profdoc.um.ac.irprofdoc.um.ac.ir/articles/a/1033148.pdfsimulation results in the 0.18 um cmos...
TRANSCRIPT
1 23
Analog Integrated Circuits and SignalProcessingAn International Journal ISSN 0925-1030Volume 74Number 2 Analog Integr Circ Sig Process (2013)74:317-330DOI 10.1007/s10470-012-9984-7
A new non-uniform adaptive-samplingsuccessive approximation ADC forbiomedical sparse signals
Maryam Zaare, Hassan Sepehrian &Mohammad Maymandi-Nejad
1 23
Your article is protected by copyright and all
rights are held exclusively by Springer Science
+Business Media New York. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you
wish to self-archive your work, please use the
accepted author’s version for posting to your
own website or your institution’s repository.
You may further deposit the accepted author’s
version on a funder’s repository at a funder’s
request, provided it is not made publicly
available until 12 months after publication.
A new non-uniform adaptive-sampling successive approximationADC for biomedical sparse signals
Maryam Zaare • Hassan Sepehrian •
Mohammad Maymandi-Nejad
Received: 6 July 2012 / Revised: 22 September 2012 / Accepted: 16 November 2012 / Published online: 7 December 2012
� Springer Science+Business Media New York 2012
Abstract This paper presents a new sampling technique
and a successive approximation analog to digital converter
(SA-ADC) which samples sparse signals in a non-uniform
adaptive way. The proposed sampling technique has the
capability to be incorporated in the structure of the
SA-ADC. The proposed SA-ADC changes the rate of
sampling in accordance with the rate of changes of the
signal. In this way, the data volume is reduced considerably
without losing the important information in the signal.
Simulation results in the 0.18 um CMOS technology shows
a power saving of up to 90.5 % and a compression ratio of
7.5 compared to the conventional sampling technique of
ECG signals.
Keywords Successive approximation ADC � Low power �Non-uniform sampling � Signal specific sampling �Biomedical signal � Electrocardiogram
1 Introduction
Bio-potential signals convey valuable information for
diagnostic purposes. In many cases it is desirable to have a
record of such signals for a relatively long period of time.
For example, in order to diagnose heart arrhythmias,
electrocardiography (ECG) recording is widely used. Since
heart arrhythmia may happen unexpectedly, long term
recording of the ECG signal is necessary for accurate
diagnosis. In these cases the ECG (or any other bio-
potential signal) recording device should be portable. This
requires the electronic recording device to consume very
little power and to have a relatively large memory to
sample and save the signal for several days (or even
weeks). Reducing the number of samples in the recording
device is critical in the reduction of the memory size as
well as the power consumption.
Figure 1 illustrates a sample of an ECG signal. The
main features of the ECG signal are shown by the P, Q, R,
S, and T letters. As can be seen in this figure the waveform
can be divided into regions of slow and fast transitions
(corresponding to ‘low activity’ and ‘high activity’ regions
in Fig. 1). The region around the QRS complex has the
fastest transitions. In other regions the signal variations are
moderate or slow. The required sampling rate depends on
the harmonics of the QRS complex and the application. For
example, the analysis of heart rate variations (HRV)
requires high resolution of time for the detection of the R
points. This demands a high sampling rate (250 Samples/s)
[1]. Sampling the whole signal with this high rate increases
the volume of data. In this case, many consecutive samples
in the slow transition regions are almost the same and do
not convey much information. If we can change the rate of
sampling in accordance with the rate of signal variations,
the data volume can be reduced considerably without los-
ing the important information in the ECG signal.
In order to reduce the volume of sampled data in a
monitoring device, the bio-signal is typically sampled and
converted to digital by an analog to digital converter
(ADC) at the Nyquist rate or more. The sampled data is
then compressed using compression technique to reduce
the size of the memory needed to save the data [2–6]. In
M. Zaare (&) � H. Sepehrian � M. Maymandi-Nejad
Integrated Systems Lab., Department of Electrical Engineering,
Ferdowsi University of Mashhad, Mashhad, Iran
e-mail: [email protected]
H. Sepehrian
e-mail: [email protected]
M. Maymandi-Nejad
e-mail: [email protected]
123
Analog Integr Circ Sig Process (2013) 74:317–330
DOI 10.1007/s10470-012-9984-7
Author's personal copy
this method, although the size of data is reduced, the power
consumption of the circuit prior to digitization is not
decreased. If the reduction of the sampling rate happens
during digitization, it can lead to lower power consumption
in the analog part of the circuit. Hence, the technique that is
used to reduce the number of samples should be compatible
with the architecture of the ADC. In bio-implantable
devices in which the sampled signal is to be transmitted out
of the body, reducing the sampling rate is also desirable
since it leads to lower transmitter power.
