KR102034955B1 - Method and apparatus for controlling transmit power in wireless communication system based on neural network - Google Patents

Abstract

An apparatus for controlling the transmission power of a transmitting terminal in a wireless communication system (WCS) is disclosed. The neural network based transmission power control apparatus according to an embodiment includes a control unit for controlling the transmission power based on a neural network and channel state information for learning an optimal transmission power by inputting a channel gain matrix.

Description

TECHNICAL AND APPARATUS FOR CONTROLLING TRANSMIT POWER IN WIRELESS COMMUNICATION SYSTEM BASED ON NEURAL NETWORK

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention generally relates to wireless communication systems, and more particularly, to a method and apparatus for controlling transmission power of a transmitting terminal in a wireless communication system.

In wireless communication systems (WCS), proper management and allocation of resources such as transmit power is important. If the transmission power of the user is not properly controlled, smooth communication may be interrupted due to frequency interference between users.

As the popularity of wireless communication increases, the population density of users of the wireless communication system increases, and accordingly, proper allocation of radio resources becomes increasingly important. Accordingly, resource allocation in a wireless communication system, that is, control of transmission power, has been extensively studied for various types of wireless communication systems (for example, D2D communication).

In general, the transmit power is not determined to be closed, but is optimized through an iterative algorithm (eg, iterative water filling or weighted least mean square error (WMMSE), etc.). This type of technique requires several iterations before convergence occurs. At this time, since the more iterations are required as the number of users increases, the calculation speed becomes very slow in an environment where there are many users.

Recently, the use of deep neural network (DNN) -based deep learning technology has been considered in many fields, and in particular, it has been shown to provide higher performance than conventional methods in the field of image analysis. Rather than having to derive complex mathematical models, simple backpropagation algorithms can solve difficult nonlinear problems. A fully trained deep neural network model can be used to suit real-time operation due to the short computation time.

Since the transmit power was generally determined using an algorithm involving a large number of iterations, the long computation time made the real time operation of the transmit power control extremely difficult. According to the present invention, a neural network-based transmission power control method is proposed which enables real-time control of transmission power for a large number of user terminals with a calculation time that is significantly faster than a transmission power determination method based on a weighted minimum mean square error (WMMSE). I would like to.

According to an embodiment, in a wireless communication system (WCS), a neural network based transmission power control apparatus controls transmission power based on neural networks and channel state information for learning optimal transmission power by inputting a channel gain matrix. It may include a control unit.

According to an embodiment, the neural network may include a deep neural network including a fully connected layer.

According to another embodiment, the neural network may include a convolutional neural network including a convolutional part, a fully connected part, and a sigmoid part.

In this case, the neural network may pre-learn a transmission power allocation scheme based on an iterative algorithm.

According to an embodiment, the neural network may use local channel state information from adjacent users as an input value and process channel state information of the remaining users as an average channel value.

According to an embodiment, the wireless communication system (WCS) is underlay D2D communication (Underlaid D2D Communication), the transmission power control device is a transmission power of the D2D user equipment (DUE) between devices The loss function of the neural network may be a parameter of a transmission rate and interference of the terminal.

According to an embodiment, the neural network may learn by considering interference generated in the transmission power of a cellular user equipment (CUE).

According to an embodiment, the neural network includes a convolution neural network including a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part. Network), and input data may be reconstructed in the first fully connected part.

According to an embodiment of the present disclosure, the second fully connected part may learn by considering both an interference between a cellular user terminal (CUE) and a user terminal between devices.

According to an embodiment, a neural network-based transmit power control method performed by a transmit power control apparatus of a wireless communication system (WCS) provides optimal transmission of a channel gain matrix through a neural network. The method may include learning power, receiving channel state information from a transmitting terminal, and controlling transmission power based on the channel state information.

According to an embodiment, the neural network may be a deep neural network including a fully connected layer.

According to another embodiment, the neural network may be a convolutional neural network including a convolutional part, a fully connected part, and a sigmoid part.

According to an embodiment, the learning of the optimal transmission power may include pre-learning a transmission power allocation scheme based on an iterative algorithm.

According to an embodiment, the learning of the optimal transmission power may use local channel state information from adjacent users as an input value and process channel state information of the remaining users as an average channel value.

According to an embodiment, the wireless communication system (WCS) is underlay D2D communication (Underlaid D2D Communication), the transmission power control device is a transmission power of the D2D user equipment (DUE) between devices In this case, the neural network may set a weighted sum-rate (WSR) of the inter-device communication terminal as a loss function.

According to an embodiment, the neural network may learn by considering interference generated in the transmission power of a cellular user equipment (CUE).

According to an embodiment, the neural network includes a convolution neural network including a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part. Network), and input data may be reconstructed in the first fully connected part.

According to an embodiment, the second fully-connected part may learn by considering both interference of a cellular user equipment (CUE) and a user terminal between devices.

In the neural network based transmission power control method according to an embodiment, unlike the conventional transmission power control method in which a complicated optimization problem has to be repeatedly solved by autonomously learning the transmission power of a user terminal through a neural network, the transmission power of a transmission terminal This can be determined quickly.

