FR2789244A1 - Equaliser with weighted decision return includes direct and recursive filters providing reliability function to reduce errors - Google Patents

Abstract

The filtering and calculation of weighted coefficients enables error transmission to be minimised by transmitting a reliability function. The equaliser with return of decisions comprises a direct filter (FF, 11) and a recursive filter (FB, 12) whose filtering coefficients are determined (13) in such a way as to optimise a predetermined quality factor. The equaliser receives the input data (x), supplying the direct filter (FF) and providing the equalised data (y') corresponding to the sum of the outputs of the direct filter (FF) and the recursive filter (FB). The recursive filter (FB) is supplied by the weighted data (y-recursif), whose value takes account of the equalised data (y) and corresponding reliability information.

Description

Weighted decision equalizer and method

corresponding equalization.

  The field of the invention is that of the reception of digital signals, in particular when the latter are transmitted via disturbed transmission channels. More specifically, the invention relates

the equalization of such signals.

  A digital signal receiver can be arbitrarily broken down into a series of modules, comprising a radio frequency (RF) head, connected to the receiving antenna, an analog / digital converter, an equalizer and a channel decoder. The role of this last module is to improve the performance of the transmission, in particular with regard to noise, as a function of the corresponding coding implemented.

works on the show.

  The purpose of the equalizer is to combat imperfections due to the transmission channel (for example echoes, fainting, "fadings", ...). In other words, its purpose is to reshape the received signal, or even "to open the eye" (with reference to the classic

"eye diagram").

  The equalization is based on three elements one (of) filter (s); - a quality criterion;

  - an algorithm for calculating the filter coefficients.

  This algorithm determines the coefficients which optimize (that is to say maximize, for example if the criterion is the opening of the eye, or minimize, not example if the criterion is a mean square error

  (EQM)) filtering, therefore equalization.

  Conventionally, the equalizer is therefore based on the implementation of a

closed loop.

  We can distinguish two main types of equalizers, depending on their use - adaptive equalizers, the quality criterion of which is determined from a reference, known to the receiver and emitted by the transmitter (to the detriment, of course, of the bit rate useful); - self-adaptive equalizers (or self-taught, or "blind" in English), not requiring references (the decisions after decoding being, for example, in a so-called operation

  "decision directed", used as references).

  We can also distinguish three main structures of equalizers: finite impulse response (RIF) filters, corresponding to conventional transversal filters; - linear infinite impulse response filters (linear RII), implementing a direct RIF filter ("forward") and a recursive RIF filter ("feedback") looped back to the output of the equalizer; - non-linear infinite impulse response filters (non-linear RII), which take into account decisions made on the value of receipts (these decisions being introduced into the recursive filter). The invention relates more precisely to the latter type of equalizers, conventionally called decision feedback equalizers (ERD, or DFE in English). The operation of the equalizer is generally broken down into two phases: a convergence phase (which may or may not use references),

and a pursuit phase.

  The invention can be used for the tracking phase, as well as for the convergence phase. Another technique (for example using

  references) can also be used for the convergence phase.

  For more details on the various known equalizers and their implementation, we can refer to the article by Odile MACCHI

  "Digital equalization in communications" (Ann. Télécomun .; 53, n0 1-

  2, 1998), which presents a synthesis on the various existing techniques.

  As this article points out, the maximum likelihood equalizer (MLSE) is the most effective. However, it has various drawbacks, and in particular a great complexity, which does not allow its implementation in practice (in particular for modulations with large

number of states).

  The ERD therefore appears to be the system offering the best ratio between

  equalization performance and complexity of implementation.

  However, it suffers from a defect inherent in its very structure: the phenomenon of propagation of errors. This phenomenon is due both to the injection of errors in the recursive part of the equalizer, and to the divergence of the algorithm for calculating the filter coefficients, when the quality criterion is distorted by these errors. It is moreover very difficult to distinguish the

  respective share of each contribution in the divergence of the ERD.

  This problem is a major problem, well known and not effectively resolved. Because of this risk, which is difficult to quantify and predict, we see that ERD is very little used in practice. The only known uses relate to very slightly disturbed channels, for example for the broadcasting of digital television on cable network (American proposal

(FCC).

