KR20070051576A - Minimize error propagation of channel equalizer apparatus and method thereof - Google Patents

Abstract

A channel equalizer and a method for minimizing error propagation are proposed. According to the present invention, a channel is estimated for a received signal, a reference tap and a position of the reference tap are selected from the estimated channel, and based on the estimated channel, the reference tap, and the position of the reference tap, a tap coefficient is calculated. Estimate and equalize the received signal using the estimated tap coefficients.

Error Propagation, Channel Estimation, Reference Tap Position, Channel Equalization

Description

Minimized error propagation of channel equalizer apparatus and method

1 is a block diagram of a channel equalizer according to an embodiment of the present invention;

2 is a block diagram of a channel estimation unit according to an embodiment of the present invention;

3 is a detailed configuration diagram of the training equalizer according to one embodiment of the present invention;

4 is a block diagram illustrating a determination unit according to an embodiment of the present invention;

5 is a flowchart provided to explain a channel equalization method according to an embodiment of the present invention;

6A and 6B are graphs illustrating root mean square error (RMSE) in Brazilian A and D channels according to an embodiment of the present invention;

7A to 7D are graphs showing signal to interface plus ratios (SINRs) of equalizers in channels A, C, D, and E of Brazil according to an embodiment of the present invention, and

8A to 8C are graphs illustrating SER (Symbol Error Rate) of the equalizer in the Brazilian E and C channels according to an embodiment of the present invention.

Description of the main parts of the drawing

301: buffer 303: channel estimator

305: training equalizer 307: feed-forward filter (FFF)

309: Subtractor 311: Determination unit

313: Feed-back filter (FBF) 315: Decision feedback equalizer

501: correlation unit 503: noise removing unit

505: LMS section 507: Selection section

The present invention relates to a channel equalizer and a method thereof, and more particularly, to a channel equalizer and a method for minimizing error propagation.

Efforts have been made to overcome performance degradation of the equalizer due to error propagation, such as burst error floor in wireless communication systems. This error propagation is mainly caused by a bad decision symbol passing through a feedback filter (FBF). Therefore, most of the proposed techniques for preventing error propagation are to reduce the probability of decision error by applying DFE-based Viterbi-based DFE. However, error propagation can also be caused by large precursors. This error propagation is a feed-forward filter in relation to the reference tap (symbol timing of the DFE) obtained from the estimated channel impulse response (CIR) and the position of the reference tap (e.g., Decision error propagation). Caused by (FFF: Feedforward filter).

Conventional reference taps can be obtained by symbol recovery performed prior to equalization. Alternatively, the reference tap could be selected during the training period of the equalizer, but this method is too complex and difficult to use given the wide area equalization for the wireless communication system. Typically the reference tap is fixed at the center of the FFF, to the far right of the FFF, or anywhere in the FFF found through extensive simulation. While the techniques for fixing the reference taps are very simple, performance depends on channel conditions such as delay spread and the size of the multipath. Thus, these techniques are not appropriate for unknown channel environments.

SUMMARY OF THE INVENTION The present invention has been proposed to meet the needs as described above, and provides a channel equalization apparatus and method for minimizing error propagation by changing the position of a reference tap according to the characteristics of a channel.

In order to achieve the above object, the channel equalization apparatus which minimizes error propagation according to the present invention comprises: a channel estimator for estimating a channel with respect to a received signal and selecting a reference tap and a position of the reference tap from the estimated channel; A training equalizer (DFE) for estimating a tap coefficient based on the estimated channel, the reference tap, and the position of the reference tap; And an equalization unit (DFE: Decision Feedback Equalizer) for equalizing the received signal by using the estimated tap coefficient.

The channel estimator may include a correlator configured to estimate a channel by applying a correlation algorithm to a received signal; A noise removing unit for removing noise of the channel estimated by the correlation unit; An LMS unit for estimating a channel by applying a moving-average variable step-size (MA-VSS) LMS algorithm to a received signal and the channel from which the noise is removed; And a selector for selecting a reference tap and a reference tap position from the channel estimated by the LMS.

Herein, the selector selects a channel having a size larger than a threshold value as a reference tap among the channels estimated by the LMS, and selects a channel that is larger than 1/2 of the size of the selected reference tap signal. It is preferable to select by position.

In order to achieve the above object, the channel equalization method of minimizing error propagation according to the present invention comprises the steps of: a) estimating a channel with respect to a received signal and selecting a reference tap and a position of the reference tap from the estimated channel; b) estimating a tap coefficient based on the estimated channel, reference tap, and position of the reference tap; And c) equalizing the received signal using the estimated tap coefficients.

Here, the step a) comprises: estimating a channel by applying a correlation algorithm to the received signal; A noise removing step of removing noise of the estimated channel; Estimating a channel by applying a moving-average variable step-size (MA-VSS) LMS algorithm to the received signal and the channel from which the noise is removed; And selecting a reference tap and a reference tap position among channels estimated by applying the LMS algorithm.

Other objects and further features of the present invention will become apparent from the following detailed description when read in conjunction with the accompanying drawings.

1 is a block diagram of a channel equalization apparatus according to an embodiment of the present invention.

Referring to FIG. 1, the channel equalizer includes a buffer 301, a channel estimator 303, a training equalizer 305, and an equalizer 315.

The buffer 301 may store the received signal.

The channel estimator 303 may receive the stored signal and estimate a channel by applying a correlation algorithm. The channel estimator 303 according to an embodiment of the present invention may select a reference tap and a position of the reference tap from the estimated channel.

In detail, the channel estimator 303 estimates a channel by applying a correlation algorithm to the received signal to remove noise of the estimated channel, and performs MA-VSS (Moving) on the received signal and the channel from which the noise is removed. Average Variable Step-Size) A channel is estimated by applying an LMS algorithm, and among the estimated channels, a reference tap and a reference tap position can be selected by a predetermined criterion.