1 is a diagram illustrating a configuration of a conventional general wireless communication system.
2 is a diagram for explaining communication between underlay devices.
3 is a block diagram illustrating a neural network based transmission power control apparatus according to an embodiment.
4 is a diagram illustrating a deep neural network in a neural network based transmission power control method according to an embodiment.
5 is a diagram illustrating a composite product neural network in a neural network based transmission power control method according to an exemplary embodiment.
6 is a diagram illustrating a composite product neural network including a pre-learning initialization process in a neural network based transmission power control method according to an embodiment.
7 is a flowchart illustrating a neural network based transmission power control method according to an embodiment.

Specific structural or functional descriptions of the embodiments according to the inventive concept disclosed herein are merely illustrated for the purpose of describing the embodiments according to the inventive concept, and the embodiments according to the inventive concept. These may be embodied in various forms and are not limited to the embodiments described herein.

Embodiments according to the inventive concept may be variously modified and have various forms, so embodiments are illustrated in the drawings and described in detail herein. However, this is not intended to limit the embodiments in accordance with the concept of the present invention to specific embodiments, it includes all changes, equivalents, or substitutes included in the spirit and scope of the present invention.

Terms such as first or second may be used to describe various components, but the components should not be limited by the terms. The terms are only for the purpose of distinguishing one component from another component, for example, without departing from the scope of the rights according to the inventive concept, the first component may be called a second component, Similarly, the second component may also be referred to as the first component.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art, and are not construed in ideal or excessively formal meanings unless expressly defined herein. Do not.

Hereinafter, exemplary embodiments will be described in detail with reference to the accompanying drawings.

1 is a diagram illustrating a configuration of a conventional general wireless communication system.

Referring to FIG. 1, in particular, the cellular communication network 100 of a conventional wireless communication system includes a base station 110 and 120 and user terminals 111, 112, 113, 121, 122, and 123 connected to the base station. Can be. As such, in a wireless communication network, communication between user terminals is via a communication network, with corresponding " serving " wireless transceivers (e.g., 3GPP 'Long Term Evolution (LTE)') of the communication network. Two user terminals connected to the same or different evolved Node B (eNodeB wireless transceivers) in an LTE Advanced (LTE Advanced) system are configured and terminated between the wireless transceivers and the user terminal. The communication channel (s) may be in communication with each other.

FIG. 2 is a diagram for explaining communication between underlay devices.

Referring to FIG. 2, it is possible to identify a recently emerging Underlaid Device 2 Device communication technique. As an alternative to traditional two-hop communication, it has recently emerged in the sense that direct communication between user terminals within relatively short distances will be efficient.

The Underlaid Device 2 Device communication system 200 communicates via the base stations 210 and 220 when the distance between the user terminals is far, and when the distance between the user terminals is close (212 and 213 or 223). And 224, the user terminals can form channels 201 and 202 to communicate with each other.

Such direct wireless communication between user terminals is based on D2D communication links established directly by two (or more) user terminals. D2D communication differs from traditional INFRA communication in that information is exchanged over physical communication channels that are directly established and terminated between user terminals rather than through a communication network.

Device-to-device (D2D) communication has recently been recognized as a core technology of cellular systems as a technique to solve the problem of increasing mobile traffic by distributing cellular data traffic to direct communication between users. In a D2D communication, a large number of users can transmit data at the same time, so adjusting the transmission power appropriately is very important in improving performance. For example, if the transmission power of the D2D user equipment (DUE) is not properly controlled, a great interference may occur between the device-to-device user terminals and performance may be degraded. Transmission power control becomes more important in underlay D2D communication in which user terminals between devices share radio resources with cellular user equipment.

3 is a block diagram illustrating a neural network based transmission power control apparatus according to an embodiment.

Referring to FIG. 3, the neural network based transmission power control apparatus 310 may include a neural network 311, a controller 312, and a communication unit 313. According to an embodiment, the neural network 311 of the neural network-based transmit power control apparatus 310 may learn an optimal transmit power by inputting a channel gain matrix.

According to an embodiment, the neural network may be a deep neural network including a fully connected layer. A detailed description of the deep neural network including a fully connected layer will be given in FIG. 4.

According to another embodiment, the neural network may be a convolutional neural network including a convolutional part, a fully connected part, and a sigmoid part. A detailed description of a convolutional neural network including a convolutional part, a fully connected part, and a sigmoid part will be provided with reference to FIG. 5.

According to another embodiment, a convolutional neural network including a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part This can be A detailed description of a convolutional neural network including a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part is shown in FIG. 6. Do it.

Here, the channel gain matrix may include channel gain between inter-device user equipments (DUEs) and channel gain between inter-device user equipments (DUEs) and cellular user terminals (CUEs).

According to an embodiment, the controller 312 may control the transmission power of the user terminal based on the channel state information. In this case, the controller 312 may transmit the determined transmission power of the user terminal to the corresponding user terminal through the communication unit 313.

According to an embodiment, the communication unit 313 may receive data related to the transmission power from the user terminals 321, 322,..., 32n.

According to another embodiment, according to one embodiment, the communication unit 313 may transmit the determined transmission power to the user terminal (321, 322, ..., 32n).

4 is a diagram illustrating a deep neural network in a neural network-based transmission power control method according to an embodiment.

