state information feedback increases significantly as the number of antennas and subbands increases. To address this, technology for delivering accurate channel state information at an appropriate feedback load is required.

[0143] CSI feedback is a critical procedure for adaptive communication technologies such as adaptive modulation, coding, and beamforming. Base stations utilize CSI feedback received from terminals to determine detailed downlink transmission methods, thereby improving overall network efficiency and system performance. However, periodically transmitting CSI feedback can cause system overhead and waste wireless resources, and pose a problem where it becomes difficult to manage as the number of users increases.

[0144] By utilizing AI / ML models to further enhance CSI feedback, it may be possible to provide benefits such as reduced overhead and improved accuracy.

[0145] As mentioned above, AI / ML is used in various fields in wireless communication systems, such as beam management, channel state quality prediction, location estimation, and mobility management.

[0146] For example, the goal of AI / ML-based Beam Management (BM) technology is to reduce measurement overhead and latency by optimizing downlink Tx beam prediction in the spatial and temporal domains for data transmission. To this end, there are two types of beam sets, Set A and Set B. Set A is the set of beams targeted for prediction, and Set B is the set of beams used for measurement. In BM-case 1, downlink transmission beam prediction for Set A is performed in the spatial domain based on the measurement results of Set B beams. On the other hand, in BM-case 2, downlink Tx beam prediction for Set A is performed in the temporal domain based on the historical measurement results of Set B beams.

[0147] Such a Set A / Set B-based AI / ML beam management structure can be configured with consideration for measurement overhead reduction or predictive beam management. Additionally, instead of periodically measuring all candidate beams, a structure that predicts the Set A beam using only limited Set B beam measurement results can be operated, which can be considered as a low-latency, high-efficiency beam operation method in the FR2 or higher band.

[0148] AI / ML-based beam management models are classified into UE-side models and NW-side models depending on who performs model training and inference. In UE-side models, the UE performs model training and inference. In NW-side models, the network performs model training and inference.

[0149] As such, as AI / ML models are actively used in various situations, the issue of performing performance evaluations on these models also becomes important. For example, UE / NW-side AI / ML models undergo life cycle management (LCM) through a performance monitoring process after training to determine whether to use the model (utilize it in the BM process or retrain the AI / ML model). The performance monitoring process may require comparing the inference results of the AI / ML model against the received beam ID using pre-configured performance metrics.

[0150] Life cycle management (LCM) refers to the process of managing AI / ML models throughout their entire lifecycle. This includes processes such as model training, inference, and updates, and is used to ensure that models operate efficiently in a 5G network environment and can be adjusted appropriately when necessary.

[0151] In the case of the NW-side model, the network can independently evaluate the model using the terminal's beam measurement results and indicate whether to activate it. However, in the case of the UE-side model, the terminal must perform the performance monitoring process based on the cooperation or control of the base station. To achieve this, it is necessary to define a protocol on the wireless interface between the base station and the terminal.

[0152] Below, we intend to specifically propose a method and device for monitoring the performance of an AI / ML model configured on a terminal in such a situation. In particular, we propose a method for utilizing a specific set of monitoring reference signals to enable the terminal to perform performance monitoring operations.

[0153] For example, the present disclosure proposes an embodiment regarding the configuration of beams used for performance monitoring of an AI / ML model (beam-managed) deployed to a terminal, and the output beams of the inference function of said model. For both the performance monitoring beams (monitoring reference signal set) and the inference beams (prediction reference signal set), the network can be configured by sending signals to the terminal. In particular, the network can direct the performance monitoring beams to the terminal through the relationship between the performance monitoring beams and the inference beams. Additionally, the network can enable an advanced monitoring mode, in which all beams configured for performance monitoring and inference are used. Otherwise, the network can configure the terminal to use only some of the beams configured for performance monitoring to reduce overhead. Depending on the configuration of the monitoring report and the inference report, beams may also be compared using the union of the two beam sets.

[0154] First, the configuration of Set A and Set B can be set through the RRC layer. For example, the base station can transmit to the terminal the range of candidate beams that may be included in Set A, the initial set of target beams for measurement included in Set B, and the maximum size or identifier (ID) of each set via an RRC Reconfiguration message. Additionally, the RRC layer can configure the purpose of Set A and Set B (e.g., for prediction, for measurement), the application scenario (BM-case 1 or BM-case 2), and whether the AI / ML-based beam prediction function is enabled. Through this, the terminal can recognize the long-term beam management structure and the scope of AI / ML application in advance.

[0155] Meanwhile, the actual beam measurement operation for Set B is performed at the MAC / PHY layer, and the measurement results can be utilized as input for AI / ML-based beam prediction. At the MAC layer, the subset of beams to be actually measured, the measurement timing, and the measurement cycle can be dynamically specified through DCI or MAC Control Elements (MAC CE). This allows the network to flexibly adjust measurement overhead according to traffic load, terminal mobility, and changes in the wireless environment. Additionally, Set A and Set B can be dynamically updated based on AI / ML-based inference results. For example, the MAC layer can instruct the reduction or replacement of the composition of measurement beams included in Set B by reflecting the inference results of AI / ML models on the terminal or network side. Along with this, the set of candidate beams for Set A can also be reconfigured based on prediction reliability or historical transmission success rates, and these dynamic updates can be performed via lightweight signals based on MAC CE or DCI. On the other hand, in the case of structural changes to Set A or Set B, such as changing the definition of the set itself or switching the AI / ML-based beam management method (BM-case 1 ↔ BM-case 2), reconfiguration through the RRC layer may be performed. This is a procedure to respond to long-term changes in the wireless environment or changes in service characteristics, and is distinct from short-term control of the MAC layer.

[0156] Consequently, the RRC layer is responsible for framework-level configuration and policy control of Set A and Set B, while the MAC layer controls the detailed operations required for measurement, prediction, and transmission within those frameworks in real-time or near-real-time. Through this cooperative structure between the RRC and MAC layers, AI / ML-based beam management can simultaneously achieve rapid beam adaptability and stable link quality while minimizing measurement overhead.

[0157] As explained, triggers for BM-case selection can be applied in relation to terminal mobility. For example, BM-case 1 can be configured to be selected when mobility is determined to be high by considering the terminal's movement speed, frequency of changes in movement direction, handover occurrence rate, or measurement event frequency. Since BM-case 1 is structured to immediately utilize near-real-time measurement results from Set B, it enables rapid spatial domain adaptation in environments where mobility is high and beam validity periods are short. Conversely, if mobility is low or repetitive movement patterns are observed, the RRC can be configured to select BM-case 2 to utilize past measurement history in the temporal domain. Additionally, BM-case switching can be controlled in relation to link quality and error performance. For example, if the Block Error Rate (BLER) exceeds a threshold for a certain period or the retransmission frequency (HARQ NACK rate) increases, the RRC layer may determine that the currently applied BM-case 2 is not sufficiently responding to link changes and instruct a switch to BM-case 1. Conversely, if the BLER remains stable and the need for remeasurement due to prediction failure is low, the system can be configured to maintain or switch to BM-case 2 to reduce measurement overhead. Additionally, frequency band characteristics, particularly whether ultra-high frequency bands such as FR3 are used, can serve as important RRC control conditions for BM-case selection. In FR3 environments, link quality is highly likely to change rapidly due to high path loss, shielding sensitivity, and increased Doppler influence; therefore, BM-case 1 can be set as the default mode when the FR3 band is active. This enables rapid beam adaptation based on near-real-time measurements. On the other hand, in FR1 or relatively stable FR2 environments, BM-case 2 can be applied to utilize time-domain prediction, thereby reducing beam measurement frequency and signal overhead.

[0158] As explained, instead of applying terminal mobility, BLER, and frequency band conditions as a single standard, the selection of BM-case 1 and BM-case 2 in the AI / ML-based beam management can be controlled by combining multiple conditions. For example, even if the FR3 band is used, BM-case 2 can be maintained if terminal mobility is low and BLER is stable; conversely, even in an FR2 environment, if mobility and error rate increase simultaneously, it can be configured to switch to BM-case 1. Through such multi-condition-based RRC settings, AI / ML-based beam management can operate flexibly according to environmental changes.

[0159] FIG. 9 is a diagram illustrating terminal operation according to one embodiment.

[0160] Referring to FIG. 9, a method for a terminal to perform a performance monitoring operation of an AI / ML model may include the step of receiving configuration information for performance monitoring of an AI / ML model from a base station, the configuration information including mapping information of a prediction reference signal set and a monitoring reference signal set (S900).

[0161] For example, the configuration information may further include reporting information for transmitting wireless resource information regarding a monitoring reference signal and performance monitoring result information. As an example, the configuration information may include wireless resource information regarding a monitoring reference signal that the terminal must measure for performance monitoring. The monitoring reference signal may be a CSI-RS, SSB, etc. The wireless resource information may include information regarding the time frequency resources at which the reference signal is transmitted. Alternatively, if the reference signal is transmitted repeatedly, the wireless resource information may further include information regarding the number of repetitions, repetition period, etc., related to the repeated transmission.

[0162] As another example, configuration information may include reporting information. The reporting information may include reporting resource information regarding which resources to use and at which time interval to transmit when the terminal transmits performance monitoring result information to the base station. Additionally, the reporting information may include time offset information for performing reporting after the monitoring reference signal is received, and information indicating the time relationship between the generation of inference result information and the transmission of performance monitoring result information.

[0163] As another example, the configuration information may include monitoring configuration identification information. The monitoring configuration identification information may include ID information to distinguish the configuration information from other configuration information (e.g., legacy CSI report configuration information). For example, the monitoring configuration identification information may be set in conjunction with the identification information of the inference configuration information to obtain inference result information for a prediction reference signal set output using an AI / ML model. The terminal may receive inference configuration information from the base station to perform operations such as the aforementioned beam prediction using an AI / ML model. The terminal may measure the received measurement reference signal (Set B) according to the inference configuration information and input the measurement result as input to the AI / ML model to generate inference result information for the prediction reference signal set (Set A).

