KR20240124218A - Method and apparatus for estimating channel using autoencoder in communication system - Google Patents

Abstract

A method of a UE according to the present disclosure may include the steps of receiving a CFI transmission period and an LVD of an encoder to perform online training from a base station, determining an encoder to perform online training based on the CFI transmission period and the LVD; receiving a first RS from the base station; generating a CFI by compressing the received first RS through the determined encoder; and transmitting the first CFI to the base station based on the CFI period.

Description

Method and Apparatus for Estimating Channel Using Autoencoder in Communication System {METHOD AND APPARATUS FOR ESTIMATING CHANNEL USING AUTOENCODER IN COMMUNICATION SYSTEM}

The present disclosure relates to improved communication technologies, and more particularly to technologies for channel estimation.

Communication networks (e.g., 5G communication networks, 6G communication networks, etc.) are being developed to provide improved communication services compared to existing communication networks (e.g., LTE (long term evolution), LTE-A (advanced), etc.). A 5G communication network (e.g., NR (new radio) communication network) can support not only a frequency band below 6 GHz but also a frequency band above 6 GHz. That is, the 5G communication network can support FR1 band and/or FR2 band. A 5G communication network can support various communication services and scenarios compared to an LTE communication network. For example, usage scenarios of a 5G communication network can include eMBB (enhanced Mobile BroadBand), URLLC (Ultra Reliable Low Latency Communication), mMTC (massive Machine Type Communication), etc.

6G communication networks can support various communication services and scenarios compared to 5G communication networks. 6G communication networks can satisfy requirements of ultra-performance, ultra-bandwidth, ultra-space, ultra-precision, ultra-intelligence, and/or ultra-reliability. 6G communication networks can support various and wide frequency bands and can be applied to various usage scenarios (e.g., terrestrial communication, non-terrestrial communication, sidelink communication, etc.).

Meanwhile, the current 5G NR uses a codebook-based CSI feedback method to estimate downlink Channel State Information (CSI). Codebook-based CSI feedback may mean a method in which a base station periodically or aperiodically transmits a downlink Channel State Information-Reference Signal (CSI-RS) to a UE, and the UE transmits CSI to the base station. In this case, the CSI may include a Channel Quality Indicator (CQI), a Precoding Matrix Indicator (PMI), and a Rank Indication (RI). The base station can estimate CSI using a predefined codebook based on the CQI, PMI, RI, etc. reported by the UE.

On the other hand, 3GPP is discussing an artificial intelligence/machine learning (AI/ML) method for the physical layer as a study item (SI) of Release 18. The AI/ML method learns output data for input data to extract specific patterns or parameters, and can use these to predict output data for arbitrary input data. In 3GPP, discussions are underway to apply the AI/ML method to technologies such as CSI feedback, beam management, and positioning accuracy enhancements.

However, the methods discussed or agreed upon in 3GPP so far only roughly propose techniques for improving AI/ML performance, and have not yet discussed specific procedures for utilizing them. Therefore, in order to support an AI/ML-based CSI feedback method instead of the existing codebook-based CSI feedback method using CQI, PMI, RI, etc., a new CSI feedback procedure that reflects the characteristics of AI/ML better than the currently discussed agreements or proposals and a definition of related parameters are needed.

The purpose of the present disclosure to solve the above problems is to provide a method and device for CSI feedback using AI/ML in a communication system.

According to an embodiment of the present disclosure for achieving the above object, a method of a user equipment (UE) may include: receiving a channel feature indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder to perform online learning from a base station; determining a first encoder to perform online learning based on the CFI transmission period and the LVD; receiving a first reference signal (RS) from the base station; generating a CFI by compressing the received first RS through the first encoder; and transmitting the first CFI to the base station based on the CFI transmission period.

The above first CFI can be determined as the product of the number of nodes in the latent space of the first encoder and the LVD.

The above LVD may be the number of bits required to express each latent variable value included in the latent space of the first encoder.

The above CFI transmission period can be determined based on at least one of the network size or channel coherent time of the first encoder.

The above LVD may be determined based on at least one of the channel resolution or latency requirement of the base station.

The method may further include: receiving re-learning related information of the first encoder from the base station; resetting the first encoder based on the re-learning related information; receiving a second RS from the base station; generating a second CFI by compressing the received second RS with the reset first encoder; and transmitting the second CFI to the base station.

The above relearning related information may include at least one of relearning instruction information of the encoder, drop-out node information of the encoder, or identifier information of the first encoder.

The identifier information of the above first encoder may further include the number of dropouts.

The method may further include: receiving learning termination instruction information of the first encoder from the base station; updating the first encoder based on information learned by the first encoder immediately before receiving the learning termination instruction information; receiving a third RS from the base station; generating a third CFI by compressing the third RS using the updated first encoder; and transmitting the third CFI to the base station.

The method may further include: receiving learning expiration information including optimal encoder information from the base station; updating the first encoder based on the optimal encoder information from the base station; setting the updated first encoder as a channel estimation encoder; receiving a fourth RS from the base station; generating a fourth CFI by compressing the fourth RS using the updated first encoder; and transmitting the fourth CFI to the base station.

The above optimal encoder information may include an encoder identifier and information on the number of drop-outs of the encoder.

A method of a base station according to an embodiment of the present disclosure may include: determining a channel feature indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder to perform online learning with user equipment (UE); transmitting the CFI transmission period and the LVD to the UE; receiving a first CFI from the UE based on the CFI transmission period; and restoring a reception RS using the received first CFI.

The method may include: obtaining a first reference signal received power (RSRP) value of the restored reception RS; checking whether the first RSRP is saturated; checking whether the first RSRP is equal to or greater than a preset threshold value when the first RSRP is saturated; and transmitting learning completion information of the encoder to the UE when the first RSRP is equal to or greater than the preset threshold value.

The above first CFI can be determined as the product of the number of nodes in the latent space of the encoder that performs the online learning and the LVD.

If the first RSRP value is equal to or within a predetermined range of RSRP values obtained from a CFI previously received from the UE, the first RSRP may be determined to be saturated.

The method may further include a step of generating re-learning related information of the encoder when the first RSRP is not saturated; and a step of transmitting the re-learning related information of the encoder to the UE; and a step of performing an encoder re-learning procedure with the UE.

The above relearning related information may include at least one of relearning instruction information of the encoder, drop-out node information of the encoder, or identifier information of the encoder.

The identifier information of the above encoder may further include the number of dropouts.

The method may further include: checking the number of dropouts corresponding to an RSRP having a highest RSRP value among learned RSRP values when relearning is not completed within a predetermined time; and transmitting encoder identifier information including the number of dropouts and learning expiration information to the UE.

The above CFI transmission period can be determined based on at least one of the network size or channel coherence time of the encoder performing the online learning.

The above LVD may be determined based on at least one of the channel resolution or latency requirement of the base station.

The method may further include: transmitting a second RS to the UE after the learning is completed; receiving a second CFI corresponding to the second RS from the UE; obtaining a second RSRP value of the received RS from the second CFI; and estimating a downlink channel to the UE using the second RSRP.

A user equipment (UE) according to an embodiment of the present disclosure comprises a transceiver configured to transmit and receive signals with a base station; and at least one processor, wherein the at least one processor:

The method may cause an encoder to perform online learning from a base station, receive a channel feature indicator (CFI) transmission period and a latent variable dimension (LVD); determine a first encoder to perform online learning based on the CFI transmission period and the LVD; receive a first reference signal (RS) from the base station; generate a CFI by compressing the received first RS through the first encoder; and transmit the first CFI to the base station based on the CFI transmission period.

The above first CFI can be determined as the product of the number of nodes in the latent space of the determined encoder and the LVD.

The CFI transmission period may be determined based on at least one of a network size or a channel coherent time of the first encoder, and the LVD may be determined based on at least one of a channel resolution or a latency requirement of the base station.

At least one processor of the above:

It may further cause the first encoder relearning related information to be received from the base station; the first encoder to be reset based on the relearning related information; the second RS to be received from the base station; the second CFI to be generated by compressing the received second RS with the reset first encoder; and the second CFI to be transmitted to the base station.

The above relearning related information includes at least one of relearning instruction information of the encoder, drop-out node information of the encoder, or identifier information of the encoder, and the identifier information of the first encoder may further include the number of drop-outs.

According to the present disclosure, when CSI feedback using an autoencoder is performed, since the UE compresses the received CSI-RS and reports it to the base station, there is an advantage of lower overhead than in the case of codebook-based CSI feedback. In addition, by appropriately determining the CSI period according to the channel situation and the hardware conditions of the UE, the autoencoder can be adaptively learned. In addition, by learning and relearning the autoencoder by dropping out nodes of the encoder, the learning of the autoencoder can be performed quickly.

Figure 1 is a conceptual diagram illustrating a first embodiment of a communication system.
FIG. 2 is a block diagram illustrating a first embodiment of a communication node constituting a communication system.
FIG. 3 is a block diagram illustrating a first embodiment of communication nodes performing communication.
FIG. 4a is a block diagram illustrating a first embodiment of a transmission path.
FIG. 4b is a block diagram illustrating a first embodiment of a receiving path.
Figure 5 is a conceptual diagram illustrating a first embodiment of a system frame in a communication system.
Figure 6 is a conceptual diagram illustrating a first embodiment of a subframe in a communication system.
Figure 7 is a conceptual diagram illustrating a first embodiment of a slot in a communication system.
Figure 8 is a conceptual diagram illustrating a first embodiment of time-frequency resources in a communication system.
Figure 9 is a conceptual diagram explaining the offline learning procedure of an autoencoder and the operation of a UE to acquire an autoencoder.
Figure 10 is a conceptual diagram explaining the encoder and decoder configuration of an autoencoder and a channel estimation scenario using an autoencoder.
Figure 11 is a conceptual diagram illustrating a dropout encoder model when dropout is instructed at a specific node of the encoder.
Figure 12 is a flowchart of operations between a base station and a UE during CFI feedback for online training.
Figure 13 is a flowchart explaining the procedure for conducting online training of an autoencoder at a base station.
Figure 14 is a flowchart explaining a case where saturation of RSRP values is determined at a base station.
Figure 15 is a flowchart illustrating the procedure for terminating online training of an autoencoder at a base station.
Figure 16 is a conceptual diagram explaining a case where the entire operation of the present disclosure is combined and operated.

The present disclosure may have various modifications and various embodiments, and thus specific embodiments are illustrated and described in detail in the drawings. However, this is not intended to limit the present disclosure to specific embodiments, but should be understood to include all modifications, equivalents, and alternatives included in the spirit and technical scope of the present disclosure.

The terms first, second, etc. may be used to describe various components, but the components should not be limited by the terms. The terms are only used to distinguish one component from another. For example, without departing from the scope of the present disclosure, the first component could be referred to as the second component, and similarly, the second component could also be referred to as the first component. The term "and/or" can mean a combination of a plurality of related listed items or any one of a plurality of related listed items.

In the present disclosure, "at least one of A and B" can mean "at least one of A or B" or "at least one of combinations of one or more of A and B." Additionally, in the present disclosure, "at least one of A and B" can mean "at least one of A or B" or "at least one of combinations of one or more of A and B."

In the present disclosure, (re)transmitting can mean “transmitting”, “retransmitting”, or “transmitting and retransmitting”, (re)setting can mean “setting”, “resetting”, or “setting and resetting”, (re)connecting can mean “connecting”, “reconnecting”, or “connecting and reconnecting”, and (re)connecting can mean “connecting”, “reconnecting”, or “connecting and reconnecting”.

When it is said that a component is "connected" or "connected" to another component, it should be understood that it may be directly connected or connected to that other component, but that there may be other components in between. On the other hand, when it is said that a component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

The terminology used in this disclosure is only used to describe specific embodiments and is not intended to be limiting of the present disclosure. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this disclosure, it should be understood that the terms "comprises" or "has" and the like are intended to specify that a feature, number, step, operation, component, part or combination thereof described in the specification is present, but do not exclude in advance the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Terms defined in commonly used dictionaries, such as those defined in common usage, should be interpreted as having a meaning consistent with the meaning they have in the context of the relevant art, and will not be interpreted in an idealized or overly formal sense unless expressly defined in this disclosure.

Hereinafter, with reference to the attached drawings, preferred embodiments of the present disclosure will be described in more detail. In order to facilitate an overall understanding in describing the present disclosure, the same reference numerals are used for the same components in the drawings, and redundant descriptions of the same components are omitted. In addition to the embodiments explicitly described in the present disclosure, operations according to combinations of embodiments, extensions of embodiments, and/or modifications of embodiments can be performed. The performance of some operations can be omitted, and the order of performing the operations can be changed.

In an embodiment, even if a method (e.g., transmitting or receiving a signal) performed by a first communication node among communication nodes is described, a second communication node corresponding thereto can perform a method (e.g., receiving or transmitting a signal) corresponding to the method performed by the first communication node. That is, if an operation of a UE (user equipment) is described, a base station corresponding thereto can perform an operation corresponding to the operation of the UE. Conversely, if an operation of a base station is described, a UE corresponding thereto can perform an operation corresponding to the operation of the base station.

The base station may be referred to as a NodeB, an evolved NodeB, a gNodeB (next generation node B), a gNB, a device, an apparatus, a node, a communication node, a BTS (base transceiver station), an RRH (radio remote head), a TRP (transmission reception point), a RU (radio unit), an RSU (road side unit), a radio transceiver, an access point, an access node, etc. The UE may be referred to as a terminal, a device, an apparatus, a node, a communication node, an end node, an access terminal, a mobile terminal, a station, a subscriber station, a mobile station, a portable subscriber station, an OBU (on-broad unit), etc.

In the present disclosure, signaling may be at least one of upper layer signaling, MAC signaling, or PHY (physical) signaling. A message used for upper layer signaling may be referred to as an "upper layer message" or an "upper layer signaling message". A message used for MAC signaling may be referred to as a "MAC message" or a "MAC signaling message". A message used for PHY signaling may be referred to as a "PHY message" or a "PHY signaling message". Upper layer signaling may refer to a transmission and reception operation of system information (e.g., a master information block (MIB), a system information block (SIB)) and/or an RRC message. MAC signaling may refer to a transmission and reception operation of a MAC control element (CE). PHY signaling may refer to a transmission and reception operation of control information (e.g., downlink control information (DCI), uplink control information (UCI), sidelink control information (SCI)).

In the present disclosure, "an operation (e.g., a transmission operation) is set" may mean that "setting information for the operation (e.g., an information element, a parameter)" and/or "information instructing performance of the operation" are signaled. "An information element (e.g., a parameter) is set" may mean that the information element is signaled. In the present disclosure, "a signal and/or a channel" may mean a signal, a channel, or "a signal and a channel," and a signal may be used to mean "a signal and/or a channel."

The communication network to which the embodiment is applied is not limited to what is described below, and the embodiment can be applied to various communication networks (e.g., 4G communication network, 5G communication network, and/or 6G communication network). Here, the communication network can be used in the same meaning as the communication system.

Figure 1 is a conceptual diagram illustrating a first embodiment of a communication system.

Referring to FIG. 1, the communication system (100) may include a plurality of communication nodes (110-1, 110-2, 110-3, 120-1, 120-2, 130-1, 130-2, 130-3, 130-4, 130-5, 130-6). In addition, the communication system (100) may further include a core network (e.g., a serving-gateway (S-GW), a packet data network (PDN)-gateway (P-GW), a mobility management entity (MME)). If the communication system (100) is a 5G communication system (e.g., a new radio (NR) system), the core network may include an access and mobility management function (AMF), a user plane function (UPF), a session management function (SMF), etc.

A plurality of communication nodes (110 to 130) can support a communication protocol specified in a 3rd generation partnership project (3GPP) standard (e.g., LTE communication protocol, LTE-A communication protocol, NR communication protocol, etc.). The plurality of communication nodes (110 to 130) may support CDMA (code division multiple access) technology, WCDMA (wideband CDMA) technology, TDMA (time division multiple access) technology, FDMA (frequency division multiple access) technology, OFDM (orthogonal frequency division multiplexing) technology, Filtered OFDM technology, CP (cyclic prefix)-OFDM technology, DFT-s-OFDM (discrete Fourier transform-spread-OFDM) technology, OFDMA (orthogonal frequency division multiple access) technology, SC (single carrier)-FDMA technology, NOMA (non-orthogonal multiple access) technology, GFDM (generalized frequency division multiplexing) technology, FBMC (filter bank multi-carrier) technology, UFMC (universal filtered multi-carrier) technology, SDMA (space division multiple access) technology, etc. Each of the plurality of communication nodes may have the following structure.

FIG. 2 is a block diagram illustrating a first embodiment of a communication node constituting a communication system.

Referring to FIG. 2, a communication node (200) may include at least one processor (210), a memory (220), and a transceiver device (230) that is connected to a network and performs communication. In addition, the communication node (200) may further include an input interface device (240), an output interface device (250), a storage device (260), etc. Each component included in the communication node (200) may be connected by a bus (270) and communicate with each other.

The processor (210) can execute a program command stored in at least one of the memory (220) and the storage device (260). The processor (210) may mean a central processing unit (CPU), a graphics processing unit (GPU), or a dedicated processor on which methods according to embodiments of the present disclosure are performed. Each of the memory (220) and the storage device (260) may be configured with at least one of a volatile storage medium and a nonvolatile storage medium. For example, the memory (220) may be configured with at least one of a read only memory (ROM) and a random access memory (RAM).

