Method for operating electronic device, method for operating base station, and method for operating wireless communication system

Abstract

A method of operating an electronic device, a method of operating a base station, and a method of operating a wireless communication system are provided. The method of operating a base station includes: transmitting a reference signal to an electronic device including a first model and a third model; receiving a first intermediate output from the electronic device; obtaining a second intermediate output by inputting the first intermediate output to a partial model from a second model other than the output layer, the second model comprising a second neural network for data decompression; transmitting the second intermediate output to the electronic device; receiving a first gradient value from the electronic device and updating a weight parameter of the partial model; and generating a second gradient value different from the first gradient value, and transmitting the second gradient value to the electronic device.

Description

[0001] This application is based on and claims the benefit of priority from Korean Patent Application No. 10-2023-0097059 filed on July 25, 2023, in the Korean Intellectual Property Office, and Korean Patent Application No. 10-2023-0187527 filed on December 20, 2023, in the Korean Intellectual Property Office, the disclosures of which are incorporated herein in their entirety by reference. Technical Field

[0002] The disclosure relates to an electronic device for feeding back channel state information (CSI) and a method of operating the electronic device, and more particularly, to an electronic device for feeding back CSI by using an autoencoder based on modified segmentation learning and a method of operating the electronic device. Background Art

[0003] After the commercialization of the fourth generation (4G) communication system, the fifth generation (5G) communication system and the quasi-5G (or pre-5G) communication system have been developed to meet the increased demand for wireless data services. The 5G communication system is a pre-5G communication system referred to as a "new radio (NR)" system in the third generation partnership project (3GPP) standard.

[0004] In order to achieve high data transmission rates, 5G communication systems operate in ultra-high frequency (mmWave) bands (e.g., 28 GHZ band, 39 GHZ band, etc.) In addition, in order to reduce the path loss of radio waves and increase the transmission distance of radio waves in the ultra-high frequency band, technologies such as beamforming, massive multiple-input multiple-output (MIMO), full-dimensional MIMO (FD-MIMO), array antennas, analog beamforming, hybrid beamforming, and large antennas have been discussed for 5G communication systems. Summary of the invention

[0005] An electronic device and a method of operating the electronic device are provided, wherein the electronic device is used to perform channel state information (CSI) feedback based on a neural network of an auto encoder for compression and decompression.

[0006] According to the disclosed aspects, a method for operating an electronic device includes: obtaining a channel estimation result based on a reference signal received from a base station; obtaining a first intermediate output by inputting the channel estimation result into a first model including a first neural network for data compression; sending the first intermediate output to the base station; receiving a second intermediate output different from the first intermediate output from the base station; obtaining a decompressed channel estimation result by inputting the second intermediate output into a third model, the third model corresponding to an output layer of a second model including a second neural network for data decompression; calculating a loss between an obtained channel estimation result and a decompressed channel estimation result by comparing the obtained channel estimation result with the decompressed channel estimation result; sending a first gradient value for minimizing the loss to the base station; receiving a second gradient value different from the first gradient value from the base station; and updating the first model based on the second gradient value.

[0007] According to the disclosed aspects, a method for operating a base station includes: sending a reference signal to an electronic device including a first model and a third model; receiving a first intermediate output from the electronic device; obtaining a second intermediate output by inputting the first intermediate output to a partial model from a second model excluding an output layer, the second model including a second neural network for data decompression; sending the second intermediate output to the electronic device; receiving a first gradient value from the electronic device and updating a weight parameter of the partial model; and generating a second gradient value different from the first gradient value, and sending the second gradient value to the electronic device.

[0008] According to the disclosed aspects, a method for operating a wireless communication system, the wireless communication system includes an electronic device and a base station, the electronic device includes an encoder for data compression, and the base station includes a decoder for data decompression, the method including: sending a reference signal to the electronic device through the base station; obtaining a channel estimation result based on the reference signal through the electronic device; inputting the channel estimation result to the encoder and obtaining an intermediate output through the electronic device; sending the obtained intermediate output to the base station through the electronic device; and inputting the intermediate output to the decoder through the base station, and obtaining a decompressed channel estimation result through the base station. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Embodiments will be more clearly understood from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0010] Figure 1 An example of a wireless communication system according to an embodiment is shown;

[0011] Figure 2 A base station according to an embodiment is shown;

[0012] Figure 3 An electronic device according to an embodiment is shown;

[0013] Figure 4 shows an example of a neural network according to an embodiment;

[0014] Figure 5 A method of operating a wireless communication system according to an embodiment is shown;

[0015] Fig. 6A A specific method of operation for training an autoencoder according to an embodiment is shown;

[0016] Figure 6B illustrates an example of signal exchange in a process of training an autoencoder according to an embodiment;

[0017] Figure 7 shows an example of an autoencoder according to an embodiment;

[0018] Figure 8 A signal exchange diagram of a wireless communication system according to an embodiment is shown;

[0019] Fig. 9 Another example of an autoencoder according to an embodiment is shown;

[0020] Fig.10 Another example of an autoencoder according to an embodiment is shown;

[0021] FIG. 11A to FIG. 11E An example of an autoencoder according to a comparative example is shown; and

[0022] Fig.12 A block diagram of a wireless communication device according to an embodiment is shown. DETAILED DESCRIPTION

[0023] Hereinafter, the embodiments are described in detail with reference to the accompanying drawings. The terms used in the disclosure are only provided for describing specific embodiments and are not intended to limit the scope of other embodiments. Unless the context clearly stipulates otherwise, the singular form includes plural indicators. The terms and words (including technical terms or scientific terms) used herein may have the same meaning as those generally understood by those skilled in the art. The terms as usually defined in the dictionary may be interpreted as having the same or similar meaning as the contextual meaning of the relevant field. Unless otherwise defined, the terms should not be interpreted as idealized or over-formalized meanings. Even if the terms are defined in the disclosure, the terms should not be interpreted as excluding disclosed embodiments in some cases.

[0024] Whether two or more elements are in physical contact with each other, the term "combination" and its derivatives represent any direct or indirect communication between two or more elements. The terms "send", "receive" and "communication" and their derivatives include both direct and indirect communication. The terms "include" and "comprise" and their derivatives represent including but not limited to. The term "or" is an inclusive term representing "and / or". The phrase "associated with..." and its derivatives represent including, included in, interconnected with, included in, included in, connected to or connected with, combined with, combined with, communicated with, collaborated with, interlaced, juxtaposed, close to, bound to or bound with, have, have the nature of, have a relationship with, etc. The term "controller" represents any device, system or part thereof that controls at least one operation. Whether local or remote, the functions associated with any particular controller can be centralized or distributed. When the phrase "at least one of..." is used with a list of items, it means that different combinations of one or more of the listed items can be used, and only one item in the list may be required. For example, "at least one of A, B, and C" includes any of the following combinations: A, B, C, A and B, A and C, B and C, and A and B and C, and any variations thereof. As an additional example, the expression "at least one of a, b, or c" may refer to only a, only b, only c, both a and b, both a and c, both b and c, all of a, b, and c, or variations thereof. Similarly, the term "set" means one or more. Thus, a set of items may be a single item or a set of two or more items.