[0164] Meanwhile, the configuration information may include mapping information between a prediction reference signal set and a monitoring reference signal set. The mapping information refers to information for indicating the reference signals that the terminal must monitor for performance monitoring of the AI / ML model, based on information regarding the reference signals predicted as a result of the terminal performing inference using the AI / ML model.

[0165] Here, the monitoring reference signal set may be identical to the prediction reference signal set or may consist of a subset of the prediction reference signal set. For example, the monitoring reference signal set consists of one or more monitoring reference signals transmitted by a base station for the terminal to monitor the performance of an AI / ML model. The monitoring reference signal set may be linked with the prediction reference signal set, which the terminal used to perform inference using the AI / ML model. This is because the prediction reference signal set and the monitoring reference signal set must be linked in order to estimate the accuracy of the actual prediction (inference) performance of the AI / ML model.

[0166] Therefore, the size of the monitoring reference signal set can be configured to be the same as the prediction reference signal set. Alternatively, the size of the monitoring reference signal set can be configured to be smaller than the prediction reference signal set.

[0167] The reference signals included in the monitoring reference signal set and the prediction reference signal set, respectively, may be interrelated. For example, the monitoring reference signal set and the prediction reference signal set may be composed of the same size and established in a 1:1 mapping relationship. Alternatively, the monitoring reference signal set may be composed of a subset of the prediction reference signal set and may consist of some of the reference signals in the prediction reference signal set.

[0168] Configuration information can be received by being included in higher-layer messages. For example, configuration information can be included in RRC messages. As one example, configuration information can be received in the form of CSI report configuration information. As another example, configuration information may be included as a detailed information element of the CSI report configuration information.

[0169] A method for a terminal to perform a performance monitoring operation of an AI / ML model may include a step of verifying a monitoring reference signal set indicated in conjunction with a prediction reference signal set output using the AI / ML model based on configuration information (S910).

[0170] For example, the terminal can identify a set of monitoring reference signals using mapping information included in the configuration information. The terminal can perform various configured operations, such as beam management, channel estimation, and RRM operations, using an AI / ML model.

[0171] For example, the terminal can receive a measurement reference signal set as Set B, generate a measurement result, and input this into an AI / ML model to derive a prediction result for a prediction reference signal set as Set A. As an example, in terms of beam management, the terminal can use the AI / ML model to predict which reference signal (beam) among the prediction reference signal set as Set A has the highest quality. As another example, in terms of channel estimation prediction, the terminal can predict the channel estimation of the prediction reference signal set as Set A in time or space. As yet another example, the terminal can control RRM operations, such as the terminal's handover, based on the prediction result of the prediction reference signal set as Set A.

[0172] However, AI / ML models configured in the terminal may experience degradation due to various factors, such as changes in the environment, changes in input data, and degradation over time. To address this, the terminal must monitor whether the performance of the AI / ML model is being maintained or to what extent it is demonstrating performance. To this end, the terminal can check a set of specified monitoring reference signals based on configuration information and use them for performance monitoring.

[0173] For example, the terminal can identify the monitoring reference signal set among the prediction reference signal sets by using mapping information in the configuration information.

[0174] For example, mapping information between a prediction reference signal set and a monitoring reference signal set can be configured as a bitmap to indicate the reference signals included in the monitoring reference signal set based on the prediction reference signal set. As it is configured in a bitmap format, the size of the bitmap is determined by the number of prediction reference signal sets. For example, the bitmap can represent the mapping relationship between the prediction reference signal set and the monitoring reference signal set. The x-th Most Significant Bit of the bitmap can be configured to correspond to the x-th resource of the prediction reference signal set, and the y-th Non-Zero Bit of the bitmap can be configured to correspond to the y-th resource entry of the monitoring reference signal set. That is, depending on the reference signals included in the prediction reference signal set, each bit of the bitmap indicates a correspondence with each prediction reference signal, and a bit with a value set to '1' can mean that it has been assigned to the monitoring reference signal set.

[0175] As another example, mapping information may include information indicating the number of reference signals included in the monitoring reference signal set relative to the prediction reference signal set. Additionally, mapping information may further include information indicating a pattern based on the configuration of the monitoring reference signal set relative to the prediction reference signal set. Through this, mapping information can indicate detailed reference signals based on the size and pattern of the monitoring reference signal set that is equal to or smaller in size than the prediction reference signal set.

[0176] A method for a terminal to perform a performance monitoring operation of an AI / ML model may include the step of transmitting performance monitoring result information for an AI / ML model using inference result information for a prediction reference signal set and quality measurement result information for a monitoring reference signal set (S920).

[0177] For example, the terminal receives and measures a monitoring reference signal transmitted by the base station based on configuration information. For example, the terminal calculates at least one indicator among L1-RSRP, RSRQ, and SINR for each monitoring reference signal. Accordingly, the terminal can perform measurements for each of two or more monitoring reference signals and calculate a quality indicator for each.

[0178] The terminal can generate performance monitoring result information for an AI / ML model using quality measurement result information including quality indicators measured for each monitoring reference signal and inference result information from the previous time inference.

[0179] For example, inference result information can be generated before quality measurement result information is produced. The inference result information may include information on predictions made by the terminal using an AI / ML model in a prediction reference signal set. This allows the terminal to generate performance monitoring result information for the AI / ML model by using the predicted value and the actual quality measurement result information as predicted values ​​before the quality measurement result information is produced.

[0180] For example, the inference result information may include the top K (where K is a natural number greater than or equal to 1) reference signal information from a set of prediction reference signals based on the prediction quality inferred through an AI / ML model. Here, K can be set to any one of 1 to 4. The value of K can be set by the base station and can be included in the inference configuration information and directed to the terminal. Here, the K prediction reference signals can be reported by the terminal to the base station as prediction result information.

[0181] As another example, the quality measurement result information may include the top M (M is a natural number greater than or equal to 1) monitoring reference signal information based on the quality measurement results of the monitoring reference signal set. M can be set to 1 or 2. The value of M can be indicated by the base station. M may also be included in the aforementioned CSI report configuration information. The terminal may generate quality measurement result information by selecting the top M reference signals with high quality based on the signal quality measurement results for each monitoring reference signal.

[0182] The terminal can generate performance monitoring result information regarding the prediction accuracy of the AI / ML model by utilizing the prediction result information described above and the subsequently measured quality measurement result information. Monitoring logic and other parameters for generating performance monitoring result information may be configured in the terminal in advance, stored through mutual agreement according to specifications, or instructed by being included in the configuration information via a higher-layer message from the base station. The higher-layer message may be CSI report configuration information.

[0183] For example, performance monitoring result information may be generated based on whether at least one of the top M monitoring reference signal information is included in the top K reference signal information. Specifically, to explain with an example, when M is set to 1 and K is set to 3, the performance monitoring result information includes information on whether the top 1 monitoring reference signal (beam) selected based on the quality measured by receiving the actual monitoring reference signal is included in the predicted reference signal (beam) reported as the top 3 of the predicted reference signal set. Alternatively, the performance monitoring result information may include information on how many or what ratio values ​​out of M are included in the K beams.

[0184] For example, let us assume a case where M is set to 2 and K is set to 2. The inference result information would have reported two reference signals predicted to be of higher quality to the base station. At this time, the reported beam indices are assumed to be #1 and #3. Subsequently, through the performance monitoring operation, the terminal calculates two reference signals measured as of higher quality based on the actual measurement results according to M as quality measurement result information. At this time, the beam indices of the two calculated monitoring reference signals are assumed to be #1 and #2. In this case, since one of the actually measured higher quality beams matches the predicted higher quality beam, the performance monitoring result information may include the number of matching beams (1) or ratio information (50%).

[0185] Additionally, the performance monitoring result information may include details on which beams matched the actual measurements and predictions. In this case, this information may be included as beam index information, or reference signal identification information may be included as well.

[0186] Through this operation, AI / ML performance monitoring can be performed by comparing the prediction results with the actual measurement results. In particular, for this comparison, the transmission and reception of signals regarding the configuration and correlation relationship between the prediction reference signal set and the monitoring reference signal set will be one of the important technical factors. Additionally, unnecessary system overhead can be reduced by using the prediction reference signal set previously executed by the terminal to indicate the monitoring reference signal set.

[0187] FIG. 10 is a diagram illustrating the operation of a base station according to one embodiment.

[0188] Referring to FIG. 10, a method for a base station to control the performance monitoring operation of an AI / ML model of a terminal may include the step of transmitting configuration information for performance monitoring of an AI / ML model, including mapping information of a prediction reference signal set and a monitoring reference signal set, to the terminal (S1000).

[0189] For example, for performance monitoring operations, the base station may transmit configuration information to the terminal. For instance, the configuration information may be CSI report configuration information and may be transmitted via a higher-layer message (e.g., RRC message).

[0190] For example, the configuration information may further include wireless resource information for two or more monitoring reference signals, reporting information for transmitting performance monitoring result information, and monitoring configuration identification information for identifying CSI report configuration information. Additionally, the configuration information may include mapping information between a prediction reference signal set and a monitoring reference signal set.

[0191] For example, the configuration information may include radio resource information regarding a monitoring reference signal that the terminal must measure for performance monitoring. The monitoring reference signal may be CSI-RS, SSB, etc. The radio resource information may include information regarding the time frequency resources at which the reference signal is transmitted. Alternatively, if the reference signal is transmitted repeatedly, the radio resource information may further include information regarding the number of repetitions, repetition period, etc., related to the repeated transmission.

[0192] As another example, configuration information may include reporting information. The reporting information may include reporting resource information regarding which resources to use and at which time interval to transmit when the terminal transmits performance monitoring result information to the base station. Additionally, the reporting information may include time offset information for performing reporting after the monitoring reference signal is received, and information indicating the time relationship between the generation of inference result information and the transmission of performance monitoring result information.

