KR102371914B1 - Sphere decoding appararus and method in large mimo system - Google Patents

Abstract

According to the present technology, disclosed are a sphere decoding device of a large-scale multiple input/output system and a method thereof. According to a specific example of the present invention, a minimum path metric of a deep path of all subtrees for a received signal vector of a neural network optimized as a tree structure is estimated, and then rearranged according to a size of the estimated minimum path metric of the top layer and a sphere decoding solution from a subtree corresponding to the order of the rearranged minimum path metric is searched, thereby fundamentally reducing computational complexity of sphere decoding.

Description

Rescue and decryption apparatus and method of large-scale multiple input/output system {SPHERE DECODING APPARARUS AND METHOD IN LARGE MIMO SYSTEM}

The present invention relates to an apparatus and method for decoding a large-scale multiple input/output system, and more particularly, to estimate the minimum path metric of a lower tree through a neural network for a received signal vector of an optimized tree structure, and to estimate the minimum path of the uppermost layer. It relates to a technique for reducing the computational complexity of old decoding by rearranging the size of metrics, determining the search radius according to the order of the rearranged minimum path metrics, and searching for the old decryption solution in the predetermined search radius.

In a multiple-input multiple-output (MIMO) system, the old decoding (sphere decoding) algorithm is known as an efficient signal detection method that has similar performance to a maximum-likelihood detection (MLD) receiver.

Recently, large-scale MIMO systems using multiple antennas in a base station for data transmission and reception have attracted great research attention in order to meet the increasing demand for ultra-high-speed data transmission in mobile communication systems.

To maximize the achievable data rate in such a large-scale MIMO system, the base station must receive as many symbols as possible simultaneously from multiple terminals to increase the multiplexing gain. In this case, the optimal receiver such as the old decoding plays an important role in accessing the channel capacity. However, the computational complexity of the old decoding greatly increases according to the number of antennas and the modulation order, so it is difficult to apply the old decoding to a large-scale MIMO system.

Recently, a deep learning (DL) technique has been applied to various fields, and the deep learning technique is a neural network-based sphere decoder and deep-based sphere decoder to select the optimal initial radius of a large-scale MIMO system. There have been various attempts to apply it to

However, in a large-scale MIMO system, information required for estimating a path metric is high-dimensional, which may increase the computational complexity of the neural network.

Accordingly, the present applicant estimates the minimum path metric of the deep path of the subtree in the tree structure to be searched, then orders the estimated minimum path metric of the highest layer in the order of size, and the subtree corresponding to the smallest minimum path metric We propose a method to fundamentally reduce the computational complexity by searching for the old decoding solution.

1. H. Q. Ngo, and E. G. Larsson, “No downlink pilots are needed in TDD massive MIMO,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 2921-2935, May 2017. 2. D. P. Kingma and J. L. Ba, “Adam: A Method for Stochastic Optimization,” A conference paper at the 3rd International Conference for Learning Representations, San Diego, 2015. 3. E. Agrell, T. Eriksson, A. Vardy, and K. Zeger, "Closest point search in lattices," IEEE Trans. Inf. Theory, vol. 48, no. 8, pp. 2201-2214, Aug. 2002.

The present invention estimates the minimum path metric of a deep path of a subtree for an input vector of a neural network optimized in a tree structure, then rearranges the estimated minimum path metric of the top layer according to the size, and according to the order of the rearranged minimum path metric An object of the present invention is to provide an old decoding device and method for a large-scale multiple input/output system that can fundamentally reduce the computational complexity of old decoding by searching for a sphere decoding solution.

The object of the present invention is not limited to the object mentioned above, and other objects and advantages of the present invention not mentioned can be understood by the following description, and will be more clearly understood by the examples of the present invention. Further, it will be readily apparent that the objects and advantages of the present invention can be realized by the means and combinations thereof indicated in the appended claims.

The old decoding apparatus of a large-scale multiple input/output system according to an embodiment of the present invention comprises:

a receiving device for generating a received signal vector and a channel matrix from an RF band signal received through a plurality of receiving antennas; and

The received signal vector is arranged in a tree structure using the R matrix characteristic obtained through QR decomposition of the generated channel matrix H , and the minimum path metric of the lower tree is estimated by learning the received signal vector of the arranged tree structure with a neural network. , which rearranges the estimated minimum path metric of the highest layer in order of size, and then sequentially searches for the old decryption solution from the lower tree corresponding to the smallest minimum path metric.