  However, a technique has been proposed by J. LABAT, O. MACCHI and C. LAOT, in the article "Adaptive Decision Feedback Equalization: can

  you skip the training period? "(IEEE trans. On com., vol. 46, no 7, July 98).

  It consists in performing an ERD equalization only when the channel is already correctly equalized by another method. This supposes a structural communication and reversible algorithm, according to a certain criterion (the EQM

in this article).

  The implementation of this technique is therefore relatively

heavy and complex.

  Another method, of course, is to transmit a sufficiently long reference sequence very regularly in order to reset the ERD during a propagation of errors. The obvious drawback of such a method is a significant consumption of flow which is unnecessary for the service. It is therefore clearly not a solution for the future, whereas the

  there is a tendency for all services to increase throughput.

  The invention particularly aims to overcome these different

disadvantages of the state of the art.

  More specifically, an objective of the invention is to provide a technique of equalization with return of decisions which limits the effect of

  propagation of errors, compared to known techniques.

  Another objective of the invention is to provide such an equalization technique, which is not very complex to implement compared to the

  technique already cited by Labat and Macchi.

  The invention also aims to provide such an equalization technique, which does not require switching with other methods

  equalization, in certain situations.

  These objectives, as well as others which will appear subsequently, are achieved, using a decision feedback equalizer, comprising a direct filter (FF) and a recursive filter (FB), the filter coefficients of which are determined so as to optimize a predetermined quality criterion, said equalizer receiving input data, supplying said direct filter (FF), and delivering equalized data, corresponding to the sum of the outputs of said

  direct filter (FF) and said recursive filter (FB).

  According to the invention, said recursive filter (FB) is supplied with weighted data, the value of which takes account of said equalized data and

  corresponding trusted information.

  Thus, the symbol (or data) introduced into the recursive filter is no longer (at least not systematically) the equalized data, which sometimes carries an error which will propagate. To avoid this propagation, this data is modified in an adequate manner, when the confidence assigned to the

  equalized data is not sufficient.

  The invention therefore introduces a weighting, a function of confidence,

  which makes it possible to fight effectively against the propagation of errors.

  It will be noted that the invention can be implemented for the phase of

  further equalization, but also for the convergence phase.

  Advantageously, the optimization of said quality criterion takes

  also has said trusted information.

  In other words, the error taken into account to control the algorithm for calculating the filter coefficients is no longer systematically the difference between the equalized data and the corresponding most probable symbol. At least when the confidence is insufficient, this error is adapted, so as to

  reduce the phenomenon of propagation of errors.

  According to a preferred approach of the invention, the equalizer comprises means for calculating said information of confidence in said equalized data, and means for using said information of

  confidence, feeding said recursive filter (FB).

  Different methods can be used for the calculation and

  for the use of trusted information.

  According to a first advantageous embodiment of the invention, said confidence information takes account of at least a distance between said equalized data and the most probable symbol corresponding to said data. In particular, said distance can be compared to at least one threshold,

  the processing of said data being a function of the result of the thresholding.

  In this particular case, that said at least one threshold can be optimized iteratively. Of course; it can also be fixed, or adjustable at the

user request.

  According to a particular embodiment of the invention, the equalizer implements a single threshold, and in that said weighted data supplying the recursive filter (FB) is: - said most probable symbol, if said distance is less than said threshold; - said equalized data, if said distance is greater than said threshold. In this case, the optimization of a quality criterion can advantageously consist in minimizing an error worth: - said most probable symbol minus said equalized datum, if said distance is less than said threshold;

  - zero, if said distance is greater than said threshold.

  According to a second advantageous embodiment of the invention, said confidence information is worth: C Min (dx +, dy, dx-, dy *) Cy = at Y - A o: dx +, dy ', dx-, and dy +, are the four distances, along the x and y axes, between said equalized datum and the border defining the zone of association of a datum with said most probable symbol; A is the most likely distance between said border and said symbol. Then, said weighted data supplying said recursive filter (FB) is preferably: Yrécursif = Cy. y '+ (l-Cy) .y o: y' is said most likely symbol;

y is said equalized data.

