DESCRIPTION

Channel estimation method and device for TDS-OFDM systems

Technical Field

The invention regards to a channel estimation method for TDS- OFDM systems according to pre-characterizing part of claim 1, and to a channel estimation device for TDS-OFDM systems.

Background Art

The Time Domain Synchronous-OFDM (TDS-OFDM) system is a specially designed OFDM system by combining both time-domain and frequency-domain processing. It is adopted in the compulsory national standard of Digital Multimedia Terrestrial Broadcasting (DMTB) in China. The TDS-OFDM offers a pseudo random (PN) sequence as guardian interval between every two frames of OFDM symbols. This special structure brings the challenge of novel channel estimation method.

Being the core part of general OFDM system, the quality of channel estimates has a direct impact on the overall system performance. The special frame structure of TDS-OFDM system re- quires novel channel estimation methods at the receiver side, instead of ones used in general OFDM system.

There are some problems and conditions to be solved. Channel estimation provides the estimate of channel impulse response to frequency domain equalizer in TDS-OFDM system. It has a direct impact on the system performance. There are several requirements for the channel estimation.

At first channel estimation has to be as accurate as possible even under low SNR situation. According to second aspect, chan-

nel estimation has to be able to cope with channels with long echoes. And as third aspect, channel estimation has to be able to handle dynamic channels.

FIG. 3 SHOWS THE CHANNEL ESTIMATION BLOCK IN THE OFDM system diagram. It recovers the channel response from the received data then feeds to the frequency domain equalizer. There is shown an arrangement of general channel estimation in the OFDM system diagram. A transmitter TX sends out sent data via an an- tenna to an antenna of a receiver RX. Path V between these antennas provides error data and multiple effects effecting sent data to become initially received data received by antenna of receiver RX. Initially received data are provided to an automatic gain control 9. Outputted data of automatic gain control 9 are forwarded to an IQ-demodulation block 10 (IQ: In- phase/Quadraturphase) . Data outputted of IQ demodulation block 10 are inputted as the received data y(i) into a block especially comprising the channel estimation block 1. Data parallel with channel estimation results outputted by channel estimation block 1 are forwarded to a frequency domain equalizer 11. Data outputted by the frequency domain equalizer 11 are forwarded to a forward error control 12, FEC.

The general channel estimation method in OFDM system can't be applied directly due to the special frame structure of TDS-

OFDM. Fig. 4 shows the TDS-OFDM frame structure for two different modes .

These two different TDS-OFDM frame structures have 4200 symbols per frame and 4725 symbols per frame, respectively. First part of symbols is cyclic shifted symbols and PN data between such cyclic shifted symbols. These PN data between cyclic shifted symbols have the length of 255 in case of 4200 frame structure or the length of 511 in case of 4725 frame structure. In other words, the frame is divided into two parts. The beginning part,

e.g. 420 symbols in the 4200 frame structure, is called guardian interval, which is a cyclic PN sequence in time domain. And the later part, e.g. 3780 in the 4200 frame structure, is the OFDM data part.

Assuming the transmitted signal is s{n) and additive noise is n(n) , which is white Gaussian noise, then the received signal y(n) is as following

L-1 y{n) = s(n)®h{n)+n{n) = σ∑a(τ)s(n-τ)+n(n), (1, r=0

where h(n) is an impulse response of a multipath channel, and a(τ) s are wide-sense stationary, narrow band complex Gaussian random processes modelling echoes with longest echo at L-symbol delays. It is assumed that they are constant during the transmission of one OFDM symbol, which is a general assumption in OFDM systems. Here "®" means linear convolution.

The channel estimation problem is how to make close estimates of these a(τ) from the received signal y(n) .

Existing state of the art solutions are based on time-domain correlation by correlating received signal y(n) in the guardian interval with the local PN sequence. If the auto-correlation function of the PN sequence is an ideal impulse, we can easily get channel response a(τ) .

By simple mathematical deduction, we can get the correlation between the received signal y{n) and local PN p(n), shown in Eq. 2.

Af-I

R»= ∑/>('>(»+0

I=O Af-I L-I

= ∑pQ)[∑φ)s(P+i-τ)+φ+0]

1=0 r=0 L-I Af-I Af-I

=∑«O)∑P(i)s(P+i-τ)+∑P ⁽ f)n ^{ n+i) r=0 i=0 1=0 n = 0,...,Af-I , (2)

where M is the PN length.

