KR20240028749A - Receiver and Receiving Method for Canceling Radio Communication Channel Noise based on Auto Encoder - Google Patents

Abstract

The disclosed embodiment is a noise removal module that obtains a noise removal symbol from which the channel noise of the wireless channel is removed by performing a neural network operation on the received signal received through a wireless channel, a transmission symbol transmitted from a transmitter, and demodulates the noise removal symbol. Provides a receiver and reception method that can perform highly reliable wireless communication and reduce learning time, including a demodulator that acquires data and a decoder that receives the demodulated data and decodes it through neural network operation to obtain recovery data. do.

Description

{Receiver and Receiving Method for Canceling Radio Communication Channel Noise based on Auto Encoder}

The disclosed embodiments relate to a receiver and a reception method, and to an autoencoder-based receiver and reception method.

Generally, when data is transmitted and received between a transmitter and a receiver in a communication system, data errors may occur due to noise present in the communication channel. As coding methods designed to enable the receiver to correct errors generated by the communication channel, error detection codes and error correcting codes (ECC) methods exist. In particular, error correction codes used for communication between transmitters and receivers may be generally referred to as channel coding. In the error correction code method, the transmitter transmits data by adding a redundant bit to the data bit to be transmitted, and the receiver uses the redundant bit to perform a decoding operation to correct errors included in the data bit. Perform.

In communication and broadcasting systems, various channel codings such as convolutional coding, turbo coding, low-density parity-check coding (LDPC coding), and polar coding. methods are being used. Accordingly, decoding techniques such as belief propagation (BP) and the minimum sum algorithm are mainly used on the receiver side. However, since the existing error correction code decoding technique is performed based on a very complex algorithm, research on decoding techniques using artificial neural networks has been actively conducted in recent years to provide more efficient and more reliable communication services.

Korean Patent Publication No. 1020200127783 (published on November 11, 2020)

The purpose of the disclosed embodiments is to provide a receiver and a reception method that can efficiently recover a received signal with high reliability by simultaneously reducing decoding loss and noise removal loss.

The disclosed embodiments include a receiver that includes a decoder that performs decoding based on an artificial neural network and an auto-encoder that performs channel noise removal, thereby improving communication performance by allowing channel noise removal and decoding to be performed separately; and The purpose is to provide a reception method.

The disclosed embodiments are a receiver and reception method that not only simplifies the structure of the auto encoder by reducing the number of channels in each layer and connecting the layers using skip connections, but also effectively removes channel noise even with a short number of learning times. The purpose is to provide.

A receiver according to an embodiment includes a noise removal module that performs a neural network operation on a received signal in which a transmission symbol transmitted from a transmitter is received through a wireless channel to obtain a noise removal symbol from which channel noise of the wireless channel is removed; and a decoder that receives the noise removal symbol and demodulates and decodes it through neural network operations to obtain recovered data.

The noise removal module may be implemented as an auto-encoder that includes at least one hidden layer, extracts features of the received signal, reduces the dimensionality, compresses it, and expands the compressed features again to obtain the noise removal symbol. .

The auto encoder may form at least one skip connection that transfers the extracted features by skipping one or more of the at least one hidden layer.

The noise removal module may be trained in advance according to the noise removal loss calculated as the difference between the transmitted symbol and the received signal.

The decoder is implemented as a Hyper-Graph-Network (HGN) decoder and may be trained in advance according to the decoding loss calculated as the difference between the transmitted data and the recovered data.

The noise removal module and the decoder invert the multi-task loss obtained by weighting the noise removal loss calculated as the difference between the transmitted symbol and the received signal and the decoding loss calculated as the difference between the transmitted data and the recovered data with a specified weight. It can be learned by spreading.

A reception method according to an embodiment includes the steps of performing a neural network operation on a received signal received through a wireless channel using a transmission symbol transmitted from a transmitter using an artificial neural network to obtain a noise removal symbol from which channel noise of the wireless channel has been removed; And obtaining recovery data by receiving the noise removal symbol and demodulating and decoding it through neural network operations using an artificial neural network.