Referring again to FIG. 1, the communication system (100) may include a plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) and a plurality of terminals (130-1, 130-2, 130-3, 130-4, 130-5, 130-6). Each of the first base station (110-1), the second base station (110-2), and the third base station (110-3) may form a macro cell. Each of the fourth base station (120-1) and the fifth base station (120-2) may form a small cell. The fourth base station (120-1), the third terminal (130-3), and the fourth terminal (130-4) may be within the cell coverage of the first base station (110-1). The second terminal (130-2), the fourth terminal (130-4), and the fifth terminal (130-5) may be within the cell coverage of the second base station (110-2). The fifth base station (120-2), the fourth terminal (130-4), the fifth terminal (130-5), and the sixth terminal (130-6) may be within the cell coverage of the third base station (110-3). The first terminal (130-1) may be within the cell coverage of the fourth base station (120-1). The sixth terminal (130-6) may be within the cell coverage of the fifth base station (120-2).

Here, each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) may be referred to as a NodeB (NB), an evolved NodeB (eNB), a gNB, an advanced base station (ABS), a high reliability-base station (HR-BS), a base transceiver station (BTS), a radio base station, a radio transceiver, an access point, an access node, a radio access station (RAS), a mobile multihop relay-base station (MMR-BS), a relay station (RS), an advanced relay station (ARS), a high reliability-relay station (HR-RS), a home NodeB (HNB), a home eNodeB (HeNB), a road side unit (RSU), a radio remote head (RRH), a transmission point (TP), a transmission and reception point (TRP), etc.

Each of the plurality of terminals (130-1, 130-2, 130-3, 130-4, 130-5, 130-6) may be referred to as a user equipment (UE), terminal equipment (TE), advanced mobile station (AMS), high reliability-mobile station (HR-MS), terminal, access terminal, mobile terminal, station, subscriber station, mobile station, portable subscriber station, node, device, OBU (on board unit), etc.

Meanwhile, each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) may operate in a different frequency band or may operate in the same frequency band. Each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) may be connected to each other via an ideal backhaul link or a non-ideal backhaul link, and may exchange information with each other via the ideal backhaul link or the non-ideal backhaul link. Each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) may be connected to a core network via an ideal backhaul link or a non-ideal backhaul link. Each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) can transmit a signal received from the core network to the corresponding terminal (130-1, 130-2, 130-3, 130-4, 130-5, 130-6), and can transmit a signal received from the corresponding terminal (130-1, 130-2, 130-3, 130-4, 130-5, 130-6) to the core network.

In addition, each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) can support MIMO transmission (e.g., single user (SU)-MIMO, multi user (MU)-MIMO, massive MIMO, etc.), coordinated multipoint (CoMP) transmission, carrier aggregation (CA) transmission, transmission in an unlicensed band, sidelink communication (e.g., device to device communication (D2D), proximity services (ProSe)), Internet of Things (IoT) communication, dual connectivity (DC), etc. Here, each of the plurality of terminals (130-1, 130-2, 130-3, 130-4, 130-5, 130-6) can perform an operation corresponding to the base station (110-1, 110-2, 110-3, 120-1, 120-2) and an operation supported by the base station (110-1, 110-2, 110-3, 120-1, 120-2). For example, the second base station (110-2) can transmit a signal to the fourth terminal (130-4) based on the SU-MIMO scheme, and the fourth terminal (130-4) can receive a signal from the second base station (110-2) by the SU-MIMO scheme. Alternatively, the second base station (110-2) can transmit signals to the fourth terminal (130-4) and the fifth terminal (130-5) based on the MU-MIMO method, and each of the fourth terminal (130-4) and the fifth terminal (130-5) can receive signals from the second base station (110-2) by the MU-MIMO method.

Each of the first base station (110-1), the second base station (110-2), and the third base station (110-3) can transmit a signal to the fourth terminal (130-4) based on the CoMP scheme, and the fourth terminal (130-4) can receive a signal from the first base station (110-1), the second base station (110-2), and the third base station (110-3) based on the CoMP scheme. Each of the plurality of base stations (110-1, 110-2, 110-3, 120-1, 120-2) can transmit and receive a signal with terminals (130-1, 130-2, 130-3, 130-4, 130-5, 130-6) within its cell coverage based on the CA scheme. Each of the first base station (110-1), the second base station (110-2), and the third base station (110-3) can control sidelink communication between the fourth terminal (130-4) and the fifth terminal (130-5), and each of the fourth terminal (130-4) and the fifth terminal (130-5) can perform sidelink communication under the control of the second base station (110-2) and the third base station (110-3).

Meanwhile, communication nodes performing communication in a communication network can be configured as follows. The communication node illustrated in Fig. 3 may be a specific embodiment of the communication node illustrated in Fig. 2.

FIG. 3 is a block diagram illustrating a first embodiment of communication nodes performing communication.

Referring to FIG. 3, each of the first communication node (300a) and the second communication node (300b) may be a base station or a UE. The first communication node (300a) may transmit a signal to the second communication node (300b). The transmission processor (311) included in the first communication node (300a) may receive data (e.g., a data unit) from a data source (310). The transmission processor (311) may receive control information from the controller (316). The control information may include at least one of system information, RRC configuration information (e.g., information configured by RRC signaling), MAC control information (e.g., MAC CE), or PHY control information (e.g., DCI, SCI).

The transmitting processor (311) can perform a processing operation (e.g., an encoding operation, a symbol mapping operation, etc.) on data to generate data symbol(s). The transmitting processor (311) can perform a processing operation (e.g., an encoding operation, a symbol mapping operation, etc.) on control information to generate control symbol(s). In addition, the transmitting processor (311) can generate synchronization/reference symbol(s) for a synchronization signal and/or a reference signal.

The Tx MIMO processor (312) can perform spatial processing operations (e.g., a precoding operation) on data symbol(s), control symbol(s), and/or synchronization/reference symbol(s). An output (e.g., a symbol stream) of the Tx MIMO processor (312) can be provided to modulators (MODs) included in the transceivers (313a to 313t). The modulators (MODs) can perform processing operations on the symbol streams to generate modulation symbols and perform additional processing operations (e.g., an analog conversion operation, an amplification operation, a filtering operation, an upconversion operation) on the modulation symbols to generate signals. The signals generated by the modulators (MODs) of the transceivers (313a to 313t) can be transmitted via the antennas (314a to 314t).

The signals transmitted by the first communication node (300a) may be received by the antennas (364a to 364r) of the second communication node (300b). The signals received by the antennas (364a to 364r) may be provided to the demodulators (DEMODs) included in the transceivers (363a to 363r). The demodulator (DEMOD) may perform a processing operation (e.g., a filtering operation, an amplification operation, a down-conversion operation, a digital conversion operation) on the signal to obtain samples. The demodulator (DEMOD) may perform an additional processing operation on the samples to obtain symbols. The MIMO detector (362) may perform a MIMO detection operation on the symbols. The receiving processor (361) may perform a processing operation (e.g., a deinterleaving operation, a decoding operation) on the symbols. The output of the receiving processor (361) may be provided to a data sink (360) and a controller (366). For example, data may be provided to the data sink (360) and control information may be provided to the controller (366).

Meanwhile, the second communication node (300b) can transmit a signal to the first communication node (300a). The transmitting processor (368) included in the second communication node (300b) can receive data (e.g., data units) from a data source (367) and perform a processing operation on the data to generate data symbol(s). The transmitting processor (368) can receive control information from the controller (366) and perform a processing operation on the control information to generate control symbol(s). In addition, the transmitting processor (368) can perform a processing operation on a reference signal to generate reference symbol(s).

The Tx MIMO processor (369) can perform spatial processing operations (e.g., precoding operations) on data symbol(s), control symbol(s), and/or reference symbol(s). An output (e.g., a symbol stream) of the Tx MIMO processor (369) can be provided to modulators (MODs) included in the transceivers (363a to 363t). The modulators (MODs) can perform processing operations on the symbol streams to generate modulation symbols and can perform additional processing operations (e.g., an analog conversion operation, an amplification operation, a filtering operation, an upconversion operation) on the modulation symbols to generate signals. The signals generated by the modulators (MODs) of the transceivers (363a to 363t) can be transmitted via the antennas (364a to 364t).

The signals transmitted by the second communication node (300b) may be received by the antennas (314a to 314r) of the first communication node (300a). The signals received by the antennas (314a to 314r) may be provided to the demodulators (DEMODs) included in the transceivers (313a to 313r). The demodulator (DEMOD) may perform a processing operation (e.g., a filtering operation, an amplification operation, a down-conversion operation, a digital conversion operation) on the signal to obtain samples. The demodulator (DEMOD) may perform an additional processing operation on the samples to obtain symbols. The MIMO detector (320) may perform a MIMO detection operation on the symbols. The receiving processor (319) may perform a processing operation (e.g., a deinterleaving operation, a decoding operation) on the symbols. The output of the receiving processor (319) may be provided to a data sink (318) and a controller (316). For example, data may be provided to the data sink (318) and control information may be provided to the controller (316).

Memories (315 and 365) can store data, control information, and/or program code. Scheduler (317) can perform scheduling operations for communications. Processors (311, 312, 319, 361, 368, 369) and controllers (316, 366) illustrated in FIG. 3 may be processors (210) illustrated in FIG. 2 and may be used to perform the methods described in the present disclosure.

FIG. 4a is a block diagram illustrating a first embodiment of a transmission path, and FIG. 4b is a block diagram illustrating a first embodiment of a reception path.

Referring to FIGS. 4A and 4B, a transmission path (410) may be implemented in a communication node that transmits a signal, and a reception path (420) may be implemented in a communication node that receives a signal. The transmission path (410) may include a channel coding and modulation block (411), an S-to-P (serial-to-parallel) block (512), an N IFFT (Inverse Fast Fourier Transform) block (413), a P-to-S (parallel-to-serial) block (414), a CP (cyclic prefix) addition block (415), and an UC (up-converter) (UC) (416). The receiving path (420) may include a DC (down-converter) (421), a CP removal block (422), an S-to-P block (423), an N FFT block (424), a P-to-S block (425), and a channel decoding and demodulation block (426). Here, N may be a natural number.

In the transmission path (410), information bits may be input to a channel coding and modulation block (411). The channel coding and modulation block (411) may perform a coding operation (e.g., a low-density parity check (LDPC) coding operation, a polar coding operation, etc.) and a modulation operation (e.g., a quadrature phase shift keying (QPSK), a quadrature amplitude modulation (QAM), etc.) on the information bits. An output of the channel coding and modulation block (411) may be a sequence of modulation symbols.

The S-to-P block (412) can convert modulation symbols in the frequency domain into parallel symbol streams to generate N parallel symbol streams. N can be an IFFT size or an FFT size. The N IFFT block (413) can perform an IFFT operation on the N parallel symbol streams to generate signals in the time domain. The P-to-S block (414) can convert the output (e.g., parallel signals) of the N IFFT block (413) into a serial signal to generate a serial signal.

The CP addition block (415) can insert a CP into a signal. The UC (416) can up-convert the frequency of the output of the CP addition block (415) to an RF (radio frequency) frequency. Additionally, the output of the CP addition block (415) can be filtered at baseband before up-conversion.

A signal transmitted from the transmission path (410) may be input to the reception path (420). An operation in the reception path (420) may be an inverse operation of the operation in the transmission path (410). The DC (421) may down-convert the frequency of the received signal to a frequency of the baseband. The CP removal block (422) may remove a CP from the signal. An output of the CP removal block (422) may be a serial signal. The S-to-P block (423) may convert the serial signal into parallel signals. The N FFT block (424) may perform an FFT algorithm to generate N parallel signals. The P-to-S block (425) may convert the parallel signals into a sequence of modulation symbols. The channel decoding and demodulation block (426) may perform a demodulation operation on the modulation symbols and perform a decoding operation on the result of the demodulation operation to restore data.

In FIGS. 4A and 4B, Discrete Fourier Transform (DFT) and Inverse DFT (IDFT) can be used instead of FFT and IFFT. Each of the blocks (e.g., components) in FIGS. 4A and 4B can be implemented by at least one of hardware, software, or firmware. For example, some of the blocks in FIGS. 4A and 4B can be implemented by software, and the remaining blocks can be implemented by hardware or a “combination of hardware and software.” In FIGS. 4A and 4B, one block can be subdivided into multiple blocks, multiple blocks can be integrated into one block, some blocks can be omitted, and blocks supporting other functions can be added.

Figure 5 is a conceptual diagram illustrating a first embodiment of a system frame in a communication system.

Referring to FIG. 5, time resources in a communication system can be divided into frame units. For example, system frames can be set sequentially in the time domain of the communication system. The length of a system frame can be 10 ms (milliseconds). A system frame number (SFN) can be set from #0 to #1023. In this case, 1024 system frames can be repeated in the time domain of the communication system. For example, the SFN of a system frame after system frame #1023 can be #0.

A system frame may include two half frames. A half frame may be 5 ms long. A half frame located at the beginning of the system frame may be referred to as "half frame #0", and a half frame located at the end of the system frame may be referred to as "half frame #1". A system frame may include 10 subframes. A subframe may be 1 ms long. The 10 subframes within a system frame may be referred to as "subframe #0-9".

Figure 6 is a conceptual diagram illustrating a first embodiment of a subframe in a communication system.

Referring to FIG. 6, one subframe may include n slots, where n may be a natural number. Accordingly, one subframe may be composed of one or more slots.

Figure 7 is a conceptual diagram illustrating a first embodiment of a slot in a communication system.

Referring to FIG. 7, one slot may include one or more symbols. One slot illustrated in FIG. 7 may include 14 symbols. The length of a slot may vary depending on the number of symbols included in the slot and the length of the symbols. Alternatively, the length of a slot may vary depending on the numerology.

In a communication system, a numerology applied to a physical signal and a channel can be variable. The numerology can be variable to meet various technical requirements of the communication system. In a communication system to which CP (cyclic prefix)-based OFDM waveform technology is applied, the numerology can include a subcarrier spacing and a CP length (or a CP type). Table 1 may be a first embodiment of a method for configuring a numerology for a CP-OFDM-based communication system. At least some of the numerologies in Table 1 may be supported depending on a frequency band in which the communication system operates. In addition, the communication system may additionally support numerology(s) not listed in Table 1.

Subcarrier spacing 15 kHz 30 kHz 60 kHz 120 kHz 240 kHz 480 kHz OFDM
Symbol length [μs] 66.7 33.3 16.7 8.3 4.2 2.1 CP length [μs] 4.76 2.38 1.19 0.60 0.30 0.15 Number of OFDM symbols in 1ms 14 28 56 112 224 448

When the subcarrier spacing is 15 kHz (e.g., μ=0), the slot length can be 1 ms. In this case, one system frame can contain 10 slots. When the subcarrier spacing is 30 kHz (e.g., μ=1), the slot length can be 0.5 ms. In this case, one system frame can contain 20 slots.

When the subcarrier spacing is 60 kHz (e.g., μ=2), the length of a slot can be 0.25 ms. In this case, one system frame can contain 40 slots. When the subcarrier spacing is 120 kHz (e.g., μ=3), the length of a slot can be 0.125 ms. In this case, one system frame can contain 80 slots. When the subcarrier spacing is 240 kHz (e.g., μ=4), the length of a slot can be 0.0625 ms. In this case, one system frame can contain 160 slots.

A symbol may be configured as a downlink (DL) symbol, a flexible (FL) symbol, or an uplink (UL) symbol. A slot composed of only DL symbols may be referred to as a "DL slot," a slot composed of only FL symbols may be referred to as a "FL slot," and a slot composed of only UL symbols may be referred to as a "UL slot."

The slot format can be semi-statically set by higher layer signaling (e.g., RRC signaling). Information indicating the semi-static slot format can be included in system information, and the semi-static slot format can be set cell-specifically. In addition, the semi-static slot format can be additionally set for each terminal by terminal-specific higher layer signaling (e.g., RRC signaling). A flexible symbol of a slot format set cell-specifically can be overridden with a downlink symbol or an uplink symbol by terminal-specific higher layer signaling. In addition, the slot format can be dynamically indicated by physical layer signaling (e.g., a slot format indicator (SFI) included in DCI). A semi-statically set slot format can be overridden by a dynamically indicated slot format. For example, a flexible symbol set semi-statically can be overridden with a downlink symbol or an uplink symbol by the SFI.

The reference signal may be a channel state information-reference signal (CSI-RS), a sounding reference signal (SRS), a demodulation-reference signal (DM-RS), a phase tracking-reference signal (PT-RS), etc. The channel may be a physical broadcast channel (PBCH), a physical downlink control channel (PDCCH), a physical downlink shared channel (PDSCH), a physical uplink control channel (PUCCH), a physical uplink shared channel (PUSCH), a physical sidelink control channel (PSCCH), a physical sidelink shared channel (PSSCH), etc. In the present disclosure, a control channel may mean a PDCCH, a PUCCH, or a PSCCH, and a data channel may mean a PDSCH, a PUSCH, or a PSSCH.

Figure 8 is a conceptual diagram illustrating a first embodiment of time-frequency resources in a communication system.

Referring to FIG. 8, a resource composed of one symbol (e.g., OFDM symbol) in the time domain and one subcarrier in the frequency domain may be defined as a "RE (resource element)". Resources composed of one OFDM symbol in the time domain and K subcarriers in the frequency domain may be defined as a "REG (resource element group)". A REG may include K REs. A REG may be used as a basic unit of resource allocation in the frequency domain. K may be a natural number. For example, K may be 12. N may be a natural number. In the slot illustrated in FIG. 7, N may be 14. The N OFDM symbols may be used as a basic unit of resource allocation in the time domain.

