Wireless communication device, method of operating same, and method of operating same

Abstract

A wireless communication apparatus, a method of operating the same, and a method of operating the same are provided. The method of operating a wireless communication device includes calculating a first channel matrix for a downlink channel based on CSI-RS, determining a level of a multi-level vector quantization process based on an uplink channel, extracting a potential vector based on the first channel matrix using a first machine learning model, in a first level of the multi-level vector quantization process, determining a level of the potential vector based on the first channel matrix, in a second level of the multi-level vector quantization process, generating a first residual potential vector based on the potential vector, where the first residual potential vector is based on a first codeword selected from a first codebook corresponding to a first stage, generating a second residual potential vector based on the first residual potential vector in a second stage corresponding to the determined stage level of the multi-stage vector quantization process, where the second residual potential vector is based on a second codeword selected from a second codebook corresponding to the determined stage level, wherein the second residual potential vector is based on a second codeword selected from a second codebook corresponding to the second stage, and the bitstream is generated based on the first codeword and the second codeword.

Description

[0001] This application is based on and claims the benefit of priority from Korean Patent Application No. 10-2023-0151949 filed in the Korean Intellectual Property Office on November 6, 2023, and Korean Patent Application No. 10-2024-0048854 filed in the Korean Intellectual Property Office on April 11, 2024, the disclosures of which are incorporated herein by reference in their entirety. Technical Field

[0002] The present disclosure relates to a wireless communication device, an operating method thereof, and a method of operating the system. Background Art

[0003] In the case of a wireless communication system, a base station may send a reference signal to a user equipment to obtain channel information between the base station and the user equipment. For example, a base station may send a channel state information reference signal (CSI-RS) for obtaining channel information between the base station and the user equipment. In some cases, the user equipment may use the CSI-RS received from the base station to identify the channel between the base station and the user equipment. Therefore, the user equipment may estimate the channel between the base station and the user equipment based on the CSI-RS. Therefore, the user equipment may report feedback information related to the estimated channel to the base station.

[0004] However, as wireless communication systems evolve towards high data rates and increased capacity, the amount of information transmission or exchange between base stations and user equipment has increased significantly, resulting in a heavy burden on the limited resources available in such systems. Therefore, there is a need in the art for a system and method that can effectively compress multiple pieces of information related to a channel estimated by a UE. Summary of the invention

[0005] The present disclosure provides a method and apparatus for compressing channel state information. Embodiments of the present disclosure may be configured to receive a channel state information reference signal (CSI-RS) from a base station of a wireless communication system at a user equipment. In some cases, the user equipment may include an encoder, wherein the encoder may be configured to perform compression of the CSI-RS based on multiple compression ratios. According to an embodiment, an encoder and a decoder may be implemented for each of the multiple compression ratios. In some cases, each of the encoder and the decoder may include a machine learning model.

[0006] In addition, embodiments of the present disclosure include a codebook for performing a multi-level vector quantization process for each of the compression ratios. According to an embodiment, each of the encoder and the decoder may support multiple compressions based on implementing a multi-level vector quantization structure. For example, the user equipment sends compressed feedback to the base station using the multi-level vector quantization process.

[0007] According to one aspect of the present disclosure, a method for operating a wireless communication device is provided, the method comprising: receiving a channel state information reference signal (CSI-RS) from a base station, calculating a first channel matrix for a downlink channel between the wireless communication device and the base station based on the CSI-RS, determining a level of a multi-stage vector quantization process based on an uplink channel between the wireless communication device and the base station, extracting a potential vector based on the first channel matrix using a first machine learning model, generating a first residual potential vector based on the potential vector in a first stage of the multi-stage vector quantization process, wherein the first residual potential vector is based on a first codeword selected from a first codebook corresponding to the first stage, generating a second residual potential vector based on the first residual potential vector in a second stage of the multi-stage vector quantization process corresponding to the determined level, wherein the second residual potential vector is based on a second codeword selected from a second codebook corresponding to the second stage, generating a bit stream based on the first codeword and the second codeword, and sending the bit stream to the base station using the uplink channel.

[0008] According to another aspect of the present disclosure, a method for operating a system is provided, the method comprising: receiving a channel state information reference signal (CSI-RS) from a base station by a wireless communication device, calculating a first channel matrix for a downlink channel between the wireless communication device and the base station based on the CSI-RS by the wireless communication device, determining a level of a multi-stage vector quantization process based on an uplink channel between the wireless communication device and the base station by the wireless communication device, generating a potential vector based on the first channel matrix using a first machine learning model of the wireless communication device, and generating a first residual potential vector based on the potential vector in a first stage of the multi-stage vector quantization process, wherein A first residual latent vector is based on a first codeword selected from a first codebook corresponding to a first stage, a second residual latent vector is generated based on a second codeword selected from a second codebook corresponding to a second stage in the multi-stage vector quantization process corresponding to the determined stage level, the wireless communication device generates a bit stream based on the first codeword and the second codeword, the wireless communication device sends the bit stream to the base station using the uplink channel, the base station receives the bit stream, the base station generates the first codeword and the second codeword based on the bit stream, and the base station estimates a second channel matrix for the downlink channel from the first codeword and the second codeword based on a second machine learning model.

[0009] According to another aspect of the present disclosure, a wireless communication device including a radio frequency integrated circuit (RFIC) and at least one processor is provided. The at least one processor is configured to receive a channel state information reference signal (CSI-RS) from a base station via the RFIC, calculate a first channel matrix for a downlink channel between the wireless communication device and the base station based on the CSI-RS, determine a level of a multi-stage vector quantization process based on an uplink channel between the wireless communication device and the base station, extract a potential vector based on the first channel matrix using a first machine learning model, generate a first residual potential vector based on the potential vector in a first stage of the multi-stage vector quantization process, wherein the first residual potential vector is based on a first codeword selected from a first codebook corresponding to the first stage, generate a second residual potential vector based on the first residual potential vector in a second stage of the multi-stage vector quantization process corresponding to the determined level, wherein the second residual potential vector is based on a second codeword selected from a second codebook corresponding to the second stage, generate a bit stream based on the first codeword and the second codeword, and send the bit stream to the base station via the RFIC.

[0010] According to another aspect of the present disclosure, a channel feedback method for a wireless communication device is provided, comprising: receiving a channel state information-reference signal (CSI-RS) from a base station; generating channel state information based on the CSI-RS; generating a potential vector based on the channel state information; determining the number of levels of a multi-stage vector quantization process based on a feedback channel of the base station, wherein the number of levels corresponds to a compression ratio of the channel state information; performing the multi-stage vector quantization process on the potential vector using the determined number of levels to obtain a quantized potential vector; and sending the quantized potential vector to the base station using the feedback channel. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Embodiments will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0012] Figure 1 illustrates operations of transmitting and receiving signals between a base station and a user equipment (UE) according to an embodiment;

[0013] Figure 2 A wireless communication device according to an embodiment is shown;

[0014] Figure 3 A wireless communication device according to an embodiment is shown;

[0015] Figure 4 A wireless communication device according to an embodiment is shown;

[0016] Figure 5A illustrates signal transmission and reception between an encoder and a decoder according to an embodiment;

[0017] Figure 5B shows a multi-stage vector quantization process of an encoder according to an embodiment;

[0018] Fig. 6A and Figure 6B Each shows sequential training of a wireless communication device according to an embodiment;

[0019] Figure 7 illustrates sequential training of a wireless communication device according to an embodiment;

[0020] Figure 8 A method of operating a wireless communication device according to an embodiment is shown;

[0021] Fig. 9 A method of operating a UE according to an embodiment is shown;

[0022] Fig.10 A method of operating a wireless communication system including a UE and a base station according to an embodiment is shown;

[0023] Fig.11 A method of operating a wireless communication device according to an embodiment is shown;

[0024] Fig.12 A block diagram showing a UE according to an embodiment; and

[0025] Fig.13 is a block diagram illustrating an electronic device according to an embodiment. DETAILED DESCRIPTION

[0026] In the case of modern wireless communication systems, the transmission of a channel state information reference signal (CSI-RS) is important for optimizing the performance and efficiency of a communication link (e.g., a communication link between a user equipment and a base station). In some examples, the CSI-RS provides detailed information about the state of the communication channel, enabling the user equipment to adjust parameters for maximizing data throughput, minimizing errors, and enhancing overall system performance. In some examples, the user equipment may report feedback information related to the estimated channel to the base station. For example, the feedback information may include a precoding matrix indicator (PMI), a rank indicator (RI), and a channel quality indicator (CQI).