The training equalizer 305 may calculate the tap coefficient of the estimated channel. Preferably, the training equalizer 305 may estimate the tap coefficient based on the channel, the reference tap, and the position of the reference tap estimated by the channel estimator. The tap coefficients estimated by the training equalizer 305 are tap coefficients 30 7 of the feed-forward filter and tap coefficients 313 of the feed-back filter.

The equalizers 307 and 315 may equalize the received signal using tap coefficients estimated by the training equalizer 305.

The equalizers 307 and 315 remove inter-symbol interference (hereinafter referred to as "ISI"). These ISIs are typically divided into precursor ISIs and post cursor ISIs.

Referring to FIG. 1, the equalizer 315 may include a feed-forward filter (“FFF”) filter 307 and a decision feedback equalization unit (“DFE”) 315. ).

FFF 307 filters the signal received from the transport channel to remove precursor ISI, while DFE mainly removes postcursor ISI. The DFE not only removes the post cursor ISI but also the free cursor ISI.

The FFF 307 uses the FFF filter tap coefficients provided by the training equalizer 305.

The DFE 315 may be, for example, a well-known minimum mean square error decision feed-back equalizer (MMSE-DFE equalizer). Referring to FIG. 1, the DFE 315 includes a subtraction unit 309, a determination unit 315, and a feed-back filter (hereinafter referred to as “FBF”) 313.

The subtraction unit 309 subtracts and outputs the signal filtered by the FFF 307 and the signal filtered by the FBF 313. The output signal is precursor ISI from the signal received at the receiver. And post cursor ISI become the removed signal.

The determination unit 315 may determine the signal output from the subtractor 309 as a symbol and output the determination value. As the determination unit 315, a well-known Schimtt trigger circuit may be used, for example. The determination unit 315 may determine the signal output from the subtractor 309 as an eight-level symbol. The equalized signal may be provided to a decoder (not shown) for further processing.

The FBF 313 receives the signal determined by the determination unit 315 and performs filtering. The FBF 313 may be expressed by, for example, the following equation.

Where g _i is the filter coefficient of the FFF 313, S _{k -N} is the output of the determination unit 311, L is the order of the FBF 313, and N is the order of the FFF. .

2 is a block diagram of the channel estimation unit 303 according to an embodiment of the present invention.

Referring to FIG. 2, the channel estimator 303 includes a correlation unit 501, a noise removing unit 503, an LMS unit 505, and a selection unit 507.

The channel estimating unit 303 according to the present invention preferably performs channel estimation under the condition that the range of channel estimation is sufficiently large and the channel is accurately estimated.

The correlator 501 estimates a channel by applying a correlation algorithm to the received signal, the noise remover 503 removes noise from the channel estimated by the correlator 501, and the LMS unit 505 A channel is estimated by applying a signal received by the buffer 301 and a channel moving-average variable step-size (MA-VSS) LMS algorithm from which the noise is removed by the noise removing unit 503, and the selecting unit 507 ) Selects a reference tap and a reference tap position among channels estimated by the LMS unit 505.

Before describing the components of the channel estimator 303 in more detail, some necessary details will be described.

The frame structure of an 8 VSB DTV system includes two data fields, each data field including a segment of one field and 313 data segments. Each data segment contains 832 symbols, in which the first segment of each field is used for field sync or training sequence (TS). As can be seen from the structure of this data field, the training sequence is relatively smaller than the data sequence DS.

The signal received by the equalizer through the white noise (AWGN) and the multipath channel may be represented by Equation 1 below.

과

은 각각, 시간

(T는 심볼 주기)에서 샘플링된 수신 신호와 전송 신호이다.

는 물리적인 멀티패스 채널, 전송 필터, 및 수신 필터를 포함한 채널 임펄스 응답(CIR)을 나타내고,

는 전체 채널 메모리 길이 를 나타내고,

은 AWGN 샘플이며, 이것은 통계적으로 i.i.d로 가정된다. 전송 시퀀스

은 2-레벨 펄스 진폭 변조(2-PAM) 심볼

(예를 들면 훈련 열(TS)과 세그먼트 동기)으로 구성되며, 8-PAM 데이터 시퀀스(DS)

은 하기 식 2와 같다.here

and

Each time

(T is a received signal and a transmitted signal sampled in a symbol period).

Represents the channel impulse response (CIR), including the physical multipath channel, the transmit filter, and the receive filter,

Represents the total channel memory length,

Is an AWGN sample, which is assumed to be iid statistically. Transmission sequence

Is a 2-level pulse amplitude modulation (2-PAM) symbol

(E.g., training sequence (TS) and segment sync), 8-PAM data sequence (DS)

Is equal to the following Equation 2.

 Here, A represents a PAM symbol.

은 기준 탭

에서 계산되며, 이 기준탭의 위치는 FFF의 가장 오른쪽으로부터 d 만큼 떨어진 곳에 위치된다. 그리고,

은 하기 식 (3)으로 표현될 수 있다.Print

Silver base tab

The position of this reference tab is located d from the rightmost side of the FFF. And,

Can be represented by the following formula (3).

,

}는 각각

, ..., -d, 및 m=1, ...,

에 대한 FFF 및 FBF의 계수이다. {

,

}는 각 필터의 탭 길이이고,

은 판정부에 의해 판정된 심볼을 나타낸다.here {

,

} Are each

, ..., -d, and m = 1, ...,

Is the coefficient of FFF and FBF for. {

,

} Is the tab length for each filter,

Represents a symbol determined by the determination unit.

Now, the components of the channel estimator 303 will be described in more detail.

The correlator 501 estimates a channel by applying a correlation algorithm defined by Equation 4 to the signal received by the buffer 301. The correlation algorithm defined as Equation 4 can satisfy the condition that the range of the channel estimation is sufficiently large, provided that the training sequence TS is sufficiently larger than the channel memory.

