Decoding method performed by base station in wireless communication system, and apparatus therefor

Abstract

The present disclosure relates to a 5G or 6G communication system for supporting a higher data transmission rate than a 4G communication system such as LTE. According to an embodiment of the present disclosure, the CSI feedback reconstruction performance of a base station that has received compressed CSI feedback from a terminal may be improved.

Description

Decoding method performed by a base station in a wireless communication system and apparatus for the same

[0001] The present disclosure relates to a wireless communication system or a mobile communication system. Specifically, it relates to a method for efficiently restoring channel state information (CSI) feedback at a base station when such feedback is received by a terminal.

[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] Meanwhile, when a terminal compresses CSI feedback and transmits it to a base station, the need for a method to efficiently restore it at the base station has arisen.

[0008] The objective of the present invention is to improve channel prediction performance by efficiently restoring a channel state information (CSI) report transmitted by a terminal at a base station.

[0009] According to one embodiment of the present disclosure for solving the above-mentioned problems, a decoding method performed by a base station in a wireless communication system is disclosed, comprising the steps of: receiving a latent vector for a channel state information (CSI) feedback signal from a terminal; storing the received latent vector in a buffer memory; and obtaining an eigenvector at any time based on at least one latent vector stored in the buffer memory for a preset time.

[0010] Meanwhile, according to another embodiment of the present disclosure, a base station in a wireless communication system comprises a transceiver, a buffer memory, a decoder, and a control unit that receives a latent vector for a channel state information (CSI) feedback signal from a terminal through the transceiver, controls the received latent vector to be stored in the buffer memory, and controls the decoder to obtain an eigenvector at any time based on at least one latent vector stored in the buffer memory for a preset time.

[0011] According to one embodiment of the present invention, the CSI feedback recovery performance of a base station that receives compressed CSI feedback from a terminal can be improved.

[0012] 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.

[0013] FIG. 1 is a diagram illustrating general channel state information (CSI) feedback,

[0014] FIG. 2 is a diagram specifically illustrating a general CSI feedback procedure,

[0015] Figure 3a is a diagram showing a general AI / ML-based CSI compression scheme,

[0016] Figure 3b is a diagram illustrating a typical CSI-RS transmission and AI / ML-based CSI compression procedure,

[0017] FIG. 4 is a diagram showing the input and output of a typical encoder and decoder,

[0018] Figure 5 is a diagram illustrating general AI-based CSI feedback,

[0019] FIG. 6 is a diagram showing a general spatial-frequency domain autoencoder,

[0020] FIG. 7 is a drawing showing a terminal-side (UE-side) encoder and a network-side temporal spatial frequency (TSF) decoder according to one embodiment of the present disclosure.

[0021] FIG. 8 is a drawing showing a UE-side encoder and a network-side single-step correlation extraction TSF decoder according to one embodiment of the present disclosure.

[0022] FIG. 9 is a drawing showing a UE-side encoder and a network-side two-step correlation extraction TSF decoder according to one embodiment of the present disclosure.

[0023] FIG. 10 is a diagram illustrating the operation of a buffer memory related to a latent vector input according to one embodiment of the present disclosure.

[0024] FIG. 11 is a drawing showing an operation related to initial input processing according to one embodiment of the present disclosure,

[0025] FIG. 12 is a diagram illustrating an operation related to feedback error processing according to one embodiment of the present disclosure,

[0026] FIG. 13 is a drawing showing a UE-side encoder and a network-side single-step correlation extraction modular TSF decoder according to one embodiment of the present disclosure.

[0027] FIG. 14 is a drawing showing a UE-side encoder and a network-side two-step correlation extraction modular TSF decoder according to one embodiment of the present disclosure.

[0028] FIG. 15 is a drawing showing a switch operation for turning on / off a TSF module according to one embodiment of the present disclosure.

[0029] FIG. 16 is a diagram showing the processing operation of a TSF module when a missing input is included as an initial input, according to one embodiment of the present disclosure.

[0030] FIG. 17 is a drawing showing a UE-side encoder and a network-side TSF neural network (NN) decoder according to one embodiment of the present disclosure.

[0031] FIG. 18 is a diagram illustrating the operation of a reconstructed CSI input-related buffer memory according to one embodiment of the present disclosure.

[0032] FIG. 19 is a drawing illustrating an operation related to initial input processing according to one embodiment of the present disclosure,

[0033] FIG. 20 is a drawing illustrating an operation related to feedback error processing according to one embodiment of the present disclosure,

[0034] FIG. 21 is a drawing showing a UE-side encoder and a network-side TSF recurrent neural network (RNN) decoder according to one embodiment of the present disclosure.

[0035] FIG. 22 is a diagram showing a TSF RNN decoder reset operation according to one embodiment of the present disclosure,

[0036] FIG. 23 is a drawing showing a UE-side encoder and a network-side multi-output decoder according to one embodiment of the present disclosure,

[0037] FIG. 24 is a diagram showing a network-side multi-output decoder including a UE-side encoder, a processing block, and a single-step correlation extraction according to one embodiment of the present disclosure.

[0038] FIG. 25 is a diagram showing a network-side multi-output decoder including a UE-side encoder, a processing block, and a two-step correlation extraction according to one embodiment of the present disclosure.

[0039] FIG. 26 is a drawing showing the operation of a multi-output decoder according to one embodiment of the present disclosure.

[0040] FIG. 27 is a drawing showing the predicted results of a simulation according to one embodiment of the present disclosure,

[0041] Figure 28 is a diagram showing the structure of a typical vanilla transformer,

[0042] FIG. 29a is a diagram showing the structure of a typical transformer-based autoencoder,

[0043] FIG. 29b is a drawing showing a transformer-based autoencoder structure according to one embodiment of the present disclosure,

[0044] FIG. 30 is a diagram showing the specific operation of multi-head attention in the structure of a typical vanilla transformer,

[0045] FIG. 31 is a diagram showing the training procedure of a general SF encoder and SF decoder,

[0046] FIG. 32a is a diagram showing a training procedure for an SF encoder and an SF / TSF decoder according to one embodiment of the present disclosure.

[0047] FIG. 32b is a diagram showing both the inference and learning procedures of an SF encoder and an SF / TSF decoder according to one embodiment of the present disclosure.

[0048] FIG. 33a is a drawing showing the predicted results of a simulation according to one embodiment of the present disclosure,

[0049] FIG. 33b is a drawing showing the predicted results of a simulation according to one embodiment of the present disclosure,

[0050] FIG. 33c is a drawing showing simulation prediction results according to one embodiment of the present disclosure,

[0051] FIG. 34 is a flowchart showing the operation of a base station according to one embodiment of the present disclosure,

[0052] FIG. 35 is a block diagram showing the configuration of a base station according to one embodiment of the present disclosure,

[0053]

[0054] FIG. 36 is a drawing showing simulation prediction results according to an embodiment of the present disclosure,

[0055] FIG. 37 is a diagram showing detailed hyperparameters of a transformer-based autoencoder of the present disclosure,

[0056] FIG. 38 is a diagram showing detailed parameters of the simulation environment of the present disclosure,

[0057] FIG. 39 is a diagram showing the complexity of the auto encoder and baseline auto encoder proposed in one embodiment of the present disclosure.

[0058] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0059] 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.

[0060] For the same reason, some components in the attached drawings have been exaggerated, omitted, or schematically depicted. Additionally, the size of each component does not entirely reflect its actual dimensions. Identical or corresponding components in each drawing have been assigned the same reference numbers.

[0061] 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.

[0062] 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 means of instruction 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).

[0063] 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.

[0064] 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 operate one or more processors. Accordingly, 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." Furthermore, the components and "parts" may be implemented to operate one or more CPUs within a device or secure multimedia card.

[0065] 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 the modifiers such as "first," "second," etc., may refer to different objects. However, the terms modified by the modifiers such as "first," "second," etc., may refer to the same object. That is, the modifiers such as "first," "second," etc., may be used to refer to the same object from different perspectives. For example, the 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.

[0066] 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.