[0193] As another example, configuration information may include monitoring configuration identification information. Monitoring configuration identification information may include ID information to distinguish the configuration information from other configuration information (e.g., legacy CSI report configuration information). For example, monitoring configuration identification information may be set in conjunction with the identification information of inference configuration information to obtain inference result information for a prediction reference signal set output using an AI / ML model. The base station may transmit inference configuration information to a terminal to perform operations such as the aforementioned beam prediction using an AI / ML model. The terminal may measure the received measurement reference signal (Set B) according to the inference configuration information and generate inference result information for the prediction reference signal set (Set A) by feeding the measurement result as input to the AI / ML model. At this time, the terminal performs performance monitoring using the inference result information and the quality measurement result for the monitoring reference signal. Therefore, the inference configuration information and the configuration information need to have a mutually linked relationship. For example, the inference configuration information and the monitoring configuration identification information of the configuration information may be set to the same value.

[0194] As another example, configuration information includes mapping information between the prediction reference signal set and the monitoring reference signal set. As explained above, performance monitoring of the model is possible only by mutually utilizing the inference result information for the prediction reference signal set and the quality measurement result information for the monitoring reference signal set during the inference process. Therefore, a mutual mapping relationship between the prediction reference signal set and the monitoring reference signal set may also be required. To this end, mapping information between the prediction reference signal set and the monitoring reference signal set may be included.

[0195] For example, the terminal can identify the monitoring reference signal set among the prediction reference signal sets by using the mapping information included in the configuration information.

[0196] For example, mapping information between a prediction reference signal set and a monitoring reference signal set can be configured as a bitmap to indicate the reference signals included in the monitoring reference signal set based on the prediction reference signal set. As it is configured in a bitmap format, the size of the bitmap is determined by the number of prediction reference signal sets. For example, the bitmap can represent the mapping relationship between the prediction reference signal set and the monitoring reference signal set. The x-th Most Significant Bit of the bitmap can be configured to correspond to the x-th resource of the prediction reference signal set, and the y-th Non-Zero Bit of the bitmap can be configured to correspond to the y-th resource entry of the monitoring reference signal set. That is, depending on the reference signals included in the prediction reference signal set, each bit of the bitmap indicates a correspondence with each prediction reference signal, and a bit with a value set to '1' can mean that it has been assigned to the monitoring reference signal set.

[0197] As another example, mapping information may include information indicating the number of reference signals included in the monitoring reference signal set relative to the prediction reference signal set. Additionally, mapping information may further include information indicating a pattern based on the configuration of the monitoring reference signal set relative to the prediction reference signal set. Through this, mapping information can indicate detailed reference signals based on the size and pattern of the monitoring reference signal set that is equal to or smaller in size than the prediction reference signal set.

[0198] A method for controlling performance monitoring operations may include the step of transmitting a monitoring reference signal directed according to configuration information to a terminal (S1010).

[0199] The base station transmits a monitoring reference signal to the terminal based on configuration information. The terminal receives and measures the monitoring reference signal. For example, the terminal calculates at least one metric among L1-RSRP, RSRQ, and SINR for each monitoring reference signal. Accordingly, the terminal can perform measurements for each of two or more monitoring reference signals and calculate a quality metric for each.

[0200] A method for controlling a performance monitoring operation may include the step of receiving performance monitoring result information for an AI / ML model generated using inference result information for a prediction reference signal set and quality measurement result information for a monitoring reference signal set from a terminal (S1020).

[0201] The terminal can generate performance monitoring result information for an AI / ML model using quality measurement result information including quality indicators measured for each monitoring reference signal and inference result information from the previous time inference.

[0202] For example, inference result information can be generated before quality measurement result information is produced. The inference result information may include information on predictions made by the terminal using an AI / ML model in a prediction reference signal set. This allows the terminal to generate performance monitoring result information for the AI / ML model by using the predicted value and the actual quality measurement result information as predicted values ​​before the quality measurement result information is produced.

[0203] For example, the inference result information may include the top K (where K is a natural number greater than or equal to 1) reference signal information from a set of prediction reference signals based on the prediction quality inferred through an AI / ML model. Here, K can be set to any one of 1 to 4. The value of K can be set by the base station and can be included in the inference configuration information and directed to the terminal. Here, the K prediction reference signals can be reported by the terminal to the base station as prediction result information.

[0204] As another example, the quality measurement result information may include the top M (M is a natural number greater than or equal to 1) monitoring reference signal information based on the quality measurement results of the monitoring reference signal set. M can be set to 1 or 2. The value of M can be indicated by the base station. M may also be included in the aforementioned CSI report configuration information. The terminal may generate quality measurement result information by selecting the top M reference signals with high quality based on the signal quality measurement results for each monitoring reference signal.

[0205] The terminal can generate performance monitoring result information regarding the prediction accuracy of the AI / ML model by utilizing the prediction result information described above and the subsequently measured quality measurement result information. Monitoring logic and other parameters for generating performance monitoring result information may be configured in the terminal in advance, stored through mutual agreement according to specifications, or instructed from the base station via a higher-layer message. The higher-layer message may be CSI report configuration information.

[0206] For example, performance monitoring result information may be generated based on whether at least one of the top M monitoring reference signal information is included in the top K reference signal information. Specifically, to explain with an example, when M is set to 1 and K is set to 3, the performance monitoring result information includes information on whether the top 1 monitoring reference signal (beam) selected based on the quality measured by receiving the actual monitoring reference signal is included in the predicted reference signal (beam) reported as the top 3 of the predicted reference signal set. Alternatively, the performance monitoring result information may include information on how many or what ratio values ​​out of M are included in the K beams.

[0207] Additionally, the performance monitoring result information may include details on which beams matched the actual measurements and predictions. In this case, this information may be included as beam index information, or reference signal identification information may be included as well.

[0208] Performance monitoring result information can be received by the base station through uplink messages such as PUSCH or PUCCH.

[0209] Through this operation, AI / ML performance monitoring can be performed by comparing the predicted results with the actual measurement results.

[0210] Below, various embodiments and specific examples that can be performed by the aforementioned terminal and base station are described with reference to the drawings. Each embodiment described below may be combined with the operation of the aforementioned terminal and base station according to any combination.

[0211] Two performance monitoring methods can be defined for UE-side AI / ML model BM-Case1 and BM-Case2. Type 1 performance monitoring is performed by the Network Worker (NW), which determines and directs LCM operations. Type 2 performance monitoring is performed by the UE, which determines and directs model selection, activation, deactivation, and switching operations.

[0212] There are two options for Type 1 performance monitoring. Option 1 is NW-side performance monitoring, where the NW calculates and reports or executes reporting triggers based on measurements transmitted by the UE. Option 2 is UE-assisted performance monitoring, where the UE calculates performance metrics and reports them to the NW or transmits them via event-based reporting triggers.

[0213] For BM-Case1 and BM-Case2, the performance monitoring process may be considered in the scenario of Option2 (UE-assisted performance monitoring). As explained above, when the NW performs performance monitoring, the terminal must transmit assistance information for the NW's performance monitoring. For example, the assistance information may be the aforementioned performance monitoring result information. When the terminal performs performance monitoring and determines and performs an LCM operation, the terminal may transmit only the LCM result to the base station.

[0214] Artificial Intelligence / Machine Learning (AI / ML)-based beam management technology is required in wireless communication systems. Specifically, the present disclosure includes a beam prediction scenario of an AI / ML model, model input data, performance monitoring, and a beam configuration method.

[0215] 1. AI / ML Model-Based Beam Management Scenarios (BM-Case1 and BM-Case2)

[0216] Regarding beam management, spatial domain prediction (BM-Case 1) and temporal domain prediction (BM-Case 2) may be considered. First, BM-Case 1 refers to performing spatial-domain downlink (DL) beam prediction for Set A beams based on the measurement results of Set B beams. In this case, the training and inference of the AI / ML model can be performed on the network (NW) side or on the user device (UE) side. The beams of Set A and Set B may belong to the same frequency range.

[0217] Second, BM-Case2 refers to performing temporal downlink beam prediction for Set A beams based on the historic measurement results of Set B beams. As with BM-Case1, the training and inference of the AI / ML model can be performed on the network side or the UE side. Predictions for F future time instances can be obtained based on the output of the AI / ML model, where each prediction is for each time instance and F is at least 1. In summary, the difference between BM-Case1 and BM-Case2 is that BM-Case1 performs spatial domain prediction of Set A beams based on Set B input beams, whereas BM-Case2 performs temporal domain prediction.

[0218] 2. Beam Set Configuration and AI / ML Model Input

[0219] The relationship between Set A (prediction target) and Set B (input target) and the model input data for beam prediction can be configured as follows.

[0220] In the case of BM-Case1, Set A and Set B may be different (Set B is not a subset of Set A), or Set B may be a subset of Set A. Here, Set A is for DL ​​beam prediction, and the codebook composition of each set can be clearly defined. As input to the AI / ML model, methods may be considered to use only L1-RSRP measurements based on Set B, use assistance information together, use Channel Impulse Response (CIR), or use L1-RSRP measurements together with the corresponding DL Tx / Rx beam ID.

[0221] In the case of BM-Case2, Set A and Set B may be distinct, Set B may be a subset of Set A (not identical), or Set A and Set B may be identical. The measurement results of K (K≥1) of the latest measurement instances are used as model input. Specific input alternatives include using only L1-RSRP measurements based on Set B, using auxiliary information together, or using L1-RSRP measurements together with the corresponding DL Tx / Rx beam IDs. Here, Set B is the set of beams whose measurements are taken as input to the AI / ML model.

[0222] 3. Performance Monitoring for UE-Side AI / ML Models

[0223] When an AI / ML model is deployed on the UE side (BM-Case1 and BM-Case2), performance monitoring can be classified into Type 1 and Type 2. In Type 1 performance monitoring, configuration or signaling for measurement and / or reporting is transmitted from the base station (gNB) to the UE. The UE can operate according to the following options.