The successive approximation analog to digital con-
verter (SA-ADC) is a good candidate for low to moderate
sampling rates and can be designed to be very low power
[7–12]. In this paper we propose a new SA-ADC which
implements a non-uniform adaptive-sampling technique.
Using this SA-ADC the data volume can be reduced for
sparse bio-signals like ECG while the power consumption
is also decreased. In Sect. 2 of this paper a few of the state-
of-the-art techniques to reduce data volume for bio-signals
are reviewed. In Sect. 3, the proposed non-uniform adap-
tive-sampling algorithm is discussed and the MATLAB
simulation results are presented. The proposed SA-ADC is
presented in Sect. 4 and simulation results are provided.
We finally conclude the paper in Sect. 5.
2 Bio-potential signal sampling techniques
Since bio-potential signals are sparse in time-domain,
sampling with a fixed frequency is not suitable. Hence,
adaptive-sampling is used in bio-potential recording sys-
tems [1, 13, 14] in which the sampling frequency changes
with the rate of signal variations.
In [1] a sampling technique for ECG signal is presented
in which the sampling rate is controlled by an activity
detector block (ACTDET). The analog to digital conver-
sion is done by an 11-bit SA-ADC. The rate of signal
variations is detected by a switched capacitor (SC) differ-
entiator. The output of the differentiator is compared with a
threshold to choose the sampling frequency which can be
either 64 Hz (fL = 64 Hz) or 1,024 Hz (fH = 1024 Hz). fHis suitable for sampling the R signal of the ECG while fL is
suitable for other slow varying parts. The threshold voltage
(Vth) is made adaptive to let the circuit follow the variations
of the ECG signal.
Using this technique the data volume is decreased by a
factor of 7.3 compared to the case where the sampling is
done by a constant frequency of 1,024 Hz [1]. Although the
ratio of the total number of data without any compression
to the number of compressed data is a relatively good, the
implementation of this technique requires more complex
digital circuitry and several relatively power hungry analog
blocks; like opamp, filter, SC differentiator and compara-
tors. Hence, this technique would not be very power-effi-
cient. Moreover, using a differentiator increases the noise
sensitivity of the circuit. Also, in the calculation of the
compression ratio, the extra data that should be saved as an
indication of the exact time of sampling (typically known
as the time stamps or time flags) are not considered. Hence,
the effective compression ratio is lower.
In [13] a non-uniform sampling technique is presented
in which two different sampling frequencies are used. In
this method the maximum and minimum points of the
signal are detected and the sampling frequency is increased
around these points. Circuit level simulations have shown
that this technique can reduce the average sampling fre-
quency and the volume of data by 50 and 35 %, respec-
tively. For the implementation of this system two
differentiators are used. Using differentiator increases the
noise susceptibility of the system.
Another sub-Nyquist sampling technique is the asyn-
chronous sampling and has been known since 1950 [15]. In
the asynchronous technique instead of sampling the signal
at fixed instants, the signal is sampled whenever it crosses a
threshold voltage. This way the sampling rate matches the
rate of change of the signal.
In the adaptive asynchronous sampling, the threshold is
changed according to the slope of the signal which may
lead to loss of some of the features of the signal. In Ref-
erence [14] a modified adaptive asynchronous sampling is
used. In this technique each peak and valley of the signal is
saved and used to calculate the high and low threshold
voltages for the next period. When the signal rises (falls)
and crosses the high (low) threshold the system starts
sampling with the highest rate. This method is able to
detect the QRS complex of the ECG and the sampled data
Fig. 1 ECG waveform with
inherent time-domain sparsity
318 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy
cannot be used to recover the whole ECG signal. This is
due to the fact that the sampling times are not saved.
Moreover, the proposed algorithm is not optimum for
detecting local extrema, e.g. the P and T waves of the ECG.
A relatively new method for sampling sparse signals
with under-Nyquist rate is compressive sampling (CS). The
CS allows full recovery of the original signal. It has been
recently considered for EEG and ECG compression
[16–20]. In [16] digital CS is proposed in which com-
pression is done after digitization with an ADC. The ulti-
mate goal in CS is to bring CS algorithm down to the
analog part of the system before the ADC. However, there
is still a long way to reach this goal. Moreover, compres-
sion techniques based on CS require a complex digital
post-processing. In fact the main reason for using CS is to
reduce the work load and complexity of the coder at the
expense of more complex decoder.
There are other sampling techniques that can theoreti-
cally decrease the volume of data but they require a
complex function to be implemented [21]. These kinds of
techniques are not desirable from a circuit point of view.
Based on the above discussions, it is clear that the best
way for sampling a sparse signal with minimum data vol-
ume and power consumption is to sample the signal with a
sub-Nyquist rate such that it can be accommodated in the
sampler circuit (before or during the ADC). The under-
sampling should be done in an adaptive way (and not
stochastically) so that the signal can be recovered without
much difficulty.