[0067] Since 5G New Radio (NR) communication systems must be able to freely reflect the diverse requirements of service providers and users, services that satisfy various requirements simultaneously must be supported. Services being considered for 5G communication systems include enhanced Mobile Broadband (eMBB), massive Machine Type Communication (mNTC), and Ultra Reliability Low Latency Communication (URLLC).

[0068] eMBB aims to provide data transmission speeds that are superior to those supported by standard 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 in the downlink and 10 Gbps in 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 LTE transmits signals using a maximum bandwidth of 20 MHz in the 2 GHz band, 5G communication systems 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.

[0069] 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, due to the nature of the service, terminals supporting mMTC are likely to be located in dead zones where cells cannot cover, such as building basements, so they require wider coverage compared to other services provided by the 5G communication system. Terminals supporting mMTC may consist of low-cost devices, and since it is difficult to frequently replace the device's battery, a very long battery life of 10 to 15 years may be required.

[0070] Finally, URLLC is a mission-critical cellular-based wireless communication service. For example, URLLC can be considered for services used in 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 example, 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.

[0071] To satisfy such diverse services, it is often necessary to support beam management or various frequency bands. In such situations, diverse channel environments may arise for each frequency band or beam, and the resource consumption for channel estimation at the terminal and reporting to the base station can be significant. To address this, there is discussion regarding CSI compression, which involves compressing and transmitting channel state information (CSI).

[0072] Channel states recovered through CSI reports based on the standard 5G NR codebook method may experience information loss during the process of transmitting the estimated channel H due to the quantization problem of the codebook. While terminals can configure codebooks in more diverse ways, the amount of data that needs to be transmitted may also increase accordingly. For this reason, a CSI compression study item was discussed in Release-18 and a work item is being discussed in Release-19, which involves compressing and transmitting CSI reports based on an auto-encoder (AE) rather than the 5G NR codebook-based channel feedback, thereby allowing the base station to efficiently decode the channel state that the terminal intends to transmit.

[0073] According to one embodiment, as illustrated in FIG. 1, a base station may transmit a CSI-reference signal (RS) to a terminal for channel measurement. When the terminal performs channel estimation using the received CSI-RS, it may acquire the estimated channel and transmit CSI feedback to the base station. At this time, the terminal may provide feedback on the CSI in the form of quantized quantities. The CSI feedback may include at least one of a precoding matrix indicator (PMI) and a channel quality indicator (CQI). The base station may transmit downlink data to the terminal based on the CSI feedback.

[0074] Specifically, the procedure for CSI feedback is explained based on Fig. 2. In order for the base station to know the downlink channel estimated by the terminal, a separate feedback process is required for the terminal to transmit channel information to the base station. The base station may transmit CSI-RS to the terminal to know the downlink channel estimated by the terminal. The terminal performs channel estimation using the received CSI-RS, and the estimated channel It can obtain. The terminal transmits PMI, which is channel information related to the precoding matrix to be used by the base station, in order to After performing eigenvalue decomposition (EVD) of, at It can be known. The terminal is During transmission, based on the pre-defined codebook specified in 3GPP technical specification (TS) 38.214 The index of the codebook most similar to can be transmitted to the base station in the form of PMI. Transmitting such a codebook index is one of the general CSI feedback methods of 5G NR.

[0075] Meanwhile, the transmission of more accurate CSI information is required for the use of multi-user (MU) multiple input multiple output (MIMO). Additionally, the Rel-16 Type 2 (eType 2) codebook, which compresses channels in the spatial and frequency domains following 3GPP Type 1 and Type 2, has been introduced. Unlike the Type 2 codebook, which generally compresses channel information in the spatial domain and provides feedback, the eType 2 codebook, which performs frequency domain compression using channel sparsity in the frequency domain, utilizes the reduced feedback bits for feedback on phase and amplitude information in more finely sub-band areas, thereby enabling more precise channel information feedback compared to the Type 2 codebook.

[0076] However, this codebook type Transmission is actually due to the limited granularity of the codebook (e.g., limited number of DFT beams, amplitude, phase information, etc.) Differences may occur. This can lead to performance degradation in MIMO systems. Furthermore, as the performance improvement of MU-MIMO systems increases the influence of CSI feedback from each terminal, more accurate CSI feedback is required in future communication systems. To this end, existing CSI feedback systems can obtain accurate CSI by increasing the number of bits used in the CSI report. However, as the number of base station antennas, bandwidth, and granularity increase, the overhead increases significantly.

[0077] To address the problem of such increased overhead, the 3GPP Rel-18 study item (SI) and Rel-19 work item (WI) utilize artificial intelligence (AI) / machine learning (ML) on the terminal It compresses itself and transmits it to the base station in the form of a low-dimensional vector z (latent vector), and the base station obtains the original from the compressed vector z AI / ML-based CSI compression to restore is being discussed.

[0078] FIG. 3a is a diagram illustrating AI / ML-based CSI compression. Reference numeral 300 represents an AI / ML-based autoencoder. As shown in FIG. 4, the encoder of the autoencoder is located at the terminal, and the decoder is located at the base station. For example, the terminal can transmit a latent vector (feature vector), which is the result of compression through the encoder, to the base station via the uplink (UL). The latent vector may be bit sequence information corresponding to the PMI. The base station then transmits this latent vector (feature vector) through the decoder. It can be restored.

[0079] Figure 3b is a diagram illustrating CSI-RS transmission and an AI / ML-based CSI compression procedure.

[0080] As illustrated in FIG. 3b, in step S310, the base station can transmit CSI-RS to the terminal. Then, in step S320, the terminal can perform encoder inference. In step S330, the terminal receives the compressed vector as described above as the output of the encoder inference. It can be transmitted to the base station in the form of a latent vector. And the base station can perform decoder inference in step S340.

[0081] The general AI / ML-based CSI compression technology described above compresses channel information in the spatial and frequency domains and feeds it back to the base station. Specifically, as illustrated in Fig. 5, the terminal measures the channel based on CSI-RS reception information and can extract an eigenvector matrix that is necessary for use as a base station precoder. Subsequently, the terminal can compress the eigenvectors for each frequency subband into latent vectors using the terminal's AI encoder model. Then, the terminal can feed back this information to the base station through a CSI report. Upon receiving the latent vectors within the CSI report, the base station can reconstruct the channel information using its base station AI decoder model.

[0082] Methods such as utilizing advanced AI models and increasing model complexity exist to enhance the performance of general AI / ML-based CSI compression technologies. However, problems may arise where the maximum performance is bounded and models exceeding a certain size cannot be implemented due to increased system overhead. It is necessary to verify the potential for performance improvement in AI / ML CSI compression technologies through compression beyond the spatial and frequency domains, or through reconstruction using domains other than spatial and frequency.

[0083] According to one embodiment, a temporal domain may exist in areas other than spatial and frequency domains. If the temporal domain is additionally considered during CSI compression at a terminal, additional standard agreements and signaling definitions considering the temporal domain may be required. Therefore, it is possible to utilize the temporal domain during the base station's channel restoration process while maintaining the standard definitions of general spatial and frequency domain channel compression-based AI / ML CSI compression technologies.

[0084] In this disclosure, a method is proposed to improve the performance of CSI compression technology by replacing the spatial frequency decoder of a base station, as shown in FIG. 6, with a temporal spatial frequency decoder.

[0085] In the case of Fig. 6, V_x is input CSI information and can be used as an input to an encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to a base station where a decoder is located. The decoder at the base station can use z_x as an input to finally output V_hat_x. V_hat_x is data that can have the same size as the original input V_x, and the corresponding auto-encoder, consisting of an encoder / SF decoder, operates to output V_hat_x similarly to V_x.

[0086] According to one embodiment, a TSF decoder is proposed to improve CSI reconstruction performance compared to an SF decoder by using historical feedback or historical reconstructed CSI received from an SF encoder of a general terminal for current CSI reconstruction.

[0087] Specifically, according to one embodiment of the present disclosure, various implementation methods of a network (e.g., base station) AI decoder that additionally utilizes a temporal domain are proposed. Furthermore, when considering the temporal domain, operations for memory storage and processing of historical feedback information of the network are also proposed in the present disclosure.