[0224] Option 1 (NW-side performance monitoring): The UE sends a report to the network, and the network calculates performance metrics.

[0225] Option 2 (UE-assisted performance monitoring): The UE calculates performance metrics and reports them to the network or reports events based on the performance metrics. Additionally, the network may instruct the UE to perform Lifecycle Management (LCM) actions, provided that the performance of the model monitoring mechanism and reporting overhead are taken into account. In Type 2 performance monitoring, instructions, requests, or reports for performance monitoring are transmitted from the UE to the base station. In some cases, such instructions, requests, or reports may not be necessary. The base station may provide configuration information to the UE for performance monitoring measurements and / or reporting. In particular, for UE-side model monitoring, the UE may make decisions regarding the selection, activation, deactivation, switching, or fallback actions of the model. Furthermore, a mechanism is provided to enable the UE to detect whether the relevant function or model is suitable or no longer suitable.

[0226] 4. Beam Configuration Methods for Beam Management AI / ML Models

[0227] The present embodiment provides a method for configuring a beam for performance monitoring of a beam-managed AI / ML model deployed to a UE, and a beam that is the output of the model's inference function. For both the performance monitoring beam and the inference beam, the network can send a signal to the UE to trigger an advanced monitoring mode in which all beams configured for performance monitoring and inference are used.

[0228] On the other hand, the network can be configured so that the UE uses a subset of the beams configured for performance monitoring to reduce overhead. Based on the configuration of the monitor report and the inference report, beams can be compared using the union of the two beam sets.

[0229] That is, the network can transmit configuration information to the terminal to indicate a monitoring reference signal set composed of a subset of the prediction reference signal set.

[0230] As described above, performance monitoring for the UE-side AI / ML model can be classified into Type 1 and Type 2, and in particular, in Type 1 performance monitoring, the network (NW) transmits configurations and signals to the UE for measurement and performance monitoring reporting.

[0231] As explained, in a VEAM management environment applying UE-side AI / ML models, BM-Case 1 and BM-Case 2 can be operated in different ways depending on the entity responsible for performance monitoring and the responsibility for Lifecycle Management (LCM). In particular, depending on whether the entity responsible for performance monitoring and LCM operations is the network or the terminal, they can be classified into Type 1 performance monitoring and Type 2 performance monitoring.

[0232] First, Type 1 performance monitoring is a method in which the network manages the performance of AI / ML models, determining LCM actions such as model activation, deactivation, or BM-Case transitions based on the performance monitoring results. When applied to BM-Case 1, the network can evaluate performance based on near-real-time measurements of the Set B beam reported by the terminal (e.g., CSI, RSRP, RSRQ, etc.). In this case, performance calculations can be performed directly by the network or in a UE-assisted manner utilizing auxiliary performance metrics calculated by the terminal, such as prediction accuracy, prediction confidence, or Top-K beam prediction success. This performance monitoring takes place at near-real-time timing in slot or mini-slot units, and the network can direct the model to remain active or perform a BM-Case transition based on the results. This structure is suitable for scenarios requiring integrated network control to respond to rapid link fluctuations.

[0233] On the other hand, when Type 1 performance monitoring is applied to BM-Case 2, the network evaluates performance based on the summary of historical measurement history of Set B beams collected from terminals, prediction failure rates, and long-term BLER trends. In this case, performance monitoring is performed at medium- to long-term timings ranging from a few frames to tens of frames, and the network can determine whether performance is degrading to trigger maintaining or switching to the BM-Case, or to retrain the model. This operational method is suitable for long-term optimization in conjunction with network analysis and automation functions such as NWDAF and SON.

[0234] Meanwhile, Type 2 performance monitoring is a method in which the terminal autonomously performs AI / ML model performance evaluation and LCM operations. When applied to BM-Case 1, the terminal immediately evaluates model performance using real-time measurements for the Set B beam and feedback regarding transmission success or failure (HARQ ACK / NACK, BLER correlation information, etc.). This evaluation can be performed in slot or mini-slot units, and the terminal can independently decide to enable or disable the model or switch to BM-Case based on the results. In this case, the terminal can minimize signaling overhead by selectively reporting only LCM results to the network instead of detailed performance information.

[0235] In addition, when Type 2 performance monitoring is applied to BM-Case 2, the terminal evaluates model performance by analyzing accumulated prediction success rates, mobility information, and environmental change history over a long time window. Based on these evaluation results, the terminal can autonomously decide whether to maintain or switch to the BM-Case, and can report the corresponding state or switch result to the network via low-frequency RRC signals only when necessary. This method can operate efficiently in environments with low mobility and repetitive patterns.

[0236] As such, BM-Case 1 and BM-Case 2 each have the characteristics of near-real-time spatial adaptation and time history-based prediction, and can be combined with Type 1 and Type 2 performance monitoring methods to support various operational scenarios.

[0237] FIG. 11 is a diagram illustrating the performance monitoring operation of a terminal according to one embodiment.

[0238] Referring to FIG. 11, the communication system is configured to include a base station (1110) and a terminal (1100). The performance monitoring procedure of an AI / ML model may be carried out by including the step of the base station (1110) transmitting configuration information to the terminal (1100) (S1105), the step of the terminal (1100) verifying a set of monitoring reference signals (S1115), the step of the base station (1110) transmitting an indicated monitoring reference signal (S1125), the step of the terminal (1100) generating performance monitoring result information (S1135), and the step of the terminal (1100) transmitting the performance monitoring result information to the base station (1110) (S1145).

[0239] Specific examples for each step are described below.

[0240] First, the base station (1110) transmits configuration information for monitoring the performance of the AI / ML model to the terminal (1100) (S1105). This configuration information can be transmitted via upper-layer signaling such as a Radio Resource Control (RRC) message, and specifically, it can be configured by being included in or linked to a Channel State Information (CSI) report configuration (CSI-ReportConfig).

[0241] This configuration information includes mapping information that defines the relationship between the Prediction Reference Signal Set, which is the target that the AI / ML model must infer, and the Monitoring Reference Signal Set, which must be measured for actual performance verification.

[0242] Here, mapping information can be implemented in various ways. In one embodiment, the mapping information may have a bitmap format. The size of this bitmap is determined by the total number of reference signals (e.g., SSB, CSI-RS, etc.) included in the prediction reference signal set. For example, the x-th most significant bit (MSB) of the bitmap corresponds to the x-th resource of the prediction reference signal set, and if the bit is a non-zero value such as '1', it may indicate that the corresponding resource is allocated as a resource entry of the monitoring reference signal set.

[0243] In another embodiment, the mapping information may include pattern information. For example, a method of selecting a beam to be monitored at specific intervals (k) from among all predictable beams (Type 1), or a method of selecting the first N beams (Type 0), may be applied. The base station (1110) can reduce measurement overhead by setting a beam subset size parameter (alpha) to the terminal (1100) to reduce the monitoring target by specifying it at a ratio of 1 / 2, 1 / 4, 1 / 8, etc. of the total beams. In particular, the performance monitoring result report may be set to a periodic, semi-persistent, or aperiodic method. Additionally, the performance monitoring result may be transmitted in a separate IE or MAC CE format distinct from the CSI report.

[0244] Additionally, the configuration information may further include time and frequency wireless resource information for the monitoring reference signal, and uplink resource information (e.g., PUCCH, PUSCH resources, reporting period, offset, etc.) for reporting performance monitoring result information to the base station (1110).

[0245] The terminal (1100) checks the monitoring reference signal set indicated in conjunction with the prediction reference signal set output using an AI / ML model based on the received configuration information (S1115).

[0246] At this time, the monitoring reference signal set may have the same configuration as the prediction reference signal set, but for efficient resource utilization, it may also be composed of a subset of the prediction reference signal set. The terminal (1100) interprets the mapping information (bitmap or pattern information) described above to identify which specific reference signals (beams) among the entire prediction target need to actually be measured and compared.

[0247] The base station (1110) transmits a monitoring reference signal instructed according to configuration information through a wireless section (S1125). The terminal (1100) receives the monitoring reference signal and performs quality measurement.

[0248] The terminal (1100) measures at least one quality indicator among L1-RSRP (Layer 1 Reference Signal Received Power), RSRQ (Reference Signal Received Quality), and SINR (Signal to Interference plus Noise Ratio) for each monitoring reference signal to generate quality measurement result information.

[0249] In parallel with or prior to this, the terminal (1100) runs an AI / ML model based on a separately set measurement reference signal (Set B) to acquire and retain inference result information for a prediction reference signal set (Set A). That is, the value predicted by the model and the actual measured value coexist within the terminal (1100).

[0250] The terminal (1100) compares the inference result information for the prediction reference signal set and the quality measurement result information for the monitoring reference signal set to generate performance monitoring result information for the AI / ML model (S1135).

[0251] Specifically, the inference result information includes the top K (Top K, where K is a natural number greater than or equal to 1, e.g., 3) predicted reference signal information based on the quality predicted by the AI / ML model. On the other hand, the quality measurement result information includes the top M (Top M, where M is a natural number greater than or equal to 1, e.g., 1) monitoring reference signal information based on the quality measured by the actual monitoring reference signal.

[0252] The terminal (1100) determines whether at least one of the top M beams actually measured is included in the top K beams predicted by the model (Hit). In one embodiment, when M=1 and K=3, if the actual best beam is included in the top 3 beams predicted by the model, the performance of the model can be determined to be valid. In another embodiment, the performance monitoring result information may be generated to include not only success / failure status but also the number of matching beams, ratio, or index information of the matching beams. The Hit-based performance evaluation can be defined by considering AI / ML model performance metrics (e.g., Top-K accuracy, confidence, prediction validity), and provides the advantage of being able to perform the determination of model validity in a lightweight manner at the terminal level.