[0025] In addition, the multiple functions described below can be implemented or supported by one or more computer programs, each of which is formed by a computer-readable program code and implemented in a computer-readable medium. The terms "application" and "program" represent one or more computer programs, software components, instruction sets, processes, functions, objects, classes, instances, related data or a part thereof that are applicable to implementation in a suitable computer-readable program code. The phrase "computer-readable program code" includes any type of computer code (including source code, object code and executable code). The phrase "computer-readable medium" includes any type of medium (such as a read-only memory (ROM), random access memory (RAM), hard drive, compact disk (CD), digital video disk (DVD) or any other type of memory) that can be accessed by a computer. "Non-transitory" computer-readable medium does not include wired, wireless, optical or other communication links that transmit temporary electrical signals or other signals. Non-transitory computer-readable media include media that can permanently store data and media that can store data and rewrite data later (such as a rewritable optical disk or an erasable memory device).

[0026] Figure 1 An example of a wireless communication system 10 according to the embodiment is shown.

[0027] Reference Figure 1 , the wireless communication system 10 may include a base station 110 and an electronic device 120. The base station 110 and the electronic device 120 may be illustrated as nodes using a wireless channel in the wireless communication system 10.

[0028] According to an embodiment, the base station 110 may include a network infrastructure that provides wireless access to the electronic device 120. The base station 110 may have a coverage area defined as a specific geographical area based on the distance traveled by the transmission signal. The base station 110 may be referred to as or correspond to an access point (AP), an evolved Node B (eNodeB) (eNB), a fifth generation node (5G node), a gNodeB (gNB), a wireless point, or other terms having equivalent technical meanings thereto.

[0029] According to an embodiment, the base station 110 may be connected to one or more transmission / reception points (TRPs). The base station 110 may send a downlink signal to the electronic device 120 or receive an uplink signal from the electronic device 120 via one or more TRPs. For example, the base station 110 may send a reference signal (in some cases, RS) to the electronic device 120 to determine the channel environment with the electronic device 120. The reference signal may include a channel state information (CSI) reference signal (RS) (in some cases, CSI-RS or CSI RS). In another example, the base station 110 may receive an "intermediate output" from the electronic device 120, and the base station 110 may obtain the CSI through the "intermediate output". The intermediate output may correspond to the output of an encoder (or an encoding neural network or a compression neural network) located in the electronic device 120 among the autoencoders. The intermediate output may be referred to as or correspond to a latent vector, a latent variable, a code, and a feature. The intermediate output includes data for obtaining a channel estimation result by being decompressed via a decoder of the base station 110. When the electronic device 120 receives a channel estimation result estimated based on a reference signal as input, the encoder may output an intermediate output. The base station 110 may receive the intermediate output from the electronic device 120 and may input the intermediate output into a decoder (or a decoding neural network or a decompression neural network) located in the base station 110 in the autoencoder to obtain a decompressed channel estimation result.

[0030] In other words, the autoencoder may correspond to a neural network for “compression” and “decompression”. The autoencoder uses the encoder model of the electronic device 120 to “compress” the channel estimation result input by the electronic device 120, so that the channel estimation result is sent to the base station 110 as an intermediate output. And, the autoencoder uses the decoder model of the base station 110 to “decompress” the intermediate output, so that the base station 110 obtains the channel estimation result estimated by the electronic device 120.

[0031] According to an embodiment, the electronic device 120 may be used by a user and communicate with the base station 110 via a wireless channel. The electronic device 120 may be referred to as or correspond to a terminal, a user equipment (UE), a mobile station, a subscriber station, a customer premises equipment (CPE), a remote terminal, a wireless terminal, a user device, or other terms having equivalent technical meanings thereto.

[0032] According to one or more embodiments, the autoencoder may be based on "modified segmentation learning". "Segmentation learning" may refer to a configuration in which the training of the encoder is performed by the electronic device 120 and the training of the decoder is independently performed by the base station 110. On the other hand, "modified segmentation learning" may refer to a configuration in which only the output layer of the decoder is separated and trained in the electronic device 120. The modified segmentation learning is described in detail below.

[0033] Figure 2 is a block diagram of a base station 110 according to an embodiment. Figure 2 The base station 200 in Figure 1 The base station 200 may include a wireless communication circuit 210, a backhaul communication circuit 220, a memory 230, and a control circuit 240.

[0034] According to an embodiment, the wireless communication circuit 210 may perform the function of sending and receiving signals through a wireless channel. According to an embodiment, the wireless communication circuit 210 may perform the conversion function between the baseband signal and the bit string according to the physical layer standard of the system. For example, when sending data, the wireless communication circuit 210 may generate a complex symbol by encoding and modulating the transmitted bit string. In addition, when receiving data, the wireless communication circuit 210 may restore the received bit string by demodulating and decoding the baseband signal. In addition, the wireless communication circuit 210 may up-convert the baseband signal to a radio frequency (RF) band signal, and then send the RF band signal through an antenna, or may down-convert the RF band signal received through an antenna to a baseband signal. In some embodiments, the wireless communication circuit 210 may include a transmit filter, a receive filter, an amplifier, a mixer, an oscillator, a digital-to-analog converter (DAC), an analog-to-digital converter (ADC), and the like.

[0035] The wireless communication circuit 210 may transmit and receive signals. For example, the wireless communication circuit 210 may transmit a synchronization signal (SS), a reference signal, system information, a message, control information, or data. In addition, the wireless communication circuit 210 may perform beamforming. The wireless communication circuit 210 may apply beamforming weights to the signals so as to provide directivity to the signals to be transmitted and the signals to be received. The wireless communication circuit 210 may change the formed beam and repeatedly transmit the signal.

[0036] According to an embodiment, the backhaul communication circuit 220 provides an interface for communicating with other nodes in the network. That is, the backhaul communication circuit 220 can convert a bit string sent from the base station 200 to another node (e.g., another access node, another base station, an upper node (higher node), a core network, etc.) into a physical signal, and can convert a physical signal received from another node into a bit string.

