Method and Apparatus for training a high compression ratio autoencoder model in a wireless communication system

Abstract

The present disclosure relates to a 5G or 6G communication system for supporting higher data transmission rates and satisfying various service requirements.

Description

Technology Field

[0001] The present disclosure relates to a wireless communication system, and more specifically to a method for training a high compression rate autoencoder model in a wireless communication system. Background Technology

[0002] 5G mobile communication technology defines wide frequency bands to enable fast transmission speeds and new services, and can be implemented not only in sub-6GHz bands such as 3.5 GHz ('Sub 6GHz') but also in ultra-high frequency bands known as millimeter wave (mmWave), such as 28GHz and 39GHz ('Above 6GHz'). Furthermore, for 6G mobile communication technology, which is referred to as a system beyond 5G, it is expected to become crucial to secure new frequency resources—including not only the 5G Sub 6GHz and ultra-high frequency bands but also the mid-frequency band (7-24 GHz), known as the Upper Mid band—and to efficiently utilize all available frequency resources as needed. This is necessary to handle the surge in data traffic driven by the spread of AI (artificial intelligence) technology and the increase in streaming services, and to improve user experience. To this end, reallocation, reuse, or sharing of existing frequency bands from 2G to 5G for 6G may also be considered. Separately, since the introduction of 5G, the telecommunications market has shown increasing interest in system operation efficiency, sustainability, and user experience improvement. Accordingly, not only are improvements in traditional communication performance, such as data transmission speed and latency, becoming increasingly important, but the introduction of new innovative technologies such as AI, reduction of operating costs, improvement of energy efficiency, expansion of service coverage, and introduction of new services are also becoming more critical.

[0003] In the early stages of 5G mobile communication technology, aiming to satisfy service support and performance requirements for enhanced mobile broadband (eMBB), ultra-reliable low-latency communications (URLLC), and massive machine-type communications (mMTC), technologies included beamforming and massive MIMO to mitigate path loss and increase propagation distance in ultra-high frequency bands; support for various numerologies (such as operating multiple subcarrier spacings) and dynamic operation of slot formats for the efficient utilization of ultra-high frequency resources; initial access technologies to support multi-beam transmission and broadband; definition and operation of band-width parts (BWPs); new channel coding methods such as low-density parity check (LDPC) codes for high-volume data transmission and polar codes for reliable transmission of control information; L2 pre-processing; and networks providing dedicated networks specialized for specific services. Standardization of network slicing and the like has been carried out.

[0004] Since the early days of 5G mobile communication technology, discussions have been held regarding the improvement and enhancement of the initial technology, taking into account the services that 5G technology was intended to support. These include V2X (vehicle-to-everything), which assists autonomous vehicles in making driving decisions and enhances user convenience based on location and status information transmitted by vehicles; NR-U (new radio unlicensed), which aims for system operation compliant with various regulatory requirements in unlicensed bands; UE power saving technology for NR terminals; non-terrestrial network (NTN), a direct terminal-satellite communication for securing coverage in areas where communication with terrestrial networks is impossible; positioning; supporting NR operation up to 71 GHz; support of reduced capability NR devices for lower cost and complexity compared to general terminals; UE power saving enhancement for improved power management in preparation for the utilization of various terminal types; and sidelink enhancement. Evolution of redundancy technology (duplex enhancements) researching a new form of duplex method called subband non-overlapping full duplex (SBFD), network energy saving that maximizes idle periods during which base stations operate in maximum power saving mode and reduces power consumption,Physical layer standardization was conducted for technologies such as network-controlled repeaters, which have improved performance compared to existing repeaters by being equipped with the function of receiving and processing side control information from the network.

[0005] Furthermore, the standardization of Intelligent Factories (Industrial Internet of Things, IIoT) to support new services through linkage and convergence with other industries; Integrated Access and Backhaul (IAB), which provides nodes to expand network service areas by integrating wireless backhaul and access links; Mobility Enhancement technologies including Conditional Handover and Dual Active Protocol Stack (DAPS) Handover; 2-step Random Access for Non-Release (RACH for NR) to simplify random access procedures; multicast and broadcast; support for multi-USIM devices that provide services to users using information from two or more SIMs; sidelink relay, which provides relay-related functions to support connections between terminals at long distances and between terminals and networks; Small Data Transmission (SDT in an inactive state), which transmits small data or signaling in an inactive state without transitioning to a connection state; and Layer 1 / L2 Triggered Mobility (LTM) / Continuous Conditional Standardization of the wireless interface architecture / protocol layer for technologies such as mobility enhancements including subsequent conditional PSCell addition / change (SCAC) and conditional handover with candidate SCGs (CHO with candidate SCGs), and XR enhancements to support XR services in NR systems has also been carried out,Systems regarding 5G baseline architectures for integrating Network Functions Virtualization (NFV) and Software-Defined Networking (SDN) technologies (e.g., service-based architecture, service-based interface); Mobile Edge Computing (MEC), which provides services based on the terminal's location; Non-Public Networks (NPN), which are accessible only to authorized terminals for non-public purposes; Disaster Roaming, which supports the use of communication services through other carrier networks in the event of a telecommunications disaster; Proximity-Based Service via 5GS; Support for Uncrewed Aerial Vehicles (UAVs) to assist with remote identification, tracking, and authorization; Architecture enhancements to support XR and interactive media services; 5GS for supporting AI / ML (artificial intelligence / machine learning) services; and Advanced Mobile Edge Computing, which provides edge computing services on roaming networks. Standardization in the architecture / service fields also proceeded.

[0006] Currently, at the physical layer, standardization is underway for technologies such as beam prediction using AI / ML technology, CSI (channel state information) prediction for improved positioning accuracy, ultra-low power terminal technology using low-power wake-up receivers, technology for transmitting LTE (long term evolution) broadcasts to 5G networks, MIMO transmission technology using multiple base stations, and ambient IoT, which transmits data by acquiring power from an external source without a battery. At the wireless interface architecture / protocol layer, standardization is underway for supporting LTM scenarios between Central Units (CUs) and conditional LTM, supporting the same XR service simultaneously among multiple devices, improving and evolving NTN coverage, supporting mobility based on AI / ML, and relaying connections between terminals across multiple hops between terminals and networks. In addition, standardization is underway in the fields of system architecture and services regarding communication optimization methods for satellites, energy usage management and efficiency of 5G systems, user plane evolution based on SBI, Ambient IoT technology, data service provision methods in IMS (IP multimedia subsystem), and avatar communication services. Once such 5G mobile communication systems are commercialized, connected devices, which are increasing explosively, will be connected to communication networks; accordingly, it is expected that there will be a need to enhance the functions and performance of 5G mobile communication systems and to integrate the operation of connected devices.To this end, additional new research is planned to be conducted on eXtended Reality (XR) to efficiently support augmented reality (AR), virtual reality (VR), mixed reality (MR), 5G performance improvement and complexity reduction using AI / ML, support for AI services, support for metaverse services, and drone communication.

[0007] Furthermore, the advancement of these 5G mobile communication systems is expected to enhance 5G performance while serving as the foundation for the ultimate evolution into 6G. In the 6G era, the three major 5G services mentioned earlier—eMBB, URLLC, and mMTC—are expected to evolve and expand into immersive communication (IC), hyper-reliable and low-latency communication (HRLLC), and massive communication (MC), respectively. In addition, new services such as AI-integrated communication, integrated sensing and communication, and ubiquitous connectivity are planned to be additionally supported. Enhanced performance requirements compared to 5G are essential for these diverse 6G services, and standardization to define these requirements is currently underway.

[0008] As such, in order to satisfy the expanded services and enhanced performance requirements of 6G, it is expected that not only will existing communication performance be improved, but system operations will also be optimized and streamlined through the introduction of AI technology, improved energy efficiency, expanded coverage, and the application of next-generation security technology, and the development of sustainable communication technology will be essential.

[0009] To this end, AI internalization technology that applies the latest AI technology across all areas from the communication system design stage to development, management, and operation to realize improved communication performance and network automation and efficiency; technology to improve user perceived performance and network operational efficiency by reducing power consumption of networks and terminals; technology to reduce power consumption in core base station components such as RF (radio frequency) and modems, as well as in channel coding and signal modulation / demodulation transmission and reception processes; multi-antenna transmission technology (eXtreme MIMO, X-MIMO) utilizing large-scale antennas to overcome propagation path loss caused by frequencies higher than the 3.5GHz band of 5G communication and provide equivalent coverage; multi-base station-based transmission and reception technology (distributed MIMO, D-MIMO) to improve quality in cell boundary areas; sub-band non-overlapping full duplex (SBFD) technology for enhancing frequency efficiency and system networks; next-generation encryption technology (post-quantum cryptography, PQC) and zero trust architecture (ZTA) technology to strengthen 6G communication security; and initial connection delay and mobility delay Intensive research is planned to be conducted on minimization techniques, the design of hardware-friendly protocol structures for ultra-high-speed data processing, and the expanded application of integrity protection techniques.

[0010] In addition, research is planned on the structure of mobile communication systems (prevention of redundant functions, function simplification, etc.), the introduction of new planes for operator service provision, measures for protecting user privacy, immersive services, enhancement of network resiliency, network sharing technology, enhanced security technology (false base station, lower layer protection, etc.), and intent-based network operation and management.

[0011] To satisfy various services, it is often necessary to support beam management or various frequency bands. In such situations, diverse channel environments may occur depending on the frequency band / beam, and the resource consumption for performing Channel Estimation at the terminal and reporting it to the base station is high. To address this, discussions began at 3GPP Rel-18 on an item called CSI Compression, which proposes compressing Channel State Information (CSI) before sending it, and related papers are being published.

[0012] Channel states recovered through CSI reports based on the existing 5G NR Codebook method suffer from information loss during the process of transmitting Estimated Channel H due to codebook quantization issues. While codebooks can be configured more diversely, the amount of data that needs to be transmitted increases accordingly. For this reason, a CSI Compression Study item was discussed in Release-18, which transmits the channel state that the terminal wants to transmit by compressing it using an Auto-Encoder (AE) based method rather than the current 5G NR codebook-based channel feedback, and allowing the base station to properly decode the channel state. Currently, a Work item is being discussed in Release-19. The problem to be solved

[0013] The purpose of the present disclosure is to propose a method for enabling the training of a high compression rate autoencoder model by applying a training method improved compared to the conventional method, while simultaneously enabling the reduction of OTA overhead during training.

[0014] The technical problems to be solved by the present disclosure are not limited to those mentioned above, and other unmentioned technical problems may be considered by those skilled in the art from the various embodiments of the present disclosure described below. means of solving the problem

[0015] The present disclosure for solving the above-mentioned problems comprises, in a method performed at a terminal in a wireless communication system, the steps of: receiving a training setting from a base station; receiving a training start from the base station based on the training setting; transmitting forward propagation to the base station; receiving back propagation from the base station; and receiving an unfreezing command from the base station, wherein the training setting includes information about a training model and an indicator indicating whether to freeze.