3 The proposed non-uniform adaptive-sampling
technique
3.1 Basic operation
In this section we present the proposed non-uniform
adaptive-sampling technique. In this method, sampling is
done based on the activity of the signal. The system has
two clock signals CLKH and CLKL with frequencies fH (the
high sampling frequency) and fL (the low sampling fre-
quency), respectively. The relationship between these
frequencies will become clear later. For the sake of illus-
tration, Fig. 2 shows a typical ECG signal sampled with
this technique. As can be seen the ECG signal is composed
of regions with slow and fast variations. This figure shows
the two clocks used in the sampling system. The signal is
monitored at a rate determined by CLKH to detect the rate
of change of the signal. During the fast transition regions
the signal is sampled by fH and during the slow transition
regions the signal is sampled by fL. The CLKsample in Fig. 2
is the overall clock which represents the sampling instants.
Clearly, the sampling rate is changing based on the rate of
variation of the ECG.
An important issue in the proposed technique is how to
detect the fast and slow transition regions. This is shown in
the flow chart of Fig. 3 which illustrates the procedure of
the proposed sampling technique. Let TH and TL be the
period of CLKH and CLKL, respectively. We also define
m as the ratio of TL to TH (m = TL/TH) and VDAC(PS) as the
value of the last sampled signal (i.e. the last digital code
converted to analog by a digital-to-analog converter). The
sampling algorithm starts by initializing the value of
VDAC(PS). The input signal at time nTH (Vi(n 9 TH)) is
compared with VDAC(PS). If the absolute value of the dif-
ference is not larger than a threshold value (Vth), then the
signal variation has not been fast enough and the signal is
not sampled. However, if the time interval between n 9 TH
and the last sampling instant is equal to TL (the condition
satisfied by Mod(n, m) = 0) the signal should be sampled.
If the absolute value of the above difference is larger than
Vth then the signal variations has been fast enough and the
signal is sampled. The value of the VDAC(PS) is also updated
for the next sampling. Note that the signal is being moni-
tored by fH, however, it is not necessarily sampled with this
frequency. In fact the sampling is done at a rate whose
average is more than fL and less than fH. To obtain the best
results in terms of the lowest distortion in the recovered
signal and not losing important information of the signal,
the three parameters fL, fH, and Vth should be chosen
carefully. fH is typically chosen large enough so that the
highest frequency harmonic present in the signal can be
sampled properly (at least the Nyquist rate of the signal).
This can be obtained from the short time fourier transform
Fig. 2 A cycle of ECG signal
sampled with the proposed
technique
Analog Integr Circ Sig Process (2013) 74:317–330 319
123
Author's personal copy
(STFT) [22]. fL is chosen based on the frequency contents
of the P and T waves of the ECG [1]. We have obtained the
best value for Vth based on system level simulations using
MATLAB, as will be explained later.
The above algorithm can be implemented by an
SA-ADC and a digital control circuit. Compared to other
signal specific sampling techniques [1], [13], the proposed
method does not require power hungry analog block and
differentiators and leads to a considerable power saving
besides a reduction in the data volume.
3.2 MATLAB simulation results
In order to show the effectiveness of the proposed sampling
technique, we have used the ECG signals from the MIT-BIH
Arrhythmia database [23]. To compare the performance of the
proposed technique with the other state of the art methods we Ta
ble
2S
yst
emle
vel
sim
ula
tio
nre
sult
s
10
Bit
AD
C@
VD
D=
1V
,f H
=1
08
0H
z,f L
=6
7.5
Hz
(tes
ted
for
48
reco
rds
of
MIT
-BIH
Arr
hy
thm
ia)
Dx
±3
0L
SB
±3
5L
SB
±4
0L
SB
±4
5L
SB
±5
0L
SB
±5
5L
SB
±6
0L
SB
±6
5L
SB
±7
0L
SB
±7
5L
SB
±8
0L
SB
Nto
tal
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
1,0
80
NP
ropose
dM
ax2
75
24
12
23
21
01
89
17
91
64
16
41
46
14
61
39
Min
11
11
07
10
41
00
79
94
94
91
89
86
85
Av
g1
49
13
81
29
12
21
17
11
21
08
10
51
02
99
96
CR
Max
6.9
97
.29
7.4
97
.79
7.9
59
8.4
98
.29
8.5
98
.79
8.9
79
9.1
9
Min
2.8
93
.29
3.4
69
3.6
79
4.1
94
.39
4.7
94
.79
5.3
95
.39
5.5
9
Av
g5
.39
5.7
96
96
.49
6.6
79
6.9
97
.29
7.4
39
7.6
97
.89
89
PR
DM
ax7
.1%
7.5
%7
.45
%7
.6%
8%
8.2
%8
.69
%2
0%
20
.2%
20
.2%
20
%
Min
1.5
%1
.75
%2
.26
%2
.1%
2.2
%1
.87
%2
.7%
2.5
%2
.5%
2.4
%2
.68
%
Av
g4
%4
.3%7
4.4
%4
.6%
4.8
%5
.1%
5.2
%5
.56
%5
.6%
5.8
%6
.1%
Fig. 3 The flow chart of the proposed sampling technique
Table 1 Classification of signal quality versus PRD
PRD Reconstructed signal quality
0 * 2 % ‘‘Very good’’ quality
2 * 9 % ‘‘Very good’’ or ‘‘good’’ quality
C9 % Not possible to determine the quality group
320 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy
use two popular criteria, i.e. compression ratio (CR) and per-
centage root-mean-square difference (PRD). CR is the ratio of
the total number of data without any compression to the
number of compressed data. PRD is defined as the following:
PRD ¼ kX �~Xk2
kXk2
� 100 ð1Þ
where X is the original signal vector and ~X is the recon-
structed signal. To calculate the PRD, the signal is regen-
erated from the sampled data using linear interpolation.