[0037] According to an embodiment, the memory 230 stores data such as basic programs, application programs, and setting information for operations of the base station 200. Throughout the disclosure, the memory 230 may represent or correspond to at least one memory or at least one memory device.

[0038] The memory 230 may include a volatile memory, a nonvolatile memory, or a combination of a volatile memory and a nonvolatile memory. According to an embodiment, the memory 230 may include a decoder model 231. The decoder model 231 may include a decoder module configured to receive a decoder from an electronic device (e.g., Figure 1 The electronic device 120 receives the intermediate output as the weight parameters of multiple layers of the neural network of the input. The decoder model 231 can decompress the intermediate output and output the decompressed channel estimation result.

[0039] The decoder model 231 may include at least one of various types of models such as a convolutional neural network (CNN) (e.g., GoogleNet, AlexNet, VGG network, etc.), a region with a convolutional neural network (R-CNN), a region proposal network (RPN), a recursive neural network (RNN), a stacked deep neural network (S-DNN), a state-space dynamic neural network (S-SDNN), a deconvolution network, a deep belief network (DBN), a restricted Boltzmann machine (RBM), a fully convolutional network, a long short-term memory (LSTM) network, and a classification network. However, the embodiments are not limited to the above models. The decoder model 231 may correspond to related software, software code, at least one code, at least one application, at least one program, or at least one computer program code.

[0040] The control circuit 240 may control the operation of the base station 200. Throughout the disclosure, the control circuit 240 may represent or correspond to at least one control circuit or at least one processor.

[0041] For example, the control circuit 240 sends and receives signals through the wireless communication circuit 210 or the backhaul communication circuit 220. In addition, the control circuit 240 may write data to the memory 230 and read data from the memory 230. In some embodiments, the control circuit 240 may include at least one processor. For example, the control circuit 240 may include at least one of a central processing unit (CPU), a neural processing unit (NPU), and a graphics processing unit (GPU). The control circuit 240 may perform training of a decoder model or inference of a decoder model according to an operation mode.

[0042] For example, in the training phase, the control circuit 240 may sequentially calculate and store intermediate parameters starting from the input layer of the decoder model based on the intermediate output received from the electronic device 120. In addition, the control circuit 240 may update the weight parameters through back propagation based on the gradient values.

[0043] In the inference phase, the control circuit 240 may sequentially calculate intermediate parameters starting from an input layer of the decoder model based on the intermediate output received from the electronic device, and the control circuit 240 may obtain a decompressed channel estimation result as a final output.

[0044] Figure 3 is a block diagram of an electronic device 300 according to an embodiment. Figure 3 , the electronic device 300 may include a communication circuit 310, a memory 320, and a control circuit 330. The electronic device 300 may be connected to Figure 1 Corresponding to the electronic device 120.

[0045] According to an embodiment, the communication circuit 310 performs the function of sending and receiving signals through a wireless channel. For example, the communication circuit 310 performs the conversion function between the baseband signal and the bit string according to the physical layer standard of the system. For example, when sending data, the communication circuit 310 can generate symbols by encoding and modulating the transmitted bit string. In addition, when receiving data, the communication circuit 310 can restore the received bit string by demodulating and decoding the baseband signal. In addition, the communication circuit 310 can up-convert the baseband signal to an RF band signal, and then send the RF band signal through an antenna, or can down-convert the RF band signal received by the antenna to a baseband signal. For example, the communication circuit 310 may include a transmission filter, a reception filter, an amplifier, a mixer, an oscillator, a DAC, an ADC, etc. The communication circuit 310 can perform beamforming. The communication circuit 310 can apply beamforming weights to the signal so as to provide directivity to the signal to be sent and the signal to be received.

[0046] According to an embodiment, the communication circuit 310 may send a signal and receive a signal. The communication circuit 310 may receive a downlink signal. The downlink signal may include an SS, an RS, system information, a configuration message, control information, or downlink data. In addition, the communication circuit 310 may send an uplink signal. The uplink signal may include a random access related signal, a reference signal (e.g., a sounding reference signal (SRS), a demodulation (DM)-RS, a phase tracking RS (PT-RS)), or uplink data including CSI feedback information.

[0047] According to an embodiment, the memory 320 may store data such as basic programs, application programs, and setting information for operations of the electronic device 300. Throughout the disclosure, the memory 320 may represent or correspond to at least one memory or at least one memory device.

[0048] The memory 320 may include a volatile memory, a nonvolatile memory, or a combination of a volatile memory and a nonvolatile memory. In addition, the memory 320 may provide the stored data in response to a request from the control circuit 330. According to an embodiment, the memory 320 may include an encoder model 321. The encoder model 321 may include weight parameters of a plurality of layers constituting a neural network, and the neural network receives the weight parameters of the plurality of layers based on the received data from the base station (e.g., Figure 1 The encoder model 321 compresses the channel estimation result to generate an intermediate output. The encoder model 321 may include at least one of various types of models (such as CNN (e.g., GoogleNet, AlexNet, VGG network, etc.), R-CNN, RPN, RNN, S-DNN, S-SDNN, deconvolution network, DBN, RBM, full convolution network, LSTM network, and classification network). However, the embodiment is not limited to the above-mentioned models.

[0049] In addition, the encoder model 321 and the decoder model constituting the autoencoder are not limited to the same model. According to one or more embodiments, the encoder model may be different from the decoder model. For example, the encoder model may be based on CNN, and the decoder model may be based on LSTM network.

[0050] The control circuit 330 may control the overall operation of the electronic device 300. Throughout the disclosure, the control circuit 330 may represent or correspond to at least one control circuit or at least one processor.

[0051] For example, the control circuit 330 may send and receive signals through the communication circuit 310. In addition, the control circuit 330 may write data to the memory 320 and read data from the memory 320. To this end, the control circuit 330 may include at least one processor or microprocessor, or may be a part of a processor. When the control circuit 330 is a part of a processor, the communication circuit 310 and a part of the control circuit 330 may be referred to as a communication processor (CP). According to an embodiment, the control circuit 330 may include at least one of a CPU, an NPU, and a GPU. The control circuit 330 may perform training of an encoder model or inference of an encoder model according to an operation mode.

[0052] For example, in the training phase, the control circuit 330 may sequentially calculate and store intermediate parameters starting from the input layer of the encoder model based on the channel estimation result. In addition, the control circuit 330 may update the weight parameters through back propagation based on the gradient value.