Referring to FIG. 4, a deep neural network of a neural network-based transmission power control apparatus learns an input matrix 410 from a fully connected part and outputs an output value through a hyperbolic tangent part 430. Can be obtained.

According to one embodiment, the input matrix 410 is a wireless channel of the UE-DUE channel and the device-to-device UE that the output of the cellular user terminal (CUE) to the communication of the device-to-device UE (DUE) It may include a channel (DUE-DUE channel). According to an embodiment, the fully connected part 420 may be a plurality of fully connected parts including the first lower fully connected part 421 and the N th lower fully connected part 422. According to one embodiment, each lower fully connected part may include a fully connected layer 423 (FC layer) and a rectified linear unit (ReLU) layer. According to one embodiment, in underlaid D2D communication, the transmission power of the device-to-device user terminal (DUE) is calculated using an in-depth learning technique to maximize the weighted sum rate (WSR). Can be done. That is, a deep neural network may be used for power control for underlay D2D communication.

According to an embodiment, the neural network of the neural network-based transmission power control apparatus may learn by considering interference generated from the transmission power of a cellular user equipment (CUE).

와

는 송신기와 수신기의 집합을 각각 나타내고, 여기서

의 인덱스 0은 셀룰러 사용자 단말(CUE) 송신기에 해당한다. 송신기 i의 송신 전력은 Pi로 표시되고, Pi는

의 범위에 있다. 셀룰러 사용자 단말(CUE)는

과 같은 고정 송신 전력

을 사용한다고 가정한다. 또한,

및 W는 각각 잡음 스펙트럼 밀도 및 대역폭이라고 가정한다.According to an embodiment, a cellular network environment in which a user terminal (DUE) and a cellular user terminal (CUE) are freely distributed among a plurality of devices in an arbitrary area S may be considered. According to an embodiment, communication between underlay devices may be considered in which inter-device user terminals (DUEs) share the same radio resources as uplink transmissions of a cellular user terminal (CUE). In this case, it may be assumed that there are N device-to-device (DUE) transceiver pairs, and the pair may simultaneously transmit data through the same frequency as the cellular user terminal (CUE) that transmits data to the base station of cellular communication.

Wow

Represents a set of transmitters and receivers, respectively

Index 0 of corresponds to a cellular user terminal (CUE) transmitter. The transmit power of transmitter i is denoted by Pi, where Pi is

Is in the range of. The cellular user terminal (CUE)

Fixed transmit power such as

Suppose we use Also,

And W are assumed to be noise spectral density and bandwidth, respectively.

와 같은 거리 관련 채널 이득에 대한 단순 경로 손실 모델을 고려할 수 있다. 송신기 i와 수신기 j 사이의 채널 이득인

는 하기 수학식 1과 같이 나타낼 수 있다.According to one embodiment,

We can consider a simple path loss model for distance-related channel gains such as The channel gain between transmitter i and receiver j

Can be expressed as in Equation 1 below.

는 송신기 i와 수신기 j 사이의 거리를 말한다.

는 송신기 i와 수신기 j 사이의 다중 경로 페이딩을 의미한다. 여기서

를 행렬로 나타낸 것이 입력 매트릭스

가 될 수 있다.Where β denotes a path loss factor and α denotes a path loss index.

Is the distance between the transmitter i and the receiver j.

Denotes multipath fading between transmitter i and receiver j. here

Where matrix is the input matrix

Can be

는 하기 수학식 2로 표현될 수 있다.According to one embodiment, the data rate of the user terminal (DUE) between the i-th device

May be represented by Equation 2 below.

는 셀룰러 사용자 단말(CUE)와 장치간 사용자 단말(DUE) 사이의 채널 이득이고,

는 장치간 사용자 단말(DUE) 사이의 채널 이득이다.here,

Is the channel gain between the cellular user terminal (CUE) and the inter-device user terminal (DUE),

Is the channel gain between the device-to-device user terminals (DUE).

According to an embodiment, the neural network-based transmission power control apparatus may consider a method of maximizing the overall weighted sum-rate (WSR) of the device-to-device user terminal (DUE). According to another embodiment, the neural network-based transmission power control device can be applied to the purpose of maximizing energy efficiency.

는 서로 다른 장치간 사용자 단말(DUE)들의 송신을 우선 순위를 정하는데 사용될 수 있는 DUE i에 대한 데이터 전송률의 가중치이다.here,

Is the weight of the data rate for DUE i that can be used to prioritize the transmission of user equipments (DUEs) between different devices.

According to an embodiment, the neural network-based transmission power control apparatus for communication between underlay devices may include a deep neural network model to find an optimized transmission power.

)이며, 송신 전력은

로 결정될 수 있다.The matrix H 410, which includes the channel gain between the device-to-device UEs and the channel gain between the device-to-device UE and the cellular user terminal (CUE), becomes an input of the deep neural network and outputs the deep neural network. Is the normalized transmit power (

) And the transmit power is

Can be determined.

