Method and apparatus for data sharing and model training for compressing and restoring CSI in wireless communication system

Abstract

The present disclosure relates to a 5G or 6G communication system for supporting a higher data transmission rate beyond a 4G communication system such as LTE. According to an embodiment of the present disclosure, CSI can be efficiently compressed and restored on the basis of AI.

Description

Data sharing and model training method and device for compressing and decompressing CSI in wireless communication systems

[0001] The present disclosure relates to a wireless communication system or a mobile communication system. Specifically, it relates to a method and apparatus for a terminal and a base station to share data and to train a model for efficiently compressing and restoring CSI based on the shared data.

[0002] Looking back at the evolution of wireless communication through successive generations, technologies have been developed primarily for human-oriented services, such as voice, multimedia, and data. Following the commercialization of 5G (5th-generation) communication systems, connected devices, which have been increasing explosively, are expected to be connected to communication networks. Examples of networked objects include vehicles, robots, drones, home appliances, displays, smart sensors installed in various infrastructures, construction machinery, and factory equipment. Mobile devices are expected to evolve into various form factors, such as augmented reality glasses, virtual reality headsets, and holographic devices. In the 6G (6th-generation) era, efforts are underway to develop improved 6G communication systems to connect hundreds of billions of devices and objects to provide diverse services. For this reason, 6G communication systems are referred to as "Beyond 5G" systems.

[0003] In the 6G communication system predicted to be realized around 2030, the maximum transmission speed is tera (i.e., 1,000 gigabit) bps, and the wireless latency is 100 microseconds (μsec). In other words, compared to the 5G communication system, the transmission speed in the 6G communication system is 50 times faster, and the wireless latency is reduced to one-tenth.

[0004] To achieve such high data transmission speeds and ultra-low latency, 6G communication systems are being considered for implementation in the terahertz band (e.g., the 95 GHz to 3 terahertz (3 THz) band). In the terahertz band, due to more severe path loss and atmospheric absorption compared to the millimeter wave (mmWave) band introduced in 5G, the importance of technology capable of guaranteeing signal reach, or coverage, is expected to increase. As key technologies to ensure coverage, radio frequency (RF) devices, antennas, new waveforms that offer better coverage than orthogonal frequency division multiplexing (OFDM), beamforming, and multi-antenna transmission technologies such as massive multiple-input and multiple-output (MIMO), full-dimensional MIMO (FD-MIMO), array antennas, and large-scale antennas must be developed. In addition, new technologies such as metamaterial-based lenses and antennas, high-dimensional spatial multiplexing technology using orbital angular momentum (OAM), and reconfigurable intelligent surface (RIS) are being discussed to improve coverage of terahertz band signals.

[0005] In addition, to improve frequency efficiency and system network, development is underway in 6G communication systems for full duplex technology, in which uplink and downlink simultaneously utilize the same frequency resources at the same time; network technology that integrates satellites and HAPS (high-altitude platform stations); network structure innovation technology that supports mobile base stations and enables network operation optimization and automation; dynamic spectrum sharing technology through collision avoidance based on spectrum usage prediction; AI-based communication technology that utilizes AI (artificial intelligence) from the design stage and internalizes end-to-end AI support functions to realize system optimization; and next-generation distributed computing technology that realizes services of complexity exceeding the limits of terminal computing capabilities by utilizing ultra-high performance communication and computing resources (mobile edge computing (MEC), cloud, etc.). In addition, attempts are continuing to further strengthen connectivity between devices, further optimize networks, promote the softwareization of network entities, and increase the openness of wireless communication through the design of new protocols to be used in 6G communication systems, the implementation of hardware-based security environments, the development of mechanisms for the safe utilization of data, and the development of technologies regarding privacy maintenance methods.

[0006] Due to the research and development of such 6G communication systems, it is expected that a new dimension of hyper-connected experience will become possible through the hyper-connectivity of 6G communication systems, which encompasses not only connections between objects but also connections between people and objects. Specifically, it is projected that 6G communication systems will enable the provision of services such as truly immersive extended reality (XR), high-fidelity mobile holograms, and digital replicas. Furthermore, services such as remote surgery, industrial automation, and emergency response, which are provided through 6G communication systems with enhanced security and reliability, will be applied in various fields including industry, healthcare, automotive, and home appliances.

[0007] The purpose of the present disclosure is to improve the accuracy of CSI feedback by reducing information loss caused by compression and restoration through efficient compression and restoration of CSI based on AI.

[0008] According to one embodiment of the present disclosure for solving the above-mentioned problems, a method performed by a terminal in a wireless communication system comprises: receiving a message from a base station comprising a first data set and a second data set for training a terminal-side encoder that performs compression of Channel State Information (CSI); and performing training based on the first data set and the second data set such that when the terminal-side encoder uses the CSI as input information, it generates information identical to the output information of the CSI compression performed by the base station-side encoder, wherein the first data set includes the CSI and the CSI compressed based on quantization, the second data set includes the CSI compressed based on Artificial Intelligence (AI), and the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the Artificial Intelligence.

[0009] According to one embodiment of the present disclosure for solving the above-mentioned problems, a method performed by a base station in a wireless communication system comprises: a step of determining whether to transmit a first data set and a second data set for training a terminal-side encoder that performs compression of Channel State Information (CSI); and, when transmission is determined, a step of transmitting a message to a terminal comprising the first data set including the CSI and the CSI compressed based on quantization, and the second data set including the CSI compressed based on Artificial Intelligence (AI); wherein the terminal-side encoder is trained to generate information identical to the output information of the CSI compression performed by the base station-side encoder based on the first data set and the second data set, and the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the Artificial Intelligence.

[0010] According to one embodiment of the present disclosure for solving the above-mentioned problems, a terminal of a wireless communication system comprises: a transceiver configured to transmit and receive a signal; and a control unit coupled to the transceiver, wherein the control unit receives a message from a base station comprising a first data set and a second data set for training a terminal-side encoder that performs compression of Channel State Information (CSI); and based on the first data set and the second data set, performs training to generate information identical to the output information of CSI compression performed by a base station-side encoder when the terminal-side encoder uses the CSI as input information, wherein the first data set includes the CSI and CSI compressed based on quantization, the second data set includes CSI compressed based on Artificial Intelligence (AI), and the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the Artificial Intelligence.

[0011] According to one embodiment of the present disclosure for solving the above-mentioned problems, a base station of a wireless communication system comprises: a transceiver configured to transmit and receive signals; and a control unit coupled to the transceiver; wherein the control unit determines whether to transmit a first data set and a second data set for training a terminal-side encoder that performs compression of Channel State Information (CSI); and when transmission is determined, transmits a message to a terminal comprising the first data set including the CSI and the CSI compressed based on quantization, and the second data set including the CSI compressed based on Artificial Intelligence (AI); wherein the terminal-side encoder is trained to generate information identical to the output information of the CSI compression performed by the base station-side encoder based on the first data set and the second data set, and the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the Artificial Intelligence.

[0012] According to one embodiment of the present disclosure, CSI compression and restoration with minimal information loss can be performed based on additional information and data sets.

[0013] The effects obtainable in the present disclosure are not limited to those mentioned in the various embodiments, and other unmentioned effects will be clearly understood by those skilled in the art to which the present disclosure pertains from the description below.

[0014] Figure 1 is a diagram illustrating a procedure for performing feedback on CSI.

[0015] Figure 2 is a diagram illustrating the procedure for performing CSI feedback based on a codebook.

[0016] Figure 3a is a diagram illustrating AI / ML-based CSI compression technology.

[0017] Figure 3b is a diagram illustrating the AI / ML-based CSI compression and CSI restoration procedures.

[0018] Figure 4 is a diagram illustrating the input and output of an encoder and a decoder.

[0019] Figure 5 is a diagram illustrating the process of a terminal compressing CSI based on AI and feeding it back to a base station.

[0020] Figure 6 is a diagram illustrating a two-sided model that performs vendor-specific training.

[0021] Figure 7 is a diagram illustrating the data set exchange and model learning procedure between a terminal and a base station.

[0022] Figure 8 illustrates the terminal-side encoder training procedure.

[0023] FIG. 9a is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to one embodiment of the present disclosure.

[0024] FIG. 9b is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to one embodiment of the present disclosure.

[0025] FIG. 9c is a diagram illustrating a procedure for exchanging a first data set and a second data set on the base station side according to one embodiment of the present disclosure.

[0026] FIG. 10 is a diagram illustrating a data set exchange and model training procedure between a terminal and a base station based on additional information according to one embodiment of the present disclosure.

[0027] FIG. 11 is a diagram illustrating a first method for terminal-side encoder training.

[0028] FIG. 12 is a drawing illustrating a second method for terminal-side encoder training according to one embodiment of the present disclosure.

[0029] FIG. 13 is a drawing illustrating a third method for terminal-side encoder training according to one embodiment of the present disclosure.

[0030] FIG. 14a is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to another embodiment of the present disclosure.

[0031] FIG. 14b is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to another embodiment of the present disclosure.

[0032] FIG. 14c is a diagram illustrating a procedure for exchanging a first data set and a second data set on the base station side according to another embodiment of the present disclosure.

[0033] FIG. 15a is a block diagram showing the configuration of a terminal according to one embodiment of the present disclosure.

[0034] FIG. 15b is a block diagram showing the configuration of a control unit of a terminal according to one embodiment of the present disclosure.

[0035] FIG. 16a is a block diagram showing the configuration of a base station according to one embodiment of the present disclosure.

[0036] FIG. 16b is a block diagram showing the configuration of a control unit of a base station according to one embodiment of the present disclosure.

[0037] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0038] In describing the embodiments, technical details that are well known in the technical field 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.

[0039] For the same reason, some components in the attached drawings have been exaggerated, omitted, or schematically depicted. Additionally, the dimensions of each component do not entirely reflect their actual dimensions. Identical or corresponding components in each drawing have been assigned the same reference numbers.

[0040] 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. The embodiments provided are merely to make the present disclosure 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.

[0041] At this time, it will be understood that each block of the process flow diagrams and combinations of the flow diagrams can be executed by computer program instructions. Since these computer program instructions can be loaded into the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create means to perform the functions described in the flow diagram block(s). Since these computer program instructions can 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 function in a specific way, the instructions stored in computer-available or computer-readable memory can also produce a manufactured item containing instruction means to perform the function 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).

[0042] 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 instance, two blocks described in succession may actually be executed substantially simultaneously, or the blocks may be executed in reverse order according to their corresponding functions.

[0043] In this embodiment, the term "part" refers to a software or hardware component, such as an FPGA or ASIC, and the "part" performs certain roles. However, the meaning of "part" is not limited to software or hardware. The "part" may be configured to reside in an addressable storage medium or configured to run one or more processors. Thus, as an 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 play one or more CPUs within the device or secure multimedia card.

