KR20230097360A - Link performance prediction apparatus based on machine learning of mimo system - Google Patents

Abstract

According to the present technology, disclosed is a machine learning-based link performance estimation device in a MIMO system. A specific implementation example of this technology trains a machine learning model RF classifier built with a training data set that uses a channel matrix and a transmission signal-to-noise ratio as inputs and a block error rate as an output to predict an index value of a quantized block error rate for the inputted channel matrix and transmission signal-to-noise ratio, thereby increasing accuracy of link performance prediction compared to the existing link performance prediction result.

Description

Link performance estimator based on machine learning in MIMO system {LINK PERFORMANCE PREDICTION APPARATUS BASED ON MACHINE LEARNING OF MIMO SYSTEM}

The present invention relates to a machine learning-based link performance estimation apparatus in a MIMO system, and more particularly, trains a machine learning model that takes a channel matrix and a transmitted signal-to-noise ratio as inputs and outputs a block error rate (BLER), A technique for estimating link performance by predicting an index of a quantized block error rate for a channel matrix and a transmission signal-to-noise ratio input to a machine learning model.

In wireless communication systems, multi-input multi-output (MIMO) technology is widely used to achieve frequency efficiency performance that increases linearly with the number of antennas in order to meet the needs of high data rates.

Such a MIMO system applies a channel coding technique, and in particular, data can be transmitted with high frequency efficiency in a MIMO system through a turbo coding technique in which iterative detection and decoding are combined.

However, the MIMO system applies LA (Link Adaptation) technology that integrates AMC (Adaptive Modulation and Coding) to meet capacity in a fading channel. At this time, the LA technology applies channel coding rate and/or modulation based on channel state information. decide the way

In order to optimally determine the channel coding rate and modulation scheme, accurate estimation of link performance such as block error rate (BLER) is required. Accordingly, channel state information for multiple subchannels is compressed into a single effective signal-to-noise ratio, and then Link performance can be estimated from the compressed effective signal-to-noise ratio and the mapped block error rate. However, link performance estimation in a MIMO system has a problem in that accuracy decreases as the number of transmit antennas increases.

Accordingly, the present applicant trains a machine learning model that takes a channel matrix and a transmission signal-to-noise ratio as inputs and a block error rate (BLER) as an output. According to the prediction, we propose a method that can fundamentally improve the accuracy of link performance estimation.

Prior Document 1: B. M. Hochwald and S. ten Brink, "Achieving near-capacity on a multiple-antenna channel," IEEE Trans. Commun. vol. 51, no. 3, p. 389-399, 2003.

Therefore, the present invention trains a machine learning model that takes a channel matrix and a transmitted signal-to-noise ratio as input and a block error rate (BLER) as an output, and trains a block quantized for the channel matrix and transmitted signal-to-noise ratio input to the trained machine learning model. This is to improve the accuracy of link performance estimation by predicting the index of the error rate.

The object of the present invention is not limited to the above-mentioned object, and other objects and advantages of the present invention not mentioned above can be understood by the following description and will be more clearly understood by the examples of the present invention. It will also be readily apparent that the objects and advantages of the present invention may be realized by means of the instrumentalities and combinations thereof set forth in the claims.

An apparatus for estimating link performance based on machine learning in a MIMO system according to an embodiment of the present invention,

In the MIMO system,

Train a machine learning model with a training data set that has a channel matrix and a transmission signal-to-noise ratio as inputs and a block error rate as outputs;

It is characterized in that it further comprises a learning device for predicting the index of the quantized block error rate for the channel matrix and the transmission signal-to-noise ratio input to the trained machine learning model.

Preferably, the machine learning model is

Includes a Random Forest classifier,

Randomly selecting V features from among a plurality of features of the training data set of the machine learning model, and generating a decision tree from the most important features among the selected V features;

It may be provided to derive labels of more than half of the class labels generated by each tree, and assign the derived class labels of more than the majority to the final class.

Preferably, the learning device,

It may be provided to predict the quantized block error rate by inputting the channel matrix and the transmission signal-to-noise ratio input to the random forest classifier of the trained machine learning model.

Preferably, the learning device,

은 다음 식 11로 정의될 수 있다.Predicted quantized block error rate

can be defined by the following Equation 11.

[Equation 11]

은 훈련데이터 세트의 출력으로 블록 오류율이고,

는 양자화 연산이다. here,

is the block error rate as an output of the training data set,

is the quantization operation.

Preferably, the random press classifier,

The index of the quantized block error rate value may be predicted as a solution of Equation 12 below.

