Method for echo cancellation in a DMT modem apparatus, DMT modem apparatus and computer program product thereof

Abstract

A method for cancelling the Echo part of a received DMT signal in presence of Far-End signal in a ADSL modem, in particular Central Office modem, where an adaptive filtration of the received signal is performed by means of a FIR Filter having multiple taps for implementing a full Least Mean Square error procedure. The procedure includes an estimation of the impulse response (w(k)) of the FIR filter and evaluating an update (w(k+1)) of the estimation in function of a correction increment parameter (μk). The method provides for performing a per-tone compensation step involving an evaluation of a per-tone increment correction (μk) that depends on the mask of tones (RT) used. The method also provides for performing a low taps compensation step reducing the number of modes to be compensated in the update to a subset (NLM) limited to the lower modes for efficiency with respect to computational complexity.

Description

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates to techniques for Least Mean Square (LMS) Echo Cancellation (EC) in a DMT (Discrete Multitone Technique) modem, in particular a ADSL Central Office (CO) modem.

2. Description of the Prior Art

In the following reference will be made to several widely known methods and circuits for telecommunications. Reference is made to publications such as N. Benvenuto and G. Cherubini ‘Algoritmi e Circuiti per le Telecomunicazioni’, Libreria Progetto, Padova Italy, S. Haykin, “Adaptive Filter Theory”, Prentice Hall, N.J., 1986, B. Widrow, “Adaptive filters”, in aspects of Network and System Theory, ed. Kalman, N.Y., 1970, G. Cariolaro, “Analisi spettrale”, Libreria Progetto, Padova, Italy, 1995, G. Cariolaro, “La Teoria unificata dei segnali”, UTET, 2001.

The new generation of ADSL modem requires a high performance adaptive echo-cancellation method, that shows improved convergence properties.

Known echo cancellations methods for cancelling the echo part of a received signal in presence of far-end signal, like the methods described in “Discrete Multitone Echo Cancellation” Minnie Ho, Cioffi, IEEE Transactions on Communications Vol. 44 NO. 7 Jul. 1996, show performance drawbacks that make them not usable in a real ADSL application.

The common approach provides for modeling the acoustical echo path using an adaptive linear combiner, which is a digital filter with adjustable coefficients. The coefficient update is performed via some adaptation method.

Recursive Least Squares (RLS) based methods have a too high computational complexity and are not reliable, whereas other time domain Least Mean Squares methods show poor results, i.e. poor convergence vs. tracking properties or no compliancy with ITU G.992.1 standard. Finally, methods based on frequency domain decision feedback with LMS error minimization, that, to decrease the computation complexity, simplify the update equation, show a degradation of convergence because of the loss of information due to such a simplification.

On the other hand “Full LMS” methods, that are efficient computational complexity LMS methods with frequency domain decision feedback, i.e. with no update simplification, are considered to show the best performance among EC LMS based methods. It is known a simplified version of such method called “Leven”, providing for reducing the computational complexity, that does not suffer of some performance degradation like other methods. The main drawback of such “Leven” method, however, is the increase of the peak excess noise that can be tolerated reducing the correction coefficient.

It is also known that the introduction of a per-tone correction can improve the convergence performance but it can sometimes create instability.

Both “Full LMS” and “Leven” methods can be considered optimum from a point of view of LMS error functional.

This is true for a reference stationary white signal.

However, the ADSL downstream DMT signal is:

• • a coloured signal. This because only some tones are used in downstream transmission;
  • a cyclostationary signal. This is due to the symbol based nature of DMT signal and the presence of Cyclic Prefix CP inserted for addressing channel distortion.

Said two features make both “Full LMS” and “Leven” sub-optimum when applied to ADSL DMT signal.

In fact, in a non-white condition the Eigen value spread of the received signal correlation matrix is increased and so a reduced convergence speed can be observed.

Further, it is quite apparent that one of the basic properties of all the adaptive theory do not hold: a DMT signal is not a stationary signal. This is mainly due to the symbol based nature of such type of modulation. For example, if only a reduced number of tones is used, then all the samples of a certain symbol are strongly correlated while no correlation occurs with samples of other symbols. It is apparent therefore that increasing the number of tones used then the correlation between different samples of the same symbol decrease reducing the level of non-stationarity of the transmitted signal.

Also the insertion of the Cyclic Prefix CP, that is a partial symbol repetition, increase the level of cyclostationarity.

A phenomenon called “low taps problem” related both to “Full LMS” and “Leven” methods was recently observed, having the following features:

• • a. Depending on alignment, the model limitation will be reached very fast or very slow.
  • b. On some alignments, the model limitation will never be reached.
  • c. Depending on which among the 32 taps of the impulse response is “hit”, this can have serious influence on convergence.

In the following, the result of some simulations is described in order to clarify the impact of “low taps problem” into the speed of convergence on a full LMS echo cancellation method.

FIGS. 1A and 1B show the per-tone increment correction with misalignment 0@552 kHz. The misalignment is the difference (in samples) between the transmitted and received DMT signal window. It has to be considered positive if the reference transmitted signal window is transmitted before the window of the received signal.

In FIG. 1A it is shown the power of residual echo derived into a de-mapper, i.e. the decoding device, of an ADSL modem, in some group of tones vs. the number symbols. It is possible to observe that at symbol 4000 there is the step variation of the echo path. The convergence is obtained in 1000 symbols.

In FIG. 1B it is shown the per-tone mean power of signal and noises over the last 500 of symbols (11500-12000), i.e. the per-tone mean power of far-end signal, received echo, residual echo, distortion and other noises into the de-mapper.

In this condition the performances are:

• • CDC (Circulant Decomposition Canceller) without echo: 388 kb/s
  • CDC with echo: 388 kb/s

This means a loss of performance of 0%.

In FIGS. 2A and 2B are shown the results in case of a misalignment, in particular in the case of Misalignment=66@552 MHz.

In FIG. 2A it is shown the power of residual echo derived into the demapper in some group of tone vs. the number symbols. At symbol 4000 there is the step variation of the echo path. The convergence is not obtained in 8000 symbols. All the groups of tones are involved in this low-convergence.

In FIG. 2B it is shown the per-tone mean power of signal and noises over the last 500 of symbols (11500-12000). The convergence was not obtained into tone 59, indicated with a circle, and the neighboring tones.

