WO1981001089A1

WO1981001089A1 - Control of coefficient drift for fractionally spaced equalizers

Info

Publication number: WO1981001089A1
Application number: PCT/US1980/001245
Authority: WO
Inventors: J Werner
Original assignee: Western Electric Co
Priority date: 1979-10-15
Filing date: 1980-09-25
Publication date: 1981-04-16
Also published as: EP0040217A1; DE3070276D1; EP0040217A4; JPS56501349A; JPH0365058B2; EP0040217B1

Abstract

Tap coefficient drift in a fractionally spaced equalizer (10) is minimized by adding signal energy (n(t)) to the received data signal (sr(t)) at frequencies where the sampled channel transfer function has substantially zero gain.

Description

CONTROL OF COEFFICIENT DRIFT FOR FRACTIONALLY SPACED EQUALIZERS

Background of the Invention The present invention relates to automatic equalizers which compensate for the distorting effects of band-limited channels on transmitted data signals.

Automatic equalizers are necessary for accurate reception of high-speed data signals transmitted over band-limited channels with unknown transmission characteristics. The equalizer is generally in the form of a transversal filter in which a sampled signal comprised of samples of the incoming data signal are multiplied by respective tap coefficients. The resulting products are added together to generate an equalizer output which is then demodulated and/or quantized to recover the transmitted data. In addition, an error signal is formed equal to the difference between the equalizer output and a reference signal which represents the transmitted data symbol. The value of the symbol that was transmitted may be known at the receiver a priori, as is the case in many equalizer start-up arrangements. Alternatively, as in the so-called adaptive type of automatic equalizer, the reference signal is derived from the decision made in the receiver (on the basis of the equalized signal value) as to what data symbol was transmitted. In either case, the error signal is used to update the tap coefficient values in such a way as to minimize a measure of the distortionprimarily intersymbol interference-introduced by the channel. The most commonly used error-directed coefficient updating algorithm is the so-called mean-squared error algorithm, which adjusts the tap coefficients so as to minimize the average of the value of the square of the error signal.

Most commercial data receivers, e.g., data modems, incorporate a synchronous, or baud, equalizer in which the received data signal is sampled at a rate equal to the symbol rate. It is, however, possible to use a so-called fractionally spaced equalizer in which the received signal is sampled at a higher rate. Data decisions, i.e., quantizations of the equalizer outputs, are still made at the symbol rate. However, the fact that equalization is carried out using a finer sampling interval provides the fractionally spaced equalizer with significant advantages over the more conventional type of equalizer. Most notable among these is insensitivity to channel delay distortion, including sampling phase errors.

There is, however, at least one significant problem unique to the fractionally spaced equalizer. In a synchronous equalizer, one set of tap coefficients is clearly optimum, i.e., provides the smallest mean-squared error. By contrast, many sets of coefficient values provide approximately the same mean-squared error in the fractionally spaced equalizer. As a consequence of this property, the presence of small biases in the coefficient updating processing hardware--such as biases associated with signal value roundoff--can cause at least some of the coefficient values to drift to very large levels, or "blow up," even though the mean-squared error remains at, or close to, its minimum value. The registers used to store the coefficients or other signals generated during normal equalizer operation can then overflow, causing severe degradation, or total collapse, of the system response. The prior art--exemplified by G.

Ungerboeck, "Fractional Tap-Spacing Equalizers and Consequences for Clock Recovery for Data Modems," IEEE Trans, on Communications, Vol. COM-24, No. 8, August 1976, pp. 856-864-- suggests that the coefficient drift can be controlled by introducing one of two alternative auxiliary terms into the conventional updating algorithm. The auxiliary term may be, for example, a predetermined small fraction of the current value of the coefficient being updated. This implements a so-called tap leakage approach. Alternatively, a spectral zero-forcing approach is suggested. Here, the auxiliary term is a predetermined small fraction of an alternating-sign sum of the current values of all the coefficients.

