1572 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO.

2, FEBRUARY 2017

An Accurate BJT-Based CMOS Temperature

Sensor With Duty-Cycle-Modulated Output
Guijie Wang, Ali Heidari, Kofi A. A. Makinwa, Fellow, IEEE, and Gerard C. M. Meijer, Senior Member, IEEE

Abstract—This paper describes the design of a precision generated a duty-cycle-modulated output signal [1] and which
bipolar junction transistor based temperature sensor im- was implemented in BICMOS technology. Nowadays, a wide
plemented in standard 0.7-µm CMOS technology. It employs range of smart temperature sensors are available in low-cost
substrate p-n-ps as sensing elements, which makes it insen-
sitive to the effects of mechanical (packaging) stress and CMOS technology [2]–[5]. However, it still makes sense to
facilitates the use of low-cost packaging technologies. The design sensors with duty-cycle-modulated outputs, because such
sensor outputs a duty-cycle-modulated signal, which can signals have a number of useful and important features:
easily be interfaced to the digital world and, after low-pass 1) Usability in both analog and digital systems: Compared
filtering, to the analog world. In order to eliminate the errors to the more widely used sigma–delta modulators, an at-
caused by the component mismatch, chopping and dynamic
element matching (DEM) techniques have been applied. tractive feature of duty-cycle modulators is that they can
The required component shuffling was done concurrently be easily and robustly connected to digital systems, such
rather than sequentially, resulting in a fast DEM scheme that as microcontrollers, as well as to analog systems, such as
saves energy without degrading accuracy. After a single- thermostats (see [6, Ch. 10]).
temperature trim, the sensor’s inaccuracy is ±0.1 °C (−20 2) Low-energy consumption: Often, the interface circuitry of
to 60 °C) and ±0.3 °C (−45 to 130 °C), respectively. Mea-
surements of sensors in different packages show that the a smart temperature sensor consumes more energy than
package-induced shift is less than 0.1 °C. Measurements of the actual sensor itself. It then makes sense to perform the
eight sensors over 367 days show that their output drift is required signal processing in an external microcon-
less than 6 mK. While dissipating only 200 µW, the sensor troller as much as possible, so that the sensor’s energy
achieves a resolution of 3 mK (rms) in a 1.8-ms measure-
consumption (and self-heating) is minimized.
ment time, and a state-of-the-art resolution figure of merit
of 3.2 pJK2 . This combination of high accuracy, high resolu- Recently, a CMOS smart temperature sensor with a duty-
tion, high speed, and low-energy consumption makes this cycle-modulated output signal has been presented [10]. In this
sensor suited for commercial and industrial applications. paper, more details are disclosed together with the results of ex-
Index Terms—Chopping, CMOS temperature sensor, tensive measurements, which were done to characterize the new
duty-cycle modulation, dynamic element matching (DEM), sensor for industrial applications. Compared to an earlier design
one-point trim. with a duty-cycle-modulated output [1], the main performance
objectives of the new design are as follows:
I. INTRODUCTION 1) better accuracy and lower sensitivity to packaging shift;
NE of the first smart sensors that could be read out by 2) higher resolution at even higher acquisition rates;
O simple microcontrollers was a temperature sensor that 3) lower energy consumption per measurement;
4) better long-term stability.
Early temperature sensors with duty-cycle-modulated outputs
Manuscript received February 4, 2016; revised May 26, 2016; ac-
cepted August 22, 2016. Date of publication September 28, 2016; date [1], [11], achieved a good accuracy by exploiting the benefits of
of current version January 10, 2017.This work was supported in part by bipolar or BiCMOS technology, e.g., good component matching
Smartec BV, The Netherlands, and in part by the Electronic and the availability of high-performance bipolar junction tran-
Instrumentation Laboratory, Delft University of Technology.
G. Wang is with Smartec BV, 4811 Breda, The Netherlands, and sistors (BJTs). Compared to CMOS technology, however, these
also with the Electronic Instrumentation Laboratory, Delft University of benefits came at the expense of higher manufacturing cost. Al-
Technology, 2628 Delft, The Netherlands (e-mail: g.wang@tudelft.nl). though the analog performance of CMOS technology is arguably
A. Heidari is with Smartec BV, 4811 Breda, The Netherlands, with
the Electronic Instrumentation Laboratory, Delft University of Technol- poorer, later work has demonstrated that it can also be used
ogy, 2628 Delft, The Netherland, and also with Guilan University, Rasht to realize accurate temperature sensors [12]–[16]. These sen-
41996-3475, Iran (e-mail: A.Heidari@tudelft.nl). sors employ dynamic error-correction techniques, such as chop-
K. A. A. Makinwa is with the Electronic Instrumentation Laboratory,
Delft University of Technology, 2628 Delft, The Netherlands (e-mail: ping, correlated-double sampling, and dynamic element match-
K.A.A.Makinwa@tudelft.nl). ing (DEM), to mitigate the effects of component mismatch,
G. C. M. Meijer is with the Electronic Instrumentation Laboratory, and employ compensation schemes to reduce the effects of the
Delft University of Technology, 2628 Delft, The Netherlands, and also
with SensArt, 2611 Delft, The Netherlands (e-mail: G.C.M.Meijer@ low current gain of the available substrate p-n-ps [12], [16].
tudelft.nl). Straightforward implementation of such techniques would re-
Color versions of one or more of the figures in this paper are available quire complex circuitry and, thus, too much chip area. Fur-
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIE.2016.2614273 thermore, the required signal processing would then require

0278-0046 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ACCURATE BJT-BASED CMOS TEMPERATURE SENSOR WITH DUTY-CYCLE-MODULATED OUTPUT 1573

Fig. 2. Charge and discharge current in this design.