In this condition the performances are:

• • CDC without echo: 388 kb/s
  • CDC with echo: 376 kb/s

This means a loss of performance of −3%.

Due to the fact that the convergence is an exponential decreasing function the speed of convergence decrease in function of the time. This problem is a severe drawback for the whole Echo Canceller approach.

Implementation in Central Office modems accounts for the most critical situation in ADSL environment, because the downstream signal is a short bandwidth reference signal and so all estimation problems become relevant.

SUMMARY OF THE INVENTION

The object of the present invention is thus to overcome the drawbacks of the prior art arrangements considered in the foregoing while providing an improved solution for echo cancellation in a DMT modem.

According to the invention such an object is achieved by means of a method having the features set forth in the annexed claims, which form an integral part of the description herein. The invention also relates to a corresponding modem apparatus and a corresponding computer program product directly loadable in the memory of a digital computer and comprising software code portions for performing the method of the invention when the product is run on a computer.

Substantially, a method for cancelling the echo part of a received signal comprising a full per-tone compensation method and a low modes compensation method, that is a simplification of the previous one, are proposed, the proposed methods involving decoupling of the non-white reference signal from cyclostationary reference signal problems. To solve the first problem it is proposed a solution that is dependent only on the tone mask used and to solve the second one it is proposed an efficient align-dependent error compensation.

BRIEF DESCRIPTION OF THE DRAWINGS

Further objects, features and advantages of the present invention will become apparent from the following detailed description and annexed drawings, which are provided by way of non limiting example, wherein:

FIGS. 1A, 1B, 2A and 2B already referred in the foregoing show simulation diagrams of an echo cancellation method according to the prior art;

FIGS. 3A, 3B, 4A and 4B show the results of a simulation of the method according to the invention;

FIGS. 5A, 5B, 6A and 6B show the results of a simulation of a simplified embodiment of the method according to the invention;

FIG. 7 shows a block diagram of a modem apparatus implementing the simplified embodiment of the method according to the invention;

FIG. 8 shows a block diagram of a modem apparatus implementing a variant to the simplified embodiment of the method according to the invention.

DETAILED DESCRIPTION OF THE DRAWINGS

It was observed that the low taps problem phenomenon can be explained with the observation that the DMT signal is a non-white cyclostationary signal. As mentioned above, it was observed that is align-dependent and this is due to the cyclostationary nature of the signal. In other words the reference signal, depending on the alignment, does not excite the same taps of the echo cancellation impulse response.

It was also observed that for an extended number of downstream tones (full overlap case) said low taps problem phenomenon cannot be measured. This is in line with the previous consideration about the signal being ciclostationary. This also means that the worst case condition is related to a reduced number of tones (i.e. ITU G.992.1 Annex B AFDD).

Also the fact that the number of the “low taps” is equal to the number of cyclic prefix samples indicates a relation.

Said phenomenon is always present in the same taps of the echo cancellation impulse response and the impact on the performance is not the same depending on the energy of the impulse response under the window with the “low taps problem”.

Since the “low taps problem” is related to the nature of the reference DMT signal and the statistical properties can a priori be known, the basic concept underlying the presently preferred embodiment of the invention is to exploit this knowledge together with the alignment to compensate in part or at all the Eigen value spread, especially into low taps.

Also it is proposed to decouple the two different problem with two different solution:

• • a real per-tone compensation that is dependent only on the mask of tones used. This compensation has to be performed into frequency domain to be computationally efficient. This is independent on the alignment;
  • a “low taps compensation” that is a compensation of some dangerous modes. It is required that the number of modes to be compensated has to be reduced, in particular to two, in order to have a high efficiency in terms of computational complexity. Said low taps compensation is proposed to be performed into frequency domain, but it can be easily derived that it can be equivalently done into the time domain. Said low taps compensation is align-dependent. The compensation vector can be obtained with a circulant shift of an original vector, so that it is not needed to access to a compensation vector for each align.

The proposed method is particularly suitable for implementation in the Central Office modem side, because at CPE (Customer Premises' Equipment) side the per-tone compensation is enough to have good performances for different alignment.

In the following a echo cancellation method that involves full compensation is described.

In the proposed full LMS method with T=T(k) is indicated the state vector matrix of the echo canceller filter, where w(k) indicates an estimated impulse response of a FIR echo path filter in the step (symbol) k, μ indicates the increment correction term and e=e(k) the error of updating.

The update w(k+1) of the FIR filter can be written as: w(k+1)=w(k)+μ·T^H·e  Eq. 1

Usually the increment correction term μ is considered a scalar correction. It can be considered, in frequency update environment, a per-tone correction.

In the proposed method it is provided to derive the energy introduced in the calculation per tone that means to derive the mean power (in each row) of FT(k)=F·TH  Eq. 2 where F is a Discrete Fourier Transform matrix, and to fix the per-tone correction A to the inverse of such quantity. The basic idea is to derive the statistical power of each tone vector and to compensate it. In other words it is proposed to extract an increment correction μ_k for each tone k: $\begin{array}{cc}{\mu }_k=\frac{{\mu }_0}{E\{\left[F{T}_{k.j}^2·1_{\mathrm{Nx}1}\right\}}& \mathrm{Eq}.3\end{array}$ where N indicates the dimension of the state vector matrix T.

In the proposed method it is considered like a square matrix correction.

In the following a stability and performance analysis of the full compensation LMS method based on the mean squared value of the estimation error e(k) is presented.

The weight-error vector is defined like difference of the estimated impulse response of the FIR filter w(k) and w₀ the optimum Wiener solution for the tap-weight vector at iteration k. ε(k)=w(k)−w₀  Eq. 4

Defined h the impulse response of the echo-channel it is also defined d=T·h, then the output of estimated can be written e(k)=d(k)−T(k)·w(k)  Eq. 5

Using equation 1 into eq. 4 and finally substituting eq. 5 it can be obtained the weight error vector expression: ε(k+1)=[I−μ·T^H·T]ε(k)+μ·T^H·e₀(k)  Eq. 6 where e₀(k) in the estimation error produced in the optimum Wiener solution e₀(k)=d(k)−T··w₀  Eq. 7

Now, the direct averaging method used in the proposed echo cancellation method is described.