[0016] According to one embodiment, the method further comprises the steps of: transmitting a second forward propagation to the base station; receiving a second back propagation from the base station; receiving the training termination from the base station; and receiving the model identifier (ID) of the terminal from the base station.

[0017] According to one embodiment, the method further includes a step of determining whether to perform a weight update for a layer where training is performed based on an indicator indicating whether to freeze.

[0018] According to one embodiment, the method further includes the step of transmitting a model identifier (identifier, ID) of the base station to the base station.

[0019] In addition, in another embodiment of the present disclosure, a method performed at a base station in a wireless communication system comprises: transmitting a training setting to a terminal; transmitting a training start to the terminal based on the training setting; receiving forward propagation from the terminal; transmitting back propagation to the terminal; and transmitting an unfreezing command to the terminal, wherein the training setting includes information about a training model and an indicator indicating whether to freeze.

[0020] In addition, in another embodiment of the present disclosure, a terminal comprises: a transceiver capable of transmitting and receiving at least one signal; and a processor coupled to the transceiver, configured to receive a training setting from a base station, receive a training start based on the training setting from the base station, transmit forward propagation to the base station, receive back propagation from the base station, and receive an unfreezing command from the base station, wherein the training setting includes information about a training model and an indicator indicating whether to freeze.

[0021] In addition, in another embodiment of the present disclosure, a base station comprises: a transceiver capable of transmitting and receiving at least one signal; and a processor coupled to the transceiver, wherein the processor is configured to: transmit a training setting to a terminal, transmit a training start to the terminal based on the training setting, receive forward propagation from the terminal, transmit back propagation to the terminal, and transmit an unfreezing command to the terminal, wherein the training setting includes information about a training model and an indicator indicating whether to freeze. Effects of the invention

[0022] According to one embodiment of the present disclosure, compared to a conventional Type 2 training method, the training epoch can be reduced by utilizing model information of a previously trained NW to show an effect of reducing OTA overhead, and the training stability and performance enhancement of a high compression rate autoencoder can be shown.

[0023] The effects obtainable from the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by those skilled in the art to which the present disclosure belongs from the description below. Brief explanation of the drawing

[0024] FIG. 1 is a diagram illustrating CSI feedback of existing 5G NR related to the present disclosure. FIG. 2 is a diagram illustrating an AI / ML-based CSI compression technology related to the present disclosure. FIG. 3 is a diagram illustrating the CSI-RS transmission and AI / ML-based CSI compression procedure related to the present disclosure. FIG. 4 is a diagram illustrating a type 2 training method of AI CSI compression related to the present disclosure. FIG. 5 is a diagram illustrating a type 2 training procedure for AI CSI compression related to the present disclosure. FIG. 6 is a drawing illustrating a training method according to one embodiment of the present disclosure. FIG. 7 is a diagram illustrating a training sequence according to one embodiment of the present disclosure. FIGS. 8A and FIGS. 8B are drawings illustrating examples of a model structure according to one embodiment of the present disclosure. FIG. 9 is a drawing illustrating another training method according to one embodiment of the present disclosure. FIG. 10 is a drawing illustrating another training method according to one embodiment of the present disclosure. FIG. 11 is a drawing illustrating a procedure according to one embodiment of the present disclosure. FIG. 12 is a drawing illustrating a procedure according to another embodiment of the present disclosure. FIG. 13 is a drawing illustrating a procedure according to another embodiment of the present disclosure. FIG. 14 is a drawing illustrating a procedure according to another embodiment of the present disclosure. FIG. 15 is a drawing illustrating the operation of a terminal according to one embodiment of the present disclosure. FIG. 16 is a drawing illustrating the operation of a network according to one embodiment of the present disclosure. FIG. 17 is a block diagram of a terminal or user equipment according to one embodiment of the present disclosure. FIG. 18 is a block diagram of a base station according to one embodiment of the present disclosure. FIG. 19 is a block diagram of a network entity performing a network function according to one embodiment of the present disclosure. Specific details for implementing the invention

[0025] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0026] In describing the embodiments, technical details that are well known in the art to which this disclosure belongs and are not directly related to this disclosure are omitted. This is intended to convey the essence of this disclosure more clearly without obscuring it by omitting unnecessary explanations.

[0027] For the same reason, some components in the attached drawings have been exaggerated, omitted, or schematically depicted. Additionally, the size of each component does not entirely reflect its actual dimensions. Identical or corresponding components in each drawing have been assigned the same or different reference numbers.

[0028] The advantages and features of the present disclosure, and the methods for achieving them, will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure is complete and to fully inform those skilled in the art of the scope of the disclosure, and the present disclosure is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components. Furthermore, in describing the present disclosure, if it is determined that a detailed description of a related function or configuration might unnecessarily obscure the essence of the present disclosure, such detailed description is omitted. Additionally, the terms described below are defined considering their functions in the present disclosure, and these may vary depending on the intentions or conventions of the user or operator. Therefore, their definitions should be based on the content throughout the specification.

[0029] In the present disclosure, it will be understood that each block of the process flow diagrams and combinations of the flow diagrams may be performed based on computer program instructions. Since these computer program instructions may be optionally loaded into at least one processor of a general-purpose computer, a computer for special purposes, or other programmable data processing equipment, the instructions performed through any one or any combination of at least one processor of the computer or other programmable data processing equipment create means for performing the functions described in the flow diagram block(s). Since these computer program instructions may also be stored in computer-available or computer-readable memory that can be directed toward the computer or other programmable data processing equipment to implement the functions in a specific manner, the instructions stored in computer-available or computer-readable memory may also produce a manufactured item containing means of instruction for performing the functions described in the flow diagram block(s). Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that perform a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in the flowchart block(s).

[0030] Additionally, each block may represent a module, segment, or part of code containing one or more executable instructions for executing a specified logical function(s). It should also be noted that in some alternative execution examples, the functions mentioned in the blocks may occur out of order. For example, two blocks (or functions) described in succession may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order according to the corresponding function.

[0031] As used in the embodiments of the present disclosure, the term “part” refers to a software or hardware component, such as a field programmable gate array (FPGA) or an application-specific integrated circuit (ASIC), and the “part” performs certain roles. However, the term including “part” is not limited to software or hardware. The “part” may be configured to reside in an addressable storage medium or may be configured to run on one or more processors. Thus, by example, the “part” includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and “parts” may be combined into a smaller number of components and “parts” or further separated into additional components and “parts.” In addition, the components and 'parts' may be implemented to utilize one or more CPUs (central processing units) within the device or secure multimedia card. Also, in the embodiments, 'parts' may include one or more processors.

[0032] As stated above, it should be noted that the blocks of each flowchart and combinations of flowcharts described in this disclosure may be executed by one or more computer programs including instructions. The entirety of one or more computer programs may be stored in a single memory device, or one or more computer programs may be divided into different parts and stored across multiple memory devices.

[0033] Additionally, any / any function or operation described in this disclosure may be processed by a single processor or a combination of processors. The single processor or combination of processors is a circuitry that performs processing and may include an application processor (AP, e.g., a central processing unit (CPU)), a communication processor (CP, e.g., a modem), a graphics processing unit (GPU), a neural network processing unit (NPU) (e.g., an artificial intelligence (AI) chip), a Wi-Fi chip, a Bluetooth® chip, a global positioning system (GPS) chip, a near-field communication (NFC) chip, a connectivity chip, a sensor controller, a touch controller, a fingerprint sensor controller, a display driver integrated circuit (IC), an audio codec (CODEC) chip, a universal serial bus (USB) controller, a camera controller, an image processing IC, a microprocessor unit (MPU), a system-on-chip (SoC), an IC, or similar circuitry.

[0034] Additionally, it should be noted that various embodiments in the claims and description of the present disclosure may be implemented in the form of hardware, software, or a combination of hardware and software.

[0035] Such software may be stored on a non-transitory computer-readable storage medium. A non-transitory computer-readable storage medium stores one or more computer programs (software modules), and said one or more computer programs include computer-executable instructions that operate an electronic device to perform a method according to the present disclosure when executed alone or collectively by one or more processors of an electronic device.

[0036] The software may be stored in a transient or non-transient storage device, for example, in the form of read-only memory (ROM) (whether or not it is erasable or rewritable), or random access memory (RAM), memory chips, devices, or integrated circuits (ICs). Additionally, the software may be stored in the form of an optically or magnetically readable medium, for example, a compact disc (CD), a digital multifunction disc (DVD), a magnetic disc, or a magnetic tape. It should be understood that the storage device and the storage medium are examples of non-transient machine-readable storage media suitable for storing programs for implementing various embodiments of the present disclosure. Accordingly, various embodiments of the present disclosure may provide a program comprising code for implementing a device or method according to any one of the claims of this specification, and a non-transient machine-readable storage medium storing such program.

[0037] In the following disclosure, determining the priority between A and B may be referred to in various ways, such as selecting the one with the higher priority according to a predetermined priority rule and performing the corresponding action, or omitting or dropping the action for the one with the lower priority.

[0038] Hereinafter, 'A or B' as described in the present disclosure may be understood as 'A and / or B', which may be understood as including 'A', or 'B', or 'A and B'.

[0039] Additionally, 'at least one of A, B, and C' described in the present disclosure may be understood to include 'A', or 'B', or 'C', or 'any combination of A, B, and C'.

[0040] Additionally, 'at least one of A, B, or C' described in the present disclosure may be understood to include 'A', or 'B', or 'C', or 'any combination of A, B, and C'.

[0041] Additionally, 'A / B' as described in the present disclosure may be understood as 'A and / or B', which may be understood as including 'A', or 'B', or 'A and B'.

[0042] Additionally, 'A, B' described in the present disclosure may be understood as 'A and / or B', which may be understood as including 'A', or 'B', or 'A and B'.

[0043] Additionally, 'A and B' described in the present disclosure may be understood as 'A and / or B', which may be understood as including 'A', or 'B', or 'A and B'.

[0044] Furthermore, the phrase "when conditions A and B are satisfied" as described in the present disclosure is not necessarily limited to cases where both conditions A and B are satisfied, but may be understood to include cases where either condition A or condition B is satisfied individually, cases where both conditions A and B are satisfied, or cases where one or more additional conditions are satisfied together.

[0045] Furthermore, throughout this specification, ordinal terms (and similar modifiers) such as 'first', 'second', 'third', etc. are used solely for the purpose of distinguishing various instances, occurrences, configurations, messages, stages, or aspects of elements, operations, or information, as described below. Unless clearly required otherwise by the context, the use of such ordinal terms does not require that the elements, operations, or information distinguished by such terms be structurally different, numerically distinct, or essentially different. For example, 'first signal' and 'second signal' may represent instances of the same signal transmitted at different times, signals containing the same core information even with some variations, or signals having different content or characteristics depending on the specific context. Similarly, 'first value' and 'second value' may represent the same magnitude measured or applied in different situations, or may represent different magnitudes. Such interpretation must be determined based on the specific technical context, function, and relationship described in the relevant parts of the specification and claims.