[0044] In the present disclosure, modifiers such as "first," "second," etc., referring to terms may be used to distinguish each term from one another when describing embodiments. The terms modified by modifiers such as "first," "second," etc., may refer to different objects. However, the terms modified by modifiers such as "first," "second," etc., may refer to the same object. That is, modifiers such as "first," "second," etc., may be used to refer to the same object from different perspectives. For example, modifiers such as "first," "second," etc., may be used to distinguish the same object in terms of function or operation. For example, the first user and the second user may refer to the same user.

[0045] Specific terms used in the following description are provided to aid in understanding the present disclosure, and the use of such specific terms may be modified in other forms without departing from the technical spirit of the present disclosure. Since 5G New Radio (NR) communication systems must be able to freely reflect the various requirements of service providers and users, services that satisfy various requirements simultaneously must be supported. Services considered for 5G communication systems may include enhanced Mobile Broadband (eMBB), massive Machine Type Communication (mNTC), and Ultra Reliability Low Latency Communication (URLLC).

[0046] eMBB aims to provide data transmission speeds that are superior to those supported by existing LTE, LTE-A, or LTE-Pro. For example, in a 5G communication system, eMBB must be able to provide a peak data rate of 20 Gbps on the downlink and 10 Gbps on the uplink from the perspective of a single base station. Furthermore, while providing these peak data rates, the 5G communication system must also provide an increased user-perceived data rate. To satisfy these requirements, it necessitates improvements in various transmission and reception technologies, including enhanced Multi-Input Multi-Output (MIMO) transmission technology. Additionally, while current LTE transmits signals using a maximum bandwidth of 20 MHz in the 2 GHz band, the 5G communication system can meet the data transmission speeds required by using a frequency bandwidth wider than 20 MHz in frequency bands of 3–6 GHz or above 6 GHz.

[0047] At the same time, mMTC is being considered to support application services such as the Internet of Things (IoT) in 5G communication systems. To efficiently provide IoT, mMTC requires support for the connection of a large number of terminals within a cell, improved terminal coverage, enhanced battery life, and reduced terminal costs. Since IoT provides communication functions by attaching to various sensors and devices, a large number of terminals within a cell (e.g., 1,000,000 terminals / km²) 2It must be able to support mMTC. In addition, terminals supporting mMTC require wider coverage compared to other services provided by the 5G communication system, as terminals supporting mMTC are likely to be located in dead zones where cells cannot cover, such as building basements, due to the nature of the service. Terminals supporting mMTC must consist of low-cost devices, and since it is difficult to frequently replace the terminal's battery, a very long battery life of 10 to 15 years is required.

[0048] Finally, URLLC is a mission-critical cellular-based wireless communication service. For example, consider services used for remote control of robots or machinery, industrial automation, unmanned aerial vehicles, remote health care, and emergency alerts. Therefore, the communication provided by URLLC must offer very low latency and very high reliability. For instance, services supporting URLLC must satisfy an air interface latency of less than 0.5 milliseconds, and simultaneously 10 -5 The following packet error rate requirements apply. Therefore, for services supporting URLLC, 5G systems must provide a Transmit Time Interval (TTI) smaller than other services, and at the same time, design requirements are needed to allocate a wide resource in the frequency band to ensure the reliability of the communication link.

[0049] To satisfy such diverse services, it may be necessary to support beam management or various frequency bands. In this case, various channel environments may exist depending on the beam or frequency band to be supported. Regarding these diverse channel environments, the terminal can perform a channel estimate and report information on the estimated channel state to the base station to provide feedback on Channel State Information (CSI).

[0050] In the present disclosure, CSI may include at least one of a rank indicator (RI), a precoding matrix indicator (PMI), or a channel quality indicator (CQI). Here, RI may represent a terminal's recommended value frequency for the number of layers desired to be used for downlink data transmission. Here, PMI may represent a terminal's recommended value for a precoding matrix desired to be used for downlink data transmission. A combination of PMI values ​​may indicate one of the codebook matrices. In this case, PMI may be reported in the structure of a type 1 codebook or a type 2 codebook. Here, CQI may represent the highest modulation and coding scheme (MCS) value for a physical downlink shared channel (PDSCH) transmission that can be received with a block error rate (BLER) of 10% or less based on the recommended RI and PMI. Here, CQI may include channel quality information. In the present disclosure, CSI may mean any one of channel information, channel status information, and information about the channel status.

[0051] The terminal can receive a CSI reference signal (RS) from the base station. Upon receiving the CSI-RS, the terminal can estimate the channel. Based on the result of the channel estimation, the terminal can obtain the CSI.

[0052] The terminal can transmit the acquired CSI to the base station to perform feedback on the CSI.

[0053] Figure 1 is a diagram illustrating a procedure for a terminal to perform feedback on CSI to a base station.

[0054] In order for a base station (or NW; network or BS; base station) to know the downlink channel estimation result of a terminal (or UE; user equipment), a separate feedback process may be required to transmit channel information from the terminal to the base station.

[0055] The base station may transmit a reference signal to the terminal for channel estimation. The reference signal may include a CSI-reference signal (RS).

[0056] The terminal can perform channel estimation based on the CSI-RS received from the base station. The terminal can transmit the CSI obtained through the channel estimation result to the base station to perform feedback on the CSI.

[0057] The base station can transmit downlink data to the terminal based on the CSI received from the terminal.

[0058] The performance of a MIMO system (or multi-user (MU)-MIMO system) can be affected by the quality of CSI feedback. Specifically, the quality of CSI feedback is related to the accuracy or quantity of information; the quality of CSI feedback is considered better when there is less information loss and accurate CSI is transmitted from the terminal to the base station. The better the quality of CSI feedback, the better the performance of the MIMO system can be. Therefore, the more accurately the CSI feedback provided by each terminal is transmitted to the base station, the more the performance of the MIMO system can be improved. For this reason, more accurate CSI feedback may be required in future communication systems.

[0059] To perform more accurate CSI feedback in existing CSI feedback systems, one might consider increasing the size of the resources used for CSI feedback (e.g., the number of bits). However, increasing the number of bits may increase the number of base station antennas, bandwidth, or granularity required to perform CSI feedback. Consequently, the overhead may increase significantly.

[0060] In this regard, the following describes the procedure for CSI feedback based on the codebook method and the limitations of the CSI feedback procedure based on the codebook method.

[0061] In 5G NR, in order to reduce the size (e.g., number of bits) of CSI resources transmitted by a terminal to a base station to perform CSI feedback, reporting based on a codebook method can be performed.

[0062] Figure 2 is a diagram illustrating the procedure for performing CSI feedback based on a codebook.

[0063] First, the base station can transmit CSI-RS to the terminal. The terminal can perform channel estimation based on the received CSI-RS, and as a result, information about the estimated channel (e.g., Alternatively, you can obtain Input CSI.

[0064] The terminal has information about the acquired estimated channel (e.g., Or based on the Input CSI), in Step 1, the base station may obtain channel information related to the precoding matrix to be used for downlink transmission to the terminal. Specifically, the channel information related to the precoding matrix may include a precoding matrix indicator (PMI). Here, the PMI may be an indicator indicating any one of at least one precoding matrix (or precoder) included in the codebook. The terminal transmits the channel information related to the precoding matrix (e.g., PMI) to the base station using an estimated channel (e.g., Eigenvalue decomposition (EVD) can be performed on ).

[0065]

[0066] Specifically, eigenvalue decomposition is This can be achieved through an equation. Here, Λ represents information about channel strength, which may refer to the signal to be transmitted in each beam direction or the relative strength of the signal along the main path of the channel. Here, V represents information about the eigenvector, which may refer to the spatial direction in which data will be transmitted from the transmitting antenna.

[0067] The terminal can transmit channel information (e.g., PMI) related to the precoding matrix to the base station. At this time, the terminal can transmit channel information for the precoding matrix based on a type 1 or type 2 codebook method. At this time, the terminal (3 rdChannel information for a precoding matrix can be transmitted based on a pre-defined codebook specified in the 3GPP technical specification (TS) 38.214 (generation partnership project). Specifically, Type 1 or Type 2 codebook methods may include Rel-16 Type 2 (e.g., eType 2) codebook methods based on spatial and frequency domains. Unlike Type 2 codebooks that generally provide channel information feedback in the spatial domain, eType 2 codebooks that report in the frequency domain using channel sparsity in the frequency domain can provide more precise channel information feedback compared to Type 2 codebooks because they utilize information feedback regarding phase or amplitude in more granular subband domains.

[0068] In step 2, the terminal may transmit the CSI to the base station to perform feedback on the CSI. In this case, the CSI may include a codebook-based PMI. At this time, the CSI may be referred to as an implicit CSI.

[0069] In step 3, the base station can acquire or determine the precoder indicated by the CSI based on the transmitted CSI based on the codebook method.

[0070] In step 4, the base station can transmit downlink data to the terminal based on the determined precoder.

[0071] As described above, a base station can receive a CSI corresponding to the downlink channel estimation result of the terminal from the terminal based on a codebook method. Upon receiving the CSI, the base station can perform recovery or reconstruction of the channel state based on the received CSI. However, when receiving the CSI based on a codebook method, there may be a problem of information loss in the recovered channel state.

[0072] Specifically, because the codebook itself is limited in granularity (e.g., limited discrete Fourier transform (DFT) beam count, amplitude information, phase information, etc.), there may be a difference between the precoder indicated by the PMI and the actual channel state between the base station and the terminal. This may result in performance degradation of the MIMO system.

[0073] To address such information loss issues, the codebook can be configured to more accurately simulate channel states by methods such as increasing the number of precorders included in the codebook. However, configuring the codebook in this manner may increase the amount of data that needs to be transmitted, which can lead to signal overhead.

[0074] To address the problem of increased overhead as described above, explicit CSI feedback methods are being actively discussed. An explicit CSI feedback method may refer to a method in which the terminal transmits the channel state (e.g., V) itself acquired to the base station, or compresses the channel state and transmits it to the base station.

[0075] In an explicit CSI feedback method, compression of the CSI can be performed using artificial intelligence (AI) or machine learning (ML). The terminal can utilize AI / ML to compress the acquired CSI and transmit or send it to the base station in the form of a low-dimensional vector (e.g., v). The base station can restore or reconstruct the original CSI from the compressed CSI.

[0076] Figure 3a is a diagram illustrating AI / ML-based CSI compression technology.

[0077] The base station transmits CSI-RS to the terminal, and the terminal can perform a channel estimate based on the received CSI-RS, and as a result, the estimated channel (e.g., You can obtain ).

[0078] The terminal is the estimated channel (e.g., For ), pre-processing (e.g., eigenvalue decomposition (EVD) or singular value decomposition (SVD)) can be performed to obtain channel information (e.g., V) for the precoding matrix.

[0079] Specifically, when pre-processing is performed through eigenvalue decomposition (EVD), H is It can be decomposed as shown, and through this, the terminal can obtain V. Here, can mean a Hermite matrix, and V is the eigenvector matrix It can provide spatial direction information of, and A is The energy in each spatial direction can be represented by the eigenvalue of, and can represent the conjugate transpose value.

