Вы находитесь на странице: 1из 4

Adaptive Blind Compensation of Gain and Timing

Mismatches in M -Channel Time-Interleaved ADCs


Christian Vogel Shahzad Saleem Stefan Mendel
Signal and Signal Processing and Signal Processing and
Information Processing Laboratory Speech Communication Laboratory Speech Communication Laboratory
ETH Zurich, Switzerland Graz University of Technology, Austria Graz University of Technology, Austria
email: c.vogel@ieee.org email: ssaleem@tugraz.at email: stefan.mendel@tugraz.at

Abstract—Gain and timing mismatches among sub-converters limit the (nM + 0)Ts + r0 Ts

performance of time-interleaved analog-to-digital converters (TIADCs). g0


In this paper we present a blind adaptive method, based on the ADC0
least-mean-square (LMS) algorithm, to compensate gain and timing
mismatches in TIADCs. Similar to other methods in the literature,
we assume a slightly oversampled input signal, but, contrary to them, (nM + m)Ts + rm Ts
analog digital compensated
we can apply our method to an arbitrary number of channels in a input output Adpative output
gm yc [n] ≈ G0 x[n]
x(t) y[n] Gain & Timing
straightforward way. We give a detailed description of the compensation ADCm MUX
and the identification part of the method and demonstrate its effectiveness Mismatch
Compensation
through numerical simulations.
(nM + (M − 1))Ts + rM −1 Ts

I. I NTRODUCTION gM −1
ADCM −1
A time-interleaved ADC (TIADC) provides an effective way to
achieve high sampling rates [1]. Each channel ADC in an M - TIADC with fs = 1/Ts
fs
channel TIADC operates with a sampling rate of M , while the
overall sampling rate is fs , i.e., M -times higher. Unfortunately, due Fig. 1. Identification and compensation of gain and timing mismatches with
to different channel characteristics, the interleaving of multiple ADCs a bandlimited oversampled input signal for an M -channel TIADC.
leads to mismatches, where in particular gain and timing mismatches
affect the performance [2].
Several mismatch identification and compensation methods have gain gm for m = 0, 1, . . . , M − 1. By assuming a bandlimited input
been introduced, e.g., [3]–[7]. However, due to their computational signal X (jΩ), i.e., X(jΩ) = 0 for |Ω| ≥ Ωb and Ωb Ts ≤ π, the
complexity their practical applicability is limited. In [8], [9] the output of the M -channel TIADC with gain and timing mismatches
authors have investigated adaptive blind calibration of gain and timing can be represented in discrete time as [12]
mismatches that have maintainable complexity but are limited to two
channels. In [10] the blind calibration technique from [9] has been “ ” M −1
X “ 2π
” “ 2π

extended to four-channel TIADCs but it only considers the calibration Y ejω = αk ej (ω−k M ) X ej (ω−k M ) (1)
of timing mismatches. k=0

In this paper we extend our previous work [5], [11] and present
with
a structure that allows to adaptively compensate gain and timing
mismatches in an M -channel TIADC using similar assumptions as “ ” M −1
1 X jω 2π
αk ejω = gm erm Hd (e ) e−jk M m (2)
in [5], [10]. Hence, we assume a slightly oversampled input signal M m=0
that exhibits sufficient energy during the adaptation process. Under
these assumptions we can use the LMS algorithm to estimate the gain and
and timing mismatches of the TIADC and compensate the error signal
due to the mismatches. Because of the relation to [5] the achievable Hd (ejω ) = jω for − π < ω ≤ π (3)
performance of this method and the method in [5] is comparable, but
the implementation complexity in this work has been considerably is the frequency response of an ideal discrete-time differentiator [13].
reduced. For an ideal TIADC without gain and timing mismatches gm = 1
For the sake of clarity, we refer to an even number of channels M and rm = 0. Since in a typical TIADC the relative time offsets rm
for all derivations in this paper. The derivations for an odd number are small compared to the sampling period Ts we can apply Taylor’s

of channels can be done accordingly. series approximation to the term erm Hd (e ) and can neglect the
higher order terms, which results in [12]
II. S YSTEM M ODEL