Table 1 shows the classification of the signal quality in
terms of the PRD [24].
We have run simulations on all 48 records of the MIT-
BIH database and for each parameter a minimum, maxi-
mum, and average (over all records) value is obtained for
PRD and CR. Table 2 shows the simulation results. In
these simulations the total number of samples is 1080. The
high and low clock frequencies (fH and fL) are 1,080 and
67.5 Hz, respectively. The value of 1,080 for fH is high
enough for diagnosis purposes. The low clock frequency of
67.5 is 1/16 of fH = 1,080. This ratio between fL and fHleads to a simpler circuit for implementing the sampling
technique. We consider a 10-bit ADC with a reference
voltage of 1 V for digitization of the signal. As mentioned
above the value of Vth is very important in the proper
Fig. 6 Record number 101 (a), 116 (b), and 231 (c) of the MIT-BIH
database and the sampling times using the proposed signal specific
sampling technique
Fig. 5 CR and PRD versus Dx
Fig. 4 The average PRD versus the average CR
Analog Integr Circ Sig Process (2013) 74:317–330 321
123
Author's personal copy
operation of the proposed method. Here we use the term Dx
as the equivalent digital value for Vth. Increasing Dx leads
to more data volume reduction and a larger CR. It also
reduces the power consumption.
The recovery of the original signal from the sampled
data requires the exact time of the samplings if a non-
uniform sampling is used. The method of saving these time
flags is important in reducing the volume of data. The
Fig. 7 Three records of MIT-
BIH shown in Fig. 7 and the
reconstructed signals
322 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy
technique that we have used for saving the time flags along
with the data is as follow: Since, the ratio of fH to fL is 16
we can assign a 4-bit digit for representing the sampling
time with respect to the previous sampled data. In this
technique each sample is saved by 14 bits (10 bits for the
data and 4 bits for the time flag). The 4-bit time flag shows
the distance of a sample from the previous one, in terms of
the number of CLKH pulses. The compression ratio that
includes the impact of these extra bits is represented in
Table 2 by CR and can be found by the equation below.
CR ¼ Ntotal �M
NProposed � ðMþ 4Þ ð2Þ
where M is the number of bits for each sample (M = 10),
Ntotal and NProposed are the total number of samples withoutFig. 8 Conventional SA-ADC
Fig. 9 The proposed structure
Analog Integr Circ Sig Process (2013) 74:317–330 323
123
Author's personal copy
any compression and the number of samples with the
proposed non-uniform sampling technique, respectively.
According to Table 2, for a Dx of ±60LSB (1LSB =
0.98 mV) the maximum and average PRD are 7.2 and
5.2 % respectively. Although the average value of PRD is
less than 9 (good range in Table 1) for Dx up to ±80LSB, we
have to make sure the signal quality is good for all the
records. Therefore, we choose Dx = ±60LSB to consider
the worst case. For this value of Dx the total number of
samples is reduced from 1,080 to 108 which corresponds to
an average CR of 7.2 and an average PRD of 5.2 %. Fig. 4
illustrates the average PRD versus the average CR and Fig. 5
shows CR and PRD in terms of Dx. According to these
figures and Table 1, there is a limit on CR and Dx if the
signal quality is to be maintained within a specific range.
Figure 6 shows three ECG records from the MIT-BIH
database and their samples using the proposed technique. In
these figures Dx is chosen equal to ±60LSB. As can be seen,
when the signal variation is fast the sampling rate has increased
while for slow varying intervals the sampling rate is low. In
Fig. 7 the original signals (record numbers 101, 116, and 231)
and the recovered signals are shown simultaneously. Clearly,
the recovered signal faithfully represents the original one.