[Equation 12]

은 양자화된 블록 오류율 값의 인덱스이고, Q는 양자화 비트이며,

는 송신 안테나의 수이고,

은 수신 안테나의 수이며,

은 입력 벡터이며,

는 실수 행렬 공간이다.here,

is the index of the quantized block error rate value, Q is the quantization bit,

is the number of transmit antennas,

is the number of receiving antennas,

is the input vector,

is the real matrix space.

According to an embodiment, a machine learning model having a channel matrix and a transmission signal-to-noise ratio as input and a block error rate (BLER) as an output is trained, and the channel matrix and transmission signal-to-noise ratio input to the trained machine learning model are quantized. By predicting the index of the block error rate, the accuracy of link performance prediction can be improved compared to the link performance prediction results using the existing PIC-ZF, MRC-MMSE, and MMIB methods.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and together with the detailed description of the present invention serve to further understand the technical idea of the present invention, the present invention is the details described in such drawings should not be construed as limited to
1 is a configuration diagram of an MMIMO system according to an embodiment.
2 is a diagram showing the configuration of the random forest classifier of FIG. 1;
3 is a diagram showing a processing algorithm of the learning device of FIG. 1;
4 is diagrams showing quantized BLER according to the number of transmit antennas and receive antennas of the MIMO system of FIG. 1;
5 to 7 are graphs showing Mean-Square Error (MSE) of a MIMO system according to an embodiment.
8 to 10 are graphs showing a cumulative distribution function (CDF) of a bias error (BE) of a MIMO system according to an embodiment.

Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only, and may be modified and implemented in various forms. Therefore, the embodiments are not limited to the specific disclosed form, and the scope of the present specification includes changes, equivalents, or substitutes included in the technical spirit.

Although terms such as first or second may be used to describe various components, such terms should only be construed for the purpose of distinguishing one component from another. For example, a first element may be termed a second element, and similarly, a second element may be termed a first element.

It should be understood that when an element is referred to as being “connected” to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle.

Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate that the described feature, number, step, operation, component, part, or combination thereof exists, but one or more other features or numbers, It should be understood that the presence or addition of steps, operations, components, parts, or combinations thereof is not precluded.

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 commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and unless explicitly defined in this specification, it should not be interpreted in an ideal or excessively formal meaning. don't

Hereinafter, the present invention will be described in detail by describing preferred embodiments of the present invention with reference to the accompanying drawings.

1 is a configuration diagram of an MMIMO system according to an embodiment, FIG. 2 is a configuration diagram of a random forest classifier of FIG. 1, and FIG. 3 is a diagram showing a processing algorithm of the learning device of FIG.

1 to 3, the MIMO system according to an embodiment includes a transmitting end 1 and a receiving end 2, and the transmitting end 1 encodes a data stream to be transmitted, interleaves it, and modulates the data stream. It transmits through transmit antennas on a fading channel.

Then, the receiving end 2 performs Iterative Detection and Decoding (IDD).

That is, the reception signal z of the narrowband block fading channel received through the reception antennas can be expressed by Equation 1 below.

[Equation 1]

은

벡터의 송신 안테나를 통해 전송된 데이터 심볼이고,

은 평균이 0이고 분산이

인

을 포함하는

벡터의 복소수 가우시안 잡음 신호이다. here,

silver

is a data symbol transmitted through the transmit antenna of the vector;

has a mean of 0 and a variance of

person

containing

The vector's complex Gaussian noise signal.

,

이고, P _T 는 데이터 심볼에 대한 평균 전송 전력이며, 전송 신호대 잡음비 SNR은

로 정의하면, 전송 데이터 심볼 s의 각 원소는 P 포인트의 복소수 콘스텔레이션(Constellation)

로부터 독립적으로 생성된다.here,

,

, P _T is the average transmission power for data symbols, and the transmission signal-to-noise ratio SNR is

If defined as, each element of the transmission data symbol s is a complex constellation of P points

generated independently from

는 송수신 간의 채널 행렬로 j번째 송신 안테나와 i번째 수신 안테나 간의 복소수 채널 이득은 h _i,j 이다.here,

is a channel matrix between transmission and reception, and the complex channel gain between the j -th transmit antenna and the i -th receive antenna is h _i,j .

The receiving end 2 includes a decoder 21 that detects a maximum-likelihood (ML) solution, which is a lattice point having the smallest Euclidean distance from the entire sphere of the transmitted signal to the received signal.

That is, the sphere decoder 21 searches for each point within a sphere defined by Equation 2 below to find a maximum-likelihood (ML) solution. Here, the ML solution is a lattice point having the smallest Euclidean distance from the entire sphere of the transmitted signal to the received signal.