In the present disclosure, RB may mean CRB (common RB). Alternatively, RB may mean PRB or VRB (virtual RB). In a communication system, CRB may mean RB constituting a set of consecutive RBs (e.g., common RB grid) based on a reference frequency (e.g., point A). Carriers and/or bandwidth portions may be arranged on the common RB grid. That is, the carrier and/or bandwidth portions may be composed of CRB(s). RBs or CRBs constituting the bandwidth portions may be referred to as PRBs, and a CRB index within the bandwidth portion may be appropriately converted to a PRB index.

Downlink data can be transmitted via PDSCH. The base station can transmit configuration information (e.g., scheduling information) of the PDSCH to the terminal via PDCCH. The terminal can obtain the configuration information of the PDSCH by receiving the PDCCH (e.g., downlink control information (DCI)). For example, the configuration information of the PDSCH can include an MCS (modulation coding scheme) used for transmitting and receiving the PDSCH, time resource information of the PDSCH, frequency resource information of the PDSCH, feedback resource information for the PDSCH, etc. The PDSCH may refer to a radio resource through which downlink data is transmitted and received. Alternatively, the PDSCH may refer to the downlink data itself. The PDCCH may refer to a radio resource through which downlink control information (e.g., DCI) is transmitted and received. Alternatively, the PDCCH may refer to the downlink control information itself.

The terminal can perform a monitoring operation for the PDCCH in order to receive the PDSCH transmitted from the base station. The base station can inform the terminal of the configuration information for the monitoring operation of the PDCCH using a higher layer message (e.g., an RRC (radio resource control) message). The configuration information for the monitoring operation of the PDCCH can include CORESET (control resource set) information and search space information.

CORESET information may include PDCCH DMRS (demodulation reference signal) information, PDCCH precoding information, PDCCH occasion information, etc. The PDCCH DMRS may be a DMRS used to demodulate the PDCCH. The PDCCH occasion may be a region in which a PDCCH can exist. That is, the PDCCH occasion may be a region in which DCI can be transmitted. The PDCCH occasion may be referred to as a PDCCH candidate. The PDCCH occasion information may include time resource information and frequency resource information of the PDCCH occasion. The length of the PDCCH occasion in the time domain may be indicated in symbol units. The size of the PDCCH occasion in the frequency domain may be indicated in RB units (for example, in PRB (physical resource block) units or CRB (common resource block) units).

The search space information may include a CORESET ID (identifier) associated with the search space, a period of PDCCH monitoring, and/or an offset. Each of the period and offset of PDCCH monitoring may be indicated in slot units. In addition, the search space information may further include an index of a symbol at which a PDCCH monitoring operation starts.

A base station can set a BWP (bandwidth part) for downlink communication. The BWP can be set differently for each terminal. The base station can inform the terminal of the configuration information of the BWP using upper layer signaling. The upper layer signaling can mean "transmission operation of system information" and/or "transmission operation of RRC (radio resource control) message." The number of BWPs set for one terminal can be 1 or more. The terminal can receive configuration information of the BWP from the base station, and can identify the BWP(s) set by the base station based on the configuration information of the BWP. When a plurality of BWPs are set for downlink communication, the base station can activate one or more BWPs among the plurality of BWPs. The base station can transmit the configuration information of the activated BWP(s) to the terminal using at least one of upper layer signaling, MAC (medium access control) CE (control element), or DCI. The base station can perform downlink communication using the activated BWP(s). The terminal can identify the activated BWP(s) by receiving configuration information of the activated BWP(s) from the base station, and perform a downlink reception operation in the activated BWP(s).

Meanwhile, the current 5G NR uses a codebook-based CSI feedback method to estimate downlink Channel State Information (CSI). Codebook-based CSI feedback may mean a method in which a base station periodically or aperiodically transmits a downlink Channel State Information-Reference Signal (CSI-RS) to a UE, and the UE transmits CSI to the base station. In this case, the CSI may include a Channel Quality Indicator (CQI), a Precoding Matrix Indicator (PMI), and a Rank Indication (RI). The base station can estimate CSI using a predefined codebook based on the CQI, PMI, RI, etc. reported by the UE.

In the future, 6G communication systems are expected to utilize unlicensed bands of millimeter waves or terahertz bands to achieve ultra-high capacity/ultra-wideband/ultra-low latency communication services. In addition, it is expected that a frequency band that is tens of times wider than the FR2 band used in 5G NR will be used. Therefore, a CSI feedback process with higher accuracy and higher overhead may be required.

On the other hand, 3GPP is discussing an artificial intelligence/machine learning (AI/ML) method for the physical layer as a study item (SI) of Release 18. The AI/ML method learns output data for input data to extract specific patterns or parameters, and can use these to predict output data for arbitrary input data. In 3GPP, discussions are underway to apply the AI/ML method to technologies such as CSI feedback, beam management, and positioning accuracy enhancements.

Discussions are ongoing on CSI feedback, beam management, and improved positioning accuracy in 3GPP SI related to AI/ML. Among them, CSI feedback using AI/ML was primarily discussed in terms of overhead reduction, improved channel estimation accuracy, and CSI prediction compared to the existing 5G NR. According to AI/ML for improved CSI feedback in the Rel-18 RAN1 #109-e meeting, in order to support CSI feedback using AI/ML, the AI/ML network is configured as a two-sided model consisting of the CSI feedback part and the CSI reconstruction part, and it was agreed that the base station and UE each use either the CSI feedback part or the CSI reconstruction part to perform channel estimation.

Also discussed in 3GPP Release 18 were various technologies proposed to support AI/ML-based CSI feedback, such as reducing overhead compared to Type 2 codebook-based CSI feedback in the current 3GPP standard, the need for AI/ML-based CSI feedback with higher accuracy than that provided by the current 3GPP standard, and determining the quantization level of AI/ML models.

However, there is no discussion on specific methods or specific procedures for improving AI/ML performance among the methods agreed upon or proposed to date in 3GPP Release 18.

Therefore, in order to support AI/ML-based CSI feedback methods instead of codebook-based CSI feedback methods using CQI, PMI, RI, etc. in the current 3GPP standards, a new CSI feedback procedure and definition of related parameters that better reflect the characteristics of AI/ML than the currently discussed agreements or proposals are required. The following explanation will examine these contents.

According to the agreement in 3GPP Release 18, the AI/ML method being discussed for CSI feedback can be composed of a model training stage and a model inference stage. The model training stage can be a stage that uses learning data as input to learn an AI model and/or ML model and extracts parameters, etc. The model inference stage can be a stage that uses the parameters obtained in the training stage of the AI model and/or ML model to obtain an inference value (output) using the input data to be inferred. The model training stage is generally a computationally intensive stage with many repetitions and can be performed for a long time because it requires repetitive operations. In addition, the performance of model inference can be determined depending on the accuracy of the parameters learned in the model training stage.

In channel estimation, the AI model and/or ML model training step is a step in which learning is performed using input data to obtain parameters to be used for channel estimation, and the model inference step can be interpreted as a step in which channel estimation is performed in an actual channel environment using the parameters obtained in the AI model and/or ML model training step.

Meanwhile, 3GPP assumed online training and offline training methods as AI/ML model training procedures to support AI/ML. Online training is a method of creating new learning data in real time and learning AI/ML models through the learning data, and offline training is a method of learning AI/ML models through already collected data.

When online training is used for channel estimation techniques, real-time channel information is utilized in online training, so accurate parameters for the current channel between the base station and the UE can be obtained. Therefore, more accurate channel training is possible than offline training between the base station and the UE. However, since AI/ML model training generally takes a long time, the AI/ML model training time can be much longer than the channel coherence time. Therefore, channel estimation is impossible in a short time through the online training method.

When offline training is used for channel estimation techniques, the base station or the server conducting the training can use previously collected data to train the AI/ML model in advance, so channel estimation is possible within the channel coherent time. However, since offline training applies parameters learned through previously collected data to the AI/ML model for AI/ML model training, accurate channel estimation may not be possible.

As described above, one way to solve the problem of channel estimation and feedback methods using AI/ML techniques is to consider a method of combining online and offline training. For example, it can be a procedure for training an AI/ML model to derive parameters that can achieve a certain level of accurate channel estimation through offline training. Therefore, offline training can be a coarse training procedure. Afterwards, a procedure for fine tuning the AI/ML model through online training can be considered to reflect the real-time channel.

However, the channel estimation procedure and method for performing offline and online training together for AI models and/or ML models have not yet been agreed upon in 3GPP. In addition, when performing a channel estimation technique that performs online and offline training together for AI models and/or ML models, the parameters obtained through offline training may be overfitted to previously collected data, and even if online training is performed in real time, a situation may occur in which parameters satisfying a certain level of accurate channel estimation for the AI model and/or ML model cannot be obtained.

Therefore, the present disclosure described below will describe a method for providing a CSI feedback procedure that reflects the characteristics of an offline-trained AI/ML model in channel estimation using an AI/ML method. Specifically, the base station (or server) considers an environment in which online training is additionally performed on the AI/ML model to ensure a certain level of accurate channel estimation in a situation in which offline training is performed on the AI/ML model using collected data. In addition, we will look into a channel estimation technique that includes a process for determining the period and dimension of a feedback signal that the UE should report to the base station in the CSI feedback process and an online training-based relearning process utilizing dropout to solve the overfitting problem.

The present disclosure described below considers a situation in which a downlink channel is estimated using an AI/ML model. However, the present disclosure is not limited to the case of estimating only a downlink channel. It can also be applied to the case of estimating an uplink channel in a manner identical or similar to the method described in the present disclosure. However, in the following description, the case of downlink channel estimation will be described for convenience of explanation.

In the present disclosure described below, parameters are defined for AI/ML-based CSI feedback according to the present disclosure instead of at least one of CSI, e.g., CQI, PMI and RI, or CSI reported by a UE to a base station in a codebook-based CSI feedback process. Then, a procedure for performing online training of an AI/ML model by the UE reporting the parameters defined in the present disclosure to the base station will be described. For example, in order to utilize an AI/ML model for CSI feedback, the base station and the UE may use an autoencoder. At this time, an environment in which the base station has a decoder of the autoencoder and the UE has an encoder of the autoencoder is considered. Since this describes the CSI feedback of a downlink channel, it is assumed that the base station has a decoder and the UE has an encoder. However, if the uplink channel is also considered, the base station and the UE may each have both an encoder and a decoder.

The autoencoder described in the present disclosure may refer to a network that compresses input data into a signal having a small dimension and then restores the compressed data back to its original data form. Such an autoencoder may be composed of an encoder, which is a part that compresses data, and a decoder that restores the compressed data.

Then, below, we will examine the case where the autoencoder according to the present disclosure is utilized for downlink channel estimation.

The base station can transmit the CSI-RS to the UE over a wireless channel. Accordingly, the UE can receive the CSI-RS from the base station over the wireless channel. Since the CSI-RS received by the UE is received over the wireless channel, the CSI-RS received by the UE may be degraded (or distorted) due to the influence of the wireless channel. In this way, the CSI-RS received by the UE over the wireless channel will be referred to as 'received CSI-RS'. The UE can compress the received CSI-RS using an encoder. Then, the UE can transmit the compressed received CSI-RS to the base station. In addition, the UE can transmit the compressed received CSI-RS and additional information to the base station, if necessary. The base station can receive the compressed received CSI-RS or the compressed received CSI-RS and additional information from the UE. Then, the base station can restore the received CSI-RS by decoding the compressed received CSI-RS through a decoder. Since the base station knows the CSI-RS transmitted by the base station, the downlink channel can be estimated using the restored received CSI-RS and the CSI-RS transmitted by the base station. Then, the base station can train the autoencoder using the estimated channel information.

According to the method presented in the present disclosure, during the learning process of the autoencoder, the UE only needs to generate compressed received CSI-RS through the encoder and report it to the base station in the channel estimation procedure. Therefore, the uplink overhead of the UE according to the present disclosure has an advantage of being much lower than the overhead of transmitting codebook-based CSI feedback according to the current standard specification of 5G NR.

In the present disclosure, an autoencoder can be trained through offline training using previously collected data. The offline training can be performed at a base station or a specific server. In the present disclosure, for convenience of explanation, it will be assumed that the offline training of the autoencoder is performed at the base station. However, the offline training of the autoencoder can also be performed at the server. If the autoencoder is trained through offline training at the server, the server can provide the trained autoencoder to the base station in advance. In addition, the UE can receive the offline trained autoencoder from the base station or the server.

Figure 9 is a conceptual diagram explaining the offline learning procedure of an autoencoder and the operation of a UE to acquire an autoencoder.

Referring to FIG. 9, a database (930), a base station (910), and a UE (920) are illustrated. In order to simplify the drawing, FIG. 9 is an exemplary drawing assuming that the base station (910) has an autoencoder (911) and that the base station (910) performs offline learning of the autoencoder (911). However, a server (not shown in FIG. 9), not the base station (910), may perform offline learning of the autoencoder and provide the offline-learned autoencoder (911) to the base station (910).

The database (930) can store data collected for offline learning of the autoencoder (911). Since the present disclosure describes an autoencoder (911) for estimating a downlink channel, the collected data can be channel information. For example, the channel information can be channel information obtained through simulation. For another example, the channel information can be various channel information actually measured between each of the base stations and the UEs from each of the plurality of base stations. The collected data stored in the database (930) can be data preprocessed in a form for learning the autoencoder according to the present disclosure.

The base station (910) may have an autoencoder (911) according to the present disclosure. The autoencoder (911) may be composed of an encoder (9111) and a decoder (9112) as described above. For convenience of explanation, the decoder of the autoencoder will be described as a decoder (9112) and the encoder of the autoencoder will be described as an encoder (9111). The base station (910) may receive data collected from a database (930) and perform offline training of the autoencoder (911). Therefore, the base station (910) may perform offline training of the autoencoder using the collected data. The autoencoder may be an autoencoder having a compression and restoration technique.

Let us briefly look at the procedure in which the encoder (9111) and decoder (9112) are trained through offline training. The encoder (9111) can encode collected data received from the database (930). The encoded data can be provided to the decoder (9112) (step S941). Accordingly, the decoder (9112) can receive compressed data from the encoder (9111) (step S941).

The decoder (9112) can decode and restore the compressed data. And the decoder (9112) can obtain channel information using the restored information. The decoder (9112) can generate control information for adjusting the encoder (9111) and the decoder (9112) based on the obtained channel information. The control information for adjusting the encoder (9111) and the decoder (9112) will be described in more detail with reference to the drawings described below. And the decoder (9112) can transmit the control information to the encoder (9111) (S942). Therefore, the encoder (9111) can receive the control information from the decoder (9112) (S942).

The encoder (9111) can be trained by applying the received control information. The encoder (9111) can re-encode the collected data based on the encoding method to which the control information is applied. The subsequent procedure can be repeated in the same manner as the operation described above. This repetition can be repeated until the decoded information in the decoder (9112) reaches a preset level.

In other words, the encoding data generated by the encoding operation of the encoder (9111) is provided to the decoder (9112), and the decoder (9112) can decode the encoded data to check whether the channel estimation is performed properly. If the channel estimation is not performed properly, in other words, if the channel estimation is not performed within a desired error range, the encoder (9111) and/or the decoder (9112) can generate control information for controlling the operations of the encoder (9111) and the decoder (9112), and apply the control information to the encoder (9111) and/or the decoder (9112). The encoder (9111) and/or the decoder (9112) to which the control information for controlling the operations of the encoder (9111) and the decoder (9112) is applied can repeat the offline training procedure while reducing the error through steps S9141 and S942.

When offline learning of the encoder (9111) and the decoder (9112) is completed, the base station (910) can transmit the autoencoder (911) to be used for channel estimation or the encoder (9111) constituting the autoencoder to the UE (920). In Fig. 9, an example is given where only the encoder (9111) constituting the autoencoder is transmitted to the UE (920). It should be noted that since the encoder (9111) in the UE (920) and the encoder (9111) in the base station (920) are the same in Fig. 9, the same reference numeral is used.

It should be noted that in FIG. 9, only the autoencoder (911) that the base station (910) and the UE (920) can store or have is exemplified. Accordingly, the base station (910) may include part or all of the configuration of the communication node (200) described above in FIG. 2. If offline learning of the autoencoder (911) is performed in a specific server and the base station (910) receives the autoencoder (911) from the server, the base station (910) may further include an interface for communicating with the server in addition to the configuration of FIG. 2. In addition, the memory (220) of the base station (910) may store the autoencoder (911). The processor (210) of the base station (910) may be an entity that performs the operation and/or control of the autoencoder (911) described below.

The UE (920) may receive and store the offline-learned autoencoder (911) or the encoder (9111) of the offline-learned autoencoder (9111) from the base station as described above. As another example, the UE (920) may have stored the offline-learned encoder (9111) in advance. As another example, the UE (920) may receive the offline-learned autoencoder (911) or the encoder (9111) of the offline-learned encoder (9111) from the server described above.

If the UE (920) has stored the offline-learned encoder (9111) in advance, the UE (920) can update the encoder (9111) based on the version information or the offline learning date information of the offline-learned encoder (9111). For example, if the UE (920) has stored the offline-learned encoder (9111) in advance, the UE (920) can receive (in advance) the version information or the offline learning date information of the autoencoder (911) that the base station (910) intends to use from the base station (910).

If the version information of the autoencoder (911) received from the base station (910) and the version information of the encoder (9111) stored in the UE (920) are the same, the UE (920) can use the stored encoder (9111) as is. On the other hand, if the version information of the autoencoder (911) received from the base station (910) and the version information of the encoder (9111) stored in the UE (920) are different (even if the version of the autoencoder (911) of the base station (910) is low), the UE (910) can receive and update the encoder (9111) from the base station (910).