4 The proposed non-uniform adaptive-sampling
SA-ADC
The technique presented above has the ability to be
incorporated in an SA-ADC easily. The SA-ADC is
Fig. 10 Clock signal and decision times
Fig. 11 Control logic block and its signals
324 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy
suitable for slow to moderate sampling rates and low power
applications. Hence, it is a good choice for bio-medical
applications [11, 12]. Moreover, in an SA-ADC before a
sampling starts the analog value of the previous sample is
saved on the capacitors of its DAC. This is what we need in
the proposed technique. Also, the addition and subtraction
of the Vth can be accomplished in the digital control block
of the SA-ADC using a digital adder.
Figure 8 shows the schematic diagram of a conventional
N-bit SA-ADC consisting of a sample and hold (S/H), a
comparator, a successive approximation register (SAR) and
a capacitive digital-to-analog converter (DAC). In the
Table 3 The truth table of the SERF adder [25]
A B Cin SUM Cout
0 0 0 0 0
0 0 1 Vdd-Vth 0
0 1 0 1 Vtp
0 1 1 0 1
1 0 0 1 \Vtp
1 0 1 Vth 1
1 1 0 0 [Vdd-Vth
1 1 1 [Vdd-Vth 1
Fig. 12 Proposed SERF adder
Fig. 13 Block diagram of
10-bit adder
Analog Integr Circ Sig Process (2013) 74:317–330 325
123
Author's personal copy
sampling phase, the analog input signal is sampled and
held on capacitor Cs. At the same time the SAR is reset and
all DAC capacitors are discharged. In the next N clock
phases, the SA-ADC uses the binary search algorithm to
find the digital code corresponding to the analog voltage
stored on Cs [10]. At the end of a conversion, the analog
value of the N-bit digital number is stored on the DAC
capacitors.
4.1 Proposed structure
The structure of the proposed SA-ADC is shown in Fig. 9.
Compared to a conventional SA-ADC, this structure has a
multiplexer, an N-bit adder, and a control logic block. All
these blocks are implemented using digital circuit. Hence,
they are not power hungry. The proposed ADC requires
two more clock cycles for each conversion. These extra
cycles are used to check whether the absolute value of the
input signal variation is larger than Vth or not. The opera-
tion of the proposed architecture is as follows. First, Vi is
sampled on the sampling capacitor Cs (Vi(nTH) or Vi(n), in
short). Next, Vi(n) is compared with the voltage stored on
the DAC capacitors, which is due to the last conversion
(VDAC(PS)). Depending on the value of Vi(n) two different
directions may be taken the ADC.
– If Vi(n) [ VDAC(PS), the comparator output is zero. The
output of the comparator is latched in the control circuit
and is applied to the carry input of the adder. Therefore,
Dx (the digital number representing Vth) is added to the
last sampled digital code (available at the output of
the multiplexer). In the next clock cycle, Vi(n) is
compared with the new value stored in the DAC, i.e.
Table 4 Transistor sizes of SERF adder
X1 X2 X3 X4 X5–X8 M1 M2 M3 M4 M5 M6–M9
WL (lm) 1.5/0.18 0.5/0.18 1.5/0.18 0.5/0.18 0.5/0.18 0.5/0.18 1.5/0.18 0.5/0.18 1.5/0.15 2.5/0.18 0.5/0.18
Fig. 14 The schematic of the
comparator
Fig. 15 The schematic of the bootstrap switch
Table 6 Transistor sizes and capacitor values of bootstrap switch
WL (lm) of M1–M3, M5, M7,
M9–M13
WL (lm) of M4, M8 C1–C2 C3
1/0.18 3/0.18 10fF 500fF
Table 5 Transistor sizes of comparator
P0–P6 N0–N2 N3, N6 N4, N5 M1, M3 M2, M4
WL (lm) 0.5/0.18 0.5/0.18 1/0.18 0.5/0.18 6.1/0.18 1.8/0.18
326 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy
VDAC(PS) ? Vth. If Vi(n) \ VDAC(PS) ? Vth then the sig-
nal variation has not been large enough and the ADC
does not convert the signal into digital. In this case the
control circuit subtracts Vth from the DAC voltage using
2’s complement of Dx. Otherwise, the input signal (Vi(n))
is converted to digital and, as a result, Vi(n) is stored in the
DAC capacitors for the next sampling.
– If Vi(n) \ VDAC(PS), the comparator output is one. The
output of the comparator is latched in the control circuit
and is applied to the carry input of the adder. In this case
Dx is subtracted from VDAC(PS) to generate VDAC(PS)-
Vth. In the next clock cycle, Vi(n) is compared with the
new value stored in the DAC. If Vi(n) [ VDAC(PS)-Vth
then the signal variation has not been large enough and
the ADC does not convert the signal into digital.
Otherwise, the input signal (Vi(n)) is converted to digital
and the value of the DAC voltage is updated.