Therefore, the receiver and reception method according to the embodiment not only implements a decoder that performs decoding in the receiver using an artificial neural network, but also includes a noise removal module implemented as an auto-encoder to remove channel noise, thereby removing and decoding channel noise. Communication performance can be improved by being performed separately in different artificial neural networks. In addition, by training the decoder and noise removal module together so that the decoding loss and noise removal loss are simultaneously reduced, the increase in learning time due to the addition of the noise removal module can be suppressed. In addition, the noise removal module implemented as an auto encoder reduces the number of channels in each layer and connects layers using skip connections, which not only simplifies the structure of the auto encoder but also effectively removes channel noise even with a short number of learning times. make it possible

Figure 1 shows a schematic configuration of a wireless communication system.
Figure 2 shows a schematic configuration of a wireless communication system according to an embodiment.
Figures 3 and 4 show an example of the detailed configuration of the noise removal module of Figure 2.
FIG. 5 shows an example of the detailed configuration of the decryption module of FIG. 2.
Figure 6 is a diagram for explaining a reception method of a receiver in a wireless communication system according to an embodiment.

Hereinafter, specific embodiments of one embodiment will be described with reference to the drawings. The detailed description below is provided to provide a comprehensive understanding of the methods, devices and/or systems described herein. However, this is only an example and the present invention is not limited thereto.

In describing one embodiment, if it is determined that a detailed description of the known technology related to the present invention may unnecessarily obscure the gist of an embodiment, the detailed description will be omitted. In addition, the terms described below are terms defined in consideration of functions in the present invention, and may vary depending on the intention or custom of the user or operator. Therefore, the definition should be made based on the contents throughout this specification. The terminology used in the detailed description is intended to describe only one embodiment and should in no way be limiting. Unless explicitly stated otherwise, singular forms include plural meanings. In this description, expressions such as “comprising” or “comprising” are intended to indicate certain features, numbers, steps, operations, elements, parts or combinations thereof, and one or more than those described. It should not be construed to exclude the existence or possibility of any other characteristic, number, step, operation, element, or part or combination thereof. In addition, terms such as "... unit", "... unit", "module", and "block" used in the specification refer to a unit that processes at least one function or operation, which is hardware, software, or hardware. and software.

Figure 1 shows a schematic configuration of a wireless communication system.

Referring to FIG. 1, the wireless communication system includes a transmitter 10 that transmits data and a receiver 30 that receives data transmitted from the transmitter 10, and the data transmitted from the transmitter 10 is transmitted through a wireless channel ( It is transmitted to the receiver 30 through 20). Here, the transmitter 10 and the receiver 30 are each included in a base station or user terminal, so that the base station or user terminal can transmit and receive data.

Transmitter 10 may include a channel encoder 11 and a modulator 12. The channel encoder 11 generates a codeword (v ^N ) by encoding the authorized transmission data (u ^k ) to correct data errors due to noise in the wireless channel 20. The channel encoder 11 can encode transmission data u ^k according to various channel coding methods, such as low-density parity check coding and polar coding. The modulator 12 modulates the codeword (v ^N ) generated by the channel encoder 11 to obtain a transmission symbol (x ^N ), and transmits the obtained transmission symbol (x ^N ) to the receiver ( 30).

While the transmission symbol (x ^N ) transmitted from the transmitter 10 is transmitted to the receiver 30 through the wireless channel 20, AWGN (Additive White Gaussian Noise), etc. are added. That is, the wireless channel 20 can be viewed as an AWGN channel.

The receiver 30 receives the received signal (y ^N ) to which the noise of the wireless channel 20 is added to the transmission symbol (x ^N ) transmitted from the transmitter 10 and transmits the transmission data (u ^k ) recovered data ( ) to obtain.