[0046] Furthermore, although terms such as "first," "second," etc., as used in this disclosure are used for various elements such as information, objects, actions, and sequences, they are not intended to limit such elements to a specific order. These terms may be understood merely as distinguishing one element from another. For example, a first element may be referred to as a second element, and likewise, a second element may be referred to as a first element.

[0047] Additionally, the terms 'first' and 'second' described in this disclosure may be understood to refer to identical or different elements. For example, if an element is information, the first information and the second information may both be information, and depending on the case, they may be the same information or different information.

[0048] Furthermore, the expressions 'if' and 'in case that' described in this disclosure or claims may be interpreted, depending on the context, as meaning 'when or upon,' 'in response to,' 'based on,' or 'according to,' and these expressions may be used interchangeably. In addition, other expressions having substantially the same meaning may be used as substitutes, provided that they do not impair the technical features of this disclosure.

[0049] Additionally, the term "not perform" as used in this disclosure or claims may be understood, depending on the context, to mean to omit or skip the corresponding step. Such a term may be replaced with other terms having the same or substantially similar meaning.

[0050] Additionally, the phrase "transmitting a message containing A and B" as described in this specification may be interpreted to include not only (i) cases where A and B are transmitted as a single message, but also (ii) cases where A and B are transmitted individually through multiple messages (e.g., transmitting a first message containing A and a second message containing B). This interpretation may also apply to cases where messages containing two or more items, such as A, B, and C, are transmitted together or individually.

[0051] In addition, 'transmitting a message containing A and transmitting a message containing B' can also be interpreted as transmitting a single message containing A and B.

[0052] In the specific embodiments of the present disclosure described below, terms or components included in the disclosure will be expressed in the singular or plural form according to the specific embodiments presented. However, the singular or plural expression is selected to suit the circumstances presented for convenience of explanation, and the present disclosure is not limited to singular or plural components; even if a component is expressed in the plural form, it may be composed in the singular form, and even if a component is expressed in the singular form, it may be composed in the plural form.

[0053] The drawings or flowcharts described below illustrate exemplary methods that may be implemented in accordance with the principles of the present disclosure, and various modifications may be made to the methods illustrated in the flowcharts of the present disclosure. For example, although illustrated as a series of steps, the various steps of each drawing or flowchart may overlap, occur in parallel, occur in a different order, or occur multiple times. In other examples, any step may be omitted or replaced with another step.

[0054] The methods and devices proposed in the embodiments of the present disclosure below are not limited to each embodiment and may be utilized as a combination of all or part of the embodiments proposed in the disclosure. Accordingly, the embodiments of the present disclosure may be applied with some modifications within the scope that does not deviate significantly from the scope of the present disclosure, at the judgment of a person skilled in the art.

[0055] In this case, any wording mentioned in different embodiments may be used interchangeably, combined, or substituted if the concepts correspond. For example, regarding the same or corresponding concepts, even if the expression 'A' is used in one embodiment and the expression 'B' is used in another embodiment, they may be understood by interchangeably, substituted, or combined.

[0056] Terms used in the following description to identify connection nodes, terms referring to network entities, terms referring to messages, terms referring to interfaces between network entities, terms referring to various identification information, etc., are examples provided for the convenience of explanation. Accordingly, the present disclosure is not limited to the terms described below, and other terms referring to objects having equivalent technical meanings may be used. Furthermore, where appropriate, such terms may be replaced with terms defined in the 3GPP (3rd generation partnership project) Technical Specifications (TS).

[0057] Hereinafter, the base station, as the entity performing resource allocation for terminals, may be at least one of gNode B, eNode B, Node B, BS (base station), wireless access unit, base station controller, or a node on a network. Additionally, the base station of the present disclosure may include a structure split into a central unit (CU) and a distributed unit (DU). In such a structure, the CU is responsible for the upper layer of the control and user plane, and the DU is responsible for wireless resource processing of the lower layer. The embodiments of the present disclosure can be equally applied to a 5G base station structure in which functions are separated into the CU and DU as described above.

[0058] The terminal may include UE (user equipment), MS (mobile station), cellular phone, smartphone, computer, or a multimedia system capable of performing communication functions.

[0059] In the present disclosure, a downlink (DL) refers to a wireless transmission path of a signal transmitted by a base station to a terminal, and an uplink (UL) refers to a wireless transmission path of a signal transmitted by a terminal to a base station.

[0060] In addition, while a 5th generation mobile communication system (5G, new radio, NR) and a 6th generation mobile communication system (6G) may be described below as examples, embodiments of the present disclosure may also be applied to other communication systems having similar technical backgrounds or channel types. For example, new advanced mobile communication systems developed after 5G and 6G may be included therein. Furthermore, the present disclosure may be applied to other communication systems (e.g., Wi-Fi systems) with some modifications made in the judgment of a person with skilled technical knowledge, without significantly departing from the scope of the present disclosure.

[0061] In the following description, the terms "physical channel" and "signal" may be used interchangeably with "data" or "control signal." For example, PDSCH (physical downlink shared channel) is a term referring to a physical channel through which data is transmitted, but PDSCH may also be used to refer to data. That is, in this disclosure, the expression "transmits a physical channel" may be interpreted as equivalent to the expression "transmits data or a signal through a physical channel."

[0062] In describing the present disclosure below, the term "upper layer signaling" may be a signaling corresponding to at least one or a combination of at least one of MIB (master information block), SIB (system information block), SIB M (M=1, 2, ...), RRC (radio resource control), MAC (medium access control), CE (control element), NAS (non-access stratum) signaling, or application layer messages. The RRC signaling may also be referred to as L3 signaling (layer 3 signaling).

[0063] Additionally, L1 signaling may be a signaling method corresponding to at least one or a combination of at least one of the following: a physical layer channel or signaling of a PDCCH (physical downlink control channel), a DCI (downlink control information), a UE-specific DCI, a group common DCI, a common DCI, a scheduling DCI (e.g., a DCI used for the purpose of scheduling downlink or uplink data), a non-scheduling DCI (e.g., a DCI not used for the purpose of scheduling downlink or uplink data), a PUCCH (physical uplink control channel), or an UCI (uplink control information). The above L1 signaling may also be referred to as physical layer signaling.

[0064] Hereinafter, the expression in the present disclosure or claims that information can be configured from a base station may mean that, depending on the context, a terminal receives said information from a base station through physical layer signaling or upper layer signaling, and such expression may be replaced with other terms having the same or substantially similar meaning.

[0065] The operating principle of the present disclosure will be explained in detail below with reference to the attached drawings.

[0067] FIG. 1 is a diagram illustrating CSI feedback of existing 5G NR related to the present disclosure.

[0068] In order for the base station to know the downlink channel estimated by the terminal, a separate feedback process is required for the terminal to transmit channel information to the base station. Specifically, when the base station transmits a CSI-reference signal (CSI-RS) to the terminal, the terminal performs channel estimation using the received CSI-RS, and the estimated channel It can obtain. The terminal transmits the precoding matrix indicator (PMI), which is channel information related to the precoding matrix to be used by the base station, in order to After performing eigenvalue decomposition (EVD) of, at It can be determined. The terminal has the estimated channel (e.g., With respect to ), channel information (e.g., V) for the precoding matrix can be obtained by performing pre-processing (e.g., eigenvalue decomposition (EVD) or singular value decomposition (SVD)). Additionally, if pre-processing is performed via singular value decomposition (SVD), H is It can be decomposed as shown, through which the terminal can obtain channel information (V) for the precoding matrix. Here, U is a left singular vector matrix that can represent the column space of H, and D is a singular value of H that can represent the energy of the channel, and can represent the row space of H. The terminal During transmission, based on the pre-defined codebook specified in 3GPP technical specification (TS) 38.214 One of the existing 5G NR CSI feedback methods is to transmit the index of the codebook most similar to it in the form of PMI.

[0069] Specifically, in step 101, the terminal encoder converts the CSI into a low-dimensional vector It can be compressed into a latent vector. The latent vector is bit sequence information corresponding to the PMI. In step 102, the terminal can transmit the explicit CSI to the network. In step 103, the network decoder the compressed vector CSI can be recovered from. In step 104, the network can determine the precoder. In step 105, the network can use the AI-resolved precoder.

[0071] FIG. 2 is a diagram illustrating an AI / ML-based CSI compression technology related to the present disclosure.

[0072] Specifically, to utilize Multi-user (MU) Multiple Input Multiple Output (MIMO), more accurate transmission of CSI information is required, and the Rel-16 Type 2 (eType 2) codebook, which compresses channels in the Spatial and Frequency domains, was introduced following 3GPP Type 1 and Type 2. Unlike the existing Type 2 codebook, which compresses channel information in the Spatial domain and provides feedback, the eType 2 codebook, which performs Frequency domain compression using channel sparsity in the Frequency domain, utilizes the reduced feedback bits for feedback of phase and amplitude information in more finely sub-band areas, thereby enabling more precise channel information feedback compared to the Type 2 codebook.

[0073] However, this codebook type Transmission is actually due to the limited granularity of the codebook (limited number of DFT beams, amplitude, phase information, etc.) Differences arise, which may lead to performance degradation in MIMO systems. Furthermore, as the performance improvement of MU-MIMO systems increases the influence of CSI feedback from each terminal, more accurate CSI feedback is required in future communication systems. To address this, existing CSI feedback systems can obtain accurate CSI by increasing the number of bits used in the CSI report; however, this results in a significant increase in overhead due to the increase in the number of base station antennas, bandwidth, and granularity. To resolve this, the 3GPP Rel-18 study item (SI) and Rel-19 work item (WI) utilize artificial intelligence (AI) / machine learning (ML) so that at the terminal Compressing itself into a low-dimensional vector It is transmitted to the base station in the form of a latent vector, and the base station receives the compressed vector from the original AI / ML-based CSI compression to restore is being discussed.

[0074] The boxed portion shown in Fig. 2 represents an AI / ML-based autoencoder. Fig. 2 depicts a configuration where the encoder of the autoencoder is located at the terminal and the decoder is located at the base station. After transmitting the latent vector (feature vector), which is the result compressed by the encoder, to the base station via the UL (Uplink), this is processed through the decoder It shows the process of restoring it.

[0076] FIG. 3 is a diagram illustrating the CSI-RS transmission and AI / ML-based CSI compression procedure related to the present disclosure.