[0027] Existing systems for compressing CSI-RS can be implemented based on reducing the amount of data required to represent the CSI-RS. In some cases, these systems can perform quantization, dimensionality reduction, and take advantage of the inherent sparsity of the channel matrix, which can result in significant degradation of the quality of the CSI-RS. For example, the user equipment of the existing system can compress the feedback by selecting an encoder based on the payload of the CSI feedback.

[0028] Furthermore, in some examples, each of the multiple encoders may include a different compression ratio. As a result, such a system may result in a trade-off between compression efficiency and the accuracy of the reconstructed channel information. In some cases, such an approach may result in the loss of key channel details, which may degrade the performance of the communication system.

[0029] In contrast, the present disclosure provides a method and apparatus for compressing channel state information. Embodiments of the present disclosure may be configured to receive a channel state information reference signal (CSI-RS) from a base station of a wireless communication system at a user equipment. In some cases, the user equipment may include an encoder, wherein the encoder may be configured to perform compression of the CSI-RS based on multiple compression ratios. According to an embodiment, an encoder and a decoder may be implemented for each of the multiple compression ratios. In some cases, each of the encoder and the decoder may include a machine learning model.

[0030] In addition, embodiments of the present disclosure include a codebook for performing a multi-level vector quantization process for each of the compression ratios. According to an embodiment, each of the encoder and the decoder may support multi-compression based on implementing a multi-level vector quantization process. For example, the user equipment sends compressed feedback to the base station based on implementing a multi-level vector quantization process.

[0031] Embodiments of the present disclosure are configured to perform compression of CSI-RS. In some cases, CSI-RS may be received by a user equipment from a base station. According to an embodiment, the user equipment may estimate a downlink channel based on CSI-RS, and a machine learning model may be used to effectively compress CSI-RS. Subsequently, the user equipment sends feedback to the base station using a multi-stage vector quantization process.

[0032] In some cases, the user equipment may select a level of a multi-level vector quantization process based at least on the bandwidth and capacity of the channel. According to an embodiment, the user equipment may generate a residual latent vector at each of the multiple levels based on selecting a codeword from a codebook. In some cases, the residual latent vector of the current level may be calculated based on the residual latent vector of the previous level, and the residual latent vector of the current level may be migrated to the next level for calculating the subsequent residual latent vector. For example, the residual latent vector may be the difference between the residual latent vector of the previous level and the codeword selected based on the residual latent vector of the previous level. In some instances, each of the selected codewords may be combined to generate a bit stream for transmission to another wireless communication device.

[0033] According to an embodiment, a user equipment may receive a CSI-RS from a base station. In some cases, the user equipment may include an encoder configured to generate a continuous potential vector. For example, the encoder includes a machine learning model, wherein the machine learning model enables the potential vector extracted from the CSI-RS to be transmitted to a quantizer. In some cases, the quantizer performs multi-level vector quantization for each level and generates a bit stream including quantized channel information for transmission to the base station. In some cases, the inverse quantizer of the base station may receive a bit stream and may be configured to generate a codeword based on index information of the codeword in the bit stream. Subsequently, the decoder of the base station may use the codeword to provide channel information, or use the machine learning model to provide a quantized potential vector.

[0034] The present disclosure describes a system and method for operating a wireless communication device. An embodiment of the present disclosure may be configured to receive a channel state information-reference signal (CSI-RS) from a base station, and calculate a first channel matrix for a downlink channel between the wireless communication device and the base station based on the CSI-RS. In some cases, the level of a multi-stage vector quantization process is determined based on an uplink channel between the wireless communication device and the base station. A first machine learning model is used to extract a potential vector based on the first channel matrix. Subsequently, a second residual potential vector is generated based on the first residual potential vector in a second level corresponding to the determined level of the multi-stage vector quantization process. In some cases, the second residual potential vector is based on a second codeword selected from a second codebook corresponding to the second level. A bit stream generated based on the first codeword and the second codeword may be sent to a base station.

[0035] Therefore, by generating latent vectors based on the machine learning model of the encoder, the embodiments of the present disclosure can achieve efficient use of available bandwidth in the wireless communication system. In addition, by implementing an encoder, decoder and codebook that can be used for each compression ratio in the compression ratio, the embodiments can prevent the selection of different encoders for different compression ratios, thereby improving the use of bandwidth and capacity of the channel. The machine learning model can be trained using a sequential training method, which leads to transfer learning and applies useful prior knowledge (i.e., information generated based on the previous stage) to the training process.

[0036] As described herein, a base station is an entity that communicates with a wireless communication device and allocates network resources to the wireless communication device, and may include at least one of a cell, a NodeB (NB), an eNodeB, a next generation radio access network (NGRAN), a wireless connection unit, a base station controller, a node on a network, and a gnodeB (gNB).

[0037] A wireless communication device is an entity that communicates with a base station or another wireless communication device and may include at least one of a node, a user equipment (UE), a next generation UE (NGUE), a mobile station (MS), a mobile equipment (ME), a device, and a terminal.

[0038] In addition, the wireless communication device may include at least one of a smart phone, a tablet personal computer (PC), a mobile phone, a video phone, an electronic book (ebook) reader, a desktop PC, a laptop PC, a netbook computer, a personal digital assistant (PDA), a portable multimedia player (PMP), an MP3 player, a medical device, a camera, and a wearable device. In addition, the wireless communication device may include a television, a digital video disk (DVD) player, an audio system, a refrigerator, an air conditioner, a vacuum cleaner, an oven, a microwave oven, a washing machine, an air purifier, a set-top box, a home automation control panel, a security control panel, a media box (e.g., Samsung HomeSync TM , Apple TV TM or Google TV TM ), game consoles (e.g., Xbox TM or PlayStation TM), electronic dictionary, electronic key, portable camera and electronic photo frame. In addition, the wireless communication device may include at least one of various medical devices (e.g., various portable medical measuring devices (glucose meter, heart rate monitor, blood pressure meter, medical thermometer, etc.), magnetic resonance angiography (MRA) machine, magnetic resonance imaging (MRI) machine, computed tomography (CT) scanner, ultrasound instrument, etc.), navigation system, global navigation satellite system (GNSS), event data recorder (EDR), flight data recorder (FDR), car infotainment device, ship electronic equipment (e.g., ship navigation system, gyrocompass, etc.), avionics, security device, car head unit, industrial or household robot, drone, automatic teller machine (ATM) of financial institution, point of sale (POS) system and Internet of Things (IoT) device (e.g., light bulb, various sensors, sprinkler, fire alarm, temperature controller, street lamp, toaster, exercise equipment, hot water tank, heater, boiler, etc.). In addition, the wireless communication device may include at least one of various multimedia systems capable of performing communication functions.

[0039] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the accompanying drawings.

[0040] Figure 1 The operations of transmitting and receiving signals between a base station and a wireless communication device according to the embodiment are shown.

[0041] Reference Figure 1 , the wireless communication system 10 may include a base station 110 and a UE 120. Although the wireless communication system 10 is illustrated as including one base station 110 and one UE 120 for ease of description, this is merely an example, and the wireless communication system 10 is not limited thereto and may be implemented to include various numbers of base stations and various numbers of UEs.

[0042] The base station 110 may be connected to the UE 120 via a wireless channel, and may provide various communication services to the UE 120. The base station 110 may provide services for user traffic via a shared channel, and may perform scheduling by collecting status information of the UE 120 (such as a buffer status, an available transmission power status, and a channel status). The wireless communication system 10 may support beamforming technology by using orthogonal frequency division multiplexing (OFDM) as a wireless access technology. In addition, the wireless communication system 10 may support an adaptive modulation and coding (AMC) method, in which a modulation scheme and a channel coding rate are determined according to a channel status of the UE 120.