은 CIR(채널 임펄스 응답: Channel Impulse Response)을 나타내고, r_a(0)은 상관관계 놈(correlation norm)

을 나타내며, (*)는 공액 복소수(complex conjugate)를 나타낸다. 한편, 상관 알고리즘만으로 채널을 추정하는 것은 상관 노이즈 때문에 채널 추정의 정확성이 떨어질 수 있다. 따라서, 본 실시 예에서는 다음에서 설명하는 과정이 채널추정부(303)에서 추가적으로 수행된다. here

Represents CIR (Channel Impulse Response), and r _a (0) is the correlation norm

(*) Represents a complex conjugate. On the other hand, estimating a channel using only the correlation algorithm may reduce the accuracy of channel estimation due to the correlation noise. Therefore, in the present embodiment, the process described below is additionally performed by the channel estimator 303.

The noise removing unit 503 removes noise from the CIR estimated by the correlation unit 501. The influence of the correlation noise can be expressed by the following expression (5).

Equation 5 inserts Equation 1 above into Equation 4.

의 크기가 상관 노이즈보다 작으면,

은 상관 노이즈 아래 묻힐수 있다(buried). 정확한 채널 추정을 위해서는, 묻힌

이 복원되어야 한다. In Equation 5, the first part represents the desired CIR, and the second and third parts represent correlation noise. if

If the magnitude of is smaller than the correlation noise,

Can be buried under correlation noise. For accurate channel estimation, buried

This must be restored.

복원을 위해서, 임계치 방법(threshold method)과 적응적 채널 추정(adaptive channel estimation)을 사용한다.According to an embodiment of the present invention,

For reconstruction, a threshold method and adaptive channel estimation are used.

First, the correlation noise may be removed by a threshold method defined by Equation 6 below.

|이 T _ch 보다 작으면,

은 제로로 설정된다. 수학식 6에서, 임계값 T _ch 의 크기가 수학식 1의

의 성능에 영향을 주지 않으면서 결정되어야 한다. Where T _ch is a predetermined threshold for removing correlation noise. If n = 0, ..., Nc-1 |

Is less than T _ch ,

Is set to zero. In Equation 6, the magnitude of the threshold value T _ch is

Should be determined without affecting the performance of the system.

By using the thresholded correlation algorithms of Equations 4 and 6 above, only plausible channel taps can be left.

은 상관 노이즈로 취급하여 제거한다. 상세하게는, 상술한 수학식 4과 수학식 6의 임계화된 상관 알고리즘을, 상기 상관부(501)가 추정한 채널에 적용하여 상관 노이즈를 제거한다.The noise removing unit 503 is smaller than the size of the threshold value T _ch .

Is treated as correlation noise and removed. In detail, the thresholded correlation algorithms of Equations 4 and 6 described above are applied to the channel estimated by the correlation unit 501 to remove correlation noise.

On the other hand, there is a need to restore discarded channel taps.

The Least-mean-square (LMS) unit 505 may restore the tap of the discarded channel by applying a Moving Average Variable step-size (MA-VSS) LMS algorithm. In the Least-mean-square (LMS) unit 505, the training sequence TS is used repeatedly to allow the LMS algorithm to fully converge.

The LMS channel unit 505 may be defined by Equation 7 below.

은, 시간 n에서dml 채널 추정기의 N_f×1 계수 백터이며,

는 식(4)가 적용된 후 남은 채널 탭으로 초기화 된다. a(n)은 N_f×1 입력 훈련렬(TS) 벡터이며, r(n)은 수신받은 신호 샘플들이다. 윗 첨자 T는 백터의 트랜스포즈(transpose)이고, e(n)은 에러 추정이다. 식 7에서, μ(n)은 하기 수학식 8에 정의된바 와 같은 MA-VSS(Moving-average variable step-size) 알고리즘을 이용하여 업데이트 된다.here

Is the N _f × 1 coefficient vector of the dml channel estimator at time n,

Is initialized to the remaining channel tap after equation (4) is applied. a (n) is an N _f x1 input training sequence (TS) vector, and r (n) is received signal samples. The superscript T is the transpose of the vector and e (n) is the error estimate. In Equation 7, μ (n) is updated using a moving-average variable step-size (MA-VSS) algorithm as defined in Equation 8 below.

인 제약조건을 나타내고, {α, β, γ}는 LMS 알고리즘의 추적(tracking) 특성을 제어하기 위한 상수들이다. L은 MA 프로세싱을 위한 창 크기(window size)로서, 그 값은 충분히 커야 한다. 에러 거동(behavior)이 평균을 함(average function)으로써 충분히 획득될 수 있어야 하기 때문이다. 한편으로는, 훈련 열(TS)이 반복되는 동안에는 스텝 사이즈가 충분히 업데이트 되도록, L값은 작아야 한다. 결국, 스텝 사이즈가 에러를 자승함(squared error)으로써 업데이트 될 때, 수학식 8에서의

부분에 대한 평균화에 의해, 수학식 1에서의 독립적인(independent) 노이즈의 효과가 부드럽게 감소될 것이다. MA-VSS 알고리즘은, LMS 채널 추정기를 위해 훈련 열(TS)의 적은 횟수의 반복과 빠른 수렴 속도를 제공한다.Where the bracket is

, {Α, β, γ} are constants for controlling the tracking characteristics of the LMS algorithm. L is the window size for MA processing, which must be large enough. This is because the error behavior must be sufficiently obtained by an average function. On the one hand, the L value must be small so that the step size is sufficiently updated while the training sequence TS is repeated. As a result, when the step size is updated by squared error,

By averaging over the part, the effect of independent noise in equation (1) will be smoothly reduced. The MA-VSS algorithm provides a small number of iterations of the training sequence (TS) and a fast convergence rate for the LMS channel estimator.

The selector 507 selects a reference tap and a reference tap position among the channels estimated by the LMS unit 505.

First, it is assumed that the tap length N _b of FBF satisfies Equation 9 below. By using this equation (9), the ISI caused by the unboiled portion of the CIR will not occur.

에서 기준 탭의 위치를 나타낸다.Where Δ is

Indicates the position of the reference tab.

The selector 507 may select the reference tap by using a simple technique as shown in Equation 10 below.

은

중에서 선택된 기준 탭을 나타낸다.