가 될 수 있다. 여기서

는 i 번째 서브 블록의 완전 연결 층의 입력,

는 i 번째 서브 블록의 가중치 및

는 i 번째 서브 블록의 바이어스이다. 각 완전 연결 층의 출력은 심층 신경망에 비선형 성을 제공하는 정류된 선형 유닛(ReLU) 층으로 공급될 수 있다. 보다 구체적으로, 정류된 선형 유닛(ReLU) 층의 입력이

일 때, 출력은

가 될 수 있다. 완전 연결 파트의 최종 출력은 하이퍼볼릭탄젠트 파트(430)로 제공되며 하이퍼볼릭 탄젠트 연산을 실행할 수 있다. 하이퍼볼릭탄젠트 파트(430)의 입력이

이고, i 번째 요소의 벡터가

일때, 출력은

로 나타낼 수 있다. 하이퍼볼릭탄젠트 파트(430)의 출력은

의 값이 0과 1 사이가 되도록 -1과 1 사이에 있고, 상기 수학식 3의 조건,

을 만족한다.According to an embodiment, the fully connected part may be composed of N lower fully connected parts connected in series. Each lower fully connected part may comprise a fully connected layer and a rectified linear unit (ReLU) layer. Matrix multiplication of weights in a fully connected layer can occur with the addition of bias. Set the number of hidden nodes for the fully connected layer of the i th lower fully connected part to Fi, and the output of the fully connected layer is

Can be here

Is the input of the fully connected layer of the i th subblock,

Is the weight of the i th subblock, and

Is the bias of the i-th subblock. The output of each fully connected layer can be fed to a rectified linear unit (ReLU) layer that provides nonlinearity to the deep neural network. More specifically, the input of the rectified linear unit (ReLU) layer

When is output

Can be The final output of the fully connected part is provided to the hyperbolic tangent part 430 and can execute the hyperbolic tangent operation. The input of the hyperbolic tangent part 430

Where the vector of the i th element

When output is

It can be represented as. The output of the hyperbolic tangent part 430 is

Is between -1 and 1 so that the value is between 0 and 1, and the condition of Equation 3,

To satisfy.

According to an embodiment, a neural network-based transmission power control apparatus must first learn through a deep neural network, and then use the learned model to determine transmission power. A matrix that is a sample of the channel gains for various locations of the inter-device user terminal (DUE) and the cellular user terminal (CUE) under different fading conditions so that the deep neural network can learn a general transmit power control strategy to follow any channel condition. H can be collected for learning. Sufficient number of channel samples for various conditions is essential to prevent overfitting of learning. The collected channel gain can be converted into decibels and then normalized so that the mean and unit variances are zero. The nature of the path loss and the effects of deep fading mean that the magnitude of the channel gains for different samples can vary significantly, and since they adversely affect deep neural network learning, preprocessing of data may be necessary for better learning.

According to an embodiment, a deep neural network of a neural network-based transmission power control apparatus may lose a weighted sum-rate (WSR) of a communication terminal between devices. According to another embodiment, the loss function of the neural network may be a parameter of the transmission rate and interference of the terminal.

Deep neural networks can be trained using stochastic gradient descent (SGD) algorithms. In this case, a loss function L such as Equation 4 may be considered for learning.

는 다중 채널 샘플에 대해 업데이트 할 수 있다.Then the weights and deviations of the deep neural network

Can be updated for multi-channel samples.

Fully trained deep neural networks can be used to control transmit power. To determine the transmit power, the channel gain can be fed to a deep neural network model that is converted into decibels, normalized and then outputs the normalized transmit power. In the case of a large number of channel samples and device-to-device UEs, the deep neural network takes a long time to learn. However, the learned deep neural network can infer the transmission power with a short calculation time to be suitable for real-time operation.

5 is a diagram illustrating a convolutional neural network (CNN) in a neural network based transmission power control method according to an embodiment.

Referring to FIG. 5, the neural network-based transmission power control apparatus combines the input matrix 510 with a composite product part 520, a fully connected part 530, and a sigmoid part 540. You can learn from it and get the output.

According to an embodiment, the convolution product 520 may pre-train an iterative algorithm based transmission power allocation method. More specifically, the composite product part 520 may be trained to have the same result as the output result of the conventional method of controlling the terminal transmission power in the deep learning structure. Here, the iterative algorithm is one of methods for determining the existing transmission power, and may include a weighted minimum mean square error (WMMSE) and an iterative water-filling scheme.

According to an embodiment of the present disclosure, the fully connected part 530 may perform deep learning using a frequency efficiency or an energy efficiency as a loss function.

According to an embodiment, the sigmoid part 540 may obtain the transmission power by using the output of the fully connected part as an input value.

According to an embodiment, a transmission power control strategy of a wireless communication system may be described as deep power control (DPC) using a composite product neural network (CNN). The deep power control (DPC) may be automatically learned through in-depth learning so that the transmission power can be appropriately determined to maximize the spectrum efficiency or energy efficiency.

The deep power control method according to an embodiment may infer an appropriate transmission power of a user so that the computational complexity is lower than that of the conventional repetition scheme. In this case, a distributed deep power control using only local channel state information for determining transmission power is proposed.