[0077] Specifically, in step 301, the gNB (320) can transmit CSI-RS to the UE (300). Subsequently, in step 302, the UE (300) can measure encoder inference. The UE (300) performs channel estimation with the received CSI-RS, and the estimated channel ...can be obtained. The UE (300) transmits a precoding matrix indicator (PMI), which is channel information related to the precoding matrix to be used by the base station. After performing eigenvalue decomposition (EVD) of, at It can be seen that. Subsequently, in step 303, the UE (300) utilizes AI / ML with the gNB (320). Compressing itself into a low-dimensional vector Encoder inference output vector in the form of a latent vector It can be transmitted. Afterwards, the latent vector (feature vector), which is the result compressed through the encoder in step 304, can be restored through the decoder. The latent vector may represent bit sequence information corresponding to the PMI.

[0079] Conventional AI / ML-based CSI compression technology compresses channel information in the spatial and frequency domains and feeds it back to the base station. The terminal measures the channel based on CSI-RS reception information and extracts the eigenvector matrix required for use as a base station precoder. Subsequently, the eigenvectors for each frequency subband are compressed into latent vectors using the terminal's AI encoder model, and this information is fed back to the base station via a CSI report. Upon receiving the latent vectors within the CSI report, the base station reconstructs the channel information using its base station AI decoder model.

[0080] Conventional AI / ML-based CSI compression technology is currently being tested and discussed for 32 transmit antenna ports and a narrow bandwidth of approximately 10 MHz for FR1. However, in addition to FR1, actual frequency bands include FR2 and FR3, which are expected to be discussed for 6G; both frequency bands utilize or are planned to utilize carrier frequencies higher than those of FR1. As carrier frequencies increase, the spacing between antennas narrows, enabling the implementation of denser antenna spacing. Consequently, it becomes possible to implement more antenna elements and ports within the same area. To enhance the performance of Multiple Input Multiple Output (MIMO) methods, an increase in the number of antennas is essential; consequently, codebooks supporting up to 128 ports—a significant improvement from the conventional 32-port system—have been discussed in the 3GPP standard. With the introduction of FR3, expected to operate in the 7 GHz band or higher, being discussed for 6G, the operation of even more antenna ports is being debated, with discussions currently underway regarding the operation of approximately 256 ports. With the increased frequency band of FR3, operation can be expected with the number of antenna ports increased from 32 to 256 and the frequency band from 10 MHz to 100 MHz. This implies that the data dimensions in the spatial domain and frequency domain, respectively, have increased. Due to the increase in antenna ports, an approximately eightfold increase in the spatial domain is required, and due to the increase in bandwidth, an approximately 1.3-fold increase in the number of subbands in the frequency domain is required compared to the conventional system.

[0081] Network CSI acquisition using CSI-RS is possible in the form of a CSI report from the terminal via over-the-air (OTA). Among these, the precoding matrix indicator (PMI) has utilized Type I and Type II codebooks defined in conventional standards; during discussions regarding Type II codebook extensions for the REL-19 standard, an agreement was reached to use conventional standards for parameters other than N1 and N2, which are related to antenna ports. Although the input dimension of FR3 has increased by approximately 11 times compared to the conventional standard, the feedback size being considered in the standard has increased by only tens of bits; this indicates the need for a compression ratio several times higher, even when accounting for the increased feedback bits. (This can be interpreted as a measure to manage overhead via OTA.)

[0082] Therefore, in order to apply conventional AI CSI compression under FR3 256 port, model training at a high compression rate is required. However, when training an autoencoder consisting of an encoder and a decoder with a dataset of approximately 800,000, which is the conventional training method, through a single training loop, it showed a very unstable training state and was confirmed to show degraded performance compared to the baseline conventional enhanced Type II codebook.

[0084] FIG. 4 is a diagram illustrating a type 2 training method of AI CSI compression related to the present disclosure.

[0085] Among the training methods for AI CSI compression models in conventional communication systems, there is a Type 2 Joint training at NW and UE method. This training method is a method in which an encoder model of an autoencoder exists at the terminal and a decoder model of an autoencoder exists at the base station (NW) in a single training loop, and training is performed by transmitting and receiving information necessary for forward propagation and back propagation via OTA.

[0086] Using this method can demonstrate the most optimized performance compared to transmitting information during the training process after learning from a single entity. Additionally, it had the advantage of allowing each terminal and base station model to possess a model suitable for the hardware optimized for each entity. However, due to the nature of this training method, which transmits and receives information via OTA, the vast amount of data required during training necessitates the use of OTA resources, presenting a problem with excessive overhead.

[0088] FIG. 5 is a diagram illustrating a type 2 training procedure for AI CSI compression related to the present disclosure.

[0089] Specifically, in step 501, the network (520) may instruct the UE (500) to start training by sending a Training start message. In step 502, the UE (500) may send forward propagation to the network (520). Forward propagation is a procedure containing information that the actual artificial intelligence model must transmit from the encoder model to the decoder model to proceed with forward propagation. This refers to the process of transmitting the computed value of the encoder model via OTA to proceed with the entire forward propagation process after the terminal encoder model training begins according to step 501. In step 503, the network (520) may send a message to the UE (500) containing information that must be transmitted to the encoder model during the back propagation transmission process, including gradients, after the forward propagation ends. Step 503 above is a process for the network's decoder to complete the FP process using the information from Step 502, perform BP (Browser-Phase) to calculate the gradient and update the model, and then transmit this information to the UE. Upon receiving this message, the UE can use it to update the encoder model. During this learning process, the learning may proceed in a form where Steps 502 and 503 are repeated, for instance, until the network determines that the learning is finished. Subsequently, if the network (502) determines that the training needs to end, in Step 504, the network (520) sends a training end message to the UE (500), thereby terminating the autoencoder training using this Type 2 method.

[0091] The present disclosure identifies that learning instability and performance degradation occur when using such conventional AI CSI compression technology in an augmented dimension, such as FR3, and aims to solve the problem. To address this problem, the present invention proposes model training using an augmented dataset and the introduction of a new training method, and can define the procedures and signals necessary for the present method to be operated in an actual communication system.

[0092] In addition, the present disclosure is intended to reduce the OTA overhead that occurs when using a Type 2 training method for conventional AI CSI compression learning. By using the present disclosure, training can be completed with fewer iterations, and as a result, the OTA overhead is reduced compared to the conventional method, and the present disclosure defines procedures and signals for this purpose.

[0093] The simulation results in Table 1 are the results obtained using the training method proposed in this disclosure, and it can be confirmed through the results that it is more stable and shows higher performance compared to the prior art.

[0094] Table 1. Simulation results

[0095]

[0096] Accordingly, the present disclosure enables the training of a high compression rate autoencoder model by applying an improved training method compared to the prior art, while simultaneously enabling the reduction of OTA overhead during training.

[0098] The present disclosure defines the signal between the NW and the UE in the invention to reduce OTA overhead during Type 2 training and to ensure stable and high performance during high compression rate autoencoder training by using a method of training with some layers of the model frozen and unfrozen through the sharing of model parameters learned in the UE in the network.

[0099] The proposed embodiment consists of embodiments (A) to (D).

[0100] (A) Use freeze and unfreeze sequentially upon initial connection,

[0101] (B) Use unfreeze or freeze only upon initial connection,

[0102] (C) Use freeze and unfreeze sequentially when updating the model,

[0103] (D) Use unfreeze during model update or freeze only during initial connection,

[0104] Before describing the embodiments of the present disclosure, examples of terms and training methods commonly used in the disclosure are explained and illustrated.

[0105] In the present disclosure, "freeze" means not performing weight updates for some layers instructed by the NW to the UE during model training, and "unfreeze" means instructing that weight updates for some layers that were previously frozen be made possible again.

[0106] The present disclosure describes a method for providing pre-trained model parameters to a UE in a network, intended to resolve the issues of prolonged training time and the inability to train high-compression autoencoders when training from scratch (randomly generated initial weights) during conventional Type 2 training. When Type 2 training is started with pre-trained model parameters, the training time is significantly shortened compared to scratch, thereby reducing OTA overhead. Furthermore, when training high-compression autoencoders, it demonstrates significantly more stable training and higher model performance compared to starting with scratch, similarly providing an effect of reducing OTA overhead when used for Type 2 training. Additionally, examples of the training method of this embodiment are defined as Training Method Examples 1 to 3 in FIGS. 6, 9, and 10, respectively, and detailed examples of the training method are as follows.

[0108] FIG. 6 is a drawing illustrating a training method according to one embodiment of the present disclosure.

[0109] The example illustrated in Fig. 6 is a method for efficient training of models having the same encoder input dimension and different feedback sizes. Here, the feedback size refers to the encoder output and decoder input dimensions of the autoencoder. The feedback size is referred to as latent or z in the model. The interior of each bold box indicates a layer responsible for extracting and restoring spatial-frequency domain features for each model. Each dotted box indicates a layer responsible for compression and restoration related to the feedback size.

[0110] The encoder and decoder of an autoencoder can be broadly classified into a layer that extracts and restores spatial-frequency information from the input and a layer that compresses and restores it based on the feedback size, respectively. In this case, to utilize the aforementioned training method, the structure and size of the encoder's feedback size compression and the decoder's feedback size restoration layers may differ depending on the autoencoder with different feedback sizes. However, the layers of the encoder extracting spatial-frequency information from the input eigenvectors and the decoder restoring it must have identical structures and sizes across encoders and decoders, respectively, for models with different feedback sizes. As another example, it may be necessary for certain layers of the eigenvector input spatial-frequency extraction and restoration layer to have the same structure and size for each specific layer across different models; in this case, the previously defined training method using freeze can be applied only to the specific layers that possess the same structure and size.

[0111] The training method illustrated in Fig. 6 can first perform full autoencoder training on a model with a large feedback size. After training is complete, the trained weights are retrieved only from the shared layer that extracts and restores spatial-frequency domain features of the input eigenvectors, and these weights (model parameters) can be shared and utilized as corresponding layer weights for another model with a small feedback size. At this time, the weights of the shared layer are frozen, and autoencoder training can be performed on the remaining weights. After training is complete, the frozen weights are unfrozen, and training (fine tuning) on ​​all autoencoder weights can be performed again. In this way, a model with a small feedback size can be properly trained, and the training time is also reduced, which can reduce OTA overhead during Type 2 training when used in future embodiments.

[0112] The training method illustrated in Fig. 6 can be performed by omitting unfreeze and only freezing, and also includes a method of freezing and omitting training and unfreezing from the beginning to train the entire model.

[0113] The operation of the training method and the characteristics and structure of the model illustrated in Fig. 6 apply to all models operating in the UE and network, and these details may be reflected when defining the reference model and reference model structure in the standard. In addition, these characteristics include the ability to apply them separately to the encoder model only or to the decoder model only, in addition to the method applied to both the encoder and the decoder as described in the paragraphs above.