The receiver 30 may include a decoder 31, and the decoder 31 may recover data (y N) from the received signal (y ^N ). ) to obtain. The existing receiver 30 further includes a demodulator (not shown) as well as a decoder, and after demodulating the received signal (y ^N ) into demodulated data, the decoder uses a decoding technique according to the encoding technique of the channel encoder 11. By decoding the demodulated data, recovery data ( ) was configured to obtain. However, as mentioned above, recently, the decoder 31 has been implemented using an artificial neural network, and the decoder 31 implemented with an artificial neural network performs demodulation and decoding by performing a neural network operation on the received signal (y ^N ) according to a learned method. Recovery data performed together ( ) is being studied to obtain. In this case, the decoder 31 implemented with an artificial neural network needs to be trained in advance in order to perform a neural network operation on the received signal (y ^N ).

Accordingly, the wireless communication system including the decoder 31 implemented with an artificial neural network further includes a decoding loss calculation module 41 that is used only during learning, so that the receiver 30 can learn the decoder 31. The decoding loss calculation module 41 includes transmission data (u ^k ) input to the channel encoder 11 of the transmitter 10 and recovery data output from the decoder 31 of the receiver 30 ( ) is authorized, and the authorized transmission data (u ^k ) and recovery data ( ) is calculated as a decoding loss and back-propagated to the decoder 31, thereby training the decoder 31 implemented with an artificial neural network.

However, in this way, the transmission data (u ^k ) and recovery data ( ) When training the decoder 31 by calculating the difference between the decoding loss as the decoding loss, the decoder 31 not only determines the difference due to the decoding of the encoding of the channel encoder 11, but also the difference due to the channel noise of the wireless channel 20. Learning is carried out by reflecting together. Therefore, reliable recovery data ( ) is not only difficult to acquire, but also takes a very long time to learn, which has the limitation of being inefficient.

FIG. 2 shows a schematic configuration of a wireless communication system according to an embodiment, FIGS. 3 and 4 show an example of the detailed configuration of the noise removal module of FIG. 2, and FIG. 5 shows a detailed configuration of the decoding module of FIG. 2. Shows an example.

A wireless communication system other than the embodiment shown in FIG. 2 also includes a transmitter 10 and a receiver 50 similar to that in FIG. 1 . However, in the case of the transmitter 10, it is configured to include a channel encoder 11 and a modulator 12 in the same manner as in Figure 1, and encodes the applied transmission data (u ^k ) to generate a codeword (v ^N ), The generated codeword (v ^N ) is modulated to generate a transmission symbol (x ^N ) and transmitted to the receiver 50 through the wireless channel 20.

Meanwhile, in the embodiment, the receiver 50 includes a decoder 51 like the receiver 30 of FIG. 1, but unlike the receiver 30 of FIG. 1, it further includes a noise removal module 53. Here, the noise removal module 53 is added while the transmission symbol (x ^N ) is transmitted through the wireless channel 20 and removes noise included in the received signal (y ^N ). That is, the receiver 50 of the embodiment first removes the noise added to the received signal (y ^N ) by the wireless channel 20 before performing modulation on the received signal (y ^N ). ) plays the role of transmitting to the decoder 51.

In an embodiment, the noise removal module 53 may be implemented as an auto encoder, a type of artificial neural network. In the auto encoder, the input layer receives the original transmission data (Original Input) (x) and extracts features, at least one hidden layer compresses the extracted features to reduce their dimension, and the output layer extracts the compressed features. It is an artificial neural network in which the neural network calculation method is learned to recover the original transmitted data (x) by expanding the dimension of the feature. Here, as a simple example for convenience of explanation, the auto encoder is shown as having only one hidden layer, but in general, the auto encoder has multiple hidden layers. As shown in FIG. 3, the autoencoder learned in this way produces damaged or distorted transmission data (Corrupt Input) ( ) is input, the transmission data ( ), extracting and compressing the features, and then restoring the compressed features, allowing the undamaged original transmission data (x) to be recovered. Autoencoder is well known as an artificial neural network that shows excellent performance in noise removal fields, such as removing noise included in image data or extracting voice signals.