“ ”
Figure 1 shows a model of a TIADC with gain and timing erm Hd (e ) ≈ 1 + rm Hd ejω . (4)
mismatches. Each channel is characterized by the relative time
offset rm Ts , where Ts is the overall sampling period, and the After substituting (4) in (2) we obtain
C. Vogel was supported by the Austrian Science Fund FWF’s Erwin “ ” “ ”
Schroedinger Fellowship J2709-N20. αk ejω = Gk + Rk Hd ejω (5)

978-1-4244-2182-4/08/$25.00 ©2008 IEEE. 49


with Compensation yc [n] Identification
≈ G0 x[n]
M −1
1 X 2π ê[n]
Gk = gm e−jk M m (6) y[n] = G0 x[n] + e[n]
f [n]
d[n] ε[n]
M m=0 LMS
X
M −1 e[n]
1 −jk 2π m ĉg [n]
Rk = gm rm e M . (7)
M m=0
m[n] f [n]
Without loss of generality, we can assume that the average value xg [n] xg [n]
of all timing mismatches rm is zero, since an overall delay does ĉg [n]
hd [n]
not introduce any mismatch effects. Accordingly, R0 is close to ĉr [n]
zero as well and can be neglected. The inverse discrete-time Fourier
m[n] f [n]
transform of (1) can therefore be written as xr [n] xr [n]
ĉr [n]
y[n] = G0 x[n] + e[n] (8)
where G0 is the average gain among all channels and e[n] the error Fig. 2. Adaptive compensation structure.
introduced by gain and timing mismatches. Since αM −k [n] = αk∗ [n],
where ∗ denotes conjugate complex, we can further express the error Finally, we can write the error (9) in vector notation as
e[n] for even M as (see [14] for a more details)
M
e[n] = cTg xg [n] + cTr xr [n]. (16)
−1
X
2

e[n] = ek [n] + e M [n] (9) III. C OMPENSATION S TRUCTURE


2
k=1 With (8) and (16) we have mathematically decomposed the output
with signal y[n] into the input signal x[n] and the error signal e[n]
„ „ « „ «« that arises due to the mismatches. In order to remove the effect of
2π 2π
ek [n]= {Gk }2 cos k n − {Gk }2 sin k n x[n] mismatches, we have to cancel the error signal e[n]. Therefore, we
M M
„ „ « „ «« use the output signal y[n] to produce a reconstructed error signal
2π 2π
+ {Rk }2 cos k n − {Rk }2 sin k n ê[n] and then subtract ê[n] from the output signal y[n] to obtain the
M M compensated output signal yc [n] as illustrated in Fig.2. This approach
×hd [n] ∗ x[n] (10) has been discussed in detail for timing mismatches in [11], [12]. The
and output signal of the compensation structure can be written with (8)
as
e M [n] = G M (−1)n x[n] + R M (−1)n hd [n] ∗ x[n]. (11)
2 2 2 yc [n]=y[n] − ê[n]
The real part of x is denoted by {x} and the imaginary part of x =G0 x[n] + e[n] − ê[n]
is denoted by {x}. For M = 2, the summation over ek [n] in (9)
≈G0 x[n] for e[n] ≈ ê[n] (17)
vanishes leaving only the real valued term e M [n].
2
By introducing the coefficient vectors To produce the reconstructed error signal ê[n] we apply (16) to the
„ output signal y[n] instead of using the signal x[n], which results
cg = {G1 }, {G1 }, . . . , {Gk }, {Gk }, . . . , with (14) and (15) in
«T ê[n] = ĉTg m[n]y[n] + ĉTr m[n]hd [n] ∗ y[n]. (18)
{G M −1 }, {G M −1 }, G M