[0114] The training method illustrated in FIG. 6 may include the following training method conditions. Training can be performed sequentially, starting with a model having a large feedback size (Encoder output / decoder input size) and progressing to models with gradually smaller feedback sizes. The feedback size of the model trained first must be greater than or equal to the feedback size of the model trained later. The training sequence for the training method illustrated in FIG. 6 includes the method illustrated in FIG. 7 below.

[0116] FIG. 7 is a diagram illustrating a training sequence according to one embodiment of the present disclosure.

[0117] Specifically, in step 701, a model with a feedback size larger than the feedback size specified in the current specification can be trained.

[0118] In step 702, the layer for extracting spatial-frequency domain corrections from step 701 can be fixed for training. Then, the remaining model of the model smaller than the feedback size can be trained.

[0119] In step 703, the entire model can be trained. The model completed in step 603 can provide the shared and fixed parameters for step 702 of the next loop. From the second loop onwards, step 701 can be omitted. Steps 701 through 703 are performed repeatedly until a model generating the target feedback size is trained.

[0120] Afterwards, fine-tuning can be performed on the frozen block in step 704. Depending on the model, step 704 may be omitted.

[0121] In other words, Step 1 - Train a model of the biggest feedback size (larger size than the current spec.), Step 2 - Fix the layers for extracting spatial-frequency domain correlation from Step 1 and train the remaining layers of the model of a smaller feedback size (from the second loop onwards, the Step 3 model from the previous loop is used), Step 3 - Train the entire mode, and the model trained in Step 3 provides the shared and fixed parameters for Step 2 of the next loop (Step 1 is omitted from the second loop onwards).

[0122] An example of applying the training sequence of the above method is as follows. Example) After training a 2000-bit model, train a 1000-bit model using it. Train a 500-bit model using the trained 1000-bit model. Train a 200-bit model using the trained 500-bit model.

[0123] However, the training sequence is not limited to the method illustrated in Fig. 7. While training can be performed continuously with gradually decreasing feedback sizes, it is also possible to train all models of different feedback sizes (which must be smaller than the large feedback size) using a portion of a single large feedback size model. In this case, the following operations may be included. The first model of Step 1 can be used in Step 2 and Step 3 of different loops, respectively. For example, after training a 2000-bit model, it can be used to train 1000-bit, 500-bit, and 200-bit models, respectively. After training continuously in the training sequence during the initial multiple loops, the Step 3 model of the last loop, which was trained continuously, can be used in Step 2 and Step 3 of different loops, respectively. For example, after training a 2000-bit model, it can be used to train a 1000-bit model. Subsequently, the 1000-bit model can be used to train 500-bit and 200-bit models, respectively.

[0124] FIGS. 8A and FIGS. 8B are drawings illustrating examples of a model structure according to one embodiment of the present disclosure.

[0125] Specifically, the structure illustrated in FIG. 8a is merely an example representing each representative model backbone and is not limited thereto. The interior of each thick line box indicates a layer that performs the role of extracting and restoring features of the spatial-frequency domain for each model. Each dotted line box indicates a layer that performs the role of compression and restoration in relation to the feedback size.

[0126] The model structure corresponding to the training method illustrated in Fig. 7 can be varied, and Fig. 8b illustrates example structures of a Transformer-based encoder, a transformer decoder, a vision transformer (vit) based encoder, a ResNet-based encoder, and a DenseNet-based encoder.

[0128] FIG. 9 is a drawing illustrating another training method according to one embodiment of the present disclosure.

[0129] The training method example illustrated in Fig. 9 is a method for efficiently training models that have different encoder input / decoder output dimensions (i.e., tx antenna port or subband or both) but the same feedback size. In this case, the input dimension is an eigenvector (V), and the size of the Tx antenna port and subband of that dimension varies by model. The inside of the bold box indicates a layer that extracts and restores features in the spatial-frequency domain for each model. Each dotted box indicates a layer for processing the encoder input and the decoder output.

[0130] The encoder and decoder of an autoencoder can be broadly divided into a layer for extracting and reconstructing spatial-frequency information and a layer for processing encoder inputs and decoder outputs. For models with different input sizes, the layers of the encoder extracting spatial-frequency information and the decoder reconstructing it must have identical structures and sizes across encoders and decoders. As another example, it may be necessary for specific layers within the spatial-frequency extraction and reconstruction layer to have the same structure and size only for those specific layers across different models; in such cases, the previously defined training method using freeze can be applied only to the specific layers that possess the same structure and size.

[0131] In the training method illustrated in Fig. 9, first, full autoencoder training can be performed on a model with a small input dimension (i.e., a 32-port TX antenna model). After training is complete, the trained weights are retrieved only from the shared layer that extracts and restores spatial-frequency domain features from the model, and these weights (model parameters) are shared and used as corresponding layer weights for another model with a large input dimension. At this time, the weights of the shared layer are frozen, and autoencoder training is performed on the remaining weights. After training is complete, the frozen weights are unfrozen, and training (fine tuning) is performed again on all autoencoder weights. In this way, a model with a large input dimension can be properly trained, and the training time is also shortened, which can reduce OTA overhead during Type 2 training when used in future embodiments. The training method illustrated in Fig. 9 can be performed by omitting the unfreeze step and only freezing, or it includes a method of freezing and skipping training, and then unfreezing from the beginning to train the entire model.

[0132] The operation of the training method and the characteristics and structure of the model illustrated in Fig. 9 apply to all models operating in UE and NW, and these details may be reflected when defining the reference model and reference model structure in the standard. In addition, these characteristics include the ability to apply them separately to the encoder model only or to the decoder model only, in addition to the method applied to both the encoder and the decoder as described in the paragraphs above.

[0133] The training method illustrated in FIG. 9 may include the following training method conditions. Training is performed sequentially from models with small Encoder input / decoder output sizes (dimensions), namely Tx antenna port and subband size, to models that gradually increase in size. The Tx antenna port and subband size of the model trained first must be smaller than or equal to the Tx antenna port and subband size of the model trained later. The training sequence includes the following method.

[0134] Step 1-Train a model of smallest Tx antenna port or subband size

[0135] Step 2-Fix the layers for extracting spatial-frequency domain correlation from Step 1 and train remaining layers of the model of larger tx antenna port or subband size (from the second loop, use the step3 model of the previous loop)

[0136] Step 3 - Train the entire model. The model completed in Step 3 can provide the shared and fixed parameters of Step 2 of the next loop (Step 1 is omitted from the second loop onwards). An example of applying this training sequence is as follows. For instance, after training an 8-port model, you can use it to train a 16-port model, use the trained 16-port model to train a 32-port model, and use the trained 32-port model to train a 64-bit model.

[0137] However, the training order is not limited to the above method. While training can be performed sequentially as the Tx antenna port and subband size gradually increase, it is also possible to train models of different sizes (which must be larger relative to the small Tx antenna port and subband size) using a portion of a single small Tx antenna port and subband size model. In such cases, the following operations may be included. The first model from Step 1 can be utilized in Step 2 and Step 3 of different loops, respectively. For example, after training an 8-port model, it can be used to train 16-port, 32-port, and 64-port models, respectively. Alternatively, after training sequentially during multiple initial loops, the Step 3 model from the last sequentially trained loop can be utilized in Step 2 and Step 3 of different loops, respectively. For example, after training an 8-port model, it can be used to train a 16-port model. Subsequently, the 16-port model can be used to train 32-port, 64-port, 256-port, and so on models, respectively.

[0139] FIG. 10 is a drawing illustrating another training method according to one embodiment of the present disclosure.

[0140] The training method illustrated in FIG. 10 is a combined method of the training method illustrated in FIG. 6 and the training method illustrated in FIG. 9, utilizing a common layer that extracts spatial-frequency domain correlation features, and a training method utilizing a layer whose structure and size change as the input and output sizes of the encoder and decoder change. The training method illustrated in FIG. 10 includes all the features of the training method illustrated in FIG. 6 and the training method illustrated in FIG. 9.

[0142] FIG. 11 is a drawing illustrating a procedure according to one embodiment of the present disclosure.

[0143] FIG. 11 is a diagram illustrating the procedure for sequential use of freeze and unfreeze during initial connection of an embodiment (A) of the present disclosure.

[0144] This embodiment is an example in which, after the terminal initially connects to the base station, it shares the configuration required for type 2 training for autoencoder training, and performs type 2 training by training the autoencoder sequentially after freezing some layers, and then training the autoencoder after unfreezing. During this operation, steps 1102, 1103, 1104, 1106, 1107, and 1108 are signals and procedures used in conventional type 2 training, and steps 1101, 1105, and 1109 are signals newly proposed and defined for the operation of this embodiment proposed in the invention.

[0145] Step 1101. The network (1120) can transmit the training configuration to the UE (1100).

[0146] - A message transmitting information about a shared model that a terminal needs to know in order to use the various training method examples proposed in this invention via OTA (definable via RRC / MAC CE / DCI)

[0147] All messages defined above, along with model parameters and model structure, can be transmitted via PDU.

[0148] Model parameters and structure are transmitted via PDU, while freeze status and NW model ID can be transmitted via control messages such as RRC.

[0149] Each detailed field defined in this message may be transmitted separately through different messages, and each message may be transmitted via different methods (i.e., RRC, MAC CE, DCI, PDU, etc.).

[0150] - Case 1) Model structure + Model parameters (+ layer location information) + freeze status (+ location information) + NW model ID (structure ID + parameter ID)

[0151] * (case1-1) Total model parameters (pre-trained in NW) → Layer location information of model parameters X,

[0152] * (case1-2) Case of transmitting some of the model parameters (pre-trained in NW) → Requires transmission of layer location information for model parameters

[0153] - Case 2) Model parameters (+ layer location information) + freeze status (+ location information) + NW model ID (parameter ID)

[0154] * When a reference structure used for Type 2 training is defined in the standard

[0155] * (case2-1) All model parameters + information on the locations of layers to be frozen,

[0156] * (case2-2) Location information of the layer to be frozen + corresponding parameter (smaller than '1' overhead)

[0157] - Freeze status can be indicated by 1 bit (e.g., 0 = freeze, 1 = freeze)

[0158] - When using Freeze, layer location information for freezing must be transmitted.

[0159] Model parameter layer location and layer location information for freezing can be specified by transmitting the first and last layer indices, e.g., 'layers 5~7', model parameters, and freeze.

[0160] Ex) First layer index 4 bits, last layer index 4 bits = 8 bits → '0001' + '0011' (based on the starting layer 0). This is a command to freeze layers 1 through 4. → ​​This example covers a maximum 16-layer model. → If the model becomes larger or smaller, this bit length may be defined differently or changed.

[0161] * When using freeze in Case 1-2 and Case 2-2, since the location of the layer to be frozen is the same as the location of the layer when transmitting model parameters, the transmission of layer location information regarding whether to freeze can be omitted.