It should be noted that the configuration of the UE (920) in FIG. 9 exemplifies a configuration including only the encoder (9111) of the autoencoder (911). The UE (920) may include part or all of the configuration of the communication node (200) described above in FIG. 2. Additionally, the UE (910) may further include an interface and/or sensor(s) for the convenience of the user. In addition, the encoder (9111) of the autoencoder (911) may be stored in the memory (220) of the UE (910). The processor (210) of the UE (920) may perform control for the operation of the autoencoder (911) described in the present disclosure.

As discussed above in Fig. 9, the offline learning of the autoencoder (911), or in other words, the procedure in which the encoder (9111) and the decoder (9112) learn through offline training, may be referred to as 'primary learning' or 'pre-learning' or 'offline learning'.

The base station (910) can learn the autoencoder (911) offline by utilizing the data collected in advance, and the offline learned encoder (9111) can be transmitted to the UE (920). In this way, the first learning or pre-learning or offline learning can be learned through coarse training for channel estimation.

Next, we will look at the encoder (9111) and decoder (9112) that make up the autoencoder (911).

Figure 10 is a conceptual diagram explaining the encoder and decoder configuration of an autoencoder and a channel estimation scenario using an autoencoder.

Referring to Fig. 10, an encoder (9111) of an autoencoder (910), a wireless channel (1001), and a decoder (9112) of the autoencoder (9111) are illustrated. Therefore, the encoder (9111) and the decoder (9112) may be an encoder and a decoder in which offline learning is performed, respectively.

As described in FIG. 9, the encoder (9111) may be carried by the UE (920), and the decoder (9112) may be carried by the base station (910). This is because the present disclosure assumes that the autoencoder (911) is used for downlink channel estimation. In the case of uplink channel estimation, the encoder (9111) may be carried by the base station (910), and the decoder (9112) may be carried by the UE (920). When both downlink channel estimation and uplink channel estimation are performed, the base station (910) and the UE (920) may each have both an offline-learned encoder (9111) and an offline-learned decoder (9112).

Referring to FIG. 10, the encoder (9111) may include an input layer (1010), a hidden layer (1020), and an output layer (1030). In FIG. 10, for convenience of explanation, the case where the hidden layer (1020) is composed of one layer is exemplified. However, the hidden layer (1020) may be composed of two or more layers. In general, as the number of hidden layers (1020) increases, the processing time of the encoder (9111) increases, while more accurate channel estimation may be possible. Conversely, as the number of hidden layers (1020) decreases, the processing time of the encoder (9111) decreases, while the accuracy of channel estimation may deteriorate.

Then, let's look at the configuration of the encoder (9111). The input layer (1010) of the encoder (9111) may be composed of multiple input nodes. For convenience of explanation, FIG. 10 illustrates a case where the input layer (1010) of the encoder (9111) is composed of six nodes. Each node constituting the input layer (1010) of the encoder (9111) in FIG. 10 may also be referred to as a neuron. In other words, one black dot in FIG. 10 may correspond to one node or one neuron.

Each node of the input layer (1010) can receive input data. Here, the data may be referred to as an input variable. The input variables provided to the input layer (1010) may be information for channel estimation. For example, the input variables may be the received CSI-RS described above. Each node constituting the input layer (1010) may be connected to nodes of the hidden layer (1020). If the hidden layer (1020) has two or more layers, each node constituting the input layer (1010) may be connected to each node of the first hidden layer. When each node constituting the input layer (1010) is connected to nodes of the hidden layer (1020), the connections between the nodes may have features based on learned weights.

Next, let's look at the hidden layer (1020). As explained above, in Fig. 10, for the convenience of explanation, the case where the hidden layer (1020) is composed of one layer is exemplified. However, the hidden layer (1020) may be composed of multiple layers. If the hidden layer (1020) is composed of multiple layers, each hidden layer may also be composed of multiple nodes. For example, if the hidden layer (1020) is composed of two layers, each node of the second hidden layer may be connected to each node of the first hidden layer. Additionally, each node of the second hidden layer may be connected to each node of the output layer (1030).

For convenience of explanation, the following description will be given using a case having one hidden layer (1020) as illustrated in Fig. 10.

As described above, each of the nodes constituting the hidden layer (1020) can be connected to the input nodes of the input layer (1010). In the example of Fig. 10, the hidden layer (1020) is illustrated as consisting of four nodes. The number of nodes in the hidden layer (1020) is not limited to the form illustrated in Fig. 10. In other words, the hidden layer (1020) may be composed of five or more nodes, or three or fewer nodes.

In addition, each of the nodes constituting the hidden layer (1020) may be connected to the nodes of the output layer (1030). Based on this configuration, each of the nodes constituting the hidden layer (1020) may provide a connection between the nodes of the input layer (1010) and the nodes of the output layer (1030). At this time, each of the nodes of the hidden layer (1020) may transmit information calculated based on weighted sum information learned offline and/or learned online when transmitting input variables from the nodes of the input layer (1010) to the nodes of the output layer (1030).

Next, we will look at the output layer (1030) of the encoder (9111).

The output layer (1030) of the encoder (9111) can output latent variables based on information received from each node of the hidden layer (1020). Here, the latent variable can mean a value output from nodes existing in the latent space, which is the output layer (1030) of the encoder (9111). In the example of Fig. 10, since the number of nodes included in the latent space is 2, if the latent space is expressed as M, the M value can be set to 2. Since the number of latent space nodes of the output layer (1030), or the M value, is less than the number of input variable nodes of the input layer (1010), the encoding operation of the encoder (9111) according to the present disclosure can be understood as a compression operation.

The operation of generating information for CSI feedback by the UE (910) using the configuration of the encoder (9111) described above is as follows.

The base station (910) can transmit a reference signal (RS) for channel estimation, such as a CSI-RS or a CSI-RS for online learning of an autoencoder, to the UE (920). Accordingly, the UE (920) can receive the CSI-RS transmitted by the base station (910). At this time, since the CSI-RS is transmitted through a wireless channel (1001), degradation or distortion may occur due to the wireless channel (1001). Accordingly, the UE (920) can compress the received CSI-RS using an encoder (9111).

The encoder (9111) can transmit the received CSI-RS provided to the input layer (1010) to the hidden layer (1020) and the output layer (1030) based on the offline learning information. At this time, the received CSI-RS is calculated based on the offline learned weighted sum information, and can be compressed and output through each node of the input layer (1010), each node of the hidden layer (1020), and each node of the output layer (1030). In other words, the encoder (9111) can compress the received CSI-RS based on the offline learning. The compressed CSI-RS can be latent variable values, which are output data of each node of the output layer (1030). The UE (920) can generate a channel feature indicator (CFI) using the latent variable values. The CFI according to the present disclosure can be expressed as in the following mathematical expression 1.

In mathematical expression 1, N can be determined as the product of the number of nodes (M) in the latent space and the latent variable dimension (LVD). The number of nodes (M) in the latent space can be determined as the number of output nodes of the output layer (1030) in the case of the encoder (9111). The LVD will be examined in more detail below.

The UE (920) can transmit the CFI to the base station (910) through the wireless channel (1230). Accordingly, the base station (910) can receive the CFI transmitted by the UE (920) through the wireless channel (1001) from the UE (920). The UE (920) can transmit the CFI to the base station (910) when online learning of the autoencoder is required and/or when estimating the channel using the autoencoder.

The base station (910) can transmit CSI-RS periodically. In addition, the base station (910) can transmit additional CSI-RS if necessary. Therefore, the UE (920) can receive CSI-RS periodically. In addition, the UE (920) can further receive additional CSI-RS. When channel estimation is required from the base station (910), the UE (920) can transmit CFI to the base station (910) instead of CSI feedback.

In addition, when the base station (910) requires online learning of the autoencoder (911), it may inform the UE (920) of the CFI reporting cycle, and the UE (920) may report the CFI to the base station (910) based on the CFI reporting cycle. At this time, the base station (910) may transmit the CSI-RS based on the CFI reporting cycle, or may transmit the CSI-RS at a preset cycle. The CFI reporting cycle will be described in more detail below.

The base station (910) can use the CFI received from the UE (920) as an input to the decoder (9112) to restore the received CSI-RS from the CFI. As described above, the received CSI-RS can be a CSI-RS that has been degraded or distorted through the wireless channel (1001) from the CSI-RS transmitted by the base station (910). Therefore, the CSI-RS restored by the base station (910) using the CFI value received from the UE (920) as an input to the decoder (9112) can theoretically be the same information as the received CSI-RS received by the UE (920).

The operation of the decoder (9112) may correspond to the reverse process of the encoding procedure of the encoder (9111). In other words, the decoder (9112) may include an input layer (1040), a hidden layer (1050), and an output layer (1060). The input layer (1040) of the decoder (9112) may be configured with the same number of nodes as the output layer (1030) of the encoder (9111) to process the CFI value. Therefore, the nodes of the input layer (1040) of the decoder (9112) may be an input latent space and may be configured with two nodes identical to the latent space of the output layer (1030) of the encoder (9111). And the hidden layer (1050) of the decoder (9112) can be composed of the same number of nodes as the hidden layer (1020) of the encoder (9111). According to the example of FIG. 10, the hidden layer (1050) of the decoder (9112) can be composed of four nodes. In the same manner, the output layer (1060) of the decoder (9112) can be composed of six nodes, which is the same as the input layer (1010) of the encoder (9111).

As illustrated in FIG. 10, the base station (910) can restore the received CSI-RS from the CFI reported by the UE (920). Then, the base station (910) can obtain a reference signal received power (RSRP) value from the restored received CSI-RS.

When conducting online learning, the base station (910) can receive CFI values from the UE (920) two or more times. Then, the base station (910) can obtain RSRP values from each of the received CFI values. The base station can compare the obtained RSRP values to determine whether the RSRP value is saturated. Here, saturation can mean a case where the RSRP value does not change by a certain level or more for a certain period of time or a certain number of times.

For example, let's look at the case where the RSRP value becomes saturated assuming that the number of times is 3.

The base station (910) can check whether the RSRP value has been acquired two or more times before the current time point (t) at which the RSRP value was acquired. Let the time point at which the current RSRP value was acquired be t, and let the previous times at which the RSRP values were acquired be t-1 and t-2, respectively. Then, if the RSRP(t-2) value acquired by the base station (910) at time point t-2, the RSRP(t-1) acquired at time point t-1, and the RSRP(t) acquired at time point t have the same value or are within a preset range, the RSRP value can be determined to be saturated. For a more specific example, if the preset range is 2 dBm, the RSRP(t-2) value is -48.4 dBm, the RSRP(t-1) value is -49 dBm, and the RSRP(t) is -48.7 dBm, the base station (910) can determine that the RSRP value is saturated. On the other hand, if the RSRP(t-2) value is -44.2 dBm, the RSRP(t-1) value is -47.4 dBm, and the RSRP(t) value is -48.7 dBm, the base station (910) can determine that the RSRP values are not saturated.

If the RSRP value is saturated, the base station (910) can complete online learning and check whether the saturated RSRP value satisfies the desired level of RSRP. Here, the desired level of RSRP can be determined based on the distance between the base station (910) and the UE (920), the channel environment, etc.

There may be various ways in which the base station (910) obtains a desired RSRP value during online learning of the autoencoder. For example, the base station (910) may receive location information of the UE from the UE (920) and determine a desired RSRP value based on the distance between the base station (910) and the UE (920). As another example, the UE (920) may feed back CSI based on the 3GPP standard during online learning of the autoencoder or immediately before online learning. Accordingly, the base station (910) may determine a desired RSRP value based on the CSI information fed back from the UE (920). In another way. The base station (910) may determine (or estimate) a desired RSRP value from the transmission power intensity of a signal for which the UE (920) reports a CFI value and the intensity information of a signal that received the received CFI value. In this case, the UE (920) may transmit the transmission power intensity for which the CFI value is reported as additional information.

In general, the learning saturation of AI/ML models can be determined by how similar the restored input data is to the actual input data based on the output data of the learning network. In other words, saturation can be determined by the restoration accuracy.

However, when performing channel estimation using the structure illustrated in Fig. 10 in an actual wireless communication environment, saturation cannot be determined based on restoration accuracy because the actual channel environment corresponding to the actual input data is unknown. Therefore, since the purpose of the present disclosure is to maximize communication quality in a wireless communication environment, whether or not the learning of the AI model or ML model is saturated can be determined using whether or not the RSRP value is saturated.

Therefore, if the RSRP value in the saturated state is within the desired RSRP range as described above in FIGS. 10 and 11, the base station completes the learning of the encoder (9111) and decoder (9112) and can use the parameters obtained through online training for channel estimation, which is the model inference step.

On the other hand, if the RSRP value in the saturated state does not satisfy the desired level of RSRP value (or range), the base station (910) may determine that the parameters learned through offline training are overfitted to previously collected data. If the base station (910) determines that the parameters learned through offline training are overfitted to previously collected data, the base station (910) may instruct the UE (920) to drop out specific node(s) of the encoder (9111) to resolve this.

If a dropout is instructed for a specific node(s) of an encoder (9111) from a base station (910), the UE (920) can drop out the corresponding node and perform online training again. In the present disclosure, dropout may mean a technique of partially omitting a neural network to solve an overfitting problem. In the present disclosure, dropout may be applied by setting a probability that a node will be dropped for a node to which dropout is to be applied. Then, a case where dropout of an encoder node is applied will be examined with reference to the attached drawing.

Figure 11 is a conceptual diagram illustrating a dropout encoder model when dropout is instructed at a specific node of the encoder.

The encoder (9111) illustrated in Fig. 11 may be an encoder stored in the UE (920). In other words, it may be an encoder received from a server or base station (910) based on what was described above in Fig. 9. However, Fig. 11 illustrates a case where a specific node is dropped out.

The input layer (1110) of the encoder (9111) illustrated in FIG. 11 has the same number of nodes as the input layer (1010) of the encoder (9111) illustrated in FIG. 10, the hidden layer (1120) of the encoder (9111) illustrated in FIG. 11 has the same number of nodes as the hidden layer (1020) of the encoder (9111) illustrated in FIG. 10, and the output layer (1130) of the encoder (9111) illustrated in FIG. 11 has the same number of nodes as the output layer (1030) of the encoder (9111) illustrated in FIG. 10.

In contrast to Fig. 10, the case in which the second node (1121) of the hidden layer (1120) is dropped out is exemplified as illustrated in Fig. 11. Accordingly, the connection between the input layer (1010) and the hidden layer (1020) of the encoder (9111) illustrated in Fig. 10 and the connection between the input layer (1110) and the hidden layer (1120) of the encoder (9111) illustrated in Fig. 11 may have different forms. In other words, each node of the input layer (1110) illustrated in Fig. 11 may have no connection relationship with the second node (1121) of the hidden layer (1120).

In comparison with Fig. 10, as illustrated in Fig. 11, due to the dropout of the second node (1121) of the hidden layer (1120), the connection between the hidden layer (1120) and the output layer (1130) may also have a different form than in Fig. 10. In other words, the second node (1121) of the hidden layer (1120) illustrated in Fig. 11 is not connected to any node of the output layer (1130). In other words, the second node (1121) indicated by a white dot in the hidden layer (1120) in Fig. 11 may not be connected to the input layer (1110) and the output layer (1130) according to the dropout instruction.

Except for the second node (1121) of the hidden layer (1120), the connections between the remaining nodes are the same as the form illustrated in FIG. 10 above. In the present disclosure, the dropout of a specific node can be caused by setting the weight to '0 (zero)', thereby causing the corresponding node to be dropped out. In other words, the connection from the input layer (1110) to the second node (1121) of the hidden layer (1120) does not provide any value since the weight is '0'. In addition, the second node (1121) of the hidden layer (1120) can apply the weight '0' to the nodes of the output layer (1130) so that it does not provide any value.

The present disclosure described below will define the procedures and parameters necessary for the operations described above to be specifically performed. In addition, based on these definitions, we will examine the procedures for performing the operations exemplified below. The processes presented in the present disclosure may include the following contents.

(1) Process of determining the transmission cycle of CFI, a CSI feedback parameter

(2) Process of determining the latent space dimension (LVD) of the latent space

(3) The base station transmits CSI-RS to the UE, and the UE generates CFI based on the received CSI-RS and reports it to the base station.

(4) The process in which the base station learns the decoder part of the base station using CFI and determines the point in time to stop learning through a saturation threshold.

(5) Process of evaluating learning performance through RSRP threshold after learning is stopped

(6) (In case of poor learning performance) A process of relearning specific node(s) of the UE's encoder through dropout, and

(7) (When learning performance is good) The process of terminating the learning of the autoencoder

The procedures (1) to (7) exemplified above will be examined in more detail below. The procedures exemplified above are procedures for learning an autoencoder, and can be operations for learning fine tuning for downlink channel estimation through online learning for an autoencoder that has undergone offline learning.

[A]. CFI Feedback Procedure for Online Training

In the following, we will discuss the CFI feedback procedure for online training. A UE that receives an offline-learned autoencoder from a base station or an offline-learned autoencoder from a server can perform CFI feedback. At this time, the base station may have all offline-learned autoencoders or decoders of offline-learned autoencoders. The UE may have all offline-learned autoencoders or encoders of offline-learned autoencoders. Since offline learning of autoencoders has already been described above, a redundant description will be omitted.