[0043] Beamforming is a signal processing technique used in antenna arrays to direct signals in a specific direction, thereby increasing signal strength and reducing interference. Beamforming works by controlling the phase and amplitude of signals transmitted from or received by an antenna array. During transmission, the signals from each antenna are adjusted so that they constructively interfere in a specific direction, forming a focused beam of energy. The focused beam can be steered by changing the phase shift, directing the signal toward a specific target or receiver. During reception, beamforming enhances signals from a specific direction while reducing noise and interference from other directions. By aligning the phase of the incoming signal from the desired direction, the system amplifies the signal while minimizing the rest. Beamforming techniques are used in applications such as wireless communications (e.g., Wi-Fi, 5G), radar, and sonar, where they improve signal strength, clarity, and range by effectively focusing energy in the desired direction.

[0044] OFDM is a digital communication technique that divides a signal into multiple closely spaced subcarriers, where each subcarrier transmits a portion of the data at the same time. The subcarriers are orthogonal, which indicates that despite the close frequency, the subcarriers do not interfere with each other, which provides efficient use of bandwidth and high data rates. OFDM is robust against signal fading and interference, making OFDM ideal for wireless communication systems such as Wi-Fi, 4G, and 5G. In addition, OFDM simplifies equalization by converting a wideband frequency selective channel into multiple flat narrowband channels.

[0045] AMC is used in wireless communication systems to optimize data transmission rate and reliability by dynamically adjusting the modulation scheme and error correction code based on current channel conditions. In some cases, AMC uses high-order modulation (e.g., 64-quadrature amplitude modulation (64-QAM)) and less redundant coding to maximize the data rate. In some cases, AMC switches to low-order modulation (e.g., quadrature phase shift keying (QPSK)) and strong coding, which reduces the data rate and improves error resistance. AMC works by continuously monitoring signal quality (e.g., signal-to-noise ratio (SNR)). Based on feedback, the system selects the most appropriate combination of modulation and coding that balances throughput and reliability. Adaptability ensures efficient use of available bandwidth, improves overall network performance, and enhances user experience by maintaining a stable connection even under fluctuating channel conditions. For example, AMC is used in modern wireless standards such as LTE, Wi-Fi, and 5G.

[0046] Refer again Figure 1For example, the wireless communication system 10 may use a wide frequency band in a 6 GHz or higher frequency band to send and receive signals. For example, the wireless communication system 10 may use a millimeter wave band (such as a 28 GHz band or a 60 GHz band) to increase the data transmission rate. As described herein, because the millimeter wave band has a relatively large signal attenuation amplitude per distance, the wireless communication system 10 may support transmission and reception based on directional beams generated using multiple antennas to ensure coverage. The wireless communication system 10 may be a system supporting multiple input and multiple output (MIMO). Therefore, the base station 110 and the UE 120 may support beamforming technology. Beamforming technology may be divided into digital beamforming, analog beamforming, hybrid beamforming, and the like.

[0047] Reference Figure 1 , the base station 110 may send a channel state information-reference signal (CSI-RS) to the UE 120. The UE 120 may use the CSI-RS to estimate the downlink channel. The UE 120 may efficiently compress the channel state information between the UE 120 and the base station 110 based on the machine learning model, and may send the compressed channel state information to the base station 110. For example, the UE 120 may send CSI-RS feedback based on multi-level vector quantization to the base station 110.

[0048] According to an embodiment, the artificial intelligence model is a machine learning model or an artificial neural network (ANN). The ANN can be a hardware component or a software component that includes connected nodes (i.e., artificial neurons) that roughly correspond to neurons in the human brain. Each connection or edge sends a signal from one node to another node (such as a physical synapse in the brain). When a node receives a signal, the node processes the signal and then sends the processed signal to other connected nodes.

[0049] ANN has many parameters, including weights and biases associated with each neuron in the network, which control the degree of connection between neurons and affect the ability of the neural network to capture complex patterns in the data. These parameters, also called model parameters or model weights, are variables that determine the behavior and characteristics of a machine learning model.

[0050] In some cases, the signals between nodes include real numbers, and the output of each node is calculated by a function of the node's inputs. For example, nodes may use other mathematical algorithms to determine their outputs (such as selecting the maximum value among the inputs as the output), or use any other suitable algorithm for activating nodes to determine their outputs. Each node and edge is associated with one or more node weights that determine how to process and transmit signals. In some cases, nodes have thresholds below which signals are not transmitted at all. In some examples, nodes are aggregated into layers.

[0051] The parameters of a machine learning model can be organized into layers. Different layers perform different transformations on their inputs. The initial layer is called the input layer, and the last layer is called the output layer. In some cases, the signal traverses certain layers multiple times. The hidden (or intermediate) layer includes hidden nodes and is located between the input layer and the output layer. The hidden layer performs a nonlinear transformation of the input entering the network. Each hidden layer is trained to generate a defined output, wherein the defined output contributes to the joint output of the output layer of the ANN. The hidden representation is a machine-readable data representation of the input, wherein the machine-readable data representation is learned from the hidden layer of the ANN and generated by the output layer. As the ANN's understanding of the input improves as the ANN is trained, the hidden representation gradually distinguishes from earlier iterations.

[0052] In some cases, a machine learning model can be trained to optimize performance. For example, the parameters of a machine learning model can be learned or estimated from training data and then used to make predictions or perform tasks based on learned patterns and relationships in the data. In some examples, the parameters are adjusted during the training process to minimize a loss function or maximize a performance metric. The goal of the training process can be to find the best values ​​for the parameters that allow the machine learning model to make accurate predictions or perform well for a given task.

[0053] Thus, node weights can be adjusted to improve the accuracy of the output (i.e., by minimizing a loss that corresponds in some way to the difference between the current outcome and the target outcome). The weights of the edges increase or decrease the strength of the signal transmitted between the nodes. For example, during the training process, the algorithm adjusts the machine learning parameters according to an optimization technique (such as gradient descent, stochastic gradient descent, or other optimization algorithm) to minimize the error or loss between the predicted output and the actual target. Once the machine learning parameters have been learned from the training data, the machine learning model can be used to make predictions on new, unseen data (i.e., during inference).

[0054] According to an embodiment, UE 120 may receive a CSI-RS from base station 110. UE 120 may calculate a first channel matrix for a downlink channel between UE 120 and base station 110 using the CSI-RS. UE 120 may determine a level s (where s is a positive integer) of multi-stage vector quantization for a potential vector based on at least one of a bandwidth, a capacity, and a resource of an uplink channel between UE 120 and base station 110. According to an embodiment, the level s of the multi-stage vector quantization may increase as at least one of a bandwidth, a capacity, and a resource of an uplink channel increases. UE 120 may extract a potential vector from the first channel matrix based on a first machine learning model.

[0055] As a first stage, UE 120 may select at least one first codeword from a first codebook based on a potential vector, and may generate a first residual potential vector. In a second stage in sequence, UE 120 may select at least one second codeword from a second codebook based on a first residual potential vector input from a previous stage. UE 120 may generate a second residual potential vector to be migrated to the next stage. This process is repeated for each stage until the sth stage, that is, a plurality of residual potential vectors corresponding to a plurality of sequential stages from the first stage to the sth stage are iteratively calculated. In some cases, the first codebook may be the same as the second codebook.

[0056] According to an embodiment, the UE 120 may divide the potential vector into a plurality of sub-potential vectors. In some cases, the UE 120 may perform the generation of the first residual potential vector and the generation of the second residual potential vector based on the sub-potential vector. For example, the UE 120 may perform the generation of the first residual potential vector and the generation of the second residual potential vector in units of sub-potential vectors. The UE 120 may generate a bit stream including index information for at least one first codeword and at least one second codeword. According to an embodiment, the UE 120 may generate a bit stream by concatenating a first bit stream for at least one first codeword and at least one second bit stream for at least one second codeword. The UE 120 may determine the size of the bit stream based on at least one of the bandwidth, capacity, and resources of the uplink channel. According to an embodiment, as at least one of the bandwidth, capacity, and resources of the uplink channel increases, the size of the bit stream may increase. The UE 120 may send the bit stream to the base station 110.

[0057] In addition, the base station 110 may receive a bit stream from the UE 120. The base station 110 may generate at least one first codeword and at least one second codeword based on the index information in the bit stream. The base station 110 may estimate a second channel matrix for a downlink channel from the at least one first codeword and the at least one second codeword based on a second machine learning model. The first machine learning model and the second machine learning model may refer to the same learning model. In some cases, the first machine learning model and the second machine learning model may respectively include an encoder and a decoder of a combined machine learning model.