[0080] In addition, if pre-processing is performed through Singular Value Decomposition (SVD), H is It can be decomposed as shown, through which the terminal can obtain V. Here, U is the Left Singular Vector Matrix, which can represent the column space of H, and D is the singular value of H, which can represent the channel energy, and can represent the row space of H.

[0081] Reference numeral 300 denotes an AI / ML-based autoencoder (AE). Here, the autoencoder may include an encoder and a decoder. In this case, the terminal may include an encoder, and the base station may include a decoder. In the present disclosure, the encoder may refer to the terminal, and the decoder may refer to the base station. However, the base station may include a base station-side encoder in addition to the decoder.

[0082] The terminal of the AE can use the CSI obtained through pre-processing (e.g., Input CSI or V) as input information. The encoder can perform compression on the received CSI, and as a result, obtain the compressed CSI (e.g., z). Here, compression can be performed based on AI.

[0083] A terminal that has acquired a compressed CSI (e.g., z) may transmit the compressed CSI (e.g., z) to a base station to perform feedback on the CSI. In this case, the terminal may transmit the compressed CSI (e.g., z) to the base station via an uplink (UL). Here, the compressed CSI (e.g., z) may be in the form of a compressed vector (e.g., a latent vector or a feature vector).

[0084] A base station that receives a compressed CSI (e.g., z) from a terminal can perform reconstruction or restoration on the compressed CSI. As a result, the base station obtains a restored CSI (e.g., ) can be obtained. Here, the restored CSI may refer to the AE result (output).

[0085] In the present disclosure, AE may mean a two-sided model. A two-sided model may mean that two vendors are involved in training a single model such as the AE. Here, a vendor may mean an entity that manufactures a terminal or a base station.

[0086] Specifically, to train a model to perform CSI feedback, an encoder that performs compression on the CSI may be located at the terminal, and a decoder that performs restoration on the compressed CSI based on the CSI feedback may be located at the base station.

[0087] Figure 3b is a diagram illustrating the AI / ML-based CSI compression and CSI restoration procedures.

[0088] In step S310, the base station may transmit a reference signal for CSI. The reference signal may include a CSI-reference signal (RS).

[0089] In step S320, the terminal can perform encoder inference based on the reference signal.

[0090] For the above encoder inference, the terminal can specifically perform a channel estimate based on the received CSI-RS, and the estimated channel (e.g., ) can be obtained. The terminal can obtain the estimated channel (e.g., For ), pre-processing (e.g., eigenvalue decomposition (EVD) or singular value decomposition (SVD)) can be performed to obtain channel information (e.g., V) for the precoding matrix.

[0091] The encoder of the terminal can use CSI obtained through pre-processing (e.g., Input CSI or V) as input information. The encoder can perform compression on the received CSI, and as a result, obtain a compressed CSI (e.g., z). Here, the compressed CSI may be in the form of a compressed vector (e.g., latent vector or feature vector). Here, compression may refer to CSI compression based on AI / ML. Here, the compressed CSI may refer to the result of encoder inference (e.g., output z).

[0092] The terminal that has acquired the compressed CSI in step S330 may transmit the compressed CSI (e.g., output z) to the base station to perform feedback on the CSI. At this time, the terminal may transmit the compressed CSI to the base station via the uplink (UL).

[0093] In step S340, the base station may perform decoder inference based on the received compressed CSI. Here, the base station may include a decoder.

[0094] For the above decoder inference, the base station decoder may use the compressed CSI (e.g., z) as input information to perform reconstruction or restoration on the compressed CSI. As a result, the base station or decoder obtains the restored CSI (e.g., You can obtain ).

[0095] Figure 4 is a diagram illustrating the input and output of an encoder and a decoder.

[0096] The encoder may refer to an encoder included in a terminal. The decoder may refer to a decoder included in a base station. However, the present disclosure is not limited thereto, and both the encoder and the decoder may be included or located within a single terminal. Additionally, both the encoder and the decoder may be included or located within a single base station. In this case, there may be one or more encoders. Additionally, there may be one or more decoders.

[0097] Compression can be performed in the encoder. Specifically, compression can be performed on the input information (e.g., encoder input) in the encoder, and as a result, z can be obtained as output information.

[0098] The output information of the encoder (e.g., z) can be the input information of the decoder. In the decoder, reconstruction or restoration may be performed. If the encoder's output information is used as the input information of the decoder, restoration of the compressed information (e.g., z) may be performed. As a result, the decoder output information (e.g., decoder output) can be obtained.

[0099] Through an auto-encoder, the terminal can perform encoding on the CSI it intends to transmit to the base station. Here, encoding may refer to compression. The terminal that has performed encoding can transmit the compressed CSI to the base station. Upon receiving the compressed CSI, the base station can perform decoding on the compressed CSI. Here, decoding may refer to restoration or reconstruction.

[0100] Figure 5 is a diagram illustrating the process of a terminal compressing CSI based on AI and feeding it back to a base station.

[0101] The base station (NW) can transmit CSI-RS to the terminal (UE). The terminal can perform channel estimation based on the received CSI-RS, and as a result, information about the estimated channel (e.g., Alternatively, you can obtain Input CSI.

[0102] The terminal provides information about the estimated channel (e.g., For ), pre-processing (e.g., eigenvalue decomposition (EVD) or singular value decomposition (SVD)) can be performed to obtain channel information (e.g., V) for the precoding matrix. Here, the precoding matrix may refer to an eigenvector matrix. In this case, eigenvectors may exist for each frequency subband of the terminal.

[0103] In Step 1, the terminal can perform compression on the CSI obtained through pre-processing (e.g., Input CSI or V). Here, the compression can be performed based on AI (or AI / ML), and the terminal may include an encoder. Specifically, AI / ML-based CSI compression may be a method for compressing channel information in the spatial and frequency domains.

[0104] Here, the encoder of the terminal may be, for example, an encoder using a neural network (e.g., a neural encoder). Here, a neural network may refer to a model that provides output information by performing calculations on input information using a neural network based on artificial intelligence. In this case, the model used by the terminal is a blind model, so the base station or the base station's decoder may not know what compression method is used or what AI it is based on.

[0105] As a result, the terminal can obtain a compressed CSI (e.g., z or explicit CSI). Here, the compressed CSI may be in the form of a compressed vector (e.g., latent vector or feature vector).

[0106] In step 2, the terminal may transmit a CSI (e.g., z or explicit CSI) to a base station to perform feedback on the CSI. Here, the base station may include a decoder.

[0107] Here, the base station decoder may be, for example, a decoder using a neural network (e.g., a neural decoder). Here, a neural network may refer to a model that provides output information by performing calculations on input information using a neural network based on artificial intelligence. In this case, the model used by the base station is a blind model, so the terminal or the terminal's encoder may not know what restoration method is used or what AI it is based on.

[0108] In Step 3, the base station can perform reconstruction on the received compressed CSI. As a result, the base station obtains the reconstructed CSI (e.g., Alternatively, you can obtain the Output CSI.

[0109] In step 4, the base station can perform decoding on the received CSI to acquire or determine a precoder.

[0110] In step 5, the base station can transmit downlink data to the terminal based on the determined precoder.

[0111] Recently, methods and devices for improving the feedback quality of CSI by reducing information loss caused by compression (and decompression) when performing compression (and decompression) on CSI are being discussed. Specifically, to increase the performance of CSI compression (and decompression), AI (or AI / ML)-based CSI compression (and decompression) can be performed.

[0112] Figure 6 is a diagram illustrating a two-sided model that performs vendor-specific training.

[0113] In the present disclosure, the two-sided model may refer to an AE. The two-sided model may refer to a form in which an encoder that performs compression on the CSI to perform CSI feedback, as disclosed in FIG. 3a, exists in the terminal, and a decoder that performs restoration on the compressed CSI based on the CSI feedback exists in the base station. Here, the vendor may refer to the entity manufacturing the terminal or the base station.

[0114] To train a single AI model (e.g., a single AE), at least one entity among the terminal-side vendor (UE vendor) and the base station-side vendor (NW vendor) may be involved, and this can be referred to as an inter-vendor collaboration issue. Specifically, when training or learning of an AI model is performed by a single entity (e.g., a single vendor), the model can be transmitted to another vendor through model transfer.

[0115] One or more vendors (e.g., UE1, UE2, and base stations) may be involved in training a single AI model. Specifically, each vendor may exchange a data set to train a single AI model. This may differ from exchanging the model itself (e.g., a reference model, the model's structure, or the model's parameters). Here, the data set for training the model may include information regarding the format of the data.

[0116] To improve the quality of AI-based (or AI / ML-based) CSI compression (and restoration), methods to increase the performance of the AI ​​model itself may be discussed. However, increasing the performance of the AI ​​model can lead to the problem of increased complexity. Specifically, there are limits to performance improvement, and the system overhead may increase relative to the enhanced performance. Due to these issues, it may be difficult to implement models exceeding a certain size.

[0117] To solve the above problems, the present disclosure may consider five methods (option) for training a model. Here, the model may refer to an AI model or a model that performs compression (and restoration) based on AI.

[0118] The five methods may be as follows.

[0119] - Option 1: Fully standardized reference model (structure and parameters)

[0120] The first option proposes a method for defining a reference model in the standard. Base stations and terminals can perform model training based on the reference model defined in the standard. Here, the reference model can be defined based on structure and parameters.

[0121] - Option 2: Standardized dataset

[0122] The second option proposes a method for defining a dataset for training a model in the standard. In this case, the base station and the terminal can perform model training based on the dataset defined in the standard.

[0123] - Option 3: Standardized reference model structure and parameter exchange between NW-side and UE-side

[0124] The third option proposes a method for defining the structure of the reference model in the standard. In this case, the base station and the terminal can perform model training based on the structure of the reference model defined in the standard and the exchanged parameters.

[0125] - Option 4: Standardized data or dataset format and dataset exchange between NW-side and UE-side

[0126] The fourth option proposes a method for defining data (or data sets) or the format of a data set in the standard. Specifically, the standard may define which data (or data sets) are to be transmitted (or exchanged) to perform model training. For example, the standard may define exchanging a data set containing CSI (e.g., target CSI or V) and CSI compressed based on quantization, and performing model training based on the exchanged data set. The base station side and the terminal side may perform model training based on the data (or data sets) or the format of the defined data set and the exchanged data set defined in the standard. For example, the format of the data set may include data types (e.g., fp32, fp16, int8, in4, etc.).

[0127] - Option 5: Standardized model format and reference model exchange between NW-side and UE-size

[0128] Option 5 proposes a method for defining the model format in the standard. Base stations and terminals can perform model training based on the model format defined in the standard and an exchanged reference model.

[0129] Figure 7 is a diagram illustrating the data set exchange and model learning procedure between a terminal and a base station.

[0130] Figure 7 is a diagram illustrating the procedure related to standardized data or dataset format and dataset exchange between the base station side and the terminal side, corresponding to the fourth option among the five methods for training a model.