Fig. 1. Operation principle of the temperature sensor. (a) Basic

principle; (b) the voltage across the capacitor C; (c) various (extrapo-
lated) currents as a function of the temperature.

hundreds of periods of the output signal, leading to a low data

rate and high-energy consumption. In this paper, it will be shown
how the required signal processing can be sped up, so that the
data rate and energy consumption can be minimized without
losing performance.
In CMOS technology, the temperature-sensing elements can
be BJTs [12]–[16], MOSFETs [17], [18], or resistors [19]–[22].
The best accuracy has been achieved with BJTs after a one-point
trim [12]–[16]. On the other hand, sensors based on MOSFETs
Fig. 3. Principle schematic circuit.
or resistors can operate from low supply voltages, even below
1 V, but achieve lower accuracy, and sometimes even require
multi-point trimming [21], [22], thus significantly increasing It should be noted that the value of D is independent of the
calibration costs. Furthermore, the substrate p-n-ps available exact value of the ST’s threshold voltages V1 and V2 and of the
in CMOS technology turn out to be quite insensitive to the capacitance C. The two currents I1 and I2 are designed to be
mechanical stress induced by low-cost plastic packaging [23]– temperature dependent: In its simplest implementation, I1 is lin-
[25]. Consequently, excellent long-term stability and robustness early proportional to absolute temperature (PTAT), while I2 is
to extreme thermal cycling can be achieved. complementary-to-absolute temperature (CTAT) [see Fig. 1(c)].
The sensor employs a continuous-time duty-cycle modula- Furthermore, if the sum Iref = I1 + I2 is designed to be tem-
tor whose system-level design is described in Section II. This perature independent [as indicated in Fig. 1(c)], then D will be
is followed, in Section III, by a description of a self-clocked a linear function of the temperature.
chopping and DEM scheme that averages component mismatch In a CMOS process, the CTAT current I2 can be derived from
over several periods of the modulator’s output. Straightforward the base–emitter voltage VBE of a substrate p-n-p, while I1 can
application of this scheme would require 256 periods. However, be derived from the difference ΔVBE between the base–emitter
it will be shown how this can be reduced to just eight without voltages of two appropriately biased p-n-ps. However, as shown
significant loss of accuracy. Details about the circuit implemen- in Fig. 1(c), the resulting duty cycle D will then vary by only
tation and signal processing are presented in Section IV and V, about 30% over the desired temperature range −45 to 130 °C.
respectively. Measurement results are described in Section VI. To increase the dynamic range of D, the currents I1 and I2 can
be implemented as the combination of a PTAT current IPTAT
II. BASIC DESIGN and a CTAT current ICTAT [11], such that
For reasons of simplicity, small chip size, low-energy con- I1 = 3IPTAT − 0.5ICTAT
sumption, and compatibility, the basic operation of the CMOS
temperature sensor (see Fig. 1) is chosen to be the same as that I2 = ICTAT − IPTAT . (2)
of previous designs [1], [11]. Under the control of a Schmitt As in [11], the sum Iref of the charging and discharging cur-
trigger (ST), a capacitor C is periodically charged by a current rents, i.e., 2IPTAT + 0.5ICTAT , was designed to have a slightly
I1 up to a threshold voltage V2 and then discharged by a current positive temperature coefficient, which effectively compensates
I2 down to a threshold voltage V1 [see Fig. 1(a)]. As can be the curvature in VBE . As shown in Fig. 2, this scheme ensures
deduced from the timing diagram shown in Fig. 1(b), the duty that D now varies from about 10% to 90% over the desired
cycle D of the resulting output signal equals temperature range –45 to 130 °C.
(V 2 −V 1 )C Fig. 3 shows a simplified block diagram of the actual CMOS
tH I2 I1 sensor. Substrate bipolar p-n-p transistors Q1 and Q2 are biased
D= = = . (1)
tL + tH (V 2 −V 1 )C
+ (V 2 −V 1 )C I1 + I2 at a 1:9 current-density ratio, and an op-amp (OP1 ) forces the
I1 I2

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
 1574 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 2, FEBRUARY 2017