은

를 선택하기 위한 임계값들

의 룩업 데이블이고, N은 임계값의 개수(個數)이다.here,

silver

Selected from among the criteria tab.

silver

Thresholds for Selecting

Is a lookup table of N, where N is the number of thresholds.

The thresholds are chosen such that the influence of large precursors is minimized, for example under Equation 11 under the Brazilian channel.

은, 채널 상기 수학식 10을 만족하는 채널 탭

이 k=1, 2,에 대해서 존재하면,

으로 선택된다. 만약 존재하지 않는다면, 가장 크기가 큰

이

으로 선택된다.

Is a channel tap that satisfies the above equation (10).

If is present for k = 1, 2,

Is selected. If not present, the largest

this

Is selected.

In order to minimize the error propagation caused by the large precursor, the position of the reference tap in the FFF can be selected by Equation 12 below.

은, 상기 조건을 만족하는

과

간의 샘플 딜레이를 나타낸다. 따라서, FFF에서의 판정 딜레이는

이다. 수학식 10를 살펴보면,

는

에 기초하여 변화됨을 알 수 있다.here,

To satisfy the above conditions

and

Sample delay in the liver. Therefore, the decision delay in FFF is

to be. Looking at Equation 10,

Is

It can be seen that based on the change.

Figure 3 is a detailed block diagram of the training equalizer according to an embodiment of the present invention.

The training equalizer 305 may calculate tap coefficients of the FFF and the FBF based on the positions of the channel, the reference tap, and the reference tap estimated by the channel estimator 303. The calculated reference tap and reference tap positions are provided in FFF, along with the FFF tap coefficients, and the FBF tap coefficients are provided in FBF. In the present specification, the training equalizer 305 according to an embodiment of the present invention implemented as shown in FIG. 3 is called GT-DFE (General Training DFE), and the training equalizer according to a conventional conventional technology is referred to as CT-DFE. Call. A channel equalizer (eg, FIG. 1) according to an embodiment of the present invention may use both GT-DFE and CT-DFE. Both may optionally be used in the channel equalizer according to one embodiment of the present invention. In the present specification, when GT-DFE and CT-DFE need to be specifically distinguished, they are called differently, but when they do not have the benefit of the division, they are referred to as training equalizers.

GT-DFE 305 includes an anti-Koseol FFF (GT-FFF), a Koseol FBF (GT-FBF), and an additional Koseol FBF (GT-aFBF).

, △,

} 와 버퍼(301)에서의

에 대응되는 수신된 신호 샘플들 r_n을 수신한다. 만약

가 영이 아니라면, 다음과 같은 수학식 13처럼 메트릭스 형태로 바꿔 쓸 수 있다.The GT-DFE 305 estimates information of the channel estimate 303 {

, △,

} And in buffer 301

Receive received signal samples r _n corresponding to. if

If is not zero, it can be rewritten in matrix form as shown in Equation 13.

은,

와

로 주어진 GT-FFF에서의 기준 탭

의 위치에서 계산된다. 따라서, {r, a, w}는 추정된 {△,

}에 의해 하기 수학식 14, 15, 16, 및 17 처럼 기술할 수 있다.In order to maintain generality, the output of the GT-DFE 305 according to an embodiment of the present invention

silver,

Wow

Tab in GT-FFF given by

Is calculated at the position of. Thus, { r, a, w } is estimated {△,

} By the following equations (14), (15), (16), and (17).

벡터이고, a는 시간 n에서

벡터이다. 상기 수학식 13의

채널 H 는, 상기 수학식 17에서 주어진 서브 메트릭스들 H ₁, H ₂, 및 H ₃로 주어진다. GT-DFE(305)에서, 채널 출력 벡터 r은 GT-FFF로 제공되어, 로부터 야기되는 포스트 커서 ISI를 제외한 포스트 커서 ISI 일부분이 제거되며, 이

는 상기 수학식 12에서

와

사이의 채널 탭들에 해당한다.Where { r, w } is

Vector and a is at time n

Vector. Of Equation 13

Channel H is given by the sub-metrics H ₁ , H ₂ , and H ₃ given in Equation 17 above. In GT-DFE 305, the channel output vector r is provided as GT-FFF, A portion of the post cursor ISI is removed except for the post cursor ISI resulting from

In Equation 12

Wow

Corresponds to channel taps in between.

는 하기 수학식 19처럼 정의된다.Here, the first and second subscripts separated by colons represent the output of the GT-FBF and the components of the vector, and the input vector of the GT-FFF.

Is defined as in Equation 19 below.

), 부분적인 포스트커서 ISI인

는

인 경우

라고 가정하면 다음 수학식 20와 같이 나타낼 수 있다.In equation (18), if the tap length of the GT-FBF is equal to the tap length of the GT-FFF (eg

), Which is a partial postcursor ISI

Is

If

If it is assumed that can be expressed as shown in the following equation (20).

은 시간 n에서의 V (n) 구성 요소이다. 상기 수학식 18에 상기 수학식 13을 대입하고 채널 매트릭스 H ₃가 완벽하게 추정될 수 있다고 가정하면, 상기 수학식 (18)을 하기 수학식 21처럼 재기술될 수 있다.From here,

Is the V (n) component at time n. Substituting Equation 13 into Equation 18 and assuming that channel matrix H ₃ can be perfectly estimated, Equation (18) can be rewritten as Equation 21 below.

채널 메트릭스

로 감소 된다. 따라서,

에서 원치 않는 채널 탭에 의해 야기된 ISI는 도 2의 GT-FFF와 GT-aFBF에 의해 제거될 수 있다. From Equations 13 and 21, the effective dimension of the channel matrix H is

Channel metrics

Is reduced to. therefore,

The ISI caused by unwanted channel taps at can be removed by the GT-FFF and GT-aFBF of FIG.