[0058] UE 120 may train the first machine learning model, and base station 110 may train the second machine learning model. For example, the first machine learning model and the second machine learning model may be trained according to repetitions from the first stage to the sth stage. As the training proceeds, the difference between the first channel matrix and the second channel matrix may decrease, and the difference between the potential vector and the quantized potential vector based on at least one first codeword and at least one second codeword may decrease. In addition, the parameters of the first machine learning model may be updated, the parameters of the second machine learning model may be updated, and the first codebook and the second codebook may be updated. In some cases, the first machine learning model and the second machine learning model are trained by calculating a loss function, and the parameters of the first machine learning model, the parameters of the second machine learning model, the first codebook and the second codebook are updated based on the loss function. According to an embodiment, UE 120 may perform training in the first stage and then perform training in the second stage. In some cases, the first machine learning model and the second machine learning model are trained by performing sequential training in multiple training phases corresponding to multiple training stages of the multi-stage vector quantization process. In some cases, a first training corresponding to the first level is performed on the first machine learning model and the second machine learning model, and a second training corresponding to the second level is performed on the first machine learning model and the second machine learning model based on the result of the first training. The first machine learning model and the second machine learning model may be trained in the first level. In addition, the first machine learning model and the second machine learning model may be trained for the second time in the first level and the second level based on the result of the first training.

[0059] UE 120 may send channel state information feedback to base station 110 based on a small amount of uplink resources. That is, when the CSI-RS feedback payload changes dynamically, UE 120 may adjust the size of a bit stream for CSI-RS feedback based on the CSI-RS feedback payload. In addition, UE 120 may use an automatic encoder (AE) (e.g., a single AE) to adjust the size of a bit stream based on various CSI-RS feedback payloads.

[0060] The UE 120 may efficiently compress the channel state information based on the first machine learning model. The UE 120 may use a single encoder to adjust the number of bits of the bit stream. The UE 120 may compress the channel state information by considering nonlinear characteristics.

[0061] Figure 2 A wireless communication device according to an embodiment is shown.

[0062] Please refer to Figure 1 To describe Figure 2 . Reference Figure 2, the wireless communication device 200 may include an encoder 210 and a multi-stage vector quantizer 220. The encoder 210 and the multi-stage vector quantizer 220 may be implemented by software and may be loaded into a memory. However, the inventive concept is not limited thereto, and in some embodiments, the encoder 210 and the multi-stage vector quantizer 220 may be implemented by hardware and may be arranged outside the memory.

[0063] The wireless communication device 200 may receive a reference signal from another wireless communication device. For example, the wireless communication device 200 may receive a CSI-RS from the base station 110. The wireless communication device 200 may estimate a downlink channel based on the CSI-RS. The wireless communication device 200 may migrate the received channel information for the downlink channel to the encoder 210. For example, the channel information may include a channel matrix for the downlink channel. The encoder 210 may extract a continuous latent vector from the channel information. For example, the encoder 210 may include a first machine learning model. The first machine learning model may extract a continuous latent vector from the channel information. The first machine learning model may be implemented by software and may be loaded in the encoder 210. However, the inventive concept is not limited thereto, and according to an embodiment, the first machine learning model may be implemented by hardware and may be arranged outside the encoder 210. The first machine learning model may include at least one of a convolutional neural network (CNN), a recurrent neural network (RNN), and a deep neural network (DNN). The first machine learning model may include various neural networks and is not limited to the above examples.

[0064] Convolutional neural networks (CNNs) are a class of neural networks that are often used in computer vision or image classification systems. In some cases, CNNs can enable processing of digital images with minimal preprocessing. CNNs can be characterized by using convolutional (or cross-correlation) hidden layers. The convolutional layer applies a convolution operation to the input before transferring the result to the next layer with a signal. Each convolution node can process data for a limited input domain (i.e., receptive domain). During the forward pass of the CNN, the filter at each layer can be convolved on the input volume, thereby calculating the dot product between the filter and the input. During the training process, the filters can be modified so that they are activated when they detect specific features within the input.

[0065] RNN is an artificial neural network designed to process sequential data by maintaining a memory of previous inputs. RNN has loops that provide information that persists across time steps, enabling RNN to capture temporal dependencies in data. In the case of RNN, each neuron receives input from the current time step and receives its own output from the previous time step. This structure allows RNN to generate predictions based on both the current input and the historical context provided by past inputs. RNN is effective for tasks involving sequential data, such as language modeling, speech recognition, and time series forecasting.

[0066] A DNN is an artificial neural network with multiple layers of neurons between the input layer and the output layer. Intermediate layers, called hidden layers, allow the network to learn and model complex patterns in the data by gradually abstracting features at different levels of representation. Each layer in a DNN includes interconnected neurons, each of which applies a weighted sum of the inputs followed by a nonlinear activation function. The deep architecture of a DNN enables the capture of intricate relationships in the data that shallow networks cannot capture, making it very effective for tasks such as image recognition, natural language processing, and speech recognition. Training a DNN involves adjusting the weights of the neurons using backpropagation, which minimizes the error between the predicted output and the actual output by propagating the error gradient through the layers. The deep structure allows DNNs to achieve high accuracy and performance on complex tasks, making them a fundamental tool in modern machine learning and artificial intelligence.

[0067] Reference Figure 2 , the multi-level vector quantizer 220 may compress the channel information by performing multi-level vector quantization on the continuous potential vectors. That is, the multi-level vector quantizer 220 may perform multi-level vector quantization step by step (i.e., through multiple stages). The multi-level vector quantizer 220 may output a bit stream including the quantized channel information. The multi-level vector quantizer 220 may adjust the number of bits of the bit stream by adjusting the number of stages. For example, the multi-level vector quantizer 220 may increase the number of bits in the bit stream by increasing the number of stages.

[0068] Figure 3 A wireless communication device according to an embodiment is shown.

[0069] Figure 3 The wireless communication device 200 shows Figure 2 A specific example of the wireless communication device 200. Figure 2 describe Figure 3 , and repeated descriptions may be omitted. Figure 3, the wireless communication device 200 may include an encoder 210 and multi-stage vector quantizers 220_1, 220_2, 220_3, ..., and 220_s.

[0070] The multi-stage vector quantizers 220_1 , 220_2 , 220_3 , . . . , and 220_s may sequentially perform vector quantization (VQ) in stages from a first stage to an s-th stage, respectively.

[0071] The multi-level vector quantizer 220_1 may perform vector quantization at the first level. The multi-level vector quantizer 220_1 may select at least one first codeword from the codebook based on the potential vector, and may generate a first residual potential vector. For example, the multi-level vector quantizer 220_1 may select at least one first codeword similar to the potential vector, and may output the difference between the potential vector and the at least one first codeword as the first residual potential vector.

[0072] In addition, the multi-level vector quantizer 220_2 may perform vector quantization at the second level. The multi-level vector quantizer 220_2 may select at least one second codeword from the codebook based on the first residual potential vector, and may generate a second residual potential vector. For example, the multi-level vector quantizer 220_2 may select at least one second codeword similar to the first residual potential vector, and may output the difference between the first residual potential vector and the at least one second codeword as the second residual potential vector.

[0073] In addition, the multi-level vector quantizer 220_s may perform vector quantization at the s-th level. The multi-level vector quantizer 220_s may select at least one s-th codeword from the codebook based on the (s-1)-th residual potential vector, and may generate an s-th residual potential vector. For example, the multi-level vector quantizer 220_s may select at least one s-th codeword similar to the (s-1)-th residual potential vector, and may output the difference between the (s-1)-th residual potential vector and the at least one s-th codeword as the s-th residual potential vector based on the selected s-th codeword.

[0074] The wireless communication device 200 can generate a bit stream based on multi-stage vector quantization by splicing the selected codewords (i.e., the selected first codeword, the selected second codeword, the selected third codeword, ..., the selected sth codeword). In addition, the wireless communication device 200 can send the bit stream to another wireless communication device (e.g., Figure 1 The user equipment 120 described in the above may send a bit stream to the base station 110).

[0075] Figure 4 A wireless communication device according to an embodiment is shown.