  Furthermore, advantageously, the optimization of a quality criterion can consist in minimizing: error = Cy. (y'- y) o: y 'is said most likely symbol;

y is said equalized data.

  Of course, the equalizer of the invention can simultaneously implement the two approaches described above, under a

adapted form.

  According to a preferred aspect of the invention, provision is made to process said equalized data differently, when these are found at

  outside the constellation defined by the possible symbols.

  We note that the risk of error remains low, even when

confidence is low.

  In particular, the data outside of said constellation can be projected on the edge closest to said constellation. After projection, a weighted data is advantageously determined in the same way as for the data inside

said constellation.

  The invention also relates to the equalization method implemented by the equalizer described above. Such a decision feedback equalization method comprises the following steps: direct filtering (FF) of input data; - recursive filtering (FB); - delivery of equalized data corresponding to the sum of the outputs of said direct and recursive filtering; - determination of the filter coefficients of said direct and recursive filtering; - determination of information representative of the confidence assigned to said equalized data; - delivery to said recursive filtering of weighted data, taking into account said equalized data and said

trusted information.

  Other characteristics and advantages of the invention will appear more

  clearly on reading the following description of an embodiment

  preferential, given by way of simple illustrative and nonlimiting example, and of the appended drawings, among which: - Figure I is a block diagram, known per se, of a DFE equalizer; - Figure 2 shows an equalizer according to the invention; - Figure 3 illustrates the principle of calculating the confidence in the equalizer of Figure 2, according to a first embodiment; - Figure 4 illustrates the principle of calculating the confidence in the equalizer of Figure 2, according to a second mode of implementation, - Figures 5A to 5E are five examples illustrating the effectiveness of equalization according to invention, compared to a conventional ERD. Figure 1 shows the principle, known per se, and already commented on

  preamble, a feedback equalizer.

  This equalizer receives a data x, corresponding to a transmitted data, having undergone the disturbances of the transmission channel, and delivers a corresponding estimated data y '. It includes a first direct filter (FF) 11, and a recursive filter (FB) 12, whose filter coefficients are calculated by a module 13 implementing an optimization algorithm, depending

of a predetermined criterion.

  The direct filter 1 1 is fed by the symbol (or data) received x, and the recursive filter 12 by the estimated symbol y '. This last symbol is delivered by an estimation module 14, which determines the most probable symbol, by

  an analysis of the sum y 15 of the output of the two filters 11 and 12.

  The calculation 13 of the coefficients implements, for example, an algorithm for minimizing the mean square error (LMS-DD),

  taking into account the error e corresponding to the difference between y and y '.

  Figure 2 illustrates an equalizer according to the invention. The elements already present in Figure 1 keep the same numbering. We see two new functions appear (which can, in certain embodiments, be combined): - calculation of the confidence 21 of the output y of the filter; - use 22 of this trust. The confidence information is notably used to adapt the value (Yrécurif) which is introduced into the recursive filter 12 and / or to adapt the value of the error e taken into account for the calculation 13 of the filtering coefficients. Thus, it is not the value y 'which is taken into account at the input of the recursive filter, but a suitable value Yrcursif, making it possible to obtain a

  more efficient equalization, including limited propagation of errors.

  By way of examples, two techniques are presented below which can be

  implemented to calculate the confidence of the filter output.

  In these two cases, the coefficients of the direct and recursive filters are

driven by an LMS-DD algorithm.

  The first technique, illustrated in Figure 3, is based on a calculation of the distance between the data y 31 actually calculated by the filter, and the nearest decision 32, i.e. the most likely symbol (and therefore , at

  a priori, the one that would be decided in the case of classic ERD).

  At the same time, we implement the following rule for using this confidence (that is to say distance d): - if the confidence is sufficient, that is to say greater than a threshold dmr ,, then the data entered in the recursive filter 12 Ycursf is equal to the decision; otherwise, the data Ycursif is equal to the output of the filter y before decision. The implementation of this rule is excessively simple, as is clearly shown in FIG. 3. Indeed, the user defines the minimum confidence threshold, called din, normalized to A (A being the distance between the

  decision and the border with the area affected by a neighboring decision).