If the transmitted signal is the local PN sequence and has the following property

s{n+i-τ) = p{(n+i-τ) _M ), nj = "\ M-Xr = O L-1, (3) then we can rewrite Eq. (2) as

R _py (n) = ∑a(τ)R _pp (n-τ)+R _pn (n) r=0 ' ^ )

= a{n)+R _pn (n)

where R _pp is the auto-correlation function of the PN sequence, which is assumed to be ideal δ function, and R _pn is the cross- relation function of PN sequence and noise, which is zero approximately.

So the channel estimates

ά(n) = R _py (n) τ = 0 L-1. (5)

Technical Problem

However, this time domain correlation method has some draw-

backs. The auto-correlation function of PN sequence in TDS-OFDM is not an ideal £ function as shown below (the case of non- cyclic PN length of 255)

So, Eq. 4 doesn't hold strictly. This will introduce extra estimation noise. The assumption of Eq. 3 cannot be made because the data section exits between two guardian intervals. This will cause performance degradation. This method will introduce extra estimation noise, when the SNR is low, because the later part of Eq. 4 cannot be ignored when the noise power is large.

It is an object of the invention to provide another channel es- timation solution by hardware, software and/or method, especially providing results without or with reduced extra estimation noise.

Technical Solution

This object is solved by a method for adaptive channel estimation in TDS-OFDM receiver having features according to claim 1, and by a device for adaptive channel estimation in TDS-OFDM receiver having features according to claim 10. Preferred aspects and embodiments are subject-matter of dependent claims.

Especially, there is provided a method for adaptive channel estimation in TDS-OFDM receiver by processing received data comprising a data frame having a guard interval, wherein received data are processed in an FIR filter for an adaptive channel estimation by adapting coefficients of the adaptive FIR filter.

Further, there is provided a channel estimation device for TDS- OFDM systems in a receiver comprising an input for received

data comprising a data frame having a guard interval, and an channel estimator device for estimation of a channel by processing the received data, wherein the channel estimator device comprises a FIR filter for an adaptive channel estimation by adapting coefficients of the adaptive FIR filter.

Advantageous Effects

It is preferred, when the guard interval is formed by a local pseudo random sequence, which is cyclic shifted to the adaptive FIR filter. Corresponding device comprises a cyclic shifter for cyclic shifting such local pseudo random sequence as the guard interval to the adaptive FIR filter.

A filter output value of the adaptive FIR filter can be used for training the error generation to get an error for coefficient update. Corresponding device comprises an error generator adapted to get such error for coefficient update by processing a filter output value of the FIR filter.

Preferably, a part of received data is selected and used for training to error generation to get an error for coefficient update. Corresponding device comprises an error generator adapted to get an error for coefficient update by processing a selected part of received data for training the error generator.

Using such methods or devices the trained coefficients are the channel estimates like in known time-domain accordingly.

Preferably, the received data are selected, which start after receiving a longest echo-length number of data upon a frame head signal. By this way the longest echo-length from a Channel Length Estimation Unit is used to decide the start point of re- ceived data for training each frame adaptively.

Especially, the longest echo-length expected to handle is set as the start point of received data for training. Thus, it is possible to set start point, if the longest echo-length is not available. In addition, end of the received data for training can be selected, when the guard interval ends.

Especially, coefficients of the adaptive FIR filter are updated adaptively via uncontrolled tap update method in case of a static channel thereby ignoring or bypassing a control unit for controlling function of tap update elements. This may be, but not exclusively, used when the channel is static. No control scheme is used and all taps can are updated normally with the error signal from the adaptive filter.

Coefficients of the adaptive FIR filter can be updated adaptively selected via a controlled tap update method in case of a dynamic or uncertain channel. In other words, a control unit is enabled. This may be, but not exclusively, used when the chan- nel is dynamic or uncertain. Only selected taps will be set in normal mode and updated using the error signal. The other taps will be set in bypass mode and not updated.

Especially, only some tap branches of a plurality of tap branches are adaptively selected to contribute to the FIR filter output. Thus, the other branches did not contribute to multiplication when providing corresponding tap values.

According to a preferred embodiment device comprises a plural- ity of tap update elements adapted to update each one of the coefficients of the adaptive FIR filter.

Especially, an update control unit can be adapted to adaptively control updating of selected coefficients (controlled mode) or all coefficients (uncontrolled mode) .

There is provided a better channel estimation solution overcoming previous drawbacks. The method and device may not require an ideal auto-correlation function of PN sequence in guardian interval. Further, method and device provide better performance in dynamic channel or channel with low SNR.