In the following description, channel estimation will be explained assuming downlink channel estimation. However, a similar or identical method can be used for uplink channel estimation. In the present disclosure, for downlink estimation, a base station transmits a CSI-RS to a UE, and the UE can generate a CFI that compresses the received CSI-RS. Then, the UE considers a situation in which online training is performed through a process of transferring the generated CFI to the base station through a CSI feedback process. In addition, in the following description, the learning of the autoencoder may mean the learning of the encoder (9111) as exemplified in FIG. 9, the learning of the decoder (9112), or the learning of both the encoder (9111) and the decoder (9112). In addition, even if learning and training are used interchangeably, they may mean a procedure in which learning of an autoencoder, an encoder, or a decoder is performed through online training.

Figure 12 is a flowchart of operations between a base station and a UE during CFI feedback for online training.

Before referring to FIG. 12, the UE (920) and the base station (910) will be examined based on what was described above in FIG. 9. Therefore, it should be noted that the reference numerals of the UE (920) and the base station (910) use the same reference numerals as described in FIG. 9. In addition, the encoder (9111) of the autoencoder may have the configuration of the encoder described in FIG. 10, and the decoder (9112) of the autoencoder may also have the configuration of the decoder described in FIG. 10. Therefore, the encoder (9111) and the decoder (9112) may be in an offline-learned state.

In step S1200, the UE (920) may perform negotiation with the base station (910) for online training of the autoencoder. The negotiation for online training of the autoencoder may be initiated by the base station (910) instructing the UE (920) to perform online training of the autoencoder or by the UE (920) requesting the base station to perform online training of the autoencoder. For online training of the autoencoder, the base station (910) and the UE (920) must have the same autoencoder. If the UE (920) does not have an autoencoder that requires online training, the base station (910) may provide the UE (920) with all of the autoencoders that require online training or the encoders of the autoencoder that the UE (920) should train.

The autoencoder may have two or more autoencoders with different network sizes as described below. Each of the autoencoders may be selected according to the required situation, and online training may be performed on at least one of the autoencoders.

In addition, in step S1200, the base station (910) may provide the UE (920) with the necessary information for autoencoder training. For example, the base station (910) may provide the UE (920) with the table information and/or the online training start time described below. In addition, the UE (920) may provide its location information to the base station (910) at the request of the base station (910) or voluntarily.

If necessary, the base station (910) can transmit a reference signal, for example, a CSI-RS, to the UE (920). The UE (920) can report an RSRP value, which is reception power information measured by the CSI-RS transmitted by the base station (910), to the base station (910). The procedure of the base station (910) transmitting the CSI-RS to the UE (920) and the UE (920) reporting the CSI to the base station (910) can be performed based on a CSI feedback procedure based on the current 3GPP standard.

In addition to the operations described above, the negotiation procedure for online training of the autoencoder at step S1200 may require other procedures, but since it is not possible to explain all of them, the description of other procedures or operations will be omitted.

In step S1210, the base station (910) can determine the CFI transmission period. In the present disclosure, the base station (910) can determine the CFI transmission period based on the network size and/or the channel coherence time of the autoencoder (911) as factors for determining the CFI transmission period. In addition to the network size and the channel coherence time of the autoencoder described in the present disclosure, additional factors may be considered when determining the CFI transmission period. In the present disclosure, only the above two contents will be examined.

a. Determine CFI transmission period based on the network size of the autoencoder.

According to one embodiment of the present disclosure, the base station (910) can determine the transmission period of the CFI according to the network size of the autoencoder (911). Here, the network of the autoencoder (911) can mean a network of the autoencoder (911) learned by the base station (910) or the server through offline training. The network size (NS) of the autoencoder (911) can be determined as the sum of the network width and the network height constituting the encoder (9111) of the autoencoder (911) or the decoder (9112) of the autoencoder.

Here, the network width can be defined as the number of nodes forming one layer in the encoder (9111) illustrated in Fig. 10. Since the decoder (9112) has a configuration that performs the reverse operation of the encoder (9111), the network width can also be defined as the number of nodes forming one layer in the decoder (9112).

And the network height may mean the number of layers that constitute the encoder (9111). In the example of Fig. 10, the encoder (9111) is illustrated as having a hidden layer (1020) composed of one layer. However, the hidden layer (1020) may be composed of two or more layers. Therefore, the network height may vary depending on the number of hidden layers (1020).

Based on the explanation above, the width of the network can mean the sum of the number of nodes forming the input layer, the number of nodes forming the hidden layer, and the number of nodes forming the output layer.

In the case of the encoder (9111) exemplified in Fig. 10, the input layer (1010), the hidden layer (1020), and the output layer (1030) are each composed of one layer. In addition, in the case of the encoder (9111) exemplified in Fig. 10, the number of nodes constituting the input layer (1010) is 6, the number of nodes constituting the hidden layer (1020) is 4, and the number of nodes constituting the output layer (1213) is 2. Therefore, the network size (NS) of the encoder (9111) exemplified in Fig. 10 can be interpreted as "6 + 4 + 2 = 12".

As described above, the number of offline-learned autoencoders can be plural. In addition, each of the offline-learned autoencoders can have a different network size. In addition, each of the offline-learned autoencoders can be included in both the base station (910) and the UE (920). In this way, the CFI transmission period according to the network size (NS) of each of the autoencoders having different sizes can be (pre)-configured to have a different period according to the network size.

Suppose there are three autoencoders with different sizes. Suppose that there is a first autoencoder with a first network size (NS #1), a second autoencoder with a second network size (NS #2), and a third autoencoder with a third network size (NS #3). And when each of the network sizes is set to NS #1 < NS #2 < NS #3, the CFI periods can be (pre)-configured to different values, as shown in Table 2 below.

Network size CFI Transmission Cycle NS #1 CFI_P #1 NS #2 CFI_P #2 NS #3 CFI_P #3

Each of the CFI transmission cycles (CFI_P #1, CFI_P #2, and CFI_P #3) illustrated in Table 2 can be a cycle with a different time interval. When the network size described above is present, the time of each of the CFI transmission cycles (CFI_P #1, CFI_P #2, and CFI_P #3) can be set as CFI_P #1 < CFI_P #2 < CFI_P #3.

The base station (910) may consider various factors when determining the network size of the autoencoder. For example, the base station (910) may determine the network size of the autoencoder according to the available power and/or channel environment of the UE (920). Here, determining the network size of the autoencoder may be understood as selecting the autoencoder. In other words, when the base station (910) selects the first autoencoder, the network size (NS #1) corresponding to the first autoencoder is selected, when the second autoencoder is selected, the network size (NS #2) corresponding to the second autoencoder is selected, and when the third autoencoder is selected, the network size (NS #3) corresponding to the third autoencoder is selected.

First, let us look at the case where the base station (910) determines the network size of the autoencoder based on the available power of the UE (920).

Based on the available power information reported in advance by the UE (920), the base station (910) can determine the network size of the autoencoder that will perform online learning with the UE (920). If the available power reported in advance by the UE (920) is low, in other words, if the remaining battery power of the UE (920) is below a preset threshold, the base station (910) can select an autoencoder with a small network size. Conversely, if the remaining battery power of the UE (920) is above a preset threshold, the base station (910) can select an autoencoder with a large network size. A large network size of the autoencoder means that the UE (920) must perform operations on many layers and many nodes when performing encoding. If the network size of the autoencoder is large, the UE (920) must perform operations on many layers and many nodes when performing encoding, which leads to a long encoding time. As a result, when the network size of the autoencoder is large, the power consumption of the UE (920) increases and the encoding time also becomes longer. In other words, the time for the UE (920) to obtain CFI becomes longer. Therefore, the base station (910) can determine the network size of the autoencoder based on the available power of the UE (920). In addition, when the network size of the autoencoder is determined, the CFI period can be determined as exemplified in Table 2. If the online learning of the autoencoder with a small network size is determined, the CFI period can be shortened, and if the online learning of the autoencoder with a large network size is determined, the CFI period can be lengthened.

Next, let us look at the case where the network size of the autoencoder is determined based on the channel environment between the base station (910) and the UE (920), as follows.

The base station (910) may have previously checked the change rate of the channel environment between the UE (920) and the base station (910) in the preceding step S1200. The base station (910) may determine an autoencoder to perform online learning based on the change rate of the channel environment. For example, if the change rate of the channel environment is fast, the base station (910) may select an autoencoder with a small network size. On the other hand, if the change rate of the channel environment is slow or there is little change in the channel environment, the base station (910) may select an autoencoder with a large network size. As described above, if the network size is large, the UE (920) has a long encoding time. Therefore, if the channel changes quickly, the base station (910) may select an autoencoder with a small network size to cope with the fast channel change. On the other hand, if the channel change is small or there is little change in the channel change, the base station (910) may select an autoencoder with a large network size.

In the above description, a method of selecting an autoencoder using the available power of the UE (920) and the rate of change in the channel environment between the UE (920) and the base station (910) has been described. However, an autoencoder may also be selected by simultaneously considering both the available power of the UE (920) and the rate of change in the channel environment between the UE (920) and the base station (910).

In this paper, we will examine the reason for mapping the network size of the autoencoder and the CFI transmission period.

The base station (910) can receive the CFI from the UE (920) during the online training process and input it into the decoder of the autoencoder for decoding. At this time, if the CFI reporting cycle is too fast, in other words, the UE (920) may not be able to compress the CSI-RS to report the CFI. In addition, even if the UE (920) reports the CFI to the base station (910), if the CFI is reported before the decoding procedure is completed in the decoder of the base station (910), the received CFI may not be properly processed.

Conversely, if the CFI reporting cycle is too slow, it may result in delays in online learning itself.

Meanwhile, in Table 2, only three network sizes and three CFI transmission periods are exemplified for convenience of explanation. However, the network sizes and CFI transmission periods according to the present disclosure are not limited to three, and each of the network sizes and the CFI periods corresponding to the network sizes may be four or more.

When using the mapping table of Table 2, the base station (910) can provide the table information of Table 2 to the UE (920) in advance. In other words, in step S1200, the base station (910) can provide the table information of Table 2 to the UE (920) in advance.

If the base station (910) provides the UE (920) with mapping information such as the table in Table 2 in advance, the base station (910) may transmit the information to the UE (920) through various signaling. For example, the mapping information such as the table in Table 2 transmitted to the UE (920) may be transmitted using a System Information Block (SIB), included in an RRC Reconfiguration message, or included in a MAC-CE message.

As another example, mapping information such as the table in Table 2 may be transmitted to the UE (920) via DCI.

As another example, mapping information such as the table in Table 2 may be transmitted to the UE (920) using the second message (Massage 2, Msg2) of the 4-step RACH procedure or message B (Massage B, MsgB) of the 2-step RACH procedure.

As another example, if an RRC signaling message is newly defined to transmit mapping information such as the table in Table 2, the newly defined RRC signaling message may be transmitted to the UE (920).

b. Determining CFI transmission period based on channel coherent time

Below, we will examine a method for determining the CFI period based on the channel coherence time.

The channel coherent time may mean a time during which there is no change in channel gain. The channel coherent time is generally due to the mobility of the UE, but may also be affected by the surrounding channel environment in addition to the mobility of the UE. The shorter the channel coherent time, the faster the UE (920) must report the CFI to the base station (910). In addition, if the channel coherent time is short, the effective time during which the UE (920) can report the CFI becomes shorter, and thus the number of times the CFI is reported also decreases. If the number of times the CFI is reported decreases, the CFI that can be used for online training decreases, and thus sufficient learning may not be performed. If the online learning of the autoencoder is insufficient, the parameters with low accuracy of the autoencoder are obtained, and accurate channel estimation cannot be performed.

Therefore, in the present disclosure, the CFI reporting cycle can be determined based on the channel coherence time (CT), and an example of the CFI reporting cycle based on the channel coherence time (CT) is as shown in Table 3 below.

Channel coherent time CFI Transmission Cycle CT #1 CFI_P #1 CT #2 CFI_P #2 CT #3 CFI_P #3

The time values of the first channel coherent time (CT #1), the second channel coherent time (CT #2), and the third channel coherent time (CT #3) illustrated in Table 3 can assume that CT #1 has the shortest time value, CT #2 has a longer time value than CT #1, and CT #3 has the longest time value. Of course, the opposite case is also possible. In the following, for the convenience of explanation, it will be explained by assuming the case where the time values are set as CT #1 < CT #2 < CT #3.

When the relationship between channel coherent time values is CT #1 < CT #2 < CT #3, the CFI transmission periods (CFI_P #1, CFI_P #2, CFI_P #3) can become longer as CFI_P #1 < CFI_P #2 < CFI_P #3. In other words, the CFI transmission period (CFI_P #1) is the shortest for the first coherent time (CT #1), and the CFI transmission period (CFI_P #3) is the longest for the third coherent time (CT #3). This means that as the channel coherent time becomes shorter, the CFI transmission period can be set shorter to obtain more inputs to the decoder of the base station (910).

Meanwhile, in Table 3, the channel coherent values are divided into three categories, but in actual implementation, only two types can be configured, or four or more types of channel coherent values can be configured. The types (number of types) of channel coherent values can be preset, or the base station (910) can flexibly determine them according to the channel environment.

When using the mapping table of Table 3, that is, when using the CFI transmission period by mapping it with the channel coherent value, the base station (910) can provide the table information of Table 3 to the UE (920) in advance. In other words, in step S1200, the base station (910) can provide the table information of Table 3 to the UE (920) in advance.

If the base station (910) provides the UE (920) with mapping information such as the table in Table 3 in advance, the base station (910) may transmit the information to the UE (920) through various signaling. For example, the mapping information such as the table in Table 3 transmitted to the UE (920) may be transmitted using SIB, included in an RRC reconfiguration message, or included in a MAC-CE message.

As another example, mapping information such as the table in Table 3 may be transmitted to the UE (920) via DCI.

As another example, mapping information such as the table in Table 3 may be transmitted to the UE (920) using the second message (Massage 2, Msg2) of the 4-step RACH procedure or message B (Massage B, MsgB) of the 2-step RACH procedure.

As another example, if an RRC signaling message is newly defined to transmit mapping information such as the table in Table 3, the newly defined RRC signaling message may be transmitted to the UE (920).

c. Determine CFI transmission period based on autoencoder size and channel coherent time.

In the above section a, which performs step S1210, the operation of determining the CFI transmission period based on the size of the autoencoder is described, and in section b, the operation of determining the CFI transmission period based on the channel coherent time is described.

These may be used individually as described above, or only one of the above may be used. In other words, when determining the network size, only the size of the autoencoder may be considered, or only the channel coherent time may be considered. In addition, the CFI transmission period may be determined by combining the method of using the network size (Table 2) and the method of using the channel coherent time (Table 3). For example, the base station (910) may determine the CFI transmission period by considering the network size described in Table 2 and the channel coherent time described in Table 3 together. In this case, a new table considering the contents of Tables 2 and 3 may be defined. Since the new table considering Tables 2 and 3 may take various forms, specific examples are not provided in this disclosure.

In another way, the base station (910) may determine the transmission period of the CFI by considering the channel coherent time, and then determine the network size of the autoencoder between the base station (910) and the UE (920) based on the determined CFI period as described in Table 2.

In addition, when determining the CFI cycle as described in Table 2 and Table 3 above, the base station (910) may request information of the UE (920) in advance at step S1200 and receive the corresponding information from the UE (920). The information that the base station (910) requests from the UE (920) may be information on available power of the UE (920) as described in Table 2 above. To obtain such information, the base station (910) may request the UE (920) through a UE Information Request message. When the UE (920) receives the UE Information Request message, the UE (920) may generate a UE Information Response message in response thereto and send it back to the base station (910) including the information requested by the base station (910).

As another example, if information such as power available to the UE (920) at a certain point in time, such as the maximum transmission power value or the remaining battery capacity at that point in time, is required, the base station (910) may request this from the UE (920) and receive the information from the UE (920). At this time, the UE (920) may use any one of uplink control information (UCI), the first message (Msg1) in the 4-step RACH procedure, or message A (MsgA) in the 2-step RACH procedure as the information requested by the base station (910).

As another example, if a new RRC signaling message is defined for the UE (920) to transmit information requested by the base station (910), the RRC signaling message can be used.

The base station (910) can determine the CFI transmission cycle in step S1210 through one or a combination of two or more of the methods described above.

In step S1220, the base station (910) can determine the latent variable dimension (LVD) of the autoencoder to perform online learning. The latent variable dimension (LVD) of the autoencoder can be determined based on the channel resolution or the latency requirement. In addition to the channel resolution and latency requirement described in this disclosure, additional factors may be considered when determining the LVD of the autoencoder. However, this disclosure will only examine cases where the above two requirements are considered when determining the LVD.

a. LVD determination based on channel resolution

In the present disclosure, the channel resolution may be the number of downlink channels that the base station (910) intends to distinguish. For example, when the base station (910) forms a downlink channel using multiple beams, the channel resolution may be determined based on the number of downlink beams. Accordingly, the channel resolution may have a higher channel resolution as the base station forms a narrower beam or intends to perform more precise channel estimation. LVD may mean a dimension required to express one latent variable value existing in a latent space. Additionally, LVD may be interpreted as the number of bits required to express one latent variable value. Therefore, the higher the LVD value, the more accurately the latent variable can be expressed. Accordingly, the mapping between the channel resolution (CR) and the latent variable dimension (LVD) may be exemplified as in Table 4 below.