resulting voltage ΔVBE = (kT /q)ln(9) across a resistor

RPTAT to generate a PTAT current IPTAT = ΔVBE /RPTAT
(∼1 μA at room temperature). Similarly, OP2 and resistor
RBE convert the base–emitter voltage VBE3 of Q3 into a
CTAT current ICTAT = VBE3 /RBE (when both switches
SBE1 and SBE2 are ON). Next, these currents are linearly
combined such that the capacitor C is charged by a current
I1 = 3IPTAT − 0.5ICTAT (S1 is ON, one of the SBE1 and SBE2
is ON) and is discharged by a current I2 = ICTAT −IPTAT (S1
is OFF, both SBE1 and SBE2 are ON).
By properly choosing the values of RPTAT and RBE , as
well as the nominal value of VBE3 , a linear duty cycle versus
temperature characteristic can be realized
Fig. 4. Circuit modifications for reaching a higher accuracy.
D = a0 + a1 ϑ (3)
where ϑ is the temperature in degree Celsius.
For compatibility with the previous design [1], the sensor (see Fig. 4), which are used to interchange Q1 and Q2 , Kelvin
was designed such that a0 = 0.32 and a1 = 0.0047 (◦ C)−1 . connections are used to accurately sense ΔVBE [14], [26].
In BICMOS technology, a straightforward implementation of The DEM and chopping-state machines are self-clocked (by
the Fig. 3 circuit resulted in good accuracy. This was because the output of the ST), and so no external clock is required. The
the current mirrors and other precision circuits could be imple- ratio RBE /RPTAT is made accurate by using large devices and
mented with well-matched bipolar transistors, while MOSFETs careful layout.
were only used as switches and digital logic. In CMOS tech- As shown in Fig. 4, seven identical PTAT current sources are
nology, however, a straightforward implantation would result used to bias Q1 and Q2 , as well as to charge and discharge
in very poor accuracy, because of the much large mismatch of the capacitor C. Therefore, a DEM cycle with at least seven
MOSFETs. Fortunately, by applying chopping and DEM, this states is required. Furthermore, four identical p-n-p transistors
problem can be solved and, as will be shown in the next sections, Q1 and Q2 were used to realize an emitter ratio of 3. A com-
the resulting accuracy exceeds that of the previous designs in plete DEM cycle of these transistors, thus, requires four steps.
bipolar or BICMOS technology. Finally, the op-amps need to be chopped and the nominally
identical RBE resistors need to be swapped. To satisfy all these
III. DESIGN FOR ACCURACY requirements, an extra dummy current source was added to the
In the circuit shown in Fig. 3, the main sources of systematic circuit (not shown in Fig. 4), and a DEM cycle with eight states
error are as follows: was chosen. During each DEM cycle, the following actions take
1) component mismatch, which induces output spread; place:
2) process spread in the base–emitter voltage VBE3 ; 1) The current sources are rotated once.
3) limited current gain β of the substrate p-n-ps; 2) The four BJTs are rotated twice.
4) nonlinearity of VBE versus temperature. 3) The RBE resistors are swapped four times.
In this section, we will discuss ways to reduce the effects of 4) The op-amps are chopped four times.
these error sources one by one. The effects of the noise will be Rotating all four groups of components one by one, so
discussed in Section VI together with the experimental results. that each possible permutation would occur, would require
256 periods. This would be quite time and energy consuming,
A. Component Mismatches and the temperature of interest could vary quite significantly
during such a long procedure. As one of the main innovations
When designing precision sensors in CMOS technology, presented in this paper, all four groups of components are rotated
component mismatch is the main nonideality to be taken into concurrently, which significantly reduces the required number
account. In our design, this will, for instance, give rise to errors of DEM states. This means that not all possible permutations of
in the gain of the various current mirrors, thus causing the ra- component configurations are implemented. However, detailed
tios between the various charging and discharging currents to analysis shows that rotating all four groups of components si-
spread. Moreover, mismatch will cause offset voltages in the multaneously is enough to cancel all first-order mismatch errors,
op-amps OP1 and OP2 . These offset voltages are connected in leaving only the much smaller second-order errors.
series with our basic signals ΔVBE and VBE , and, thus, directly
reduce sensor accuracy. In order to prevent this, the effects of
B. Process Spread in VBE3
op-amp offset and 1/f noise are mitigated by the use of chop-
ping. In addition, errors in current-mirror gain, as well as in The second most important error source is the effect of pro-
the ratio of resistances and in the emitter areas of the substrate cess spread on the base–emitter voltage VBE3 [11] of Q3 . In
p-n-ps are mitigated by applying DEM. Since some ill-defined the selected CMOS technology, the maximum deviation in this
voltages will be dropped across the DEM switches in block SB2 voltage amounts to about ±15 mV, resulting in an unacceptable

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ACCURATE BJT-BASED CMOS TEMPERATURE SENSOR WITH DUTY-CYCLE-MODULATED OUTPUT 1575

temperature error of about ±4 K. This deviation can be cor-

rected by trimming both the bias current and the emitter area of
Q3 . When resistor spread is also taken into account, the required
peak-to-peak trimming range is about 10 K. To correct for this,
a trimming scheme with a worst case (largest) trimming step of
about 50 mK (see Section IV) has been implemented. Fig. 5. Output signal of the temperature sensor.

C. Limited Current Gain

respectively. The large voltage swing range (about VCC − 2 V)
Another source of error and spread is the finite current gain at the input of the ST ensures that its input-referred noise has
β of the substrate p-n-ps (about 25 at room temperature). To negligible impact on the duty cycle. By using a large capacitor
reduce this error, an extra current source (see Fig. 4), which is (C ∼ 150 pF), the modulator’s oscillation frequency is designed
not included in the DEM scheme, and a substrate p-n-p (Q4 ) are to be low enough (less than 7 kHz) to ensure that the error caused
used to implement a simple beta-compensation scheme. by the ST’s own switching time (a few nanoseconds) is less than
10 mK.
D. Nonlinearity of VBE
In order to partially compensate for the nonlinearity of VBE D. Calibration
versus temperature, the current sum (I1 + I2 ) has been designed
An 8-bit trimming network, consisting of a switchable array
to be slightly proportional to temperature [11], [26].
of p-n-ps, was used to adjust the base–emitter voltage of Q3
(in Fig. 4) and to compensate for process spread. The base–
IV. CIRCUIT DESIGN
emitter voltage can be trimmed, starting from its minimum
In this section, key aspects of the circuit-level implementation value, with increments in the range of 0–40 mV, with a worst
of the CMOS temperature sensor are presented. case (largest) step size of about 50 mK. After calibration, the trim
code is stored by zapping Zener diodes that form a reliable and
A. Current Mirrors and Current Sources low-cost on-chip memory.
To generate the PTAT currents (see Fig. 4), a wide-swing
cascoded current mirror [27] has been used. Its large output V. SIGNAL AVERAGING
impedance ensures that IPTAT remains constant during the Due to the application of DEM and chopping in the circuit,
charging and discharging of the modulator’s timing capaci- the presence of component mismatch means that the duty cycle
tor C. This ensures that the sensor has a low supply-voltage of the sensor’s output will vary from period to period. However,
sensitivity. the output signal repeats every eight periods, which is the period
of a full DEM cycle. In order to achieve an accurate temperature
B. Op-Amps measurement, the sensor’s output must be properly processed,
The finite gain of the amplifiers OP1 and OP2 causes errors as will be discussed in this section.
in IPTAT and ICTAT , respectively. In order to keep the resulting Fig. 5 shows the output signal of the temperature sensor over
temperature-sensor errors below, for instance, 50 mK, the gains a full DEM cycle. A microcontroller can measure the time inter-
of OP1 and OP2 must be larger than 90 and 70 dB, respectively. vals tL1 , tH1 ; tL2 , tH2 , etc., relative to its own clock frequency.
Moreover, the amplifiers must be able to handle input voltages As discussed in Section III-A, systematic errors due to compo-
(VBE ) down to about 0.3 V at 130 °C. Both requirements are met nent mismatch and offset are compensated by averaging them
by implementing OP1 and OP2 as folded-cascode amplifiers over eight successive periods of the duty-cycle modulator. This
with PMOS input pairs [28]. signal-processing step is performed by the user, who must, there-
fore, be aware that the use of incomplete DEM cycles will result
C. Schmitt Trigger in a significant loss of accuracy. However, there are various
ways in which the sensor’s output can be averaged, as will be
The ST is based on the use of two inverters in series, with discussed below.
a positive feedback path that controls the threshold voltages of
the first inverter [29]. In order to make the swing range as large A. First Type of Averaging
as possible, the threshold voltages V1 and V2 [see Fig. 1(b)] are
chosen as close as possible to the corresponding rail voltages, A first type of averaging involves computing the duty cycle
which are of each period and then averaging the results. This yields the
average value Davg1 as
V1 = VTH N
8