Present channel estimation method and device are based on time domain and utilize the special guardian interval. Time domain correlation is the usual way for simple implementation, but it introduces severe noise when the SNR is low or when the guardian interval doesn't have an ideal autocorrelation function. Instead, there is used a controlled LMS (least mean square error) method and device with special tap update scheme for chan- nel estimation. Method and device provide much better performance than the correlation method under low SNR or dynamic situations with little extra hardware cost. Also the method doesn't require an ideal autocorrelation of guardian interval, which is the case in the DMTB standard.

Description of Drawings

An embodiment will be disclosed in more details with respect to enclosed drawing. There are shown in:

Fig. 1 a preferred device or arrangement of components being arranged to receive data and to execute a channel estimation, (I added outputs of data y(i) and channel a(j) to clear to confusion between fig.l and fig.3)

Fig. 2 an interval of data for training,

Fig. 3 a channel estimation block in the OFDM system diagram,

Fig. 4 TDS-OFDM frame structures for two different modes, and

Fig. 5 channel estimation results for AWGN channel under low SNR=I.8dB.

Mode for Invention

Fig. 1 shows a preferred device or a preferred arrangement of components being arranged to receive data and to execute a channel estimation. Especially, instead of hardware components there can be used a software algorithm. There are inputted received data y(i) into such a channel estimator device 1. A number i of received data y(i) corresponds to data positions L, L+l, ..., 420 out of a sequence of received data y(i) for training to be used for channel estimation.

The received data y(i) are inputted into an error generator 2. The error generator 2 outputs error values e. The error values e are inputted into each of a plurality of tap update elements 3(0), 3(1), ..., 3(254) of a coefficients update device 3. In present case a number i of tap update elements 3(0), 3(1), ..., 3(254) correspond to the number i of received data y(i), i.e. 255.

A number i of tab values a(0), a(l), ..., a (254) outputted by the plurality of tab update elements 3(0), 3(1), ..., 3(254) are inputted into a corresponding number i of multiplier elements 4(0), 4(1), ..., 4(254) of a multiplier 4. Multiplied data outputted of the plurality of multiplier elements 4(0), 4(1), ..., 4(254) are inputted into an adder 5. The adder 5 adds the data values to output a filter output value fo . Filter output value fo is inputted into a further input of the error generator 2. Thus, the error generator 2 generates the actual error value e depending from received data value y(i) and actually inputted filter output value fo.

In addition, a local PN generator 6 generates a local pseudo random (PN) sequence. Values of such pseudo random sequence are inputted into a cyclic shifter 9. Cyclic shifter 9 shifts inputted values by one per shifting step. Cyclic shifter 9 out- puts a plurality of i shifter values p(i), which are shifted per shifting step. Shifter values p(i) = p(0), p(l), ..., p (254) each are inputted into a corresponding of the tab update elements 3(0), 3(1), ..., 3(254). In other words, first shifter value p(0) has to be inputted into first tab update element 3(0) and last shifter value p(254) is inputted into last tab update element 3(254) when taking a first cycle step. In addition these shifter values p(0), p(l), ..., p(254) each are inputted into a further multiplier element input of a corresponding of the multiplier elements 4(0), 4(1), ..., 4(254).

To control function of the tab update elements 3(0), 3(1), ..., 3(254) there is provided a flag signal "flag(i)" having a number i of individual flags flag(0), flag(l), ..., flag(254) corresponding to the number i of received data y(i) for training. Thus, each of the tab update elements 3(0), 3(1), ..., 3(254) is controlled by a corresponding flag signal element flag(0), flag(l) , ..., flag(254) .

The elements of the flag signal flag(i) are provided by an up- date control unit 8. Further, a channel length estimation device 7 provides a channel length value L. Such channel length value L provides the first element of received data y(i) within a greater sequence of received data elements composed of cyclic data, PN data, and data content data.

Thus, received data (y(i)) are processed in an FIR filter (2 - 8) for an adaptive channel estimation, wherein the FIR filter (2 - 8) is composed by several components and method steps according to especially Fig. 1.

When reviewing Fig. 3, however, within these component elements, devices, hardware blocks, and/or software blocks for executing channel estimation 1 is provided like shown by Fig. 1 (correct, after I added two outputs in fig.l). Data parallel with channel estimation results outputted by channel estimation block 1 are forwarded to the frequency domain equalizer 11.