[0053] In the inference phase, the control circuit 330 may sequentially calculate intermediate parameters starting from the input layer of the encoder model based on the channel estimation result, and may generate an intermediate output as a final output and deliver the intermediate output as the final output to the base station 110. The base station 110 may obtain a decompressed channel estimation result based on the decoder model using the intermediate output as an input.

[0054] Figure 4 is an example of a neural network NN according to an embodiment.

[0055] Reference Figure 4 , the neural network NN can be used with Figure 2 The decoder model 231 and Figure 3 Each of the encoder models 321 in corresponds. The neural network NN may include a plurality of layers L1 to Ln (n may be an integer greater than 1). Each of the plurality of layers L1 to Ln may include a linear layer or a nonlinear layer. According to an embodiment, at least one linear layer and at least one nonlinear layer may be combined with each other and referred to as a single layer. For example, the linear layer may include a convolutional layer or a fully connected layer, and the nonlinear layer may include a sampling layer, a pooling layer, and an activation layer.

[0056] For example, the first layer L1 may include a convolutional layer, and the second layer L2 may include a sampling layer. The neural network NN may also include activation layers and hidden layers that perform different types of operations.

[0057] Each of the multiple layers L1 to Ln may receive input image data or a feature map generated in a previous layer as an input feature map, and generate an output feature map by calculating the input feature map. In addition, the feature map represents data expressing various characteristics of the input data. Each of the feature maps FM1, FM2, and FM3 (or referred to as the first feature map FM1 to the third feature map FM3) may be set as, for example, a 2-dimensional matrix or a 3-dimensional matrix. The first feature map FM1 to the third feature map FM3 may have a width W (or column), a height H (or row), and a depth D that may correspond to the x-axis, y-axis, and z-axis of the coordinate system, respectively. In addition, the depth D may be referred to as the number of channels.

[0058] The first layer L1 may generate a second feature map FM2 by convolving the first feature map FM1 with a weight map WM. The weight map WM may filter the first feature map FM1 and is therefore referred to as a filter or a kernel. For example, the depth of the weight map WM (i.e., the number of channels) may be the same as the depth of the first feature map FM1 (i.e., the number of channels). The weight map WM and the first feature map FM1 of the same channel may be convolved with each other. The weight map WM is shifted by passing through the first feature map FM1 as a sliding window. The amount of shift may be referred to as a "stride length" or "stride". During each shift, each of the weights in the weight map WM may be multiplied and added to all feature values ​​in the first feature map FM1 in an area overlapping with the weight map WM. When the first feature map FM1 and the weight map WM are convolved with each other, one channel of the second feature map FM2 may be created. Although in Figure 4 A single weight map WM is shown in FIG, but basically, a plurality of weight maps may be convolved with the first feature map FM1 to generate a plurality of channels of the second feature map FM2. In other words, the number of channels of the second feature map FM2 may correspond to the number of weight maps.

[0059] The second layer L2 may generate a third feature map FM3 by changing the spatial size of the second feature map FM2. For example, the second layer L2 may include a sampling layer. The second layer L2 may perform upsampling or downsampling, and may select a portion of the data of the second feature map FM2. For example, a 2-dimensional window WD may be shifted on the second feature map FM2 in units of the size of the window WD (e.g., a 4×4 matrix), and a value at a specific position (e.g., the first row and the first column) in the area overlapping the window WD in the second feature map FM2 is selected. The second layer L2 may output the selected data as data for the third feature map FM3. In another example, the second layer L2 may include a pooling layer. In this case, the second layer L2 may select the maximum value (or the average value of the eigenvalues) in the area overlapping the window WD in the second feature map FM2. The second layer L2 may output the selected data as data for the third feature map FM3.

[0060] Therefore, a third feature map FM3 having a changed spatial size may be generated from the second feature map FM2. The number of channels of the third feature map FM3 may be equal to the number of channels of the second feature map FM2. In addition, according to an embodiment, the calculation speed of the sampling layer may be faster than the calculation speed of the pooling layer. The sampling layer may improve the quality of the output image (for example, in terms of peak signal-to-noise ratio (PSNR)). For example, in the calculation using the pooling layer, a maximum value or an average value may be calculated. Therefore, the calculation time using the pooling layer may be longer than the calculation time using the sampling layer.

[0061] According to an embodiment, the second layer L2 is not limited to a sampling layer or a pooling layer. That is, the second layer L2 may include a convolution layer similar to the first layer L1. The second layer L2 may generate a third feature map FM3 by convolving the second feature map FM2 with a weight map. In this case, the weight map (the weight map on which the convolution operation is performed on the second layer L2) may be different from the weight map WM on which the convolution operation is performed on the first layer L1.

[0062] The Nth feature map may be created in the Nth layer by a plurality of layers including a first layer L1 and a second layer L2. The Nth feature map may be input to a reconstruction layer "located at the back end of an artificial neural network through which output data is output". The reconstruction layer may generate an output image based on the Nth feature map. In addition, the reconstruction layer may receive the Nth feature map and a plurality of feature maps (such as a first feature map FM1 and a second feature map FM2). The reconstruction layer may generate an output image based on the plurality of feature maps.

[0063] The third layer L3 may generate output data (REC) by combining features (CL) of the third feature map FM3. For example, the output data may include data obtained when compressed input data is decompressed.

[0064] Figure 5 is a flow chart illustrating a method of operating the wireless communication system 10 according to an embodiment.

[0065] Reference Figure 5 In operation S510, a wireless communication system (eg, Figure 1 An electronic device (eg, Figure 1 The electronic device 120 in the example may perform channel estimation. For example, the electronic device 120 may obtain a signal from a base station (e.g., Figure 1 The base station 110 in the embodiment receives a reference signal for channel estimation and obtains a channel estimation result based on the reference signal.

[0066] In operation S520, an autoencoder (or autoencoder model) that performs compression and decompression on CSI feedback may be trained. For example, the autoencoder may include an encoder model for compressing input data and a decoder model for decompressing the compressed input data. The encoder model may be provided in the electronic device 120. The decoder model may be provided in the base station 110. In addition, as Figure 3 As shown in , the encoder model 321 is included in the electronic device 300 , and the decoder model 231 is included in the base station 200 .

[0067] The electronic device 120 may train an encoder model by using the channel estimation result obtained in operation S510 as an input. The encoder model may generate an intermediate output corresponding to the channel estimation result. The electronic device 120 may send the intermediate output to the base station 110 and the base station 110 may train a decoder model.

[0068] According to an embodiment, the head model of the autoencoder may be included in the electronic device 120 during the training process. The head model may include at least one output layer of the decoder model. That is, the head model may include at least one layer in the decoder model that generates a decompressed channel estimation result.