[0162] Cases 1-2 and 2-2 have smaller OTA overhead compared to cases 1-1 and 2-1. Since the purpose of transmitting the pre-trained layer is for the terminal to utilize it, it appears that the layer to be frozen and its location will be identical during actual operation.

[0163] - NW model ID

[0164] For Case 1, where both structure and parameter are shared, provide the structure ID + parameter ID; for Case 2, where only the parameter is shared, provide only the parameter ID.

[0165] - Each ID is in a format identifiable in the network. It can be configured as a bitmap index.

[0166] Example) Case1) Structure ID 4bit + parameter ID 4bit, Case2) parameter ID 4bits

[0167] Step 1102. The network (1120) can send a Training start message to the UE (1100).

[0168] - Instruction message to start Type 2 training

[0169] Step 1103. The UE (1100) can transmit Forward propagation (FP) information to the network (1120).

[0170] - According to Step 1101, the process of transmitting the compute values ​​of the encoder model requiring FP execution via OTA to proceed with the entire forward propagation process after the terminal encoder model training begins.

[0171] Step 1104. The network (1120) can transmit back propagation (BP) information to the UE (1100).

[0172] - A process for the NW decoder to complete the FP process using the information from step 1102, perform BP which is a process of calculating the gradient to update the model, and then transmit this information to the UE. The UE that receives this message can use it to update the encoder model and repeat steps 1102 and 1103 to train the model.

[0173] Step 1105. The network (1120) can send an Unfreezing command to the UE (1100).

[0174] - Instructions for performing fine-tuning on all layers after training the remaining layers using the pre-trained layers received from the NW

[0175] - 1-bit message possible (DCI / MAC-CE)

[0176] - When the terminal receives this message, the entire layer is unfrozen to proceed with fine-tuning.

[0177] Convert all layers to trainable state

[0178] Step 1106. The UE (1100) can transmit Forward propagation (FP) information to the network (1120).

[0179] - According to Step 1105, the process of transmitting the compute values ​​of the encoder model requiring FP execution via OTA to proceed with the entire forward propagation process after the terminal encoder model training begins.

[0180] Step 1107. The network (1120) can transmit back propagation (BP) information to the UE (1100).

[0181] - A process for the NW decoder to transmit information to the UE after completing the FP process using the information from Step 1106, performing BP (which calculates gradients to update the model), and receiving this information. Upon receiving this message, the UE uses it to update the encoder model. Steps 1106 and 1107 are repeated to train the model.

[0182] Step 1108. The network (1120) can send a Training end message to the UE (1100).

[0183] - Message sent by the network to the terminal to terminate training

[0184] - If the terminal wishes to end training, it may request this from the NW.

[0185] Step 1109. The network (1120) can transmit the UE model ID to the UE (1100).

[0186] - Bit index message to assign an ID to the trained terminal model after Type 2 training is complete

[0187] - RRC / MAC CE / DCI messages available

[0188] - Example 1) UE model ID can be defined independently and is 8-bit indexable.

[0189] - Example 2) Structure ID 4-bit + parameter ID 4-bit -> 8-bit UE model ID message possible (Similar to NW model ID, each ID can be defined within the field in the form of a parameter corresponding to the structure)

[0191] FIG. 12 is a drawing illustrating a procedure according to another embodiment of the present disclosure.

[0192] FIG. 12 is a diagram illustrating the procedure for using unfreeze or freeze alone during initial connection in embodiment (B) of the present disclosure.

[0193] This embodiment is an example of a type 2 training that, after the terminal initially connects to a base station, shares the configuration required for type 2 training for autoencoder training, and performs the training by freezing some layers and then terminating the training, or by training the autoencoder after unfreezing and then terminating the training.

[0194] In this operation, steps 1202, 1203, 1204, and 1205 are signals and procedures used in conventional type 2 training, while steps 1201 and 1206 are signals newly proposed and defined for the operation of this embodiment proposed in the invention. The difference between this embodiment and embodiment (A) is that, unlike embodiment (A), this embodiment completes model training in a single training process. Details of the signal procedures below that overlap with the preceding embodiment are omitted, and the differences are explained in detail.

[0195] Step 1201. The network (1220) can transmit the training configuration to the UE (1200).

[0196] - Freeze status for this message '0' → Initial connection + unfreeze

[0197] - Freeze status '1' for this message and end training at the training end after the first training → Initial connection + freeze

[0198] Step 1202. The network (1220) can send a Training start message to the UE (1200).

[0199] Step 1203. The UE (1200) can transmit forward propagation (FP) information transmission to the network (1220).

[0200] Step 1204. The network (1220) can transmit back propagation (BP) information to the UE (1200).

[0201] Step 1205. The network (1220) can send a Training end message to the UE (1200).

[0202] Step 1206. The network (1220) can transmit the UE model ID to the UE (1200).

[0204] FIG. 13 is a drawing illustrating a procedure according to another embodiment of the present disclosure.

[0205] FIG. 13 is a diagram illustrating the procedure for sequentially using freeze and unfreeze during a model update in embodiment (C) of the present disclosure.

[0206] This embodiment is an operation in which a terminal transmits a NW model ID possessed by the terminal to a base station as one of the UE capabilities to determine whether the NW requires a model update and to perform the corresponding process. This operation is an example of sequentially performing model training using freeze and unfreeze during type 2 training, similar to embodiment (A). Among these operations, steps 1303, 1304, 1305, 1307, 1308, and 1309 are signals and procedures used in conventional type 2 training, and steps 1301, 1302, 1306, and 1310 are signals newly proposed and defined for the operation of this embodiment proposed in the invention.

[0207] Step 1301. The UE (1300) can transmit the NW Model ID to the network (1320).

[0209] *- Message for NW to determine whether to update the model using the model ID information of the CSI compression encoder model held by the terminal

[0210] * Example 1. If the NW model ID is outdated, a model update (training) can be requested from the terminal.

[0211] Example 2. Although the NW model ID is not outdated, an update or additional training request can be made to the terminal when monitoring performance degradation is detected or when training a model of a different feedback size is required.

[0212] - Can be transmitted as one of the UE capabilities

[0213] - A message that transmits the value of the NW model ID information, a detailed field of 'Training configuration', held by the terminal to the NW.

[0214] Step 1302. The network (1320) can transmit the training configuration to the UE (1300).

[0215] - If the NW model ID in 'Step 1301' is outdated, it is necessary to transmit a new model structure + model parameters or model parameters and other detailed fields.

[0216] - If the NW model ID in 'Step 1301' is not outdated, it is possible to transmit either the model structure + model parameters or only detailed fields other than the model parameters and NW model ID (since the existing NW model will be maintained and only training will proceed).

[0217] Step 1303. The network (1320) can send a Training start message to the UE (1300).

[0218] Step 1304. The UE (1300) can transmit Forward propagation (FP) information to the network (1320).

[0219] Step 1305. The network (1320) can transmit back propagation (BP) information to the UE (1300).

[0220] Step 1306. The network (1320) can send an Unfreezing command to the UE (1300).

[0221] Step 1307. The UE (1300) can transmit Forward propagation (FP) information to the network (1320).

[0222] Step 1308. The network (1320) can transmit back propagation (BP) information to the UE (1300).

[0223] Step 1309. The network (1320) can send a Training end message to the UE (1300).

[0224] Step 1310. The network (1320) can transmit the UE model ID to the UE (1300).

[0226] FIG. 14 is a drawing illustrating a procedure according to another embodiment of the present disclosure.

[0227] FIG. 14 is a diagram illustrating the procedure for using only unfreeze during model update or freeze during initial connection in embodiment (D) of the present disclosure.

[0228] This embodiment is an operation in which a terminal transmits a NW model ID possessed by the terminal to a base station as one of the UE capabilities to determine whether the NW requires a model update and to perform the corresponding process. Similar to Embodiment (B), this operation is an example of performing only one model training process using freeze or unfreeze during type 2 training. In this operation, steps 1403, 1404, 1405, and 1406 are signals and procedures used in conventional type 2 training, while steps 1401, 1402, and 1407 are signals newly proposed and defined for the operation of this embodiment proposed in the invention. Steps 1401 and 1402 of this embodiment are identical to steps 1301 and 1302 of Embodiment (C), and the subsequent procedure (steps 1403~) is configured identically to steps 1202~ of Embodiment (B).

[0229] Step 1401. The UE (1400) can transmit the NW Model ID to the network (1420).

[0230] Step 1402. The network (1420) can transmit the training configuration to the UE (1400).

[0231] Step 1403. The network (1420) can send a Training start message to the UE (1400).

[0232] Step 1404. The UE (1400) can transmit forward propagation (FP) information transmission to the network (1420).

[0233] Step 1405. The network (1420) can transmit back propagation (BP) information to the UE (1400).

[0234] Step 1406. The network (1420) can send a Training end message to the UE (1400).

[0235] Step 1407. The network (1420) can transmit the UE model ID to the UE (1400).

[0237] FIG. 15 is a drawing illustrating the operation of a terminal according to one embodiment of the present disclosure.

[0238] In step 1501, training settings can be received from the base station. The training settings may include information about the training model and an indicator indicating whether to freeze.

[0239] Additionally, at step 1502, the terminal may receive a training start from the base station based on the received training settings. Subsequently, at step 1503, the terminal may transmit forward propagation to the base station, and at step 1504, the base station may transmit back propagation to the terminal. Here, steps 1503 and 1504 may be performed repeatedly until training is completed.

[0240] Afterwards, in step 1505, the terminal can receive an unfreezing command from the base station.

[0241] Subsequently, the terminal can transmit a second forward propagation to the base station, and the base station can transmit a second back propagation to the terminal. Here, the second forward propagation and the second back propagation can also be performed repeatedly until training is completed.

[0242] Subsequently, the terminal receives the training termination from the base station and can receive the terminal's model identifier (identifier, ID) from the base station.

[0243] In addition, the terminal can determine whether to perform a weight update for the layer where training is performed based on an indicator indicating whether to freeze.

[0244] In addition, the terminal may transmit the base station's model identifier (identifier, ID) to the base station before receiving the above training settings.

[0246] FIG. 16 is a drawing illustrating the operation of a network according to one embodiment of the present disclosure.

[0247] In step 1601, the base station may transmit training settings to the terminal. The training settings may include information about the training model and an indicator indicating whether to freeze.

[0248] Additionally, in step 1602, the base station may transmit a training start based on the received training settings. Subsequently, in step 1603, the terminal may transmit forward propagation to the base station, and in step 1604, the base station may transmit back propagation to the terminal. Here, steps 1603 and 1604 may be performed repeatedly until training is completed.

[0249] Afterwards, in step 1605, the base station can send an unfreezing command to the terminal.

[0250] Subsequently, the terminal can transmit a second forward propagation to the base station, and the base station can transmit a second back propagation to the terminal. Here, the second forward propagation and the second back propagation can also be performed repeatedly until training is completed.