이다. 결과적으로, GT-DFE의 수렴 속도는 증가한다. GT-DFE의 출력

하기 수식 22처럼 나타낼 수 있다.Channel dimension reduction affects the signal to interference plus noise ratio (SINR) and convergence speed of the signal output by the GT-DFE. That is, the SINR of GT-DFE is higher than that of CT-DFE.

to be. As a result, the convergence speed of the GT-DFE increases. GT-DFE output

It can be expressed as Equation 22 below.

와

에 대한

는 GT-FFF와 GT-FBF 각각의 계수이다. GT-FBF의 계수는 도 2에 나타낸 것처럼

에 대하여

로 설정된다.here,

Wow

For

Is the coefficient of GT-FFF and GT-FBF respectively. The coefficient of GT-FBF is as shown in FIG.

about

Is set to.

를 위한

로 설정될 수 있다. 상기 수학식 22에서, 상기 수학식 20과 관련된

는, GT-FFF 이전에, 하기 수학식 23과 같이 정의되는 시프트 특성에 의해 제거될 수 있다. k = △ + 1, ...,

for

It can be set to. In Equation 22, Equation 20

May be removed by a shift characteristic defined as in Equation 23 before GT-FFF.

이고, 상기 수학식 22는 하기 수학식 24처럼 재 기술될 수 있다.Equation 23 is defined on the premise that the tap length of GT-FBF is longer than the tap length of GF-FFF. If channel estimation is perfect,

Equation 22 may be rewritten as in Equation 24 below.

가 제로(0)인 경우, GT-DFE의 출력은 CT-DFE의 값과 동일함에 주목하라. 이러한 사실로부터, GT-DFE의 구조는

에 따라 달라질 수 있으므로, CT-DFE의 구조보다 플렉서블하다는 것을 알 수 있다.In Equation 24,

Note that if is zero, the output of the GT-DFE is equal to the value of the CT-DFE. From this fact, the structure of the GT-DFE

Since it can vary depending on, it can be seen that more flexible than the structure of the CT-DFE.

가

로 초기화된 후에 GT-DFE가 수렴하도록 훈련 열(TS)는 여러 번 반복된다. 훈련 동안에, GT-DFE의 계수들은, 하기 수학식 25와 같은 LMS 알고리즘을 사용하여 업데이트 된다.To overcome the shortcomings of short training trains, the GT-FFF

end

After initialization, the training sequence TS is repeated several times for the GT-DFE to converge. During training, the coefficients of the GT-DFE are updated using an LMS algorithm, such as

는 각각 GT-FFF의

계수 벡터이고, GT-aFBF 의

계수 벡터이다. a(n)은 GT-aFBF 의

입력 벡터이고, 스텝 사이 즈

은 상기 수학식 8의 MA-VSSS 알고리즘에 의해 업데이트 된다.here,

Each of the GT-FFF

Coefficient vector, and the value of GT-aFBF

Coefficient vector. a (n) is the GT-aFBF

Input vector, step size

Is updated by the MA-VSSS algorithm of Equation 8.

는 FFF의 계수들을 위해서 저장되고(save), GT-aFBF 의 계수

는 사용되지 않는다. 상기 수학식 3의 FBF의 계수는 하기 수학식 26의 관계식을 사용하여 계산된다.After the training is completed, the coefficient of GT-FFF

Is saved for the coefficients of the FFF, and the coefficients of the GT-aFBF

Is not used. The coefficient of FBF in Equation 3 is calculated using the relation in Equation 26 below.

의 크기는 에러 전파를 방지되도록 하는 조건을 대략적으로 만족하면서 작아지게 된다.From the structure of the GT-DFE of FIG. 3, it can be seen that the influence of the large post cursors associated with the sub-matrix H2 in Equation 17 is eliminated by the GT-aFBF and GT-FFF. As a result, in Equation 26

The size of becomes small while approximately satisfying the condition for preventing error propagation.

When the coefficients of the FFF 307 and the FBF 313 are initialized by the GT-DFE, the received signal samples stored in the buffer 301 are supplied to the FFF 307 and the DFE 315 and restored. Fine tuning is performed by the signals corresponding to the training sequence TS among the received signals. In order to update the coefficients of the FFF 307 and the FBF 313, a MA-VSS LMS algorithm is applied during the training period (“TS mode”), and a modified stop and go algorithm (reference: M Ghosh, blind Decision Feedback Equalization for Terrestrial Television Receivers, " in Processings of the IEEE , vol. 86, no. 10, Oct. 1998.) This data is used for the data equalization period (" DS mode "). In the N Go algorithm, the fixed step size is replaced by the MAVSS algorithm.

In the case of using the determination unit of FIG. 4, which is an embodiment of the present invention described below, error propagation can be reduced by correcting past incorrect determination symbols as well as reducing the probability of determination error.

4 is a block diagram illustrating a determination unit according to an embodiment of the present invention.

The determination unit according to an embodiment of the present invention is a combined structure of an FBF and a Viterbi decoder that minimizes the post cursor ISI by reducing the probability of determination error on the subtractor output yn and correcting past determination symbols in the FBF.

Referring to FIG. 4, the FBF uses 12 Viterbi coders in parallel, and the input of each Viterbi decoder is reconstructed by a symbol de-interleaver 601. The symbol deinterleaver 601 performs the exact opposite function of the trellis coder interleaver, passing all other symbols that are not interleaved with the trellis coder. The filter storages 603a to 603T of FIG. 4 perform a function as a symbol interleaver after a Viterbi decode operation, which places past decision symbols of any Viterbi decoder at the first position of each filter bank to determine past determination. This is done by correcting the symbols. Here, each of the filter banks is composed of 12 symbol spaced-register V.

Each of the Viterbi decoders of FIG. 4 outputs a tentative decision symbol by performing the following main steps at time n.

1) branch methric generation

2) Survivor path and path metric update for all states

3) Trace back the most likely path

In any Viterbi decoder, the symbols determined through the third step are represented by Equation 27 below.

은 도 4의 i번째 비터비 디코더의 판정 심볼 벡터를 나타낸다. i번째 비터비 디코더는 도 4의 i번째 트렐리스 인코더에 대응된다.