Preferably, the phrase decoder,

an optimizer for arranging a received signal vector in a tree structure by using the R matrix characteristic obtained through the QR decomposition technique of the generated channel matrix and optimizing the received signal vector of the tree structure through a search space constraint;

a learning unit for estimating the minimum path matrix of the deep path of all subtrees by learning the neural network through the optimized received signal vector;

a minimum path metric sorting unit for arranging a minimum path metric of an estimated subtree of the uppermost layer in order of size; and

and a search unit configured to set a search radius according to the order of the estimated minimum path metric, and to search for an old decoding solution in the set search radius.

Preferably, the phrase decoder,

Further comprising a minimum path metric deriving unit for deriving a minimum path metric for the received signal vectors of the remaining layers except the top layer,

The minimum path metric alignment unit,

The minimum path metric deriving unit may be provided to generate a smallest path metric with respect to the derived minimum path metric and to sort the generated minimum path metric in order of magnitude.

Preferably, the phrase decoder,

The method may further include an early search termination unit that sets a search radius according to the order of the sorted minimum path metrics and stops the search for the old and decrypted solution when a predetermined early search termination condition is satisfied.

Preferably, the minimum path metric alignment unit,

It may be provided to sort in ascending order according to the size of the generated minimum path metric.

A method for recovering and decoding a large-scale multiple input/output system according to an embodiment of the present invention comprises:

(a) a reception step of generating a reception signal vector and a channel matrix from RF band signals received through a plurality of reception antennas; and

(b) The received signal vector is arranged in a tree structure using the R matrix characteristic obtained through QR decomposition of the generated channel matrix, and the minimum path metric of the lower tree is obtained by learning the received signal vector of the arranged tree structure with a neural network. It is characterized in that it includes an old decoding step of estimating, rearranging the estimated minimum path metric of the highest layer in order of size, and then sequentially searching for an old decoding solution from a subtree corresponding to the smallest minimum path metric.

Preferably, step (b) comprises:

arranging the received signal vector in a tree structure using the R matrix characteristic obtained through the QR decomposition technique of the generated channel matrix and optimizing the received signal vector of the tree structure through a search space constraint;

estimating a minimum path metric of a deep path of a subtree by training a neural network through an optimized received signal vector;

arranging the estimated minimum path metric of the subtree of the uppermost layer in order of size; and

The method may include setting a search radius according to the order of the sorted minimum path metrics and searching for a deciphering solution in the set search radius.

Preferably, step (b) comprises:

Further comprising the step of deriving a minimum path metric for the subtrees of the remaining layers except the top layer,

The step of arranging the minimum path metric in order of size is

It may be provided to generate a minimum path metric with the smallest size for each of the estimated minimum path metric of the highest layer and the derived minimum path metric of the remaining layers except for the top layer, and to sort the generated minimum path metric in order of size. .

Preferably, step (b) comprises:

The method may further include setting a search radius in the order of the smallest sorted minimum path metric, searching the old decrypted solution within the set search radius, and stopping the search for the old decrypted solution when a predetermined early termination condition is satisfied.

According to this feature, the minimum path metric of the subtree is estimated for the input vector of the neural network optimized as a tree structure, then rearranged in the order of the size of the estimated minimum path metric of the top layer, and according to the order of the rearranged minimum path metric It has the effect of fundamentally reducing the complexity of the old decryption operation by searching the old decryption solution within the specified search radius.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and serve to further understand the technical spirit of the present invention together with the detailed description of the present invention to be described later, so the present invention is a matter described in such drawings should not be construed as being limited only to
1 is a block diagram of a large-scale MIMO system to which an embodiment is applied.
2 is a detailed configuration diagram of a rescue/decryption apparatus of a system according to an exemplary embodiment.
3 is an exemplary view showing a received signal vector of an arrangement structure according to an embodiment.
4 is an exemplary diagram illustrating a structure of a neural network according to an embodiment.
5 is a diagram illustrating a recovery/decryption process according to an embodiment.
6 is a diagram showing BER performance of the old decoding apparatus according to an embodiment.
7 is a first exemplary diagram illustrating the computational complexity of the old decoding apparatus according to an embodiment.
8 is a second exemplary diagram illustrating the computational complexity of the old decoding apparatus according to an embodiment.
9 is an overall flowchart illustrating a rescue/decryption process according to another embodiment.

Hereinafter, embodiments of the present invention will be described in more detail with reference to the drawings.

Advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only these embodiments allow the disclosure of the present invention to be complete, and common knowledge in the art to which the present invention pertains It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims.

Terms used in this specification will be briefly described, and the present invention will be described in detail.

The terms used in the present invention have been selected as currently widely used general terms as possible while considering the functions in the present invention, which may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, and the like. In addition, in a specific case, there is a term arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the corresponding invention. Therefore, the term used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than the name of a simple term.