These approaches, while providing an upper limit for the coefficient values, are not wholly satisfactory from other standpoints. For example, it is desirable in any transversal filter type of automatic equalizer to have as many of the coefficient values at or as close to zero as possible. This, means that the numerical computations associated with coefficient updating will involve the manipulation and storage of smaller numbers than would otherwise be the case. This, in turn, minimizes the complexity and expense of the computational hardware. In addition, keeping as many of the coefficient values at or as close to zero as possible best conditions the system to withstand the effects of, and to recover from, phase hits and other transmission disturbances. The above-described approaches for dealing with coefficient drift, while providing an upper limit for the coefficient values, allow a large number of the coefficients to assume values which are not at or close to zero. Thus, system performance suffers.

A more efficacious technique for controlling coefficient drift is taught in U. S. patent application Serial No. 16,495. As in Ungerboeck, a tap leakage term is introduced into the coefficient updating algorithm. Here, however, the magnitude of the tap leakage term is independent of any coefficient value; it is illustratively a constant. This approach has been found to substantially avoid the above-outlined drawbacks of the Ungerboeck approach.

On the other hand, the tap leakage of U.S. patent application Serial No. 16,495 (like Ungerboeck) necessarily introduces a certain amount of noise into the equalization process inasmuch as it changes the coefficients from the values which the error-directed algorithm specifies. This has not been found to be a significant effect in, for example, the so-called T/2 equalizer which receives two line samples per symbol interval. However, T/p equalizers, p > 2, tend to exhibit greater tendency toward tap coefficient drifting. This necessitates increasing the magnitude of the tap leakage term, introducing further noise in the equalization process and thereby increasing the likelihood of an incorrect data decision. Summary of the Invention

This invention is directed toward controlling coefficient drift in fractionally spaced equalizers while minimally affecting the accuracy of the equalization process itself.

In accordance with the invention, energy is added to the sampled signal at frequencies at which the sampled channel transfer function has substantially zero gain. This forces the equalizer to have a unique transfer function and, thus, a unique set of coefficient values for the channel being equalized. The coefficient drift problem is thus substantially eliminated.

The added energy can be in any of numerous forms. However, excellent results are obtained when it is in the form of one or more sweeping sinusoids. Brief Description of the Drawing

The invention will be clearly understood from a consideration of the following detailed description and accompanying drawing in which:

FIG. 1 is a block diagram of a data receiver including circuitry embodying the principles of the invention;

FIGS. 2 - 5 are signal amplitude vs. frequency plots helpful in explaining the operation of baud-sampled equalizer; FIGS. 6 - 10 are signal amplitude vs. frequency plots helpful in explaining the operation of fractionally spaced equalizers and, in particular, the coefficient drift phenomenon; and FIGS. 11 - 13 are signal amplitude vs. frequency plots depicting several forms for the signal energy added to the received data signal in accordance with the invention. Detailed Description FIG. 1 depicts a receiver 10 for data signals transmitted from a transmitter (not shown) over a bandlimited channel, e.g., voiceband telephone circuit. The data signals are illustratively quadrature amplitude moduulated (QAM) data signals wherein four paralleled information bits are transmitted during each of a succession of symbol intervals of duration T. The symbol rate is thus 1/T, yielding a binary transmission rate of 4/T bits per second. During each symbol interval, the four bits to be transmitted are encoded into two signal levels each of which can take on one of the four values [+1, -1, +3, -3]. The two signal levels transmitted during the m^th symbol interval comprise data symbol A_m--a complex quantity having real and imaginary components a_m and â_m, respectively. Components a_m and â_m, in turn, amplitude modulate respective 1800 Hz in-phase and quadrature-phase carrier waves. The modulated signals, when added together, form a QAM signal s (t) of the form

s (t) = Re [ ∑A_mg (t-mT) e^{j ω} c^t]

where g(t) is a real function and ω_c is the radian carrier frequency. Signal s(t) is then transmitted to receiver 10.