Channel Resolution Latent Variable Dimension (LVD) CR #1 LVD #1 CR #2 LVD #2 CR #3 LVD #3

In Table 4, the channel resolution (CR) is assumed such that the first channel resolution (CR #1) has the smallest channel resolution, the second channel resolution (CR #2) is higher than the first channel resolution (CR #1) and lower than the third channel resolution (CR #3), and the third channel resolution (CR #3) has the highest channel resolution. Then, the channel resolutions can have a relationship such as CR #1 < CR #2 < CR #3.

As explained above, the higher the channel resolution, the more bits the latent variable dimension (LVD) can have. Therefore, depending on the relationship of the channel resolution, the LVD values in Table 4 can also have the relationship LVD #1 < LVD #2 < LVD #3.

Table 4 exemplifies a case where three channel resolutions are distinguished, but this is only one embodiment, and only two channel resolution values may be used, or four or more channel resolution values may be used. The number of channel resolutions according to the present disclosure may be preset, or the base station (910) may adaptively change it according to the channel environment.

In addition, the mapping information between the channel resolution and LVD as exemplified in Table 4 can be transmitted in advance by the base station (910) to the UE (920). For example, in step S1200 above, the base station (910) can transmit in advance the mapping information between the channel resolution and LVD as exemplified in Table 4 to the UE (920). At this time, the base station (910) can transmit the mapping information between the channel resolution and LVD to the UE (920) through various signaling.

For example, mapping information such as Table 4 transmitted to UE (920) may be transmitted using a System Information Block (SIB), included in an RRC Reconfiguration message, or included in a MAC-CE message.

As another example, mapping information such as Table 4 can be transmitted to the UE (920) via DCI.

As another example, mapping information such as Table 4 may be transmitted to the UE (920) using the second message (Massage 2, Msg2) of the 4-step RACH procedure or message B (Massage B, MsgB) of the 2-step RACH procedure.

As another example, if an RRC signaling message is newly defined to transmit mapping information such as Table 4, it may be transmitted to the UE (920) using the newly defined RRC signaling message.

b. LVD determination based on delay requirements

Below, we will examine the case where the latent variable dimension (LVD) is determined according to the latency requirement.

As explained above, LVD refers to the dimension required to express one latent variable value existing in the latent space, and can also be interpreted as the number of bits required to express the latent variable value. As the LVD increases, the amount (number of bits) of CFI that the UE (920) reports to the base station (910) through the CSI feedback increases. As the amount of CFI increases, the time required for online training of the autoencoder may increase. As the time required for online training increases in this way, there may be cases where the latency requirement required by the UE (920) cannot be satisfied. Here, the latency requirement may be a delay requirement due to the characteristics of the UE (920) itself, or a delay requirement according to the type of service provided to the UE (920). For example, if the UE (920) is a specific node in a time-sensitive network, a delay requirement may exist in the time-sensitive network. In this case, the UE (920) may have a delay requirement according to the characteristics of the time-sensitive network node.

As another example, when a URLLC (Ultra Reliable Low Latency Communication) service is provided to a UE (920), it may have a delay requirement required according to the characteristics of the service. In this way, there may be a delay requirement based on the characteristics of the network (or characteristics of the UE) in which the UE (920) operates and a delay requirement in the service.

In the present disclosure, LVD can be determined based on such delay requirements, and the mapping between delay requirements and LVD can be exemplified as shown in Table 5 below.

Delay requirements Latent Variable Dimension (LVD) LR #1 LVD #1 LR #2 LVD #2 LR #3 LVD #3

Looking at each of the delay requirements in Table 5, assuming that LR #1 has the shortest delay, LR #2 has a longer delay than LR #1, and LR #3 has the longest delay, each of the delay requirements can have the relationship LR #1 < LR #2 < LR #3. The LVD values corresponding to the delay requirements can also have the relationship LVD #1 < LVD #2 < LVD #3. In other words, the shorter the delay requirement, the smaller the latent variable dimension (LVD), and the longer the delay requirement, the larger the latent variable dimension (LVD).

Table 5 exemplifies a case where three delay requirements are distinguished, but this is only one embodiment, and it may be configured with only two delay requirements, or it may be configured with four or more delay requirements. The number of these delay requirements and channel resolutions may be preset, and the base station (910) may adaptively change them according to the channel environment.

In addition, the mapping information between the delay requirement and LVD as exemplified in Table 5 can be transmitted in advance from the base station (910) to the UE (920). For example, in step S1200 above, the base station (910) can transmit in advance the mapping information between the delay requirement and LVD as exemplified in Table 5 to the UE (920). At this time, the mapping information between the delay requirement and LVD can be transmitted from the base station (910) to the UE (920) through various signaling.

For example, mapping information such as Table 5 transmitted to UE (920) may be transmitted using SIB, included in an RRC Reconfiguration message, or included in a MAC-CE message.

As another example, mapping information such as Table 5 may be transmitted to the UE (920) via DCI.

As another example, mapping information such as Table 5 may be transmitted to the UE (920) using the second message (Massage 2, Msg2) of the 4-step RACH procedure or message B (Massage B, MsgB) of the 2-step RACH procedure.

As another example, if an RRC signaling message is newly defined to transmit mapping information such as Table 5, it may be transmitted to the UE (920) using the newly defined RRC signaling message.

c. LVD decision based on channel resolution and delay requirements

In the above section a, which performs step S1220, the operation of determining LVD based on channel resolution is described, and in section b, the operation of determining LVD based on delay requirements is described.

These can be used individually as described above, or only one of the above can be used. In other words, when determining the latent variable dimension (LVD), only the channel resolution can be considered, or only the delay requirement can be considered. Also, when determining the LVD, both the case of using the channel resolution (Table 4) and the case of using the delay requirement (Table 5) can be considered and decided.

In addition, when determining the latent variable dimension described in Table 4 and Table 5, information of the UE (920) may be required. Therefore, in this case, the base station (910) may request the UE (920) for information necessary for LVD determination in step S1200 described above, and may receive information necessary for latent variable dimension determination from the UE (920).

Various forms of signaling can be used as a method for UE (920) to provide UE information to base station (910). For example, UE (920) can use either uplink control information (UCI) or UE assistance information.

As another example, if the UE (920) can support learning of an autoencoder, it may report this in advance to the base station (910) using the first message (Msg1) of the 4-step RACH procedure or message A (MsgA) of the 2-step RACH procedure.

As another example, if a new RRC signaling message is defined to transmit information requested by the base station (910), the UE (920) may provide UE information using the newly defined RRC signaling message.

As another example, the base station (910) may request UE information from the UE (920) and receive the information from the UE (920). At this time, the base station (910) may transmit a UE Information Request message to the UE (920), and the UE (920) may include the information requested in the UE Information Request message in a UE Information response message and report it to the base station (910).

In step S1230, the base station (910) can transmit the CFI transmission period information determined in step S1210 and the latent variable dimension information determined in step S1220 to the UE (920).

An example of information transmitted from the base station (910) to the UE (920) in step S1230 can be illustrated in Table 6 below.

Transmission information Base station → UE CFI Transmission Cycle=CFI_P #2, LVD=LVD #3 CSI-RS

As illustrated in Table 6, the base station (910) can transmit a CFI transmission period to the UE (920). The CFI transmission period may be determined by considering only the network size of the autoencoder as described in Table 2 above, may be determined by considering only the channel coherent time as described in Table 3, or may be determined by considering both the autoencoder and the channel coherent.

Additionally, as illustrated in Table 6, the base station (910) may transmit latent variable dimension (LVD) information to the UE (920). The LVD may be determined by considering only the channel resolution as described in Table 4 above, or by considering only the delay requirements as described in Table 5, or by considering both the channel resolution and the delay requirements together.

In the embodiment of the present disclosure, it is assumed that the base station (910) considers the size of the autoencoder and the channel coherent time together when determining the CFI period. In addition, in the embodiment of the present disclosure, it is assumed that the base station (910) considers the channel resolution and delay requirements together when determining the LVD.

According to another embodiment of the present disclosure, the base station (910) may determine the CFI transmission period and LVD by additionally considering various factors such as frequency resource utilization, the number of UEs connected to the base station (910), RSRP of each UE, available power of each UE, etc. in addition to the above-mentioned factors in determining the CFI determination and the LVD determination.

For example, when the number of UEs attempting to communicate with the base station (910) is large or the frequency resource utilization rate is high, the base station (910) may already have a large number of frequency resources being used for communication. Therefore, the amount of frequency resources that can be used by the base station (910) may not be large. In this case, the base station (910) may set a CFI value with a long period and set a low LVD in order to train the autoencoder while providing a communication service without interference between all frequency bands or UEs.

As another example, a shorter CFI transmission period and a higher LVD can be determined to increase the reliability of the CFI signal as the RSRP of the UE that wants to communicate with the base station (910) is lower.

As another example, the base station (910) can determine the network size based on the available power of each UE reported by the UEs, and determine the CFI transmission period and LVD based on the determined network size.

In step S1230, the base station (910) can use various forms of signaling when transmitting CFI transmission cycle and LVD information to the UE (920). Such signaling can use signaling methods as described in Tables 2 to 5 above.

At step S1230, the UE (920) can receive CFI transmission cycle and LVD information from the base station (910).

In step S1240, the UE (920) can determine an encoder to learn based on the CFI transmission period and LVD information received from the base station (910). As described above in step S1200, the UE (920) can receive two or more autoencoders (or encoders of autoencoders) that have been learned offline from the base station (910) or the server. At this time, the UE (910) can receive autoencoders having two or more network sizes. Therefore, in step S1230, the autoencoder to perform online learning can be determined based on the CFI transmission period information and LVD information received from the base station (910).

To briefly explain this again, the CFI transmission period can be determined based on the network size and channel coherent time as described above. Therefore, the UE (920) can know the network size of the autoencoder from the CFI transmission period information. In addition, when the LVD information is determined, the number of nodes of the output layer (1030) of the encoder (9111) as illustrated in FIG. 10 can be determined. Based on the above information, the UE (920) can select an autoencoder to perform online learning.

In step S1250, the base station (910) can transmit a CSI-RS for online training (or learning) of the autoencoder to the UE (920). At this time, the CSI-RS may be a CSI-RS defined separately for online training of the autoencoder, or may be a CSI-RS transmitted periodically. Therefore, in step S1250, the UE (920) can receive a CSI-RS from the base station (910).

In step S1260, the UE (920) can use the received CSI-RS as input data of the encoder (9111) of the autoencoder (910) for online training. The CSI-RS received by the UE (920) can be the received CSI-RS described above in FIG. 10. In other words, since the received CSI-RS has passed through the wireless channel (1001), it can be in a state where attenuation and distortion occur compared to the CSI-RS transmitted by the base station (910). The UE (920) can generate a CFI by using the received CSI-RS as an input of the encoder (9111). The CFI can be determined as in the mathematical expression 1 described above, and if expressed again, it can be expressed as in the mathematical expression 2 below.

According to mathematical expression 2, CFI can be generated based on the number of nodes (M) of the latent space of the autoencoder and the LVD information indicated in step S1230. At this time, the number of nodes (M) of the latent space can be determined based on the decision of the autoencoder. In other words, when the network size is determined, the number of nodes (M) of the latent space of the autoencoder can be determined accordingly. In the example of Fig. 10, the number of nodes (M) of the latent space is 2, and the LVD is expressed as the number of bits representing each node constituting each latent space. Therefore, the CFI expressed as mathematical expression 2 can be expressed as the product of 2 and LVD in the case of Fig. 10.

In step S1270, the UE (920) can transmit the CFI to the base station (910) based on the CFI transmission cycle. At this time, the CFI transmission cycle may be the CFI transmission cycle that the base station (910) transmitted to the UE (920) in step S1230. Therefore, the base station (910) can receive the CFI from the UE (920) based on the CFI transmission cycle in step S1270. In step S1270, the base station (910) can receive the CFI from the UE (920).

At this time, the CFI transmitted by the UE (920) to the base station (910) can be reported in various forms. For example, since the CFI is a report on CSI-RS, it can be transmitted to the base station (910) through a measurement report message of the CSI reporting procedure, which is a CSI-RS feedback procedure according to the current 3GPP standard.

Alternatively, if a new RRC signaling message is defined for CFI reporting, the CFI can be transmitted to the base station (910) via the newly defined RRC signaling.

Let's compare the CFI reported by the UE (920) to the base station (290) at step S1270 with the method currently specified in the 3GPP standard.

According to the current 3GPP standard, a UE must generate CSI expressed in RI, CQI, and PMI values based on CSI-RS received from a base station, and perform Type 1 codebook-based CSI feedback or Type 2 codebook-based CSI feedback. On the other hand, the method according to the present disclosure can generate CFI using an encoder capable of expressing the received CSI-RS itself in a low dimension, and feed it back to the base station. Therefore, the CFI feedback procedure according to the present disclosure can be defined as a new feedback procedure for AI/ML-based CSI feedback.

According to the CFI transmission method that the UE (920) transmits to the base station (910) at step S1270, it can be one method that can map the latent variable to the CFI. The mapping relationship between the latent variable and the CFI can be determined in various ways.

[B] Training and retraining procedure of autoencoder through online training

The present disclosure described below will describe a procedure for learning an autoencoder through online training and a procedure for relearning the autoencoder in channel estimation using AI/ML. The autoencoder for online training may be learned in advance through offline training. In addition, the base station may have the entire autoencoder or a part of the autoencoder (for example, a decoder in the case of downlink channel estimation) in advance. The UE may also have the entire autoencoder or a part of the autoencoder (for example, an encoder in the case of downlink channel estimation) in advance. Since the case where the base station and/or the UE have the entire or a part of the autoencoder and the configuration thereof have been described above in FIG. 9, the same description will be omitted.

As above, when the base station and the UE have autoencoders, the base station and the UE can perform fine-tuning of the autoencoder through online training. The procedure for obtaining information for this fine-tuning procedure is described in Section [A].

In section [B] according to one embodiment of the present disclosure, a procedure for learning an autoencoder through online training for fine-tuning at a base station using the acquired information and a procedure for retraining the autoencoder will be described. Therefore, the operation described in section [B] may be an operation based on the information acquired in section [A]. However, the operation described in section [B] may also utilize information acquired through a CSI reporting procedure according to the current 3GPP standard. In the following description, a procedure for learning and retraining an autoencoder through online training will be described based on the description in section [A].

Figure 13 is a flowchart explaining the procedure for conducting online training of an autoencoder at a base station.

Before explaining Fig. 13, it is assumed that the base station and the UE are the base station (910) and the UE (920) described in Figs. 9 to 12 above. Therefore, it should be noted that the reference numerals of the UE (920) and the base station (910) use the same reference numerals as described in Fig. 9. In addition, when explaining Fig. 13, it will be explained assuming that the procedure of Fig. 12 has been performed above.

In step S1300, the base station (910) can receive a CFI from the UE (920) and input the received CFI into the decoder of the autoencoder that the base station (910) has. The received CFI can be a value that the UE (920) reported to the base station (910) in step S1270 in the procedure of FIG. 12 above. In step S1300, the base station (910) can obtain the received CSI-RS of the UE (920) by decoding the received CFI using the decoder. Therefore, the base station (910) can obtain the RSRP value from the received CSI-RS. The base station (910) can determine whether the obtained RSRP is saturated. Here, the RSRP saturation can mean a phenomenon in which the RSRP does not increase even though the base station continues online training using the decoder.

The base station (910) can use information in the table 7 below to determine RSRP saturation status.

Condition (Saturation count: COUNTSAT,
Maximum saturation count: MAXCOUNT_SAT) Whether training (or learning) is in progress COUNTSAT < MAXCOUNT_SAT Conduct training (or learning). COUNTSAT = MAXCOUNT_SAT Saturation is determined and training (or learning) is stopped.

In Table 7, the conditions for determining whether learning progresses can use the saturation count (COUNTSAT) value and the maximum saturation count (MAXCOUNT_SAT) value. If the saturation count value is less than the maximum count value, training or learning can continue, and if the saturation count value is equal to the maximum count value, it can be determined to be saturated and training or learning can be stopped. This will be examined in more detail with reference to the attached Figure 14.

Figure 14 is a flowchart explaining a case where saturation of RSRP values is determined at a base station.

In step S1400, the base station (910) can set the saturation count (COUNTSAT) value to '0'. Then, in step S1402, the base station (910) can restore the reception CSI-RS by using the received CFI information as an input to the decoder of the base station (910) as described above. Then, the base station (910) can obtain the RSRP value by using the restored reception CSI-RS.

In step S1404, the base station (910) can compare the acquired RSRP value with the previous RSRP value. If the received CFI is the first reception, the acquired RSRP value may be the first RSRP value. Therefore, if there is no previous RSRP (if the first CFI value is acquired) or it is not the same as the previous RSRP value, the base station (910) can proceed to S1400.

If the first RSRP value is not obtained after receiving the CFI, a previous RSRP value exists. Therefore, the base station (910) can compare the RSRP value obtained through decoding with the RSRP value obtained in the previous training, and check whether the two values are the same in step S1404.

If the two values are the same as a result of the verification in step S1404, the base station (910) can proceed to step S1406 and increase the saturation count value by 1. On the other hand, if the two values are not the same as a result of the verification in step S1406, the base station (910) can proceed to step S1400. At this time, in order to prevent infinite repetition in the case where the RSRP values are not the same and proceed to step S1400, a repetition count limit value can be set. Since the method of using a repetition count limit value to prevent infinite repetition is a widely known method, further description will be omitted.

The above explanation assumes that the two RSRP values are equal, but the two values may be considered equal if the RSRP values are within a specific range.