In the entire specification, when a part "includes" a certain element, this means that other elements may be further included, rather than excluding other elements, unless otherwise stated. Also, as used herein, the term “unit” refers to a hardware component such as software, FPGA, or ASIC, and “unit” performs certain roles. However, "part" is not meant to be limited to software or hardware. A “unit” may be configured to reside on an addressable storage medium and may be configured to refresh one or more processors.

Thus, by way of example, “part” includes components such as software components, object-oriented software components, class components and task components, processes, functions, properties, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays and variables. The functionality provided within components and “parts” may be combined into a smaller number of components and “parts” or further divided into additional components and “parts”.

One embodiment estimates a deep path minimum path metric of a subtree with respect to an input vector of a neural network in which the received signal vectors are arranged in a tree structure, and then rearranges them in the order of the size of the estimated minimum path metric of the highest layer, and the sorted minimum path The old decoding solution is searched according to the order of the metrics.

Hereinafter, with reference to the accompanying drawings, the embodiments of the present invention will be described in detail so that those of ordinary skill in the art can easily implement them. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description will be omitted.

Hereinafter, an apparatus for decoding a large-scale multiple input/output system according to an embodiment will be described with reference to the accompanying drawings.

1 is a block diagram showing the configuration of a receiving end of a large-scale multiple input/output system to which an embodiment is applied, FIG. 2 is a detailed configuration diagram of the old decoding device 20 shown in FIG. 1, and FIG. 3 is the optimization unit of FIG. It is an exemplary view showing a received signal vector of an array structure, FIG. 4 is an exemplary view showing a neural network structure of the learning unit of FIG. 2, and FIG. 5 is a view showing an operation process of the old decoding apparatus of FIG.

First, in the case of a large-scale multiple input/output system having N _t transmit antennas and N _r receive antennas, the receive signal vector y of the receiver 10 generated from RF band signals received through a plurality of receive antennas is expressed by Equation 1 below. can

[Equation 1]

의 N_r x 1 크기의 복소수 부가 화이트 가우시안 노이즈 벡터이고, x는 QAM(Quadrature Amplitude Modulation) 콘스텔레이션

에서 추출된 N_t x 1 크기의 복소수 전송 신호 벡터이다.Here, the reception signal vector y is a complex reception signal vector of size N _r x 1, H is a complex channel matrix of size N _r × N _t , and v is a zero mean and variance matrix.

is a complex additive white Gaussian noise vector of size N _r x 1, and x is a quadrature amplitude modulation (QAM) constellation.

It is a complex number transmission signal vector of size N _t x 1 extracted from .

The received signal vector y of the receiving end 10 is transmitted to the old decoding device 20 .

As shown in FIG. 2 , the old decoding device 20 includes an optimization unit 110 , a learning unit 120 , a minimum path metric derivation unit 130 , a minimum path metric aligning unit 140 , and a search for an old decoding solution. At least one of the units 150 may be included.

에서 수신신호 벡터 y까지의 유클리드(Euclidean) 거리가 가장 작은 격자점

은 최적의 ML(maximum-likelihood) 검출 기법을 통해 검색되고, 이때 격자점

은 다음 식 2로 나타낸다.full sphere of transmit signal vector x

The lattice point with the smallest Euclidean distance from to the received signal vector y

is searched through the optimal ML (maximum-likelihood) detection technique, where lattice points

is represented by Equation 2 below.

[Equation 2]

상의 후보 벡터에 대한 ML 검출 시 높은 연산 복잡도가 요구된다.these phrases

High computational complexity is required when ML detection for a candidate vector of an image is performed.

Accordingly, the optimizer 110 has a limited search space to maintain ML performance while having low computational complexity. That is, the Sphere Decoding (SD) solution for the search radius d can be expressed by Equation 3 below.

[Equation 3]

And when the optimizer 110 performs QR decomposition (QRD) on the generated channel vector H , the channel vector H can be obtained by Equation 4.

[Equation 4]

Here, the R matrix is an upper triangular matrix of size N _r x N _t obtained through QR decomposition of the channel vector H , and the Q = [Q1, Q2] matrix is a unit matrix of size N _r x N _t . , Q1 and Q2 each consist of the first N _t columns and the last N _r -N _t columns of Q. And the search space constraint is expressed by Equation 5.

[Equation 5]

및

이다. 검색 공간 제약을 반영하여 구복호 구복호 해는 식 6로 재공식화할 수 있다.here,

and

am. Reflecting the search space constraints, the old-and-decrypted old-and-decrypted solution can be reformulated as Equation 6.

[Equation 6]

That is, using the R matrix characteristic of the channel matrix H decomposed by the QR decomposition detection technique for the channel matrix, the received signal vector is arranged in a tree structure, and at this time, it branches from the root node to the leaf node ( branch), the old decoding solution is searched using the all-path metric accumulating the matrix. And the full path metric can be used to optimize early termination and initial search radius.