In receiver 10, the received QAM signal s_r (t) passes through automatic gain control circuit 8 where it emerges as signal s_r'(t). The latter is applied to an input circuit 11 and, more particularly, to analog bandpass filter 12 thereof. The function of filter 12 is to filter out any energy in signal s_r'(t) outside of the transmission band of interest--in this example the band 300-3000HZ. In accordance with the present invention, the output signal q(t) of filter 12 is added in an adder 13 to a signal n(t) generated by out-of-band signal generator 18. Signal n(t) is discussed in more detail hereinbelow. Input circuit 11 further includes a phase splitter 14, a sampler in the form of an analog-to-digital (a/d) converter 17 and sample clock 19. Phase splitter 14 responds to the output signal q'(t) of adder 13 to generate two versions of signal q'(t). One of these is q"(t), which may be identical to q'(t) or may be a phase-shifted version of it. The other, represented as

"(t), is the Hilbert transform of q" (t). Signals q" (t) and

" (t) may be regarded as the real and imaginary components of a complex signal Q" (t). Signals q" (t) and

" (t) are passed to a/d converter 17. The latter is operated by clock 19 p times per symbol interval to generate a sampl d signal in the form of equalizer input samples q_k and

_k, k = 1,2 .... of signals q" (t) and

" (t). Equalizer input samples q_k and may be thought of as components of a complex

equalizer input sample Q_k.

Equalizer input sample components q_k and _k

pass on to transversal filter equalizer 22. The latter generates an output once every T seconds. In particular, the output of equalizer 22 during the m^th receiver symbol interval of duration T is complex passband equalizer output U_m having components u_m and û_m. Equalizer 22 generates its outputs by forming linear combinations of the equalizer input sample components in accordance with the relations

In these expressions r_m and r^{^} _mare (Nx1) matrices, or vectors, respectively comprised of the N most recent real and imaginary equalizer input sample components, N being a selected interger. That is

In addition, c_m and d_m are (N x 1) vectors each comprised of an ensemb l e of N tap co effi ci ent s having values associated wi th the m^th receivrr interval (The superscript "T" used in the above expressions indicates the matrix transpose operation wherein the (N x 1) vectors c_m and d_m are transposed into (1 x N) vectors for purposes of matrix multiplication.) The values of the coefficients in these cectors are determined in the manner described below. Vectors c_m and d_m may be thought of as the real and imaginary components of a complex coefficient C_m

Passband equalizer output U_m is demodulated by demodulator 25 to yield baseband equalizer output Y_m .

The latter and passband equalizer output U_m respectively represent baseband and passband versions of transmitted symbol A_m. Baseband equalizer output Y_m has real and imaginary components y_m and y^{^} _m , the demodulation process b eing expres s ed as

θ^* _m being an estimate of the current carrier phase. For purposes of generating y_m and y^{^} _m in accordance with the above expressions, demodulator 25 receives representations of cos θ^* _m and sin θ^* _m from a carrier source 27.

Baseband equalizer output Y_m is quantized in decision circuit 31. Tire resulting output A^* _m is a decision as to the value of the transmitted symbol A_m .

In particular, the real and imaginary parts of A^* _m, a^* _m and â^* _m are decisions as to the data signal values represented by the real and imaginary components am and âm of transmitted symbol A_m. Decision circuit 31, more particularly, forms decision a^* _m(â^* _m) by identifying the one of the four possible data signal values

[+1, -1, +3, -3] which is closest to the value of equalizer output component y_m(y^{^} _m).

Decision A^* _m is also used to generate an error signal for use in updating coefficient vectors c_m and d_m. In particular, decision components a^* _m and â^* _m are combined in decision remodulator 35 with sin θ^* _m and cos θ^* _m from carrier source 27 to form remodulated, or passband, decision A_p*_m. The real and imaginary components of A_p*_m, a_p*_m and â_p*_m are formed in accordance with

Passband decision A is subtracted from passband equalizer output U_m in subtractor 38 to yield passband error E_pm having components e_pm and ê_pm g i ven by

Error signal components e_pm and ê_pm are extended to coefficient store and update unit 23 within equalizer 22 for purposes of updating the values of the coefficients in coefficient vectors c_m and d_m in preparation for the next, (m+1)^st, symbol interval. The so-called mean-squared error stochastic updating algorithm is illustratively used, yielding the updating rules

α being a predetermined constant. These rules can be written in complex notation as

C_m+1 = C_m - αR_kE_pm .