와

는 송신기 세트 및 수신기 세트를 각각 나타내고,

에 속하는 수신기 k는

에 속하는 송신기 k로부터 데이터를 수신하고,

에 속하는 다른 송신기 i로부터 간섭을 수신할 수 있다. 이때, 입력

는 하기 수학식 5와 같이 나타낼 수 있다.According to an embodiment, it is possible to assume a wireless communication system in which a plurality of users are randomly distributed in the DxD region. In this case, it is assumed that there are N single antenna transceiver pairs including one transmitter and one receiver, and all transmissions occur simultaneously on the same frequency. Thus, communication of a transceiver pair may receive interference from the transmission of other transceiver pairs. For example,

Wow

Represents a transmitter set and a receiver set, respectively,

The receiver k belonging to

Receive data from transmitter k belonging to

May receive interference from another transmitter i belonging to. At this time, input

Can be expressed as Equation 5 below.

는 송신기 i와 수신기 j의 거리 관련 채널 이득이고,

는 송신기 i와 수신기 j의 다중경로 페이딩을 나타낸다.

는 독립적이고 동일하게 분포된 순환 대칭 복소 가우스(CSCG)로, 평균이 0이고 분산이 1인 확률 변수로 모델링 된다고 가정한다. 단순 경로 손실 모델은

로 표현된다. 여기서,

는 경로 손실 계수,

는 경로 손실 지수,

는 송신기 i와 수신기 j 간의 거리이다.

Is the distance related channel gain of transmitter i and receiver j,

Denotes the multipath fading of transmitter i and receiver j.

Is an independent and equally distributed cyclic symmetric complex Gaussian (CSCG), which is assumed to be modeled as a random variable with mean 0 and variance 1. The simple path loss model

It is expressed as here,

Is the path loss factor,

Is the path loss index,

Is the distance between transmitter i and receiver j.

The attainable frequency efficiency of the transmitter i can be written as Equation 6 below.

Here, when the transmit power of the transmitter i is Pi, 0 <Pi <Pmax, P is a vector of Pi, N0 is a noise spectral density, and W is a bandwidth.

The energy efficiency of the transmitter i may be expressed by Equation 7 below.

Where Pc is the power dissipated in the circuit of the transmitter. Here, maximization of frequency efficiency and maximization of energy efficiency are considered, two goals for controlling transmission power. In this case, as shown in Equation 8, optimization of transmission power may be expressed in consideration of frequency efficiency.

In addition, as shown in Equation 9, optimization of transmission power may be expressed in consideration of energy efficiency.

According to an embodiment, the convolution product 520 may include a plurality of lower blocks 521, 522, 523, 524, 525,... 526 connected in series. In this case, the lower block is a composite product layer (522, 524, 526) for performing a two-dimensional spatial composite product of the input data and a rectified linear unit (ReLU) layer (521, 523, 525) for injecting nonlinearity into the composite product neural network. It may include.

일 때, 출력은 max (

, 0)이 된다.The depth of the convolutional layer is set to Ci for the i th lower block. The stride, the step size used in the multiplication product filter, is set to 1 and zero padding is used so that the size of the output remains the same as the size of the input. The output of each convolutional layer is fed to a rectified linear unit layer to prevent negative values. More specifically, the input of the rectified linear unit layer

When is output, max (

, 0).

인 3 차원 행렬이라면 이것은

로 표시되는 벡터로 재구성될 수 있다. 그러면 완전 연결 파트의 출력은

가 된다. 여기서

는 완전 연결의 가중치이고,

는 완전 연결의 바이어스이다.

의 크기가

이고,

의 크기가

인 경우, 완전 연결 파트의 출력의 형상은

이 된다. 완전 연결 파트(530)의 출력은 시그모이드 파트(540)에 입력되어, 시그모이드 파트의 입력이

일 때, 시그모이드 파트의 i 번째 출력은

가 된다. 따라서, 송신 전력은 하기 수학식 10과 같이 결정될 수 있다.In the fully connected part, the output of the convolutional part is combined and reduced to N outputs, which can be used to determine the transmit power. The output of the product part is

If it is a three-dimensional matrix that is

It can be reconstructed into a vector denoted by. The output of the fully connected part

Becomes here

Is the weight of the full connection,

Is the bias of the full connection.

The size of

ego,

The size of

If, the shape of the output of the fully connected part is

Becomes The output of the fully connected part 530 is input to the sigmoid part 540, so that the input of the sigmoid part is

When is the i th output of the sigmoid part

Becomes Therefore, the transmission power may be determined as shown in Equation 10 below.

는 합성곱 파트의 기능을 나타낸다.here

Denotes the function of the composite product part.

를 얻고 합성곱 신경망의 입력이 될 수 있다.In a deep power control (DPC) according to an embodiment, an input matrix H, which is a sample of channel information, may be collected first. Since both the transmitter and receiver location and multipath fading are different for all channel samples, the optimal transmit power may be different for each channel sample. Thus, deep power control (DPC) can learn about any channel condition so that a general transmit power control strategy can be obtained. In the proposed technique, channel samples are converted to decibels and then normalized.

Can be input to the composite-product neural network.

When the channel sample is prepared, pre-learning may be performed before actual learning. According to an embodiment, the neural network may pre-learn a transmission power allocation method based on an iterative algorithm based on at least one of a weighted minimum mean square error (WMMSE) and iterative water filling, in the composite product part. have.