Accordingly, the receiver 50 of the embodiment also implements the noise removal module 53 as an auto encoder to remove channel noise of the wireless channel 20. However, since the autoencoder is also a type of artificial neural network, it needs to be trained in advance. In this embodiment, the decoder 51 and the noise removal module 53 implemented with an artificial neural network are separately provided, so the noise removal module 53 learns to remove only the channel noise of the wireless channel 20 regardless of decoding. It can be. That is, unlike the decoder 51 in FIG. 1, there is no need to perform decoding and channel noise removal simultaneously. Therefore, the autoencoder implemented with the noise removal module 53 can exhibit the required performance even if it is implemented with a simple structure. As an example, the noise removal module 53 may be implemented with a simple structure including a relatively large filter and multiple layers (Layer1 to Layer3) with a small number of channels.

In Figure 4, as an example, it is assumed that the noise removal module 53 includes one input layer and three hidden layers (Layer1 to Layer3). The input layer that receives the received signal (y ^N ) is configured to output 15 channels by performing a one-dimensional convolution (1-D conv.) operation with a filter of size (5 Among Layer1 to Layer3), the first hidden layer (Layer2) corresponding to the encoding process is composed of a filter of size (5 It can be configured to output channels. And the second hidden layer (Layer2) corresponding to the decoding process performs a 1-D transpose convolution (1-D trans. conv.) operation with a filter of size (5 It is expanded to 15 channels again, and the third hidden layer (Layer3) performs a one-dimensional transpose convolution operation with a filter ^of size (5 Convert it to a one-dimensional vector form with the corresponding size and create a noise removal symbol ( ) is output. Here, each layer can use the tanh function as the activation function.

In the embodiment, each layer of the noise removal module 53 implemented as an auto encoder commonly uses a filter with a relatively large size (here, (5 quickly remove noise symbol ( ) can be obtained.

In addition, in the embodiment, the noise removal module 53 determines a relatively accurate feature location in the transmission symbol (x ^N ) despite the noise included in the reception signal (y ^N ), and selects a noise removal symbol similar to the transmission symbol (x ^N ). ( ), at least one skip connection that transfers features by skipping layers may be further formed to obtain. Here, as an example, a first skip connection that transmits the output of the first hidden layer (Layer1) to the output of the second layer (Layer1) and a second skip connection that transmits the output of the input layer to the output of the third layer (Layer3) This can be formed.

Meanwhile, the decoder 51 uses a noise removal symbol ( ) is approved in the form of LLR (Log-likelihood Ratio) and demodulated and decoded to recover data ( ) to obtain. The decoder 51 may also be implemented with an artificial neural network, similar to the decoder 31 in FIG. 1 . The decoder 51 may be implemented with various artificial neural networks, but in the embodiment, it is assumed to be implemented with a Hyper-Graph-Network (HGN) decoder.

As shown in (a) of Figure 5, the HGN decoder is composed of a plurality of variable nodes (v) and a plurality of check nodes (c), so that the message is transmitted to a plurality of variable nodes and a plurality of check nodes (c). It is a decoder composed of a Trellis graph format in which multiple check nodes are expanded from the Tanner graph, which checks parity by repeatedly passing between check nodes. As shown in (b) of Figure 5, in the HGN decoder, each check node is expanded to be connected to one variable node, and the variable node is replaced with a g network (g) to form a fully connected network. Here, a plurality of check nodes and the g network (g) are alternately and repeatedly arranged in a specified number to perform a recurrent neural network operation, and an additional network (f) is included to learn the weights of the g network (g).

Since the HGN network is a known technology, detailed description is omitted here.

As shown in FIG. 2, in the receiver 50 in the embodiment, not only the decoder 51 but also the added noise removal module 53 is implemented with an artificial neural network, so both the decoder 51 and the noise removal module 53 are implemented as an artificial neural network. It must be learned in advance. Accordingly, the wireless communication system includes a decoding loss calculation module 61 for training the decoder 51 during training, and a denoising loss calculation module for calculating denoising loss for training the noise removal module 53. (62) is further included. Here, the noise rejection loss (L _Denoise ) is calculated from the transmission symbol (x ^N ) and the noise rejection symbol ( ) can be calculated as the difference between

However, if the decoder 51 and the noise removal module 53 are individually trained, the learning time is long. Accordingly, in the embodiment, the multi-task loss calculation module 63 is further included to determine the decoding loss (L _Decode ) calculated in the decoding loss calculation module 61 and the noise removal loss (L _Denoise ) calculated in the noise removal loss calculation module 62. ), the decoder 51 and the noise removal module 53 can be trained together by calculating the multi-task loss (L) and back-propagating the calculated multi-task loss (L).