Equation 6 is a stochastic difference equation in the weight-error vector ε(k). To study the convergence behavior of such stochastic algorithm in an average sense, it is possible to invoke a direct-averaging method, known from H. J. Kushner, “Approximation and Weak Convergence Methods for Random Processes with application to Stochastic System Theory”, MIT Press, Cambridge Mass., 1984. According to this method equation 6, for a small step size parameter μ, is close to the solution of another stochastic difference equation whose system matrix is equal to the ensemble average:

E[I−μ·T ^H ·T]=I−μ·R _T  Eq. 8

• • where R_T is the correlation matrix of the state matrix of the filter T. More specifically eq. 6 can be replaced with ε(k+1)=[I−μ·R_T]ε(k)+μ·T^H·e₀(k)  Eq. 9

The correlation matrix R_T is always nonnegative definite. Also the correlation matrix R_T is definite positive except for noise-free signal-processing problems.

Now, the stability of the Least Mean Square procedure is discussed.

To determine the condition for the stability of the proposed LMS update, a method which is similar to the steepest-descent algorithm stability approach is used. The natural modes of the method are examined.

To study the stability of this method the feedback loop term of equation 9 can be considered

ε(k+1)=[I−μ·R_T]ε(k)  Eq. 10

Using the unitary similarity transformation, the correlation matrix can be expressed as follow R_T=Q·Λ·Q^H  Eq. 11

The matrix Q has as its columns an orthogonal set of eigenvector associated with the Eigen value of the matrix R_T.

Also Q is a unitary matrix ( ) Q^H·Q=I=>Q^−q=Q 15 H and is a diagonal matrix whose elements are the Eigen values of the matrix R_T.

Substituting eq. 11 into eq. 10 it is obtained ε(k+1)=[I−μ·Q·Λ·Q^H]ε(k)  Eq. 12

Defining a new set of coordinates ∓as follow v(k)=Q^H·ε(k)  Eq. 13

• • and premultiplying both sides of equation 12 it is obtained v(k+1)=[I−μ_Q·Λ]v(k)  Eq. 14
  • where μ_Q=Q^H·μ·Q  Eq. 15

The matrix μ_Q can be seen like a diagonal matrix with element μ_Q,n. Indeed it is enough to compensate all the modes.

From eq. 14 the nth natural mode of the method can be written as v_n(k+1)=(1−μ_Q,n·λn)v_n(k), n=0, . . . , N−1  Eq. 16

Eq. 16 is a geometric series. For stability the magnitude of the geometric ratio must be less then 1.

For all the real and positive Eigen value a necessary and sufficient condition for the convergence or the stability is $\begin{array}{cc}0<{\mu }_{Q.n}<\frac{2}{{\lambda }_n}& \mathrm{Eq}.17\end{array}$

The n-th time constant can be expressed in terms of the step size parameter and Eigen value $\begin{array}{cc}{\tau }_n=-\frac{1}{\ln \left(1-{\mu }_{Q.n}·{\lambda }_n\right)}& \mathrm{Eq}.18\end{array}$

Applying the same speed of convergence (same time constant) to all the modes it is obtained μ_Q=μ₀·Λ⁻1  Eq. 19

• • and using eq. 15 and 19 it is finally obtained for the increment correction matrix μμ=μ₀·Q·Λ⁻¹·Q^H  Eq. 20
  • where 0<μ₀<2. If some null Eigen values are present they are not considered into the Eq. 20 inversion.

It is clear that all this operations are possible with the knowledge of the correlation matrix R_T.

Therefore it is shown in the following the method of derivation of the correlation matrix for a DMT signal.

The major problem of a DMT signal is that it is not a stationary signal.

The correlation of a signal x the function is r_X(τ,t)=E[x(t)·x*(t−τ)] Eq. 21

• • with t∈Z(T) since it is considered on a digital signal transmitter.

A DMT signal is a cyclostationary signal. This means r_x(τ,t)=r_x(τ,t+T_c)  Eq. 22

• • where Tc is the period of one DMT symbol. In Eq. 22 the presence of the cyclic prefix is not kept into account. It can be observed that the period of the DMT symbol can be related to the number of tones of the inverse fast fourier transform IFFT (N_T/2) and the length of cyclic prefix (v): T_C=T(v+N_T).

Into the following considerations will be done the hypothesis of uncorrelation between samples of different symbols.

The DMT Signal Just after demodulator s(t) is expressed as $\begin{array}{cc}s\left(t\right)=\frac{1}{N_T}\sum _{m=-\infty }^{+\infty }\sum _{k=0}^{N_T-1}{a}_m^k\sum _{n=-V}^{N_T-1}{e}^{\mathrm{j2\pi }kn/N_T}\delta \left(t-nT-m\left(N_T+v\right)T\right)& \mathrm{Eq}.23\end{array}$

Without loss of generality it is imposed T=1. Due to correlation periodic property it can be derived r_x(τ,t) for −v<t<N_T.

This means that it is possible to refer to only one symbol (m=0 for example). Imposing also, α_m^k=(α_m^NT−k)* with k=0, . . . , N_T/2−1, eq. 23 yields a real baseband signal.

From eq. 23 it is obtained $\begin{array}{cc}s\left(t\right)=\frac{1}{N_T}\sum _{k=0}^{N_T-1}{a}_m^k·e^{\mathrm{j2\pi }kt/N_T}\mathrm{for}-v\le t<N_T,m=0& \mathrm{Eq}.24\end{array}$

This equation is rewritten keeping into account the symmetry (hermitian) of the information per-tone.

Into the hypothesis that no information will be inserted into DC and Nyquist tones it is obtained $\begin{array}{cc}s\left(t\right)=\frac{1}{N_T}\sum _{k=0}^{N_T/2-1}{a}_k·e^{\mathrm{j2\pi }kt/N_T}+a_{N_T-k}·e^{\mathrm{j2\pi }\left(N_T-k\right)t/N_T}=\frac{1}{N_T}\sum _{k=0}^{N_T/2-1}{a}_k·e^{\mathrm{j2\pi }kt/N_T}+a_k^{*}·e^{\mathrm{j2\pi }kt/N_T}& \mathrm{Eq}.25\end{array}$

Eq. 25 can be rewritten in function of only sine and cosine. $\begin{array}{cc}s\left(t\right)=\frac{2}{N_T}\sum _{k=0}^{N_T/2-1}\mathrm{Re}\left[a_k\right]·\cos \left(2\pi kt/N_T\right)-\mathrm{Im}\left[a_k\right]·\sin \left(2\pi kt/N_T\right)& \mathrm{Eq}.26\end{array}$

Defined $\begin{array}{cc}G\left(t\right)={\left\{g_k\left(t\right)\right\}}_{k=0,\dots ,N_T-1}\mathrm{and}{g}_k\left(t\right)=\cos \left(2\pi kt/N_T\right)F\left(t\right)={\left\{f_k\left(t\right)\right\}}_{k=0,\dots ,N_T-1}\mathrm{and}{f}_k\left(t\right)=\sin \left(2\pi kt/N_T\right)A_{\mathrm{in}}={\left\{a_m^k\right\}}_{k=0,\dots ,N_T-1}& \mathrm{Eqs}.27\end{array}$

• • all the previous vectors have to be considered column vectors.