After increasing the saturation count value by 1 in step S1606, the base station (910) can proceed to step S1408. In step S1408, the base station (910) can compare whether the saturation count value is equal to the maximum saturation count value (MACCOUNT_SAT). If the saturation count value is equal to the maximum saturation count value, the base station (910) can determine that RSRP is saturated and terminate the routine of FIG. 14.

On the other hand, if the saturated count value is not equal to the maximum count value, the base station (910) may proceed to step S1402 and repeat the training according to the present disclosure.

Here, the maximum saturation count value can be actively set by the base station (910) based on various factors such as the channel environment between the base station (910) and the UE (920) or the mobility of the UE (920). For example, in an environment where the channel changes rapidly or the mobility of the UE (920) is large, the base station (910) can set the maximum count value to a small value in order to quickly progress the learning speed. On the other hand, in a case where there is little channel change and the mobility of the UE (920) is not large, the base station (910) can set the maximum count value to a large value in order to more accurately learn.

Referring again to FIG. 13, at step S1300, the base station (910) can determine whether RSRP is saturated based on the above-described steps. If RSRP is saturated, the base station can proceed to step S1510.

Meanwhile, in FIG. 13, a method for a base station (910) to determine whether RSRP is saturated based on CFI reported by a UE (920) is exemplified, but the UE (920) may also determine whether RSRP is saturated. This is because when the UE (920) generates CFI based on the received CSI-RS and the CFI generates the same value, it may determine that RSRP is saturated. As another example, the UE (920) may directly measure the RSRP of the received CSI-RS and determine that RSRP is saturated.

In this way, when the UE (920) determines whether RSRP saturation has occurred, the base station (910) can provide maximum saturation count (MAXCOUNT_SAT) value information to the UE (920). At this time, the maximum saturation count value information can be provided to the UE (920) through various forms of signaling.

For example, the base station (910) may provide the maximum saturation count value information to the UE (920) via SIB, DCI, MAC-CE, the second message (Msg2) of the 4-step RACH procedure, message B (MsgB) of the 2-step RACH procedure, or an RRC Reconfiguration message. As another example, if a new RRC signaling message is defined to deliver information about the maximum saturation count (MAXCOUNT_SAT) value to the UE (920), the base station (910) may also transmit information about the maximum count value to the UE (920) using the new RRC signaling message.

If the UE (920) determines whether RSRP is saturated, the UE (920) can notify the base station (910) if RSRP is saturated. At this time, various forms of signaling can be used as a method for the UE (920) to notify the base station (910) of the RSRP saturation state.

For example, the UE (920) may notify the base station (910) of an RSRP saturation state using either UCI or UE Assistance Information.

As another example, the UE (920) may notify the base station (910) of the RSRP saturation state using either the first message (Msg1) of the 4-step RACH procedure or message A (Msg A) of the 2-step RACH procedure.

As another example, if a new RRC signaling message is defined to indicate RSRP saturation, the UE (920) can indicate RSRP saturation to the base station (910) using the newly defined RRC signaling message.

As another example, when a base station (910) requests RSRP saturation status information from a UE (920) using a UE Information Request message, the UE (920) may inform the base station (910) of the RSRP saturation status information using a UE Information Response message.

In the above, the method by which the base station (910) or UE (920) determines the RSRP saturation state has been described. In the following description, for convenience of explanation, it is assumed that the base station (910) determines the RSRP saturation state.

If RSRP saturation occurs by performing step S1300, the base station (910) can proceed to step S1310.

In step S1310, the base station (910) can check whether the RSRP value is at a satisfactory level. A method for checking whether the level is satisfactory can be done by comparing the RSRP threshold (RSRP _TH ) with a saturated RSRP value. At this time, the RSRP _th parameter can mean a certain level of RSRP value set by the base station (910) or the UE (920), and can be set through various criteria such as the accuracy of channel estimation desired by the base station (910) and/or the UE (920) or a hardware problem of the UE (920). At this time, in order to check the accuracy of channel estimation, etc., the UE (920) can report the RSRP or CSI directly measured from the CSI-RS received by the base station (910) in step S1200 described above in FIG. 12. Therefore, the base station (910) can determine the RSRP threshold based on the RSRP or CSI reported in step S1200.

For example, if the base station (910) and/or the UE (920) desires a high accuracy channel estimation, the base station (910) and/or the UE (920) may set a high RSRP threshold and perform re-learning for the desired performance. As another example, if frequent learning is not possible due to a hardware problem of the UE (920), the remaining battery level of the UE (920), or channel congestion, the base station (910) and/or the UE (920) may set a low RSRP threshold to reduce the number of re-learnings. The learning results based on the RSRP threshold in this way can be exemplified as shown in Table 8 below.

Condition (RSRP Threshold: RSRP _th ) Learning Outcomes RSRP ≥ RSRP _th End of learning. RSRP < RSRP _th Need to relearn.

According to Table 8 exemplified above, the base station (910) can terminate learning if the RSRP value is greater than or equal to the RSRP threshold (RSRP _th ). On the other hand, if the RSRP value is lower than the RSRP threshold (RSRP _th ), the base station (910) can determine that re-learning is necessary.

In the example of Fig. 13, the case where the base station (910) determines whether RSRP is satisfied and decides whether to relearn is exemplified. However, the UE (920) may also determine whether RSRP is satisfied. In other words, the entity that determines whether to relearn the autoencoder may be the base station (910) or the UE (920).

When the base station (910) is the entity that determines whether to relearn the autoencoder, the base station (910) can determine whether to relearn by determining whether RSRP is satisfied. At this time, the base station (910) can determine relearning based on the comparison result of the RSRP value reported by the UE (920) and the RSRP threshold value.

On the other hand, if the UE (920) is the entity that determines whether to retrain the autoencoder, the UE (920) can compare the RSRP measured by itself with the RSRP threshold value to determine retraining. At this time, if the RSRP threshold value is determined by the base station (910), the RSRP threshold value can be provided to the UE (920) in advance by the base station (910).

If the UE (920) is the entity that determines whether to retrain the autoencoder and the RSRP threshold is determined by the base station (910), the method for the base station (910) to provide RSRP threshold information to the UE (920) can utilize various signaling methods.

For example, RSRP threshold information transmitted to UE (920) may be transmitted using SIB, included in an RRC Reconfiguration message, or included in a MAC-CE message.

As another example, RSRP threshold information can be transmitted to the UE (920) via DCI.

As another example, the RSRP threshold information may be transmitted to the UE (920) using the second message (Massage 2, Msg2) of the 4-step RACH procedure or message B (Massage B, MsgB) of the 2-step RACH procedure.

As another example, if an RRC signaling message is newly defined to transmit RSRP threshold information, it may be transmitted to the UE (920) using the newly defined RRC signaling message.

In this way, if the UE (920) is the entity that determines whether autoencoder re-learning is necessary, if it determines that autoencoder re-learning is necessary, it can notify the base station (910) that autoencoder re-learning is necessary. The UE (920) can use various forms of signaling messages to notify the base station (910) of information for notifying that autoencoder re-learning is necessary.

For example, the UE (920) may transmit information to the base station (910) to indicate the need for autoencoder retraining using either uplink control information (UCI), the first message (Msg1) in the 4-step RACH procedure, or message A (MsgA) in the 2-step RACH procedure.

As another example, the UE (920) may transmit information to the base station (910) to notify that autoencoder retraining is necessary using UE Assistance Information.

As another example, if a new RRC signaling message is defined for the UE (920) to transmit information to the base station (910) to inform that autoencoder retraining is needed, the RRC signaling message can be used.

If the base station (910) determines the entity that determines whether re-learning of the autoencoder is necessary in advance as the UE (920) at step S1200 of FIG. 12 or a separate step, the base station (910) can determine this to the UE (920). In this way, notifying that the entity that determines whether re-learning of the autoencoder is necessary is the UE (920) can be understood as the same as requesting whether re-learning of the autoencoder is necessary to the UE (920). Accordingly, in this case, the base station (910) can request re-learning determination information from the UE (920) using a UE Information Request message. Then, the UE (920) can determine whether re-learning of the autoencoder is necessary as described above, and report information on whether or not to re-learn the autoencoder to the base station (910) using a UE Information Response message.

If the desired RSRP is not satisfied at step S1310, that is, if the saturated RSRP is less than the RSRP threshold, the base station (910) may proceed to step S1320. If the desired RSRP is satisfied, that is, if the saturated RSRP is greater than or equal to the RSRP threshold, the base station (910) may proceed to step S1320.

Step S1320 is a state in which re-learning of the autoencoder is required because the saturated RSRP does not satisfy the desired RSRP, i.e., the threshold RSRP value. The base station (910) can generate re-learning related information.

In step S1330, the base station (910) can transmit the generated autoencoder re-learning related information to the UE (920).

An example of information related to autoencoder retraining generated in step S1320 and transmitted to UE (920) in step S1330 may be information as shown in Table 9 below.

Direction of information transmission Transmission information Base station → UE Encoder re-training Encoder Reconfiguration: Dropout Node Information Encoder identifier (ID) = Encoder_ID_n

As exemplified in Table 9, the transmitted information may include encoder re-training information indicating encoder retraining, encoder reconfiguration information, and an encoder identifier.

The encoder identifier (ID) exemplified in Table 9 may be information for informing the UE (920) of which encoder the current base station (910) is using to perform online learning. The UE (920) can determine which encoder the base station (910) indicates through the encoder identifier received from the base station (910).

In addition, the encoder re-training parameters illustrated in Table 9 may mean information that the decoder learned through the CFI transmitted by the current UE (920) does not satisfy a certain level of RSRP and thus begins re-training. Accordingly, the encoder re-training parameters may indicate that the training procedure of the encoder and decoder having the autoencoder identifier is performed again.

And the encoder reset parameters exemplified in Table 9 may include dropout node information as described above in FIG. 11. The dropout node information may include node information to which dropout is to be applied among the nodes of each layer constituting the encoder (9111) stored in the UE (920) and dropout probability information of the corresponding node. In general, among the nodes constituting the encoder (9111), a layer to which dropout is to be applied may be a node of the hidden layer (1120).

If the hidden layer (1120) is composed of one layer or two or more layers, all nodes of a specific hidden layer (1120) must not be dropped out. This is because if all nodes of a specific hidden layer (1120) are dropped out, the encoding operation is not performed. To this end, the present disclosure may define the minimum value of the nodes constituting a specific layer and the maximum value of the nodes constituting a specific layer.

To explain this with an example, let's assume that the hidden layer (1120) is composed of two layers, a first hidden layer connected to the input layer (1110) and a second hidden layer connected to the output layer (1130), and let's assume that the minimum and maximum values of the nodes constituting the first and second hidden layers are the same.

Then, the number of nodes that the first hidden layer can be configured by dropout can have the following relationship, and the second hidden layer can also have the same relationship.

<Relationship between the number of hidden layers by the maximum value, minimum value, and drop probability of the hidden layer>

The minimum value of the nodes that make up the first hidden layer ≤ the number of nodes in the first hidden layer to be made by the drop probability ≤ the maximum value of the nodes that make up the first hidden layer

Dropout node information can be determined by the above relationship.

Meanwhile, the last "n" in the parameter of the encoder identifier can be the number to which dropout is applied. If dropout is not applied, it consists of only the encoder identifier, and if it is an encoder to which dropout is applied once, the number of dropouts can be specified as "n", such as "Encoder_ID_1". Therefore, the number of "n" can increase depending on the number of dropouts.

The method of transmitting the encoder reset information may not be set in step S1320, but may be set in advance. For example, when the base station (910) provides the encoder (9111) to the UE (920) after performing offline training, the base station (910) may transmit in advance information about the order of nodes to be dropped for nodes existing in all layers required for additional encoder re-learning. As another example, when the base station (910) provides the encoder (9111) to the UE (920) after performing offline training, the base station (910) may transmit to the UE (920) information about the order of nodes to be dropped for nodes in each layer. The transmission of such information may be transmitted to the UE (920) in advance in step S1200 described above in FIG. 12.

In this way, if the base station (910) transmits drop-out node information to the UE (920) in advance, and if re-learning is required, the base station (910) can transmit only the encoder re-training and encoder identifier to the UE (920). Then, the UE (920) can confirm that the encoder needs to be reset from the encoder re-training parameters, and can drop out the corresponding nodes based on the drop-out node information promised in advance.

When the first drop out is indicated, the base station (910) can notify the UE (920) that the first drop out is indicated by transmitting the encoder identifier set to “Encoder_ID_1” as described above.

When the base station (920) transmits autoencoder retraining related information, such as Table 9, to the UE (910), various forms of signaling can be used. For example, the autoencoder retraining related information, such as Table 9, transmitted to the UE (920) can be transmitted using SIB, included in an RRC Reconfiguration message, or included in a MAC-CE message.

As another example, information related to autoencoder retraining, such as Table 9, can be transmitted to the UE (920) via DCI.

As another example, information related to autoencoder retraining, such as Table 9, may be transmitted to the UE (920) using the second message (Massage 2, Msg2) of the 4-step RACH procedure or message B (Massage B, MsgB) of the 2-step RACH procedure.

As another example, if an RRC signaling message is newly defined to transmit autoencoder retraining related information as in Table 9, the newly defined RRC signaling message may be transmitted to the UE (920).

At step S1330, the UE (920) can receive autoencoder retraining related information based on the method described above.

In step S1340, the UE (920) may reset the autoencoder based on the information related to autoencoder retraining. In the present disclosure, what the UE (920) resets may be the encoder of the autoencoder. In other words, it may be an operation of setting the dropout of a specific layer of the encoder described in FIG. 11.

In step S1350, the UE (920) can transmit Encoder Reconfiguration Complete information to the base station (910) through the encoder dropout. The Encoder Reconfiguration Complete information can be information indicating that the dropout for the encoder of a specific autoencoder is complete. Therefore, it can include the identifier information of the autoencoder. In other words, the UE (920) can report to the base station (910) which autoencoder has been reset and the number of times the dropout has been performed. Through this, the base station (910) can confirm whether it is the autoencoder that it has instructed to the UE (920) and whether the number of times the dropout has been instructed is the same.

Since the present disclosure describes an example for downlink, the autoencoder reported by the UE (920) can be specifically an encoder. In addition, if at least one node among the nodes constituting a specific layer of the encoder is dropped out, the input and output to the node can have their weights set to "0". In other words, both the input and output can be set to "0".

When the UE (920) transmits an autoencoder reset completion message to the base station (910) at step S1350, the autoencoder reset completion message can be transmitted to the base station (910) through various signaling.

For example, the UE (920) may transmit an autoencoder reset complete message to the base station (910) using either UCI or UE assistance information.

As another example, the UE (920) may report the autoencoder reconfiguration completion message to the base station (910) in advance using the first message (Msg1) of the 4-step RACH procedure or message A (MsgA) of the 2-step RACH procedure.

As another example, if a new RRC signaling message is defined to transmit information requested by the base station (910), the UE (920) may provide UE information using the newly defined RRC signaling message.

As another example, the base station (910) may request transmission of an autoencoder reconfiguration complete message to the UE (920) by transmitting a UE Information Request message. In this case, the UE (920) may report the autoencoder reconfiguration complete message to the base station (910) by including it in a UE Information response message.

At step S1350, the base station (910) can receive an autoencoder reset completion message.

In step S1360, the base station (910) may perform an autoencoder retraining procedure based on receiving an autoencoder re-reset completion message. The autoencoder retraining procedure may be a procedure that starts again from the CSI-RS transmission step described above in FIG. 12.

[C] Procedure for Forced Termination of Online Training

Meanwhile, when performing the procedures of FIGS. 12 and 13, the time for learning the autoencoder through online training between the base station (910) and the UE (920) may continue to increase. For example, when the desired RSRP is not satisfied, the learning time of the autoencoder may continue to increase. When performing online learning between the base station (910) and the UE (920), traffic occurs between the base station (910) and the UE (920), which may increase the congestion of the wireless channel and the frequency usage. In addition, since the UE (920) is generally powered by a battery, the battery consumption of the UE (920) also increases. Therefore, even when the desired RSRP is not satisfied, it may be desirable to stop the learning or relearning procedure at an appropriate time.

In this disclosure, we will look into setting online training expiration information and the procedure for terminating online training when online training expires.

Figure 15 is a flowchart illustrating the procedure for terminating online training of an autoencoder at a base station.

Before explaining Fig. 15, it is assumed that the base station and the UE are the base station (910) and the UE (920) described in Figs. 9 to 13. Therefore, it should be noted that the reference numerals of the UE (920) and the base station (910) use the same reference numerals as described in Fig. 9.

At step S1500, the base station (910) can start online training of the autoencoder. Starting online training of the autoencoder can be a procedure in which initial online training is performed based on the procedure described above in FIGS. 12 and 13.

In step S1502, the base station (910) can obtain an RSRP value in the online training procedure of the autoencoder and store the obtained RSRP and the corresponding autoencoder information. Here, the RSRP value can be obtained based on the CFI value reported by the UE (920) as described above.

In step S1504, the base station (910) can check whether the acquired RSRP satisfies the desired RSRP value. In other words, step S1504 can be a procedure for checking whether it is greater than or equal to the RSRP threshold value described in FIG. 12 above. If the acquired RSRP value satisfies the desired RSRP value, the base station (910) can terminate the online training of the autoencoder.

On the other hand, if the RSRP value acquired in step S1504 does not satisfy the desired RSRP value, the base station (910) can proceed to step S1506 to check whether the online training forced termination condition is satisfied. Here, the online training forced termination condition can be defined with various values.

Now, let's look at examples of conditions that force termination of online training.