Accordingly, the learning unit 120 first estimates the path metric based on the neural network in order to search for the old decoding solution using the full path metric.

Here, the neural network is designed to estimate the minimum path metric of the subtree of the top layer, the layer N _t . That is, before starting the tree search, the DPP deep path prediction-based sphere decoding scheme of an embodiment estimates the minimum path metric of the deep path to the rib node of each subtree of the layer N _t , and then The minimum path metric is used to set the ordering of the minimum path metric size, early termination, and transmit symbol search radius.

개 노드 중 하나를 루트(root)로 가진다. Referring to FIG. 3 , the optimizer 110 arranges the received signal vector in a tree structure by using the R matrix characteristic derived from the QR decomposition of the channel matrix for the old decoding employing the QPSK (Quadrature Phase Shift Keying) modulation method. and is provided as an arranged subtree, and each subtree is a hierarchy of N _t

It has one of the nodes as the root.

The received signal vectors arranged in the tree structure shown in FIG. 3 are transmitted to the learning unit 120 , and the neural network structure of the learning unit 120 is as shown in FIG. 4 . The received signal vector arranged in the tree structure shown in FIG. 3 is provided to the neural network shown in FIG. 4 as an input vector, and the neural network learns the input vector to estimate the minimum path metric of the subtree of the top layer N _t . . Accordingly, the learning unit 120 may learn the neural network through the optimized received signal vector to estimate the minimum path metric of the deep paths of all sub-trees.

일 때 N_t 계층의 q 번째 노드에 의해 루트된 서브 트리의 최소 경로 메트릭

은 하기 식 6와 같이 공식화할 수 있다.That is, the target vector of the neural network is

The minimum path metric of the subtree rooted by the qth node of the N _t hierarchy when

can be formulated as in Equation 6 below.

[Equation 6]

는 계층 N_t의 후보 심볼이고, 전송신호 벡터 x 내의 N_t-1 심볼의 포함된 벡터는

로 나타낸다. 이에

=

이다.here,

is a candidate symbol of the layer N _t , and the vector included in the N _t -1 symbol in the transmission signal vector x is

is indicated by Therefore

=

am.

와 채널 정보 H의 QR 분해된 Q행렬의 허미션(Hermitian)행렬을 곱한 행렬과 상 삼각행렬

은 고차원이므로 신경망의 연산 복잡도는 크게 증가된다. 따라서 학습부(120)는 높은 입력 벡터의 차원에 의해 대규모의 신경망의 학습셋이 요구되고 이에 추정 정확도가 낮아질 수 있다.Since such a path metric is dependent on the channel matrix and the received signal vector, the input vector of the neural network NN for estimating the path metric includes channel and received signal coefficient information. Received signal vector and channel status in large-scale MIMO system

The matrix multiplied by the Hermitian matrix of the QR decomposed Q matrix of the channel information H and the upper triangular matrix

Since is of high dimensionality, the computational complexity of the neural network is greatly increased. Therefore, the learning unit 120 requires a large-scale training set of the neural network due to the high dimension of the input vector, and thus estimation accuracy may be lowered.

Accordingly, for QR-decomposed channel information, in order to reduce the number of input vectors of the neural network, the path metric can be expressed as Equation 7.

[Equation 7]

,

및

을 기반으로 경로 메트릭은 연산될 수 있다. already know

,

and

A path metric may be calculated based on .

,

및

은 신경망의 입력으로 사용될 수 있다.

은

및

보다 작은 수를 포함하고 있고, 선행문헌 1에 개시된 빔의 점근적으로 유리한 전파(asymptotically favorable propagation) 및 채널 경화(channel-hardening) 효과로 인해 대규모 MIMO 시스템에서 대각선 행렬에 접근될 수 있다. 즉,

이다. That is, to estimate the path metric instead of the received signal and channel matrix.

,

and

can be used as input to the neural network.

silver

and

It contains a smaller number, and due to the asymptotically favorable propagation and channel-hardening effect of the beam disclosed in Prior Document 1, the diagonal matrix can be accessed in a large-scale MIMO system. in other words,

am.

이라고 가정하면,

는 고정된 행렬로 근접할 수 있고, 수신 신호와 채널 정보에 대해 독립적이다. here,

Assuming that

can be approximated by a fixed matrix, and is independent of the received signal and channel information.

는 식 1의 수신 신호 및 채널 행렬의 정규화에 의해 얻어질 수 있고, 이에 신경망의 입력 벡터에서 제외될 수 있다. 경로 메트릭은 노이즈 분산

에 종속적이고, 노이즈 분산

는 신경망의 최소 경로 메트릭의 추정 정확도를 개선할 수 있다. also,

can be obtained by normalization of the received signal and channel matrix of Equation 1, and thus can be excluded from the input vector of the neural network. The path metric is noise variance

is dependent on the noise distribution

can improve the estimation accuracy of the minimum path metric of the neural network.