Substituting eqs. 27 which introduce a vector notation into eq. 26 it is derived the DMT Signal Just after demodulator s(t) $\begin{array}{cc}s\left(t\right)=\frac{2}{N_T}\left({G\left(t\right)}^T·\mathrm{Re}\left[A_0\right]-{F\left(t\right)}^T·\mathrm{Im}\left[A_0\right]\right)& \mathrm{Eq}.28\end{array}$

Now, the method for deriving the Correlation function Just after modulator is described

The correlation of a DMT signal is derived under the hypothesis

• • uncorrelation of samples of different symbols;
  • uncorrelation between real and imaginary information per-tone E{Re[ak]Im[ak]}=0;
  • uncorrelation between information on different tones;
  • the statistical power for real and imaginary part is the same for each tone. This means that each real/imaginary power with the second hypothesis is the half of the statistical power of each complex signal per-tone.

From the previous hypothesis it can be concluded that:

• • the correlation can be defined between −∓[t<N_T due to its periodic property. It is possible to refer for the moment to a correlation function not null only in this period;
  • the correlation (into one period) into the hypothesis of uncorrelation between information of different symbols is $r_x\left(\tau ,t\right)=\{\begin{array}{cccc}0& & \mathrm{otherwise}& \\ \ne 0& -v\le t-\tau <N_T& \mathrm{and}& -v\le t<N\end{array}$

The correct correlation of the signal can be finally written as which $r_{\mathrm{xp}}\left(\tau ,t\right)=\sum _{m=-\infty }^{+\infty }{r}_x\left(\tau ,t-{\mathrm{mT}}_C\right)$ keeps into account the periodicity of the signal.

It is proposed to derive the not null part of the correlation. $s\left(t\right)=\frac{2}{N_T}\sum _{k=0}^{N_T/2-1}\mathrm{Re}\left[a_k\right]·\cos \left(2\pi kt/N_T\right)-\mathrm{Im}\left[a_k\right]·\sin \left(2\pi \mathrm{kt}/N_T\right)$

From Eq. 28, under the hypothesis of uncorrelation between real and imaginary information per-tone the cross multiplication can be considered null. $\begin{array}{cc}E\left[s\left(t\right)·s\left(t-\tau \right)\right]=\frac{2}{N_T}E\left\{G^T\left(t\right)·\mathrm{Re}\left[A_m\right]·{\mathrm{Re}\left[A_m\right]}^H·G^{*}\left(t-\tau \right)+F^T\left(t\right)·\mathrm{Im}\left[A_m\right]·{\mathrm{Im}\left[A_m\right]}^H·F^{*}\left(t-\tau \right)\right\}& \mathrm{Eq}.29\end{array}$

From Eq. 29 using hypothesis of uncorrelation between information on different tones the central multiplication can be considered a diagonal matrix. Defined E[Re[α_k]·Re[α_k]*]=σ_kr²  Eq. 30 E[Im[α_k]·Im[α_k]*]=σ_k,j²

From Eq. 29 and 30 it is obtained $\begin{array}{cc}E\left[s\left(t\right)·s\left(t-\tau \right)\right]=\frac{4}{N_T^2}\{G^T\left(t\right)·\mathrm{diag}\left[\left\{{\sigma }_{k,r}^2\right\}\right]·G^{*}\left(t-\tau \right)+& \mathrm{Eq}.31\\ F^T\left(t\right)·\mathrm{diag}\left[\left\{{\sigma }_{k,i}^2\right\}\right]·F^{*}\left(t-\tau \right)\}& \\ \mathrm{or}& \\ E\left[s\left(t\right)·s\left(t-\tau \right)\right]=\frac{4}{N_T^2}\sum _{k=0}^{N_T/2-1}{\sigma }_{k,r}^2·\cos \left(2\pi kt/N_T\right)·& \mathrm{Eq}.32\\ \cos \left(2\pi k\left(t-\tau \right)/N_T\right)+{\sigma }_{k,i}^2·\sin \left(2\pi kt/N_T\right)·\sin (2\pi k\left(t-\tau \right)/N_T& \end{array}$

Under the fourth hypothesis, that the statistical power for real and imaginary part is the same for each tone, it is easily obtained $\begin{array}{cc}{r}_x\left(\tau ,t\right)=E\left[s\left(t\right)·s\left(t-\tau \right)\right]=\frac{2}{N_T^2}\sum _{k=0}^{N_T/2-1}{\sigma }_k^2·\cos \left(2\pi k\tau /N_T\right)& \mathrm{Eq}.33\end{array}$

• • where with σ_k is indicated the per-tone power.

It can be observed that into its domain of validity, that is function of both t and τ, it is related only to τ. For this reason into the next considerations it will be referred to r_x(τ)=r_x(τ,t) for −v<t−τ<N_T and −v<t<N_T.

The cyclostationarity is due to the validity of eq. 32 and 33, i.e. the correlation is null outside the symbol.

An efficient derivation procedure of eq. 33 will be shown in a further part of the description.

In the following the derivation of the correlation function of a filtered DMT signal in the proposed method will be described.

A DMT signal filtered by a time-invariant FIR with impulse response h is considered.

The output of the filter is cyclostationary if the input is cyclostationary.

The correlation is derived into time to time function because is more efficient of a Spectral Density passage.