는 비터비 디코더의 역 추적(trace back) 길이를 나타내며,

는 i번째 비터비 디코더의 j 번째 패스(path)의 역 추적에서의 판정된 심볼을 나타낸다.

≥15에 대하여 비터비 디코더는 최적화된 비터비 디코더에 비교할 때 무시할 만한 성능 열화를 가져온다. 그러나,

≥의 조건이 8VSB DTV 시스템에 적용되면, 판정 딜레이 12(

-1)가 12개의 비터비 디코더에 의해 생긴다. 이로 인하여, DFE의 초기 수렴(Initial convergence)에 영향을 미친다. 따라서, 본 발명에서는 판정 딜레이로 인한 영향을 없애기 위해서 초기 12(

-1) 심볼 주기 동안

=1인 비터비 디코더를 사용한다. 도 4에서, 각각의 디터비 디코더로(311)부터 FBF(313)까지의 정정 과정은, 심볼 디 인터리버가 송신기에서의 트렐리스 코드 인터리버와 완벽히 동기화 된다는 전제하에, 통상적으로 다음 세 가지 경우로 수행된다.here

Denotes a decision symbol vector of the i-th Viterbi decoder of FIG. The i-th Viterbi decoder corresponds to the i-th trellis encoder of FIG. 4.

Denotes the trace back length of the Viterbi decoder,

Denotes the determined symbol in the backward tracking of the j th path of the i th Viterbi decoder.

For ≥ 15, the Viterbi decoder results in negligible performance degradation when compared to the optimized Viterbi decoder. But,

If the condition of ≥ is applied to the 8VSB DTV system, decision delay 12 (

-1) is generated by twelve Viterbi decoders. This affects the initial convergence of the DFE. Therefore, in the present invention, in order to eliminate the effects of the decision delay, the initial 12 (

-1) during symbol period

Use a Viterbi decoder with = 1. In FIG. 4, the correction process from each diterbi decoder 311 to the FBF 313 is generally performed in the following three cases, provided that the symbol deinterleaver is completely synchronized with the trellis code interleaver at the transmitter. Is performed.

중 첫 번째 판정 심볼

은, 비-트렐리스 코딩된 심볼들이

번째 필터 뱅크를 통과할 때까지, 첫 번째 필터 뱅크로 전달되기만 한다. 경우 1은, 12(

-1) 심볼 동안 FBF에서의 정정없이

=1 인 비터비 디코더와 유사한 과정을 가진다.Case 1: This procedure is applied after the first segment of any field is received. During the segment sync input of the second segment, the arm of the symbol deinterleaver of FIG. 3 rotates to match the reordering of the transmitter. Thereafter, the I-th Viterbi decoder for yn is operated in response to the output symbol of the i-th trellis encoder,

First judgment symbol

Is a non-trellis coded symbol

It only passes to the first filter bank until it passes through the first filter bank. Case 1 is 12 (

-1) without correction in FBF during symbol

It has a similar process to the Viterbi decoder with = 1.

-필터 뱅크를 통과한 후에, 후속하는 입력 심볼들에 대한 i번째 비터비 디코더의 판정 심볼 벡터

가 FBF로 전달되어 마지막 832-12(

-1) 심볼들 동안에 과거 판정 심볼을 하기 수학식 28과 같이 정정한다.Case 2: non-trellis coded symbols

After passing through the filter bank, the decision symbol vector of the i-th Viterbi decoder for subsequent input symbols

Is passed to FBF and the last 832-12 (

-1) During past symbols, the past decision symbol is corrected as in Equation 28 below.

은 i번째 트렐리스 코더의 출력 심볼들에 대한 과거 판정 심볼 벡터이고,

은 시간 n에서 k번째 필터 뱅크의 첫 번째 심볼이다. 따라서,

에 대하여

이면,

은

에 의해 정정된다.Where formula

Is the past decision symbol vector for the output symbols of the i th trellis coder,

Is the first symbol of the k-th filter bank at time n. therefore,

about

If,

silver

Corrected by

Case 3: This procedure applies from the third segment of any field and depends on the structure of the transmitter's multiplexer, for example the segment's sync and the three segment's patterns inserted between the segments. That is, when segment synchronization is received, the four Viterbi decoders cycle without input and the four Viterbi decoders that are skipped change the correction ordering on the next cycle. In the process of correcting the FBF by the skipped i-th Viterbi decoder, the effect of segment synchronization is reflected as in Equation 29 below.

은 시간 n에서 k 번째 필터 뱅크의 첫 번째 위치에 위치하는 세그먼트 동기를 나타낸다. 상기 식 29는, 28식에서

부터

까지에 대응하는 과거 판정 심볼들을 필트 뱅크 하나 더 만큼 예를 들면

부터

까지 시프트하는 것을 보여준다. 따라서, 식 28의 정정 과정은, 세그먼트 동기 입력에 앞서서 판정된 심볼들에는 적용될 수 없으며, 수행될 딜레이된 정정과정은 다음과 같다.here,

Denotes segment synchronization located at the first position of the k th filter bank at time n. Equation 29 is represented by 28 equations

from

Past determination symbols corresponding to up to one more filter bank, for example

from

Show shift up to. Therefore, the correction process of Equation 28 cannot be applied to the symbols determined before the segment synchronization input, and the delayed correction process to be performed is as follows.

이면,

은

에 의해 정정된다. 그리고, k번째 필터 뱅크에서는 정정이 수행되지 않는다. 이 것은 상기 식 29에서 ‘0’으로 나타난다. 이 딜레이된 정정 과정은 세그먼트 동기가

번째 필터 뱅크를 통과할 때까지 계속되며, 그 후에 식 28의 정정과정은, 모든 비터비 디코드들에 대하여 이상적으로 적용된다. 경우 3의 정정 패턴은, 필드의 311 데이터 세그먼트를 통해서 반복된다. 경우 1부터 경우 3까지의 정정 과정은 또한 연속적인 데이터 필드에 대하여 정확히 반복된다.For (k + 1) = j,

If,

silver

Corrected by The correction is not performed in the k-th filter bank. This is represented by '0' in Equation 29. This delayed correction process

It continues until it passes through the first filter bank, after which the correction process of equation 28 is ideally applied for all Viterbi decodes. The correction pattern of case 3 is repeated through 311 data segments of the field. The correction process from Case 1 to Case 3 is also repeated exactly for successive data fields.