Accordingly, the input vector e of the neural network of the learning unit 120 is set by Equation 8 below.

[Equation 8]

Here, the magnitude of the input vector e is 2N _t +2.

As shown in FIG. 4 , the learning unit 120 adopts a Gaussian radial basis function network (G-RBFN) system for estimating a minimum path metric including one hidden layer, and the radial basis function of the G-RBFN system is It is represented by following formula (9).

[Equation 9]

,

,

과 폭

을 가지는 가우시안 함수를 채용하고, 여기서 가우시안 함수는 은닉층의 각 계층의 활성 함수로 사용된다. 은닉층 내의 계층의 수는

로 설정된다. The G-RBFN scheme is central for path metric estimation

over width

Adopt a Gaussian function having , where the Gaussian function is used as an activation function of each layer of the hidden layer. The number of layers in the hidden layer is

is set to

Equation 10 is the mean squared error (MSE) loss function used to optimize the parameter vector θ constituting the bias and weight between the input layer and the hidden layer and the hidden layer and the output layer in the neural network.

[Equation 10]

, 및

각각은 m번째 학습 데이터에 대한 타겟 벡터와 출력 벡터이다. 학습의 최적화 알고리즘으로 스케일링된 켤레 그레디언트(SCG: scaled conjugate gradient)가 이용되며 학습 비(learning rate)는 0.0001로 설정된다.Here, M is the number of training,

, and

Each is a target vector and an output vector for the m-th training data. A scaled conjugate gradient (SCG) is used as an optimization algorithm for learning, and a learning rate is set to 0.0001.

는 최상 계층 N_t의 모든 하위 트리의 추정된 최소 경로 메트릭이고, 최소 경로 메트릭 도출부(130)의 최소 경로 메트릭은 계층 N_t 계층을 제외한 나머지 계층 N_t-1 이하의 하부 트리의 최소 경로 메트릭이다.An output vector for any m-th training data in the neural network of the learning unit 120 .

is the estimated minimum path metric of all subtrees of the uppermost layer N _t , and the minimum path metric of the minimum path metric deriving unit 130 is the minimum path metric of the subtrees below N _t-1 of the remaining layers except for the layer N _t layer. am.

If the estimated minimum path metric of the subtree is small, the probability of the subtree including the final solution of the old decoding increases. Therefore, it is efficient in terms of computational complexity to first search a subtree with a smaller estimated path metric.

As shown in FIG. 1, the ordering of the subtree corresponding to the minimum path metric is applied to neural network-based ordering (NN-aided ordering) to the top layer N _t , and Schnorr-Euchner reconstruction is applied to the remaining layers except for the top layer N _t . Call (SE-SD) ordering is applied.

That is, the minimum path metric of the lower tree of the remaining layers N _{t -1} or less except for the layer N t is derived by the minimum path metric deriving unit 130 . The minimum path metric of the subtree below the remaining layer N _t-1 is derived by Schnorr-Euchner old decoded (SE-old decoded) ordering, and in the present specification, Schnorr-Euchner old decoded (SE-old decoded) ordering is described in the prior literature. 3, and should be understood at the level of a person skilled in the art.

The minimum path metrics of all subtrees calculated by the minimum path metric deriving unit 130 are transmitted to the minimum path metric arranging unit 140 .

을 도출하고 도출된 가장 작은 크기의 최소 경로 메트릭

을 순서화한다. The minimum matrix path aligning unit 140 is a minimum path metric of the smallest size among the estimated minimum path metrics of the tree structure.

, and the derived smallest size smallest path metric

order the

내의 추정된 최소 경로 메트릭를 오름차순으로 정렬하여

을 생성한다. 즉,

이다.At this time, the ordering of the neural network is the output vector of the neural network.

By sorting in ascending order the estimated minimum path metric within

create in other words,

am.

This ordered and estimated minimum path metric is transmitted to the search unit 150 for deciphering and decoding. The old decryption solution search unit 150 searches for the old decrypted solution by sequentially setting an initial search radius according to the sort order of the sorted minimum path metric.

In this case, if the initial search radius is small, the computational complexity of the old decoding is reduced and the possibility that the actual old decoding solution exists outside the sphere region is increased, so the bit error rate (BER) performance may be degraded. Conversely, if the initial search radius is large, BER performance is improved, but computational complexity is increased. Therefore, according to an embodiment, a method for reducing the initial search radius while maintaining near-optimal BER performance is proposed.