[0069] During the training process, the decoder model of the base station 110 may include only the remaining layers except for the at least one output layer. Figure 2 The signal output by the decoder model 231 in the base station 110 may include an intermediate output instead of the decompressed channel estimation result. Hereinafter, the intermediate output outputted from the encoder model of the electronic device 120 is referred to as the "first" intermediate output. The other intermediate output outputted from the decoder model of the base station 110 is referred to as the "second" intermediate output.

[0070] The base station 110 may deliver the second intermediate output to the electronic device 120. The electronic device 120 may receive the second intermediate output from the base station 110 and use the second intermediate output as an input to train the head model. The output of the head model may be a result of decompressing the channel estimation result estimated by the electronic device 120 and compressed by the encoder model.

[0071] The electronic device 120 may calculate the loss through compression and decompression by comparing the output of the head model with the pre-estimated channel estimation result. The electronic device 120 may generate a gradient value for updating the weight parameter based on the loss. However, since the previous (or previous) layer of the head model exists in the decoder model of the base station 110, the electronic device 120 may send the gradient value to the base station 110. The base station 110 may update the weight parameter of the decoder model based on the gradient value.

[0072] Similarly, because the previous layer of the decoder model exists in the encoder model of the electronic device 120, the gradient value may be provided to the electronic device 120 again. In addition, the gradient value for updating the weight parameter of the decoder model may be referred to as a "first" gradient value, and the gradient value for updating the weight parameter of the encoder model may be referred to as a "second" gradient value. The electronic device 120 may receive the second gradient value from the base station 110 and update the weight parameter of the encoder model based on the second gradient value. When the training is completed, the electronic device 120 may provide the head model of the electronic device 120 to the base station 110. This is because the base station 110 can obtain the overall decoder model by concatenating the head model and the decoder model.

[0073] In operation S530, the electronic device 120 may use the inference of the auto encoder to send CSI feedback to the base station 110. After the training of the auto encoder is completed in operation S520 and the electronic device 120 sends the head model to the base station 110, operation S530 may be performed. The electronic device 120 may receive a reference signal for CSI, perform channel estimation based on the reference signal, and obtain a channel estimation result. The electronic device 120 may input the channel estimation result into the encoder model of the electronic device 120. The encoder model may generate an intermediate output by compressing the channel estimation result. The electronic device 120 may send the intermediate output to the base station 110. The base station 110 may input the intermediate output into the decoder model and obtain the decompressed channel estimation result. In addition, the decoder model to which the intermediate output is input may correspond to a model created by connecting the head model sent by the electronic device 120 in operation S520 with the decoder model trained by the base station 110.

[0074] Fig. 6A is a flowchart illustrating a specific method of operation for training an autoencoder according to an embodiment, Figure 6B is an example of a signal exchange in a process of training an autoencoder according to an embodiment, and Figure 7 is an example of an automatic encoder according to an embodiment. In one example, the UE side may refer to the electronic device side, and the network side may refer to the base station side.

[0075] Reference Fig. 6A In operation S610, an electronic device (eg, Figure 1 The electronic device 120 in the embodiment can train the encoder model. Figure 6B In order to train the encoder model, the electronic device 120 may provide the channel estimation result H as an input to the encoder model. The encoder model may correspond to model 601 (hereinafter, referred to as encoder model 601). Figure 1The electronic device 120 may provide the channel estimation result H as an input to the neural network for compression, and may train the encoder model 601 by sequentially calculating and storing intermediate variables starting from the input layer of the neural network for compression. The channel estimation result H may pass through the encoder model 601 and be converted into a first intermediate output Z. Figure 6B , the first intermediate output Z may be sent from the electronic device 120 to the base station 110 .

[0076] In operation S620, the base station 110 may train a partial decoder model, which may include a model including the remaining layers except for at least one output layer in the decoder model for performing decompression. Figure 6B , the partial decoder model may correspond to the model 602 (hereinafter, referred to as the partial decoder model 602). At least one output layer may correspond to the head model described below, and may be trained by the electronic device 120. The base station 110 may receive a first intermediate output Z from the electronic device 120. The first intermediate output Z may correspond to the output of the encoder model 601 of the electronic device 120. In order to train the partial decoder model 602, the base station 110 may use the first intermediate output Z as an input of the partial decoder model 602. The base station 110 may provide the first intermediate output Z as an input to the neural network of the partial decoder model 602, and may train the partial decoder model 602 by sequentially calculating and storing intermediate variables starting from the input layer of the partial decoder model 602. The output of the partial decoder model 602 may include a second intermediate output different from the first intermediate output Z, instead of the decompressed channel estimation result. Refer together Figure 6B , the second intermediate output may correspond to the signal Z′ (hereinafter, referred to as the second intermediate output Z′).

[0077] In operation S630, the electronic device 120 may train a head model, which may include at least one output layer of the decoder model. Figure 6B , the head model may correspond to the model 603 (hereinafter, referred to as the head model 603). That is, the head model 603 may be combined with the partial decoder model 602 of the base station 110 to form an overall decoder model for decompressing the first intermediate output Z into which the channel estimation result has been compressed. The electronic device 120 may receive the second intermediate output Z' from the base station 110, and use the second intermediate output Z' as an input to train the head model 603. The head model 603 may output the decompressed channel estimation result. Figure 6B, the decompressed channel estimation result may correspond to the signal H' (hereinafter, referred to as the decompressed channel estimation result H'). That is, the decompressed channel estimation result H' may include data obtained by compressing the channel estimation result estimated by the electronic device 120 and then decompressing the channel estimation result estimated by the electronic device 120, and therefore may not be completely identical to the channel estimation result H. The difference between the channel estimation result H and the decompressed channel estimation result H' may be defined as a loss or error. The electronic device 120 may update the weight parameters of the head model 603, the partial decoder model 602, and the encoder model 601 by error back propagation, and thereby complete the training (or learning) of the autoencoder once (or 1 epoch). The electronic device 120 may send the first gradient value obtained by differentiating the loss function to the base station 110. Refer together Figure 6B , the first gradient value may correspond to the signal G (hereinafter, referred to as the first gradient value G). For example, the first gradient value may be generated based on the back-propagation algorithm. The base station 110 may receive the first gradient value G and update the weight parameters of the partial decoder model 602 to minimize the loss. The base station 110 may sequentially update the weight parameters of the layers of the partial decoder model 602 in the opposite direction. For example, the base station 110 may update the weight parameters in the reverse order from the layer outputting the second intermediate output Z' of the partial decoder model 602 to the layer receiving the first intermediate output Z. The base station 110 may generate a second gradient value and send the second gradient value to the electronic device 120. The second gradient value is used to update the weight parameters of the encoder model 601 of the electronic device 120. Refer to Figure 6B The second gradient value may correspond to the signal G′ (hereinafter, referred to as the second gradient value G′). For example, the second gradient value may be generated based on back propagation of the partial decoder model 602 of the base station 110 .