The output y (domain U) of a filter h ( ) with input signal x (domain I) is (generalized Haar integration) y(t)=·_ldu·h(t−u)×x(u)  Eq. 34

Defined s_y(t1, t2)=E[y(t1)·y*(t2)]  Eq. 35 s_x(t1, t2)=E[x(t1)×(t2)]

• • it is obtained s_y(t1,t2)=∫_ldu1∫_ldu2·h(t1−u1)·s_x(u1,u2)·h*(t2−u2)  Eq. 36

It is more interesting to rewrite eq. 36 using eq. 21 definition r_y(τ,t)=s_y(t,t−τ)=∫_ldu2·h*(t−τ−u2)∫_ldu1·h(t−u1)·r_x(u1−u2,u1)  Eq. 37

Introducing now the hypothesis of cyclostationarity of input signal (domain∈Z(T)/Z(T_c)) the previous relation can be written r_yp(τ,t)=∫_ldu2·h*(t−τ−u2)∫_ldu1·h(t−u1)·r_sp(u1−u2,u1)  Eq 38

Keeping into account that I=Z(T) and I_C=Z(T)/Z(T_c) it is possible to rewrite eq. 38 as $\begin{array}{cc}{r}_{\mathrm{yp}}\left(\tau ,t\right)=\sum _{\mathrm{u2}=-\infty }^{+\infty }{h}^{*}\left(t-\tau -\mathrm{u2}\right)·\sum _{\mathrm{u1}=-v}^{N_T-1}h\left(t-\mathrm{u1}\right)·r_{\mathrm{xp}}\left(\mathrm{u1}-\mathrm{u2},\mathrm{u1}\right)& \mathrm{Eq}.39\end{array}$

For a more efficient derivation of eq. 39 it is possible to refer to r_x( , ).

It can be observed that $\begin{array}{cc}{r}_y\left(\tau ,t\right)=\sum _{\mathrm{u2}=-y}^{N_T-1}{h}^{*}\left(t-\tau -\mathrm{u2}\right)·\sum _{\mathrm{u1}=-y}^{N_T-1}h\left(t-\mathrm{u1}\right)·r_x\left(\mathrm{u1}-\mathrm{u2},\mathrm{u1}\right)& \mathrm{Eq}.40\end{array}$

And, to keep account the periodicity of the output, $\begin{array}{cc}{r}_{\mathrm{yp}}\left(\tau ,t\right)=\sum _{k=-\infty }^{+\infty }{r}_y\left(\tau ,t-k·T_C\right)& \mathrm{Eq}.41\end{array}$

An efficient derivation of eq. 40 is proposed further on in the description.

Now, for what concerns the derivation of state vector correlation matrix R_T, defined the square matrix T with dimension N the state vector matrix of EC FIR filter. $T=\left[\begin{array}{ccccc}{x}_D& x_{D-1}& x_{D-2}& \cdots & x_{D-N+1}\\ x_{D+1}& x_D& x_{D-1}& & \\ x_{D+2}& x_{D+1}& x_D& ⋰& \\ ⋮& & ⋰& ⋰& x_{D-1}\\ x_{D+N-1}& & & x_{D+1}& x_D\end{array}\right]$

• • and R_T=E[T^H·T] Eq. 42
  • it is easy, now, to derive R_T correlation matrix $\begin{array}{cc}{\left[R_T\right]}_{i,j}=\sum _{k=0}^{N-1}{r}_{\mathrm{yp}}\left(i-j,k+D-\min \left(i,j\right)\right)& \mathrm{Eq}.43\end{array}$

It can be seen that correlation matrix R_T is symmetric.

In FIGS. 3A and 3B are shown the result of simulation of the full compensation method, using the update defined in Eq. 1, where the increment correction matrix l is derived from Eq. 20 and the correlation matrix RT is derived from Eq. 43.

The complexity of this increment scaling is, in the proposed implementation due to the decimation before the FIR echo canceller, 256² multiplication to be added to the standard update using Full-LMS complexity.

In FIGS. 3A and 3B a full compensation increment correction matrix for a misalignment of 66@552 Mhz is shown. In FIG. 3A it is shown the power (amplitude) of the increment matrix per tap in time domain (taps 0-255).

In FIG. 3B it is shown the power of the increment matrix per tone in frequency domain (tones 0-128)

In FIGS. 4A and 4B the full compensation increment correction is shown. In FIG. 4A it is shown the power of residual echo derived into the de-mapper in different group of tones, 0 to 32, 33 to 59 and 60 to 64 in particular, in function of the number symbols. In FIG. 3B it is shown the per-tone mean power of far-end signal, received echo, residual echo, distortion and other noises into the de-mapper.

In this condition the performances are:

• • CDC without echo: 388 kb/s
  • CDC with echo: 388 kb/s

That means a loss of performance of 0%.

The performances of this method are really better then the performances of per-tone increment update version shown with reference to FIGS. 2A and 2B. With the proposed method the convergence can be obtained in less then 600 of symbol instead of more then 8000 symbols (in 8000 symbols the convergence was not yet obtained, see FIG. 2A, in other simulation to obtain the same level of residual echo are needed approximately 40000 symbols).

The proposed full LMS compensation method gives very good results but it is also very complex.

The increment correction matrix is alignment dependent and requires storing a lot of information. The alignment at Central Office modem cannot be known a priori, so ideally 256² real number should be stored for each alignment. It was proved that 16 bit are just enough to obtain good performances behavior. This is the major problem to be solved for its implementation.

As it can be seen in FIG. 3A the full compensation increment correction matrix μ can be seen like the sum of two components:

• • the first one is due to the standard update, in which the compensation pushes the diagonal components, i.e. it is increased the tap to tap contribution. It is apparent that also the neighboring taps are correlated with this one so it is possible to understand that the compensation increment correction matrix μ decrease with the distance from the main diagonal
  • the second component is the ‘cross’ over the full compensation increment correction matrix μ that can be noticed in FIG. 3A. It is readily understandable that such a ‘cross’ pushes some taps over the impulse response and this explains the reason of the non-convergence in some taps without full compensation.

Also it can be easily observed that the position of the ‘cross’ is dependent on the alignment and the delay of Echo Path.

In the simulation the impact of low-mode problem is more complex, it depends also on the level of impulse response energy in the ‘low taps’. It is clear that if in those taps there is no energy then no problem of convergence can be found and vice versa.

From the analytic study it is clear that low taps problem is due to cyclic prefix influence onto the correlation properties. Indeed this problem is related to the taps corresponding to the Cyclic Prefix taps.

Therefore in the following a low-complexity embodiment of the proposed method, for the 32-taps problem, it will be described. Said embodiment can be regarded as a simplification of the Full LMS Compensation method disclosed above.

The basic idea underlying said embodiment is to support the optimal per-tone correction and to add one or more correction terms to obtain the convergence.