  Around the possible decided symbol, we draw a square 33, or a circle of dmin dimension. If the data y at the output of the equalizer is inside this square, then the data entered in the recursive filter 12 is: Yrcursif = decision otherwise: Yrécursif = Y

  At this level, other rules can, of course, be used.

  With regard to the error e taken into account in the algorithm for calculating the filter coefficients, the rule can be as follows: if d <dm ,,, n then: error = Yestimé - Y otherwise: error = 0 Les performances of this equalizer can be improved by managing the external points of the constellation in a different way (this remark is valid regardless of the embodiment, with the corresponding adaptations). Indeed, when the filter output is outside the constellation, even at a distance greater than dmin, the risk of error is very low, and in this case the output is projected onto the edge of the constellation. More precisely, taking as an example of modulation an MAQ 64, we know that the distance between an axis and the external points is 75 (25 being the distance between two consecutive points). According to the approach proposed above, when the distance of the sample considered is greater than 78 + dmin, the sample is folded back to the distance of 75 along the axis (or axes) considered. In other words: if Dx> 75 + dmn, then Dx = 75 if Dy> 75 + dmin, then Dy = 78 Then, we apply the rules on Yrécursif and on

error e.

  The value of dmjni can be set by the user. We can also foresee that an optimal value of dnin is automatically sought,

  using an iterative algorithm.

  The second proposed embodiment, illustrated in Figure 4, consists in calculating the confidence using the following equation C = Min (dx +, dy ', dx-, dy +) A The rule proposed according to this second method is advantageously the following one: Yrécursif = Cy.yesti ,, m + (1-cy) 'Y error = cy. (yestimé - Y) Again, we can treat differently the points being outside the constellation, by carrying out the following projection (case of a

MAQ 64):

  if Dx> 7 / then Dx = 73 if Dy> 75 then Dy = 73

  Then, we apply the rule proposed above.

  Of course, other embodiments can be envisaged.

  In particular, it is possible to "mix" the approaches defined above, for example by providing, in the embodiment illustrated by FIG. 3, several squares (or circles), corresponding to several thresholds, a particular treatment.

  being provided for each of these thresholds.

  We can also plan to apply the first rule when the

  distance is less than dmin, and the second rule otherwise.

  Figures SA and 5B illustrate simulations showing the effectiveness of

the technique of the invention.

  The simulation context comes from the document "association of equalization, synchronization and channel decoding functions for modulations with high spectral efficiency" by D. Mottier (thesis INSA Rennes; November 1997). It concerns the transmission of a 64 QAM modulation, according to the DVB-cable scheme through two stationary channels

different (Figures 5A and 5B).

  The first (FIG. 5A) corresponds to a slightly disturbed channel, comprising an echo of 10 dB at 30 Ts. The second channel (FIG. 5B) is a very strongly disturbed channel, and therefore very difficult to equalize, defined by the following impulse response:

  H = (0.8264; -0.1653; 0.8512; 0.1636; 0.81)

  For these two channels, a disturbance 51 was voluntarily added during the phase of pursuit of the equalizers, in order to generate an error propagation phenomenon in the case 52 of the conventional ERD. All the illustrated equalizers, optimized by an LMS-DD, converged using

of a learning sequence.

  We have represented the results obtained with an optimal ERD 53 (supervised ERD), which is a "virtual" equalizer for which the data used for the recursive filter and the algorithm are the transmitted data, which is generally considered as the ERD having optimal performance,

  with which to compare a real equalizer.

  We see in these two figures that the classic ERD 52 starts very

  rapidly spreading errors, when non-stationarity 51 appears.

  On the other hand, in both cases, the technique of the invention 54 (ERDP, for weighted ERD), remains stable. The first embodiment is used here

  described, with a value of dmn "= 0.6.

  FIG. 5C illustrates the same situation as FIG. 5B, when the second embodiment described is used. Again, we see that there is

  no divergence of the curve 54, unlike the classic ERD 52.

  FIG. 5D presents an example of processing an MAQ 16 modulation subjected successively: - to a noise 55 of 16 dB (at the instant corresponding to the iteration

000);

  - to an echo 56 (at the instant corresponding to iteration 20,000).

  Again, we see that, despite these two disturbances

  important, there is no discrepancy due to the propagation of errors.