[0078] The electronic device 120 may receive the second gradient value G' and update the weight parameters of the encoder model 601 to minimize the loss. The electronic device 120 may sequentially update the weight parameters of the layers of the encoder model 601 in the reverse direction. For example, the electronic device 120 may update the weight parameters in the reverse order from the layer outputting the first intermediate output Z of the encoder model 601 to the layer receiving the channel estimation result.

[0079] According to an embodiment, the electronic device 120 may repeat operations S610 to S630 multiple times. For example, the electronic device 120 may repeat operations S610 to S630 until the loss between the channel estimation result H and the decompressed channel estimation result H' is no longer reduced.

[0080] In operation S640, the electronic device 120 may send the head model to the base station 110. When the difference between the channel estimation result H and the decompressed channel estimation result H' no longer decreases or is less than or equal to the threshold, the electronic device 120 may determine that the training of the automatic encoder for compressing and decompressing the channel estimation result has been completed. The electronic device 120 may send the head model 603 to the base station 110 in response to the completion of the training (or after the completion of the training).

[0081] After receiving the header model 603, the base station 110 may generate an overall decoder model 710 by connecting the header model 603 with the partial decoder model 602 previously stored in the base station 110. According to an embodiment, the overall decoder model 710 is not trained in the base station 110, but the header model 603 as a part of the overall decoder model 710 is trained in the electronic device 120. Therefore, the loss as the difference between the channel estimation result H and the decompressed channel estimation result H' may be calculated inside the electronic device 120. That is, if the header model 603 is not placed in the electronic device 120, the channel estimation result H may be provided to the base station 110 to calculate the loss. However, the channel estimation result H may include personal information that can identify the user of the electronic device 120. If other people listen to the channel estimation result H provided to the base station 110, there may be an adverse effect on privacy and information security. Therefore, the header model 603 is placed in the electronic device 120 to improve data privacy.

[0082] Reference Figure 7 , the electronic device 120 may obtain a channel estimation result H by performing channel estimation based on a reference signal for CSI. The electronic device 120 may input the channel estimation result H into the encoder model 601 and generate compressed data of the channel estimation result H. The encoder model 601 may receive an input of the channel estimation result H and output a first intermediate output Z. The first intermediate output Z may correspond to the compressed data. The electronic device 120 may provide the first intermediate output Z to the base station 110. The base station 110 may provide the first intermediate output Z as an input to the overall decoder model 710. The overall decoder model 710 may receive the first intermediate output Z and output a decompressed channel estimation result H'. That is, the electronic device 120 may directly provide the channel estimation result to the base station 110 by CSI feedback based on the auto encoder.

[0083] Figure 8 is a signal exchange diagram of a wireless communication system according to an embodiment. Figure 8 , the base station 110 may transmit a reference signal to the electronic device 120 (operation 810), and the electronic device 120 may perform channel estimation based on the reference signal (operation 815).

[0084] The electronic device 120 may perform training of the first model based on the channel estimation result (operation 820). The first model may be Figure 6B The electronic device 120 may provide the channel estimation result as an input to the first model for compressing the channel estimation result. The electronic device 120 may generate a first intermediate output by sequentially calculating and storing intermediate parameters starting from the input layer of the first model. The electronic device 120 may send the generated first intermediate output to the base station 110 (operation 825). The first intermediate output may correspond to intermediate data for obtaining the channel estimation result when decompression is performed by the second model.

[0085] The base station 110 may receive the first intermediate output and train a second model based on the first intermediate output (operation 830). The second model may be Figure 6B The base station 110 may provide the first intermediate output as an input to the second model for decompressing the compressed channel estimation result. The base station 110 may generate a second intermediate output by sequentially calculating and storing intermediate parameters starting from the input layer of the second model. The base station 110 may send the generated second intermediate output to the electronic device 120 (operation 835). The second intermediate output may correspond to intermediate data of “the decompressed channel estimation result can be obtained when passing through the output layer of the decoder model”.

[0086] The electronic device 120 may perform training of a third model based on the second intermediate output and calculate a loss (operation 840). The third model may be Figure 6B The electronic device 120 may calculate the parameters by inputting the second intermediate output into the layer of the third model, and thus may generate a decompressed channel estimation result. However, if the channel estimation result H is compressed and then decompressed, at least a portion of the data may change. The electronic device 120 may calculate the loss by comparing the initially estimated channel estimation result with the decompressed channel estimation result, and may calculate the first gradient value to minimize the loss. The electronic device 120 may send the first gradient value to the base station 110 (operation 845). The base station 110 may update the second model based on the first gradient value (operation 850). The updating step may correspond to the step of adjusting the weight parameters of the second model according to back propagation, and the updating step may be performed in reverse order from the last layer of the second model to the first layer. The base station 110 may calculate the second gradient value to minimize the loss. The base station 110 may send the second gradient value to the electronic device 120 (operation 855). For example, the first gradient value may be different from the second gradient value.

[0087] The electronic device 120 may update the first model based on the second gradient value (operation 860). The updating step may correspond to the step of adjusting the weight parameters of the first model according to back propagation, and the updating step may be performed in reverse order from the last layer to the first layer of the first model. Therefore, the autoencoder training may be performed once. When the training is performed a specific number of times and the learning is terminated, the electronic device 120 may provide the third model to the base station 110 (operation 865).

[0088] Fig. 9 is another example of an automatic encoder according to an embodiment. Fig. 9 , wireless communication systems (e.g., Figure 1 10) in may also include an electronic device. For example, the wireless communication system may include a first electronic device UE1, a second electronic device UE2, and a base station. The base station may communicate with the first electronic device UE1 and the second electronic device UE2 at the same time.

[0089] According to an embodiment, each of the first electronic device UE1 and the second electronic device UE2 may perform training of the auto encoder in parallel. In other words, the first channel estimation result H 1 To train the encoder model of the first electronic device UE1, and based on the second channel estimation result H 2 To train the encoder model of the second electronic device UE2.