가 가중치 최소 평균 자승 에러(WMMSE) 체계의 전송 전력이라고 가정하면, 합성곱 신경망은 하기 수학식 11과 같이, 손실 함수

를 최소화하기 위해 사전 학습될 수 있다.During prior learning, the convolutional neural network can learn to regenerate the transmit power of conventional power control schemes, eg, weighted least mean square error (WMMSE), for given channel samples.

Assuming that is the transmission power of the weighted least mean square error (WMMSE) system, then the multiplicative neural network is a loss function,

Can be pre-learned to minimize

According to one embodiment, the pre-learned power control deep power control (DPC) may be an approximation of the weighted minimum mean square error (WMMSE) scheme. The role of pre-learning is to initialize the weights and biases of the convolutional neural network before actual learning. That is, even in the worst case, the transmit power of the weighted least mean square error (WMMSE) is regenerated to provide an initialization point so that the deep power control (DPC) can at least achieve a performance similar to the weighted least mean square error (WMMSE). have.

Learned deep power control (DPC) can be used to determine transmit power based on current channel state information (CSI), which is called inference in deep learning.

Although the training of a convolutional neural network involves a large number of calculations, the inference of deep power control (DPC) can be performed with relatively low overhead, which allows real-time computation of deep power control (DPC).

According to one embodiment, the prior learning and learning of the composite product neural network can be performed offline. That is, a server or cloud platform capable of parallel computing can be used for learning separately from a wireless communication system. The randomly selected user or central hub may use the trained multiplicative neural network to determine the transmit power based on the channel gains received from other users.

According to an embodiment, the neural network of the transmission power control apparatus may use local channel state information from neighboring users as an input value, and process channel state information of the remaining users as an average channel value. As the number of users of deep power control (DPC) increases, the signal overhead required to collect channel state information may increase. To solve this problem, distributed deep power control (DPC) is disclosed that does not require full channel state information. The transmit power can be calculated using local channel state information from neighboring users, which can reduce overhead compared to schemes based on overall channel state information.

는 전체 채널 상태 정보를 사용하여 학습된 합성곱 신경망으로 공급될 수 있다.In distributed deep power control, approximated

Can be fed to the learned multiplication neural network using the full channel state information.

행렬이 계산되고, 수집되지 않은 값에 대해서는 0으로 처리할 수 있다. 송신기 j는이 근사 된

를 학습된 합성곱 신경망에 공급함으로써 송신 전력 Pj를 결정할 수 있다.To this end, receiver j measures local channel state information for other transmitters and transmits this information to the corresponding transmitter j to receive normalized local channel state information. after that,

The matrix is computed and can be treated as 0 for uncollected values. Transmitter j has been approximated

Can be determined by supplying the trained multiplicative neural network.

6 is a diagram illustrating a composite product neural network including a pre-learning initialization process in a neural network based transmission power control method according to an embodiment.

Referring to FIG. 6, the neural network based transmission power control apparatus according to an embodiment may include an input matrix 610 including a first fully connected part 620, a product multiplication part 640, a second fully connected part 650, and the like. A convolutional neural network including the sigmoid part 660 may be learned and output a result value.

According to one embodiment, the first fully connected part 620 may include a reconstruction unit 630. Here, the first fully connected part 620 and the reconstructor 630 may learn and reconstruct an optimal input with respect to the input matrix.

According to an embodiment, the convolution product 640 may extract spatial features from the reconstructed input data.

According to an embodiment, the second fully connected part 650 may merge the extracted spatial features.

According to one embodiment, sigmoid part 660 may determine normalized transmit power.

According to one embodiment, it may be assumed that the cellular system with the communication function between the underlay device. Here, transceiver pairs (TPs) between the N single antenna devices may transmit data to each other, and the cellular user terminal (CUE) may transmit data to the base station. At this time, in the communication between the underlay devices, the uplink transmission and the device-to-device transmission of the cellular system share the same radio resource. In this case, it is assumed that the inter-device user terminals (DUE) and the cellular user terminal (CUE) are randomly distributed in the area S and the base station is located in the center of the area S, and the frequency effect for optimizing the transmission power See for more information.

는 트랜시버의 세트를 나타내고,

의 인덱스 0은 셀룰러 사용자 단말(CUE)의 전송부에 대응할 수 있다. 이때, DUE TP i의 전송 전력은 Pi로 표현되며,

조건을 만족하고 셀룰러 사용자 단말(CUE)의 전송 전력은

로 고정된다. W와 N0는 각각 대역폭 및 잡음 스펙스럼 밀도이다. 다중 경로 페이딩 및 거리에 관련된 채널 이득으로 구성된, 송신기 TPi와 수신기 TPj 간의 채널 이득이

가 될 수 있다.here,

Represents a set of transceivers,

Index 0 of may correspond to the transmission unit of the cellular user terminal (CUE). In this case, the transmit power of the DUE TP i is represented by Pi,

Conditions and the transmit power of the cellular user terminal (CUE)

Is fixed. W and N0 are the bandwidth and noise spectrum density, respectively. The channel gain between transmitter TPi and receiver TPj, consisting of multipath fading and channel gain relative to distance,

Can be

라고 할 때,

로 표현될 수 있다. 이때, 셀룰러 사용자 단말의 주파수 효율 SE0는 하기 수학식 13으로 표현될 수 있다.According to one embodiment, in determining the transmission power of the neural network-based transmission power control device, interference by transmission power between user equipments between devices may be considered. Interference with cellular user terminals by inter-device user terminals

When I say

It can be expressed as. In this case, the frequency efficiency SE0 of the cellular user terminal may be expressed by Equation 13 below.