[0088] FIG. 7 is a diagram showing a terminal-side (UE-side) encoder and a network-side temporal spatial frequency (TSF) decoder according to one embodiment of the present disclosure.

[0089] In the case of Fig. 7, V_x is the input CSI information of the current cycle and can be used as the input to the encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to the base station where the decoder is located. The preceding process is executed in previous cycles as well, and accordingly, the base station or base station decoder can store z_x-1 and z_x-2, which are previously transmitted information, in a pre-defined memory. The base station decoder, the TSF decoder, can finally output V_hat_x by using the previously transmitted z_x-1 and z_x-2 along with the currently transmitted z_x as inputs. V_hat_x is data that can have the same size as the original input V_x, and the corresponding auto-encoder, composed of an encoder and a TSF decoder, operates to output V_hat_x similarly to V_x.

[0090] According to the first embodiment, a method is proposed in which a base station utilizes historical feedback received from a terminal to restore current feedback. According to this method, improvement in the channel restoration performance of the decoder can be expected by considering the temporal domain correlation between feedback information. By considering the addition of the temporal domain to the NW receiver decoder, improved performance compared to a general SF decoder can be achieved. Furthermore, there is an advantage that such performance can be improved solely by changing the base station decoder, without changing the AI / ML model structure or model parameters (weights) of the terminal encoder.

[0091] Specifically, according to the first-1 and first-2 embodiments, the step of extracting correlation between latent vectors may differ. According to the first-1 embodiment, single-step correlation extraction may be a feature as illustrated in FIG. 8. The base station may restore the latent vector at each time point into an eigenvector or a matrix having a dimension greater than or equal to the latent vector, and then extract correlation in the temporal domain, spatial domain, and frequency domain using NN2 (neural network).

[0092] For example, Figure 8 illustrates a detailed example of how a TSF decoder operates. V_x is the input CSI information of the current cycle and can be used as an input to an encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to a base station that includes a decoder. The preceding process is executed in previous cycles as well, and accordingly, the base station or base station decoder can store z_x-1 and z_x-2, which are previously transmitted information, in a pre-defined memory. The TSF decoder, which is the base station decoder, can finally output V_hat_x by using the previously transmitted z_x-1 and z_x-2 along with the currently transmitted z_x as inputs. At this point, the TSF decoder can process the zs collectively to output z_x, z_x-1, ... as h_x, h_x-1, and h_x-2, respectively, which are information with larger dimensions, through NN1. Subsequently, these outputs can pass through NN2 to be finally output as V'_hat_x. NN2 is expected to play the role of extracting the temporal relationship between the current and past transmitted hs. V_hat_x is data that can have the same size as the original input V_x, and the corresponding autoencoder, consisting of an encoder and a decoder, operates to output V_hat_x similarly to V_x.

[0093] Meanwhile, Figure 9 is another detailed example of how a TSF decoder operates. V_x is the input CSI information of the current cycle and can be used as the input to the encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to the base station where the decoder is located. The preceding process is executed in the previous cycle as well, and accordingly, the base station or base station decoder can store the previously transmitted information z_x-1 and z_x-2 in a pre-defined memory. The base station decoder, the TSF decoder, can finally output V_hat_x by using the previously transmitted z_x-1 and z_x-2 along with the currently transmitted z_x as inputs. At this time, the TSF decoder integrates and processes the zs to cause NN1 to output z'_x, which has the same size as z_x, z_x-1, ..., and enables the output to pass through NN2 to finally output V_hat_x. NN1 is expected to play the role of extracting the temporal relationship between the current and past transmitted zs. V_hat_x is data having the same size as the original input V_x, and the corresponding autoencoder, consisting of an encoder and a decoder, operates to output V_hat_x similarly to V_x.

[0094] According to the first-2nd embodiment, as illustrated in FIG. 9, two-step correlation extraction may be featured. The base station may extract a latent vector_hat by extracting the temporal correlation between latent vectors using NN1, and then go through a procedure to restore the eigenvector by extracting components in the spatial and frequency domains using NN2.

[0095] In FIGS. 8 and 9 illustrating the above-mentioned 1-1 and 1-2 embodiments, NN represents a neural network, V_x represents an eigenvector at time x, z_x represents a latent vector at time x (bit sequence), and h_x represents a matrix with temporal, spatial, and frequency axes at time x.

[0096] According to the first embodiment, since correlation extraction is performed after increasing the dimension of the latent vector, there is an advantage of obtaining greater performance gain in the simulation.

[0097] Meanwhile, according to the first and second embodiments, since temporal domain correlation extraction is performed on a low-dimensional latent vector, there is an advantage of having a relatively small computational complexity.

[0098] The following proposes a specific operation of the first embodiment. FIG. 10 is a diagram illustrating the operation of a latent vector (z) buffer memory.

[0099] To utilize historical feedback in the NW decoder, it is necessary to define a buffer memory for storing the latent vector (z) and the operation of said buffer memory. The said buffer memory can perform operations of storing and reading z in memory in chronological order. The size of the buffer memory can be specified as a specific size. The buffer memory performs a first-in, first-out (FIFO) operation so that if the buffer memory space is exceeded, the feedback z stored at the earliest time (z in Fig. 10) x-n It is possible to make ) delete. When using information within the buffer memory as input in the decoder, the information can be processed by concatenating it in chronological order.

[0100] FIG. 11 is a diagram illustrating operations related to initial input processing according to an embodiment of the present disclosure. According to an embodiment, when using an initial input, the historical z value may not exist in the buffer memory. Therefore, if null information exists in the initial input and in the buffer memory, the base station may process it instead with padding and use it for model input.

[0101] In this case, the padding may be zero-padding, for example, added to maintain the size of the input to the decoder. z_0 in Fig. 11 represents the initial input, and the padding may fill the empty space within the remaining buffer memory. In this case, z_0 and the padding are concatenated to form a single data, which can be used as the initial input to the TSF decoder.

[0102] Meanwhile, FIG. 12 is a diagram illustrating an operation related to feedback error processing according to an embodiment of the present disclosure. When a feedback error (such as a decoding failure) occurs for z transmitted from a terminal, z at that point in time may not be able to be stored in the buffer memory of the base station. Therefore, as shown in FIG. 12, there may be two methods: inputting (or storing) padded information into the memory at that point in time, or emptying the memory at that point in time and processing the data by padding the empty memory when inputting the model. FIG. 12 illustrates an example in which z_x-1, which is an encoder output corresponding to V_x-1, fails to transmit and fails to decode at the base station. As shown in FIG. 12, the original data cannot exist at the location of z_x-1 within the buffer memory, and to handle this, it refers to an operation of padding or skipping to place an empty space at the location corresponding to z_x-1.

[0103] Meanwhile, in the second embodiment of the present disclosure, a method of using a TSF decoder and a general SF decoder together is proposed.

[0104] As shown in FIGS. 13 and 14, the TSF module uses historical z as input, and the output of the TSF module meets the output of the existing SF decoder, passes through a processing block, and can be restored to a channel that takes the temporal domain into account.

[0105] Figure 13 shows a case including a single-step correlation extraction modular TSF decoder as described in Figure 8, and Figure 14 shows a case including a two-step correlation extraction modular TSF decoder as described in Figure 9.

[0106] Figure 13 illustrates a detailed example of the operation of a TSF decoder. V_x is the input CSI information of the current cycle and can be used as the input to the encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. Subsequently, z_x can be transmitted to a base station containing a decoder. A feature of the embodiment illustrated in Figure 13 is that a switch is located therein, and when the switch is turned off, the TSF module below the switch is short-circuited, allowing only the encoder-decoder operation to be performed. The encoder-decoder operation is a process of recovering V_hat_x using z_x as input to the decoder; at this time, the processing block is simply passed through or skipped because h'_x is not input, so V'_hat_x has a value exactly identical to V_hat_x.