[Equation 2]

는 구(sphere)의 중심이며, r은 검색 반경이다. here,

is the center of the sphere and r is the search radius.

The old decoder 21 modifies the LSD (List Sphere Decoder) disclosed in Prior Document 1 to provide an Nc candidate list for data symbol s inside the sphere.

으로 정의되고, 초기 반지름 r은 하기 식 3에 의거 선택된다.In LSD, the center of the sphere is

, and the initial radius r is selected based on Equation 3 below.

[Equation 3]

는 실제 데이터 심볼 s을 높은 확률로 추출할 수 있도록 선택된다. here,

is selected so that the actual data symbol s can be extracted with high probability.

In addition, the receiving end 2 performs deinterleaving 22 and channel decoding 23 based on the derived reconstruction solution, and then performs interleaving 24 on the channel decoding result. Accordingly, the receiving terminal 2 is implemented with an Iterative Detection and Decoding (IDD) structure.

On the other hand, the receiving end 2 further includes a learning device 100 that predicts the index of the quantized block error rate using the random forest classifier of the machine learning model.

및 전송 신호대 잡음비 SNR를 입력으로 하고 블록 오류율(BLER)을 출력으로 하는 기계학습모델의 랜덤 포레스트(Random Forest) 분류기를 훈련하고 훈련된 기계학습모델로 입력된 채널 행렬 및 전송 신호대 잡음비에 대한 양자화된 블록 오류율의 인덱스를 예측하도록 구비되고, 출력된 양자화된 블록 오류율의 인덱스로 링크 성능은 추정될 수 있다. The learning device 100 is a channel matrix

and a random forest classifier of a machine learning model that takes the transmitted signal-to-noise ratio SNR as input and the block error rate (BLER) as output, and quantizes the input channel matrix and transmitted signal-to-noise ratio to the trained machine learning model. It is provided to predict the block error rate index, and the link performance can be estimated with the output quantized block error rate index.

Depending on the link performance estimation result, a modulation and coding scheme (MCS) level of the transmitter 1 may be determined as an optimal state.

Here, the random forest (RF) classifier of the machine learning model is a tree-based machine learning algorithm and can derive accurate prediction results by combining prediction results of multiple machine learning algorithms. Here, each decision tree of the tree-based machine learning algorithm is trained by a randomly selected subset of the entire training data set.

That is, referring to FIG. 2, the trees Tree-1, ... Tree-N of the RF classifier are executed in parallel, and the parallelly performed trees Tree-1, ... Tree-N are classes Class A, Class B is voted on, and the final class with the majority of votes in the decision tree is determined. As the number of RF classifier trees increases, computational complexity increases, but more accurate and stable prediction results can be output.

를 포함한다. 여기서, 기계학습모델의 입력

은 L개의 특징으로 구성된 벡터이고 출력 y _m 은 m번째 학습 데이터 세트의 클래스 레이블로 1부터 K 값이다. The machine learning model of the learning device 100 is M training data

includes Here, the input of the machine learning model

is a vector of L features and the output y _m is the class label of the mth training data set and is a value from 1 to K.

를 무작위로 선택하고, 여기서

이고,

으로 일반화된다. The RF classifier selects V features out of a total of L features, i.e.

is randomly selected, where

ego,

generalized to

In addition, the RF classifier must generate a decision tree, and uses the Gini index as a split criterion during classification to obtain the most important features for the best split of the tree. At this time, a low value of the Gini index means that it plays a more important role in dividing the training data.

을 결정한 다음 무작위로 만들어진 트리가 V 특징에 대해 클래스 레이블이 완전히 생성될 때까지 반복된다.After this, the RF classifier has the most important features

After determining , the randomized tree is iterated until class labels are completely generated for V features.

That is, the RF classifier generates each tree with class labels Class A, Class B, ... Class N as output for N classification trees, Tree-1, ... Tree-N, and N classification trees. Predicted class labels Class A, Class B, ... Class N are assigned a final class to the class labels that obtain a majority (Majority Voting).

Hereinafter, a series of processes of estimating the link performance of a MIMO system with the RF classifier of the trained machine learning model will be described with reference to FIG. 3 .

채널 행렬과 전송 신호대 잡음비 SNR을 입력으로 하고 블록 오류율 BLER 값을 출력으로 하는 훈련된 기계학습모델의 훈련데이터 세트에 의거 양자화된 블록 오류율의 인덱스를 추정한다.A MIMO system to which an embodiment is applied

The index of the quantized block error rate is estimated based on the training data set of the trained machine learning model that takes the channel matrix and the transmitted signal-to-noise ratio SNR as input and the block error rate BLER value as output.