[0076] Please refer to Figure 2 To describe Figure 4 . Reference Figure 4, the wireless communication device 300 may include a dequantizer 320 and a decoder 310. The decoder 310 and the dequantizer 320 may be implemented by software and may be loaded into a memory. However, the inventive concept is not limited thereto, and according to some embodiments, the decoder 310 and the dequantizer 320 may be implemented by hardware and may be arranged outside the memory.

[0077] The wireless communication device 300 may receive a signal from another wireless communication device (eg, Figure 2 The wireless communication device 200 receives a bit stream based on multi-stage vector quantization.

[0078] The wireless communication device 300 may input the received bitstream to the inverse quantizer 320. The inverse quantizer 320 may estimate the quantized latent vector using the index information for the codeword in the bitstream. That is, the inverse quantizer 320 may output the codeword using the index information for the codeword in the bitstream.

[0079] The decoder 310 may estimate channel information from a codeword (e.g., a quantized latent vector). For example, the decoder 310 may include a second machine learning model. The second machine learning model may extract channel information from the codeword. The second machine learning model may be implemented by software and may be loaded in the decoder 310. However, the inventive concept is not limited thereto, and according to some embodiments, the second machine learning model may be implemented by hardware and may be arranged outside the decoder 310. The second machine learning model may include at least one of a CNN, an RNN, and a DNN. In addition, the second machine learning model may include various neural networks and is not limited to the examples set forth herein. Figure 2 The first machine learning model of the encoder 210 and Figure 4 The second machine learning model of the decoder 310 may use the same neural network.

[0080] Figure 5A Signal transmission and reception between an encoder and a decoder according to an embodiment are shown. Figure 5B A multi-stage vector quantization process of an encoder according to an embodiment is shown.

[0081] Please refer to Figure 3 and Figure 4 To describe Figure 5A and Figure 5B . Reference Figure 5A , the wireless communication system 20 may include an encoder 210, a decoder 310, concatenation units 230_1 to 230_(s-1), quantizers 220_1 to 220_s, and residual units 240_1 to 240_(s-1).

[0082] The AE can perform learned data compression, in which a pair of encoder 210 and decoder 310 are parameterized as a DNN. Figure 5A , the encoder 210 may output a continuous latent vector based on the first machine learning model.

[0083] [Expression 1]

[0084] The encoder 210 can be parameterized as Here, H represents a channel matrix, z represents a continuous latent vector, and s represents a multi-stage vector quantization level (or a level of a multi-stage vector quantization process).

[0085] The quantizer 220_1 at the first stage may be based on a subcodebook The codebook may be a set of sub-codebooks. For example, a codebook may be represented by a set of sub-codebooks, such as, . , can be a divisor of K, and It can be the s-th sub-codebook.

[0086] The continuous latent vector z can be split into p Each sub-vector can be approximated as an additive combination of multiple codewords, as shown in Equation 1.

[0087] [Equation 1]

[0088] p It refers to the index of the segmented sub-vector. Represents a collection of level indices. j is for the vector quantization level s The index of . This can be achieved through one-hot encoding. is the accumulated codebook. Include 1 as an element to provide the element as output. represents the M×K codebook matrix The M×1 codebook vector in, represents S sparse binary vector, and The support set of is given by .

[0089] The wireless communication device 200 may determine the maximum level of multi-level vector quantization according to at least one of the bandwidth, capacity, and resources of the feedback link (eg, uplink). The wireless communication device 200 may determine the maximum level of multi-stage vector quantization according to at least one of the bandwidth, capacity, and resources of the feedback link which may be expressed using Inequality 1: In the case of a UE, when there is a feedback rate constraint of B bits, the feedback rate can be expressed based on the information used to represent the level up to The vector quantization level is determined by the total number of bits of the codeword selected for each of the P segments .

[0090] [Inequality 1]

[0091] here, , can be a divisor of K, and It can be the s-th sub-codebook. represents the feedback rate. B represents the maximum bit rate of the feedback link.

[0092] Vector quantization level It can be determined by Equation 2.

[0093] [Equation 2]

[0094] Perform multi-level vector quantization up to level To discretize the latent vector into .

[0095] The multi-level vector quantizers 220_1 to 220_s perform the multi-level vector quantization from the multi-level vector quantization level s=1 to the multi-level vector quantization level s=1. The processing of multi-level vector quantization can be represented as described in this article.

[0096] In some cases, the discrete representation z can be expressed based on the L2 norm distance. As described herein, z can be expressed using Equation 3.

[0097] [Equation 3]

[0098] Codebook-based C Quantized latent vector. q Indicates that the codebook includes C The code words in .

[0099] The S term can be approximated using Equation 4 and Equation 5 ( S -term approximation).

[0100] [Equation 4]

[0101] [Equation 5]

[0102] optimal S The term approximation can be viewed as a generalization of the vector quantization objective.

[0103] Reference Figure 5B In operation S101, the multi-level vector quantizers 220_1 to 220_s may each use a level s The codebook at selects the codeword closest to the residual latent vector. As described herein, the residual latent vector may be referred to as the residual. s The method of selecting the codeword that is closest to the residual latent vector from the codebook can be expressed using Equation 6.

[0104] [Equation 6]

[0105] represents the codeword selected at level s. Denotes the residual latent vector at stage s-1. In some cases, Equation 6 may be modified to Equation 7.

[0106] [Equation 7]

[0107] here, Based on codebook Select the codeword.

[0108] In operation S103, the multi-stage vector quantizers 220_1 to 220_s may each cause a codeword index of the selected codeword to be included in the codeword index set. Operation S103 may be represented by Equation 8.

[0109] [Equation 8]

[0110] Display level s The codeword index at . represents the codeword index at stage s-1. Equation 8 can be modified to Equation 9.

[0111] [Equation 9]

[0112] is the set of selected codewords. The set of selected codewords may be updated as shown in Equation 10.

[0113] In operation S105, the multi-level vector quantizers 220_1 to 220_s may update the quantized latent vectors to level s. The updating process may be represented by Equation 10.

[0114] [Equation 10]

[0115] In the case of undivided sub-vectors, Equation 10 can be modified to Equation 11.

[0116] [Equation 11]

[0117] In operation S107, each of the multi-stage vector quantizers 220_1 to 220_s may update a residual latent vector, wherein the updated residual latent vector may be expressed as Equation 12.

[0118] [Equation 12]

[0119] In the case of undivided sub-vectors, Equation 12 can be modified to Equation 13.

[0120] [Equation 13]

[0121] The codebooks may be the same or different according to the partition and level. In the case where the codebooks are the same, the efficiency may be improved.

[0122] When p When each of the divided sub-vectors performs operations S101 to S107 described herein, the corresponding p A set of codeword indices. p The codeword index sets may be concatenated and transmitted via the feedback channel after performing channel coding. p The bit stream generated by concatenating a set of codeword indices can be expressed by Expression 14.

[0123] [Expression 14]

[0124] Refer again Figure 5A , the decoder 310 may receive the updated residual latent vector. The decoder 310 may estimate the channel information from the updated residual latent vector based on the second machine learning model (using Expression 15). The wireless communication device 300 may use the estimated channel information to optimize the communication system.

[0125] [Expression 15]

[0126] Decoder 310 may be parameterized as θ.

[0127] The encoder 210 and the decoder 310 may determine a single set of models and codebooks through training, wherein the single set may support multiple compression ratios.

[0128] According to an embodiment, the encoder 210 and the decoder 310 may perform one-shot training. As used herein, one-shot training means that the encoder 210 and the decoder 310 may be trained to combine the first machine learning model, the second machine learning model, and the codebook according to the overhead of the feedback link to adjust the compression ratio of the feedback information. The loss function of the one-shot training may be represented by Equation 16.

[0129] [Equation 16]

[0130] The importance of each of the plurality of compression ratios may be based on (or controlled by) a weight w. , and It is often used to compress channels with different feedback speeds. , and Commonly used to compress channels with different feedback rates.

[0131] The loss function for each compression ratio can be expressed by Equation 17.

[0132] [Equation 17]

[0133] here, is the stopping gradient operator. In Equation 17, the first term represents the reconstruction loss for training a pair of encoder 210 and decoder 310. The second term represents the regularization term for making the output of encoder 210 converge to the closest codeword. Here, is the regularization rate. The third term is used to The codeword closest to the output of encoder 210 is trained under the norm distance measurement. Stop gradient Can be the identity function in the forward pass and can be set to 0 in the backward pass.