Present solution is to use the controlled LMS (least mean square error) method described below for channel estimation. There is fed the local PN sequence to an adaptive filter and the received signal y(i) data is used as decisions to update the taps and tap update elements 3 of the adaptive filter. Especially, this LMS based method does not require a PN sequence with ideal auto-correlation function. In addition, by a special tap-update scheme, only the taps have to be selected, where there are possibly echoes ev, to do the LMS training. This can achieve much better performance than applying LMS to all of the taps equally, because generally there are more taps to be trained than the number of training data, i.e. received date y(i), available when updating the filter taps equally during one frame. By the controlled tap update scheme, the number of tap to be trained is reduced, thus the estimation quality is improved.

Thus, Fig. 1 shows the structure of proposed controlled LMS channel estimation.

Mainly, it is an adaptive FIR filter with 255 taps in the case of non-cyclic PN length of 255. The taps are updated with the error signal e between the filter output fo and the received signal y{i) according to the flags flag(i) set by update control unit 8. These trained taps, i.e. tap values a(0), a(l), ... , a (254), are the estimates of channel responses.

In this method, the local PN sequence at the receiver side-p(k)

is cycle-shifted in the filter registers by cyclic shifter 9 upon each received data y(i) . But not all of the received data y(i) will be used for training of each tap update element 3(i). Instead, there should be choosed the received data y(i)/ which are supposed to contain PN and noise only. Usually, these received data y(i) start after receiving the longest-echo number of data upon a frame head signal, and end when the guardian interval ends as shown in Fig. 2.

Fig. 2 presents the interval of data for training and an exem- plarily method to provide channel length value L with respect to different starting points depending from existing echo values ev within received data y(i) .

The length of data y(i) for training can be changed dynamically according to the longest echo length detection shown in figure 1. When the channel is long, the data length for training becomes shorter accordingly. The PN sequence in the filter register in Fig. 1 will be cycle-shifted accordingly upon each re- ceived data for training.

Assuming the longest echo is within 255 symbols, there are only 420 - 225 = 165 received data y(i) for training of 255 filter taps for each frame. This won't cause much trouble when the channel is static, in which the taps could be trained over several frames.

But in dynamic channel, the training data is not sufficient to ensure a satisfying training over one frame period. Thus, fur- ther there is proposed the controlled training method by updating the tap values a(i) where there are echoes possibly. This is based on the fact that echoes will distribute in spikes along the path V instead of being equally distributed in the practical environment.

There is set the flag signal flag(i) for each tap update element 3(i) in the above adaptive filter shown in right part of Fig. 1. The flag signal flag(i) will control two operations at the same time. One is the update of corresponding tap update elements 3(i) and the other is the contribution to filter sum. When the Update Control Unit 8 detects that there are echoes possibly at an arbitrary tap, the specific flag signal flag(i) is set to 1, and normal adaptive operations will be carried out. That corresponding tap value a(i) will be updated and tap multiplication result from that branch will be counted into the filter sum output value fo . Once the flag signal flag(i) is set to 0, which means that there are no possible echoes, the specific tap value a(i) will neither be updated, nor contribute to the filter sum output value fo . These flag signals flag(i) are set by control unit 8 using certain fast rough channel estimation based on modified time-domain correlation method.

This controlled LMS channel estimation method can estimate echoes up to 255-symbol delay for frame mode with 4200 symbols or 511-symbol delay for frame mode with 4725 symbols. The advantage of this method over time-domain correlation method is that this method provides more accurate channel estimates under low SNR or dynamic channels. Even for high SNR or static channels, it outperforms the correlation method because it does not re- quire an idea impulse auto-correlation function of the PN sequence in the guardian interval from the transmitter tx. In addition, it has low hardware cost for implementation.

Fig. 4 shows two different TDS-OFDM frame structures having 4200 symbols per frame and 4725 symbols per frame, respectively. The PN data between cyclic shifted symbols have the length i of 255 in case of 4200 frame structure or the length i of 511 in case of 4725 frame structure. However, when amending channel length value L and number i of received data y(i) for training other frame structure can be used within such method

for channel estimation, too.

Fig. 5 shows on left hand side data processed by present least mean square method and on the right hand side data processed by prior art time-domain correlation method. There are shown echo amplitude values over path delay in symbols. As can be seen, present LMS method provides better result. Especially, figure 5 show the performance comparison of the new method and the general time-domain correlation method under low SNR (Signal-to- Noise-Ratio) case. The transmitter works in the 4200 frame mode with QPSK (Quadrature Phase Shift Keying) modulation. The SNR is set to be 1.8dB. The channel model is selected to be an Additive White Gaussian Noise model. The channel estimate using controlled LMS method has much lower noise floor than the one using time-domain correlation method.