V2 = VCC − |VTH P | (4) Davg1 = (tHi /(tHi + tLi ))/8 (5)
i=1
where VTH N and VTH P are the threshold voltages of the in-
put invertor of an NMOS transistor and a PMOS transistor, where i is the order of the period in one DEM cycle (see Fig. 5).

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
 1576 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 2, FEBRUARY 2017

Fig. 6. Simulated residual errors obtained with D av g 1 and D av g 2 for

an offset voltage V o s1 = 1 mV of OP 1 .

Fig. 7. Chip photo and packages of the temperature sensor.

B. Second Type of Averaging
A simpler method is to first sum the “High” and “Low” time
5) Next, the duty cycle Davg3 is calculated with an equation
intervals and then compute the average Davg2 with a single
similar to (6)
division as follows:
8 8 8

 
Davg2 = tHi / (tHi + tLi ). (6) Davg3 = (NHi /(NHi + NLi ))/8. (7)
i=1 i=1 i=1

This second type of averaging is equivalent to using an ana- Note that for M DEM cycles, only eight divisions are re-
log low-pass filter to convert the sensor’s duty-cycle-modulated quired. So, as compared to using Davg1 , this approach reduces
output into a dc voltage, which can then be read out by, for the calculation time by roughly a factor M, while its accuracy is
instance, a multimeter. It can also be used in very simple ana- as good as when using Davg1 .
log temperature-control systems (see [6, Ch. 7]). However, this
type of averaging does not completely cancel the mismatch- VI. FABRICATION AND TEST RESULTS
induced errors, because in the calculation of (6), each period has The temperature sensor is fabricated in standard 0.7-μm
a different weight. The longer periods will have larger weights CMOS technology of ON Semiconductor. The die size is
than the shorter ones and, thus, will contribute more error to 1.7 mm × 1.3 mm (see Fig. 7). In total, the chip has 13 pads.
the final “averaged” result. Fig. 6 shows the simulated resid- Nine of them are used to store the trimming code determined
ual temperature errors caused by 1-mV offset in OP1 (see Fig. by wafer-level calibration at room temperature. The other four
4) for the two averages Davg1 and Davg2 , respectively. Here, pads are available to the user and are VCC , GND, OUT, and PD
both residual errors have been normalized to 0 at 27 °C. Note (an optional pad for POWER DOWN). The sensors have been
that the average Davg2 results in much more error, especially packaged in TO18, TO92, TO220, SOT223, and SOIC-8.
at low temperatures. However, for the limited range of −10 to The sensor has been trimmed at wafer level to counteract
+110 °C, the error is still less than 0.1 °C, which is acceptable the effects of process spread. This is significantly less expen-
in many applications. sive than trimming individually packaged devices, but relies on
the sensor being insensitive to packaging shift. Without trim-
C. Third Method of Averaging ming the sensor exhibits about ±5 K error, which is not
Better temperature-sensing resolution can be obtained by av- acceptable.
eraging the sensor’s output over more than one DEM cycle,
because this will reduce the noise bandwidth (see Section VI). A. Test Setup
If this is desired, a third type of averaging can be used to re-
To characterize the duty cycle versus temperature accurately,
duce the number of divisions required, while still obtaining
the sensor’s output is compared with that of a reference sensor.
high accuracy. This involves calculating the duty cycle Davg3
In our case, a Pt100 platinum resistor, with an inaccuracy of less
as follows:
than ±20 mK over the full temperature range of −45 to 130 °C,
1) Suppose that the numbers NL1 , NH1 , NL2 , NH2 , . . . ,
was used as the reference sensor. To ensure that the sensor tem-
NL8 , NH8 represent the 16 time intervals tL1 , tH1 , . . . ,
peratures were as close as possible to that of the reference sensor,
tL8 , tH8 of the sensor’s output over one DEM cycle.
a special setup was built, as shown in Fig. 8 for TO18 and TO92
2) For the first eight periods, the values of NL1 to NH8 are
packaged samples. An ARM processor (STM32F103CBT6)
stored in separate registers.
with a 72-MHz counter was used to digitize the time intervals
3) For the ninth period (= the first period of the second
of the sensor output and to calculate the average duty cycle.
DEM cycle), the number NL9 is added to NL1 , while
the number NH9 is added to NH1 . In a similar manner,
B. DEM and Averaging
this is done for the other 14 time intervals of the second
DEM cycle. The sensor’s output is a rail-to-rail square-wave voltage. The
4) Step 3 is repeated for all other DEM cycles. frequency varies from about 500 Hz to 7 kHz, depending on the

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ACCURATE BJT-BASED CMOS TEMPERATURE SENSOR WITH DUTY-CYCLE-MODULATED OUTPUT 1577

periods in one of the ways described in Section V. Finally,

the temperature is calculated using (3).
To show the importance of averaging over complete groups
of eight periods, Fig. 9 shows a typical real-time measurement
result for each period (dashed line) without any averaging, and
the moving average, using (5) over eight periods (solid line).
Note that with the dashed line, the results repeat every eight
readings, corresponding to the eight states of one DEM cycle,
Fig. 8. Measurement setup for the temperature characterization of the where errors induced by component mismatch are still present.
sensors (in TO18 and TO92 packages). The temperature error of each single period varies from −2.9
to 3.2 K, and these values vary from sample to sample, de-
pending on the specific mismatch. In contrast, averaging over
eight periods reduces mismatch-induced errors significantly to
−0.018 K.
It can be shown that the measurement can be started at an
arbitrary transient (upgoing or downgoing) in a DEM cycle.
So, no synchronization is required because any series of eight
periods will cover a full DEM cycle. The measurement results
discussed in the following section are all based on using the
Fig. 9. Measured temperature reading at room temperature (T re f =
25.356 ◦ C). average over eight periods or integer numbers of eight periods.