[0251] Subsequently, the base station can transmit the training termination to the terminal and transmit the terminal's model identifier (identifier, ID) to the terminal.

[0252] In addition, the terminal can determine whether to perform a weight update for the layer where training is performed based on an indicator indicating whether to freeze.

[0253] In addition, prior to transmitting the above training settings, the base station's model identifier (identifier, ID) can be received from the terminal.

[0255] FIG. 17 is a block diagram of a terminal or user equipment (1700) according to one embodiment of the present disclosure.

[0256] The terminal (1700) is an electronic device capable of wireless communication and may have various form factors. Examples of the terminal may include at least one of a user device (UE), a mobile station (MS), a cellular phone, a smartphone, a computer, a tablet, a wearable device, an Internet of Things (IoT) device, or other devices / systems capable of performing wireless communication with a base station (BS) and / or other terminals via a wireless channel.

[0257] Referring to FIG. 17, the terminal (1700) may include at least one transceiver (1701) (hereinafter, transceiver), at least one processor (1702) (hereinafter, processor), and at least one memory (1703) (hereinafter, memory). According to at least one or a combination thereof of methods corresponding to embodiments of the present disclosure, the transceiver (1701), processor (1702), and memory (1703) of the terminal (1700) may be operated. However, the components of the terminal (1700) are not limited to the examples of components shown in FIG. 17. In other embodiments, the terminal (1700) may include additional components in addition to the aforementioned components, or some components may be omitted. Also, in some embodiments, any combination of the transceiver (1701), processor (1702), or memory (1703) may be integrated into a single component.

[0258] The transceiver (1701) may be a basic communication circuit or communication circuitry that enables the terminal (1700) to perform wireless communication with a node or entity of a network. For example, the transceiver (1701) may enable the terminal (1700) to transmit and receive signals to and from a base station via cellular wireless communication, or to transmit and receive signals to and from another terminal via cellular wireless communication. For example, the transceiver (1701) may be 3G (3rd generation), 4G (4th generation), LTE (long-term evolution), 5G (5th generation), NR (new radio), 6G (6th It can support at least one of various cellular wireless communication technologies including (generation), and the various cellular wireless communication technologies supported by the transceiver (1701) may include all subsequent evolved generations of wireless communication.

[0259] According to one embodiment, the terminal (1700) may include a plurality of transceivers, and for example, when supporting EN-DC (E-UTRA (evolved-universal terrestrial radio access) - NR dual connectivity), it may include a first transceiver supporting 4G LTE wireless communication and a second transceiver supporting 5G NR wireless communication. According to another embodiment, when the terminal (1700) supports NR-DC (NR Dual Connectivity), the terminal (1700) may include a plurality of transceivers supporting 5G NR wireless communication. According to another embodiment, if the terminal (1700) supports short-range wireless communication, the terminal (1700) may separately include a transceiver that supports at least one of a group of wireless communication protocol standards such as those defined by Bluetooth®, wireless LAN or WLAN (wireless local area network) network (including, but not limited to, IEEE (institute of electrical and electronics engineer) 802.11-2016 standard or modifications thereof such as 802.11ah, 802.11ad, 802.11ay, 802.11ax, 802.11az, 802.11ba and 802.11be).

[0260] According to one embodiment, the transceiver (1701) may include various circuit structures used to transmit and receive signals to and from a base station via a wireless channel. The signals may include control information and data. For example, the transceiver (1701) may be configured to include a radio frequency (RF) transmitter that up-converts and amplifies the frequency of a transmitted signal, and an RF receiver that low-noise amplifies a received signal and down-converts the frequency. The transceiver (1701) may output the signal received via the wireless channel to a processor (1702) and transmit the signal output from the processor (1702) via the wireless channel.

[0261] A processor (1702) may control the overall operation of a terminal (1700) according to an embodiment of the present disclosure. The processor (1702) may be implemented as one or more IC (integrated circuit (or circuitry)) chips and may execute various data processing operations. The processor (1702) may include at least one electrical circuit and may execute instructions (or programs, code, data, etc.) stored in memory (1703) individually, collectively, or in any combination. Additionally, the processor (1702) may include a single-core processor or a multi-core processor, and in a specific implementation, may be composed of a processor assembly including a plurality of processing circuits.

[0262] The processor (1702) is electrically, operatively, and / or communicatively coupled to the transceiver (1701) so as to control the transceiver (1701).

[0263] The processor (1702) may include at least one processor (or, processing circuitry), and at least one processor may perform the following operations individually, collectively, or in any combination. For example, the processor (1702) may include a communication processor (CP) that controls communication operations and an application processor (AP) that controls the execution of an upper layer (e.g., an application layer). In a specific embodiment, at least one part of the processor (1702) may be included in one chip (or, IC), and another part of the processor (1702) may be included in a separate chip (or, IC). Alternatively, at least one processor may be included in other components, e.g., a transceiver (1701) or a memory (1703).

[0264] The processor (1702) may perform, cause, or control the operation of a terminal to perform at least one or a combination of the methods according to the embodiments of the present disclosure. For example, the processor (1702) may control the operation of a terminal to process a downlink signal received from a base station or to generate an uplink signal and transmit it to a base station. To this end, the processor (1702) may control other components of the terminal (1700) to perform various operations by executing computer programs, code, or instructions stored in memory (1703).

[0265] Memory (1703) is a hardware storage device capable of storing information temporarily or permanently and may include one or more storage media. For example, memory (1703) may include a memory assembly comprising one or more storage media. For example, the one or more storage media may include a hard drive, flash memory, permanent memory such as ROM (read-only memory), semipermanent memory such as RAM (random access memory), cache memory, or any combination thereof.

[0266] The memory (1703) can be electrically, operatively, and / or communicatively coupled to the processor (1702) and can be accessed by the processor (1702).

[0267] A computer program, code, or instruction that can be executed by a processor (1702) may be stored in the memory (1703). According to one embodiment, the computer program, code, or instruction that can be executed by the processor (1702) may be stored in a single memory device or may be separated and distributed across two or more memory devices. The processor (1702) may perform various functions according to the embodiments of the present disclosure by executing the instruction stored in the memory (1703).

[0268] According to one embodiment of the present disclosure, the operation of the terminal (1700) may be caused to be performed based on at least one processor (or processing circuit) configured to perform the features of the present disclosure individually, collectively, or in any combination based on the execution of instructions (or computer program or code) stored in memory (1703), based on processing circuitry not configured to execute instructions, and / or based on components of a processing circuitry not configured to execute instructions.

[0270] FIG. 18 is a block diagram of a base station (1800) according to one embodiment of the present disclosure.

[0271] The base station (1800) can perform wireless communication with at least one terminal within the area of ​​the base station (1800) via a wireless channel. The base station (1800) can perform communication with a node or entity of the network via wired or wireless communication.

[0272] Referring to FIG. 18, a base station (1800) may include at least one transceiver (1801) (hereinafter, transceiver), at least one processor (1802) (hereinafter, processor), and at least one memory (1803) (hereinafter, memory). According to at least one or a combination thereof of methods corresponding to embodiments of the present disclosure, the transceiver (1801), processor (1802), and memory (1803) of the base station (1800) may be operated. However, the components of the base station (1800) are not limited to the examples of components shown in FIG. 18. In other embodiments, the base station (1800) may include additional components in addition to the aforementioned components, or some components may be omitted. Also, in some embodiments, any combination of the transceiver (1801), processor (1802), or memory (1803) may be integrated into a single component.

[0273] The transceiver (1801) may be a communication circuit or communication circuitry that enables the base station (1800) to perform wireless communication with a node or entity of the network. For example, the transceiver (1801) may enable the base station (1800) to transmit and receive signals to and from a terminal (1700) via cellular wireless communication or to transmit and receive signals to and from another network entity via wireless communication. For example, the transceiver (1801) may be 3G (3rd generation), 4G (4th generation) LTE (long-term evolution), 5G (5th generation) NR (new radio), 6G (6th generation) Various cellular wireless communication technologies, including (generation), etc., can be supported, and the various cellular wireless communication technologies supported by the transceiver (1801) may include all subsequent evolved generations of wireless communication. According to one embodiment, the transceiver (1801) may include various circuit structures used to transmit and receive signals to and from a terminal via a wireless channel. The signals may include control information and data. For example, the transceiver (1801) may be configured to include an RF (radio frequency) transmitter that up-converts and amplifies the frequency of a transmitted signal, and an RF receiver that low-noise amplifies a received signal and down-converts the frequency. The transceiver (1801) may output the signal received via the wireless channel to a processor (1802) and transmit the signal output from the processor (1802) via the wireless channel.

[0274] Meanwhile, according to one embodiment of the present disclosure, a base station (1800) may communicate with an entity or node of a network via wired or wireless communication. For example, the base station (1800) may communicate via wired or wireless communication with an entity or node of an adjacent base station or core network via a backhaul network. Although not shown in the drawings, when the base station (1800) performs wired communication, the base station (1800) may include a separate network interface for wired communication in addition to the transceiver (1801). The network interface may be referred to as network interface circuitry, communication interface circuitry, etc.

[0275] A processor (1802) can control the overall operation of a base station (1800) according to an embodiment of the present disclosure. The processor (1802) may be implemented as one or more IC (integrated circuit or circuitry) chips and may execute various data processing operations. The processor (1802) may include at least one electrical circuit and may execute instructions (or programs, code, data, etc.) stored in memory (1803) individually, collectively, or in any combination. Additionally, the processor (1802) may include a single-core processor or a multi-core processor, and in a specific implementation, may be composed of a processor assembly including a plurality of processing circuits.

[0276] The processor (1802) is electrically, operatively, and / or communicatively coupled to the transceiver (1801) so as to control the transceiver (1801).

[0277] The processor (1802) may include at least one processor (or processor circuitry), and at least one processor may perform the following operations individually, collectively, or in any combination. In a specific embodiment, at least one part of the processor (1802) may be included in one chip (or IC), and another part of the processor (1802) may be included in a separate chip (or IC). Alternatively, at least one processor may be included in other components, such as a transceiver (1801) or a memory (1803).

[0278] The processor (1802) may perform, cause, or control the operation of a base station to perform at least one or a combination of the methods according to the embodiments of the present disclosure. For example, the processor (1802) may control the operation of a base station to generate a downlink signal and transmit it to a terminal, or to process an uplink signal received from a terminal. Alternatively, the base station may transmit and receive signals with an adjacent base station, transmit a signal received from a terminal to an upper node of the network, or receive a signal from an upper node of the network and transmit it to a terminal. To this end, the processor (1802) may control other components of the base station (1800) to perform various operations by executing computer programs, codes, and instructions stored in memory (1803).

[0279] Memory (1803) is a hardware storage device capable of storing information temporarily or permanently and may include one or more storage media. For example, memory (1803) may include a memory assembly comprising one or more storage media. For example, the one or more storage media may include a hard drive, flash memory, permanent memory such as ROM (read-only memory), semi-permanent memory such as RAM (random access memory), cache memory, or any combination thereof.

