Integrated Smart Sensors

INTEGRATED SMART SENSORS
INTEGRATED SMART SENSORS

DESIGN AND CALIBRATION
by
Gert van der Horn

University of Delft
and
Johan L. Huijsing
University of Delft
SPRINGER-SCIENCE+BUSINESS MEDIA, B.V.

A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN 978-1-4419-5016-1 ISBN 978-1-4757-2890-3 (eBook)

DOI 10.1007/978-1-4757-2890-3
Printed on acid-free paper
AII Rights Reserved

© 1998 Springer Science+Business Media Dordrecht
Originally published by Kluwer Academic Publishers in 1998
Softcover reprint of the hardcover 1st edition 1998
No part of the material protected by this copyright notice may be reproduced or
utilized in any fonn or by any meanS, electronic or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written pennission from the copyright owner.
Table of Contents
1 lntroduction
1.1 Introduction ..............................................................
1
1
1.2 Sensors and actuators ............................................... 2
1.3 Integrated smart sensors ........................................... 5
1.4 Measurement errors and correction .......................... 6
l.4.1 Process variations and time variations .......................... 7
1.4.2 Calibration ..................................................................... 8
1.4.3 Se1f-testing ................................................................... Il
l.4.4 Se1f-calibration ............................................................ 13
l.4.5 Redundancy improvement Of feedforward testing ...... 15
l.4.6 ConcJuding remarks ..................................................... 16
1.5 Objective and Organization ................................... 16
1.5.1 Objective ...................................................................... 16
1.5.2 Organization ................................................................ 17
1.5.3 Design constraints ....................................................... 17
References ........................................................................... 19
2 lntegrated Smart Sensor Concept 21

2.1 Introduction............................................................ 21
2.2 Silicon sensors..... ................................................... 22
2.2.1 Temperature sensor ..................................................... 22
2.2.2 Pressure sensor ............................................................ 23
2.2.3 Measurement errors ..................................................... 25
INTEGRATED SMART SENSORS: Design and Calibration v

Table of Contents
2.3 Analog Interface circuits ......................................... 27

2.3.1 Temperature sensor readout ........................................ 27
2.3.2 Pressure sensor readout ............................................... 32
2.4 Analog-to-Digital conversion ................................. 33
2.4.1 Sigma-delta converter ................................................. 34
2.5 Digital bus or microcontroller interface ................. 39
2.5.1 Microcontroller compatible interfaces ........................ 39
2.5.2 Smart Sensor bus interfaces ........................................ 40
2.6 Integrating a calibration function in the smart sensor
concept ................................................................................. 43
2.6.1 Error sources in the smart sensor signal path .............. 43
2.6.2 Conventional sensor calibration techniques ................ 44
2.6.3 Correction of the sensor signal transfer ...................... 45
2.6.4 Calibration memory .................................................... 46
References ........................................................................... 47
3 Calibration and Iinearimtion

Techniques 51
3.1 Introduction ............................................................. 51
3.1.1 Errors in the sensor transfer curve .............................. 51
3.1.2 Systematic and random errors ..................................... 53
3.1.3 Linear and nonlinear errors ......................................... 55
3.1.4 Normalized sensor transfer functions .......................... 56
3.2 Linearization ........................................................... 58
3.2.1 Sensor characteristic linearization ............................... 59
3.2.2 Linearization based on a look-up table ....................... 60
3.2.3 Linearization based on piecewise-linear interpolation 63
3.2.4 Linearization based on piecewise-polynomial or spline
interpolation ................................................................ 67
3.2.5 Linearization based on curve fitting ............................ 70
3.2.6 Linearization based on error minimization ................. 73
3.2.7 Conc1usion ................................................................... 75
VI INTEGRATED SMART SENSORS: Design and Calibration