[0107] Meanwhile, when the Switch is on, z_x is passed as input to the TSF module and can be used for output reconstruction along with z values ​​received at past points in time, such as z_x-1 and z_x-2, which exist in existing memory. At this time, the TSF decoder can process the z values ​​collectively to output z_x, z_x-1, ... as h_x, h_x-1, and h_x-2, respectively, which are information with larger dimensions, passing through NN1. Subsequently, these outputs can pass through NN2 to be finally output as h'_x. NN2 is expected to play the role of extracting the temporal relationship between the current and past transmitted h values. V_hat_x is data that can have the same size as the original input V_x, and the corresponding autoencoder, consisting of an encoder and a decoder, operates to output V_hat_x similarly to V_x. Afterward, h'x is used as input to the processing block along with V_hat_x, and the final V'_hat_x can be output.

[0108] Figure 14 is another detailed example of how a TSF decoder operates. V_x is the input CSI information of the current cycle and can be used as the input to the encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to the base station where the decoder is located. A feature of the embodiment illustrated in Figure 14 is that a switch is located, and when the switch is turned off, the TSF module below the switch is short-circuited so that only the encoder-decoder operation can be performed. The encoder-decoder operation is a process of recovering V_hat_x using z_x as input to the decoder; at this time, the processing block is simply passed through or skipped because h'_x is not input, so V'_hat_x has a value exactly identical to V_hat_x.

[0109] Meanwhile, when the Switch is on, z_x is passed as input to the TSF module and can be used for output reconstruction along with z values ​​received at past points in time, such as z_x-1 and z_x-2, which exist in existing memory. At this time, the TSF decoder can process the zs collectively to output z'_x—information with the same dimensions as z_x, z_x-1, etc.—through NN1. Subsequently, this output can pass through NN2 to be finally output as h_x. NN1 is expected to play the role of extracting the temporal relationship between the current and past transmitted zs. V_hat_x is data that can have the same size as the original input V_x, and the corresponding autoencoder, consisting of an encoder and a decoder, operates to output V_hat_x similarly to V_x. Afterward, h_x is used as an input to the processing block along with V_hat_x, and the final V'_hat_x can be output.

[0110] The function of the 'processing block' included in FIGS. 13 and 14 can perform at least one operation, such as acting as a switch to block one of the two inputs, acting as a switch to block one of the two output circuits, performing an operation to calculate the weighted sum of the two inputs, and a Neural Network to calculate weights for calculating the weighted sum of the two inputs.

[0111] The second embodiment utilizes a TSF module while using an existing SF decoder, so the terminal can use the encoder in the same way as before, and the base station maintains the operation of the existing SF decoder while demonstrating performance improvement, which is an advantage. The scenarios in which this embodiment is applicable include cases where the time correlation between channels is constant, such as operating short-period CSI reports, terminals with uniform mobility, or operations of rank 2 or higher. Scenarios in which this embodiment is not applicable include cases where a specific threshold is exceeded when using metrics such as the correlation of z in memory or the Mean Absolute Error (MAE).

[0112] The second-1 embodiment illustrated in FIG. 13 shows a method of performing correlation extraction in one step, and the second-2 embodiment illustrated in FIG. 14 shows a method of performing correlation extraction in two steps, in the time domain and in the spatial and frequency domains.

[0113] Meanwhile, FIG. 15 is a diagram illustrating a switch operation for turning on / off a TSF module according to one embodiment of the present disclosure. According to one embodiment, when one or more spaces in the buffer memory are empty or padded, the TSF module operation may be skipped by the switch operation before an input is transferred to the TSF module.

[0114] Specifically, the conditions for switch off (skip TSF module) may include cases where the correlation between feedbacks within the buffer memory is below a certain threshold (e.g., when the MAE of z within the buffer memory exceeds a specific threshold), or cases where there are missing values ​​within the buffer memory. In this case, under the switch off condition, the operation of the TSF module is not performed, but z within the buffer memory x As a FIFO operation, it is necessary to perform the operation of storing in buffer memory in chronological order.

[0115] Meanwhile, the conditions for switch on (use TSF module) may include cases where there are no missing values ​​in memory and the correlation between feedback z is above a certain threshold, or cases where there are missing values ​​in memory but the correlation between feedback excluding the missing values ​​is above a certain threshold.

[0116] FIG. 16 is a diagram showing the processing operation of a TSF module when a missing input is included as an initial input, according to one embodiment of the present disclosure.

[0117] As illustrated in FIG. 16, if one or more spaces within the buffer memory are empty or padding, h xIt can be operated with padding (e.g., zero-padded). For example, if the current z_x is the initial input, the buffer memory space corresponding to a past time point other than z_x is empty, and the empty space of the said buffer memory can be padded and concatenated to be applied as input to the TSF module. In this case, h_x as the output of the TSF module may be zero-padded data, and accordingly, it may not affect the operation of the processing block. And the final V'_hat_x may be identical to V_hat_x.

[0118] According to an embodiment, if a switch does not exist, zero-padded or padded with a specific value input to the processing block in the embodiment shown in FIG. 16 may not affect the operation with V_x_hat (reconstructed CSI). Therefore, a value equal to the V_x_hat value may be the final output of the processing block.

[0119] Meanwhile, although switch on / off operations can be performed, h, which is the operation result of the TSF module x If it is zero-padded or padded with a specific value, it may not affect the operation with V_x_hat even when the switch is on. Therefore, the final output of the processing block may be the same value as V_x_hat.

[0120] Meanwhile, FIG. 17 is a diagram showing a UE-side encoder and a network-side TSF neural network (NN) decoder according to a third embodiment of the present disclosure.

[0121] The third embodiment can use the result of the SF decoder as the input to a TSF NN for considering temporal correlation. Unlike the previously described embodiment, the third embodiment is characterized by using the pre-processed SF decoder output reconstructed CSI, rather than the latent, as the input to the model.

[0122] In the case of FIG. 17, V_x is input CSI information and can be used as an input to an encoder. The encoder included in the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to a base station where a decoder is located. The base station decoder can use z_x as an input to finally output V_hat_x. At this time, the SF dec of a past time point or feedback period outputs V_hat_x-1, V_hat_x-2, …, and the output can be stored in a buffer memory. Therefore, the reconstructed information of the current and past V_hat_x, V_hat_x-1, … can be applied as an input to the TSF NN artificial intelligence model to be output as V'_hat_x, which has improved reconstruction performance.

[0123] The specific operation of the third embodiment is described below. Specifically, FIG. 18 is a diagram illustrating the operation of a buffer memory related to a reconstructed CSI input. In order to use V_x_hat in the decoder of the network, it is necessary to define a buffer memory for storing V_x_hat and the operation of said buffer memory. The said buffer memory can perform the operation of storing and reading V_x_hat from memory in chronological order. The size of the buffer memory can be specified as a specific size and can perform a First in First out (FIFO) operation; if the buffer memory space is exceeded, V_x_hat stored at the earliest time (V_x-n_hat in FIG. 17) can be deleted. When information within the buffer memory is used as input to the decoder, the information can be processed by concatenating it in chronological order.

[0124] FIG. 19 is a diagram illustrating operations related to initial input processing according to an embodiment of the present disclosure. According to an embodiment, when using an initial input, the historical V_hat value may not exist in the buffer memory. Therefore, if null information exists in the initial input and in the buffer memory, the base station may process it instead using padding and use it for model input. In this case, the padding may be, for example, zero-padding, and may be added to maintain the size of the input to the decoder. V_hat_0 in FIG. 19 represents the initial input, and the padding may fill the empty space in the remaining buffer memory. In this case, V_hat_0 and the padding are concatenated to form a single data, and the formed single data can be used as the initial input of the TSF NN.

[0125] Meanwhile, FIG. 20 is a diagram illustrating an operation related to feedback error processing according to an embodiment of the present disclosure. When a feedback error (such as a decoding failure) occurs in z transmitted from a terminal, V_x_hat at that point in time may not be able to be stored in the buffer memory of the base station. Therefore, as illustrated in FIG. 20, there may be two methods: inputting (or storing) padded information into the memory at that point in time, or emptying the memory at that point in time and processing the data by padding the empty memory when the model is input. FIG. 20 represents a case in which z_x-1, which is an encoder output corresponding to V_x-1 according to an embodiment, fails to transmit and fails to decode at the base station, and the original data cannot exist at the location of V_hat_x-1 in the buffer memory. To process this, it may mean an operation to place an empty space at the location corresponding to V_hat_x-1 by padding or skipping.