Here, multi-task loss (L) can be calculated according to Equation 1.

(Here w ₁ and w ₂ represent loss weights.)

The decoding loss calculation module 61, the noise removal loss calculation module 62, and the multi-task loss calculation module 63 can be viewed as components constituting a learning module for training the decoder 51 and the noise removal module 53. You can. Accordingly, when the learning of the decoder 51 and the noise removal module 53 is completed based on the multi-task loss (L), the loss calculation module 61, the noise removal loss calculation module 62, and the multi-task loss calculation module 63 ) can be omitted.

As a result, the receiver 50 of the wireless communication system according to the embodiment not only has the decoder 51 implemented with an artificial neural network, but also has an additional noise removal module 53 implemented with an auto encoder, so that the received signal (y ^N ), first remove noise and create a noise removal symbol ( ), demodulated and decrypted to recover data ( ), so recovery data ( ), as well as recovery data ( ) can be obtained.

In the illustrated embodiment, each component may have different functions and capabilities in addition to those described below, and may include additional components other than those described below. Additionally, in one embodiment, each component may be implemented using one or more physically separate devices, one or more processors, or a combination of one or more processors and software, and, unlike the example shown, may be implemented in specific operations. It may not be clearly distinguished.

The transmitter and receiver shown in FIG. 2 may be implemented in a logic circuit using hardware, firmware, software, or a combination thereof, and may also be implemented using a general-purpose or special-purpose computer. The device may be implemented using hardwired devices, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), etc. Additionally, the device may be implemented as a System on Chip (SoC) including one or more processors and a controller.

In addition, the transmitter and receiver hardware elements may be mounted on a computing device or server in the form of software, hardware, or a combination thereof. A computing device or server includes all or part of a communication device such as a communication modem for communicating with various devices or a wired or wireless communication network, a memory for storing data to execute a program, and a microprocessor for executing a program to perform calculations and commands. It can refer to a variety of devices, including:

Figure 6 is a diagram for explaining a reception method of a receiver in a wireless communication system according to an embodiment.

In an embodiment, the receiver of the wireless communication system further includes a decoder 51 as well as a noise removal module 53, as shown in FIG. 2, the decoder 51 is implemented as an HGN decoder, and the noise removal module 53 ) can be implemented as an autoencoder. Accordingly, the decoder 51 and noise removal module 53, which are respectively implemented as an HGN decoder and an auto-encoder, which are a type of artificial neural network, need to be trained in advance to perform the required neural network calculation. Accordingly, in the reception method of the receiver of the wireless communication system of FIG. 6, there is a learning step 70 for performing learning on the decoder 51 and the noise removal module 53, and recovery data ( ) will be described separately in the reception step (80) of obtaining.

In the learning step 70, learning data is first applied to the transmitter 10 as transmission data (u ^k ), and the transmitter 10 encodes and modulates the applied learning data to transmit a transmission symbol (x ^N ) through a wireless channel ( It is transmitted to the receiver 50 through 20 (71). The receiver 50 receives the received signal (y ^N ) transmitted through the wireless channel 20 and removes noise to create a noise removal symbol ( ), and obtain the noise removal symbol ( ) is demodulated and decoded to restore the learning data obtained as the received data, thereby restoring the recovery data ( ) obtain (72). And the noise removal loss (L _Denoise ) for learning the noise removal module 53 is divided into the transmission symbol (x ^N ) and the noise removal symbol ( ) is calculated as the difference between (73). In addition, transmission data (u ^k ) and recovery data ( ) Calculate the decoding loss (L _Decode ) by the difference between (74). And by calculating the multi-task loss (L) according to Equation 1 based on the calculated noise removal loss (L _Denoise ) and decoding loss (L _Decode ) and back-propagating it, the decoder 51 and the noise removal module 53 Learn (75). Then, it is determined whether learning is complete (76). Learning may be terminated when learning to backpropagate the multi-task loss (L) is repeated more than a specified number of times or when the calculated multi-task loss (L) becomes less than or equal to the specified reference loss.