)보다 작도록 제한하면서 장치간 사용자 단말들의 전체 주파수 효율을 극대화하기 위함이다.Here, the purpose of the transmission power control of the user terminal between the device is, as shown in equation (14), the amount of interference generated by the cellular user terminal is interference constraint (

In order to maximize the overall frequency efficiency of the user terminals between devices while limiting to less than).

According to one embodiment, the first fully connected part 620 may reconstruct the input data. In this case, the normalized channel gain may be used to alleviate the problem caused by the large scale difference between the maximum value and the minimum value of the channel gain.

According to one embodiment, the first fully connected part 620 may find an optimal reconstruction of the input data. In this case, the number of hidden nodes of the first fully connected part 620 may be equal to the number of normalized channel gain data. The output of the first fully connected part may be reconstructed, reconstructed into a two-dimensional matrix, and provided as an input of the convolution product part 640.

According to an embodiment, the convolution product 640 may include a plurality of lower blocks. The lower block of the convolution product part 640 may include a convolution product layer and a rectified linear unit (ReLU) layer. At this time, by applying to the composite product part 640, the weight of the deep neural network can be effectively reduced.

According to one embodiment, in the second fully connected part 650, the output of the composite product part may be summed into N outputs. For example, the three-dimensional matrix that is the output of the convolutional product part may be reconstructed into a one-dimensional vector and provided to the second fully connected part.

The neural network based transmission power control apparatus according to an embodiment should learn a deep neural network model requiring a large number of channel samples. Therefore, samples of the channel gain must be collected first. In this case, channel gain samples for different locations of the device-to-device (DUE) and the cellular user terminal (CUE) may be used to enable the deep neural network to learn a general transmit power control strategy for any channel condition. The learned deep neural network can be applied at any user location.

According to an embodiment, if a channel sample is prepared, an initialization step may be performed before learning. Unlike previous work, where weights and biases are randomly initialized, deep neural networks are initialized to regenerate existing transmit power control schemes for a given channel sample. The proposed initialization procedure can ensure that the deep neural network has an initial point that provides performance similar to the existing scheme.

The transmission power of the device-to-device user terminal (DUE) is a weighted minimum mean square error (WMMSE) transmission power that maximizes the frequency efficiency of the device-to-device user terminal (DUE) without considering interference caused by the cellular user terminal (CUE). Can be assumed to be initialized. Thus, the parameters of the DNN can be updated to minimize the loss function.

After initialization, the deep neural network may be trained to maximize the overall frequency efficiency of the inter-device user terminals (DUEs) while limiting the amount of interference due to cellular user terminals (CUE).

7 is a flowchart illustrating a neural network based transmission power control method according to an embodiment.

According to an embodiment, a neural network-based transmit power control method performed by a transmit power control apparatus of a wireless communication system (WCS) provides optimal transmission of a channel gain matrix through a neural network. The method may include learning power, receiving channel state information from a transmitting terminal, and controlling transmission power based on the channel state information.

According to an embodiment, in operation S710, the neural network based transmission power control apparatus may learn an optimal transmission power by inputting a channel gain matrix through a neural network.

In this case, the neural network may be a deep neural network including a fully connected layer.

According to another embodiment, the neural network may be a convolutional neural network including a compound product part, a fully connected part, and a sigmoid part.

According to an embodiment, the learning of the optimal transmission power may include pre-learning a transmission power allocation scheme based on a weighted minimum mean square error (WMMSE) in the composite product part.

According to an embodiment, the learning of the optimal transmission power may use local channel state information from adjacent users as an input value and process channel state information of the remaining users as an average channel value.

According to an embodiment, the wireless communication system (WCS) is underlay D2D communication (Underlaid D2D Communication), the transmission power control device is a transmission power of the D2D user equipment (DUE) between devices And the neural network may set a weighted sum-rate (WSR) of the inter-device communication terminal as a loss function. According to another embodiment, the loss function of the neural network may be a parameter of the transmission rate and interference of the terminal.

According to an embodiment, the neural network may learn by considering interference generated from transmission power of a cellular user equipment (CUE).

According to an embodiment, the neural network includes a convolution neural network including a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part. Network), and input data may be reconstructed in the first fully connected part.

According to an embodiment of the present disclosure, the second fully connected part may learn by considering both an interference between a cellular user terminal (CUE) and a user terminal between devices.

According to an embodiment, in step S720, the neural network based transmission power control apparatus may receive channel state information from a transmitting terminal. In this case, the channel state information may include only local channel state information from adjacent users, not overall channel state information.

According to an embodiment, in operation S730, the neural network based transmission power control apparatus may control the transmission power based on the channel state information. In this case, the neural network based transmission power control apparatus according to an embodiment may control the transmission power of the user terminal between devices in consideration of interference generated in the transmission power of the cellular user terminal or the user terminal between the devices.