[0126] Meanwhile, in the fourth embodiment of the present disclosure, a method is proposed to use the reconstructed CSI, which is the output of the SF decoder, as the input to the TSF recurrent neural network (RNN).

[0127] As illustrated in FIG. 21, on the NW side, at least one reconstructed CSI, which is the output of the SF decoder, can be used as the input to the TSF RNN. According to this fourth embodiment, all past SF decoder outputs can influence the current TSF RNN output. Therefore, if the channel changes abruptly, a reset process may be required to prevent previous outputs from affecting the current output. In the embodiment illustrated in FIG. 21, V_x is input CSI information and can be used as the input to the encoder. The encoder located at the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bitstring and can output a bitstring of z_x. z_x can be transmitted to the base station where the decoder is located. The base station decoder can use z_x as input to produce the final output of V_hat_x. At this time, the SF decoder at a past time point or feedback cycle is V_hat_x-1, V_hat_x-2, … , would have been output, and the output content can be stored in a buffer memory. The embodiment illustrated in FIG. 21 is a structure that continuously inputs V'_hat values ​​from a previous point in time, and may be in the form of a buffer memory having only one space capable of storing only one V'_hat. Accordingly, the current and past V'_hat_x-1 and the reconstructed information can be applied as inputs to a TSF RNN artificial intelligence model together with V_hat_x, which is the current SF decoder output, to output V'_hat_x with enhanced reconstruction performance. At this time, V'_hat_x-1 may contain continuous channel information from a past point in time.

[0128] Figure 22 is a diagram illustrating such a TSF RNN reset operation. In an RNN structure where all past channels can affect the current output, an operation for model reset is required to avoid being affected by past channels.

[0129] When a change in the channel state is detected, the base station may perform a TSF RNN reset operation. To perform such a reset of the present model, the base station may perform the reset by padding the input to the TSF RNN (e.g., zero-padding). In this case, the output value may be V'_hat_x, which is identical to the current V_hat_x used as input.

[0130] Meanwhile, the fifth embodiment of the present disclosure proposes a method for outputting both reconstructed CSI by SF and TSF domains.

[0131] FIG. 23 is a diagram illustrating a UE-side encoder and a network-side multi-output decoder according to the 5-1 embodiment. According to the 5-1 embodiment, a single decoder can simultaneously output V_x_hat, which is the result in the spatial-frequency domain, and V'_x_hat, which is the result in the temporal-spatial-frequency domain. In the decoder, NN3 uses only the current bitstring z_x as input, but NN1 can consider the correlation in the temporal domain by utilizing the past latent vector stored in the receiver's buffer memory. FIG. 23 is a detailed example of how the decoder operates. The decoder illustrated in FIG. 23 is characterized by being capable of performing both SF decoder operations and TSF decoder operations. V_x is the input CSI information of the current cycle and can be used as the input to the encoder. The encoder located at the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bitstring and can output the bitstring of z_x. Subsequently, z_x can be transmitted to the base station where the decoder is located. The preceding process is executed in previous cycles as well, and accordingly, the base station or base station decoder can store the previously transmitted information, z_x-1 and z_x-2, in a predefined memory. The base station decoder can use the currently transmitted z_x along with the previously transmitted z_x-1 and z_x-2 as inputs to produce the final output V'_hat_x. At this time, NN1 can process the zs collectively to output z'_x, which has the same size as z_x, z_x-1, ..., and this output can pass through NN2 to be finally output as V'_hat_x. NN1 is expected to play the role of extracting the temporal relationship between the current and previously transmitted zs. V'_hat_x may be data having the same size as the existing input V_x.NN3 can reconstruct V_hat_x using only the current z_x. V_hat_x can be data of the same size as the original input V_x. Since V'_hat_x utilizes past time relationships, it can exhibit higher reconstruction performance compared to V_hat_x, which does not use previously transmitted data.

[0132] FIG. 24 is a diagram illustrating a network-side multi-output decoder including a UE-side encoder, a processing block, and single-step correlation extraction according to the 5-2 embodiment. According to the 5-2 embodiment, a processing block may be added between NN1 and NN2 in the decoder. Specifically, the current z xConsidering the temporal correlation, operations such as weighted sums are applied to z'_x_hat, the output computed by NN1 in the latent vector domain, to output the final V'_x_hat. In the 5-2 embodiment, since a skip connection structure is used, there is the advantage that only the residuals excluding z_x need to be computed when z'_x_hat is output from NN1, and there is the advantage of low computational load as the operation extracts the temporal correlation in the latent vector domain, which is of relatively small dimension. Figure 24 is a detailed example of the decoder operation. The decoder illustrated in Figure 24 is characterized by being capable of performing both SF and TSF operations. V_x is the input CSI information of the current period and can be used as the input to the encoder. The encoder located at the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output the bit string of z_x. Subsequently, z_x can be transmitted to the base station where the decoder is located. The preceding process is executed in previous cycles as well, and accordingly, the base station or base station decoder stores the previously transmitted information z_x-1 and z_x-2 in a predefined memory. The base station decoder can use the currently transmitted z_x along with the previously transmitted z_x-1 and z_x-2 as inputs to produce the final output V'_hat_x. At this time, NN1 can output z'_x, which has the same size as z_x, z_x-1, ..., by processing the zs collectively. The above output can be input into the processing block along with z_x to produce z''_x. The above output can pass through NN2 to be finally output as V'_hat_x. NN1 is expected to play the role of extracting the temporal relationship between the current and previously transmitted zs. The processing block can facilitate learning by summing z_x. V'_hat_x is data that has the same size as the original input V_x.NN3 can reconstruct V_hat_x using only the current z_x. V_hat_x is data with the same size as the original input V_x. Because V'_hat_x utilizes past time relationships, it can demonstrate higher reconstruction performance compared to V_hat_x, which does not use previously transmitted data.

[0133] FIG. 25 is a diagram illustrating a network-side multi-output decoder including a UE-side encoder, a processing block, and two-step correlation extraction according to the 5-3 embodiment. According to the 5-3 embodiment, a structure is proposed in which a processing block is added between NN1 and NN2. Current z x h in a dimensionally increased state after being computed in NN1 in the latent vector domain x It can be output as. In NN2, h xh'_x, considering the temporal correlation between the past h_(x-1), …, can be output. Then, in the processing block, operations such as the weighted sum of h'_x and V_x_hat are applied to output the final V'_x_hat. According to the 5-3rd embodiment, since a skip connection structure is used, there is an advantage that when V'_x_hat is output from NN1, only the residuals excluding V_x_hat need to be calculated. Figure 25 is a detailed example of how a decoder operates. The decoder illustrated in Figure 25 is characterized by being capable of performing both SF and TSF operations. V_x is the input CSI information of the current cycle and can be used as the input to the encoder. The encoder located at the terminal is a pre-trained artificial intelligence model that compresses V_x into a single bit string and can output a bit string of z_x. z_x can be transmitted to the base station where the decoder is located. The preceding process is executed in the previous cycle as well, and accordingly, the base station or base station decoder stores the previously transmitted information z_x-1 and z_x-2 in a predefined memory. The base station decoder can produce the final output V'_hat_x by using the previously transmitted z_x and z_x-1 and z_x-2 as inputs along with the currently transmitted z_x. At this time, the zs are processed collectively to produce h_x, h_x-1, and h_x-2, which are information having a larger dimension than z_x and z_x-1, respectively, which can be output through NN1. These outputs can then pass through NN2 to be output as h'_x. NN2 is expected to play the role of extracting the temporal relationship between the current and previously transmitted hs. The above outputs are applied as inputs to the processing block along with V_hat_x, so that the final output V'_hat_x can be produced. V'_hat_x is data having the same size as the original input V_x.Because V'_hat_x utilizes past time relationships, it can demonstrate higher recovery performance compared to V_hat_x, which does not use previously transmitted data.