2 2 2
As the particular values of cg and cr are generally unknown, they
have been replaced by their estimates ĉTg and ĉTr , respectively.
cr = {R1 }, {R1 }, . . . , {Rk }, {Rk }, . . . ,
Since in a typical TIADC the signal energy of the signal G0 x[n]
«T is significantly larger than the energy of the error signal e[n], we
{R M −1 }, {R M −1 }, R M (12) can neglect e[n] for a first order analysis [2]. Therefore, we can
2 2 2
simplify (18) by substituting (8) and neglecting e[n], which gives
and the modulation vector
„ „ « „ « ê[n] = ĉTg m[n]G0 x[n] + ĉTr m[n]hd [n] ∗ (G0 x[n]). (19)
2π 2π
m[n]= 2 cos 1 n , −2 sin 1 n , . . . ,
M M Substituting the error e[n] given in (16) and its reconstruction ê[n]
„ « „ «
2π 2π given in (19) into (17) results in
2 cos k n , −2 sin k n , . . . , „ «T
M M 1
„„ « « „„ « « yc [n]=G0 x[n] + G0 cg − ĉg xg [n]
M 2π M 2π G0
2 cos −1 n , −2 sin −1 n , „ «T
2 M 2 M
«T 1
+G0 cr − ĉr xr [n]. (20)
(−1) n
(13) G0
From the first order analysis in (20) we can conclude that the
we can define the signal vectors compensated output is G0 x[n] if the estimated coefficients are
xg [n] = m[n]x[n] (14) 1
ĉg = cg (21)
G0
and 1
xr [n] = m[n]hd [n] ∗ x[n]. (15) ĉr = cr . (22)
G0

50
xg [n] TABLE I
S IMULATED GAIN AND TIMING MISMATCH VALUES
1
G0 cg
xr [n] d[n] ε[n] ADC gk rk
1
G0 cr
LMS ADC0 1.01 −0.007Ts
ADC1 0.98 0.002Ts
ADC2 0.99 −0.003Ts
ADC3 1.02 0.008Ts
ĉr [n] e[n]