[0131] Specifically, the present disclosure assumes that a data set is defined in a standard. In this case, the data set may be characterized by including a CSI (e.g., target CSI or V) and a compressed CSI (e.g., a CSI compressed based on quantization) for performing CSI feedback. The base station side and the terminal side may exchange the data set defined in the standard (e.g., CSI and compressed CSI).

[0132] In the present disclosure, the model may refer to an apparatus for performing CSI feedback, comprising an encoder (and a decoder). Specifically, the model in the present disclosure may perform compression (and decompression) of CSI using AI or based on AI. In this case, the model may aim to improve the quality of feedback on CSI by reducing information loss caused by CSI compression (and decompression). To this end, the model may perform compression on input information. Additionally, the model may perform decompression on the compressed information.

[0133] In the present disclosure, a training procedure refers to a process of performing learning or training on a model. Specifically, information regarding input and output information may be exchanged, and learning or training on the model may be performed so that the same output information can be obtained for the same input information. For example, in the case of a model that performs compression, learning or training may be performed so that when compression is performed on received input information (e.g., original information), the compression is performed in the same way as the received output information (e.g., compressed information). Additionally, for example, in the case of a model that performs restoration, learning or training may be performed so that when restoration is performed on received input information (e.g., compressed information), the restoration is performed in the same way as the received output information (e.g., restored information).

[0134] Reference numeral 710 illustrates a model training procedure at the base station side. Here, the model may include an encoder that performs compression and a decoder that performs restoration based on an AI model or AI.

[0135] A base station may include either an encoder (e.g., a NW encoder or an encoder) or a decoder (a NW decoder or a decoder). Alternatively, a base station may include both an encoder and a decoder. An encoder is a device that performs encoding and may perform compression on input information. A decoder is a device that performs decoding and may perform restoration or reconstruction on input information.

[0136] Specifically, the base station encoder can use CSI (e.g., target CSI or V) as input information. Here, CSI is the output information of the base station encoder (e.g., When ) is used as input information for the base station decoder, it can represent the target CSI that the base station decoder intends to restore identically. In this case, the more identically it restores (i.e., the smaller the difference between the output information of the base station decoder and the input information of the base station encoder), the better the performance of the base station model can be evaluated.

[0137] A base station encoder using CSI (e.g., target CSI or V) as input information can perform compression on the CSI to train the terminal encoder. In this case, the compression can be performed based on quantization. As a result, the base station encoder [receives] the CSI compressed based on quantization (e.g., ) can be obtained as output information. In this case, 'q' may mean quantized.

[0138] The base station decoder uses compressed CSI based on quantization corresponding to the result of the base station encoder (e.g., ) can be used as input information. Compressed CSI based on quantization (e.g., A base station-side decoder using ) as input information can perform reconstruction or restoration on the compressed CSI. In this case, reconstruction or restoration can be performed based on de-quantization. As a result, the base station-side decoder obtains the reconstructed or restored CSI (e.g., ) can be obtained as output information.

[0139] At this time, the CSI (e.g., target CSI or V) and the restored CSI (e.g., The performance of the model can be evaluated through a comparison of ). Specifically, the more identically the reconstruction is performed (i.e., the smaller the difference between the output information of the base station decoder and the input information of the base station encoder), the better the performance of the base station model can be said to be.

[0140] The model training or learning process at the base station may be performed before exchanging the dataset with the terminal. Alternatively, the model training or learning process at the base station may be performed before CSI compression for CSI feedback at the terminal is performed.

[0141] In step 720, the base station may transmit or deliver a data set to the terminal. This procedure may mean an exchange of data sets between the base station and the terminal.

[0142] Here, the dataset may include CSI and CSI compressed based on quantization. For example, the dataset is It can be composed of. As mentioned above, the CSI may refer to the target CSI.

[0143] Reference numeral 730 illustrates a model training procedure at the terminal side. Here, the model may include an AI model or an encoder that performs compression based on AI.

[0144] The terminal may include an encoder (e.g., a UE encoder or encoder). The encoder is a device that performs encoding and can perform compression on the input information.

[0145] The terminal can perform training on a terminal-side encoder for performing CSI feedback based on a data set received from a base station.

[0146] Specifically, the terminal-side encoder may use CSI (e.g., target CSI or V) from the received data set as input information. In this case, the terminal-side encoder using CSI as input information may perform compression on the CSI. In this case, the terminal-side encoder may use the compressed CSI (e.g., based on quantization generated by the base station-side encoder) It can be trained to acquire or generate ). In this case, compression can be performed based on quantization. As a result, the terminal-side encoder can generate compressed CSI (e.g., based on quantization) ) can be obtained as the output. Here, q' may mean quantized.

[0147] When compression is performed based on quantization, the size of the CSI feedback resource is reduced, which can reduce overhead. For example, CSI feedback can be transmitted in the format of floating-point values. In this case, the size of the feedback can consist of 100 fp 32 values. If quantization is not performed on the CSI feedback (e.g., target CSI or V), 3200 bits may be required to perform the feedback. However, in the case of CSI feedback that has been quantized (e.g., 2-bit quantization performed) (e.g., CSI compressed based on quantization), only 200 bits may be required to perform the feedback. Compared to the case of CSI feedback that has not been quantized, the overhead of the feedback can be significantly reduced in the case of quantized CSI feedback due to the reduction in the number of bits required.

[0148] However, in the case of quantized CSI feedback, information loss may occur compared to non-quantized CSI feedback. This can lead to a decrease in the accuracy of model training. Consequently, a decline in model performance may occur.

[0149] FIG. 8 is a diagram illustrating a terminal-side encoder considering quantization.

[0150] In order to solve the problem of reduced performance of the aforementioned AL / ML model, in one embodiment of the present disclosure, when the terminal-side encoder performs CSI compression, information loss due to quantization can be taken into account.

[0151] In one embodiment of the present disclosure, a data set transmitted by a base station to a terminal to perform training for a terminal-side encoder includes a CSI (e.g., target CSI or V) and a CSI compressed based on quantization (e.g., It proposes a method to transmit additional information as well as )

[0152] In one embodiment of the present disclosure, a CSI (e.g., target CSI or V) and a compressed CSI based on quantization (e.g., A data set including ) can be defined as a first data set. Additionally, in one embodiment of the present disclosure, additional information exchanged to reduce information loss due to quantization can be defined as a second data set. Specifically, the second data set is CSI that has been compressed based on artificial intelligence (AI) but has not been compressed based on quantization (e.g., CSI compressed based on AI or It may include ). Or, additional information may include information about a quantizer. Here, information about a quantizer may include any one of a quantizer, a step size of scalar quantization, or a codebook of vector quantization. Here, the quantizer may refer to a quantizer used in the base station encoder or a quantizer included in the base station-side encoder.

[0153] Both the first data set and the second data set may refer to information for training a terminal-side model. Although the present disclosure distinguishes between the first data set and the second data set, they do not necessarily need to be treated separately. The first data set and the second data set may be included in a single data set for exchange or transmission.

[0154] Figure 8 illustrates the terminal-side encoder training procedure.

[0155] Reference numeral 810 illustrates a training procedure for a terminal-side encoder. To perform training for the terminal-side encoder, the terminal may receive a first data set and a second data set. Upon receiving these, the terminal may perform training for the terminal-side encoder based on the first data set and the second data set.

[0156] Specifically, to perform training on the terminal-side encoder, the terminal encoder may use CSI (e.g., target CSI or V) from the received dataset as input information. The terminal-side encoder using CSI as input information may perform compression on the CSI to perform encoder training. In this case, compression may refer to the encoding process. As a result of compression, the terminal-side encoder [receives] the compressed CSI (e.g., based on quantization ) can be obtained as the result (output). Here, q' may mean quantization. The encoding process is described in detail below.

[0157] Reference numeral 820 illustrates a more detailed training procedure for a terminal-side encoder according to one embodiment of the present disclosure.

[0158] The terminal-side encoder may include a compressor and a quantizer. Here, the compressor is a component that performs compression and may refer to a model (or AI model) that performs compression based on AI. The AI ​​model may use a neural network. Here, a neural network may refer to a model that provides output information by performing calculations on input information using a neural network based on artificial intelligence. In this case, the base station may not know which compression method is used or which AI is based on it. Here, the quantizer may refer to a component that performs quantization.

[0159] To perform training for a terminal-side encoder, the terminal may receive a first data set and a second data set from a base station. Upon receiving these, the terminal may perform training for the terminal-side encoder based on the first data set and the second data set.

[0160] Specifically, training can be performed in the compression unit and the quantizer to perform training for the terminal-side encoder.

[0161] In the compression section, when CSI (e.g., target CSI or V) from the first data set is used as input information, CSI compressed based on AI from the second data set (e.g., It can be trained to obtain ) as the output. Here, 'raw' may mean that it is not quantized. In this case, only AI-based compression can be performed in the compression unit. That is, compression based on quantization may not be performed in the compression unit.

[0162] In the quantizer, AI-based compressed CSI from the second dataset (e.g., When ) is used as input information, among the first data set, the CSI compressed based on quantization (e.g., ) can be obtained as result information. Or, in the quantizer, output information of the compression unit (e.g., CSI compressed based on AI or When ) is used as input information, among the first data set, the CSI compressed based on quantization (e.g., ) can be obtained as result information. Here, 'q' may mean quantization. In this case, compression based on quantization can be performed in the quantizer. In this case, quantization can be performed based on information about the quantizer in the second data set.

[0163] That is, at the terminal-side encoder, compressed CSI based on quantization based on the first data set and the second data set (e.g., Not only ) but also AI-based compressed CSI (e.g., Training can be performed to acquire ).

[0164] However, the terminal-side encoder training is not limited to the above embodiments, and training for the terminal-side encoder may be performed through embodiments disclosed thereafter. Alternatively, training for the terminal-side encoder may be performed through a combination of the above embodiments and subsequent embodiments, or through a combination of subsequent embodiments.

[0165] For the data sharing and model training method and apparatus considering quantization presented in the present disclosure, one embodiment of the present disclosure proposes a signaling method related to data exchange and an apparatus for signaling.

[0166] FIG. 9a is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to one embodiment of the present disclosure.

[0167] In step 910, the terminal can report its UE capability to the base station.

[0168] Specifically, the terminal may report to the base station the terminal's capability related to terminal-side encoder training or terminal-side encoder inference. In this case, the report may be transmitted in the form of a message, and said message may include information regarding the terminal's capability related to terminal-side encoder training or inference.

[0169] Information or settings regarding the capabilities of the terminal may be as follows. Alternatively, a message for reporting terminal capabilities may include at least one of the following information or settings.

[0170] - AI support status

[0171] Regarding the capabilities of the terminal, the terminal may transmit information regarding AI support or an indicator indicating AI support to the base station. In this case, AI support may be determined based on the level of implementation of the terminal. In this case, the terminal may not transmit information regarding which AI it supports to the base station. The base station may not be able to know which AI the terminal supports.