[0134] FIG. 26 is a diagram illustrating the operation of a multi-output decoder according to an embodiment of the present disclosure. FIG. 26 is a diagram illustrating which result to use from V_x_hat and V'_x_hat derived according to the aforementioned embodiments 5-1 to 5-3. To this end, the base station may analyze the correlation between historical z within the buffer memory and use the analysis result. For example, as shown in FIG. 26, the base station z x to z x-n Time correlation can be performed. According to one embodiment, z_x represents currently transmitted information, and z_x-1, ..., z_x-n represent information received from past time n and transmission cycles, which may refer to information stored in the buffer memory. In the embodiment of FIG. 26, the base station outputs two pieces of information considering SF and TSF, and can make a decision to select which output to use in advance through time correlation analysis between the z information in the buffer memory. The base station may select V_x_hat if the result of the time correlation is below or less than a preset or determined threshold. Additionally, the base station may select V'_x_hat if the result of the time correlation is above or greater than a preset or determined threshold. According to one embodiment, the base station may use metrics such as Hamming distance, MAE, and MSE for time correlation.

[0135] FIG. 27 is a diagram showing the predicted results of a simulation according to one embodiment of the present disclosure. Referring to FIG. 27, it can be seen that when a temporal spatial frequency decoder is used in a base station according to an embodiment proposed in the present disclosure, there is an improvement in performance compared to a general spatial frequency decoder.

[0136] Meanwhile, FIG. 28 is a diagram showing the structure of a general vanilla transformer. For example, the structure illustrated in FIG. 28 may be a diagram of a general transformer. The part illustrated on the left side of FIG. 28 may correspond to the encoder of the transformer. In one embodiment of the present disclosure, a method utilizing multi-head attention (self-attention) included in the encoder is proposed. In the diagram illustrated in FIG. 28, the content on the left corresponds to the encoder of the vanilla transformer, and the content on the right corresponds to the decoder. Input information corresponding to the input value to the encoder passes through the Input embedding block, and the corresponding output through the Input embedding block can be added to the positional encoding value and used as the input to the attention block. Positional encoding serves to assign an order to the input data of the transformer. The attention block may be composed of sub-blocks for multi-head attention, add&norm, and feedforward. The information input to the attention block may pass through N series-connected attention blocks defined in advance. Subsequently, the output and input of the aforementioned input information are added through a multi-head attention process that computes the correlation between the inputs, and the resulting sum can pass through a normalization block. In the subsequent process, the data passes through a feedforward block to increase its dimensionality before returning to its original dimension, and the final output of the attention block can be used as the input for the decoder's multi-head attention block. During this process, the encoder's attention block is repeated N times, and the input and output data sizes can be maintained identically.The decoder block on the right differs from the encoder block in that it can include a masked multi-head attention block, which obscures some data to prevent the recognition of future information at the present time. Furthermore, as the vanilla transformer is a large language model requiring prediction of the final output, it passes through a softmax function to produce the final output probability for each data point or word. The decoder's attention block also appears repeated N times in series, and the number of repetitions for the attention blocks in the encoder and decoder may differ.

[0137] FIG. 29a is a diagram illustrating the structure of a general transformer-based autoencoder. However, it is understood that in this disclosure, a model in which some layers or hyperparameters of FIG. 29a are modified may also be applied. The input CSI V_in, which is the input information of FIG. 29a, can pass through a Reshape block, then through a linear embedding block, be summed with the positional encoding value, and input into an attention block. The attention block illustrated in FIG. 29a may consist of a multi-head attention block, a layer normalization block, and an FCN block. The input and output of the multi-head attention block may be summed together and input into the layer normalization block. The output of the layer normalization block is input into the FCN block, and the input and output of the FCN block are summed and undergo layer normalization. Subsequently, the output may sequentially input / output FCN + reshape + flatten + tanh and pass through a quantization block for quantization. Afterwards, the final encoder output z is output, and z can be passed as input to the decoder.

[0138] The input z of the decoder sequentially passes through the FCN+LeakyRelu+reshape blocks and can be input into the linear embedding block. The output of the linear embedding block can be combined with the positional encoding value and input into the attention block of the decoder. The attention block illustrated in FIG. 29a can be composed of multi-head attention, layer normalization, and FCN blocks. The input and output of multi-head attention can be combined and input into layer normalization. The output of layer normalization is input into the FCN, and the input and output of the FCN are combined and pass through layer normalization. The output can sequentially pass through FCN+reshape+normalization to be output as the final output, reconstructed CSI(V_out).

[0139] Meanwhile, FIG. 29b is a diagram showing the structure of a transformer-based autoencoder proposed in the present invention according to one embodiment. However, it is understood that in the present disclosure, a model in which some layers or hyperparameters of FIG. 29b are modified may also be applied.

[0140] The input CSI V, which is the input information in Fig. 29b, is in the form of a complex number and is input into a pre-processing block for processing in an artificial intelligence model, and can be output as real number data [Re(V), Im(V)]. Subsequently, the output passes through an embedding block and can be input into a reshape block. The output of the reshape block can be input into an attention block. The attention block written in the figure consists of multi-head attention, layer normalization, and dense blocks, and the input and output of multi-head attention are summed and input into layer normalization. The output of layer normalization is input into dense, and the input and output of dense are summed and go through layer normalization. Subsequently, the output can pass through a quantization block for quantization by sequentially inputting / outputting dense, reshape, flatten, and tanh. Afterwards, z can be output as the final encoder output, and z can be passed as the input to the decoder. The attention block of the above encoder can be overridden by a predefined L_enc.

[0141] The decoder's input z sequentially passes through the dense and reshape blocks and can be input into the embedding block. The output of the embedding block passes through the reshape block and can be input into the decoder's attention block. The attention block described in the diagram consists of multi-head attention, layer normalization, and dense blocks, and the input and output of multi-head attention are summed and input into layer normalization. The output of layer normalization is input into dense, and the input and output of dense are summed and pass through layer normalization. The output can be output after sequentially passing through dense, reshape, and normalization. Since the output is in the form of a real number and the actual reconstructed final result is data in the form of a complex number, the output can finally pass through the post-processing block to be output as the final output, reconstructed CSI(V_hat).

[0142] FIG. 30 is a diagram illustrating the specific operation of multi-head attention in the structure of a general vanilla transformer. According to one embodiment, FIG. 30 illustrates an example in which multi-head attention operates in SF domain CSI compression. For example, it may correspond to an embodiment in which an antenna domain is used for embedding and a frequency domain (subband) region is treated as a sequence for computation. Due to the characteristics of multi-head attention, the size of the weight parameter is related to the embedding size and may be independent of the sequence length. Accordingly, the method proposed in this disclosure utilizes the characteristics of multi-head attention to extend the temporal domain along the sequence axis to use the input, and proposes a model training method suitable for this.

[0143] Meanwhile, the present disclosure may also include a model structure in which the size of the weight parameter does not change according to the input length of the temporal domain, rather than a model based on Transformer or multi-head attention. Such a model may also demonstrate performance improvement similar to the embodiments of the present disclosure by using the training method described below.

[0144] Figure 31 is a diagram showing the training procedure of a general SF encoder and SF decoder.

[0145] Meanwhile, FIG. 32a is a diagram illustrating the training procedure of an SF encoder and an SF / TSF decoder according to an embodiment of the present disclosure. The present disclosure proposes a decoder capable of covering both the SF domain and the TSF domain as a temporal-domain adaptive unified (TAU) decoder. The TAU decoder proposed in the present disclosure may be proposed as a Transformer-based temporal-domain robust (TTR) decoder.

[0146] For example, FIG. 32b is a diagram illustrating the inference and learning procedures of an SF encoder and an SF / TSF decoder according to one embodiment of the present disclosure. The present disclosure proposes a decoder capable of covering both the SF domain and the TSF domain as a Transformer-based temporal-domain robust (TTR) decoder.

[0147] According to the method proposed in this disclosure, a single fixed decoder can process inputs of various lengths. Therefore, the decoder can enable high-resolution CSI recovery by utilizing temporal-domain correlation. According to the method proposed in this disclosure, since the same terminal encoder and the same base station decoder can perform all the proposed operations, there are no additional standard signaling issues, and implementation is possible through the enhancement of the base station receiver. Since the base station model is also fixed as a single unit, a separate procedure for changing the base station model can be omitted, which can be an advantage in terms of network operations.