[0134] In addition, Equation 17 can be expressed by Equation 18.

[0135] [Equation 18]

[0136] In Equation 18, the first term is used to update the encoder 210, the decoder 310, and the codebook so that the error in channel information recovery is minimized. The second term is used to update the encoder 210 so that the error between the output of the encoder 210 and the quantized latent vector is minimized. The third term is used to update the codeword of the codebook so that the error between the output of the encoder 210 and the quantized latent vector is minimized.

[0137] and The encoder 210 and the decoder 310 are used at the level Optimize compression parameters. is with until level The corresponding multi-level codebook.

[0138] According to an embodiment, the encoder 210 and the decoder 310 may perform sequential training. Fig. 6A , Figure 6B and Figure 7 Further details regarding the sequential training process are described.

[0139] The first potential vector of the quantized latent vector may be processed in one of the base station 120, the wireless communication device 300, and the decoder 310. p The first phase of decoding may be represented by Equation 19. Decoding may include an inverse quantization process. The inverse quantizer 320 may perform operations according to Equation 19.

[0140] [Equation 19]

[0141] In addition, the complete quantized latent vector can be reconstructed by Equation 20.

[0142] [Equation 20]

[0143] The reconstructed quantized latent vector may be input to the decoder 310. In addition, the decoder 310 may reconstruct channel state information.

[0144] Fig. 6A and Figure 6B A sequential training of a wireless communication device according to an embodiment is shown.

[0145] As used herein, sequential training refers to a training method that performs training while adding vector quantization stages sequentially (i.e., one by one). In addition, sequential training is a training method that migrates a previous training result to a next stage, and performs training by using (e.g., using intact) the previous training result in the next stage. Sequential training can reduce (e.g., avoid) task-specific learning based on adopting transfer learning. In addition, sequential training can extract useful prior knowledge and effectively apply it to subsequent training processes.

[0146] In the case of transfer learning, the wireless communication device 200 first generates pre-trained parameters at the previous stage. Subsequently, the wireless communication device 200 may update the pre-trained parameters at the current stage. The generation and update of parameters may be repeated for each stage. Fig. 6A and Figure 6B An example is provided to provide a description of the generation and update of parameters.

[0147] Fig. 6A and Figure 6B A portion of a sequential training process is shown. Fig. 6A shows the training in the first level, and Figure 6B The wireless communication device 200 and the wireless communication device 300 can perform training in the second stage by performing reference Figure 2 and Figure 3 Describes the operations to perform training.

[0148] Reference Fig. 6A , the first machine learning model of the encoder 210 and the second machine learning model of the decoder 310 may be trained when the multi-stage vector quantization level is 1. Therefore, the parameters of the first machine learning model of the encoder 210, the parameters of the second machine learning model of the decoder 310, and the codebook may be updated.

[0149] Reference Figure 6B , the first machine learning model of the encoder 210 and the second machine learning model of the decoder 310 may perform training when the multi-stage vector quantization level is 2. Therefore, the parameters and codebook when the multi-stage vector quantization level is 1 may be optimized. Therefore, the wireless communication device 200 may use the parameters of the previous stage to update the parameters in the next stage.

[0150] Figure 7 A sequential training process of a wireless communication device according to an embodiment is shown.

[0151] Please refer to Figures 2 to 6B To describe Figure 7 . Reference Figure 7 , the wireless communication device 200 transfers knowledge from the pre-trained R1 model to the initialization stage of the R2 specific training stage.

[0152] Reference Figure 7 , the wireless communication system 20 may include an encoder 210, a decoder 310, a splicing unit 230_1, quantizers 220_1 and 220_2, and a residual unit 240_1. Figure 7 It is shown that the training results of the first stage are migrated to the second stage. Specifically, the parameters of the encoder 210, the parameters of the decoder 310 and the codebook updated in the first stage can be migrated to the second stage.

[0153] After performing transfer training for each compression task (e.g., transferring parameters from the first stage to the second stage), the wireless communication device 200 may generate a general model and a general codebook based on the fine-tuning operation. In some cases, the level may be optimized. s The -1 parameter enables task-specific learning of the loss function. In addition, the pre-trained parameters can be used as initialization parameters for training in stage s.

[0154] According to an embodiment, the wireless communication device 200 and the wireless communication device 300 may perform initialization, such as for , and . Available at Subcodebook , and the sub-codebook can be randomly initialized .

[0155] According to the embodiment, the wireless communication device 200 and the wireless communication device 300 may be based on making the objective function Minimize to (e.g., together with information initialization) perform the Training, where the objective function Available for optimization level Performance in.

[0156] After performing transfer learning through each of the s stages, the parameters converge (e.g., become close) to the optimal point for the objective of stage s. Therefore, the parameters of the machine learning model and the codebook are initialized as and Afterwards, fine-tuning can be performed on each of the compression tasks to optimize the trade-off between general performance and task-specific performance.

[0157] Each of the parameters can be fine-tuned based on the weight, as shown in Equation 21.

[0158] [Equation 21]

[0159] here, is the loss function for fine-tuning. In Equation 21, the first term describes the training for various compression ratios and refers to the loss function for one-shot training. In addition, when the weight When , the unquantized reconstruction error can be included in Equation 21. In some cases, the weight The learning of latent space candidates can be accelerated while preserving the low-dimensional manifolds of the channel. The second term in Equation 21 can be used to adjust various parameters of the encoder 210 and the decoder 310 .

[0160] Figure 8 A method of operating a wireless communication device according to an embodiment is shown.

[0161] Reference Figure 8In operation S201, the wireless communication device 200 may receive a reference signal. In some examples, the wireless communication device 200 may receive a channel state information reference signal (CSI-RS) from a base station.

[0162] In operation S203 , the wireless communication device 200 may calculate a first channel matrix for a reception channel based on a reference signal (eg, CSI-RS).

[0163] In operation S205, the wireless communication device 200 may determine the level of the multi-stage vector quantization process for the potential vector based on at least one of the bandwidth, capacity, and resources of the transmission channel. s In some cases, the wireless communication device 200 may determine the level of the multi-stage vector quantization process based on an uplink channel between the wireless communication device and a base station.

[0164] In operation S207, the wireless communication device 200 may extract a potential vector from the first channel matrix based on the first machine learning model. In some cases, the first machine learning model (such as, for example, Figures 2 to 7 The first machine learning model described herein can extract a latent vector based on the first channel matrix.

[0165] In operation S209, the wireless communication device 200 may select a first codeword from a first codebook corresponding to the first stage in a first stage of a multi-stage vector quantization process, and may generate a first residual latent vector based on the latent vector and the selected first codeword.

[0166] In operation S211, the wireless communication device 200 may select a second codeword from a second codebook corresponding to the second stage in a second stage corresponding to the determined stage level s of a multi-stage vector quantization process, and generate a second residual latent vector based on the first residual latent vector and the selected second codeword.

[0167] In operation S213 , the wireless communication device 200 may generate a bit stream based on the first codeword and the second codeword.

[0168] In operation S215, the wireless communication device 200 may transmit a bit stream. In some cases, transmission is performed using an uplink channel based on a base station.

[0169] Fig. 9 A method of operating a UE according to an embodiment is shown.

[0170] Please refer to Figure 1 To describe Fig. 9 . Reference Figures 2 to 7 Provides further details about each of the operations. Fig. 9 In operation S301 , the UE 120 may receive a CSI-RS from the base station 110 .

[0171] In operation S303 , the UE 120 may calculate a first channel matrix for a downlink channel between the UE 120 and the base station 110 based on the CSI-RS.

[0172] In operation S305, the UE 120 may determine a level of multi-stage vector quantization for the potential vector based on at least one of a bandwidth, a capacity, and a resource of an uplink channel between the UE 120 and the base station 110. s In some cases, the wireless communication device (ie, UE 120) may determine the level of the multi-stage vector quantization process based on an uplink channel between the wireless communication device and a base station.

[0173] In operation S307, UE 120 may extract a latent vector from the first channel matrix based on the first machine learning model. In some cases, the first machine learning model may extract a latent vector based on the first channel matrix. According to an embodiment, UE 120 may split the latent vector into a plurality of sub-latent vectors, and may generate a first residual latent vector based on the plurality of sub-latent vectors.