[0172] - Capability related to AI computing

[0173] Information regarding capabilities related to AI computing may include either Input / Output (I / O) memory bandwidth or information related to AI-driven Floating Point Operations Per Second (FLOPS).

[0174] Here, I / O memory bandwidth can represent information about how quickly a terminal can transfer data from memory to the processor.

[0175] Here, AI-driven FLOPS can represent information about how many floating-point operations a terminal can process per second.

[0176] - Capability related to AI model size

[0177] Information regarding capability related to model size may include any one of information related to AI model parameters, memory, or storage.

[0178] Here, information related to AI model parameters can, for example, indicate the number of parameters that the terminal can process.

[0179] Here, information related to memory may represent a model that the terminal has stored in memory or is currently running while stored in memory.

[0180] Here, information related to storage space may represent information about the storage space available on the terminal.

[0181] - Information on supported model representation formats (MRF)

[0182] In this case, information regarding the supported MRF may refer to information related to the Open Neural Network Exchange (ONNX). Here, ONNX may be an example of a model format that the terminal supports or can process.

[0183] Information related to terminal capabilities may be categorized and transmitted or sent. For example, information related to terminal capabilities may be transmitted or sent in the form of a table index.

[0184] Table 1 below shows an example of a case where information related to terminal capabilities is classified. Specifically, it shows information on FLOPS and memory by category of terminal capabilities. Here, the category of terminal capabilities may refer to capabilities related to AI computing capabilities. Here, memory refers to capabilities related to AI model size and may represent information on the storage space available on the terminal.

[0185] Terminal Capability Category FLOPSAI memory (MB) Category 0 100 Categories 1 Category 2 200 Categories 3 Category 4 500

[0186] When information related to terminal capability is classified and transmitted, the terminal capability report message may indicate which category the terminal belongs to through the message. For example, the terminal may indicate to the base station that it belongs to Category 0, and in this case, the base station [indicates] that the terminal per second It can be seen that OPS processing is possible and the storable memory is 100MB. For example, the terminal can instruct the base station that it belongs to Category 1, and at this time, the base station [instructs] the terminal per second It can be seen that OPS processing is possible and the storable memory is 100MB. For example, the terminal can instruct the base station that it belongs to Category 2, and at this time, the base station [instructs] the terminal per second It can be seen that OPS processing is possible and the storable memory is 200MB. For example, the terminal can instruct the base station that it belongs to Category 3, and at this time, the base station [instructs] the terminal per second It can be seen that OPS processing is possible and the storable memory is 200MB. For example, the terminal can instruct the base station that it belongs to Category 4, and at this time, the base station [instructs] the terminal per second It can be seen that OPS processing is possible and the storage memory is 500MB.

[0187] In step 920, the terminal may transmit a request message to the base station to request a second data set. This step is optional, and even if the terminal does not transmit a request message for the second data set to the base station, the base station may transmit a data set containing the second data set to the terminal.

[0188] When a terminal transmits a request message for a second data set to a base station, it may transmit it in the form of a 1-bit message. Specifically, the 1 bit may be transmitted in the form of an indicator. For example, if the bit is set to 0, the terminal may indicate to the base station that the second data set is not needed, and if the bit is set to 1, the terminal may request the second data set from the base station. Alternatively, it is possible to transmit the message with the bit set to 1 only when the second data set is requested, and not transmit the message when the second data set is not requested. Alternatively, if the message bit is set to 1, the terminal may indicate to the base station that the second data set is not needed, and if the message bit is set to 0, the terminal may request the second data set from the base station.

[0189] Table 2 below illustrates an example where information related to terminal capabilities is transmitted or conveyed in the form of a table index. Here, the category of the second data set indicates the category number to which the terminal belongs. Here, the list of the second data set indicates a list of second data sets that the terminal requests for terminal-side encoder training or requires for terminal-side encoder training, depending on the category to which the terminal belongs.

[0190] Category of the 2nd dataset List of the 2nd dataset Category 0x Category 1 {AI-based compressed CSI} Category 2 {Information on the quantizer} Category 3 {AI-based compressed CSI, Information on the quantizer} Category 4 {AI-based compressed CSI, Information on the quantizer, Target performance}

[0191]

[0192] When information related to the capabilities of a terminal is transmitted in the form of a table index, the terminal capability report message may indicate which category the terminal belongs to through the message. For example, the terminal may indicate to the base station that it belongs to category 0, and in this case, the terminal may indicate that a second data set is not needed (none). For example, the terminal may indicate to the base station that it belongs to category 1, and in this case, the terminal may request a second data set containing CSI compressed based on AI from the base station. For example, the terminal may indicate to the base station that it belongs to category 2, and in this case, the terminal may request a second data set containing information about a quantizer from the base station. For example, the terminal may indicate to the base station that it belongs to category 3, and in this case, the terminal may request a second data set containing at least one of CSI compressed based on AI and information about a quantizer from the base station. For example, the terminal may instruct the base station that it belongs to category 4, and the terminal may request the base station to provide a second data set containing at least one of AI-based compressed CSI, information about the quantizer, and information about the target performance.

[0193] In step 930, the base station may transmit a message to the terminal containing a first data set and a second data set. This process may imply a data exchange.

[0194] The base station may transmit a response message containing a second data set to the terminal in response to a request message. Alternatively, even without a request for the second data set from the terminal, the base station may transmit the second data set to the terminal by determining or deciding whether to transmit the second data set based on a report on the terminal's capabilities.

[0195] The method by which a base station transmits a first data set and a second data set to a terminal may be as follows. Alternatively, the method by which a base station signals the first data set and the second data set to a terminal may be as follows.

[0196] In the following, data may refer to either of the data included in the first data set and the second data set. In the following, a data set may refer to either the first data set or the second data set, or a data set including both the first data set and the second data set.

[0197] - Delivery of datasets online (Over-the-air or Online) or offline

[0198] The offline method may be preferred when the data size is large. For example, CSI (e.g., target CSI or V), AI-based compressed CSI (e.g., ) or compressed CSI based on quantization (e.g., The data may be large in size and may be transmitted offline. For example, a base station may instruct a terminal to download a data set. In this case, the terminal may download the data set in an offline manner. For example, the terminal may download the data set in an offline manner from the base station operator's memory device.

[0199] However, the above transmission method is merely an example, and the transmission method may not be limited to the size of the data set, and an online method may be used even if the data size is large. For example, a base station may transmit only the address of the server from which the data set will be downloaded to the terminal and instruct the terminal to download the data set. An online method may be used in which the terminal downloads the data in an online environment based on the address.

[0200] - Online (Over-the-air or Online) method

[0201] The online method may be preferred when the data size is small. For example, information related to target performance or quantizers may be small in size and can be transmitted to the terminal via the online method.

[0202] The format or type of the data set transmitted by the base station to the terminal may be as follows.

[0203] - Floating point values

[0204] When a base station transmits a data set to a terminal in the format of floating-point values, the floating-point values ​​may be, for example, fp32 or fp16. For example, in the case of fp32, the floating-point may mean 32-bit, and in the case of fp16, the floating-point may mean 16-bit. Using such formats may have the advantage of providing high precision and minimizing data loss.

[0205] - Quantized values

[0206] When a base station transmits a data set to a terminal using quantized values, the quantized values ​​may be, for example, int8, int4, int2, etc. For example, int8 may mean that the integer of the transmitted dataset is 8-bit, int4 may mean that the integer of the transmitted dataset is 4-bit, and int2 may mean that the integer of the transmitted dataset is 2-bit. Using this format may have the advantage of reducing memory usage and the consumption of resources (e.g., bandwidth) used for transmission, but it may have the disadvantage of some information loss.

[0207] - Deep Neural Network (DNN) framework for quantizers

[0208] The base station can transmit a data set to the terminal through a DNN framework. The framework can be, for example, Open Neural Network Exchange (ONNX).

[0209] The base station can inform the terminal whether the quantizer of the terminal-side encoder is trainable. Information indicating whether the quantizer of the terminal-side encoder is trainable may be transmitted together with or included in the dataset. Such information may be in the form of an indicator.

[0210] If the base station informs the terminal that the quantizer of the terminal-side encoder is untrainable, the terminal may use the quantizer received from the base station as is. Information indicating that training is untrainable may be transmitted in a form such as trainable=False. If the base station informs the terminal that the quantizer of the terminal-side encoder is trainable, the terminal may design and use a new quantizer. Information indicating that training is possible may be transmitted in a form such as trainable=True.

[0211] In step 940, the terminal can perform training on the terminal-side encoder. Based on the received first data set and second data set, the terminal can perform AI-based compressed CSI (e.g., ) or compressed CSI based on quantization (e.g., Training for the encoder can be performed to acquire any one of the following. Or compressed CSI based on AI (e.g., Compressed CSI based on ) and quantization (e.g., Training for the encoder can be performed to acquire all of them. Specific details regarding the above-mentioned terminal-side encoder training method will be described later.

[0212] FIG. 9b is a diagram illustrating the exchange of a first data set and a second data set on the terminal side and the terminal side encoder training procedure according to one embodiment of the present disclosure.

[0213] In step 955, the terminal can report its UE capability to the base station.

[0214] Specifically, the terminal may report to the base station the terminal's capability related to terminal-side encoder training or terminal-side encoder inference. In this case, the report may be transmitted in the form of a message, and said message may include information regarding the terminal's capability related to terminal-side encoder training or inference.

[0215] In step 960, the terminal may transmit a request message to the base station to request a second data set. This step is optional, and even if the terminal does not transmit a request message for the second data set to the base station, the base station may transmit a data set containing the second data set to the terminal.

[0216] In step 965, the terminal may receive a message from the base station containing a first data set and a second data set. This process may imply a data exchange.

[0217] In step 970, the terminal can perform training on the terminal-side encoder. Based on the received first data set and second data set, the terminal can perform AI-based compressed CSI (e.g., ) or compressed CSI based on quantization (e.g., Training for the encoder can be performed to acquire any one of the following. Or compressed CSI based on AI (e.g., Compressed CSI based on ) and quantization (e.g., Training for the encoder can be performed to acquire all of them.

[0218] FIG. 9c is a diagram illustrating a procedure for exchanging a first data set and a second data set on the base station side according to one embodiment of the present disclosure.

[0219] In step 975, the base station can receive a report of the terminal's capability (UE capability) from the terminal.

[0220] Specifically, the terminal may report to the base station the terminal's capability related to terminal-side encoder training or terminal-side encoder inference. In this case, the report may be transmitted in the form of a message, and said message may include information regarding the terminal's capability related to terminal-side encoder training or inference.

[0221] In step 980, the base station may receive a request message from the terminal to request a second data set. This step is optional, and the base station may transmit a message containing the second data set to the terminal even if it does not receive a request message from the terminal.

[0222] At step 985, the base station may determine or decide whether to transmit the second data. Specifically, the base station may determine whether to transmit the second data set based on a request message. For example, the base station may transmit a response message containing the second data set in response to the request message. Alternatively, without a request message from the terminal, the base station may determine or decide whether to transmit the second data set based on a report regarding the terminal's capabilities.