[0090] Based on the first channel estimation result H 1 and the second channel estimation result H 2 The partial decoder model of the base station is trained using both the intermediate output Z1 received from the first electronic device UE1 and the intermediate output Z2 received from the second electronic device UE2. Because the partial decoder model is trained using both the intermediate output Z1 received from the first electronic device UE1 and the intermediate output Z2 received from the second electronic device UE2, the partial decoder model of the base station can be applied together. When the training is completed, the first electronic device UE1 and the second electronic device UE2 can both complete the decoder model by sending the head model to the base station. Therefore, the overall decoder model of the base station may include all of the partial decoder model that can be commonly applied to the first electronic device UE1 and the second electronic device UE2, the first head model for decompressing the intermediate output Z1 of the first electronic device UE1, and the second head model for decompressing the intermediate output Z2 of the second electronic device UE2. The base station may include an overall decoder model that is variably adjustable by only changing the head model for multiple electronic devices while maintaining the partial decoder model. For example, when the intermediate output Z1 is received from the first electronic device UE1, the base station may change the head model to the first head model, and input the intermediate output Z1 to the overall decoder model, thereby decompressing the channel estimation result compressed by the first electronic device UE1, and outputting the decompressed channel estimation result H'. 1For example, when receiving the intermediate output Z2 from the second electronic device UE2, the base station may change the head model to the second head model and input the intermediate output Z2 to the overall decoder model, thereby decompressing the channel estimation result compressed by the second electronic device UE2 and outputting the decompressed channel estimation result H' 2 .

[0091] Fig.10 is another example of an automatic encoder according to an embodiment. Fig.10 , wireless communication systems (e.g., Figure 1 10) may include a first electronic device UE1, a second electronic device UE2, a first base station BS1, and a second base station BS2. Each of the first base station BS1 and the second base station BS2 may communicate with the first electronic device UE1 and the second electronic device UE2 at the same time.

[0092] According to an embodiment, each of the first electronic device UE1 and the second electronic device UE2 may perform training of the auto encoder in parallel. In other words, the first channel estimation result H 1 To train the encoder model of the first electronic device UE1, and based on the second channel estimation result H 2 To train the encoder model of the second electronic device UE2.

[0093] Based on the first channel estimation result H 1 and the second channel estimation result H 2 The partial decoder model of each of the first base station BS1 and the second base station BS2 is trained by both. Because the partial decoder model of the first base station BS1 is trained using both the intermediate output Z1 received from the first electronic device UE1 and the intermediate output Z2 received from the second electronic device UE2, the partial decoder model of the first base station BS1 can be commonly applied to the first electronic device UE1 and the second electronic device UE2.

[0094] Because the partial decoder model of the second base station BS2 is trained using both the intermediate output Z1 received from the first electronic device UE1 and the intermediate output Z2 received from the second electronic device UE2, the partial decoder model of the second base station BS2 can be commonly applied to the first electronic device UE1 and the second electronic device UE2. When the training is completed, the first electronic device UE1 and the second electronic device UE2 can complete the decoder model by sending the head model to the base station respectively. Therefore, the overall decoder model of each of the first base station BS1 and the second base station BS2 may include: all of the partial decoder model that can be commonly applied to the first electronic device UE1 and the second electronic device UE2, the first head model for decompressing the intermediate output Z1 of the first electronic device UE1, and the second head model for decompressing the intermediate output Z2 of the second electronic device UE2. Each of the first base station BS1 and the second base station BS2 may include an overall decoder model that is variably adjustable by only changing the head model for multiple electronic devices while maintaining the partial decoder model.

[0095] For example, when receiving the intermediate output Z1 from the first electronic device UE1, the first base station BS1 and the second base station BS2 may change the head model to the first head model and input the intermediate output Z1 into the overall decoder model, thereby decompressing the channel estimation result compressed by the first electronic device UE1 and outputting the decompressed channel estimation result H' 1 and H' 2 For example, when receiving the intermediate output Z2 from the second electronic device UE2, the first base station BS1 and the second base station BS2 may change the head model to a second head model and input the intermediate output Z2 into the overall decoder model, thereby decompressing the channel estimation result compressed by the second electronic device UE2.

[0096] FIG. 11A to FIG. 11E is an example of an autoencoder of a comparative example. Although the autoencoder according to the embodiment is trained in both the base station and the electronic device, Fig.11A (Comparative Example) The autoencoder may be trained in the base station. In this case, the electronic device may send the channel estimation result to the base station so that the channel estimation result is input to the encoder model. Therefore, there is a disadvantage in terms of data privacy. In addition, after the training is completed in the base station, the encoder model itself may be sent to the electronic device. The transmission of the encoder model itself requires a large amount of data, which may be a burden on the network. For example, the channel estimation result after decompression Can be output by the decoder model as output.

[0097] Although the autoencoder according to the disclosed embodiment is trained in both the base station and the electronic device, Fig. 11BThe autoencoder of (Comparative Example) can be trained in the electronic device. In this case, after the training is completed in the electronic device, the decoder model itself can be sent to the base station. The transmission of the decoder model itself requires a large amount of data, which may be a burden on the network.

[0098] Fig. 11C The autoencoder of (Comparative Example) may be based on segmentation learning similar to that of the embodiment. That is, the encoder model may be trained in the electronic device, and the decoder model may be trained in the base station. In the autoencoder according to the embodiment, the head model is set in the electronic device so that the electronic device can directly calculate the loss as the difference between before and after compression during the training process. However, in Fig. 11C In the automatic encoder of , it can be seen that in order to calculate the difference between before and after compression, a data set including a channel estimation result is sent to a base station. Since the electronic device can send the channel estimation result to the base station, there is a disadvantage in terms of data privacy.

[0099] Fig.11D The autoencoder of (Comparative Example) can be based on segmentation learning similar to that of the embodiment. Fig.11D , the base station may train both the encoder model and the decoder model based on the channel estimation results, and the electronic device may train the encoder model independently of the base station. A shared data set may be required to calculate the loss before and after compression of the channel estimation results. For example, the base station may need to provide the channel estimation results and the intermediate output to the electronic device. The transmission of the data set requires a large amount of data, which may be a burden on the network.

[0100] Fig.11E The autoencoder of (Comparative Example) can be based on segmentation learning similar to that of the embodiment. Fig.11E , the electronic device may train both the encoder model and the decoder model based on the channel estimation result, and the base station may train the decoder model independently of the electronic device. A shared data set may be required to calculate the loss before and after compression of the channel estimation result. For example, the electronic device may need to provide the channel estimation result and the intermediate output to the base station. The transmission of the data set requires a large amount of data, which may be a burden on the network. Fig.11E In , the NW decoder model can represent the network side decoder.