[0174] In operation S309, the UE 120 may select at least one first codeword from a first codebook corresponding to the first stage in a first stage of a multi-stage vector quantization process, and generate a first residual latent vector based on the latent vector and the selected at least one first codeword.

[0175] In operation S311, the UE 120 may select at least one second codeword from a second codebook corresponding to the second stage in the multi-stage vector quantization process corresponding to the determined stage level s, and may generate a second residual latent vector based on the selected at least one second codeword and the first residual latent vector. According to an embodiment, the first codebook may be different from the second codebook. According to another embodiment, the first codebook may be the same as the second codebook.

[0176] In operation S313, the UE 120 may generate a bit stream based on the at least one first code word and the at least one second code word. For example, a bit stream including index information for the at least one first code word and the at least one second code word may be generated. According to an embodiment, the UE 120 may generate a bit stream by splicing a first bit stream for the at least one first code word and at least one second bit stream for the at least one second code word.

[0177] In operation S315, UE 120 may transmit the bit stream to base station 110. In some cases, UE 120 may transmit the bit stream using an uplink channel.

[0178] Fig.10 A method of operating a wireless communication system including a UE and a base station according to an embodiment is shown.

[0179] Please refer to Figure 8 and Fig. 9 right Fig.10 The wireless communication device 300 may operate as a base station. The wireless communication system may include the wireless communication device 200 and the base station 110. Fig.10 In operation S401 , the wireless communication device 200 may receive a CSI-RS from the base station 110 .

[0180] In operation S403 , the wireless communication device 200 may calculate a first channel matrix for a downlink channel between the wireless communication device 200 and the base station 110 using the CSI-RS.

[0181] In operation S405, the wireless communication device 200 may extract a latent vector from the first channel matrix based on the first machine learning model.

[0182] In operation S407, the wireless communication device 200 may perform multi-stage vector quantization. The wireless communication device 200 may perform Figure 8 The operations of operation S209, operation S211 and operation S213 correspond to each other as a multi-stage vector quantization process.

[0183] In operation S409, the wireless communication device 200 may transmit CSI-RS feedback to the base station 110. For example, the wireless communication device 200 may transmit a bit stream based on multi-stage vector quantization to the base station 110.

[0184] In operation S411, the base station 110 may generate at least one first codeword and at least one second codeword based on index information in the CSI-RS feedback.

[0185] In operation S413, the base station 110 may estimate a second channel matrix for a downlink channel from at least one first codeword and at least one second codeword based on a second machine learning model.

[0186] The wireless communication device 200 and the base station 110 may be based on the repeated references herein. Fig.10 The operations described above can be used to perform one-time training or sequential training. According to an embodiment, according to the first level to the second level, s The wireless communication device 200 may train the first machine learning model, and the base station 110 may train the second machine learning model. In some cases, the parameters of the first machine learning model and the parameters of the second machine learning model may be updated, and the first codebook and the second codebook may be updated (e.g., as training progresses) so that the difference between the first channel matrix and the second channel matrix decreases, and the difference between the potential vector and the quantized potential vector decreases, wherein the quantized potential vector may be based on at least one first codeword and at least one second codeword.

[0187] According to an embodiment, the wireless communication device 200 and the base station 110 may undergo sequential training in the first stage and the second stage in the order stated. In addition, the parameters of the first machine learning model and the parameters of the second machine learning model may be updated, and the first codebook and the second codebook may be updated (e.g., as training progresses) so that the difference between the first channel matrix and the second channel matrix decreases, and the difference between the potential vector and the quantized potential vector decreases, wherein the quantized potential vector may be based on at least one first codeword and at least one second codeword.

[0188] Fig.11 A method of operating a wireless communication device according to an embodiment is shown.

[0189] Please refer to Figures 1 to 4 To describe Fig.11 .

[0190] Reference Fig.11 In operation S501 , the wireless communication device 200 may receive a CSI-RS from the base station 110 .

[0191] In operation S503 , the wireless communication device 200 may generate channel state information by estimating a channel based on the CSI-RS.

[0192] In operation S505 , the wireless communication device 200 may generate a potential vector based on the channel state information.

[0193] In operation S507, the wireless communication device 200 may determine the number of stages for quantizing the potential vector. In some cases, the number of stages corresponds to the compression ratio of the channel state information based on the bandwidth of the feedback channel. In some cases, the wireless communication device 200 may determine the number of stages for multi-stage vector quantization processing based on the feedback channel of the base station, wherein the number of stages corresponds to the compression ratio of the channel state information. In some cases, as the compression ratio decreases, the number of stages may increase. In some cases, as the bandwidth of the feedback channel increases, the number of stages may increase. In some cases, the wireless communication device 200 may determine the number of stages based on the state of the feedback channel.

[0194] In operation S509, the wireless communication device 200 may perform multi-stage vector quantization by ensuring that the potential vector sequentially undergoes a stage consistent with the determined stage number. In each stage, the wireless communication device 200 may select a codeword closest to the potential vector from the same codebook. In some cases, the wireless communication device 200 may perform a multi-stage vector quantization process on the potential vector using the determined stage number to obtain a quantized potential vector.

[0195] In operation S511, the wireless communication device 200 may transmit feedback including the quantized latent vector to the base station 110 via the feedback channel. For example, the wireless communication device 200 may transmit the quantized latent vector to the base station using the feedback channel.

[0196] Fig.12 A block diagram of a UE according to an embodiment is shown.

[0197] Reference Fig.12 , the wireless communication device 400 may include a processor 410, an RFIC 420, and a memory 430. Although one processor 410, one RFIC 420, and one memory 430 are shown for ease of description, each of the processor 410, the RFIC 420, and the memory 430 may be provided in plural numbers, and may not be limited to Fig.12 The processor 410 may control the RFIC 420 and the memory 430, and may be configured to implement the operation method and operation flow chart of the wireless communication device 400 according to the present disclosure.

[0198] The wireless communication device 400 may include multiple antennas, and the RFIC 420 may transmit and receive radio signals via one or more antennas. At least some of the multiple antennas may each correspond to a transmitting antenna. The transmitting antenna may transmit a radio signal to an external device (e.g., another UE or another base station) instead of the wireless communication device 400. The remaining antennas of the multiple antennas may each correspond to a receiving antenna. The receiving antenna may receive a radio signal from an external device.

[0199] According to an embodiment, the wireless communication device 400 may include an RFIC 420 and at least one processor 410. The at least one processor 410 may receive a CSI-RS from a base station via the RFIC 420. The at least one processor 410 may calculate a first channel matrix for a downlink channel between the wireless communication device 400 and the base station using the CSI-RS. The at least one processor 410 may determine a level of multi-level vector quantization for a potential vector based on at least one of a bandwidth, a capacity, and a resource of an uplink channel between the wireless communication device 400 and the base station. s (in, s is a positive integer).

[0200] In some cases, the level of multi-level vector quantization sIt may increase with at least one increase of the bandwidth, capacity and resources of the uplink channel. At least one processor 410 may extract a potential vector from a first channel matrix based on a first machine learning model. In the first stage, at least one processor 410 may select at least one first codeword from a first codebook based on the potential vector, and may generate a first residual potential vector based on the selected first codeword. In each of the sequential levels from the second stage to the sth stage, at least one processor 410 may select at least one current level codeword from a codebook corresponding to the current stage based on the residual potential vector of the previous stage input from the previous stage. In addition, at least one processor 410 may generate a residual potential vector of at least one current level to be migrated to the next level based on the selected codeword of at least one current level.

[0201] In addition, at least one processor 410 may generate a bit stream including index information for at least one first codeword and at least one second codeword. At least one processor 410 may be configured to send the bit stream to the base station via RFIC 420. The first codebook may be the same as the second codebook. At least one processor 410 may generate a bit stream by splicing a first bit stream for at least one first codeword and at least one second bit stream for at least one second codeword. In some cases, the size of the bit stream may increase as at least one of the bandwidth, capacity, and resources of the uplink channel increases. At least one processor 410 may be used to divide the potential vector into a plurality of sub-potential vectors. At least one processor 410 may perform the generation of the first residual potential vector and the generation of the third residual potential vector based on the sub-potential vector.

[0202] Fig.13 is a block diagram illustrating an electronic device according to an embodiment.