The full compensation method can be rewritten like sum of each Eigen value/vector contribution of the correlation matrix. Eq. 1 can be written using eq. 20 $\begin{array}{cc}w\left(k+1\right)=w\left(k\right)+{\mu }_0·\sum _{n=0}^{N-1}{\lambda }_n^{-1}·q_n·q_n^H·e_T& \mathrm{Eq}.44\end{array}$

• • where λn is the Eigen value of the correlation matrix and q_n is its corresponding eigenvector associated to a certain natural mode.

Since from observation of FIGS. 3A and 3B turns out that the per-tone optimal correction is able to push for a lot of natural modes but not some low modes, i.e. the modes that generate low-taps problem, it is intended to add some contribution able to push said low modes.

The frequency domain per-tone optimal increment matrix μ_f,PT can be defined as μ_f,PT=diag({μ_k}_{k=0, . . . , N−1})  Eq. 45 where the increment μ_k with k=0, ..., N−1, is defined further on the in the description, in particular with reference to eq. 57.

The low-mode compensation update can be written $\begin{array}{cc}w\left(k+1\right)=w\left(k\right)+{\mu }_0·F^{-1}{\mu }_{f.\mathrm{PT}}·F·e_T+{\mu }_0·\sum _{n=0}^{N_{\mathrm{LM}}-1}{\alpha }_n^{-1}·p_n·p_n^H·e_T& \mathrm{Eq}.46\end{array}$ where α_n is the Eigen value of the correlation matrix and p_n is its corresponding eigenvector that is associated to the low modes and N_LM is the number of Low Modes.

With this method it is reduced both the computation complexity and the memory needed to store the Eigen value/vector.

Here the Eigen values/vector of the residual matrix μ_res=μ_full−F⁻¹μ_f,PT·F  Eq. 47

• • are studied for different misalignments. In FIG. 5A some Eigen values of the residual matrix μ_res for different value of Misalignment-Delay of Echo Path are shown. In FIG. 5B the eigenvector corresponding to the first and biggest Eigen value for different values of Misalignment-Delay of Echo Path (into @2.2 MHz) is shown.

As it can be seen in FIG. 5A, there is a clear difference between the first 2 Eigen values, indicated with 1 EV and 2 EV, an all the others. The first Eigen value is usually related to the low taps problem but the second Eigen value is related to another phenomenon, which do not permit the normal convergence in the marginal taps (first and last 20 taps). The inconvenient due to such second phenomenon can be dealt with the same approach.

Also the third Eigen value/vector is related to the original low-taps problem but the Eigen value is really smaller.

It can be also observed that this problem relevant only with a misalignment between −26 and +449@2.2 MHz (TBC). In this region the value of the corresponding first Eigen value is very flat.

Also it can be observed from FIG. 5B that in this region, the first eigenvector is approximately obtained with a circulant shift.

That is not true between +450 and +519.

In FIGS. 6A and 6B are shown the results of simulation of the low taps compensation method obtained using eq. 1 update, where the increment correction matrix is derived from eq. 20 and the correlation matrix is derived from eq. 43.

The complexity of this increment scaling is, in our implementation due to the decimation before the FIR echo canceller, 256² multiplication to be added the standard update (Full-LMS) complexity. It can be observed that the additional complexity is more than double than the complexity of the base-update.

In FIG. 6A it is shown for the Low mode compensation correction with a misalignment of 66@552 kHz, the power of residual echo derived into the de-mapper in some group of tone vs the number symbols.

In FIG. 6B it is shown the Low mode compensation correction with a misalignment of 66@552 kHz, the per-tone mean power of far-end signal, received echo, residual echo, distortion and other noises into the de-mapper.

In this condition the performances are:

• • CDC without echo: 388 kb/s;
  • CDC with echo: 388 kb/s

That means a loss of performance of 0%.

The performances are a little worse than the full compensation case. This is due to:

• • the frequency per tone compensation is not the same (we have used the mean case);
  • the fact that not all the low taps modes are pushed.

However performances are better than those of known methods, i.e. the method shown with reference to FIG. 2. The convergence can be considered obtained after 2000 symbols.

In the following efficient derivation methods for the above equations are described

An efficient proposed derivation of eq. 33 is based on the only hypothesis that all used tones have the same power level for D1<σ_k²=σ² k<D2 with D2<N_T. In this case eq. 33 can be written $\begin{array}{cc}{r}_x\left(\tau \right)=\frac{{\sigma }^2}{N_T^2}\sum _{k=\mathrm{D1}}^{\mathrm{D2}}\left(e^{j2\pi \mathrm{kt}/N_T}+\mathrm{cc}\right)& \mathrm{Eq}.48\end{array}$

• • where cc indicates complex conjugate. Collecting an exponential term and solving the exponential finite series it is obtained $\begin{array}{cc}{r}_x\left(\tau \right)=\frac{2·{\sigma }^2}{N_T^2}\frac{\sin \left(2\pi \left(\mathrm{D2}-\mathrm{D1}+1\right)\tau /2N_T\right)}{\sin \left(2\mathrm{\pi \tau }/2N_T\right)}\cos \left(2\pi \left(\mathrm{D2}+\mathrm{D1}\right)/2N_T\right)& \mathrm{Eq}.49\end{array}$
  • or using the definition of periodic since $\begin{array}{cc}{r}_x\left(\tau \right)=\frac{2·{\sigma }^2\left(\mathrm{D2}-\mathrm{D1}+1\right)}{N_T^2}·{\mathrm{Sinc}}_{\mathrm{D2}-\mathrm{D1}+1}\left(2\pi \left(\mathrm{D2}-\mathrm{D1}+1\right)\tau /2N_T\right)·\cos \left(2\pi \left(\mathrm{D2}+\mathrm{D1}\right)/2N_T\right)& \mathrm{Eq}.50\end{array}$

An efficient derivation of eq. 40 can be obtained, considering the correlation function of the signal just after the modulator (the non-periodic one) function only of τ parameter (−v<t−T<N_T and −v<t<N_T).

In conclusion eq. 40 can be rewritten as $\begin{array}{cc}{r}_y\left(\tau ,t\right)=\sum _{\mathrm{u2}=-v}^{N_T-1}{h}^{*}\left(t-\tau -\mathrm{u2}\right)·\sum _{\mathrm{u1}=-v}^{N_T-1}h\left(t-\mathrm{u1}\right)·r_x\left(\mathrm{u1}-\mathrm{u2}\right)& \mathrm{Eq}.51\end{array}$

The complexity of this derivation is (NT+v)² for each couple t, τ. Eq. 1 cannot be used directly into the computation of correlation matrix due to complexity reason.