[0253] As another example, the terminal (1100) may operate in Normal Monitor Mode or Advanced Monitor Mode according to instructions from the base station (1110). In Normal Monitor Mode, results are generated only for a configured subset to reduce reporting overhead, and if Advanced Monitor Mode is instructed because model performance degradation is suspected, comparison results may be generated for the entire beam or a more extended beam set. Here, the Normal Monitor Mode or Advanced Monitor Mode can be defined as default monitoring or on-demand / advanced monitoring, and, for example, unnecessary measurements and reporting can be prevented by triggering extended monitoring when the network suspects model performance degradation.

[0254] The terminal (1100) transmits the generated performance monitoring result information to the base station (1110) (S1145). This information may be transmitted via PUCCH (Physical Uplink Control Channel) or PUSCH (Physical Uplink Shared Channel), and may also be transmitted in the form of MAC CE (Medium Access Control Element).

[0255] The base station (1110) can analyze the received performance monitoring result information to determine the current performance status of the AI / ML model on the terminal (1100) side and use it to determine whether to retrain, update, or fallback to a legacy method of the model.

[0256] FIG. 12 is a diagram illustrating a performance monitoring operation according to a mode of a terminal according to another embodiment.

[0257] Referring to FIG. 12, an embodiment is described for dynamically operating Normal Monitor Mode and Advanced Monitor Mode to efficiently manage system overhead. There are no limitations on the terms Normal Monitor Mode and Advanced Monitor Mode. Additionally, the Normal Monitor Mode described herein may refer to a mode in which the aforementioned set of monitoring reference signals is composed of a subset of the set of predicted reference signals. Similarly, the Advanced Monitor Mode may refer to a case where the set of monitoring reference signals is identical to the set of predicted reference signals.

[0258] The communication system includes a base station (1210) and a terminal (1200).

[0259] The base station (1210) transmits configuration information for monitoring AI / ML model performance to the terminal (1200) (S1205). The configuration information may be transmitted via upper layer signaling, such as an RRC message. For example, the configuration information includes mapping information between a prediction reference signal set (Set A) and a monitoring reference signal set. In particular, in this embodiment, this mapping information may include multiple reference signal set setting information that can be applied variably depending on the monitoring mode, or may define a sub-set configuration that has low overhead by default.

[0260] For example, mapping information may include bitmap or pattern information indicating that a monitoring reference signal set is composed of a subset of a prediction reference signal set. Alternatively, configuration information may include a beam subset size parameter (alpha) to be used for monitoring, thereby setting a monitoring target reduced to 1 / 2, 1 / 4, 1 / 8, etc. of the entire beam.

[0261] Here, the beam subset size parameter α is a parameter used to determine the size of the monitoring reference signal set (Set B) relative to the total prediction reference signal set (Set A). In other words, α can be used as a reference value to control the ratio of beams actually used for measurement and comparison among the total predicted target beams during the AI / ML-based beam management and performance monitoring process. Specifically, when the total number of predictable beams is denoted as N_total, the number of beams to be monitored, N_monitor, can be defined as shown in Equation 1 below.

[0262] [Mathematical Formula 1]

[0263]

[0264] Accordingly, when α is 0, approximately half of the beams can be set as monitoring targets, when α is 1, approximately 1 / 4 of the beams can be selected as monitoring targets, and when α is 2, approximately 1 / 8 of the beams can be selected as monitoring targets.

[0265] As such, the α value plays the role of progressively reducing or expanding the size of the monitoring reference signal set and functions as a key parameter for balancing measurement overhead and performance verification accuracy. For example, when the α value is relatively small, including more beams as monitoring targets can improve the prediction accuracy and performance verification resolution of the AI / ML model, but the resulting measurement and reporting overhead may increase. Conversely, when the α value is relatively large, the number of monitoring target beams decreases, reducing measurement overhead, but the resolution for finely verifying model performance may be lowered. Therefore, the beam subset size parameter α can be adaptively set by comprehensively considering the wireless environment, service requirements, terminal mobility, and network load, and can be utilized as an important control means for efficiently operating AI / ML-based beam management and performance monitoring.

[0266] In addition, pattern-based selection parameters may be applied as a specific selection method for the monitoring reference signal set. For example, by defining an interval parameter k, one beam can be selected for every k interval from the total beam indices. For instance, when k=2, beam indices 0, 2, 4, 6, etc., can be selected as monitoring targets. In this case, the pattern-based selection can specify the first beam to start the pattern application through the start index (startIndex), and the number of selected beams (N_monitor) can be determined based on a value calculated by α or an explicitly set value.

[0267] In another embodiment, a bitmap-based selection parameter (bitmap size / mask) may be applied. In this case, the length of the bitmap may be set to be equal to the number of beams included in the prediction reference signal set (Set A), and each bit indicates whether the corresponding beam is included in monitoring. Specifically, if the bit value is '1', the beam is selected as a monitoring target, and if the bit value is '0', the beam is excluded from monitoring. In this regard, the α parameter may be used as a higher-level parameter that indirectly controls the density of '1's in the bitmap, and if α is relatively large, the ratio of '1's within the bitmap may be set to decrease. That is, α controls the scale of the monitoring target at a relatively coarse level, while the bitmap performs fine control at the individual beam level, and these roles can be applied in a divided manner.

[0268] Here, as an example, the bitmap setting may utilize an interval-based pattern (every k-th beam) method. In this method, a monitoring reference signal set (Set B) is configured by selecting one beam at regular intervals from among a plurality of beam indices included in a prediction reference signal set (Set A) according to a preset interval parameter k. For example, if k is set to 2, beams distributed at even intervals, such as 0, 2, 4, and 6 among the total beam indices, can be selected as monitoring targets. Accordingly, beams capable of representing the entire beam space can be efficiently sampled, and the effect of reducing signaling and measurement overhead for monitoring is provided.

[0269] In another embodiment, a first N beams selection method may be applied. In this method, a monitoring reference signal set (Set B) is configured by selecting N consecutive beams from the start index among the beams included in the prediction reference signal set (Set A). For example, when the beam indices are sorted, the top N beams or consecutive beam blocks corresponding to a specific range may be designated as monitoring targets. This method is suitable for efficiently performing performance verification for a specific direction, a cell center area, or an area where specific service characteristics are concentrated.

[0270] Compared to bitmap-based mapping methods, such pattern-based mapping methods can construct a monitoring reference signal set using relatively less signaling information, thereby reducing signaling overhead. On the other hand, bitmap-based mapping methods allow for more precise control by explicitly controlling the inclusion of individual beam units. Furthermore, in AI / ML-based beam management and performance monitoring, the size and bitmap range of the monitoring reference signal set (Set B) can be configured differentially depending on the monitoring mode. In Normal Monitor Mode, a monitoring reference signal set (Set B) reduced by the beam subset size parameter α can be constructed. In this case, Set B is set as a subset of the entire prediction reference signal set (Set A) and can be represented in the form of a partial bitmap or a sparse bitmap. This Normal Monitor Mode can be applied as the default mode during periods when the AI / ML model is operating normally. On the other hand, in Advanced Monitor Mode, the monitoring reference signal set (Set B) can be configured to be identical to the prediction reference signal set (Set A). That is, all beams included in Set A are designated as monitoring targets, and this can be represented in the form of a full bitmap.

[0271] The base station (1210) instructs the terminal (1200) to use Normal Monitor Mode (S1215). The instruction may be transmitted via separate L1 signaling (e.g., DCI) or MAC CE, or may be set to Default Mode by initial configuration information.

[0272] The terminal (1200) performs performance monitoring only on a configured subset of beams, rather than the entire predicted target beam, in accordance with the general monitoring mode instructions. The terminal (1200) generates performance monitoring result information by comparing the quality measurement results for the reduced monitoring reference signal set with the inference results of the corresponding AI / ML model.

[0273] Subsequently, the terminal (1200) transmits performance monitoring result information for the generated subset beam to the base station (1210) (S1225). This allows the basic performance trends of the AI / ML model to be continuously identified while minimizing the measurement load and uplink reporting overhead of the terminal (1200).

[0274] Here, although the transmission of configuration information in step S1205 and the operation of instructing the general monitoring mode in step S1215 are described as separate steps, information for instructing the general monitoring mode may be explicitly or implicitly merged into the configuration information and performed as a single step. For example, if only some prediction reference signals in the mapping information of the configuration information are instructed as monitoring reference signals, it may be determined that the general monitoring mode has been implicitly instructed.

[0275] Likewise, steps S1205 and S1255 can also be merged into a single step. In this case, if only the entire prediction reference signal in the mapping information of the configuration information is indicated as a monitoring reference signal, it may be determined that an advanced monitoring mode has been implicitly indicated.

[0276] The base station (1210) may determine, based on the analysis of subset-based result information received from the terminal (1200), that performance degradation of the AI / ML model is suspected or that more precise verification is required. In this case, the base station (1210) instructs the terminal (1200) to use the Advanced Monitor Mode (S1255). The mode switching between the Normal Monitor Mode and the Advanced Monitor Mode can be performed by L1 signaling (e.g., DCI) or MAC CE, or can be indicated through mapping information within the RRC configuration information. For example, if the RRC mapping information is configured such that Set B includes only some beams of Set A, it can be determined that the Normal Monitor Mode is applied, and if all beams included in Set A are indicated as Set B, it can be interpreted that the Advanced Monitor Mode is applied. In this way, the linkage structure between the monitoring mode and the Set B configuration and the mode switching can flexibly adjust the monitoring range according to the performance status of the AI / ML model.

[0277] The terminal (1200) that receives this instruction expands the monitoring target. Specifically, the terminal (1200) measures the monitoring reference signal for all beams (Full Beam) included in the prediction reference signal set or a beam set expanded beyond the general monitoring mode, and compares it with the inference result.