[0101] Fig.12 1 is a block diagram of a wireless communication device 1200 according to an embodiment. Fig.12 , the wireless communication device 1200 may include a modem and a radio frequency integrated circuit (RFIC) 1260. The modem may include an application specific integrated circuit (ASIC) 1210, an application specific instruction set processor (ASIP) 1230, a memory 1250, a main processor 1270, and a main memory 1290. Fig.12 The wireless communication device 1200 may include or correspond to the electronic device 120 or the electronic device 300 according to an embodiment.

[0102] RFIC 1260 may be connected to antenna Ant and receive signals from the outside or send signals to the outside using a wireless communication network. ASIP 1230 may include an integrated circuit customized for a specific purpose, and may support a dedicated instruction set for a specific application and execute instructions of the instruction set. Memory 1250 may communicate with ASIP 1230 and store multiple instructions executed by ASIP 1230 as a non-temporary storage device. For example, as a non-limiting example, memory 1250 may include any type of memory (such as random access memory (RAM), read-only memory (ROM), tape, disk, optical disk, volatile memory, non-volatile memory, and combinations thereof) that can be accessed by ASIP 1030.

[0103] The main processor 1270 may control the wireless communication device 1200 by executing a plurality of instructions. Throughout the disclosure, the main processor 1270 may represent or correspond to at least one control circuit or at least one processor.

[0104] For example, the main processor 1270 may control the ASIC 1210 and the ASIP 1230, and may process data received through a wireless communication network or process user input of the wireless communication device 1200. The main memory 1290 may communicate with the main processor 1270 as a non-transitory storage device and store a plurality of instructions executed by the main processor 1270. For example, as a non-limiting example, the main memory 1290 may include any type of memory (such as RAM, ROM, tape, disk, optical disk, volatile memory, non-volatile memory, and combinations thereof) accessible by the main processor 1270. Throughout the disclosure, the main memory 1290 may represent or correspond to at least one memory or at least one memory device.

[0105] While the disclosure has been particularly shown and described with reference to disclosed embodiments, 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 for operating an electronic device, the method comprising: Obtaining a channel estimation result based on a reference signal received from a base station; Obtaining a first intermediate output by inputting the channel estimation result into a first model including a first neural network for data compression; sending the first intermediate output to a base station; receiving from the base station a second intermediate output different from the first intermediate output; Obtaining a decompressed channel estimation result by inputting the second intermediate output into a third model, the third model corresponding to an output layer of the second model including a second neural network for data decompression; By comparing the obtained channel estimation result with the decompressed channel estimation result, a loss between the obtained channel estimation result and the decompressed channel estimation result is calculated; sending a first gradient value for minimizing the loss to a base station; receiving, from the base station, a second gradient value different from the first gradient value; as well as The first model is updated based on the second gradient value.

2. The method according to claim 1, further comprising: identifying that a difference between the channel estimation result and the decompressed channel estimation result is less than or equal to a threshold; as well as The third model is sent to the base station based on identifying that the difference between the channel estimation result and the decompressed channel estimation result is less than or equal to the threshold.

3. The method according to claim 1, wherein: The base station includes a second model.

4. The method according to claim 3, wherein: The second intermediate output corresponds to the output of the second model of the base station.

5. The method according to claim 3, wherein: The second gradient value is generated based on back propagation of the second model of the base station.

6. The method according to any one of claims 1 to 5, wherein: The encoder corresponding to the first model corresponds to at least one of a convolutional neural network, a recurrent neural network, a stacked-based deep neural network, and a long short-term memory network.

7. The method according to any one of claims 1 to 5, wherein: The decoder obtained by connecting a partial model of the second model except the output layer with the third model corresponds to at least one of a convolutional neural network, a recurrent neural network, a stack-based deep neural network, and a long short-term memory network.

8. A method of operating a base station, the method comprising: sending a reference signal to an electronic device including the first model and the third model; receiving a first intermediate output from an electronic device; Obtaining a second intermediate output by inputting the first intermediate output to a partial model excluding an output layer from a second model, the second model including a second neural network for data decompression; sending the second intermediate output to the electronic device; receiving a first gradient value from an electronic device and updating a weight parameter of the partial model; as well as A second gradient value different from the first gradient value is generated and the second gradient value is transmitted to the electronic device.

9. The method according to claim 8, wherein: The first model includes a first neural network for data compression, and Among them, the third model corresponds to the output layer of the second model.

10. The method according to claim 9, wherein: The first intermediate output corresponds to data output by the first model using a channel estimation result obtained based on the reference signal as input.

11. The method according to claim 9, wherein: The first gradient value is generated based on a back propagation algorithm to minimize a loss calculated by comparing a channel estimation result obtained by the electronic device based on the reference signal with a decompressed channel estimation result output by a third model using the second intermediate output as an input.

12. The method according to claim 11, further comprising: receiving a third model from the electronic device, The third model is sent based on the difference between the identified channel estimation result and the decompressed channel estimation result being less than or equal to a threshold.

13. The method according to any one of claims 8 to 12, wherein: The encoder corresponding to the first model corresponds to at least one of a convolutional neural network, a recurrent neural network, a stacked-based deep neural network, and a long short-term memory network.

14. The method according to any one of claims 8 to 12, wherein: The decoder obtained by connecting the partial model with the third model corresponds to at least one of a convolutional neural network, a recursive neural network, a stacked-based deep neural network, and a long short-term memory network.

15. A method of operating a wireless communication system, the wireless communication system comprising an electronic device and a base station, the electronic device comprising an encoder for data compression, the base station comprising a decoder for data decompression, the method comprising: sending a reference signal to the electronic device via the base station; Obtaining a channel estimation result based on the reference signal by an electronic device; By electronic means, the channel estimation result is input into an encoder and an intermediate output is obtained; sending the obtained intermediate output to a base station via electronic means; as well as Through the base station, the intermediate output is input to the decoder, and through the base station, a decompressed channel estimation result is obtained.

16. The method according to claim 15, further comprising: By means of electronics, the encoder is trained, and By electronic means, the output layer of the decoder is trained.

17. The method according to claim 16, further comprising: Through the base station, the remaining layers of the decoder except the output layer of the decoder are trained.

18. The method according to claim 15, further comprising: Based on the completion of the training of the output layer of the decoder, the output layer of the decoder is transmitted to the base station.

19. The method according to any one of claims 15 to 18, wherein: The encoder corresponds to at least one of a convolutional neural network, a recurrent neural network, a stacked-based deep neural network, and a long short-term memory network.

20. The method according to any one of claims 15 to 18, wherein: The decoder corresponds to at least one of a convolutional neural network, a recurrent neural network, a stack-based deep neural network, and a long short-term memory network.

Images