5 is a flowchart provided to explain a channel equalization method according to an embodiment of the present invention.

5 and 1, the buffer 301 stores a received signal (S701). The correlator 501 of the channel estimator 303 estimates a channel by applying a correlation algorithm to the received signal stored in the buffer 301 (S703). The noise removing unit 503 removes correlation noise of the channel indicated by applying a correlation algorithm (S705). After the LMS unit 505 performs channel estimation by applying a MA-VSS LMS algorithm with reference to the received signal stored in the buffer 301 to the channel from which the correlation noise is removed (S707), the selection unit 507 selects the LMS unit. From among the channels estimated by 505, a channel having a size larger than the threshold is selected as the reference tap, and a position where a channel larger than 1/2 of the size of the selected reference tap signal is located as the reference tap position. (S709). The training equalizer 305 estimates the tap coefficient based on the channel, the reference tap, and the position of the reference tap estimated by the channel estimator 303 (S711). If the output SINR of the training equalizer 305 is 15 dB or more (S713: Y), the equalizer 315 performs channel equalization using the tap coefficient of the training equalizer 305 on the received signal stored in the buffer 301. (S715). The determination unit 311 determines the reliability of the output of the equalized signal (S717). On the other hand, when the output SINR of the training equalizer 305 is less than 15 dB (S713: N), the selecting unit 507 of the channel estimating unit 303 may select another reference tap from among the channels estimated by the LMS unit 505. The position is selected (S709).

In the static and modified Brazilian channels as shown in Table 1, the examples of the present invention were computer simulated.

In Table 1, note that path4 of the modified Brazilian channel has 0 dB magnitude and 0.05 Hz Doppler phase shift, while path4 of the Brazilian channel has -0.1 dB magnitude and 0 degree phase. That is, pate4 created a flat fading environment whose size is 0.05 Hz cycles. In modified channel D, path1 located at 0.15 Hz creates a flat fading environment with a Doppler phase shift of 0.05 Hz. Performance was measured by root mean square error (RMS) on dB scale, SINR on dB scale, and symbol error rate (SER). The SINR and SER results are averaged from 50 independent measurements over 40 fields (approximately 10 ⁷ symbols). In general, RMSE is better than mean-square-error in comparing the performance of channel estimators. The magnitude offset of the estimated CIR affects calculating the coefficients of the equalizer based on the channel estimate. SINR is a good way to compare the convergence speed of an equalizer. SER can be used to measure the SNR required to satisfy a threshold of visibility (TOV) of 0.2.

6A and 6B are graphs illustrating Root Mean Square Error (RMSE) in Brazilian A and D channels according to an embodiment of the present invention. 6A to 6B and 8A to 8C, A, B, C, D, E, F, and G respectively mean the following.

A: DFE

B: DFE with VD-TBl-1

C: DFE with VD-TBl-15

D: CT-DFE based DFE with VD-TBL-1

E: CT-DFE based DFE with VD-TBL-15

F: GT-DFE based DFE with VD-TBL-1

G: GT-DFE based DFE with VD-TBL-15

Here, VD-TBL-1 indicates that the length of the traceback of the Viterbi decoder is 1, and VD-TBL-15 indicates that the length of the traceback of the Viterbi decoder is 15.

로 행함) 등으로 RMSE의 성능을 비교하였다. MA-VSS LMS 알고리즘의 파라미터의 파라미터는, 다음과 같다.6A and 6B, four different channel estimation techniques, namely (a) correlation, (b) MA-VSS LMS algorithm, (c) correlation and MA-VSS LMS algorithm, and (d) of the present invention One embodiment (four different thresholds in equation (6)

The performance of RMSE was compared. The parameters of the parameters of the MA-VSS LMS algorithm are as follows.

)을 사용하지 않은 경우(예를 들면 (a), (c))에는 상관 노이즈 때문에 성능 열화가 발생한다. MA-VSS LMS 알고리즘의 RMSE는 불충분한 훈련 열 길이에 기인한 성능 손실 때문에 브라질 D와 같은 채널에서 증가된다. 한편, 본원 발명의 실시예들에서는, MA-VSS LMS 알고리즘의 계수들을 thresholded 상관 알고리즘 적용후에 잔존한 탭들로 초기화하였으므로, 훈련 열의 길이가 짧음으로 인한 단점을 극복하였다. As can be seen in Figures 6a and 6b, the threshold (

) Is not used (for example, (a) and (c)), performance degradation occurs due to correlation noise. The RMSE of the MA-VSS LMS algorithm is increased in channels such as Brazil D due to performance loss due to insufficient training column length. On the other hand, in the embodiments of the present invention, since the coefficients of the MA-VSS LMS algorithm are initialized with the remaining taps after the thresholded correlation algorithm is applied, the shortcomings due to the short training string are overcome.

의 크기에 따라서 다른 성능을 보여줌을 알 수 있다.

가 0.2 인 경우는

={0.1, 0.3, 0.4}인 경우보다, 입력되는 SNR이 증가됨에 따라서, RMSE가 더 작아진다.

=0.1 인 경우와

={0.3, 0.4}인 경우는, 각각 잔류 상관 노이즈와 버려진 큰 채널 탭들로부터 기인한 MA-VSS LMA 알고리즘의 불충분한 초기화 때문이다. 따라서,

=0.2 인 경우가 이상적이다.Referring to Figure 6b, embodiments of the present invention,

It can be seen that it shows different performance depending on the size of.

Is 0.2

As the input SNR is increased, the RMSE becomes smaller than the case of = {0.1, 0.3, 0.4}.