ĉg [n] coefficients ĉg and ĉr that are


Fig. 3. Model of the identification structure illustrating the system identifi- ĉg [n]=ĉg [n − 1] + μg xg [n]ε[n]
cation of gain ĉg and timing ĉr coefficients.
ĉr [n]=ĉr [n − 1] + μr xr [n]ε[n] (29)
where μg and μr are the step size parameters [15]. As we see from
IV. I DENTIFICATION S TRUCTURE the derivations, the principle of the method is independent from
The principle architecture is shown in Fig.2, where the left side the number of channels. For each additional channel we need two
shows the compensation structure as derived in the last section and high-pass filters, two modulators, four multipliers, and corresponding
the right side shows the identification structure. To compensate the modifications of the LMS implementation to identify and compensate
mismatches we have to identify the coefficients cg and cr . The gain and timing mismatches. Hence, the complexity linearly scales
presented identification is based on the observation that oversampling with the number of channels.
a wideband input signal x(t) with a TIADC with mismatches should
V. S IMULATIONS
result in a region, denoted as mismatch band, where no input
signal energy is present [5], [9], [10]. However, due to gain and To illustrate the performance of the proposed adaptive compensa-
timing mismatches among the channels a certain amount of signal tion structure we have simulated a 4-channel TIADC with gain and
energy is in the mismatch band. By minimizing the signal energy timing mismatches as shown in Tab. I. The input signal has been
in the mismatch band we can adaptively identify timing and gain bandlimited, i.e., X(jΩ) = 0 for |Ω| ≥ Ωb and Ωb Ts = 0.7π,
mismatches, for which the least-mean-square (LMS) algorithm is white Gaussian noise (WGN) with variance σ = 1. From the input
an appropriate candidate. The high-pass filters f [n] attenuate the signal 131072 samples have been taken. The step size parameters μr
input signal energy and only pass the energy in the mismatch band. and μg of the LMS were chosen to be 0.01. The differentiator hd [n]
The variable coefficients ĉg [n] and ĉr [n] are adapted by the LMS and the high pass filters f [n] have been designed using the Matlab
algorithm to minimize the signal ε[n] in the mean-square sense. function ’firpm’, where the number of taps have been 33 and the cut-
To explicitly see the system identification problem, we will derive off frequency has been 0.7π for both filters. The overall performance
the system identification model shown in Fig. 3. The desired signal of the compensation structure has been evaluated by the signal-to-
d[n] is given by high-pass filtering the TIADC output signal y[n], noise ratio (SNR) for the uncompensated output
PN −1 !
n=0 |G0 x[n]|
2
d[n]=y[n] ∗ f [n] SNR = 10log10 PN −1 (30)
n=0 |G0 x[n] − y[n]|
2
=G0 x[n] ∗ f [n] + e[n] ∗ f [n]
=e[n] ∗ f [n] (23) and the compensated output
PN −1 !
where we have substituted (8) for y[n]. Substituting (16) in (23) gives |G0 x[n]|2
SNR = 10log10 PN −1n=0 . (31)
n=0 |G0 x[n] − yc [n]|
2
d[n] = cTg xg [n] + cTr xr [n] (24)
Figure 4 shows the energy density spectrum of the uncompensated
with
output signal y[n] (last 2048 samples of the 131072 samples), where
xg [n]=m[n]x[n] ∗ f [n] the SNR is about 35.1 dB. In Fig. 5 the energy density spectrum
xr [n]=m[n]x[n] ∗ hd [n] ∗ f [n]. (25) of the compensated output yc [n] is shown (last 2048 samples).
The computed value of the SNR is about 66 dB; hence, leading
Applying the high-pass filter to the reconstructed error signal ê[n] to an improvement of almost 30.9 dB. It should be noted that the
in (19) results in the signal two spectra show the outputs after the LMS algorithm has already
converged. The convergence behavior of the timing coefficients ĉr [n]
e[n] = ê[n] ∗ f [n] = G0 ĉTg xg [n] + G0 ĉTr xr [n]. (26) and the gain mismatch coefficients ĉg [n] are shown in Fig. 7 and
The error signal of the identification structure is given by Fig. 6, respectively. The convergence behavior of the coefficients are
compared to the values from theory (21), which corresponds well.
ε[n] = d[n] − e[n]. (27)
VI. CONCLUSION
Substituting (24) and (26) into (27) yields
We have introduced a structure to identify and compensate gain
„ «T „ «T
1 1 and timing mismatches for TIADCs. The identification is based on
ε[n] = G0 cg − ĉg xg [n] + G0 cr − ĉr xr [n]. the well-known LMS algorithm for which many efficient implementa-
G0 G0
(28) tions exist. We have shown that we can nearly double the initial SNR
The error ε[n] will be zero if (21) is fulfilled. Assuming the of a 4-channel TIADC that is limited by gain and timing mismatches.
independence of xg [n] and xr [n], we can use the LMS to find the As the derivations have further shown, the method can be used for an

51
0
0.005

−10

−20 0
Energy Density Spectrum [dBc]

−30

Estimated Time Error


−40 −0.005

−50
−0.01
−60

−70
−0.015
−80
LMS Adaptation
−90 Ideal Values
0 0.1 0.2 0.3 0.4 0.5 −0.02
Normalized frequency Ω/Ω 0 2 4 6 8 10 12
s Number of Samples 4
x 10

Fig. 4. Energy density spectrum of the uncompensated output y[n] after Fig. 7. The convergence behavior of the timing mismatch coefficients ĉr [n]
convergence of the LMS algorithm (last 2048 samples). The SNR is 35.1 dB. compared to their expected values G1 cr = [−0.001, 0.0016, −0.005].
0

0
R EFERENCES
−10
[1] W. C. Black Jr. and D. A. Hodges, “Time-interleaved converter arrays,”
−20 IEEE Journal of Solid State Circuits, vol. 15, no. 6, pp. 1024–1029,
Energy Density Spectrum [dBc]