[0148] According to one embodiment of the present disclosure, the aforementioned multi-head attention can be used. In this case, the size of the weight parameters may be related only to the embedding dimension. The multi-head attention model can cover inputs of any length having a specific embedding dimension. When using the multi-head attention model, since embedding is performed along the tx antenna axis, the length of the input data in the temporal domain becomes irrelevant to the determination of the model's weight parameter size. By utilizing this feature, a single decoder can perform the role of a general SF decoder and, by processing inputs of various lengths, can also be utilized to perform the role of a TSF decoder.

[0149] During training, a loss function that optimizes the sum of the SGCS of CSI reconstruction values ​​for multiple inputs considering the temporal domain can be used. This differs from the fact that, generally, a loss function that optimizes only the SGCS of a single output for the current input is used.

[0150] However, the loss function described in FIGS. 32a and 32b is merely an example, and other loss functions or metric functions other than SGCS-based (e.g., based on MSE, NMSE, etc.) may be used. The definition of SGCS used in this disclosure is as shown in Equation 1 below.

[0151]

[0152] In the above mathematical formula 1, N_s and N_l represent the number of sub-bends and rank, respectively. The specific loss function used for model training in the present disclosure is as shown in the following mathematical formula 2.

[0153]

[0154] In the above mathematical formula 2, S represents the total number of samples, T per sample represents the CSI feedback of the last cycle, N represents the CSI feedback of the last cycle that considers temporal correlation extraction during learning, and k represents the number of historical feedbacks considered for calculating temporal correlation. Furthermore, in addition to the general learning method of optimizing the current output based on the current input, the present disclosure proposes a method of additionally utilizing past feedback during learning to help optimize the current output by additionally utilizing past feedback.

[0155] One embodiment of the present disclosure may use a system model in which the number of transmitting antennas N_t is greater than 1 and the number of receiving antennas N_r is less than 1, considering a frequency division duplex (FDD) multiple input multiple output (MIMO) system. Orthogonal frequency division multiplexing (OFDM) is configured over N_c subcarriers and consists of N_1 ranks, and the received signal of the n-th subcarrier and the l-th layer can be defined by the following Equation 3.

[0156]

[0157] of the above mathematical formula 3 (N_r by a complex number of size N_t) represents the channel matrix of the subcarrier, and p_(n,l) (a complex number of size N_t) represents the precoding matrix of the n-th subcarrier and the l-th layer. / epsilon_n(a complex number of size N_r) represents the noise.

[0158] To maximize the precoding gain, it is necessary to use a precoding matrix that satisfies the following mathematical equation 4.

[0159]

[0160] The left side of the above mathematical equation 4 represents the optimal precoding matrix, which means the precoding matrix with the greatest directionality with the channel. Therefore, the terminal needs to transmit the corresponding precoding matrix v_(n,l) (a complex number of size N_t). Since frequency resources need to be utilized at the subband level rather than the subcarrier level according to the actual 3GPP system, the eigenvector representing the optimal precoding matrix of the l-th layer of the m-th subband is expressed as v_(m,l) (a complex number of size N_t). For each m-th subband, the precoding matrices of all N_l layers can be expressed as shown in the following mathematical equation 5.

[0161]

[0162] In addition, the precoding matrix for the entire subband can be expressed by the following mathematical formula 6.

[0163]

[0164] In one embodiment of the present invention, the channel has a center frequency f_c and a moving speed v, and when the speed of light is c, the coherence time of the channel can be expressed by the following mathematical formula 7.

[0165]

[0166] The technique proposed in one embodiment of the present invention enables higher performance CSI reconstruction by utilizing channel coherence time to additionally utilize feedback information received from a past base station decoder as the input to a current decoder.

[0167] The AI ​​encoder proposed in one embodiment of the present invention can be expressed by the following mathematical formula 8, where the subscript t in mathematical formula 8 indicates information at time step t, f_(enc) indicates the encoder function, and z indicates the output of the encoder.

[0168]

[0169] The decoder function when k historical feedbacks are used as inputs to the decoder can be defined by Equation 9 below. In Equation 9, tk represents the information at the tk time step, and f_(dec) represents the decoder function. Left side (N_sN_l by N_t, a complex number of size) represents the decoder output, which is the reconstructed CSI when k historical feedbacks are used as inputs.

[0170]

[0171] The value of k can be up to the Nth time step of a past time point that falls within a time shorter than T_c, allowing a total of natural numbers from 0 to N to be used for learning. In one embodiment of the present invention, the decoder input proposed is input in a manner that concatenates the current feedback with the past feedback (Cat(z_(tk), ..., z_t)). When k=0, it means that past information is not used, and the TTR decoder proposed in one embodiment of the present invention can process all possible inputs for all k situations. Although V, which is used as the input to the terminal encoder, is data composed of complex numbers, actual artificial intelligence models can only process information in the form of real numbers; therefore, when using the encoder input of one embodiment of the present invention, the real and imaginary parts of V are separated and represented as real numbers, [Re(V), Im(V)] (real numbers of size N_sN_l by 2N_t), is used as input. Re() and Im() each represent a function that extracts only the real and imaginary parts of the information within the parentheses and converts them into real numbers.

[0172] When training the autoencoder proposed in one embodiment of the present invention, the parameters of the artificial intelligence model ( ) is updated. The model parameters of the autoencoder ( ) and its encoder parameters and ( ) and decoder parameters( ) can be defined by the following mathematical formula 10.

[0173]

[0174] In one embodiment of the present invention, when the function of the proposed autoencoder is designed as f, the autoencoder including a decoder that uses the historical feedback of the final k lightning bolts as input can be defined with model parameters as shown in Equation 11 below.

[0175]

[0176] The inference of Fig. 32b above is performed for one k value, and the TTR decoder proposed in one embodiment of the present invention is characterized by the ability to restore CSI with improved performance for all k values ​​between 0 and N during the inference process due to the influence of the loss function used during the learning process.

[0177] Through the learning method proposed in the present disclosure, it is possible to handle existing SF domain inputs and also process inputs with added historical feedback, thereby enabling higher resolution CSI recovery.

[0178] FIGS. 33a to 33c are drawings showing the predicted results of a simulation according to an embodiment of the present disclosure. The simulation results of FIGS. 33a and 33b represent the SGCS performance evaluation of a general SF domain decoder and a TTR decoder. In FIGS. 33a and 33b, the circle (○) markers represent the results for the TTR decoder proposed in the present disclosure, and the triangle (△) markers represent the results for a general SF domain decoder.

[0179] In addition, the same color in FIGS. 33a and FIGS. 33b means that it is the result of using a decoder with the same computational complexity.

[0180] As shown in FIGS. 33a and 33b, it can be confirmed that the TTR decoder proposed in this disclosure demonstrates performance superiority in all measurement domains at the same complexity. Furthermore, since the computational complexity and model size of the SF domain decoder increase together, the TTR decoder has the advantage of maintaining a small number of parameters. In FIGS. 33a and 33b, L1 of Rank 2 refers to layer 1, and L2 refers to layer 2.

[0181] In addition, FIG. 33c is a graph showing the SGCS of Rank 2, with the layer 1 and layer 2 cases from left to right, respectively. The solid line represents the TTR Decoder proposed in the present invention, and the dotted line represents the performance of the baseline technique. The enlarged subplot of FIG. 33c is a diagram enlarged based on a feedback size of 431 bits.

[0182] Meanwhile, FIG. 34 is a flowchart illustrating the operation of a base station according to an embodiment of the present disclosure. In step S3410, the base station may receive a latent vector for a channel state information (CSI) feedback signal from a terminal. Then, in step S3420, the base station may store the received latent vector in a buffer memory. In step S3430, the base station may obtain an eigenvector at any time based on the latent vector stored in the buffer memory for a preset time.

[0183] Meanwhile, FIG. 35 is a block diagram showing the configuration of a base station according to one embodiment of the present disclosure.