It is proposed to define the new function $\begin{array}{cc}{u}_x^N\left(t,u\right)=\sum _{\mathrm{u1}=-v}^{N-1}h\left(t-\mathrm{u1}\right)·r_x\left(\mathrm{u1}-u\right)& \mathrm{Eq}.52\end{array}$

By observing that the value of this function into (t+1,u+1) is strongly related to the one into (t+1,u+1), it is possible to rewrite Eq. 51 $\begin{array}{cc}{r}_y\left(\tau ,t\right)=\sum _{\mathrm{u2}=-v}^{N_T-1}{h}^{*}\left(t-\tau -\mathrm{u2}\right)·u_x^{N_T}\left(t,\mathrm{u2}\right)& \mathrm{Eq}.53\end{array}$ Referring to Eq. 53 into the (t+1) parameter and proving that is strongly related to the value into t and further considering the even symmetry of the r_x( ) function then it is obtained for the correlation function r_y(τ,t+1)=r_y(τ,t)+h*(t−τ+v+1)·u_x^NT(t+1,−v)−h*(t−τ−N_T+1)·u_x^Nr(t,N_T−1)+h(t+1+v)·u_s^NT−1(t−τ,−v−1)−h(t+1−N_T)·u_x^NT−1(t−τ,N_T−1)  Eq. 54

This equation is an iterative solution into the t parameter.

This derivation is very efficient, some of the terms are function of t, other of (t−τ). This can help to reduce the complexity because this terms can be derived for all the (t, τ) couples.

Other methods to reduce the complexity are related to the hypothesis that it is possible to make on the impulse response (h(.)) for example:

• 1. Causal: h(n)=0 for n<0
• 2. Finite length: h(n)=0 for n≧N_h

Observing that r_y(τ,t) correlation function is not zero for

• 1. t−τ≧−v
• 2. t−τ−N_T+1≧N_h this bound can be used to derive the correlation in a more efficient way: $\begin{array}{cc}{r}_y\left(\tau ,-v\right)=h\left(0\right)·\sum _{\mathrm{u2}=-v}^{N_T-1}{h}^{*}\left(-\tau -\mathrm{u2}-v\right)·r_x\left(-v-\mathrm{u2}\right)=h\left(0\right)·u_x^{N_T}\left(-v-\tau ,-v\right)& \mathrm{Eq}.55\end{array}$

For what concerns per tone compensation, with full LMS method, there is a different value of energy put for different tones. This is due to PSD shaping of DMT ADSL signal.

A good increment correction have to derive proportionally to the statistical power of input vector. The proposed solution, as mentioned above, is to make this for each per-tone vector.

In other words it has to be derived RF=E└F·T^H·T·F⁻¹┘  Eq. 56 and to fix $\begin{array}{cc}{\mu }_k=\frac{{\mu }_0}{{\mathrm{RF}}_{k,k}}k=0,\dots ,N-1& \mathrm{Eq}.57\end{array}$

Since it is known that DMT signal is a non-stationary signal, so generally Eq. 56 is alignment dependent. In this case it is desirable a scaling factor that have to be considered fixed (so not align-dependent). In this case it is considered the mean spectral analysis {overscore (r)}_sp(τ)=∫_Icdt·r_xp(τ,t)  Eq. 58

Being the correlation (not the periodic one) not null for −v[t−τ<N from eq. 33, 41 and 58 for a DMT signal is obtained $\begin{array}{cc}{\stackrel{\_}{r}}_{\mathrm{xp}}\left(\tau \right)=\{\begin{array}{cc}\left(1-\frac{\left|\tau \right|}{N+v}\right)·r_x\left(\tau \right)& \left|\tau \right|<N+v\\ 0& \mathrm{otherwise}\end{array}& \mathrm{Eq}.59\end{array}$

The mean correlation after the Low Pass Filter of echo-emulation (with h impulse response) path is easily {overscore (r)}_yp(τ)=∫_Icdu·{overscore (r)}_xp(u)·c_h(τ−u)  Eq. 60 where c_h(τ)=∫_Udv·h(v+τ)·h*(v)  Eq. 61

A reduced complexity update method can be also used, that will be referred to as “Lambda Even” update. Such a Lambda Even update is described in the European Patent Application No. 02291217.4 in the name of the same Applicant and it was proved in to have good convergence properties, in connection with the full and low modes compensation procedure to improve the performance.

In FIG. 7 an ADSL transceiver of DMT modem is shown, comprising a transmission branch 100 and a reception branch 200 cross connected by an echo path 400, usually through the hybrid transformer.

On the transmission branch 100 in a block 101 a inverse Fast Fourier transform operation from N real to N/2 complex is performed on the encoded signal to be transmitted and the CP Cyclic Prefix is added. The signal is then transmitted on the channel via digital to analog conversion in a block 103, preceded by filtration stage 102 comprising a biquadratic filter and shaping filters. Interpolation filters can be used as an alternative. The signal transmitted is defined in ITU-T G.992.1, G992.3, G992.5 standards.

Upstream the block 102 the transmitted signal is taken and sent to an adaptive filter 300. Signal is taken before the filtration stage 102 in order to increase the speed of convergence.

On the reception branch 200 a received signal is passed through an analog to digital converter indicated with 201, passed through a filtration block 202 comprising a biquadratic filter to reach a time domain equalizer (TEQ) 205. Downstream the time domain equalizer 205 is disposed a subtraction node 206 where the output of adaptive filter 300 is subtracted from the received signal to cancel echo. It has to be noted that the DSP front-end is not very important. What is important is that the echo signal is removed before the time domain equalizer 205 to avoid to insert TEQ variation in a adaptive equalization solution.

Subsequently, in a block 207 the Cyclic Prefix is removed and a Fast Fourier transform operation from N real to N/2 complex is performed, followed by a frequency equalization (FEQ) performed in a block 208.

Finally the received signal reaches a de-mapper (DEMAP) 209 that performs de-codification of the symbols. The input and the output of de-mapper 209 are however taken and subtracted in a subtraction node 210, which output is then feed to a multiplication node 316 of the adaptive filter 300.