및 전송 신호대 잡음비 SNR

m=1, 2, … M 은 무작위로 훈련데이터 세트를 생성되고, 생성된 훈련데이터 세트는 다음 식 4로 나타낼 수 있다.For example M channel matrices

and transmit signal-to-noise ratio SNR

m=1, 2, … M randomly generates a training data set, and the generated training data set can be represented by Equation 4 below.

[Equation 4]

과 출력 y_m은 입력 x_m에 대응되는 실제 블록 오류율 BLER 값이다. 여기서,

와

는 복소수의 실수부와 허수부이고, 총

개의 특징의 입력 x_m의 특징

의 크기는

이다.here, enter

and the output y _m is the actual block error rate BLER value corresponding to the input x _m . here,

and

is the real and imaginary parts of the complex number, and the total

Features of input x _m of features

the size of

am.

는 시뮬레이션을 통해 얻을 수 있다.Also a vector of actual BLER values

can be obtained through simulation.

를 Q 비트로 양자화한다. 여기서, BLER

은 로그함수에 의해 양자화되며, 이에 양자화된 BLER값

은 하기 식 5에 의해 얻을 수 있다.To apply the training data set to the RF classifier, BLER

is quantized with Q bits. Here, BLER

is quantized by the logarithmic function, and the quantized BLER value

can be obtained by the following formula 5.

[Equation 5]

는 양자화 연산기이다. here,

is a quantization operator.

의 BLER 값을 고려하면, 양자화된 BLER 값

은

이다.Also range

Considering the BLER value of , the quantized BLER value

silver

am.

의 인덱스 m은 RF 분류기에 의해 추정될 수 있다.Quantized BLER values in the training data set

The index m of can be estimated by the RF classifier.

That is, for input x _m , the solution obtained by the RF classifier can be formulated as in Equation 6 below.

[Equation 6]

은 양자화된 BLER 값의 인덱스이고 양자화된 BLER 값

은

으로 나타낼 수 있다.here

is the index of the quantized BLER value and is the quantized BLER value

silver

can be expressed as

4 is a graph showing the predicted block error rate BLER and the actual BLER in Additive White Gaussian Noise (AWGN). Referring to FIG. 4, the BLER predicted through training using the RF classifier of the machine learning model corresponds to the actual BLER It can be confirmed that it is close, and it can be seen that the accuracy of the link performance prediction based on machine learning according to an embodiment is higher than the existing link performance prediction result.

에 대해 [-2, 0]의 SNR 범위에서 PIC-ZF, MRC-MMSE 및 MMIB 방식의 평균 MSE는 2 Х 2 MIMO의 경우 각각 0.160, 0.215, 0.262인 반면, 일 실시예에 의거한 MSE는 양자화 비트 Q = 4, 6 및 8 각각에 대해 각각 0.130, 0.119 및 0.107임을 알 수 있으며, 이에 도 5 및 도 6를 참조하면, 일 실시예에 의거한 2 Х 2 및 3 Х 3 MIMO 시스템의 경우 전체 BLER 범위에서 MSE는 기존 방식의 MSE와 비교하여 우수하거나 유사 성능이 달성됨을 확인할 수 있다.In addition, each of FIGS. 5 to 7 shows the mean-square error (MSE) based on an embodiment and the existing PIC-ZF (Perfect Interference Cancellation and Zero-Forcing), MRC-MMSE (Maximum Ratio Combining and Minimum Mean-Square Error), MMIB (Mean Mutual Information per Bit) method. Referring to FIG. 4, the size is a step size of 0.5.

In the SNR range of [-2, 0], the average MSEs of PIC-ZF, MRC-MMSE, and MMIB schemes are 0.160, 0.215, and 0.262, respectively, in the case of 2 Х 2 MIMO, whereas the MSE according to an embodiment is quantized It can be seen that bit Q = 0.130, 0.119 and 0.107 for each of 4, 6 and 8, respectively. Referring to FIGS. 5 and 6, in the case of a 2 Х 2 and 3 Х 3 MIMO system according to an embodiment, the entire It can be seen that in the BLER range, MSE achieves superior or similar performance compared to conventional MSE.

In addition, referring to FIG. 7, in the case of PIC-ZF, MRC-MMSE, and MMIB, average MSEs are 0.173, 0.237, and 0.292, whereas quantization bits Q = 0.0986 for each of 4, 6, and 8 according to an embodiment, It can be seen that they are 0.0769 and 0.0650, and thus it can be seen that the difference between the MSE according to an embodiment and the MSE of the conventional method gradually increases in the 4Х4 MIMO system.