As mentioned above, the proposed architecture requires
two more clock cycles. This is illustrated in Fig. 10. These
extra cycles are used to implement the above mentioned
decisions. As shown in this figure, in the first half-cycle the
input voltage is sampled. In the second half-cycle the
sampled voltage is compared with VDAC(PS) and the result
of the comparison is latched in a DFF to be used as the
carry input of the adder. In the third half-cycle the adder
adds ±Vth to the value of the DAC voltage. In the last
(fourth) half-cycle the comparator compares the sampled
input signal with VDAC(PS) ± Vth. If the sampled signal is to
be converted to digital the DAC is reset and the next N
clock cycles are used for conversion (assuming N-bit
converter). Otherwise, the voltage of the DAC is restored
to its previous value and the comparator turns off for the
rest of the N clock cycles.
All the above functions are implemented by the control
logic block. The schematic of the control block and the
timing diagram of the control signals are shown in Fig. 11.
The control signals Clk1, add_select, CLKcomp, res1,
Comp_ctrl should be generated using appropriate digital
gates (not shown in Fig. 11 for conciseness). When the
Fig. 16 Waveforms of a Sampling Clock, Clksample, b DAC output voltage, VDAC, c supply current of the comparator, d DAC current taken from
Vref and e supply current of the adder
Analog Integr Circ Sig Process (2013) 74:317–330 327
123
Author's personal copy
Comp_ctrl signal is 1, the comparator is on for the com-
parisons needed at the beginning of each monitoring cycle.
The signal CLKsample indicates the times that the compar-
ator should be on for a conversion by applying the supply
voltage to the comparator (Comp_enable). The output of
the comparator is latched in a DFF by CLKcomp. The carry
input (Cin) of the adder and a signal called Conv_En is
generated from the output of the DFF. Cin goes to the
adder and Conv_En indicates whether the ADC should
convert the signal or not by generating the CLKsample
control signal. The Select signal is used for the control
signal of the multiplexer (Fig. 9). The Reset signal is
generated for resetting the DAC, if necessary.
4.2 ADC circuit and simulation results
The proposed non-uniform adaptive-sampling SA-ADC is
implemented in the 0.18 lm CMOS technology for
10-bits of resolution. The supply voltage is 1 V. For the
digital adder the sense energy recovery full adder (SERF)
reported in [25] is a suitable candidate due to its low
power consumption. However, the main problem of this
adder is that for some combination of inputs (Table 3) the
outputs can not reach GND or Vdd owing to the fact that
the outputs are generated by pass transistors. This can
lead to extra power dissipation in the succeeding stages.
To solve this problem we have modified the adder as
shown in Fig. 12. In the modified structure, pass transis-
tors are replaced by transmission gates and XOR function
is implemented besides the XNOR function. The modified
SERF full adder cell is used to make a 10-bit adder as
shown in Fig. 13. Table 4 shows the size of all the SERF
adder transistors.
The schematic of the comparator is illustrated in Fig. 14
[7]. The inputs of this comparator are applied to both
PMOS and NMOS transistors. Hence, the input range is rail
to rail. The transistor sizes of the comparator are shown in
Table 5. The SA-ADC needs a switch to sample the input
signal. In order to avoid limiting the input signal swing, we
Table 7 Simulation results of the proposed SA-ADC
Proposed adaptive-
sampling off
Proposed
adaptive-sampling
on
Average sampling rate 1,000 Hz 95 Hz
Data Size 10 bit 14 bit
Data Rate 10,000 bit/S 1,330 bit/S
Average power over one
ECG period
570 nW 54 nW
Total memory required
for 3-day recording
324 Mbyte 43.1 Mbyte
Power saving 90.5 %
Area 90,832 lm2 Area overhead: 6.5 %
Table 8 Performance comparison with other designs
Parameter [1] [20] This work
Technology 0.5 lm 90 nm 180 nm
Supply voltage 2 V 0.6 V 1 V
Power
consumption
5.2 lW
(including the
SPI power)
1.9 lW 0.58 lW
Power saving NA NA 90.5 %
Area Conventional
SA-ADC:
723,600 lm2
90,000 lm2 Conventional
SA-ADC:
84,957 lm2
ACTDET:
918000 lm2Extra logic:
5,875 lm2
Sampling rate Between 64 and
10,24 Hz
\20 kHz 95 Hz
CR 7 10 7.5
PRD NA 10 dBa 2.67 %
Simulation/
Measurement
Measurement Measurement Simulation
a This is SNDR and is defined as the reference signal energy divided
by the error energy between the reconstructed signal and the reference
Fig. 17 Record number 231 of the MIT-BIH database. Sampling
times (a), original and recovered signals (b)
328 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy
have used a bootstrap switch. The schematic of the boot-
strap switch is shown in Fig. 15 [26]. Table 6 illustrates
transistor sizes and capacitor values of bootstrap switch.