C. Accuracy Versus Temperature

To characterize device spread, 70 calibrated samples in
(metal) TO-18 packages were tested over the temperature range
from −45 to 130 °C. As an example, Fig. 10(a) shows the mea-
sured temperature errors of 70 samples from two batches after
computing the average duty cycle according to (5) or (7) (Davg1
or Davg3 ). This figure shows clearly the remaining systematic
nonlinearity, which is mainly due to incomplete curvature cor-
rection and to the exponential increase of leakage currents at
high temperatures.
Fig. 10(b) shows the measured total error for the case when,
with the same measurement data, the average duty cycle is cal-
culated with the simpler (6) (Davg2 ). In agreement with Fig. 6,
there is significantly more error at low temperatures. However,
for the temperature range from 0 to 90 °C, this error is still less
than 0.2 °C. Note that this error is larger than the simulated error
Fig. 10. Systematic error versus temperature for 70 samples from two
batches when the duty cycle is calculated (a) with (5) or (7) (D av g 1 or
shown in Fig. 6 and varies from sample to sample, depending on
D av g 3 , V C C = 5 V) and (b) with (6) (D av g 2 ). the specific amounts and combinations of offset and component
mismatch that are present.
The measurement results depicted in Fig. 10(a) show that
the spread between the samples is very small. Therefore, the
systematic nonlinearity can further be reduced by fitting the
sensor’s residual nonlinearity with a higher order polynomial.
After a least squares fit on the measurement results shown in
Fig. 10(a), the relationship, between the average duty cycle D
and temperature ϑ, is found to satisfy the following third-order
polynomial:
Fig. 11. Same data as depicted in Fig. 10(a), but when using a
third-order correction of the results, according to (8). D = a0 + a1 ϑ + a2 ϑ2 + a3 ϑ3 (8)

supply voltage and temperature, although the exact variation is where a0 = 0.32; a1 = 4.68 × 10−3 /(◦ C); a2 = 7.03 ×
subject to process spread. Only the duty cycle contains accurate 10−8 /(◦ C)2 ; a3 = 1.10 × 10−9 /(◦ C)3 and ϑ = temperature
temperature information. in °C.
As explained in Section III, DEM and chopping have been Fig. 11 shows that the residual inaccuracy after computing
applied to achieve an accurate result. To benefit from this, the Davg1 and applying (8) is less than ±0.2 °C from −45 to
sensor’s duty cycle should be averaged over eight successive 130 °C.

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
 1578 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 2, FEBRUARY 2017

(standard deviation). For a supply voltage of 3.3 V, a supply

current of 60 μA, and a measurement time tm of 1.8 ms (eight
periods), the energy E for one measurement is only 356 nJ. The
sensor’s resolution FoM is 3.2 pJK2 , which is much smaller than
that of other products in the market (see Table II). Even including
recent research results, this FoM represents the state-of-the-art
for BJT-based temperature sensors [10], [30].
In sensor systems, where a slower microprocessor is used
Fig. 12. Measured resolution (standard deviation) versus to digitize the time intervals, the quantization noise due to
measurement time (72-MHz sampling frequency). the limited sampling speed should also be taken into account
(see [30]).

E. DC Supply-Voltage Sensitivity
Fig. 13 shows the change of the sensor’s error versus the dc
supply voltage for three temperatures, referred to the errors at
VCC = 5 V. Over the whole temperature range, the output varies
by less than 0.1 °C over the supply-voltage range from 2.5 to
5.5 V.
Fig. 13. Variation of sensor output with change of dc supply voltage, F. Packaging and Packaging Shift
relative to its output with a 5-V supply.
TABLE I To meet different market requirements, the new tempera-
MEASURED PACKAGING SHIFT FOR THREE TYPES OF PACKAGES COMPARED ture sensor has been packaged in various types of packages
WITH THE SHIFT IN THE PRODUCTS DESCRIBED IN [1] (see Fig. 7). Due to differences in the thermal expansion co-
efficients of the various materials involved (silicon die, die
TO-18 TO-921 TO-2201 TO-18 TO-921 attachment, and metal substrate), some mechanical stress re-
this work this work this work [1] [1] mains after the high-temperature packaging process. Plastic
Number of sensors 36 18 9 18 18 packages induce much more mechanical stress than metal-
Mean error (°C) 0.004 0.044 0.053 −0.026 0.38 can packages (i.e., TO-18) because they use an epoxy resin
Spread (3σ ) (°C) 0.06 0.13 0.06 0.31 0.35 that completely covers the chip [24], [25]. Due to the piezo-
1
The plastic packages employ a stress-relieving die coating.
junction effect, this packaging-induced stress will change the
base–emitter voltage VBE3 (see Fig. 4) and, thus, induce ex-
tra error in the sensor’s output. As shown in [23] and [25],
vertical p-n-ps are much less stress sensitive than vertical n-p-
ns. So, the sensor presented here should exhibit less packag-
ing shift and better stability than the one described in [11].
Packaging shift, as due to mechanical stress, has been in-
vestigated for the five different packages. The average val-
ues of this shift at room temperature for TO-18, TO-92,
and TO-220 are listed in Table I, together with values for a
Fig. 14 Variation of temperature output over 367 days for eight samples previous BiCMOS design in TO-18 and TO-92 packages [1].
at 22 °C. Test results for other plastic packages (SOT223 and SOIC) are
similar to those for TO92 and TO220 in Table I. These results
D. Noise show that the room-temperature error induced by the metal-
can package (TO-18) is almost negligible. The plastic packages
At a stable temperature of about 25 °C, the sensor’s noise
induce a positive shift, which is much smaller than that of a
was measured by logging the results of 360 000 measurements,
previous product [1]. These results demonstrate the remarkable
where each measurement is based on averaging over eight peri-
improvement that can be achieved when vertical p-n-ps rather
ods. A microcontroller with a 72-MHz sampling frequency was
than n-p-ns are used as sensing elements.
used to digitize the time intervals. As shown in Fig. 12, for the
minimum measurement time tm of 1.8 ms (eight periods), the
resolution is about 3 mK (rms). The sensor’s energy efficiency G. Stability
can be benchmarked with the help of the resolution figure of Long-term stability tests have been performed by a certi-
merit (FoM) F, which is defined as follows [30]: fied qualification company (Tempcontrol I.E.P. B.V.). Eight sen-
sors (never powered-on after wafer calibration) were inserted in
F = E · s2 (9)
a metal tube, filled with thermal conductive compound. This
where E is the energy consumed during one complete mea- tube was put in a water bath whose temperature was regu-
surement (one DEM cycle) and s is the sensor’s resolution lated at 22 °C with an inaccuracy <0.5 mK. Over 367 days,