[0278] The terminal (1200) transmits performance monitoring result information for the entire beam generated for the extended target to the base station (1210) (S1265).

[0279] As a result, the wireless communication system provides a method to balance resource efficiency (general mode) and performance verification accuracy (advanced mode) by dynamically adjusting the range of the monitoring reference signal set.

[0280]

[0281] FIG. 13 is a diagram illustrating an operation for indicating a set of monitoring reference signals according to configuration information according to one embodiment.

[0282] Referring to FIG. 13, a prediction reference signal set (1300) that is the subject of inference for an AI / ML model and a bitmap-type mapping information (1350) that indicates the target that the terminal must measure for actual performance verification are illustrated.

[0283] For example, the prediction reference signal set (1300) represents the entire set of candidates for which an AI / ML model performs inference, and in the example of FIG. 13, it includes a total of 10 reference signals (RS #1 to RS #10). These reference signals may be SSB (Synchronization Signal Block) or CSI-RS (Channel State Information-Reference Signal) resources.

[0284] At this time, the mapping information (1350) within the configuration information transmitted by the base station to the terminal can be implemented in a bitmap format. The size (Bit length) of this bitmap is determined according to the number of reference signals included in the prediction reference signal set (1300). In the case of FIG. 13, since there are 10 prediction reference signals, it can be seen that the corresponding bitmap (1350) is also composed of 10 bits.

[0285] Each bit of the bitmap (1350) corresponds 1:1 with each resource of the prediction reference signal set (1300). Specific mapping rules can be set as follows.

[0286] Resource location mapping: The x-th most significant bit (MSB) or sequential bit location of the bitmap corresponds to the x-th resource of the prediction reference signal set (1300). For example, the first bit of the bitmap corresponds to RS #1 and the second bit corresponds to RS #2.

[0287] Monitoring Target Indicator: If a specific bit value in the bitmap is '1' (non-zero), it means that the predicted reference signal corresponding to that location has been selected as the actual monitoring reference signal. Conversely, if the bit value is '0', the corresponding reference signal is excluded from the monitoring target.

[0288] Therefore, analyzing the bit sequence "0 0 1 0 1 0 1 0 1 0" of the specific mapping information illustrated in FIG. 13 yields the following: RS #1 and RS #2 (bit value 0) are excluded from the monitoring target. RS #3 (bit value 1) is included in the monitoring reference signal set. RS #4 (bit value 0) is excluded. RS #5 (bit value 1) is included in the monitoring reference signal set. RS #7 and RS #9 (bit value 1) are each included in the monitoring reference signal set.

[0289] As a result, according to this bitmap setting, the terminal determines a sub-set containing only four reference signals of {RS #3, RS #5, RS #7, RS #9} out of a total of 10 prediction targets as the monitoring reference signal set.

[0290] Additionally, the y-th non-zero bit of the bitmap corresponds to the y-th resource entry of the monitoring reference signal set. For example, the first non-zero bit (y=1) is located at the 3rd position of the entire bit sequence, indicating that the 1st entry of the monitoring reference signal set is the 3rd resource (RS #3) of the prediction reference signal set. Likewise, the second non-zero bit (y=2) is located at the 5th position of the entire bit sequence, indicating that the 2nd entry of the monitoring reference signal set is the 5th resource (RS #5) of the prediction reference signal set.

[0291] This method allows the base station to clearly define the linkage between the prediction reference signal set (Set A) and the monitoring reference signal set, thereby enabling the terminal to accurately determine which predicted value and which measured value to compare (e.g., the predicted value of RS #3 vs. the measured value of RS #3) when generating performance monitoring result information.

[0292] This bitmap-based mapping method provides high flexibility. The base station can freely select any reference signal and designate it as a monitoring target depending on the channel environment or the situation of the terminal. In addition, by designating only a part of the entire set as a subset as shown in Fig. 13, the measurement overhead and reporting overhead of the terminal can be efficiently reduced while effectively verifying the performance of the AI / ML model.

[0293] As mentioned above, mapping information may be indicated by pattern information and count information.

[0294] Additionally, as an embodiment, bitmap-based mapping information may be configured individually for each service group. Here, a service group may refer to a set of services having different QoS requirements, latency sensitivity, reliability requirements, or traffic characteristics, and may include, for example, a URLLC group for ultra-low latency services, an eMBB group for bandwidth-intensive services, or a specialized group for XR / sensing linked services.

[0295] A base station can configure individual bitmaps corresponding to each service group, thereby configuring different monitoring reference signal sets (Set B) for each service group even if they share the same predicted reference signal set (Set A). For example, for service groups where latency and reliability are critical, a relatively high-density bitmap can be configured to include more reference signals as monitoring targets, whereas for service groups where overhead minimization is critical, a sparse bitmap can be configured to designate only a limited number of reference signals as monitoring targets. As a specific example, bitmaps can be configured by service ID.

[0296] Service ID #1 (e.g., Latency / Reliability Sensitive Service)

[0297] Bitmap: 1 1 1 1 1 1 1 1 1 1

[0298] In this case, all reference signals from RS #1 to RS #10 are set as monitoring targets, and Set B has the same configuration (Full bitmap) as Set A. Through this, precise performance verification across the entire beam range can be performed for the corresponding service.

[0299] Service ID #2 (e.g., General Data Service)

[0300] Bitmap: 0 0 1 0 1 0 1 0 1 0

[0301] In this case, only RS #3, RS #5, RS #7, and RS #9 are selected as monitoring targets, forming a reduced subset (Set B). This allows for the reduction of measurement and reporting overhead while maintaining basic performance trends.

[0302] Service ID #3 (e.g., low-priority / overhead sensitive service)

[0303] Bitmap: 0 0 0 1 0 0 0 1 0 0

[0304] In this case, only RS #4 and RS #8 are included as monitoring targets, and sparse bitmap-based minimal monitoring is performed.

[0305] In this way, by maintaining the same prediction reference signal set (Set A) while setting different monitoring bitmaps (Set B) for each service ID, the terminal can perform differentiated performance monitoring according to service characteristics using a single AI / ML model. Additionally, the base station can flexibly adjust the monitoring range for each service by updating only the bitmap corresponding to the service ID.

[0306] Such service group-specific bitmap settings can be configured in conjunction with service identifiers (QoS flow, 5QI, or logical service ID, etc.) at the RRC layer, and the terminal can perform different performance monitoring operations for each service group based on the received configuration information. Through this, the terminal can perform performance verification optimized for service characteristics while using a single AI / ML model.

[0307] Consequently, the service group-specific bitmap configuration structure further expands the flexibility of bitmap-based mapping, enabling differentiated performance monitoring based on service characteristics and providing the effect of more precise control over the balance between measurement and reporting overhead and AI / ML model validation accuracy.

[0308] FIG. 14 is a diagram illustrating an example of configuring a monitoring reference signal set according to one embodiment.

[0309] Referring to FIG. 14, an example of a specific configuration type for selecting a sub-set for monitoring and reporting from a prediction reference signal set (Set A) is described.

[0310] To reduce the reporting overhead of the terminal and perform efficient performance monitoring, the base station may select only a portion of the all possible beams to configure as a monitoring reference signal set. This subset configuration can be determined by the beam subset size parameter (α) and the selection pattern type.

[0311] The explanation will focus on Type 0 configuration (1400) and Type 1 configuration (1450). For example, a monitoring reference signal set may be configured as a subset of a prediction reference signal set.

[0312] For example, the beam subset size parameter α can be set to a 2-bit value and determines the ratio of the monitored beams to the total number of beams. Specifically, the number of beams to be used for performance monitoring reporting is determined according to the following formula.

[0313] Based on the output value of the beam management model and the beam subset size parameter, settings can be made to report monitoring results to the network.

[0314] The beam subset size parameter α is It is a 2-bit value that determines the number of beams to be used in the performance monitoring report.

[0315] Therefore, the subset of beams to be used in the performance monitoring report can be 1 / 2, 1 / 4, 1 / 8, or 1 / 16 of the total beam size output by rounding the UE-side AI / ML.

[0316] The subset selection pattern type parameter can be one of two types.

[0317] For type 0, the beam is selected from the first N beams, and N is It is equal to the value determined through, and α is the beam subset size parameter.

[0318] For type 1, the selected beam will satisfy Equation 2, and the first k-1 beam is not selected.

[0319] [Mathematical Formula 2]

[0320]

[0321] As shown in Fig. 14, the difference between Type 0 and Type 1 configurations can be observed when the output size of the beam management AI / ML model is 37 and the beam subset size parameter is set to 0. N can be determined to be 18 and k to be 2.

[0322] In this way, the size of the monitoring reference signal set can be reduced to 1 / 2, 1 / 4, 1 / 8, or 1 / 16 of the entire prediction reference signal set.

[0323] As explained, in the example of Fig. 14, the output size of the beam management AI / ML model, i.e., the total number of predictable beams N total We can assume a case where this is 37 and the beam subset size parameter α is set to 0. In this case, according to the formula, N monitor = is determined to be 18, and in the Type 0 configuration, the first 18 beams are selected, while in the Type 1 configuration, 18 beams distributed evenly according to the spacing parameter k = 2 can be observed as a difference.

[0324] As such, the monitoring reference signal set (Set B) can be configured with a reduced size compared to the prediction reference signal set (Set A), and its size and selection method can be explicitly indicated by the α parameter and selection pattern type. These indications can be implemented in the form of a bitmap format, 2-bit information representing the ratio, or a parameter representing the number of monitoring targets. Through this, the network can flexibly adjust the balance between measurement overhead and performance verification accuracy during AI / ML-based beam management and performance monitoring.

[0325] FIG. 15 is a diagram illustrating examples of a prediction reference signal set and a monitoring reference signal set according to one embodiment.