The system or apparatus described above may be implemented as a hardware component, a software component, and / or a combination of hardware components and software components. For example, the systems, devices, and components described in the embodiments may include, for example, processors, controllers, arithmetic logic units (ALUs), digital signal processors, microcomputers, field programmable arrays (FPAs). ), A programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions, may be implemented using one or more general purpose or special purpose computers. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of explanation, one processing device may be described as being used, but one of ordinary skill in the art will appreciate that the processing device includes a plurality of processing elements and / or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as parallel processors.

The software may include a computer program, code, instructions, or a combination of one or more of the above, and configure the processing device to operate as desired, or process it independently or collectively. You can command the device. Software and / or data may be any type of machine, component, physical device, virtual equipment, computer storage medium or device in order to be interpreted by or to provide instructions or data to the processing device. Or may be permanently or temporarily embodied in a signal wave to be transmitted. The software may be distributed over networked computer systems so that they may be stored or executed in a distributed manner. Software and data may be stored on one or more computer readable recording media.

The method according to the embodiments may be embodied in the form of program instructions that may be executed by various computer means and recorded on a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the media may be those specially designed and constructed for the purposes of the embodiments, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Although the embodiments have been described by the limited embodiments and the drawings as described above, various modifications and variations are possible to those skilled in the art from the above description. For example, the described techniques may be performed in a different order than the described method, and / or components of the described systems, structures, devices, circuits, etc. may be combined or combined in a different form than the described method, or other components. Or even if replaced or substituted by equivalents, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are within the scope of the claims that follow.

310: neural network based transmission power control device
311: neural network
312: control unit
313: communication unit
321, 322, ..., 32n: user terminal

Claims (18)

In the neural network-based transmission power control apparatus in a wireless communication system (WCS),
Optimal transmission using one model of deep gain network, convolutional neural network, and convolutional neural network that performs pre-learning Neural networks for learning power; And
A control unit for controlling the transmission power based on the channel state information,
The neural network uses the local channel state information from adjacent users as an input value, and processes the channel state information of the remaining users as an average channel value to learn the optimal transmission power based on the neural network. The method of claim 1,
The deep neural network may include a fully connected layer.
Neural network-based transmission power control device. The method of claim 1,
The convolutional neural network includes a convolutional part, a fully connected part, and a sigmoid part.
Neural network-based transmission power control device. The method of claim 3,
The neural network is characterized in that the pre-learning a transmission power allocation method based on an iterative algorithm
Neural network-based transmission power control device. delete The method of claim 1,
The wireless communication system (WCS) is Underlaid D2D Communication (Underlaid D2D Communication),
The apparatus for controlling transmission power controls the transmission power of a D2D user equipment (DUE) between devices,
The loss function of the neural network is based on the transmission rate and the interference of the terminal.
Neural network-based transmission power control device. The method of claim 6,
The neural network learns in consideration of interference generated in the transmission power of a cellular user equipment (CUE).
Neural network-based transmission power control device. The method of claim 1,
A convolutional neural network performing the pre-learning initialization process may include a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part. Including,
Input data is reconstructed in the first fully connected part.
Neural network-based transmission power control device. The method of claim 8,
The second fully connected part learns in consideration of both the interference between the cellular user terminal (CUE) and the user terminal between the devices.
Neural network-based transmission power control device. In the neural network-based transmission power control method performed by a transmission power control apparatus of a wireless communication system (WCS),
The channel gain matrix is input through a neural network, and one of a deep neural network, a convolution neural network, and a convolution neural network that performs a pre-learning initialization process. Learning an optimal transmission power using a model;
Receiving channel state information from a transmitting terminal; And
Controlling transmission power based on the channel state information;
The learning of the optimal transmission power, the neural network-based transmission power control method of using the local channel state information from the adjacent users as an input value, the channel state information of the remaining users as an average channel value. The method of claim 10,
The deep neural network may include a fully connected layer.
Neural network based transmission power control method. The method of claim 10,
The convolutional neural network includes a convolutional part, a fully connected part, and a sigmoid part.
Neural network based transmission power control method. The method of claim 12,
The learning of the optimal transmission power may include pre-learning a transmission power allocation scheme based on an iterative algorithm.
Neural network based transmission power control method. delete The method of claim 10,
The wireless communication system (WCS) is Underlaid D2D Communication (Underlaid D2D Communication),
The apparatus for controlling transmission power controls the transmission power of a D2D user equipment (DUE) between devices,
The loss function of the neural network is based on the transmission rate and the interference of the terminal.
Neural network based transmission power control method. The method of claim 15,
The neural network learns in consideration of interference generated in the transmission power of a cellular user equipment (CUE).
Neural network based transmission power control method. The method of claim 10,
A convolutional neural network performing the pre-learning initialization process may include a first fully connected part, a composite product part, a second fully connected part, and a sigmoid part. Including,
Input data is reconstructed in the first fully connected part.
Neural network based transmission power control method. The method of claim 17,
The second fully connected part learns in consideration of both the interference between the cellular user terminal (CUE) and the user terminal between the devices.
Neural network based transmission power control method.

Images