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: ACCURATE BJT-BASED CMOS TEMPERATURE SENSOR WITH DUTY-CYCLE-MODULATED OUTPUT 1579

TABLE II
COMPARISON WITH OTHER STATE-OF-THE-ART TEMPERATURE SENSORS

This design SMT160 [1] D18B20 [2] LMT01 [3] TMP107 [4] ADT7320 [5]

Supply voltage (V) 2.7 to 5.5 4.75 to 7.2 3.0 to 5.5 2.0 to 5.5 1.7 to 5.5 2.7 to 5.5
Supply current (μA) 42 to 75 180 1000 to 1500 (at 5 V) 34 to 125 200 to 400 210 to 300
Number of pins 3 or 4 3 3 2 8 12
Temperature range (°C) −45 to 130 −45 to 130 −55 to 125 −50 to 150 −55 to 125 −40 to 125
Output signal DCM DCM Digital Digital Digital Digital
Measurement time (ms) 1 to 22 0.25 to 2 94 to 750 100 12 to 18 240
Supply voltage Sensitivity(°C /V) 0.1 0.1 0.1 0.13 0.1 0.1
Best accuracy (°C) 0.11 (−20 to 60) 0.7 (−10 to 100) 0.5 (−10 to 85) 0.5 (−20 to 90) 0.4 (−20 to 70) 0.5 (−10 to 105)
1
(Temperature range (°C)) 0.3 (−45 to 130) 1.2 (−45 to 130) 2 (−55 to 125) 0.7 (−50 to 150) 0.7 (−55 to 125) 0.66 (−40 to 125)
Resolution (°C) 0.003 0.005 0.0625 0.0625 0.0156 0.0078
measurement time (ms) 1.8 20 750 100 15 240
Resolution FoM (pJK 2 ) 3.2 430 1.46 × 10 7 4.38 × 10 4 2.42 × 10 3 8.28 × 10 3