[0280] The memory (1803) can be electrically, operatively, and / or communicatively coupled to the processor (1802) and can be accessed by the processor (1802).

[0281] A computer program, code, or instruction that can be executed by a processor (1802) may be stored in the memory (1803). According to one embodiment, the computer program, code, or instruction that can be executed by the processor (1802) may be stored in a single memory device or may be separated and distributed among two or more memory devices. The processor (1802) may perform various functions according to the embodiments of the present disclosure by executing the instruction stored in the memory (1803).

[0282] According to one embodiment of the present disclosure, the operation of a base station (1800) may be caused to be performed based on at least one processor (or processing circuitry) configured to perform the features of the present disclosure individually, collectively, or in any combination based on the execution of instructions (or computer programs or code) stored in memory (1803), based on processing circuitry not configured to execute instructions, and / or based on components of a processing circuitry not configured to execute instructions.

[0284] A terminal or a base station can perform various communication procedures related to the control plane or user plane by interacting with network entities based on communication through a wireless channel. For example, a terminal can communicate with network entities such as an access and mobility management function (AMF) or a session management function (SMF) through a base station. Alternatively, the base station can perform at least one communication procedure by directly transmitting and receiving signals or relaying them with network entities. The structure of the above-mentioned network entities will be explained in more detail through the drawings below.

[0285] FIG. 19 is a block diagram of a network entity (1900) performing network functions according to one embodiment of the present disclosure.

[0286] A network entity (1900) may include one or more network functions (NF) that constitute a core network (e.g., 5G (5th generation) core, 5GC) in a communication system, or entities (devices, devices, nodes, or servers, etc.) that perform part of a network function. In this case, multiple NFs may be implemented within a single network entity, or a single NF may be distributed and implemented across multiple network entities. Additionally, when an NF is implemented within a network entity, the NF may be implemented in the form of software, and in such cases, a program for running the NF may be loaded into the memory of the network entity (1900).

[0287] A single NF can be implemented as one or more instances and can operate by being distributed across the same network entity or multiple network entities. Here, the instance is a software unit that logically executes a specific network function and may be separate from physical hardware resources. Additionally, one or more NFs may be implemented as a single network slice to operate in order to satisfy the specifications required by a specific service.

[0288] The above NF may include any one of an access and mobility management function (AMF), a session management function (SMF), a local session management function (L-SMF), a user plane function (UPF), a local user plane function (L-UPF), a policy control function (PCF), unified data management (UDM), a unified data repository (UDR), a network exposure function (NEF), a network repository function (NRF), an application function (AF), a network slice selection function (NSSF), a network data analytics function (NWDAF), a network slice admission control function (NSACF), an authentication server function (AUSF), or a data network (DN).

[0289] Referring to FIG. 19, a network entity (1900) may include at least one network interface (1901), at least one processor (1902) (hereinafter, processor), and at least one memory (1903) (hereinafter, memory). As described above, the NF may be implemented in the form of a physical device such as the network entity (1900), or may be implemented and executed in the form of a virtualized instance. When the NF is implemented in the form of an instance, it may not necessarily include physical components as illustrated in FIG. 19. In such cases, the instance may be composed of one or more logical functional units and may be logically represented.

[0290] According to at least one or a combination thereof of the methods corresponding to the embodiments of the present disclosure, the network interface (1901), processor (1902), and memory (1903) of the network entity (1900) may be operated. However, the components of the network entity (1900) are not limited to the examples of components shown in FIG. 19. In other embodiments, the network entity (1900) may include additional components in addition to the aforementioned components, or some components may be omitted. Also, in one embodiment, the network interface (1901), processor (1902), or memory (1903) may be implemented as a single component.

[0291] The network interface (1901) is a collective term for the transmitting and receiving parts of a network entity and may be a communication circuit for transmitting and receiving signals with a terminal (user equipment, UE), a base station, or other network entities. In this case, the communication circuit may include both a communication circuit for wireless communication and a communication circuit for wired communication. For example, the network interface (1901) may include circuits, logic, hardware, etc. configured to exchange control plane messages or user plane messages with a terminal, a base station, or other core network entities via wireless or wired communication. The network interface (1901) may operate using various protocols (e.g., NAS (Non-Access Stratum) protocols). Depending on the convenience of description and technical implementation, the network interface (1901) may be referred to as a communication circuitry, a network interface circuitry, or a communication interface circuitry.

[0292] A processor (1902) may control the overall operation of a network entity (1900) according to an embodiment of the present disclosure. In one embodiment, the processor (1902) may be implemented as one or more integrated circuit (IC) chips and may execute various data processing operations. The processor (1902) may include at least one electrical circuit and may execute instructions (or programs, code, data, etc.) stored in memory (1903) individually, collectively, or in any combination. Additionally, the processor (1902) may include a single-core processor or a multi-core processor, and in a specific implementation, may be composed of a processor assembly including a plurality of processing circuits. Additionally, it should be noted that the processor (1902) may not necessarily be composed of physical hardware when the network function (1900) is implemented in an instance form according to another embodiment.

[0293] According to one embodiment, the processor (1902) is electrically, operatively, and / or communicatively coupled to the network interface (1901) so as to control the network interface (1901).

[0294] The processor (1902) may include at least one processor (or processor circuitry), and at least one processor may perform the following operations individually, collectively, or in any combination. In a particular embodiment, at least one part of the processor (1902) may be included in one chip (or IC), and another part of the processor (1902) may be included in a separate chip (or IC). Alternatively, at least one processor may be included in other components, such as a network interface (1901) or memory (1903).

[0295] A processor (1902) may perform or control operations of a network entity (1900) to perform at least one or a combination thereof of methods according to embodiments of the present disclosure. For example, the processor (1902) may control operations of the network entity (1900) to exchange control plane messages or user plane messages with terminals, base stations, or other core network entities via wireless or wired communication using various protocols (e.g., NAS protocols). To this end, the processor (1902) may control other components of the network entity (1900) to perform various operations by executing computer programs, code, or instructions stored in memory (1903).

[0296] Memory (1903) is a hardware storage device capable of storing information temporarily or permanently and may include one or more storage media. For example, memory (1903) may include a memory assembly comprising one or more storage media. For example, the one or more storage media may include a hard drive, flash memory, permanent memory such as ROM (read-only memory), semipermanent memory such as RAM (random access memory), cache memory, or any combination thereof.

[0297] According to one embodiment, the memory (1903) may be electrically, operatively, and / or communicatively coupled to the processor (1902) and may be accessed by the processor (1902).

[0298] A computer program, code, or instruction that can be executed by a processor (1902) may be stored in the memory (1903). According to one embodiment, the computer program, code, or instruction that can be executed by the processor (1902) may be stored in a single memory or may be separated and distributed across two or more memories. The processor (1902) may perform various functions according to the embodiments of the present disclosure by executing the instruction stored in the memory (1903).

[0299] According to one embodiment of the present disclosure, the operation of a network entity (1900) may be caused to be performed based on at least one processor (or processing circuit) configured to perform the features of the present disclosure individually, collectively, or in any combination based on the execution of instructions (or computer program or code) stored in memory (1903), based on processing circuitry not configured to execute instructions, and / or based on components of a processing circuitry not configured to execute instructions.

[0300] Meanwhile, although specific embodiments have been described in the detailed description of the present disclosure, it is understood that various modifications are possible within the scope of the present disclosure. Therefore, the scope of the present disclosure should not be limited to the described embodiments, but should be defined by the claims set forth below as well as equivalents thereof.

Claims

Claim 1 A method performed at a terminal in a wireless communication system, comprising: receiving a training setting from a base station; receiving a training start from the base station based on the training setting; transmitting forward propagation to the base station; receiving back propagation from the base station; and receiving an unfreezing command from the base station, wherein the training setting includes information about a training model and an indicator indicating whether to freeze. Claim 2 A method according to claim 1, further comprising: a step of transmitting a second forward propagation to the base station; a step of receiving a second back propagation from the base station; a step of receiving the training termination from the base station; and a step of receiving the model identifier (ID) of the terminal from the base station. Claim 3 A method according to claim 1, further comprising the step of determining whether to perform a weight update for a layer where training is performed based on an indicator indicating whether to freeze. Claim 4 A method according to claim 1, further comprising the step of transmitting the model identifier (identifier, ID) of the base station to the base station. Claim 5 A method performed at a base station in a wireless communication system, comprising: transmitting a training setting to a terminal; transmitting a training start to the terminal based on the training setting; receiving forward propagation from the terminal; transmitting back propagation to the terminal; and transmitting an unfreezing command to the terminal, wherein the training setting includes information about a training model and an indicator indicating whether to freeze. Claim 6 A method according to claim 5, further comprising: receiving a second forward propagation from the terminal; transmitting a second back propagation to the terminal; transmitting the training termination to the terminal; and receiving the model identifier (ID) of the terminal to the terminal. Claim 7 A method according to claim 5, further comprising the step of determining whether to perform a weight update for a layer on which training is performed based on an indicator indicating whether to freeze. Claim 8 A method characterized by further including the step of receiving a model identifier (identifier, ID) of the base station from the terminal in claim 5. Claim 9 A terminal comprising: a transceiver capable of transmitting and receiving at least one signal; and a processor coupled to the transceiver, configured to receive a training setting from a base station, receive a training start based on the training setting from the base station, transmit forward propagation to the base station, receive back propagation from the base station, and receive an unfreezing command from the base station, wherein the training setting includes information about a training model and an indicator indicating whether to freeze. Claim 10 In claim 9, the control unit is further configured to transmit a second forward propagation to the base station, receive a second back propagation from the base station, receive the training termination from the base station, and receive the model identifier (ID) of the terminal from the base station. Claim 11 In claim 9, the control unit is further configured to determine whether to perform a weight update for a layer where training is performed based on an indicator indicating whether to freeze. Claim 12 In claim 9, the terminal is further configured such that the control unit transmits the model identifier (identifier, ID) of the base station to the base station. Claim 13 A base station comprising: a transceiver capable of transmitting and receiving at least one signal; and a processor coupled to the transceiver, wherein the processor is configured to: transmit a training setting to a terminal, transmit a training start to the terminal based on the training setting, receive forward propagation from the terminal, transmit back propagation to the terminal, and transmit an unfreezing command to the terminal, wherein the training setting includes information about a training model and an indicator indicating whether to freeze. Claim 14 In claim 13, the control unit is further configured to receive a second forward propagation from the terminal, transmit a second back propagation to the terminal, transmit the training termination to the terminal, and receive the model identifier (ID) of the terminal from the terminal, characterized in that it is a base station. Claim 15 In claim 13, the base station is further configured to receive a model identifier (identifier, ID) of the base station from the terminal.

Images