에 의해 연산된 BE(bias error)의 CDF(cumulative distribution function)를 기존의 PIC-ZF, MRC-MMSE, MMIB 방식에 대핸 보인 그래프로서, 도 7에 도시된 PIC-ZF 및 MRC-MMSE 방식의 CDF에 따르면, 실제 BLER 값과 예측 값 간의 차이의 약 78%가 BE 범위 [-0.2, 0.2]에 있는 반면 일 실시예에 의거한 CDF에 따르면, 실제 BLER 값과 예측 값 간의 차이의 82%가 BE 범위 [-0.2, 0.2]내에 있음을 확인할 수 있다.8 to 10 are 2 Х 2, 3 Х 3 and 4 Х 4 in each MIMO system

As a graph showing the CDF (cumulative distribution function) of the BE (bias error) calculated by , for the existing PIC-ZF, MRC-MMSE, and MMIB methods, the CDF of the PIC-ZF and MRC-MMSE methods shown in FIG. , about 78% of the difference between the actual BLER value and the predicted value is in the BE range [-0.2, 0.2], whereas according to the CDF according to an embodiment, 82% of the difference between the actual BLER value and the predicted value is BE It can be confirmed that it is within the range [-0.2, 0.2].

Referring to FIG. 10, while there is about 80% of the difference between the actual BLER value and the predicted value of the PIC-ZF and MRC-MMSE schemes in the same BE range, the difference between the actual BLER value and the predicted value according to an embodiment is about It can be seen that 90% are present.

Accordingly, an embodiment trains an RF classifier of a machine learning model built with a training data set having a channel matrix and a transmitted signal-to-noise ratio as input and a block error rate as an output, and a block error rate quantized for the input channel matrix and transmitted signal-to-noise ratio. By predicting the index value of , it is possible to improve the accuracy of link performance prediction compared to the existing link performance prediction result.

As described above, although the embodiments have been described with limited drawings, those skilled in the art can apply various technical modifications and variations based on the above. For example, the described techniques may be performed in an order different from the method described, and/or components of the described system, structure, device, circuit, etc. may be combined or combined in a different form than the method described, or other components may be used. Or even if it is replaced or substituted by equivalents, appropriate results can be achieved.

Predict the index value of the quantized block error rate for the input channel matrix and transmitted SNR by training the RF classifier of the machine learning model built with the training data set with the channel matrix and transmitted SNR as input and the block error rate as output. As a result, in MIMO systems that can improve the accuracy of link performance prediction, it can bring about great progress in terms of operational accuracy and reliability, and furthermore, performance efficiency for machine learning-based link performance estimation devices. can be applied, and as the core technology of the MIMO system is secured, environmental monitoring in related industries can be actively utilized, and the possibility of commercialization or business of the wireless communication system is sufficient and it can be clearly implemented in reality. It is an invention with potential for use.

1: transmitting end
2: receiving end
21: Old Bok Unit
22: learning device

Claims (5)

In the MIMO system,
Train a machine learning model with a training data set that has a channel matrix and a transmission signal-to-noise ratio as inputs and a block error rate as outputs;
A machine learning-based link performance estimation apparatus in a MIMO system, further comprising a learning apparatus for predicting an index of a quantized block error rate for a channel matrix and a transmission signal-to-noise ratio input to the trained machine learning model. The method of claim 1, wherein the machine learning model
Include a random prest classifier,
Randomly selecting V features from among a plurality of features of the training data set of the machine learning model, and generating a decision tree from the most important features among the selected V features;
An apparatus for estimating link performance based on machine learning in a MIMO system, characterized in that it is provided to derive labels of more than half of the class labels generated by each tree, and assign the class labels of more than the majority of the derived class labels to the final class. The method of claim 2, wherein the learning device,
A machine learning-based link performance estimation device in a MIMO system, characterized in that it is provided to predict a quantized block error rate with a channel matrix and a transmission signal-to-noise ratio input to a random prest classifier of a trained machine learning model. The method of claim 3, wherein the learning device,
Predicted quantized block error rate

Is a machine learning-based link performance estimator in a MIMO system, characterized in that defined by the following Equation 11.
[Equation 11]

here,

is the block error rate, which is the output of the training data set,

is the quantization operation. The method of claim 4, wherein the random forest classifier,
An apparatus for estimating link performance based on machine learning in a MIMO system, characterized in that it is provided to derive the index of the quantized block error rate value as a solution of Equation 12 below.
[Equation 12]

here,

is the index of the quantized block error rate value, Q is the quantization bit,

is the number of transmit antennas,

is the number of receiving antennas,

is the input vector,

is the real matrix space.

Images