1
The accuracy in these cells is the result using (8) and D a v g 1 or D a v g 3 .

as shown in Fig. 14, the sensor’s output drift was found to be (3.2 pJK2 ). This makes this sensor highly suited for low-energy
less than ±6 mK. applications. After wafer calibration at room temperature, the
sensor’s accuracy is better than 0.1 °C (−20 to 60 °C) and 0.3 °C
H. Performance Summary and Comparison With Other
(−45 to 130 °C), respectively. Package-induced errors were
Products
found to be less than 0.1 °C at room temperature. Measure-
Table II summarizes the main features of this design com- ments over 367 days show a long-term drift less than 6 mK.
pared to existing products with similar accuracy or with a similar
number of pins. A full list of specifications of the final product ACKNOWLEDGMENT
can be found in [31]. It can be seen that it achieves better ac-
curacy, resolution, and resolution FoM, as well as higher speed. The authors would like to thank J. van Wensveen of Temp-
One of the main reasons for its excellent resolution FoM is that control Industrial Electronic Products B.V., The Netherlands,
it outputs a quasi-analog signal whose time intervals are then for performing the long-term stability test.
digitized by a microcontroller. Thanks to this feature, the sen-
sor’s energy consumption, and, hence, any self-heating effects REFERENCES
are extremely low.
[1] Smartec BV. Datasheet SMT 160-30 Digital Temperature Sensor. (1992).
[Online]. Available: http://www.smartec-sensors.com
VII. CONCLUSION [2] Maxim Integrated. Datasheet of D18B20. (2015). [Online]. Available:
http://datasheets.maximintegrated.com/en/ds/DS18B20.pdf
The details of a BJT-based temperature sensor designed and [3] Texas Instruments. Datasheet of LMT01. (2015). [Online]. Available:
fabricated in a 0.7-μm CMOS technology were discussed. The http://www.ti.com/lit/ds/symlink/lmt01.pdf
[4] Texas Instruments. Datasheet of TMP107. (2015). [Online]. Available:
use of a duty-cycle-modulated output signal enables easy in- http://www.ti.com/lit/ds/symlink/tmp107.pdf
terfacing with both digital and analog systems. Its high output [5] Analog Devices. Datasheet of ADT7320. (2012). [Online]. Available:
data rate makes this sensor well suited to both temperature http:// www.analog.com/media/en/technical-documentation/data-sheets/
ADT7320.pdf
monitoring and controlling. A high accuracy is achieved by [6] G. C. M. Meijer Ed., Smart Sensor Systems, ch. 7. New York, NY, USA:
applying DEM and chopping, which significantly reduces the Wiley, 2008.
errors caused by the CMOS component mismatching. Simul- [7] V. C. Gungo and G. P. Hancke, “Industrial wireless sensor networks:
challenges, design principles and technical approaches,” IEEE Trans. Ind.
taneously rotating all the component groups during the DEM Electron., vol. 56, no. 14, pp. 4258–4265, Oct. 2009.
process, instead of rotating them one after one, reduces the [8] P. Li, Y. Wen, Z. Zhang, and S. Pan, “A high-efficiency management cir-
number of DEM periods significantly, making the measurement cuit using multi winding up conversion current transformer for power-
line energy harvesting,” IEEE Trans. Ind. Electron., vol. 62, no. 10,
speed much faster. By applying DEM, the mismatch-induced pp. 6327–6335, Oct. 2015.
errors from the various groups of components can be reduced to [9] J. P. Carmo, L. M. Gonçalves, and J. H. Correia, “Thermoelectric micro
the second order. To achieve this result, the results of a complete converter for energy harvesting systems,” IEEE Trans. Ind. Electron.,
vol. 57, no. 3, pp. 861–867, Mar. 2010.
DEM cycle must be averaged. Three different types of averag- [10] A. Heidari, G. Wang, K. A. A. Makinwa, and G. C. M. Meijer, “A BJT-
ing were discussed. They provided different tradeoffs between based CMOS temperature sensor with a 3.6 pJK2 resolution FOM,” in Dig.
computational speed/complexity and sensing accuracy. Apply- Tech. Papers IEEE Int. Solid-State Circuits Conf., Feb. 2014, pp. 224–225.
[11] G. C. M. Meijer, R. van Gelder, V. Nooder, J. van Drecht, and
ing substrate p-n-ps instead of n-p-ns reduced errors caused by H. M. M. Kerkvliet, “A three-terminal integrated temperature transducer
package-induced stress. As a result, the presented sensor has a with microcomputer interfacing,” Sens. Actuators, vol. 18, pp. 195–206,
much better long-term stability than earlier designs. It also facil- Jun. 1989.
[12] M. A. P. Pertijs, K. A. A. Makinwa, and J. H. Huijsing, “A CMOS tem-
itates the use of low-cost packaging technologies, which com- perature sensor with a 3σ inaccuracy of ±0.1°C from −55 °C to 125 °C,”
pared to metal-can packaging, typically induce large amounts IEEE J. Solid-State Circuits, vol. 40, no. 12, pp. 2805–2815, Dec. 2005.
of mechanical stress. [13] K. Souri, Y. Chae, and K. A. A. Makinwa, “A CMOS temperature sen-
sor with a voltage calibrated inaccuracy of ±0.15°C (3σ) from −55 °C
Measurement results showed that this temperature sensor to 125 °C,” IEEE J. Solid-State Circuits, vol. 48, no. 1, pp. 292–301,
achieves a state-of-the-art resolution FoM, which is better than Jan. 2013.
Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.
 1580 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 64, NO. 2, FEBRUARY 2017