In the reception step (80) where learning is completed and the wireless communication system performs communication, the receiver 50 determines whether the received signal (y ^N ) is received (81). When the received signal (y ^N ) is received, the learned noise removal module 53 performs a neural network operation on the received signal (y ^N ) to remove channel noise of the wireless channel 20 included in the received signal (y ^N ). Noise reduction symbol ( ) obtain (82). And the noise removal symbol ( ) is demodulated and decoded by neural network operation according to the method learned for the transmission data input to the transmitter 10 (recovery data ( ) obtain (83).

In FIG. 6, it is described that each process is executed sequentially, but this is only an illustrative explanation, and those skilled in the art can change the order shown in FIG. 6 and execute it without departing from the essential characteristics of the embodiments of the present invention. Alternatively, it can be applied through various modifications and modifications by executing one or more processes in parallel or adding other processes.

Although the present invention has been described in detail through representative embodiments above, those skilled in the art will understand that various modifications and other equivalent embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the attached claims.

10: Transmitter 11: Channel encoder
12: modulator 20: wireless channel
50: Receiver 51: Demodulator
52: Decoder 53: Noise removal module
61: Decoding loss calculation module 62: Noise removal loss calculation module
63: Multi-task loss calculation module

Claims (10)

In the receiver of a wireless communication system,
a noise removal module that performs a neural network operation on the received signal received through a wireless channel to obtain a noise removal symbol from which the channel noise of the wireless channel is removed; and
A receiver including a decoder that receives the noise removal symbol and demodulates and decodes it through neural network operation to obtain recovered data. The method of claim 1, wherein the noise removal module
A receiver implemented as an auto-encoder that includes at least one hidden layer, extracts the features of the received signal, reduces the dimensionality, compresses it, and expands the compressed features again to the dimensionality to obtain the noise removal symbol. The method of claim 2, wherein the auto encoder
A receiver in which at least one skip connection is formed to deliver features extracted by skipping one or more of the at least one hidden layer. The method of claim 1, wherein the noise removal module
A receiver that is trained in advance according to noise removal loss calculated as the difference between the transmitted symbol and the received signal. The method of claim 1, wherein the decoder
A receiver implemented as a Hyper-Graph-Network (HGN) decoder and trained in advance according to the decoding loss calculated as the difference between the transmitted data and the recovered data. The method of claim 1, wherein the noise removal module and the decoder
A receiver that is learned by backpropagating a multi-task loss obtained by weighting a noise removal loss calculated as a difference between the transmitted symbol and the received signal and a decoding loss calculated as a difference between the transmitted data and the recovered data with a specified weight. In a receiving method of a receiver of a wireless communication system,
Obtaining a noise removal symbol from which the channel noise of the wireless channel is removed by performing a neural network operation using an artificial neural network to calculate the received signal from the transmitter through a wireless channel; and
A reception method comprising obtaining recovery data by receiving the noise removal symbol and demodulating and decoding it through neural network operations using an artificial neural network. The method of claim 7, wherein the step of obtaining the noise removal symbol is
A reception method performed using an auto-encoder that includes at least one hidden layer, extracts features of the received signal, reduces and compresses the dimensions, and expands the compressed features again to obtain the noise removal symbol. The method of claim 8, wherein the auto encoder
A receiving method in which at least one skip connection is formed to deliver features extracted by skipping one or more of the at least one hidden layer. The method of claim 7, wherein the receiving method is
Before receiving the received signal, further comprising a learning step for training an artificial neural network for obtaining the noise removal symbol and an artificial neural network for obtaining the recovery data,
The learning step is
Calculate noise cancellation loss as the difference between the transmitted symbol and the received signal,
Calculate decoding loss as a difference between the transmitted data and the recovered data,
A reception method that performs learning by backpropagating a multi-task loss obtained by weighting the noise removal loss and the decoding loss with a specified weight.