을 기반으로 결정된다. Therefore, in order to satisfy the constraint on the probability of the actual old decryption solution existing inside the sphere in the old decoding algorithm, the initial search radius is

is determined based on

은 다음 식 11로부터 정해진다.i.e. the initial search radius

is determined from the following equation (11).

[Equation 11]

는 역 불완전 감마 함수이고 1-ε은 구 내에 존재하는 실제 해의 확률이며 여기서, 1-ε은 0.999로 가정한다.here,

is the inverse incomplete gamma function and 1-ε is the probability of an actual solution in the sphere, where 1-ε is assumed to be 0.999.

에서 가장 크기가 작은 최소 경로 메트릭

에 해당하는 하위 트리에 대해 심층 경로를 통해 구복호 해를 검색을 수행한 다음

을 제외한 최소 경로 메트릭

에 해당하는 하위 트리의 구복호 해를 검색한다. 이러한 구복호 해 검색은 구복호의 조기 종료 조건을 만족할 때까지 계속된다.The search unit 150 is an ordered minimum path metric of the lower nodes of the layer N _t .

smallest path metric in

After performing a search for the old decryption solution through a deep path for the subtree corresponding to

Minimum route metric excluding

Search for the old decryption solution of the subtree corresponding to . This search for the old decryption solution is continued until the condition for the early termination of the old decryption is satisfied.

=

을 고려하여 초기 검색 반경

를 다음 식 12로 설정한다.The old decoding solution search unit 150 is the smallest minimum path metric among the estimated minimum path metrics among all possible paths in the tree structure.

=

taking into account the initial search radius

is set to Equation 12 below.

[Equation 12]

이 정확하게 추정된 경우, λ₁=1을 가지는 초기 검색 반경은

을 만족하고 이에 구 내에 하나의 구복호 해가 존재되므로, 최적 ML 성능이 달성된다.Here, λ ₁ is a design parameter. smallest path metric

If this is correctly estimated, the initial search radius with λ ₁ =1 is

, and there is one old decoding solution in the sphere, so that the optimal ML performance is achieved.

의 추정 에러가 포함될 수 있으며, 이에 최적 ML 성능과 근접된 성능을 제공하기 위해 λ₁>1로 선택되면, 초기 검색 반경은

이다. 따라서, 트리 검색의 연산 복잡도를 감소하기 위해 초기 검색 반경은

대신에

즉, λ₁

로 설정된다. However, the smallest minimum path metric

Estimation error of may be included, and if λ ₁ >1 is selected to provide performance close to the optimal ML performance, the initial search radius is

am. Therefore, to reduce the computational complexity of tree search, the initial search radius is

Instead of

That is, λ ₁

is set to

On the other hand, the old decoding apparatus 20 further includes an early search termination unit 160, and the early search termination unit 160 is a search for searching for an old decoding solution based on the estimated minimum path metric in order to reduce computational complexity. When the radius satisfies a predetermined early termination condition, an early termination algorithm for old decoding may be performed.

는 도출된 구복호 해의 거리로 업데이트된다. p번째 하부 트리 내의 구복호 해의 검색이 완료되면 검색 반경

은 현재 구복호의 최상해에 의해 결정된다. That is, the search radius when the old decryption solution is derived through the search of the p -th subtree of the first layer.

is updated with the distance of the derived old decoding solution. When the search for the old decryption solution in the p -th subtree is completed, the search radius

is determined by the best solution of the current reconstruction.

At this time, the early search termination unit 160 determines whether the early search termination condition of Equation 13 below is satisfied.

[Equation 13]

는 설계 파라미터이다. 이러한 식 16의 조기 검색 종료 조건을 만족하는 경우 가장 최근 구복호 해는 최종해로 채택되고 이에 신경망 지원 순서화 알고리즘은 종료된다. 이에 신경망 지원 순서화 기반의 구복호 DPP-구복호의 연산 복잡도는 기존의 구복호 SE-SD의 연산 복잡도보다 감소된다. here,

is a design parameter. If the early search termination condition of Equation 16 is satisfied, the most recent old decryption solution is adopted as the final solution, and the neural network-supported ordering algorithm is terminated. Accordingly, the computational complexity of the neural network-supported ordering-based old decoding DPP-old decoding is reduced compared to that of the existing old decoding SE-SD.

이 추정되고, 단계 2 및 3에서 계층 Nt의 하위 트리의 최소 경로 메트릭

중 가장 크기가 작은 최소 경로 메트릭을 생성하고 생성된 최소 경로 메트릭의 크기에 따라 오름차순으로 재정렬한다.5 is a diagram showing an algorithm of a deep path estimation old decoding scheme. In step 1, a minimum path metric using a neural network

This is estimated, and the minimum path metric of the subtree of the hierarchical Nt in steps 2 and 3

Generates the smallest path metric among the smallest and rearranges them in ascending order according to the size of the generated smallest path metric.