Figure 16 illustrates sampling clock, output voltage of
the DAC, the supply current of the comparator and the
adder, as well as the current taken from Vref. The Clksample
signal shows the periods in which the sampling is done.
Whenever Clksample is high a conversion has happened and
when it is low, the input is not converted to digital. During
the periods that Clksample is low, the input has been less
than ±60LSB. The current waveforms in this figure show
the times that each block is not active. When a block is
active a current is taken from the supply or Vref.
In order to check the performance of the SA-ADC an
ECG signal from the MIT-BIH data base (record 231) is
applied to the ADC. Fig. 17 depicts the original and the
recovered signals. Simulation results of the SA-ADC are
shown in Table 7. As can be seen, the average sampling
rate is reduced to 95 Hz. The average power consumption
of the proposed ADC is lowered by 90.5 % and the number
of bits of the required memory is reduced by 86.7 %.
Simulated specifications of the proposed SA-ADC are
compared with the state-of-the-art implementations in
Table 8. The proposed architecture exhibits considerable
power saving, while the area overhead is 6.5 %. The power
consumption reported in Table 8 is the total power of the
circuit for converting of one sample.
5 Conclusions
A new non-uniform signal specific sub-Nyquist sampling
technique for sparse signals has been proposed. The pre-
sented algorithm monitors the signal with a high frequency
and adaptively adjusts the sampling rate based on the
activity of the signal. Hence, the average sampling rate is
reduced, resulting in considerable power saving and
smaller memory size requirement. Circuit level simulations
of the proposed technique show a compression ratio of 7.5
for the ECG signal and a power saving of 90.5 %. Total
memory required for 3-day recording in conventional and
proposed system is 324 and 43.1 Mbyte, respectively.
References
1. Yazicioglu, R.F., Kim, S., Torfs, T., Kim H., Hoof, C.V., A 30uW
analog signal processor ASIC for portable biopotential signal
monitoring. IEEE Jornal of Solid-State Circuits, 46(1), Jan. 2011.
2. Jalaleddine, S., Hutchens, C., Strattan, R., & Coberly, W. (1990).
ECG data compression techniques—A unified approach. IEEETransactions on Biomedical Engineering, 37, 329–343.
3. Chen, J., & Itoh, S. (1998). A wavelet transform-based ECG
compression method guaranteeing desired signal quality. IEEETransactions on Biomedical Engineering, 45(12), 1414–1419.
4. Rajoub, B. A. (2002). An efficient coding algorithm for the
compression of ECG signals using the wavelet transform. IEEETransactions of Biomedical Engineering, 49(4), 355–362.
5. Michaeli, T., & Eldar, Y. C. (2012). Xampling at the rate of
innovation. Signal Processing, IEEE Transactions on, 60(3),
1121–1133.
6. Bendifallah, A., Benzid, R., & Boulemden, M. (2011). Improved
ECG compression method using discrete cosine transform.
Electronics Letters, 47(2), 87–89.
7. Hong, H., & Lee, G. (2007). A 65fJ/conversion-step 0.9-V
200-kS/s rail-to-rail 8-bit successive approximation ADC. IEEEJournal Solid-State Circuits, 42(10), 2161–2168.
8. Lotfi, R., Majidi, R., Maymandi-Nejad, M., Serdijn W.A. (2009).
An ultra-low-power 10-Bit 100-kS/s successive-approximation
analog-to-digital converter. Proceedings IEEE International
Symposium on Circuits and Systems (ISCAS), 1117–1120.
9. Elzakker M., et al. (2008) A 1.9 lW 4.4fJ/conversion-step 10b
1MS/s charge-redistribution ADC. IEEE Solid-State Circuits
Conference Digest of Technical Papers, 244–246.
10. Sepehrian, H., Saberi, M., Lotfi, R. (2011) A signal-specific
successive-approximation analog-to-digital converter. Proceeding
IEEE International Symposium on Circuits and Systems,
1624–1627.
11. Liew, W.-S., Yao, L., Lian, Y. (2008). A moving binary search
SAR-ADC for low power biomedical data acquisition system.
Proceeding IEEE Asia Pacific Conference on Circuits and Sys-
tems, 646–649.
12. Zou, X., et al. (2008). A 1-V 1.1-lW sensor interface IC for
wearable biomedical devices. Proceeding IEEE International
Symposium on Circuits and Systems, 2725–2728.
13. Rieger, R., Chen, S. (2006). A signal based clocking scheme for
A/D converters in body sensor networks, TENCON 2006. 2006
IEEE Region 10 Conference, 1–4.
14. Agarwal, R., & Sonkusale, S. R. (2011). Input-feature correlated
asynchronous analog to information converter for ECG moni-
toring. IEEE Transactions on Biomedical Circuits Systems, 5(5),
459–467.