  In the four examples above, we are in the tracking phase, the convergence phase having been previously carried out (for example using a reference sequence). In the example of FIG. 5E, on the other hand, it is shown that a self-taught operation can be obtained according to the invention. This is an MAQ16 modulation subject to a noise of dB. The convergence and tracking phases are ensured by highlighting

  the second embodiment.

Claims (13)

  1 - Decision equalizer, comprising a direct filter (FF) (11) and a recursive filter (FB) (12), the filter coefficients of which are determined (13) so as to optimize a predetermined quality criterion, said equalizer receiving input data (x), supplying said direct filter (FF), and delivering equalized data (y '), corresponding to the sum of the outputs of said direct filter (FF) and said recursive filter (FB), characterized in that said recursive filter (FB) is supplied with weighted data (Yrc ,, ursif), the value of which takes account of said equalized data (y) and   corresponding trusted information.   2- Equalizer with return of decisions according to claim 1, characterized in that the optimization (13) of said quality criterion also takes into account of said trusted information.   3- Equalizer with return of decisions according to any one of the   Claims 1 and 2, characterized in that it comprises means (21) for   calculation of said confidence information in said equalized data, and means (22) of using said confidence information, supplying said recursive filter (FB).   4 - Equalizer with return of decisions according to any one of the   Claims 1 to 3, characterized in that said confidence information   takes into account at least a distance (d) between said equalized datum and the   most likely symbol corresponding to said data.   - Equalizer with return of decisions according to claim 4, characterized in that said distance (d) is compared to at least one threshold (dran,), the   processing of said data being a function of the result of the thresholding.   6- Equalizer with decision feedback according to claim 5, characterized   in that said at least one threshold is iteratively optimized.   7- Equalizer with return of decisions according to any one of the   claims 5 and 6, characterized in that it implements a single threshold, and   in that said weighted data supplying the recursive filter (FB) is: said most probable symbol, if said distance is less than said threshold; - said equalized data, if said distance is greater than said threshold.   8- Equalizer with return of decisions according to claims 2 and 7,   characterized in that the optimization of a quality criterion consists in minimizing an error worth: - said most probable symbol minus said equalized datum, if said distance is less than said threshold;   - zero, if said distance is greater than said threshold.   9- Equalizer with return of decisions according to claim 4, characterized in that said information of confidence is worth: C = Min (dx +, dy ', dx-, dy +) y A o: dx +, dy-, dx-, and dy + , are the four distances, along the x and y axes, between said equalized datum and the border defining the zone of association of a datum with said most probable symbol; A is the distance between said border and said most symbol likely.   - Equalizer with decision feedback according to claim 9, characterized in that said weighted data supplying said recursive filter (FB) is: Yrécursif = Cy. y '+ (l-Cy) .y o: y' is said most likely symbol; y is said equalized data.   11- Equalizer with return of decisions according to any one of the   claims 9 and 10 characterized in that the optimization of a criterion of   quality consists in minimizing: error = Cy. (y '- y) o: y' is said most likely symbol; y is said equalized data.   12- Equalizer with return of decisions characterized in that it implements   simultaneously implementing a calculation according to any one of the claims   4 to 8 and a calculation according to any one of claims 9 to 11.   13- Equalizer with return of decisions according to any one of the   Claims 4 to 12, characterized in that said treatments are treated differently   equalized data, when these are outside the constellation   defined by the possible symbols.   14- Decision equalizer according to claim 13, characterized in that the data outside of said constellation are   projected on the edge closest to said constellation.   - Equalizer with return of decisions according to claim 14, characterized in that, after projection, a weighted data item of the same is determined   so that for the data located inside said constellation.   16- Process for equalization with return of decisions, characterized in that it comprises the following stages: - direct filtering (FF) (it 1) of input data; - recursive filtering (FB) (12); - delivery of equalized data (y ′) corresponding to the sum (15) of the outputs of said direct (11) and recursive (12) filtering; - Determination (13) of the filter coefficients of said direct and recursive filtering; characterized in that it comprises the steps of: - determination (21) of information representative of the confidence assigned to each equalized datum; - delivery to said recursive filtering of weighted data (Yrécursif), taking into account said equalized data (y ') and said trusted information.