The problem to which the present invention is directed is illustrated in FIGS. 2 - 9. FIG. 2, in particular, shows the positive frequency portion of the magnitude |G(ω) I of the transfer function of a typical voiceband telephone transmission channel. In this discussion, the "transmission channel" is assumed to include the transmitter, the transmission medium and all receiver circuitry through and including the receiver phase splitter. The negative frequency portion of |G(ω)| need not be considered since the complex signal Q" (t) at the output of phase splitter 14 is "analytic," i.e., has no negative frequencies. The radian carrier frequency of the QAM signal to be transmitted is ω_c. The transmission channel extends from (ω_c-(1+β)π/T) to (ω_c+ (1+β) π/ T ) rad/sec, meaning that within that band the gain provided by the channel is at least some predetermined minimum, e.g., -40 db. Outside of that band, the gain is less than that minimum and is presumed to be zero. The parameter 3 is the so-called percent roll-off and has a value between zero and unity given by β= (ω_coT/π-1) where the frequency (ω_c+ω_co) ^is the upper frequency limit of the transmission channel.

The transfer function G_s(ω) of the "sampled channel," defined as the combination of the transmission channel with all receiver circuitry up through and including the sampler, i.e., a/d converter 17, is arrived at by a "folding" operation to yield

It is thus seen that the sampled channel transfer function is a superposition in the frequency domain of repetitions of the unsampled transfer function translated in frequency by multiples of 2πp/T.

For baud equalization, of course, p = 1. As shown in FIG. 3, |G_s(ω)| for this case is non-zero for all ω because the tails (shown in dotted line) of each translated G (ω ) overlap and add with those of adjacent translated G(ω)'s. A baud equalizer is properly equalized when the equalized channel transfer function, i.e., the overall transfer function of the sampled channel-plus-equalizer combination, has constant gain (and linear phase characteristic) at all frequencies. The magnitude IH_s(ω)| of an ideal such transfer function for a baud-sampled receiver is shown in FIG. 5. Thus, for any given channel, there is a unique, optimum equalizer transfer function--corresponding to a unique ensemble of tap coefficient values--which provides the best equalization, i.e., the smallest mean-squared error for baud sampling. The magnitude |F_s(ω)| of that optimum equalizer transfer function for the channel under consideration is shown in FIG. 4.

For fractionally spaced equalization, by contrast, p > (1+β) so that the tails of adjacent repetitions of |G (ω)I do not overlap. They are, rather, separated by what may be referred to as "no-energy bands." This is illustrated in FIG. 6 which shows the transfer function magnitude |G_s'(ω)| of G(ω) sampled at ρ/T samples per second, p > (1+β).

A consequence of the nonoverlapping of the repetitions of G (ω) is that a number of fractionally space equalizer transfer functions--each corresponding to a different coefficient ensemble--provide substantially the same, minimum mean-squared error.

The magnitudes of two such equalizer transfer functions |F_s'₁(ω) | are shown |F_s'₂(ω)| are shown in FIGS. 7 and 8, respectively. The magnitudes |H_s'₁(ω) | |H_s'₂(ω) | of the resulting equalized channel transfer functions are shown in FIGS. 9 and 10. From the standpoint of minimizing the mean-squared error, the equalizer transfer functions of FIGS. 7 and 8 are equivalent; since there is no signal energy in the no-energy bands, the fact that |H_s'₂(ω) | is non-zero in these bands is irrelevant--at least in theory.