It is important to perform a decision of the received symbol even if it is a hard decision. The algorithm implemented in the de-mapper 209 is very robust and can work even in presence of high probability of wrong decision.

Such an adaptive filter 300 comprises a filtering part 301 and an update part 302. The filtering part 302 comprises a biquadratic filter 303 and a FIR filter 304 with 256 taps.

The signal is taken upstream the FIR filter 304 and sent to the update part 302, firstly to a block 305 performing a Fast Fourier Transform operation from N real to N/2 complex, that yields an even array of values Aeven and an odd array of values Aodd, that represent the odd and even output of the Fast Fourier Transform on on 2*N samples. As it is based on a decimated transmitted signal, N sample are used. The odd array Aodd undergoes a conjugation operation in a block 306, which result is fed to a multiplication node 313. The even array Aeven undergoes a conjugation and hermitian reduction operation in a block 309, which result is fed to a multiplication node 314.

The output of subtraction node 209 is mixed in the multiplication mode 316 with the result of a block 316 that calculates the norm of the output of the frequency equalizer 208. The output of node 316 is fed to a block 307 that performs in association with a block 308 the classical echo cancellation according to a full LMS efficient method.

The output of block 307 is multiplied in a multiplication node 313 by the output of block 306 and fed to a block 308, that provides for removing the aliasing components in reception inserted by the last decimation stage contained in block 202, by replicating a complex FFT. The result is fed then to a multiplication node 315.

To said multiplication node 315 is also sent the output of a further multiplication node 314, that multiplies the output of block 309 with output of a block 310 that performs an operation of hermitian extension on the output of node 316.

The output of node 315 is fed to a block 311 in which the calculation of the increment correction, calculating the compensation for the two lower modes, is performed and feed such increment correction to a block 312 where the new update is calculated. The output of block 312 with update is then fed to the FIR filter 304.

In FIG. 8 it shown an ADSL transceiver implementing the Lambda Even update with low tap compensation.

As it can be seen the lambda odd array evaluation is suppressed.

From the transmitted signal taken upstream the FIR filter 304 only the even array Aeven is obtained via a folding operation in block 325. Then a real Fast Fourier transform operation is performed in block 326. Then he even array undergoes a conjugation and hermitian reduction operation in the block 309, which result is fed to a multiplication node 314.

The output of subtraction node 209 is mixed in the multiplication mode 316 with the result of a block 317 that calculates the norm of the output of the frequency equalizer 208.

The output of node 314 is then fed to the block 311 in which the calculation of the increment correction is performed and feed such increment correction to a block 312 where the new update is calculated. The output of block 312 with update is then feed to the FIR filter 304.

The proposed method, by introducing low modes compensation increases the performances of the echo canceller adaptive method.

The proposed method advantageously obtains a speed of convergence that is approximately the same for different alignments. Further, the efficiency of the adaptive method is increased, as can be seen comparing the mean quantitative of excess noise cancelled per update normalized by the computational complexity of each update. It was shown that this method can be ten times more efficient during tracking than all other methods known from the prior art.

The advantage in efficiency so obtained can be traded both to reduce the computational complexity, thus obtaining a reduced digital area for a hardware implementation or a less complex Digital Signal Processing unit in a software implementation, or to increase the number of update in the same time unit.

Without prejudice to the underlying principle of the invention, the details and embodiment may vary, also significantly with respect to what has been discussed just by way of example without departing from the scope of the invention, ad defined by the claims that follow.

For instance the solution described herein can be applied to all ADSL modem, in Central Office or Customer Premises' Equipment.

Claims

1. A method for canceling the echo part of a received DMT signal in the presence of a far-end signal in ADSL modem, comprising: adaptive digital filtration of the DMT signal using multiple taps, implemented through a full Least Mean Square error procedure, including estimating an impulse response and evaluating an update of the estimation using a correction increment parameter; and performing a per-tone compensation that includes evaluation of a per-tone increment correction that depends on a mask of tones used.

2. The method of claim 1 further comprising: evaluating a per-tone optimal increment correction matrix; performing a low taps compensation reducing the number of modes to be compensated in the update to a subset of lower modes; and adding to the per-tone optimal increment correction matrix one or more compensation terms to obtain convergence.

3. The method of claim 2 wherein adding one or more compensation terms comprises adding two compensation terms.

4. The method of claim 1 wherein the per-tone compensation is performed in the frequency domain.

5. The method of claim 2 wherein the low taps compensation is performed in the frequency domain.

6. The method of claim 2 wherein the low taps compensation is performed in the time domain.

7. The method of claim 2 wherein the low taps compensation is align-dependent.

8. The method of claim 1 wherein a compensation vector can be obtained with a circulant shift of an original vector.

9. The method of claim 1 wherein evaluating the mask of tones used in the per-tone compensation is performed assuming: uncorrelation of samples of different symbols; uncorrelation between real and imaginary information per-tone; uncorrelation between information on different tones; and the statistical power for the real and imaginary parts are the same for each tone.

10. The method of claim 1 wherein evaluating the mask of tones used in the per-tone compensation step assumes that all used tones have the same power.

11. The method of claim 1 wherein the per-tone compensation comprises evaluation of a per-tone increment correction by deriving each per-tone vector proportionally to the statistical power of an input vector.

12. The method of claim 1 wherein the update is calculated using a Lambda even update procedure that provides for calculating a subarray of the Eigen values of the correlation matrix.

13. A DMT modem apparatus for canceling the echo part of a received DMT signal in the presence of a far-end signal in ADSL modem, comprising: an adaptive filter including: adaptive digital filtration means using multiple taps, implemented through a full Least Mean Square error procedure, including estimating an impulse response and evaluating an update of the estimation using a correction increment parameter; and means for performing a per-tone compensation that includes evaluation of a per-tone increment correction that depends on a mask of tones used.

14. The apparatus of claim 13 wherein the DMT modem comprises a Central Office modem.

15. A computer program product directly loadable in the memory of a digital computer and comprising software code portions for performing, when the product is run on a computer, a method for canceling the echo part of a received DMT signal in the presence of a far-end signal in ADSL modem, the method comprising: adaptive digital filtration of the DMT signal using multiple taps, implemented through a full Least Mean Square error procedure, including estimating an impulse response and evaluating an update of the estimation using a correction increment parameter; and performing a per-tone compensation that includes evaluation of a per-tone increment correction that depends on a mask of tones used.