중 재정렬된 최소 경로 메트릭 중 가장 크기가 작은 재정렬 최소 경로 메트릭

을 기반으로 결정된다. 이 후 단계 5 내지 16에서, 구복호 해 검색 진행은 각 하위 트리에서 수행된다. And in step 4, the initial search radius is the minimum path metric

The smallest reordered least path metric of the smallest reordered least path metric

is determined based on After that, in steps 5 to 16, the search process is performed in each subtree.

Here, in steps 6 to 10, the optimal old decoding solution of the p -th subtree is determined, and in steps 12 to 14, whether the old decoding algorithm based on neural network support ordering is terminated early is determined.

Simulation results

In the DPP-recovered decoding of an embodiment, the neural network structure includes 100,000 training data in 16x16 and 24x24 MIMO systems with QPSK modulation, and the SNR is set to a uniformly distributed random value in the range [4, 14], and DPP-recovered It is assumed that the design parameters λ ₁ and λ ₂ for minimizing the arc operation complexity are determined in Table 1 below in correspondence with the SNR.

[Table 1]

6 is an exemplary diagram showing the BER performance of DPP-SD and SE-SD according to an embodiment. there is.

7 and 8 are graphs showing the ratio of the computational complexity of DPP-SD to that of SE-SD. Referring to FIGS. 7 and 8, DPP-SD according to an embodiment shows the SNR of SE-old decoding at 11 dB or less. It can be seen that the computational complexity is low. In addition, it can be seen that the reduction ratios of the computational complexity of the DPP-recovered decoding and the SE-SD computational complexity of an embodiment for each of the 16×16 MIMO system and the 24×24 MIMO system are 36.7-43.2% and 50-59.2%.

9 is an overall flowchart illustrating an old decoding process used in the old decoding apparatus of the large-scale multiple input/output system shown in FIG. 1 . The old decoding method according to an embodiment is described. It may be performed in the old decryption device.

In the following description, the old decryption method according to an embodiment will be described with reference to an example performed by the old decryption apparatus shown in FIGS. 1 and 2, but must be performed by the old decryption apparatus of FIGS. 1 and 2 should not be limited to

Referring to FIG. 9 , the old decoding method according to an embodiment is performed in the old decoding apparatus 20 , and the optimizer 110 of the old decoding apparatus 20 receives the RF band signals received through a plurality of reception antennas. A received signal vector and a channel matrix are generated (step 201), and the received signal vector is arranged in a tree structure using the R matrix characteristic obtained through the QR decomposition technique of the generated channel matrix, and the received signal of the tree structure is constrained by the search space. Optimize the vector (step 202).

And, the learning unit 120 of the old decoding device 20 according to an embodiment learns the neural network through the optimized received signal vector of the optimizer 110 to minimize the deep path of all sub-trees of the uppermost layer (root layer). Estimate the path metric (step 203)

In addition, the minimum path metric deriving unit 130 of the old decoding apparatus 20 according to an embodiment derives the minimum path metric of the subtrees of the remaining layers except for the highest layer (step 204).

In addition, the minimum path metric sorting unit 140 of the old decoding apparatus 20 according to an embodiment sorts the minimum path metric of the subtree of the uppermost layer according to the size (step 205).

The old decryption solution search unit 150 of the old decryption apparatus 20 according to an embodiment sets a search radius according to the order of the sorted minimum path metrics, and searches for the old decrypted solution from the set search radius (step 207).

Thereafter, the early search termination unit 160 of the old decryption apparatus 20 according to an embodiment stops the search for the old decryption solution when the early termination condition is satisfied with the search radius set to the rearranged minimum path metric and the predetermined design parameters ( step 208).

Therefore, according to one embodiment, for the input vector of the neural network optimized in the tree structure, the minimum path metrics of the deep paths of all subtrees of the uppermost layer are estimated, then rearranged according to the size of the estimated minimum path metrics, and the reordered By searching the Sphere Decoding (SD) solution according to the order of the minimum path metric, the computational complexity of the old decoding can be fundamentally reduced.

Although the present invention has been described with reference to the embodiment shown in the drawings, this is merely an example, and those skilled in the art to which various modifications and equivalent other embodiments are possible. will understand Therefore, the technical protection scope of the present invention should be defined by the following claims.