For example, you can define the maximum time allocated to online training as "OnlineTrainingTimeMax" and check whether the condition for forcibly terminating online training is met by checking whether the time set to that time has elapsed.

As another example, it is also possible to set a maximum number of relearning times for forced termination of online training and use a method to check whether the condition for forced termination of online training is met by checking whether the maximum number of relearning times has elapsed. In the following, for convenience of explanation, a method for checking whether the condition for forced termination of online training is met by using the maximum time allocated to online training will be described.

The OnlineTrainingTimeMax value may be set by the base station (910), or may be set by the UE (920) based on its battery capacity or hardware requirements. If the UE (920) sets the OnlineTrainingTimeMax value, the UE (920) may transmit the OnlineTrainingTimeMax value to the base station (910) when the online learning of the autoencoder is initiated, or before the online learning of the autoencoder is initiated, so that the base station (910) may check the time expiration.

As another example of how the UE (920) determines the OnlineTrainingTimeMax value, the base station (910) may transmit to the UE (920) a set of possible OnlineTrainingTimeMax values according to the situation of the UE (920), and the UE (920) may select one value from the set and transmit it to the base station (910).

The OnlineTrainingTimeMax value can be set based on various criteria, such as the channel environment between the base station (910) and the UE (920) or the mobility of the UE (920). For example, in an environment where the channel changes rapidly or when the mobility of the UE (920) is large, a small OnlineTrainingTimeMax value can be set to quickly respond to changes in the channel by accelerating the learning speed for more accurate channel estimation.

The UE (920) may use various forms of signaling when transmitting the OnlineTrainingTimeMax value set by the UE (920) to the base station (910). For example, the UE (920) may transmit the OnlineTrainingTimeMax value set by using UCI and UE assistance information to the base station (910). As another example, the UE (920) may transmit the OnlineTrainingTimeMax value set by using the first message (Msg1) of the 4-step RACH procedure or the message A (MsgA) of the 2-step RACH procedure to the base station (910). As another example, the UE may use a newly defined RRC signaling message when a new RRC signaling message is defined to transmit the set OnlineTrainingTimeMax value to the base station (910).

If the base station (910) requests the UE (920) to report information on the maximum time value that can be allocated to online training as the OnlineTrainingTimeMax value through a UE information request message, the UE (910) may also reply to the base station (910) through a UE information response message with information on the maximum time value that it can allocate to online training.

If the online training forced termination condition is set in at least one of the methods described above, the base station (910) can check whether the online training forced termination condition is met in step S1506. In other words, the base station (910) can check whether the online training forced condition is met by checking whether the timer set to the OnlineTrainingTimeMax value has expired.

If the condition for forced termination of online training is met as a result of the inspection at step S1506, the base station (910) can proceed to step S1508, and if the condition for forced termination of online training is not met, the base station (910) can proceed to step S1512.

When proceeding to step S1508, the base station (910) can determine the training model with the highest RSRP value among the autoencoder models learned through online training so far as a model for downlink channel estimation.

In step S1510, the base station (910) can transmit the determined channel estimation model to the UE (920). At this time, the channel estimation model information can be encoder identifier information. In other words, identifier information for indicating a specific autoencoder among multiple autoencoders and encoder identifier information including the number of dropouts indicating which dropout was performed as described above can be transmitted to the UE (920).

An example of information transmitted by the base station (910) to the UE (920) in step S1510 can be exemplified as shown in Table 10 below.

Transmission information Base station → UE Training expiration information Encoder identifier = Encoder_ID_3

The training expiration information exemplified in Table 10 above may be information that the base station (910) notifies (or instructs) the UE (920) that the online training of the autoencoder has ended. In addition, as described above, the encoder identifier may indicate a specific autoencoder (or an encoder of the autoencoder) among multiple autoencoders, and the last number "3" may indicate a case where three dropouts have been performed. Through this, the optimal encoder can be selected among the encoders that have performed online training up to now.

In addition, the training expiration information and the encoder information as in Table 10 may be transmitted from the base station (910) to the UE (920) using various forms of signaling as described above. For example, the training expiration information and the encoder information as in Table 10 may be transmitted from the base station (910) to the UE (920) using any one of the SIB, MAC-CE, or RRC Reconfiguration message. As another example, the training expiration information and the encoder information as in Table 10 may be transmitted from the base station (910) to the UE (920) via DCI. As yet another example, the training expiration information and the encoder information as in Table 10 may be transmitted from the base station (910) to the UE (920) using the second message (Msg2) of the 4-step RACH procedure or the message B (MsgB) of the 2-step RACH procedure. As another example, if a new RRC signaling is defined for transmitting training expiration information and encoder information, such as Table 10, the training expiration information and encoder information may be transmitted from the base station (910) to the UE (920) using the newly defined RRC signaling message.

Meanwhile, if the condition for forced termination of online training is not satisfied as a result of the verification at step S1506, the base station (910) may proceed to step S1512 and perform an online re-learning procedure. The online re-learning procedure may include steps S1250 to S1270 described above in FIG. 12, and may be a procedure of steps S1300 to S1350 described in FIG. 13.

The base station (910) can store the RSRP value obtained in the corresponding round and the corresponding model information and, so to speak, the number of dropouts information whenever online retraining is completed. Using this, the base station (910) can identify the RSRP with the highest value among the RSRPs obtained through online training of the autoencoder in step S1508, and can obtain the identifier of the autoencoder with the highest RSRP value and the number of dropouts information in step S1510.

[D] Normal termination procedure for online training

Next, let's look at the procedure for normal termination of online learning. This can be the case when the saturated RSRP value of the base station (910) satisfies the desired RSRP value in step S1310 of Fig. 13.

In Fig. 13, it is simply explained in the form of termination of online learning, but the base station (910) can notify the UE (920) of termination of online learning. At this time, if information for the base station (910) to notify the UE (920) of termination of online learning is exemplified, it can be exemplified as in Table 11 below.

Transmission information Base station → UE Encoder training complete

As illustrated in Table 11, the base station (910) can transmit encoder training completion information to the UE (920). The encoder training completion information can mean that the encoder at the current trained time can be used. Accordingly, when the UE (920) receives the encoder training completion information, it can update the encoder to the encoder in the online training state performed immediately before receiving the encoder training completion information. In addition, the base station (910) can also store the encoder and decoder of the autoencoder to maintain the same state based on the completion of the encoder training.

The encoder training completion information can be transmitted from the base station (910) to the UE (920) via various forms of signaling. For example, the encoder training completion information as in Table 11 can be transmitted from the base station (910) to the UE (920) using any one of the SIB, MAC-CE, or RRC Reconfiguration message. As another example, the encoder training completion information as in Table 11 can be transmitted from the base station (910) to the UE (920) via DCI. As yet another example, the encoder training completion information as in Table 11 can be transmitted from the base station (910) to the UE (920) using the second message (Msg2) of the 4-step RACH procedure or the message B (MsgB) of the 2-step RACH procedure. As another example, if a new RRC signaling is defined for transmitting encoder training completion information such as Table 11, the encoder training completion information may be transmitted from the base station (910) to the UE (920) using the newly defined RRC signaling message.

Meanwhile, the UE (920) and the base station (910) can perform model inference using the acquired autoencoder model when estimating a downlink channel transmitted from the base station (910) using the parameters acquired during the current online training process.

The model inference procedure can be a procedure similar to the online training process, which is a step for performing actual channel estimation. In other words, the base station (910) can transmit the CSI-RS, and the UE (920) can perform a procedure for obtaining CFI through an encoder for the received CSI-RS and then reporting the obtained CFI to the base station (910). Then, the base station (910) can obtain the received CSI-RS through a decoder and obtain an RSRP value, thereby determining it as downlink CSI.

The contents of each section described above can be applied through simple combination, partial combination, and extended combination.

Figure 16 is a conceptual diagram explaining a case where the entire operation of the present disclosure is combined and operated.

Step S1610 is an operation performed at the base station (910), and step S1611 may be an operation corresponding to step S1210 described above in FIG. 12. In step S1611, the base station (910) may determine a CFI transmission period based on the network size and channel coherent time.

In addition, step S1612 may be an operation corresponding to step S1220 described above in FIG. 12. In step S1612, the base station (910) may determine LVD based on channel resolution and delay requirements.

Step S1620 may be an operation corresponding to steps S1230 and S1250 described above in FIG. 12. In step S1620, the base station (910) may transmit a CFI transmission period, LVD to the UE (920), and transmit a CSI-RS to the UE (920).

Step S1630 may be an operation corresponding to steps S1240, S1260, and S1270 described above in FIG. 12. In step S1630, the UE (920) may determine a CFI dimension, compress the received CSI-RS through an encoder, and report the CFI to the base station (910).

Step S1640 may be an operation corresponding to steps S1300 and S1310 described above in FIG. 13. In step S1640, the base station (910) may stop learning when RSRP is saturated and compare the saturated RSRP value with the RSRP threshold value.

If the saturated RSRP value in step S1640 is greater than or equal to the RSRP threshold value, step S1650 may be performed, and if the saturated RSRP value in step S1640 is less than the RSRP value, step S1660 may be performed.

Step S1650 can perform the operation described in section [D]. In other words, in step S1650, the base station (910) can transmit learning completion information to the UE (920) after learning is completed.

Step S1660 may correspond to steps S1320 and S1330 described in FIG. 13 described above. Therefore, in step S1660, the base station (910) may transmit relearning information, encoder reset information, and encoder identifier information to the UE (920).

In addition, step S1670 may correspond to steps S1340 and S1350 in FIG. 13 described above. Therefore, in step S1670, the UE (920) may reset the encoder based on the relearning information, the encoder reset information, and the encoder identifier information, and then report the encoder reset completion information to the base station (910).

In addition, if learning is completed within the allocated time in step S1670, the UE (920) can report encoder reset completion information to the base station (910) after resetting the encoder. If learning is not completed within the allocated time in step S1670, step S1680 can be performed.

Step S1680 may be a case where the learning described in the [C] section and FIG. 15 above is not completed within the allocated time. Therefore, in step S1680, the base station (910) may transmit learning completion information and the most optimal encoder identifier to the UE (920).

The above-described Figure 16 exemplifies a case where all the procedures described in the present disclosure are configured as a single procedure. However, some of the procedures in Figure 16 may be omitted or modified. Some of these modified examples have been described in the drawings described above. Since the present disclosure cannot cover all forms of modifications, it should be noted that various forms of modifications and combinations may be possible based on the contents described in the present disclosure.

The operation of the method according to the present disclosure can be implemented as a computer-readable program or code on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices that store information that can be read by a computer system. In addition, the computer-readable recording medium can be distributed over network-connected computer systems so that the computer-readable program or code can be stored and executed in a distributed manner.

Additionally, the computer-readable recording medium may include hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. The program instructions may include not only machine language codes generated by a compiler, but also high-level language codes that can be executed by the computer using an interpreter, etc.

While some aspects of the present disclosure have been described in the context of an apparatus, that may also represent a description of a corresponding method, wherein a block or device corresponds to a method step or a feature of a method step. Similarly, aspects described in the context of a method may also be described as a feature of a corresponding block or item or a corresponding device. Some or all of the method steps may be performed by (or using) a hardware device, such as, for example, a microprocessor, a programmable computer or an electronic circuit. In some embodiments, at least one or more of the most significant method steps may be performed by such a device.

A programmable logic device (e.g., a field-programmable gate array) may be used to perform some or all of the functions of the methods described in this disclosure. A field-programmable gate array may operate in conjunction with a microprocessor to perform one of the methods described in this disclosure. In general, the methods are preferably performed by some hardware device.

Although the present disclosure has been described with reference to preferred embodiments thereof, it will be understood by those skilled in the art that various modifications and changes may be made to the present disclosure without departing from the spirit and scope of the present disclosure as set forth in the claims below.

Claims (20)

In a method of user equipment (UE),
A step of receiving a channel feature indicator (CFI) transmission period and a latent variable dimension (LVD) of an encoder to perform online learning from a base station;
A step of determining a first encoder to be learned online based on the CFI transmission period and the LVD;
A step of receiving a first reference signal (RS) from the base station;
A step of generating a CFI by compressing the received first RS through the first encoder; and
A step of transmitting the first CFI to the base station based on the CFI transmission cycle,
The above first CFI is determined by the product of the number of nodes in the latent space of the first encoder and the LVD.
UE's method. In claim 1,
The above LVD is the number of bits required to express each latent variable value included in the latent space of the first encoder.
UE's method. In claim 1,
The above CFI transmission period is determined based on at least one of the network size or channel coherent time of the first encoder.
UE's method. In claim 1,
The above LVD is determined based on at least one of the channel resolution or latency requirement of the base station.
UE's method. In claim 1,
A step of receiving re-learning related information of the first encoder from the base station;
A step of resetting the first encoder based on the relearning related information;
A step of receiving a second RS from the base station;
A step of generating a second CFI by compressing the received second RS with the reset first encoder; and
Further comprising the step of transmitting the second CFI to the base station,
UE's method. In claim 5,
The above relearning related information includes at least one of relearning instruction information of the encoder, drop-out node information of the encoder, or identifier information of the first encoder,
The identifier information of the first encoder further includes the number of dropouts.
UE's method. In claim 1,
A step of receiving learning termination instruction information of the first encoder from the base station;
A step of updating the first encoder based on information learned up to immediately before receiving the learning termination instruction information;
A step of receiving a third RS from the above base station;
A step of generating a third CFI by compressing the third RS using the updated first encoder; and
Further comprising the step of transmitting the third CFI to the base station,
UE's method. In claim 1,
A step of receiving learning expiration information including optimal encoder information from the base station;
A step of updating the first encoder based on the optimal encoder information from the base station;
A step of setting the above updated first encoder as a channel estimation encoder;
A step of receiving a fourth RS from the above base station;
A step of generating a fourth CFI by compressing the fourth RS using the updated first encoder; and
Further comprising the step of transmitting the above fourth CFI to the base station,
The above optimal encoder information includes an encoder identifier and information on the number of drop-outs of the encoder.
UE's method. In the method of the base station,
A step of determining the transmission period of a Channel Feature Indicator (CFI) and a latent variable dimension (LVD) of an encoder to perform online learning and a user equipment (UE);
A step of transmitting the CFI transmission cycle and the LVD to the UE;
A step of receiving a first CFI from the UE based on the CFI transmission cycle; and
A step of restoring the received RS using a decoder for the received first CFI,
Method of base station. In claim 9,
A step of obtaining a first reference signal received power (RSRP) value of the restored receiving RS;
A step for checking whether the first RSRP is saturated;
A step of checking whether the first RSRP is greater than or equal to a preset threshold value when the first RSRP is saturated;
A step of transmitting learning completion information of the encoder to the UE when the first RSRP is greater than or equal to the preset threshold value,
The above first CFI is determined by the product of the number of nodes in the latent space of the encoder performing the online learning and the LVD.
Method of base station. In claim 10,
If the RSRP values obtained from the CFI previously received from the UE and the first RSRP value are equal to or within a predetermined range, the first RSRP is determined to be saturated.
Method of base station. In claim 11,
a step of generating re-learning related information of the encoder when the first RSRP is not saturated; and
A step of transmitting re-learning related information of the above encoder to the UE;
Further comprising the step of performing the above UE and encoder re-learning procedure,
Method of base station. In claim 12,
The above relearning related information includes at least one of relearning instruction information of the encoder, drop-out node information of the encoder, or identifier information of the encoder.
The identifier information of the above encoder further includes the number of dropouts.
Method of base station. In claim 13,
A step of checking the number of dropouts corresponding to the RSRP having the highest RSRP value among the learned RSRP values if relearning is not completed within a predetermined time; and
Further comprising a step of transmitting encoder identifier information including the number of dropouts and learning expiration information to the UE.
Method of base station. In claim 9,
The above CFI transmission period is determined based on at least one of the network size or channel coherence time of the encoder performing the online learning.
Method of base station. In claim 9,
The above LVD is determined based on at least one of the channel resolution or latency requirement of the base station.
Method of base station. In claim 9,
A step of transmitting a second RS to the UE after the above learning is completed;
A step of receiving a second CFI corresponding to the second RS from the UE;
A step of obtaining a second RSRP value of a receiving RS from the second CFI; and
Further comprising a step of estimating a downlink channel to the UE using the second RSRP.
Method of base station. In user equipment (UE),
A transceiver configured to transmit and receive signals from a base station; and
Comprising at least one processor, said at least one processor comprising:
Receive a channel feature indicator (CFI) transmission period and latent variable dimension (LVD) of an encoder to perform online learning from the base station;
Determine a first encoder to be learned online based on the above CFI transmission period and the above LVD;
Receive a first reference signal (RS) from the base station;
By compressing the received first RS through the first encoder, a CFI is generated; and
Causing the first CFI to be transmitted to the base station based on the above CFI transmission cycle,
The above first CFI is determined by the product of the number of nodes in the latent space of the determined encoder and the LVD.
UE. In claim 18,
The above CFI transmission period is determined based on at least one of the network size or channel coherent time of the first encoder, and
The above LVD is determined based on at least one of the channel resolution or latency requirement of the base station.
UE. In claim 18,
At least one processor of the above:
Receive information related to relearning the first encoder from the base station;
Resetting the first encoder based on the above relearning related information;
Receive a second RS from the above base station;
Generating a second CFI by compressing the received second RS with the reset first encoder; and
further causes the second CFI to be transmitted to the base station,
The above relearning related information includes at least one of relearning instruction information of the encoder, drop-out node information of the encoder, or identifier information of the encoder, and
The identifier information of the first encoder further includes the number of dropouts.
UE.

Images