[0223] In step 990, the base station may transmit a message to the terminal containing a first data set and a second data set. This process may imply a data exchange.

[0224] FIG. 10 is a diagram illustrating the exchange of a first data set and a second data set and model training procedure between a terminal and a base station according to one embodiment of the present disclosure.

[0225] Reference numeral 1010 illustrates a model training procedure at the base station side. Here, the model may include an encoder that performs compression and a decoder that performs restoration based on an AI model or AI. The encoder and decoder of the base station may constitute an auto encoder.

[0226] Specifically, the base station encoder may use CSI (e.g., target CSI or V) as input information. The base station encoder using CSI (e.g., target CSI or V) as input information may perform compression on the CSI to train the terminal encoder. In this case, the final result of the CSI compression may be information output based on quantization. As a result, the base station encoder [receives] the CSI compressed based on quantization (e.g., ) can be obtained as output information. In this case, 'q' may mean quantized.

[0227] To perform training on the terminal-side encoder, the base station-side encoder uses AI-based compressed CSI (e.g., It can acquire or generate ) and deliver or transmit to the terminal.

[0228] Specifically, the base station encoder may include a compressor and a quantizer. Here, the compressor is a component that performs compression and may refer to a model (or AI model) that performs compression based on AI. The AI ​​model may use a neural network. Here, a neural network may refer to a model that provides output information by performing calculations on input information using a neural network based on artificial intelligence. In this case, the terminal may not be able to know which compression method is used or which AI is based on it. Here, the quantizer may refer to a component that performs quantization.

[0229] In the compression section, when CSI (e.g., target CSI or V) is used as input information, the AI-based compressed CSI (e.g., ) can be obtained or generated as an output. Here, 'raw' may mean that it has not been quantized. In this case, only AI-based compression can be performed in the compression unit. That is, compression based on quantization may not be performed in the compression unit.

[0230] In the quantizer, AI-based compressed CSI (e.g., When ) is used as input information, CSI compressed based on quantization (e.g., ) can be obtained or generated. Here, 'q' may mean quantization. In this case, compression based on quantization can be performed in the quantizer.

[0231] The base station decoder uses compressed CSI (e.g., based on quantization corresponding to the encoding result of the base station encoder) ) can be used as input information. Compressed CSI based on quantization (e.g., A base station-side decoder using ) as input information can perform reconstruction or restoration on the compressed CSI. In this case, reconstruction or restoration can be performed based on de-quantization. As a result, the base station-side decoder obtains the reconstructed or restored CSI (e.g., Training can be performed to acquire or generate ) as output information.

[0232] In step 1020, the base station may transmit or deliver a first data set and a second data set to the terminal. The procedure may mean the exchange of data sets between the base station side and the terminal side.

[0233] The first data set is a CSI (e.g., target CSI or V) and a compressed CSI based on quantization (e.g., It may include ). Or, the first data set may additionally include information about target performance.

[0234] The second dataset is AI-based compressed CSI (e.g., It may include at least one of the information about the quantizer.

[0235] Reference numeral 1030 illustrates a model training procedure at the terminal side. Here, the model may include an AI model or an encoder that performs compression based on AI.

[0236] The terminal can perform training for a terminal-side encoder to perform CSI feedback based on a first data set and a second data set received from a base station.

[0237] Specifically, training can be performed in the compression unit and the quantizer to perform training for the terminal-side encoder.

[0238] In the compression section, when CSI (e.g., target CSI or V) from the first data set is used as input information, CSI compressed based on AI from the second data set (e.g., It can be trained to obtain ) as the output. Here, 'raw' may mean that it is not quantized. In this case, only AI-based compression can be performed in the compression unit. That is, compression based on quantization may not be performed in the compression unit.

[0239] In the quantizer, AI-based compressed CSI from the second dataset (e.g., When ) is used as input information, among the first data set, the CSI compressed based on quantization (e.g., ) can be obtained as result information. Or, in the quantizer, output information of the compression unit (e.g., CSI compressed based on AI or When ) is used as input information, among the first data set, the CSI compressed based on quantization (e.g., ) can be obtained as result information. Here, 'q' may mean quantization. In this case, compression based on quantization can be performed in the quantizer. In this case, quantization can be performed based on information about the quantizer in the second data set.

[0240] That is, the terminal-side encoder according to the present disclosure is a compressed CSI based on quantization based on a first data set and a second data set (e.g., Not only ) but also AI-based compressed CSI (e.g., Training can be performed to acquire ).

[0241] In one embodiment of the present disclosure, the following three methods for performing model training are proposed. Through the following three methods, the present disclosure aims to minimize information loss. Specifically, the loss of information is to be minimized by minimizing the value of the loss function.

[0242] To perform model training through the following three methods, forward propagation and backward propagation processes may be carried out. Specifically, forward propagation may refer to the process of acquiring output information by inputting CSI (e.g., target CSI or V) as input information. The output information in the forward propagation process is a compressed CSI based on quantization (e.g., ) or AI-based compressed CSI (e.g., It may mean at least one of ). The output information from the forward propagation process can be used to calculate the value of the loss function.

[0243] The backpropagation process can refer to the process of updating a model based on information obtained through forward propagation. During the backpropagation process, the value of the loss function can be calculated to perform updates to the model. In this case, the value of the loss function can be calculated based on the output information from the forward propagation process.

[0244] Below, three methods are proposed to minimize the value of the loss function.

[0245] Method 1: Compressed CSI based on quantization generated by a base station-side encoder (e.g., Training can be performed on the terminal-side encoder to acquire or generate ) (Train UE-side encoder to generate quantized CSI feedback that NW-side encoder generates), and the specific operation for this will be explained through FIG. 11.

[0246] FIG. 11 is a diagram illustrating a first method for terminal-side encoder training.

[0247] The loss value to be minimized during training for the terminal-side encoder can be as follows.

[0248]

[0249] Here, It can refer to loss functions such as mean squared error (MSE) or mean absolute error (MAE).

[0250] Here, is CSI compressed based on quantization generated by the terminal-side encoder (e.g., It can mean ), is CSI compressed based on quantization generated by the base station encoder (e.g., It may mean ). Accordingly, the loss function for the first method above may mean the difference between the results of compressing the same CSI based on quantization between the base station and the terminal.

[0251] In the case of terminal-side encoder training for the first method, the terminal may perform training for the terminal-side encoder based on a first data set received from a base station. At this time, a second data set may not be considered.

[0252] Specifically, the terminal-side encoder can use CSI (e.g., target CSI or V) as input information. In this case, the terminal-side encoder using CSI as input information can perform compression on the CSI. In this case, the terminal-side encoder can use the compressed CSI (e.g., based on quantization generated by the base station-side encoder) It can be trained to acquire or generate ).

[0253] In this case, compression can be performed based on quantization. As a result, the terminal-side encoder compresses the CSI based on quantization (e.g., ) can be obtained as the output. Here, q' may mean quantized.

[0254] Second method: CSI compressed based on AI generated by the compression unit of the base station encoder (e.g., Training can be performed on the terminal-side compression unit so that the terminal-side encoder can acquire or generate raw CSI feedback (Train neural network of UE-side encoder to generate raw CSI feedback that neural network of NW-side encoder generates), and the specific operation for this will be explained through FIG. 12.

[0255] FIG. 12 is a drawing illustrating a second method for terminal-side encoder training according to one embodiment of the present disclosure.

[0256] The loss value to be minimized in the compression section of the terminal-side encoder can be as follows.

[0257]

[0258] Here, It may refer to a loss function such as mean squared error (MSE) or mean absolute error (MAE). Accordingly, the loss function for the second method described above may refer to the difference in the results of compressing the same CSI based on AI between the base station and the terminal.

[0259] Here, is CSI compressed based on AI generated by the compression unit of the terminal-side encoder (e.g., It can mean ), is CSI compressed based on AI generated by the compression unit of the base station-side encoder (e.g., It can mean ).

[0260] In the case of the terminal-side encoder training method for the second method, the terminal can perform training for the terminal-side encoder based on the first data set and the second data set received from the base station. However, in the case of the terminal-side encoder training method for the second method, compressed CSI based on the quantization included in the first data set (e.g., ) may not be considered. Or, information about the quantizer in the second data set may not be considered.

[0261] For example, in the case of the terminal-side encoder training method for the second method, the encoder of the terminal may include a quantizer but may not use it. Alternatively, information regarding the quantizer in the second data set may not have been received from the base station, or even if received, it may not be considered.

[0262] Specifically, training may be performed in the compression unit to perform training on the terminal-side encoder. In the compression unit, when a CSI (e.g., target CSI or V) from the first data set is used as input information, a CSI compressed based on AI from the second data set (e.g., It can be trained to obtain ) as the output. Here, 'raw' may mean that it is not quantized. In this case, only AI-based compression can be performed in the compression unit. That is, compression based on quantization may not be performed in the compression unit.

[0263] In the case of the terminal-side encoder training method for the second method, the terminal-side encoder has less information loss in the CSI (e.g., CSI where compression is performed based on AI or Because it is based on ), training can be performed to better mimic the base station-side encoder.

[0264] Third method: A neural network can be trained by considering both loss function 1 and loss function 2 (Train neural network to reduce both losses in 1 and 2), and the specific operation for this will be explained through FIG. 13.

[0265] Here, loss function 1 is the loss function associated with the first method, It can be the same as. Also, loss function 2 is the loss function associated with the second method, It can be like this.

[0266] FIG. 13 is a drawing illustrating a third method for terminal-side encoder training according to one embodiment of the present disclosure.

[0267] The Total Loss value to be minimized by the terminal-side encoder may be as follows.

[0268] Total Loss =

[0269] Here, It can refer to loss functions such as mean squared error (MSE) or mean absolute error (MAE).

[0270] Here, is CSI compressed based on quantization generated by the terminal-side encoder (e.g., It can mean ), is CSI compressed based on quantization generated by the base station encoder (e.g., It can mean ). Here, is a CSI compressed based on AI generated by the compression unit of the terminal-side encoder (e.g., It can mean ), is CSI compressed based on AI generated by the compression unit of the base station-side encoder (e.g., It may mean ). Accordingly, the loss function for the third method above may mean the sum of the difference between the result of compressing the same CSI based only on AI between the base station and the terminal, and the difference between the result of compressing by also considering the quantization unit.

[0271] In the case of the terminal-side encoder training method for the third method, the terminal can perform training for the terminal-side encoder based on the first data set and the second data set received from the base station. At this time, non-quantized CSI (e.g., CSI compressed based on AI or ) and quantized CSI (e.g., CSI compressed based on quantization or All of ) can be considered.

[0272] Specifically, training can be performed in the compression unit and the quantizer to perform training for the terminal-side encoder.