15. Ellis, P. H. (1959). Extension of phase plane analysis to quantized
systems. IRE Transactions on Automatic Control, 4, 43–59.
16. Mamaghanian, H., Khaled, N., Atienza, D., & Vandergheynst, P.
(2011). Compressed sensing for real-time energy-Efficient ECG
compression on wireless body sensor nodes. IEEE TransactionsBiomedical Engineering, 58(9), 2456–2466.
17. Allstot, E.G., Chen, A.Y., Dixon, A.M.R., Gangopadhyay, D.,
Allstot, D.J. (2010) compressive sampling of ecg bio-signals:
quantization noise and sparsity considerations. Proceeding IEEE
Biomedical Circuits and Systems Conference, 41–44.
18. Aviyente, S. (2007). Compressed sensing framework for EEG
compression. Proceedings of the IEEE Workshop on Statistical
Signal Proccessing, 181–184.
19. Enay, S.S., Chaparro, L.F. Sun, M., Sclabassi, R.J. (2008).
Compressive sensing and random filtering of EEG signals using
slepian basis. Proceedings of the EURASIP EUSIPCO’08.
20. Chen, F., Chandrakasan, A. P., & Stojanovic, V. M. (2012).
Design and analysis of a hardware-Efficient compressed sensing
architecture for data compression in wireless sensors. IEEEJournal of Solid-State Circuits, 47(3), 744–756.
21. Feizi, S., Goyal, V. K., & Medard, M. Time-stampless adaptive
nonuniform sampling for stochastic signals. http://arxiv.org/abs/
1110.3774. Accessed 17 Oct 2011.
22. Goodwin, M. M. (1997). Adaptive signal models: Theory, algo-
rithms, and audio applications, PhD dissertation, university of
California, Berkeley.
Analog Integr Circ Sig Process (2013) 74:317–330 329
123
Author's personal copy
23. MIT-BIH arrhythmia database. [Online]. http://www.physionet.
org/physiobank/database/mitdb/.
24. Zigel, Y., Cohen, A., & Katz, A. (2000). The weighted diagnostic
distortion (WDD) measure for ECG signal compression. IEEETransactions on Biomedical Enggineering, 47, 1422–1430.
25. Moradi, F., Wisland, D.T., Mahmoodi, H., Aunet, S., Cao, T.V.,
Peiravi, A. (2009) Ultra low power full adder topologies. Pro-
ceeding IEEE International Symposium Circuits and Systems,
3158–3161.
26. Abo, A. M., & Gray, P. R. (1999). A 1.5-V, 10-bit, 14.3-MS/s
CMOS pipeline analog-to-digital converter. IEEE Journal ofSolid-State Circuits, 34(5), 599–606.
Maryam Zaare was born in
Mashhad, Iran, in 1983. She
received the B.S. and M.S. degrees
in electrical engineering, in 2006
and 2008, respectively, from
Ferdowsi University of Mashhad,
Mashhad, Iran, where she is cur-
rently working toward the Ph.D.
degree. Her research interests
include design of analog and
mixed-signal integrated circuits
and systems, data converters, and
biomedical circuits and systems.
Hassan Sepehrian was born in
Mazandaran, Iran, in 1984. He
received the B.S. from Shahed
University of Tehran, Iran, in
2008 and M.S. degrees from
Ferdowsi University of Mash-
had, Mashhad, Iran, in 2010,
both in electronics engineering.
He is currently working in NRP
Company in Tehran, Iran. His
research interests include design
of analog and mixed-signal
integrated circuits and systems,
data converters, and biomedical
circuits and systems.
Mohammad Maymandi-Nejadreceived the B.Sc. degree from
Ferdowsi University of Mash-
had, Mashhad, Iran, in 1990 and
the M.Sc. degree from Khajeh
Nassir Tossi University of
Technology, Tehran, Iran, in
1993, and the Ph.D. degree from
the University of Waterloo,
Waterloo, ON, Canada, in 2005,
all in electronics engineering.
From 1994 to 2001, he was a
Lecturer in the Department of
Electrical Engineering, Fer-
dowsi University of Mashhad,
where he was engaged in teaching and research and also conducted
several industrial projects in the field of automation and computer
interfacing. He is currently an Assistant Professor in the same
department. His current research interests include low-voltage, low-
power analog ICs and their applications in biomedical circuits and
systems. Dr. Maymandi-Nejad received the Strategic Microelectron-
ics Council of Information Technology Academia Collaboration
(ITAC) Industrial Collaboration Award in 2005 for his work on a
wireless bioimplantable device for monitoring blood pressure of
transgenic mice.
330 Analog Integr Circ Sig Process (2013) 74:317–330
123
Author's personal copy