[14] S. Shalmany, D. Draxelmayr, and K. A. A. Makinwa, “A micro power Ali Heidari was born in Langroud, Iran, in 1971.
battery current sensor with ±0.03% (3σ) inaccuracy from −40 °C to He received the B.S. degree from Iran Univer-
+85 °C,” in Dig. Tech. Papers IEEE Int. Solid-State Circuits Conf., Feb. sity of Science and Technology, Tehran, Iran, in
2013, pp. 386–387. 1992, the M.S. degree in electrical engineering
[15] A. L. Aita, M. A. P. Pertijs, K. A. A. Makinwa, J. H. Huijsing, and from Tehran University, Tehran, in 1995, and the
G. C. M. Meijer, “Low-power CMOS smart temperature sensor with a Ph.D. degree in electrical engineering from Delft
batch-calibrated inaccuracy of ±0.25 °C (±3σ) from −70 °C to 130 °C,” University of Technology (TU Delft), Delft, The
IEEE Sensors J., vol. 13, no. 5, pp. 1840–1848, May 2013. Netherlands, in 2010.
[16] X. Pu et al., “A ±0.4°C accurate high-speed remote junction temperature He is a Member of the Academic Staff in
sensor with digital Beta correction and series-resistance cancellation in 65 the Department of Electrical Engineering, Guilan
nm CMOS,” in Proc. Symp. VLSI Circuits, Jun. 2013, pp. 214–215. University, Rast, Iran. He is also a Guest Re-
[17] K. Souri, Y. Chae, F. Thus, and K. A. A. Makinwa, “A 0.85 V 600 nW searcher with TU Delft, where he is working on integrated temperature
All-CMOS temperature sensor with an inaccuracy of ±0.4°C (3σ) from sensors in the CMOS technology. His research focuses on analog- and
−40 °C to 125 °C,” in Dig. Tech. Papers IEEE Int. Solid-State Circuits mixed-signal integrated circuit design.
Conf., Feb. 2014, pp. 222–223.
[18] C. Azcona, B. Calvo, N. Medrano, and S. Celma, “1.2V-0.18 μm CMOS
temperature sensor with quasi-digital output for portable systems,” IEEE
Trans. Instrum. Meas., vol. 64, no. 9, pp. 2565–2573, Sep. 2015.
[19] R. Quan, U. Sonmez, F. Sebastiano, and K. A. A. Makinwa, “A 4600 μm2
1.5 °C (3σ) 0.9 kS/s thermal-diffusivity temperature sensor with VCO- Kofi A. A. Makinwa (M’97–SM’05–F’11) re-
based readout,” in Dig. Tech. Papers IEEE Int. Solid-State Circuits Conf., ceived the B.Sc. and M.Sc. degrees from
Feb. 2015, pp. 488–489. Obafemi Awolowo University, Ife, Nigeria, in
[20] S. Huber, A. Laville, and C. Schott, “A bridge-type resistive temperature 1985 and 1988, respectively, the M.E.E. degree
sensor in CMOS technology with low stress sensitivity,” in Proc. IEEE from the Philips International Institute, Eind-
Sensors Conf., Nov. 2014, pp. 1455–1458. hoven, The Netherlands, in 1989, and the Ph.D.
[21] P. Chen, C. Chen, Y. Peng, K. Wang, and Y. Wang, “A time-domain degree from Delft University of Technology, Delft,
SAR smart temperature sensor with curvature compensation and a 3σ The Netherlands, in 2004.
inaccuracy of 0.4 °C∼ +0.6 °C over a 0 °C to 90 °C range,” IEEE J. From 1989 to 1999, he was a Research Sci-
Solid-State Circuits, vol. 45, no. 3, pp. 600–609, Mar. 2010. entist with Philips Research Laboratories, Eind-
[22] P. Park, D. Ruffieux, and K. A. A. Makinwa, “A thermistor-based temper- hoven, where he worked on interactive displays
ature sensor for a real-time clock with ±2ppm frequency stability,” IEEE and on front ends for optical and magnetic recording systems. He is cur-
J. Solid-State Circuits, vol. 50, no. 7, pp. 1571–1580, Jul. 2015. rently an Antoni van Leeuwenhoek Professor with the Faculty of Elec-
[23] G. C. M. Meijer, G. Wang, and F. Fruett, “Temperature sensors and voltage trical Engineering, Computer Science and Mathematics, Delft University
references implemented in CMOS technology,” IEEE Sensors J., vol. 1, of Technology, which he joined in 1999. His main research interests in-
no. 3, pp. 225–234, Oct. 2001. clude design of precision mixed-signal circuits, sigma–delta modulators,
[24] J. F. Creemer, F. Fruett, G. C. M. Meijer, and P. J. French, “The smart sensors, and sensor interfaces. This has resulted in ten books, 25
piezo-junction effect in silicon sensors and circuits and its relation to patents, and more than 200 technical papers.
piezo-resistance,” IEEE Sensors J., vol. 1, no. 2, pp. 98–108, Aug. 2001. Prof. Makinwa is a member of the Program Committees of the Inter-
[25] F. Fruett and G. C. M. Meijer, The Piezo-Junction Effect in Silicon Inte- national Solid-State Circuits Conference (ISSCC), the VLSI Symposium,
grated Circuits and Sensors. Boston, MA, USA: Kluwer, 2002. the European Solid-State Circuits Conference (ESSCIRC), and the Ad-
[26] M. A. P. Pertijs and J. Huijsing, Precision Temperature Sensors in CMOS vances in Analog Circuit Design Workshop. He has also served as a
Technology. New York, NY, USA: Springer, 2006, pp. 267–268. Guest Editor of the IEEE JOURNAL OF SOLID-STATE CIRCUITS (JSSC) and
[27] D. A. Neamen, Electronic Circuit Analysis and Design. Homewood, IL, as a Distinguished Lecturer of the IEEE Solid-State Circuits Society. For
USA: Irwin, 1996, pp. 606–607. his doctoral research, he received the 2005 Simon Stevin Gezel Award
[28] M. C. Schneider and C. Galup-Montoro, CMOS Analog Design Using All- from the Dutch Technology Foundation. He is a co-recipient of several
Region MOSFET Modeling. Cambridge, U.K.: Cambridge Univ. Press, best paper awards, from the JSSC, ISSCC, Transducers, and ESSCIRC,
2010, pp. 306–320. among others. He is an Alumnus of the Young Academy of the Royal
[29] I. M. Filanovsky and H. Baltes, “CMOS Schmitt trigger design,” IEEE Netherlands Academy of Arts and Sciences and an elected member of
Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 41, no. 1, pp. 46–49, the IEEE Solid-State Circuits Society AdCom, the society’s governing
Jan. 1994. board.
[30] K. A. A. Makinwa, “Smart temperature sensors in standard CMOS,” in
Proc. Euro-Sens. Conf., Sep. 2010, pp. 930–939.
[31] Smartec BV. Datasheet SMT172 Digital Temperature Sensor. (2015).
[Online]. Available: http://www.smartec-sensors.com
Gerard C. M. Meijer (M’94–SM’98) received
the M.Sc. and Ph.D. degrees in electrical engi-
neering from Delft University of Technology (TU
Delft), Delft, The Netherlands, in 1972 and 1982,
respectively.
Since 1972, he has been a Member of the
Research and Teaching Staff, Delft University
Guijie Wang was born in LuoYang, China. She of Technology, where he is currently a Profes-
received the B.S. degree in physics and the M.S. sor, involved in research and teaching on ana-
degree in electronic engineering from NanKai log electronics and electronic instrumentation.
University, Tianjin, China, in 1984 and 1987, re- Since 1984, he has been a Consultant to in-
spectively, and the Ph.D. degree in electronic dustrial companies and research institutes. In 1996, he cofounded the
engineering from Delft University of Technology, company SensArt, where he is a Consultant in the field of sensor sys-
Delft, The Netherlands, in 2005. tems. In addition to many journal and conference papers, he is also an
She is a Research and Design Engineer with author and editor of books in the field of sensor systems, published by
Smartec BV, Breda, The Netherlands. Her main IOP, Kluwer, Springer, and Wiley.
focus is on BJT-based CMOS temperature sen- Prof. Meijer is a Member of the IEEE Industrial Electronics Society
sors; the recent result has been successively and a Senior Member of the IEEE Solid-State Circuits Society. In 1999,
commercialized as an important product (Smartec, SMT172). She is the Dutch Technology Foundation STW awarded him with the honorary
also busy with technical support for customers. She is currently working degree “Simon Stevin Meester,” and in 2001, he was awarded the Antoni
on the BJT-based temperature sensor in a 0.18-μm CMOS technology. van Leeuwenhoek Chair at TU Delft.

Authorized licensed use limited to: Walailak University provided by UniNet. Downloaded on November 09,2023 at 01:58:55 UTC from IEEE Xplore. Restrictions apply.