Estimate the minimum path metrics of the deep paths of all subtrees of the top layer for an input vector of a neural network optimized with a tree structure, then reorder them according to the size of the estimated minimum path metrics, and recover according to the order of the reordered minimum path metrics A very large improvement in the accuracy and reliability of operation of the old decoding device and method of a large-scale multiple input/output system, which can fundamentally reduce the computational complexity of old decoding by searching for an SD (Sphere Decoding) solution, and further in terms of performance efficiency It is an invention with industrial applicability because it has sufficient potential for commercialization or sales of terminals related to 5G mobile communication, the next-generation growth engine, as well as to the extent that it can be clearly implemented in reality.

Claims (10)

a receiving device for generating a received signal vector and a channel matrix from an RF band signal received through a plurality of receiving antennas; and
The received signal vector is arranged in a tree structure using the R matrix characteristic obtained through QR decomposition of the generated channel matrix H , and the minimum path metric of the lower tree is estimated by learning the received signal vector of the arranged tree structure with a neural network. , including an old decoder that rearranges the estimated minimum path metric of the top layer in order of size, and then sequentially searches the old decryption solution from the subtree corresponding to the smallest minimum path metric,
The neural network is
Derived from the characteristics of the upper triangular matrix R obtained through QR decomposition of the channel matrix H

,

, and

vector based

with as input, the route metric

It is provided to learn by outputting
The old decryptor is
Set the search radius according to the order of the sorted minimum path metric, and a predetermined early search termination condition

, here,

is the design parameter,

is the search radius,

The old decryption apparatus of a large-scale multiple input/output system, characterized in that it is provided to stop the search for the old decryption solution when the minimum path metric, , is satisfied. The method of claim 1, wherein the old decoder comprises:
an optimizer for arranging a received signal vector in a tree structure by using the R matrix characteristic obtained through the QR decomposition technique of the generated channel matrix and optimizing the received signal vector of the tree structure through a search space constraint;
a learning unit for estimating the minimum path metric of the deep path of all subtrees by learning the neural network through the optimized received signal vector;
a minimum path metric sorting unit for arranging a minimum path metric of an estimated subtree of the uppermost layer in order of size; and
and a deciphering and deciphering solution search unit for setting a search radius according to the order of the estimated minimum path metric and searching for a deciphered and deciphered solution from the set search radius. The method of claim 2, wherein the old decoder comprises:
Further comprising a minimum path metric deriving unit for deriving a minimum path metric for the received signal vectors of the remaining layers except the top layer,
The minimum path metric alignment unit,
and generating a minimum path metric having the smallest size with respect to the derived minimum path metric of the minimum path metric deriving unit and arranging the generated minimum path metric in order of size. delete The method of claim 2, wherein the minimum path metric alignment unit comprises:
The reconstruction/decryption apparatus of a large-scale multiple input/output system, characterized in that it is arranged in ascending order according to the size of the generated minimum path metric. (a) a reception step of generating a reception signal vector and a channel matrix from RF band signals received through a plurality of reception antennas; and
(b) The received signal vector is arranged in a tree structure using the R matrix characteristic obtained through QR decomposition of the generated channel matrix, and the minimum path metric of the lower tree is obtained by learning the received signal vector of the arranged tree structure with a neural network. a step of estimating, rearranging the estimated minimum path metric of the highest layer in order of size, and then sequentially searching for the old and decoded solution from the subtree corresponding to the smallest minimum path metric,
The neural network is
Derived from the characteristics of the upper triangular matrix R obtained through QR decomposition of the channel matrix H

,

, and

vector based

with as input, the route metric

It is provided to learn by outputting
Set the search radius according to the order of the sorted minimum path metric, and a predetermined early search termination condition

, here,

is the design parameter,

is the search radius,

The old decryption method of a large-scale multiple input/output system, characterized in that it is provided to stop the search for the old decryption solution when the minimum path metric, , is satisfied. The method of claim 6, wherein (b) step,
arranging the received signal vector in a tree structure using the R matrix characteristic obtained through the QR decomposition technique of the generated channel matrix and optimizing the received signal vector of the tree structure through a search space constraint;
estimating a minimum path metric of a deep path of a subtree by training a neural network through an optimized received signal vector;
arranging the estimated minimum path metric of the subtree of the uppermost layer in order of size; and
and setting a search radius according to the order of the sorted minimum path metrics, and searching for the old decryption solution in the set search radius. The method of claim 6, wherein (b) step,
Further comprising the step of deriving a minimum path metric for the subtrees of the remaining layers except the top layer,
The step of arranging the minimum path metric in order of size is
A minimum path metric with the smallest size is generated for each of the estimated minimum path metric of the top layer and the derived minimum path metric of the remaining layers except for the top layer, and the generated minimum path metric is arranged in order of size. A method of recovery and decryption of large-scale multiple input/output systems. delete A recording medium in which a program for executing the method of any one of claims 6 to 8 is recorded and judged by a computer.

Images