There is a problem, however, as previously noted there is no unique optimum equalizer transfer function. Thus, as the tap coefficient values are updated over time, small biases in the coefficient updating processing hardware--such as biases associated with signal value round-off--can cause at least some of the coefficient values to drift. This corresponds to a drift in the equalizer transfer function. For example, the transfer function of FIG. 7 may be the one which exists right after equalizer start-up, but may drift to that of FIG. 8 after, say, five minutes. Ultimately, some of the coefficient values may drift to such large levels that the registers used to store the coefficients or other signals generated during normal equalizer operation can then overflow, causing severe degradation, or total collapse, of the system response.

Unlike the tap leakage arrangements, which represent time domain approaches to the coefficient drift problem, the present invention fights the coefficient drift in the frequency domain. In accordance with the present invention, energy is added to the sampled signal in the no-energy bands. That is, the energy is added at frequencies at which the sampled channel transfer function has substantially zero gain. The effect of adding energy in the no-energy bands is that if the coefficient values now start to drift, causing the gain of the equalizer transfer function to increase in the no-energy bands, the contribution of the added energy to the mean-squared error will be nonnegligible. The coefficient updating algorithm responds by adjusting the coefficients so as to minimize the error. In so doing, it forces the equalizer to have a unique transfer function, one which, like the transfer function of FIG. 7, has substantially zero gain in the no-energy bands. Since a unique transfer function implies a unique set of coefficient values, the coefficient drift problem is substantially eliminated.

Adding energy to the sampled signal may be accomplished by introducing the signal in digital form at the output of the a/d converter. Alternatively, as in the present illustrative embodiment, it may be introduced in analog form ahead of the a/d converter. In the latter case, the signal need have energy only in the no-energy bands within the range (0 - 2πp/τ) rad/sec, i.e., at frequencies at which the transfer function of the transmission channel has substantially zero gain. This is because the folding which occurs upon sampling results in energy being added to the sampled signal in each no-energy band of the sampled channel transfer function, as desired, not just within the range (0 - 2τrp/T). In a system in which the sampler input is an analytic signal, as in the present illustrative embodiment, energy is added in at least one of the frequency bands (0 - (ω_c(1+β) π/T)) and ((ω_c+(1+β)π/T) - 2πp/T) rad/sec. Where the sampler input is not analytic, but is real, energy is added in at least one of the frequency bands (0 - (ω_c-(1+β) π/T)) and ((ω_c+(1+β) π / T) - πp/T) rad/sec. In receiver 10, in particular, the abovediscussed added signal energy, denominated n(t), is illustratively provided in analog form from out-of-band signal generator 18. Signal n(t) is added to signal q(t) via adder 13 to generate signal q'(t).

The amplitude of signal n(t) should be sufficiently large to provide the necessary amount of coefficient drift control. An amplitude of 10-15 db below the amplitude of the AGC output signal has been found to be appropriate.

The spectrum of signal n(t) may take any of several forms. For exmaple, as shown in FIG. 11, signal n (t) may be comprised of a sum of sinusoids, each having a selected frequency. The frequency gap between adjacent sinusoids should be fairly small. Otherwise, the equalizer will generate zeros on the frequency axis where the sinuosids are located, but it will be unable to compensate for the build-up of gain between the zeros. In an actual embodiment of the invention having ω_c = 2π.1800, p = 6, and T = 1/1600, a spacing of 150 Hz between adjacent sinusoids was found to be adequate. Alternatively, as shown in FIG. 12, signal n(t) may be comprised of one or more sweeping sinusoids, i.e., sinusoids whose frequencies are varied in a predetermined manner. In many applications this will be the simplest to implement. With ω_c, p and T having the values indicated above, a sweeping frequency of 10 Hz was found to be effective.

Alternatively, signal n(t) may be comprised of random noise as shown in FIG. 13. In this case, however, special care must be taken to ensure that the sum of the AGC output and n(t) does not exceed the dynamic range of the subsequent circuitry.

As mentioned above, the additive signal of the present invention may be introduced in digital form at the a/d converter output. In this regard, it should be noted that if the radian frequency 2πp/T happens to fall within a no-energy band, a sinusoid of amplitude A can be introduced at that frequency by adding the quantity A to the real component q_k of each equalizer input sample.