[0326] Referring to FIG. 15, the present disclosure may use configuration information transmitted by a base station to a terminal for both a monitored beam set and an inference beam set. By using a combination of the two beam set configurations, a measured beam can be mapped to an inference beam for AI / ML model performance monitoring. That is, through the mapping relationship between the prediction reference signal set (Inference Report Beam Subset) and the monitoring reference signal set (Monitor Report Beam Subset), comparable beams that are the subject of actual performance evaluation can be identified.

[0327] 1500 represents the Example Inference Report Beam Subset configured for reporting inference results of AI / ML models, and 1550 represents the Example Monitor Report Beam Subset configured for performing monitoring reports through actual measurements.

[0328] The base station can set configuration information for reporting AI / ML model inference results (Inference Configuration) and monitoring configuration information for performance verification (Monitoring Configuration) to the terminal, respectively.

[0329] The inference reporting beam subset (1500) is a set of beams set to be reported to base stations from the entire beam (Set A) in which the AI / ML model performs predictions. The inference reporting beam subset (1500) represents beams selected for reporting and beams that are not reported.

[0330] The monitoring report beam subset (1550) is a set of monitoring beams configured to allow the terminal to perform measurement in an actual wireless channel environment and perform monitoring operations according to the results. The monitoring report beam subset (1550) represents beams selected as reporting targets and beams that are not reported.

[0331] As described above, the indication for the beam to be reported among the beam sets can be defined by a bitmap or the aforementioned beam subset size parameter (α) and a selection pattern (Type 0 or Type 1), respectively. For example, both sets may be set to the same Type 1 (dispersive) pattern, or they may be set to different patterns.

[0332] To accurately assess the performance of an AI / ML model, it must be possible to compare the values ​​predicted by the model with the actual measured values ​​on a one-to-one basis. As indicated by the bidirectional arrow in Figure 15, the prediction reference signal and the monitoring reference signal can be mapped to comparable beams.

[0333] Using configuration information transmitted by the base station to the terminal, a beam to be compared can be determined through a union or mapping relationship between a monitoring beam set and an inference beam set. For example, a comparison can be performed by mapping only between a beam of a specific index selected as a reporting target in the inference reporting beam subset (1500) and a beam of the same index (or mapped index) selected as a reporting target in the monitoring reporting beam subset (1550).

[0334] The terminal calculates performance metrics only for comparable beams. That is, beams that are included in the inference set but excluded from the monitoring set (or vice versa) may be excluded from direct comparison in performance evaluation.

[0335]

[0336] Meanwhile, the terminal can generate performance monitoring result information using quality measurement result information obtained by measuring the quality of the received monitoring reference signal and inference result information regarding the predicted reference signal set. The performance monitoring result information can be composed of various logics as described above. For example, the performance monitoring result information can be generated based on whether at least one of the top M monitoring reference signal information is included in the top K reference signal information. Specifically, to explain with an example, when M is set to 1 and K is set to 3, the performance monitoring result information includes information on whether the top 1 monitoring reference signal (beam) selected based on the quality measured by receiving the actual monitoring reference signal is included in the predicted reference signal (beam) reported as the top 3 of the predicted reference signal set. Alternatively, the performance monitoring result information may include information on how many out of M or what ratio values ​​are included in the K beams. The terminal can transmit the performance monitoring result information to the base station.

[0337] As described above, the present disclosure allows a base station to effectively reduce the measurement load and uplink reporting overhead of a terminal by configuring only a subset, rather than the entire prediction target, as monitoring reference signals when monitoring the performance of the terminal's AI / ML model. In particular, the embodiments according to the present disclosure establish an accurate correspondence between the predicted reference signals and the actual measured reference signals through a mapping relationship based on bitmap or pattern information, thereby clearly identifying the consistency between the model's inference results and the actual channel environment. Furthermore, the present disclosure can maintain a balance between resource efficiency and verification accuracy by dynamically switching between a general monitoring mode and an advanced monitoring mode depending on the situation. Additionally, the present disclosure can achieve reliability and beam management optimization of the entire wireless communication system by rapidly detecting and responding to performance degradation or errors in the terminal-side AI model through a method of comparing the top K predicted beams with the top M actually measured beams.

[0338] The terminal and base station configurations capable of performing the operations of each of the aforementioned embodiments are briefly described once again.

[0339] FIG. 16 is a drawing for explaining a terminal configuration according to one embodiment.

[0340] Referring to FIG. 16, a terminal (1600) performing a performance monitoring operation of an AI / ML model may include a receiver (1630) that receives configuration information for performance monitoring of an AI / ML model, including mapping information of a prediction reference signal set and a monitoring reference signal set, from a base station; a control unit (1610) that checks a monitoring reference signal set indicated in conjunction with a prediction reference signal set output using the AI / ML model based on the configuration information; and a transmitter (1620) that transmits performance monitoring result information for the AI / ML model using inference result information for the prediction reference signal set and quality measurement result information for the monitoring reference signal set.

[0341] The receiver (1630) receives configuration information for monitoring the performance of an AI / ML model from a base station. The configuration information includes mapping information between a prediction reference signal set and a monitoring reference signal set.

[0342] Specifically, the configuration information received by the receiver (1630) can be transmitted via Higher Layer Signaling, such as an RRC message, and may be in the form of CSI Report Configuration Information (CSI-ReportConfig) or an information element included therein.

[0343] Additionally, the configuration information may include time and frequency wireless resource information, whether repeated transmission is required, etc., for a monitoring reference signal (e.g., CSI-RS, SSB, etc.) that the terminal (1600) must measure for performance monitoring. Furthermore, the receiver (1630) may receive additional reporting information including reporting resource information and time offset information for transmitting performance monitoring result information. The receiver (1630) may receive monitoring configuration identification information. The identification information may be set in conjunction with inference configuration information for inference of an AI / ML model. Through this, the terminal (1600) can form a logical link between the inference operation and the monitoring operation.

[0344] The control unit (1610) interprets the mapping information within the configuration information to identify the monitoring reference signal that is the actual measurement target. At this time, the monitoring reference signal set may be identical to the prediction reference signal set or may be composed of a sub-set of the prediction reference signal set.

[0345] In one embodiment, the control unit (1610) can identify a monitoring target using mapping information in the form of a bitmap. The size of the bitmap is determined according to the number of prediction reference signal sets. The control unit (1610) specifically identifies which of the entire prediction targets is the actual measurement target by mapping the x-th most significant bit (MSB) of the bitmap to the x-th resource of the prediction reference signal set and mapping the y-th non-zero bit to the y-th resource entry of the monitoring reference signal set.

[0346] Additionally, the control unit (1610) runs an AI / ML model to generate inference result information for a prediction reference signal set (Set A) based on a measurement reference signal (Set B). The control unit (1610) controls the receiver (1630) to receive the indicated monitoring reference signal and calculates quality indicators such as L1-RSRP, RSRQ, and SINR to generate quality measurement result information.

[0347] The control unit (1610) generates performance monitoring result information by comparing the inference result information for the prediction reference signal set and the quality measurement result information for the monitoring reference signal set. For example, the control unit (1610) compares the top K reference signals (Top K, K is a natural number greater than or equal to 1) predicted through an AI / ML model with the top M monitoring reference signals (Top M, M is a natural number greater than or equal to 1) actually measured. The control unit (1610) determines whether at least one of the top M beams actually measured is included in the top K beams predicted (whether it is a hit), the number of matches or the ratio, etc., and constructs the final result information.

[0348] The transmitting unit (1620) transmits performance monitoring result information for the AI / ML model generated by the control unit (1610) to the base station.

[0349] The transmitter (1620) transmits the corresponding result information using the designated uplink resource (PUCCH or PUSCH) and time according to the previously received report information. This result information may include whether the performance is suitable, the index of the matching beam, or statistical information comparing the measured value and the predicted value.

[0350] In addition to this, the control unit (1610) controls the overall operation of the terminal (1600) in accordance with the performance monitoring operation for the AI / ML model required to perform the aforementioned invention.

[0351] The transmitting unit (1620) and the receiving unit (1630) are used to transmit and receive signals, messages, and data necessary to perform the aforementioned invention with the base station.

[0352]

[0353] FIG. 17 is a drawing for explaining the configuration of a base station according to one embodiment.

[0354] Referring to FIG. 17, a base station (1700) controlling the performance monitoring operation of an AI / ML model of a terminal may include a transmitter (1720) that transmits configuration information for performance monitoring of an AI / ML model, including mapping information of a prediction reference signal set and a monitoring reference signal set, to the terminal and transmits an indicated reference signal to the terminal, and a receiver (1730) that receives performance monitoring result information for an AI / ML model generated using inference result information for a prediction reference signal set and quality measurement result information for a monitoring reference signal set from the terminal.

[0355] The control unit (1710) determines the necessary details for performing performance monitoring of the AI / ML model of the terminal and generates configuration information accordingly. The transmission unit (1720) transmits the configuration information generated by the control unit (1710) to the terminal.

[0356] The configuration information includes mapping information between the Prediction Reference Signal Set and the Monitoring Reference Signal Set. The transmitter (1720) can transmit the configuration information through Higher Layer Signaling, such as an RRC message, and specifically, it may be in the form of CSI Report Configuration Information (CSI-ReportConfig).

[0357] The control unit (1710) allocates time and frequency radio resources for the transmission of a monitoring reference signal (e.g., CSI-RS, SSB, etc.) that the terminal is to measure for performance monitoring, configures this into radio resource information, and notifies the terminal through the transmitter (1720). In addition, the control unit (1710) sets up report information including uplink resources (PUCCH, PUSCH, etc.), reporting time (Timing), offset, etc., for receiving performance monitoring result information generated by the terminal, and includes this information in the configuration information.

[0358] In particular, the control unit (1710) sets mapping information that defines the relationship between the prediction reference signal set, which is the inference target of the AI / ML model, and the monitoring reference signal set, which is the actual measurement target. This mapping information can be configured so that the monitoring reference signal set is identical to the prediction reference signal set or is a sub-set of the prediction reference signal set.