[0273] In the compression section, when CSI (e.g., target CSI or V) from the first data set is used as input information, CSI compressed based on AI from the second data set (e.g., It can be trained to obtain ) as the output. Here, 'raw' may mean that it is not quantized. In this case, only AI-based compression can be performed in the compression unit. That is, compression based on quantization may not be performed in the compression unit.

[0274] In the quantizer, AI-based compressed CSI from the second dataset (e.g., When ) is used as input information, among the first data set, the CSI compressed based on quantization (e.g., ) can be obtained as result information. Or, in the quantizer, output information of the compression unit (e.g., CSI compressed based on AI or When ) is used as input information, among the first data set, the CSI compressed based on quantization (e.g., ) can be obtained as result information. Here, 'q' may mean quantization. In this case, compression based on quantization can be performed in the quantizer. In this case, quantization can be performed based on information about the quantizer in the second data set.

[0275] FIG. 14a is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to another embodiment of the present disclosure.

[0276] In step 1405, the terminal can report its UE capability to the base station.

[0277] Specifically, the terminal may report to the base station the terminal's capability related to terminal-side encoder training or terminal-side encoder inference. In this case, the report may be transmitted in the form of a message, and said message may include information regarding the terminal's capability related to terminal-side encoder training or inference.

[0278] In step 1410, the terminal may transmit a request message to the base station to request a second data set. This step is optional, and even if the terminal does not transmit a request message for the second data set to the base station, the base station may transmit a data set containing the second data set to the terminal.

[0279] In step 1415, the base station may transmit a message to the terminal containing a first data set and a second data set. This process may imply data exchange.

[0280] The base station may transmit a response message containing a second data set to the terminal in response to a request message. Alternatively, even without a request for the second data set from the terminal, the base station may transmit the second data set to the terminal by determining or deciding whether to transmit the second data set based on a report on the terminal's capabilities.

[0281] The base station can inform the terminal whether the quantizer of the terminal-side encoder is trainable. Information indicating whether the quantizer of the terminal-side encoder is trainable may be transmitted together with or included in the dataset. Such information may be in the form of an indicator.

[0282] In step 1420, the terminal can perform training on the terminal-side encoder. Based on the received first data set and second data set, the terminal can perform AI-based compressed CSI (e.g., ) or compressed CSI based on quantization (e.g., Training for the encoder can be performed to acquire any one of the following. Or compressed CSI based on AI (e.g., Compressed CSI based on ) and quantization (e.g., Training for the encoder can be performed to acquire all of them.

[0283] In addition, if training for the quantizer of the terminal-side encoder is possible, training for the terminal-side quantizer can be performed in step 1420.

[0284] If the distribution of data acquired at the terminal differs from the distribution of data used at the base station (UE data distribution shift), the CSI (e.g., target CSI or V) or the AI-based compressed CSI (e.g., If appropriate quantization cannot be performed for ), the terminal may perform training on the quantizer of the terminal-side encoder. Or, if the base station does not use an appropriate quantizer, the terminal may perform training on the quantizer of the terminal-side encoder. Or, even if the quantizer used by the base station is available, if the performance of the quantizer of the terminal-side encoder is better (for example, if the CSI compressed based on quantization at the terminal has less information loss than the CSI compressed based on quantization at the base station), the terminal may perform training for a new quantizer.

[0285] If the quantizer of the terminal-side encoder is newly trained or if the quantizer of the terminal-side encoder is changed, the terminal may change information about the quantizer. For example, in the case of a scalar quantizer, the step size may be changed. Or, in the case of a vector quantizer, the centroid or codebook may be changed.

[0286] In step 1425, the terminal may transmit or exchange information about a newly trained or modified quantizer or quantizer with the base station. Specifically, in the case of a scalar quantizer, the terminal may report or transmit a modified step size. Or in the case of a vector quantizer, it may report or transmit a modified centroid or a modified codebook.

[0287] In step 1430, the base station can design a de-quantizer of the base station-side decoder based on information about the newly trained or modified quantizer received.

[0288] The base station may redesign the quantizer of the base station-side encoder so that it can perform the same quantization process as the quantizer of the terminal-side encoder based on information about the quantizer (e.g., changed step size, changed centroid, or changed codebook). In addition, the base station-side inverse quantizer may also be redesigned to correspond to the redesigned base station-side quantizer.

[0289] FIG. 14b is a diagram illustrating the exchange of a first data set and a second data set and a terminal-side encoder training procedure according to another embodiment of the present disclosure.

[0290] In step 1435, the terminal can report its UE capability to the base station.

[0291] Specifically, the terminal may report to the base station the terminal's capability related to terminal-side encoder training or terminal-side encoder inference. In this case, the report may be transmitted in the form of a message, and said message may include information regarding the terminal's capability related to terminal-side encoder training or inference.

[0292] In step 1440, the terminal may transmit a request message to the base station to request a second data set. This step is optional, and even if the terminal does not transmit a request message for the second data set to the base station, the base station may transmit a data set containing the second data set to the terminal.

[0293] In step 1445, the terminal may receive a message from the base station containing a first data set and a second data set. This process may imply a data exchange.

[0294] When a terminal receives a data set from a base station, it may receive information regarding whether the quantizer of the terminal-side encoder is trainable. Information indicating whether the quantizer of the terminal-side encoder is trainable may be transmitted together with or included in the data set. Such information may be in the form of an indicator.

[0295] In step 1450, the terminal can perform training on the terminal-side encoder. Based on the received first data set and second data set, the terminal can perform AI-based compressed CSI (e.g., ) or compressed CSI based on quantization (e.g., Training for the encoder can be performed to acquire any one of the following. Or compressed CSI based on AI (e.g., Compressed CSI based on ) and quantization (e.g., Training for the encoder can be performed to acquire all of them.

[0296] In addition, if training for the quantizer of the terminal-side encoder is possible, training for the terminal-side quantizer can be performed in step 1450.

[0297] If the quantizer of the terminal-side encoder is newly trained or if the quantizer of the terminal-side encoder is changed, the terminal may change information about the quantizer. For example, in the case of a scalar quantizer, the step size may be changed. Or, in the case of a vector quantizer, the centroid or codebook may be changed.

[0298] In step 1455, the terminal may transmit or exchange information about a newly trained or modified quantizer or quantizer with the base station. Specifically, in the case of a scalar quantizer, the terminal may report or transmit a modified step size. Or in the case of a vector quantizer, it may report or transmit a modified centroid or a modified codebook.

[0299] FIG. 14c is a diagram illustrating a procedure for exchanging a first data set and a second data set on the base station side according to another embodiment of the present disclosure.

[0300] In step 1465, the base station can receive a report from the terminal regarding the terminal's capability (UE capability).

[0301] Specifically, the terminal may report to the base station the terminal's capability related to terminal-side encoder training or terminal-side encoder inference. In this case, the report may be transmitted in the form of a message, and said message may include information regarding the terminal's capability related to terminal-side encoder training or inference.

[0302] In step 1470, the base station may receive a request message from the terminal to request a second data set. This step is optional, and the base station may transmit a message containing the second data set to the terminal even if it does not receive a request message from the terminal.

[0303] In step 1475, the base station may determine or decide whether to transmit the second data. Specifically, the base station may determine whether to transmit the second data set based on a request message. For example, the base station may transmit a response message containing the second data set in response to the request message. Alternatively, without a request message from the terminal, the base station may determine or decide whether to transmit the second data set based on a report regarding the terminal's capability.

[0304] In step 1480, the base station may transmit a message to the terminal containing a first data set and a second data set. This process may imply a data exchange.

[0305] When the base station transmits a message containing a first data set and a second data set to a terminal, it may indicate whether the quantizer of the terminal-side encoder is trainable. Information indicating whether the quantizer of the terminal-side encoder is trainable may be transmitted together with the data sets or included in the data sets. Such information may be in the form of an indicator.

[0306] In step 1485, the base station may receive information about a newly trained or modified quantizer or quantizer from the terminal. Specifically, in the case of a scalar quantizer, the terminal may report or transmit a modified step size. Or in the case of a vector quantizer, it may report or transmit a modified centroid or a modified codebook.

[0307] In step 1490, the base station can design a de-quantizer of the base station-side decoder based on information about the newly trained or modified quantizer received.

[0308] The base station may redesign the quantizer of the base station-side encoder so that it can perform the same quantization process as the quantizer of the terminal-side encoder based on information about the quantizer (e.g., changed step size, changed centroid, or changed codebook). In addition, the base station-side inverse quantizer may also be redesigned to correspond to the redesigned base station-side quantizer.

[0309] FIG. 15a is a block diagram showing the configuration of a terminal according to one embodiment of the present disclosure.

[0310] Referring to FIG. 15a, the terminal may include a transceiver (1510), a control unit (1520), and a storage unit (1530). In the present disclosure, the control unit (1520) may be defined as a circuit or application-specific integrated circuit or at least one processor.

[0311] The transceiver (1510) can transmit and receive signals with other network entities. The transceiver (1510) can, for example, transmit CSI to a base station. Or, for example, the transceiver (1510) can receive a message from a base station that includes a first data set and a second data set.

[0312] The control unit (1520) can control the overall operation of the terminal according to the embodiment proposed in the present disclosure. For example, the control unit (1520) can control the signal flow between each block to perform operations according to the flowchart described above. Specifically, the control unit (1520) can control the compression of the CSI according to the embodiment of the present disclosure.

[0313] The storage unit (1530) can store at least one of the information transmitted and received through the transmission and reception unit (1510) and the information generated through the control unit (1520). For example, the storage unit (1530) can store CSI as a buffer memory according to one embodiment of the present disclosure.

[0314] FIG. 15b is a block diagram showing the configuration of a control unit of a terminal according to one embodiment of the present disclosure.

[0315] Referring to FIG. 15b, the control unit (1520) of the terminal may include an encoder (1540).

[0316] The encoder (1540) can perform compression operations according to the embodiments proposed in the present disclosure. The encoder (1540) can perform compression on the input and output z as a result. In the present disclosure, the encoder can perform the operations described above. Specifically, the encoder can perform training on the terminal-side encoder based on a first data set and a second data set according to one embodiment of the present disclosure. At this time, the terminal-side encoder can perform training by any one of the first method, the second method, or the third method. For example, the terminal-side encoder can be trained to obtain a compressed CSI (e.g., a target CSI or V) as output information when the CSI (e.g., a target CSI or V) is used as input information.

[0317] FIG. 16a is a block diagram showing the configuration of a base station according to one embodiment of the present disclosure.

[0318] Referring to FIG. 16a, the base station may include a transceiver (1610), a control unit (1620), and a storage unit (1630). In the present disclosure, the control unit (1620) may be defined as a circuit or application-specific integrated circuit or at least one processor.

[0319] The transmitting and receiving unit (1610) can transmit and receive signals with other network entities. The transmitting and receiving unit (1610) can, for example, transmit CSI-RS to a terminal. Or, for example, the transmitting and receiving unit (1610) can transmit a message including a first data set and a second data set to a terminal.