December 1980.
−30 [2] C. Vogel, “The impact of combined channel mismatch effects in time-
interleaved ADCs,” IEEE Transactions on Instrumentation and Measure-
−40 ment, vol. 54, no. 1, pp. 415–427, February 2005.
[3] J. Elbornsson, F. Gustafsson, and J. E. Eklund, “Blind adaptive equal-
−50
ization of mismatch errors in a time-interleaved A/D converter system,”
−60
IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 51,
no. 1, pp. 151–158, January 2004.
−70 [4] J. Elbornsson, F. Gustafsson, and J.-E. Eklund, “Blind equalization of
time errors in a time-interleaved ADC system,” IEEE Transactions on
−80 Signal Processing, vol. 53, no. 4, pp. 1413–1424, April 2005.
[5] C. Vogel, “A frequency domain method for blind identification of timing
−90 mismatches in time-interleaved ADCs,” in Proceedings of the 24th IEEE
0 0.1 0.2 0.3 0.4 0.5
Normalized frequency Ω/Ωs Norchip Conference, November 2006, pp. 45–48.
[6] V. Divi and G. Wornell, “Scalable blind calibration of timing skew
Fig. 5. Energy density spectrum of the compensated output yc [n] after in high-resolution time-interleaved ADCs,” in Proceedings of IEEE
convergence of the LMS algorithm (last 2048 samples). The SNR is 66 dB, International Symposium on Circuits and Systems, ISCAS, May 2006,
which is an improvement of 30.9dB compared to the uncompensated output. pp. 3390–3393.
[7] M. Seo, M. J. W. Rodwell, and U. Madhow, “Generalized blind
mismatch correction for a two-channel time-interleaved ADC: Analytic
approach,” in Proceedings of IEEE International Symposium on Circuits
0.02 and Systems, ISCAS, May 2007, pp. 109–112.
[8] S. Jamal, D. Fu, M. Singh, P. Hurst, and S. Lewis, “Calibration of
0.015 sample-time error in a two-channel time-interleaved analog-to-digital
converter,” IEEE Transactions on Circuits and Systems I: Regular
0.01 Papers, vol. 51, no. 1, pp. 130–139, January 2004.
Estimated Gain Error

[9] S. Huang and B. Levy, “Adaptive blind calibration of timing offset


0.005 and gain mismatch for two-channel time-interleaved ADCs,” IEEE
Transactions on Circuits and Systems I: Regular Papers, vol. 53, no. 6,
0 pp. 1278–1288, June 2006.
[10] ——, “Blind calibration of timing offsets for four-channel time-
−0.005
interleaved ADCs,” IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 54, no. 4, pp. 863–876, April 2007.
−0.01
[11] S. Saleem and C. Vogel, “LMS-based identification and compensation
of timing mismatches in a two-channel time-interleaved analog-to-
−0.015 LMS Adaptation digital converter,” in Proceedings of the 25th IEEE Norchip Conference,
Ideal Values November 2007.
−0.02
0 2 4 6 8 10 12 [12] S. Tertinek and C. Vogel, “Reconstruction of two-periodic nonuniformly
Number of Samples x 10
4
sampled bandlimited signals using a discrete-time differentiator and a
time-varying multiplier,” IEEE Transactions on Circuits and Systems II,
Fig. 6. The convergence behavior of the gain mismatch coefficients ĉg [n] vol. 54, no. 7, pp. 616–620, July 2007.
compared to their expected values G1 cg = [0.005, 0.01, 0]. [13] A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Discrete-Time Signal
0
Processing. Prentice Hall, 1999.
[14] S. Mendel and C. Vogel, “On magnitude response mismatch compensa-
tion for M-channel time-interleaved ADCs,” in Proceedings of the 2007
IEEE International Symposium on Circuits and Systems , ISCAS, May
arbitrary number of channels, whereas the implementation complexity 2007, pp. 3375–3378.
of the structure linearly scales with the number of channels. [15] S. Haykin, Adaptive filter theory. Prentice Hall, 2002.

52

Вам также может понравиться