The foregoing merely illustrates the principles of the present invention. For example, in some applications coefficient drift may be satisfactorily eliminated without adding energy in narrow no-energy bands, e.g., the band (0 - (ω_c-(1+β) π/T)).

It will thus be appreciated that numerous arrangements embodying the principles of the invention may be devised by those skilled in the art without departing from their spirit and scope.

Claims

1. Method comprising the steps of forming a sampled signal by forming samples of a received data signal which was transmitted over a transmission channel at 1/T symbols per second, said samples being formed at a rate of p/T samples per second, p > (1 + β), where 3 is the percent rolloff of said data signal, multiplying ones of said samples by predetermined coefficients at said symbol rate, forming a decision as to the value of each transmitted data symbol in response to the sum of the resulting products and forming a corresponding error signal, and updating each of said coefficients in response to said error signal, characterized by the step of adding to said sampled signal a second signal having energy at frequencies at which the transfer function for said sampled signal has substantially zero gain to minimize the tendency for the values of said coefficients to drift over time.

2. The method of claim 1 wherein said adding step is characterized by the step of adding to said received signal a signal having energy within the range (0 - 2πp/T) rad/sec at frequencies at which the transfer function of said transmission channel has substantially zero gain.

3. The method of claims 1 or 2 wherein said second signal is characterized by a plurality of sinusoids.

4. The method of claims 1 or 2 wherein said second signal is characterized by at least one sinusoid, the frequency of which is varied in a predetermined manner.

5. The method of claims 1 or 2 characterized by said second signal being a noise signal.

6. The method of claims 1 or 2 wherein said data signal is characterized by a modulated signal of radian frequency ω_c and wherein the highest frequency at which the transfer function of said transmission channel has at least a predetermined minimum gain is characterized by (ω_c+ (1+β) π/T).

7. The invention of claims 1 or 2 wherein in said updating step are characterized by said coefficients being updated so as to minimize the average of the value of the square of said error signal.

8. Apparatus for performing any of the methods of claims 1 through 7 including means for forming a sampled signal comprised of sampling means (17, 19) for forming samples of a received data signal which was transmitted over a transmission channel at 1/T symbols per second, said transmission channel and said sampling means comprising a sampled channel and said sample forming means operating at a rate of p/T samples per second, p > (1 + β), where β is the present rolloff of said data signal, equalizer means (22) operative at said symbol rate for multiplying ones of said samples by predetermined coefficients, and decision circuit means (25, 27, 31, 35, 38) responsive to the sum of the resulting products for forming a decision as to the value of each transmitted data symbol and for forming a corresponding error signal, said equalizer means including means ( 23 ) for updating each of said coefficients in response to said error signal, characterized by means (13, 18) for adding to said sampled signal a second signal having energy at frequencies at which the transfer function of said sampled channel has substantially zero gain to minimize the tendency for the values of said coefficients to drift over time.

9. Apparatus of claim 8 wherein said adding means is characterized by means for adding to said received signal a signal having energy within the range (0 - 2πp/T) rad/sec at frequencies at which the transfer function of said transmission channel has substantially zero gain.

10. Apparatus of claims 8 or 9 wherein said second signal is characterized by a plurality of sinusoids, each having a selected frequency.

11. Apparatus of claims 8 or 9 wherein said second signal is characterized by at least one sinusoid, the frequency of which is varied in a predetermined manner.

12. Apparatus of claims 8 or 9 wherein said second signal is characterized by a noise signal.

13. Apparatus of claims 8 or 9 wherein said data signal is characterized by a modulated signal of radian frequency ω_c and wherein the highest frequency at which the transfer function of said transmission channel is characterized by at least a predetermined minimum gain is (ω_c + (1+β)π/T).

14. Apparatus of claim 13 wherein said updating means is characterized by means for updating each of said coefficients so as to minimize the average of the value of the square of said error signal.