[0184] FIG. 36 is a diagram showing the downlink throughput performance of a TTR Decoder and a baseline at high complexity according to one embodiment of the present disclosure.

[0185] FIG. 37 is a diagram illustrating the detailed hyperparameters of a transformer-based autoencoder proposed in the invention according to one embodiment of the present invention.

[0186] FIG. 38 is a diagram showing detailed parameters used in the simulation environment of the present invention.

[0187] FIG. 39 is a diagram showing the complexity and model size of a TTR decoder and a baseline decoder according to an embodiment of the present invention.

[0188] Meanwhile, FIG. 35 is a drawing illustrating the structure of a base station according to one embodiment of the present invention.

[0189] Referring to FIG. 35, the base station may include a transceiver (3510), a control unit (3520), and a storage unit (3530). In the present invention, the control unit (3520) may be defined as a circuit or application-specific integrated circuit or at least one processor.

[0190] The transmitting and receiving unit (3510) can transmit and receive signals with other network entities. The transmitting and receiving unit (3510) can, for example, receive CSI feedback from a terminal.

[0191] The control unit (3520) can control the overall operation of the base station according to the embodiment proposed in the present invention. For example, the control unit (3520) can control the signal flow between each block to perform operations according to the flowchart described above. Specifically, the control unit (3520) can control the operation to restore the CSI feedback of the base station according to the embodiment of the present invention.

[0192] The storage unit (3530) can store at least one of the information transmitted and received through the transmission and reception unit (3510) and the information generated through the control unit (3520). For example, the storage unit (3530) can be a buffer memory according to one embodiment of the present disclosure and can store a latent vector for a CSI feedback signal received from a terminal.

[0193] 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.

[0194] 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.

[0195] 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 such 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 the components, regardless of order or importance, and are used only to distinguish one component from another and do not limit the components. When it is mentioned that a certain (e.g., 1st) component is "(functionally or telecommunicationally) connected" or "connected" to another (e.g., 2nd) component, said certain component may be directly connected to said other component or connected through another component (e.g., 3rd component).

[0196] 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).

[0197] 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.

[0198] 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.

[0199] 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. In a decoding method performed by a base station in a wireless communication system, A step of receiving a latent vector for a channel state information (CSI) feedback signal or a bit signal corresponding to the latent vector from a terminal; The step of storing the received latent vector or a bit corresponding to the latent vector in a buffer memory; and A method comprising the step of obtaining reconstructed CSI (channel state information) at any time based on at least one latent vector stored in the buffer memory for a preset time.

2. In Paragraph 1, The step of obtaining the above-mentioned reconstructed CSI is, A step of obtaining a higher-dimensional matrix for each of a plurality of latent vectors stored in the buffer memory during the above-determined time; and A method further comprising the step of obtaining the reconstructed CSI based on correlations in the temporal domain, spatial domain, and frequency domain for each of the plurality of latent vectors.

3. In Paragraph 1, The step of obtaining the above-mentioned reconstructed CSI is, A step of obtaining one latent vector based on the correlation with respect to the temporal domain for a plurality of latent vectors stored in the buffer memory for a preset time; and A method characterized by further including the step of obtaining the reconstructed CSI based on the spatial domain and the frequency domain for the above-mentioned latent vector.

4. In Paragraph 1, A step of obtaining a first historical reconstructed CSI based on correlations in the spatial domain and frequency domain for a latent vector received at an arbitrary point in time; A step of obtaining correlations for the temporal domain, spatial domain, and frequency domain for a plurality of latent vectors received during a preset time prior to the arbitrary point in time above and the arbitrary point in time above; and The method further includes the step of obtaining the reconstructed CSI based on the first historical reconstructed CSI and the obtained correlation, wherein A method characterized in that if the correlation between the plurality of latent vectors received during a preset time prior to the arbitrary point in time and the arbitrary point in time is less than a threshold value, the acquired correlation is not used to acquire the reconstructed CSI.

5. In Paragraph 1, For each latent vector received at multiple time points, a step of obtaining a historical reconstructed CSI based on the spatial domain and the frequency domain; A step of sequentially storing the acquired historical reconstructed CSI for each latent vector received at multiple points in time into a buffer memory; and A method further comprising the step of obtaining the reconstructed CSI based on at least one historical reconstructed CSI stored in the buffer memory for a preset time.

6. In Paragraph 1, For the received latent vector above, a step of obtaining a first historical reconstructed CSI based on the spatial domain and the frequency domain; A method further comprising the step of determining at least one of the historical reconstructed CSI obtained at the arbitrary time and the first historical reconstructed CSI based on preset conditions.

7. In Paragraph 6, The above-mentioned determining step is, A step of performing a time domain correlation between at least one latent vector stored in the buffer memory during the above preset time; and A method further comprising the step of determining the first historical reconstructed CSI when the correlation result for the time domain is smaller than a threshold, and determining the historical reconstructed CSI obtained at the arbitrary time when the correlation result for the time domain is larger than a threshold.

8. In a base station of a wireless communication system, At least one transceiver; Buffer memory; At least one processor connected to the above at least one transceiver so as to be able to communicate; and The base station is connected to communicate with at least one processor and is capable of executing individually or in any combination of the at least one processor. A latent vector for a channel state information (CSI) feedback signal or a bit signal corresponding to the latent vector is received from a terminal through the at least one transceiver, and The received latent vector or the bit corresponding to the latent vector is stored in the buffer memory, and A memory storing an instruction to obtain reconstructed CSI (channel state information) at any time based on at least one latent vector stored in the buffer memory for a preset time; Base station including 9. In Paragraph 8, The above command is, the base station Acquire a higher-dimensional matrix for each of the plurality of latent vectors stored in the buffer memory during the above preset time, and A base station characterized by obtaining the reconstructed CSI based on correlations in the temporal domain, spatial domain, and frequency domain for each of the above-mentioned multiple latent vectors.

10. In Paragraph 8, The above command is, the base station For a plurality of latent vectors stored in the buffer memory for a preset time, one latent vector is obtained based on the correlation with respect to the temporal domain, and A base station characterized by obtaining the reconstructed CSI based on the spatial domain and frequency domain for the above-mentioned latent vector.

11. In Paragraph 8, The above command is, the base station For a latent vector received at an arbitrary point in time, a first historical reconstructed CSI is obtained based on correlations in the spatial domain and frequency domain, and Correlation is obtained for a plurality of latent vectors received during the aforementioned arbitrary point in time and a preset time prior to the aforementioned arbitrary point in time for the temporal domain, spatial domain, and frequency domain, and The reconstructed CSI is obtained based on the first historical reconstructed CSI and the obtained correlation, wherein A base station characterized in that if the correlation between the plurality of latent vectors received during the arbitrary point in time and the preset time prior to the arbitrary point in time is less than a threshold value, the acquired correlation is not used to acquire the reconstructed CSI.

12. In Paragraph 8, The above command is, the base station For each latent vector received at multiple time points, historical reconstructed CSI is obtained based on the spatial domain and frequency domain, and For each latent vector received at multiple points in time, the acquired historical reconstructed CSI is sequentially stored in the buffer memory, and A base station characterized by acquiring the reconstructed CSI based on at least one historical reconstructed CSI stored in the buffer memory for a preset time.

13. In Paragraph 8, The above command is, the base station For the received latent vector above, a first historical reconstructed CSI is obtained based on the spatial domain and the frequency domain, and A base station characterized by determining at least one of the historical reconstructed CSI obtained at the arbitrary time and the first historical reconstructed CSI based on preset conditions.

14. In Paragraph 13, The above command is, the base station Performing a time domain correlation between at least one latent vector stored in the buffer memory during the above preset time, and A base station characterized by determining the first historical reconstructed CSI when the correlation result for the above time domain is smaller than a threshold value.

15. In Paragraph 13, The above command is, the base station Performing a time domain correlation between at least one latent vector stored in the buffer memory during the above preset time, and A base station characterized by determining the historical reconstructed CSI obtained at the arbitrary time when the correlation result for the above time domain is greater than a threshold value.

Images