With = 0.1 and

The case of = {0.3, 0.4} is due to insufficient initialization of the MA-VSS LMA algorithm resulting from residual correlation noise and discarded large channel taps, respectively. therefore,

Ideally, = 0.2.

7A to 7D are graphs showing signal to interface plus ratios (SINRs) of equalizers in Brazilian A, C, D, and E channels according to an embodiment of the present invention. 7A to 7D, A: DFE, B: CT-DFE-based DFE, and C: GT-DFE-based DFE are shown.

Convergence characteristics of the equalizer, for example start-up SINR, convergence rate, and SINR after convergence were compared in the following cases, respectively.

GT-DFE based equalizer

B) CT-DFE based equalizer

C) Equalizer without GT-DFE

에 따라서 변화된다. GT-DFE, CT-DFE 및 DFE의 MA-VSS LMS 알고리즘의 파라미터 값들은 하기 표2로 요약하였다.In the CT-DFE based equalizer, after the reference tap is selected according to the channel characteristics, the reference tap is located at the rightmost side of the FFF of the CT-DFE. 400 tabs FFF and 620 tabs FBF are used. 400 tabs GT-FFF and CT-FFF are also used, and the tap length of GT-aFBF

Will change accordingly. The parameter values of the MA-VSS LMS algorithm of GT-DFE, CT-DFE and DFE are summarized in Table 2 below.

In Table 2, parameters with a (*) mark were not used in the CT-DFE. In order for the training of the GT-DFE and CT-DFE to converge, the signal samples corresponding to the training sequence were repeated 18 times. The trace back length of the Viterbi decoder used is 15. In Figs. 7A to 7D described later, the SINR curve of the equalizer is given as a function of the number of symbols, and each point represents an average of 1000 or more symbols.

7A to 7D, the position of the reference tap should be carefully selected for the initialization of the DFE, and in CT-DFE, the SINR loss due to error propagation is greater than the SINR gain without FBF, and the GT-DFE starts fast. -up, convergence speed, high power SINR As a result, initialization of the coefficients is necessary to widen the equalizer, and the GT-DFE is more suitable for 8VSB DTV than the CT-DFE.

8A to 8C are graphs illustrating SER (Symbol Error Rate) of the equalizer in the Brazilian E and C channels according to an embodiment of the present invention.

The SER performances of the GT-DFE based equalizer, CT-DFE based equalizer, and equalizer were compared as a function of the input SNR. SER was calculated for 10,000 symbols.

Referring to FIG. 8A, the SER forms a curve in the Brazilian E channel. In FIG. 8A, the GT-DFE based DFE shows a lower SER than the DFE when the input SNR is greater than 23 dB. In particular, it is noted that the performance of E (GT-DFE based DFE) is better than that of G (DFE). It can be seen that when the magnitude of the input SNR is 23 dB or more, the input SNR is improved by about 1 dB compared to the case where the traceback length of Viterbi decode is 15 (E, G) is 1 (A, F). In contrast, the CT-DFE based DFE shows the worst SER performance due to error propagation, as shown in FIG. 7D. In FIG. 8A, the GT-DFE based DFE satisfies TOV when the input SNR is 22 dB or more.

8B and 8C show SER performance in modified Brazilian C and D channels. In these channels, the reference taps and their positions in the GT-DFE are the same as those of the Brazilian channel C and D channels. 8B shows that the GT-DFE based DFE shows the best performance when compared to the others when the traceback length of the Viterbi decoder is 15, whereas the GT- when the traceback length of the Viterbi decoder is 1 DFE based DFE shows similar performance to DFE when the traceback length of the Viterbi decoder is one. Performance degradation of the CT-DFE based DFE can be analyzed from FIG. 7B. Referring to FIG. 8C, an equalizer having a traceback length of 15 in a Viterbi decoder exhibits better performance than an equalizer having a traceback length of 1 in a Viterbi decoder. The effect of the reference tap can be minimized by making the traceback length of the Viterbi decoder 15. 8B and 8C, it can be seen that the TOV is very well satisfied when the input SNR is 20 dB.

As described above, according to the channel equalizer and the method of the present invention, error propagation is minimized by changing the position of the reference tap in accordance with the characteristics of the channel.

Although the embodiments have been described with reference to the accompanying drawings, the method is not limited to the above-described embodiments and various other changes and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. All such changes and modifications are intended to be included within the scope of the invention as defined by the appended claims.

Claims (5)

A channel estimator for estimating a channel with respect to the received signal and selecting a reference tap and a position of the reference tap from the estimated channel; A training equalizer (DFE) for estimating a tap coefficient based on the estimated channel, the reference tap, and the position of the reference tap; And And an equalization unit (DFE: Decision Feedback Equalizer) for equalizing the received signal by using the estimated tap coefficients. The method of claim 1, The channel estimator, A correlator configured to estimate a channel by applying a correlation algorithm to the received signal; A noise removing unit for removing noise of the channel estimated by the correlation unit; An LMS unit for estimating a channel by applying a moving-average variable step-size (MA-VSS) LMS algorithm to a received signal and the channel from which the noise is removed; And And a selector for selecting a reference tap and a reference tap position from among the channel estimated by the LMS. The method of claim 2, The selection unit, From among the channels estimated by the LMS, a channel having a size larger than a threshold value is selected as a reference tap, and a location where a channel larger than 1/2 of the size of the selected reference tap signal is located is selected as a reference tap position. Channel equalizer with minimal error propagation. a) estimating a channel for the received signal and selecting a reference tap and a position of the reference tap from the estimated channel; b) estimating a tap coefficient based on the estimated channel, reference tap, and position of the reference tap; And and c) equalizing the received signal using the estimated tap coefficients. The method of claim 4, wherein Step a) is Estimating a channel by applying a correlation algorithm to the received signal; A noise removing step of removing noise of the estimated channel; Estimating a channel by applying a moving-average variable step-size (MA-VSS) LMS algorithm to the received signal and the channel from which the noise is removed; And Selecting a reference tap and a reference tap position from among estimated channels by applying the LMS algorithm.

Images