[0359] In one embodiment, the control unit (1710) may configure mapping information using a bitmap format. The size of the bitmap is determined according to the number of prediction reference signal sets. The control unit (1710) sets each bit value such that the x-th most significant bit (MSB) of the bitmap corresponds to the x-th resource of the prediction reference signal set, and the y-th non-zero bit corresponds to the y-th resource entry of the monitoring reference signal set. Through this, the base station (1700) can clearly instruct the terminal exactly which beam or resource among the entire prediction target should be measured and compared.

[0360] The transmitter (1720), under the control of the control unit (1710), transmits a monitoring reference signal indicated by the configuration information above to the wireless section. For example, the transmitter (1720) transmits CSI-RS or SSB over a set time and frequency resource, and may perform beam sweeping or repetitive transmission as needed. This is to enable the terminal to receive the signal and measure quality indicators such as L1-RSRP and SINR.

[0361] The receiver (1730) receives performance monitoring result information for an AI / ML model from the terminal. This performance monitoring result information is information generated by the terminal using inference result information for a prediction reference signal set and quality measurement result information for a monitoring reference signal set.

[0362] Specifically, the result information includes the result of a comparison between the top K reference signals predicted by an AI / ML model and the top M monitoring reference signals actually measured. For example, the receiver (1730) receives information regarding whether at least one of the top M beams actually measured is included in the predicted top K beams (hit status), the number of matches, or the ratio.

[0363] The control unit (1710) analyzes the performance monitoring result information received through the receiver (1730) to determine the current performance status of the AI / ML model on the terminal side. If the result information indicates performance degradation of the model, the control unit (1710) may generate additional control commands to instruct the model to be retrained or to switch the monitoring mode from Normal Mode to Advanced Mode. Alternatively, the control unit (1710) may decide to stop AI-based beam management and fallback to a legacy beam management procedure.

[0364] In addition to this, the control unit (1710) controls the overall operation of the base station (1700) according to the management operation of AI / ML model information necessary to perform the aforementioned invention.

[0365] The transmitting unit (1720) and the receiving unit (1730) are used to transmit and receive signals, messages, and data necessary to perform the aforementioned invention with the terminal.

[0366] Meanwhile, the above-described disclosure may be implemented by a control unit, a transmitter, and a receiver of a terminal and a base station. The control unit, the transmitter, and the receiver may each be composed of a memory and a processor, and may include an antenna as needed.

[0367] The aforementioned embodiments may be supported by standard documents disclosed in at least one of the wireless access systems IEEE 802, 3GPP, and 3GPP2. That is, steps, configurations, and parts in the embodiments that are not described to clearly reveal the technical concept may be supported by the aforementioned standard documents. Furthermore, all terms disclosed in this specification may be explained by the standard documents disclosed above.

[0368] The embodiments described above may be implemented through various means. For example, the embodiments may be implemented by hardware, firmware, software, or a combination thereof.

[0369] In the case of implementation by hardware, the method according to the embodiments may be implemented by one or more ASICs (Application Specific Integrated Circuits), DSPs (Digital Signal Processors), DSPDs (Digital Signal Processing Devices), PLDs (Programmable Logic Devices), FPGAs (Field Programmable Gate Arrays), processors, controllers, microcontrollers, or microprocessors.

[0370] In the case of implementation by firmware or software, the method according to the embodiments may be implemented in the form of a device, procedure, or function that performs the functions or operations described above. Software code may be stored in a memory unit and executed by a processor. The memory unit may be located inside or outside the processor and may exchange data with the processor by various means already known.

[0371] Additionally, terms such as "system," "processor," "controller," "component," "module," "interface," "model," or "unit" described above may generally refer to computer-related entities, hardware, combinations of hardware and software, software, or running software. For example, the aforementioned components may be, but are not limited to, processes driven by a processor, processors, controllers, control processors, objects, execution threads, programs, and / or computers. For example, both the application running on the controller or processor and the controller or processor may be components. One or more components may reside within a process and / or execution thread, and the components may be located on a single device (e.g., a system, a computing device, etc.) or distributed across two or more devices.

[0372] The foregoing description is merely an illustrative explanation of the technical concept of the present disclosure, and those skilled in the art to which the present disclosure pertains may make various modifications and variations within the scope of the essential characteristics of the technical concept. Furthermore, since these embodiments are intended to explain, not limit, the scope of the technical concept is not limited by these embodiments. The scope of protection of the present disclosure shall be interpreted by the claims below, and all technical concepts within an equivalent scope shall be interpreted as being included within the scope of rights of the present disclosure.

[0373]

[0374] CROSS-REFERENCE TO RELATED APPLICATION

[0375] This patent application claims priority pursuant to Section 119(a) of the U.S. Patent Act (35 USC §119(a)) to Patent Application No. 10-2025-0009767 filed in Korea on January 22, 2025 and Patent Application No. 10-2026-0011911 filed in Korea on January 21, 2026, all of which are incorporated by reference into this patent application. Additionally, this patent application claims priority in countries other than the United States for the same reasons as above, all of which are incorporated by reference into this patent application.

Claims

1. A method for a terminal to perform a performance monitoring operation of an AI / ML model, A step of receiving configuration information for performance monitoring of an AI / ML model, including mapping information of a prediction reference signal set and a monitoring reference signal set, from a base station; A step of verifying the monitoring reference signal set indicated in conjunction with the prediction reference signal set output using the AI / ML model based on the above configuration information; and A method comprising the step of transmitting performance monitoring result information for the AI / ML model using inference result information for the prediction reference signal set and quality measurement result information for the monitoring reference signal set.

2. In Paragraph 1, The above configuration information is, A method further comprising wireless resource information for the above monitoring reference signal and reporting information for the transmission of the above performance monitoring result information.

3. In Paragraph 1, The above monitoring reference signal set is, A method composed of the same as the above prediction reference signal set or a subset of the above prediction reference signal set.

4. In Paragraph 1, The mapping information between the above prediction reference signal set and the monitoring reference signal set is, A method comprising a bitmap for indicating a reference signal included in the above monitoring reference signal set based on the above prediction reference signal set.

5. In Paragraph 4, The mapping information between the above prediction reference signal set and the monitoring reference signal set is, A method in which the size of the bitmap is determined according to the number of the above-mentioned prediction reference signal set.

6. In Paragraph 4, The x-th Most Significant Bit of the above bitmap corresponds to the x-th resource of the above prediction reference signal set, and A method in which the y-th non-zero bit of the bitmap is configured to correspond to the y-th resource entry of the monitoring reference signal set.

7. In Paragraph 1, The above inference result information is, A method comprising the top K (K is a natural number greater than or equal to 1) reference signal information among the prediction reference signal set based on the prediction quality inferred through the above AI / ML model.

8. In Paragraph 7, The above quality measurement result information is, A method including top M (M is a natural number greater than or equal to 1) monitoring reference signal information based on the quality measurement results of the above monitoring reference signal set.

9. In Paragraph 8, The above performance monitoring result information is, A method generated based on whether at least one of the top M monitoring reference signal information is included in the top K reference signal information.

10. A method for a base station to control the performance monitoring operation of an AI / ML model of a terminal, A step of transmitting configuration information for performance monitoring of an AI / ML model, including mapping information between a prediction reference signal set and a monitoring reference signal set, to a terminal; A step of transmitting a monitoring reference signal directed according to the above configuration information to the terminal; and A method comprising the step of receiving performance monitoring result information for the AI / ML model generated using inference result information for the prediction reference signal set and quality measurement result information for the monitoring reference signal set from the terminal.

11. In Paragraph 10, The above configuration information is, A method further comprising wireless resource information for the above monitoring reference signal and reporting information for the transmission of the above performance monitoring result information.

12. In Paragraph 10, The above monitoring reference signal set is, A method composed of the same as the above prediction reference signal set or a subset of the above prediction reference signal set.

13. In Paragraph 10, The mapping information between the above prediction reference signal set and the monitoring reference signal set is, A method comprising a bitmap for indicating a reference signal included in the above monitoring reference signal set based on the above prediction reference signal set.

14. In Paragraph 13, The mapping information between the above prediction reference signal set and the monitoring reference signal set is, A method in which the size of the bitmap is determined according to the number of the above-mentioned prediction reference signal set.

15. In Paragraph 13, The x-th Most Significant Bit of the above bitmap corresponds to the x-th resource of the above prediction reference signal set, and A method in which the y-th non-zero bit of the bitmap is configured to correspond to the y-th resource entry of the monitoring reference signal set.

16. In a terminal that performs performance monitoring of an AI / ML model, A receiver that receives configuration information for performance monitoring of an AI / ML model, including mapping information of a prediction reference signal set and a monitoring reference signal set, from a base station; A control unit that checks the monitoring reference signal set indicated in conjunction with the prediction reference signal set output using the AI / ML model based on the above configuration information; and A terminal comprising a transmitter that transmits performance monitoring result information for the AI / ML model using inference result information for the prediction reference signal set and quality measurement result information for the monitoring reference signal set.

17. In Paragraph 16, The above monitoring reference signal set is, A terminal configured to be identical to the above-mentioned prediction reference signal set or a subset of the above-mentioned prediction reference signal set.

18. In Paragraph 16, The mapping information between the above prediction reference signal set and the monitoring reference signal set is, A terminal configured as a bitmap for indicating a reference signal included in the above monitoring reference signal set based on the above prediction reference signal set.

19. In Paragraph 18, The mapping information between the above prediction reference signal set and the monitoring reference signal set is, A terminal in which the size of the bitmap is determined according to the number of the above-mentioned prediction reference signal set.

20. In Paragraph 18, The x-th Most Significant Bit of the above bitmap corresponds to the x-th resource of the above prediction reference signal set, and A terminal configured such that the y-th non-zero bit of the bitmap corresponds to the y-th resource entry of the monitoring reference signal set.

Images