[0203] Reference Fig.13 , the electronic device 500 may include a memory 510, a processor unit 520, an input / output control unit 540, a display unit 550, an input device 560, and a communication processing unit 590. Here, the memory 1010 may be provided in plural. The description of each component may be made as follows.

[0204] The memory 510 may include a program storage unit 511 and a data storage unit 512, wherein the program storage unit 511 stores a program for controlling the operation of the electronic device 500 and the data storage unit 512 stores data generated during the execution of the program. The data storage unit 512 may store data required for the operation of the application program 513 and the channel information compression program 514. The program storage unit 511 may include the application program 513 and the channel information compression program 514. Here, the program in the program storage unit 511 is a collection of instructions and may be referred to as an instruction set. The application program 513 includes an application program that operates on the electronic device 500. That is, the application program 513 may include instructions for applications driven by the processor 522. The channel information compression program 514 may include instructions for (for example, as referred to in Figures 1 to 12 A procedure for compressing channel state information (described in detail in the specification).

[0205] The peripheral device interface 523 may control the connection between each of the processor 522 and the memory interface 521 and the input / output peripheral device of the base station. By using at least one software program, the processor 522 enables the base station to provide a service corresponding to the software program. Here, by executing at least one program stored in the memory 510, the processor 522 may provide a service corresponding to the program.

[0206] The input / output control unit 540 may provide an interface between input / output devices such as the display unit 550 and the input device 560 and the peripheral device interface 523. The display unit 550 displays station information, input characters, dynamic pictures, static pictures, etc. For example, the display unit 550 may display application information about an application driven by the processor 522.

[0207] The input device 560 may provide input data generated by the selection of the electronic device 500 to the processor unit 520 via the input / output control unit 540. Here, the input device 560 may include a keyboard including at least one hardware button, a touch pad for sensing touch information, etc. For example, the input device 560 may provide touch information (such as touch, touch movement, or touch release) sensed by the touch pad to the processor 522 via the input / output control unit 540. The electronic device 500 may include a communication processing unit 590 configured to perform a communication function for voice communication and data communication.

[0208] The processing discussed above is intended to be exemplary and not restrictive. Those skilled in the art will understand that the steps of the processing discussed herein may be omitted, modified, combined and / or rearranged, and any additional steps may be performed without departing from the scope of the present invention. More generally, the above disclosure is intended to be exemplary and not restrictive. Only the appended claims are intended to limit what the present invention includes. In addition, it should be noted that the features and limitations described in any one embodiment may be applied to any other embodiment herein, and the flowcharts or examples associated with one embodiment may be combined with any other embodiment in a suitable manner, completed in a different order, or completed in parallel. In addition, the systems and methods described herein may be executed in real time. It should also be noted that the above-described systems and / or methods may be applied to other systems and / or methods, or used according to other systems and / or methods.

[0209] While the inventive concept has been particularly shown and described with reference to embodiments thereof, it will be understood that various changes in form and details may be made therein without departing from the spirit and scope of the appended claims.

Claims

1. A method of operating a wireless communication device, the method comprising: receiving a channel state information reference signal CSI-RS from a base station; Calculate a first channel matrix for a downlink channel between the wireless communication device and the base station based on the CSI-RS; determining a level of a multi-stage vector quantization process based on an uplink channel between the wireless communication device and the base station; extracting a latent vector based on a first channel matrix using a first machine learning model; generating a first residual latent vector based on the latent vector in a first stage of the multi-stage vector quantization process, wherein the first residual latent vector is based on a first codeword selected from a first codebook corresponding to the first stage; generating a second residual latent vector based on the first residual latent vector in a second stage of the multi-stage vector quantization process corresponding to the determined stage level, wherein the second residual latent vector is based on a second codeword selected from a second codebook corresponding to the second stage; generating a bitstream based on the first codeword and the second codeword; and The bit stream is sent to the base station using the uplink channel.

2. The method according to claim 1, wherein: The step of generating a second residual latent vector comprises: A plurality of residual latent vectors corresponding to a plurality of sequential levels from a first level to a second level are iteratively calculated.

3. The method according to claim 1, wherein: The step of generating the bit stream comprises: A first bit stream corresponding to the first code word and a second bit stream corresponding to the second code word are concatenated.

4. The method according to claim 1, wherein: The stage level of the multi-stage vector quantization process is determined based on at least one of bandwidth, capacity, and resources of the uplink channel.

5. The method according to claim 1, wherein: The size of the bit stream is determined based on at least one of a bandwidth, a capacity, and a resource of the uplink channel.

6. The method according to claim 1, further comprising: The latent vector is divided into a plurality of sub-latent vectors, wherein a first residual latent vector is generated based on the plurality of sub-latent vectors.

7. A method of operating a system, the method comprising: The wireless communication device receives a channel state information reference signal CSI-RS from a base station; Calculating, by the wireless communication device, a first channel matrix for a downlink channel between the wireless communication device and the base station based on the CSI-RS; Determining, by the wireless communication device, a level of a multi-stage vector quantization process based on an uplink channel between the wireless communication device and the base station; extracting a latent vector based on a first channel matrix using a first machine learning model of the wireless communication device; generating a first residual latent vector based on the latent vector in a first stage of the multi-stage vector quantization process, wherein the first residual latent vector is based on a first codeword selected from a first codebook corresponding to the first stage; generating a second residual latent vector based on a second codeword selected from a second codebook corresponding to the second stage in the multi-stage vector quantization process corresponding to the determined stage level; generating, by the wireless communication device, a bit stream based on the first codeword and the second codeword; The wireless communication device transmits the bit stream to the base station using the uplink channel; Receiving the bit stream by the base station; generating, by the base station, a first codeword and a second codeword based on the bit stream; and A second channel matrix for the downlink channel is estimated by the base station from the first codeword and the second codeword based on a second machine learning model.

8. The method according to claim 7, wherein: The step of generating a second residual latent vector comprises: A plurality of residual latent vectors corresponding to a plurality of sequential levels from a first level to a second level are iteratively calculated.

9. The method according to claim 7, wherein: The step of generating the bit stream comprises: A first bit stream corresponding to the first code word and a second bit stream corresponding to the second code word are concatenated.

10. The method according to claim 7, wherein: The stage level of the multi-stage vector quantization process is determined based on at least one of bandwidth, capacity, and resources of the uplink channel.

11. The method according to claim 7, wherein: The size of the bit stream is determined based on at least one of a bandwidth, a capacity, and a resource of the uplink channel.

12. The method according to claim 7, further comprising: The wireless communication device divides the latent vector into a plurality of sub-latent vectors, wherein a first residual latent vector is generated based on the plurality of sub-latent vectors.

13. The method according to claim 7, wherein: The first machine learning model and the second machine learning model respectively include an encoder and a decoder of the combined machine learning model.

14. The method according to claim 7, further comprising: The first machine learning model and the second machine learning model are trained by calculating the loss function, and the parameters of the first machine learning model, the parameters of the second machine learning model, the first codebook and the second codebook are updated based on the loss function.

15. The method according to claim 7, further comprising: The first machine learning model and the second machine learning model are trained by performing sequential training in multiple training stages corresponding to the multiple training stages of the multi-stage vector quantization process.

16. The method according to claim 15, wherein: The steps of performing the sequential training include: performing a first training corresponding to the first level on the first machine learning model and the second machine learning model; and A second training corresponding to the second level is performed on the first machine learning model and the second machine learning model based on the result of the first training.

17. A channel feedback method for a wireless communication device, the method comprising: receiving a channel state information reference signal CSI-RS from a base station; generating channel state information based on the CSI-RS; generating a potential vector based on the channel state information; Determining the number of stages for multi-stage vector quantization processing based on a feedback channel of the base station, wherein the number of stages corresponds to a compression ratio for the channel state information; performing the multi-level vector quantization process on the latent vector using the determined number of levels to obtain a quantized latent vector; and The quantized latent vector is sent to the base station using the feedback channel.

18. The method according to claim 17, wherein: The number of stages increases as the compression ratio decreases.

19. The method according to claim 17, wherein: The number of stages increases as the bandwidth of the feedback channel increases.

20. The method according to claim 17, wherein: The steps of performing the multi-stage vector quantization process include: A plurality of residual latent vectors corresponding to the determined number of stages are iteratively calculated.

Images