[0320] The control unit (1620) can control the overall operation of the base station according to the embodiment proposed in the present disclosure. For example, the control unit (1620) can control the signal flow between each block to perform operations according to the flowchart described above. Specifically, the control unit (1620) can perform restoration of the compressed CSI according to the embodiment of the present disclosure.

[0321] The storage unit (1630) can store at least one of the information transmitted and received through the transmission and reception unit (1610) and the information generated through the control unit (1520). For example, the storage unit (1530) can be a buffer memory according to one embodiment of the present disclosure and can store a data set to be transmitted to a terminal.

[0322] FIG. 16b is a block diagram showing the configuration of a control unit of a base station according to one embodiment of the present disclosure.

[0323] The encoder (1640) can perform compression operations according to the embodiments proposed in the present disclosure. The encoder (1640) can perform compression on the input and output z as a result. In the present disclosure, the encoder can perform the operations described above. Specifically, the encoder can perform compression on the CSI according to one embodiment of the present disclosure.

[0324] The decoder (1650) can perform an operation for restoration according to an embodiment proposed in the present disclosure. The decoder (1650) can perform restoration on z compressed in the encoder and output an output as a result. In the present disclosure, the decoder can perform the operation described above. Specifically, the decoder can perform restoration or reconstruction on the compressed CSI according to one embodiment of the present disclosure.

[0325] In the specific embodiments of the present disclosure described above, the components included in the disclosure are expressed in a singular or plural form according to the specific embodiments presented. However, the singular or plural expression is selected to suit the situation 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 of a singular form, and even if a component is expressed in the singular form, it may be composed of a plural form.

[0326] 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.

[0327] The various embodiments of the present disclosure and the terms used therein are not intended to limit the technology described in the present disclosure to specific embodiments and should be understood to include various modifications, equivalents, and / or substitutions of said embodiments. In connection with the description of the drawings, similar reference numerals may be used for similar components. A singular expression may include a plural expression unless the context clearly indicates otherwise. In the present disclosure, expressions such as "A or B," "at least one of A and / or B," "A, B or C," or "at least one of A, B and / or C" may include all possible combinations of items listed together. Expressions such as "first," "second," "first," or "second" may modify said components regardless of order or importance and are used only to distinguish one component from another and do not limit said components. Where it is stated that a certain (e.g., first) component is "(functionally or telecommunicationally) connected" or "connected" to another (e.g., second) component, said certain component may be directly connected to said other component or connected through another component (e.g., third component).

[0328] As used in this disclosure, the term "module" includes a unit composed of hardware, software, or firmware, and may be used interchangeably with terms such as logic, logic block, component, or circuit. A module may be a component formed integrally, or a minimum unit or part thereof that performs one or more functions. For example, a module may be composed of an application-specific integrated circuit (ASIC).

[0329] Various embodiments of the present disclosure may be implemented as software (e.g., a program) comprising instructions stored in a machine-readable storage medium (e.g., internal memory or external memory) that is readable by a machine (e.g., a computer). The machine may include a terminal according to various embodiments, which is a device capable of calling instructions stored from the storage medium and operating according to the called instructions. When the instructions are executed by a processor, the processor may perform a function corresponding to the instructions directly or by using other components under the control of the processor. The instructions may include code generated or executed by a compiler or an interpreter.

[0330] A device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-transitory' means merely that the storage medium does not contain a signal and is tangible, without distinguishing whether data is stored semi-permanently or temporarily on the storage medium.

[0331] Methods according to the various embodiments disclosed herein may be provided as included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed online in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)) or through an application store (e.g., Play Store™). In the case of online distribution, at least a portion of the computer program product may be temporarily stored or temporarily created in a storage medium such as the memory of a manufacturer's server, an application store's server, or a relay server. Each component (e.g., a module or program) according to the various embodiments may be composed of a singular or multiple entities, and some of the aforementioned sub-components may be omitted, or other sub-components may be further included in the various embodiments. Generally or additionally, some components (e.g., a module or program) may be integrated into a single entity to perform the same or similar functions as those performed by each of the respective components prior to integration. Operations performed by a module, program, or other component according to various embodiments may be executed sequentially, in parallel, iteratively, or heuristically, or at least some operations may be executed in a different order, omitted, or other operations may be added.

Claims

1. A method performed by a terminal in a wireless communication system, A step of receiving a message from a base station comprising a first data set and a second data set for training a terminal-side encoder that performs compression of Channel State Information (CSI); and Based on the first data set and the second data set, the method includes the step of performing learning to generate output information of CSI compression performed by the base station encoder when the terminal-side encoder uses the CSI as input information. The first data set above includes CSI compressed based on the CSI and quantization, and The second dataset above includes CSI compressed based on artificial intelligence (AI), and A method characterized in that the output information of the CSI compression performed at the base station side encoder includes the CSI compressed based on the artificial intelligence.

2. In claim 1, the step of performing learning for the terminal-side encoder is: A method characterized by further including the step of the compression unit of the terminal-side encoder performing learning to obtain a compressed CSI based on the artificial intelligence using the CSI as input information.

3. In paragraph 2, if the output information of the CSI compression performed by the base station-side encoder further includes CSI compressed based on the quantization, the step of performing learning for the terminal-side encoder is: The quantizer of the terminal-side encoder further includes the step of performing learning to acquire a compressed CSI based on the quantization using the CSI compressed based on the AI ​​as input information. A method characterized in that the base station side decoder is trained to restore the output information of the CSI compression performed by the base station side encoder into information identical to the CSI.

4. In Paragraph 1, A step of transmitting a report message to the base station to report the capability of the terminal; and The method further includes the step of transmitting a request message to the base station to request the second data set based on the report message. If the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the quantization: The second data set further includes information about the quantizer included in the base station-side encoder, and University learning on the above terminal-side encoder is performed based on the above quantizer, and A method characterized by including information about the above quantizer, any one of the quantizer, the step size of scalar quantization, or the codebook of vector quantization.

5. In a method performed by a base station in a wireless communication system, A step of determining whether to transmit a first data set and a second data set to train a terminal-side encoder that performs compression of Channel State Information (CSI); and When transmission is determined, the method includes the step of transmitting to a terminal a message comprising the first data set including the CSI compressed based on the CSI and quantization, and the second data set including the CSI compressed based on artificial intelligence (AI). The terminal-side encoder is trained to generate output information of CSI compression, which is compressed by the base station-side encoder, based on the first data set and the second data set. A method characterized in that the output information of the CSI compression performed at the base station side encoder includes the CSI compressed based on the artificial intelligence.

6. In Paragraph 5, The compression unit of the above terminal-side encoder performs learning to obtain a compressed CSI based on the artificial intelligence using the above CSI as input information, and If the output information of the CSI compression performed by the base station-side encoder further includes CSI compressed based on the quantization, The quantizer of the terminal-side encoder is trained to acquire a compressed CSI based on the quantization using the CSI compressed based on the AI ​​as input information, and A method characterized in that the base station side decoder is trained to restore the output information of the CSI compression performed by the base station side encoder into information identical to the CSI.

7. In Paragraph 5, A step of receiving a report message from the terminal to report the capability of the terminal; and The method further includes the step of receiving a request message for requesting the second data set from the terminal based on the report message. If the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the quantization: The second data set further includes information about the quantizer included in the base station-side encoder, and University learning on the above terminal-side encoder is performed based on the above quantizer, and A method characterized by including information about the above quantizer, any one of the quantizer, the step size of scalar quantization, or the codebook of vector quantization.

8. In a terminal of a wireless communication system, Transceiver configured to transmit and receive signals; and It includes a control unit coupled to the above-mentioned transmitting and receiving unit, and the control unit: Receiving a message from a base station comprising a first data set and a second data set for training a terminal-side encoder that performs compression of Channel State Information (CSI), and Based on the first data set and the second data set, when the terminal-side encoder uses the CSI as input information, learning is performed to generate output information of CSI compression performed by the base station-side encoder. The first data set above includes CSI compressed based on the CSI and quantization, and The second dataset above includes CSI compressed based on artificial intelligence (AI), and A terminal characterized in that the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the artificial intelligence.

9. In paragraph 8, the control unit is: A terminal characterized in that the compression unit of the above-mentioned terminal-side encoder performs learning to obtain a compressed CSI based on the above-mentioned artificial intelligence using the above-mentioned CSI as input information.

10. In claim 9, if the output information of the CSI compression performed by the base station-side encoder further includes CSI compressed based on the quantization, the control unit: The quantizer of the terminal-side encoder uses the CSI compressed based on the AI ​​as input information and performs learning to acquire the CSI compressed based on the quantization, A terminal characterized in that the base station side decoder is trained to restore the output information of the CSI compression performed by the base station side encoder into information identical to the CSI.

11. In Paragraph 8, Transmitting a report message to the base station to report the capability of the terminal, and A request message for requesting the second data set is transmitted to the base station based on the report message, and If the output information of the CSI compression performed by the base station-side encoder includes the CSI compressed based on the quantization: The second data set further includes information about the quantizer included in the base station-side encoder, and University learning on the above terminal-side encoder is performed based on the above quantizer, and A terminal characterized by including information regarding the above-mentioned quantizer, any one of the quantizer, the step size of scalar quantization, or the codebook of vector quantization.

12. In a base station of a wireless communication system, Transceiver configured to transmit and receive signals; and It includes a control unit coupled to the above-mentioned transmitting and receiving unit, and the control unit: Determining whether to transmit a first data set and a second data set to train a terminal-side encoder that performs compression of Channel State Information (CSI), and When transmission is determined, a message is transmitted to the terminal comprising the first data set including the CSI compressed based on the CSI and quantization, and the second data set including the CSI compressed based on artificial intelligence (AI). The terminal-side encoder is trained to generate output information of CSI compression, which is compressed by the base station-side encoder, based on the first data set and the second data set. A base station characterized by the output information of the CSI compression performed by the base station-side encoder including the CSI compressed based on the artificial intelligence.

13. In Paragraph 12, The compression unit of the above terminal-side encoder performs learning to obtain a compressed CSI based on the artificial intelligence using the above CSI as input information, and If the output information of the CSI compression performed by the base station-side encoder further includes CSI compressed based on the quantization, The quantizer of the terminal-side encoder is trained to acquire a compressed CSI based on the quantization using the CSI compressed based on the AI ​​as input information, and A base station characterized in that the base station side decoder is trained to restore the output information of the CSI compression performed by the base station side encoder into information identical to the CSI.

14. In paragraph 12, if the output information of the CSI compression performed by the base station-side encoder includes CSI compressed based on the quantization: The second data set further includes information about the quantizer included in the base station-side encoder, and University learning on the above terminal-side encoder is performed based on the above quantizer, and A base station characterized by including information regarding the above-mentioned quantizer, any one of the quantizer, the step size of scalar quantization, or the codebook of vector quantization.

15. In paragraph 12, the control unit above: Receive a report message from the above terminal to report the capability of the above terminal, and A base station characterized by receiving a request message for requesting the second data set from the terminal based on the report message.

Images