Table of Contents
3.3 Progressive polynomial calibration method ........... 76

3.3.1 One-dimensional functions .......................................... 77
3.3.2 Two-dimensional functions ......................................... 84
3.3.3 Multi-dimensional functions ....................................... 88
3.4 Conclusion ............................................................. 89
References ........................................................................... 90
4 Signal
Calibration using Analog
Processing 93
4.1 Introduction ............................................................ 93
4.2 Conventional sensor calibration ............................. 94
4.2.1 Resistor trimming ........................................................ 94
4.2.2 Programmable resistor-arrays ...................................... 96
4.3 Analog calibration circuits ..................................... 96
4.3.1 Classical offset and full-scale calibration .................... 98
4.3.2 Classical cross-sensitivity calibration .......................... 99
4.3.3 Improved analog calibration circuit... ........................ 102
4.4 Classical pulse-modulated calibration .................. 104
4.5 Analog implementation of a polynomial
calibration . . . . . . . . . . . . . . . . . . . . . . . .. 106
4.5.1 System configuration ................................................. 106
4.5.2 Analog current multiplier .......................................... 108
4.5.3 Programmable current divider DAC ......................... 111
4.5.4 VI-converters and IV-converters ............................... 113
4.5.5 System implementation ............................................. 116
4.5.6 System operation ....................................................... 117
4.5.7 Realization ................................................................. 121
4.6 Conclusion ............................................................ 124
References ................ 126
o .... o .......... o .... o o .. o .. o ....................... o ....
INTEGRATED SMART SENSORS: Design and Calibration Vll

Table of Contents
5 Calibration using Sigma-Delta

Analog-to-Digital Conversion 127
5.1 Introduction ........................................................... 127
5.2 Sigma-delta AD-converters .................................. 128
5.2.1 Principle of oversampling ......................................... 129
5.2.2 Principle of noise shaping ............. ........... ............. .... 130
5.2.3 Principle of decimation ............................................. 136
5.2.4 Transfer from analog input to digital output ............. 139
5.3 Alteration of the sigma-delta transfer ................... 140
5.3.1 Digital sigma-delta bitstream generator .................... 140
5.3.2 Modulating the input signal for gain calibration ....... 143
5.3.3 Modulating an additional signal for offset
calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.3.4 Sigma-delta calibration configurations ..................... 148
5.4 Implementation of the sigma-delta AD-converter 152
5.4.1 Differential integrator amplifier ................................ 153
5.4.2 Clocked comparator .................................................. 155
5.4.3 Sigma-delta converter ............................................... 157
5.4.4 Digital decimation filter implementation .................. 159
5.5 Smart temperature sensor realizations .................. 160
5.5.1 Smart temperature sensor with calibration and IS 2 bus in-
terface ............................................... ......................... 161
5.5.2 Temperature sensor with linearity calibration ........... 165
5.6 Conclusion ............................................................ 168
References... .... ........ ...... ..... ....... ...... ..... ....... ........... ........... 169
viii INTEGRATED SMART SENSORS: Design and Calibration

Table of Contents
6 Calibration using Digital

Signal Processing 173
6.1 Introduction ........................................................... 173
6.2 Hardware implementation ..................................... 174
6.3 Software implementation ...................................... 178
6.3.1 Sensor system configuration ...................................... 178
6.3.2 Calibration algorithm ................................................. 179
6.3.3 Calibration program ................................................... 189
6.3.4 Future work ............................................................... 191
6.4 Conclusion ............................................................ 191
References ......................................................................... 192
7 Summary and Conclusions 193
Index 199
INTEGRATED SMART SENSORS: Design and Calibration IX

Introduction
1
1.1 Introduction
The (signal processing and storage) capacity ofthe human brain enables us
to become powerful autonomous beings, but only if our brains operate in
conjunction with (at least some of) our senses and muscles. Using these
organs, we can interact with our environment, learn to adapt, and improve
important aspects of our life. Similarly, the signal processing capabilities
of modern electronics (computers) could be combined with electronic
sensors and actuators to enable interaction with, and adaptation to, the
(non-electrical) environment. This willlead to smarter and more powerful
automated tools and machines. To facilitate and stimulate such a
development, easy-to-use low-cost sensors are needed. The combination
of electronic interface functions and a sensor in an integrated smart
sensor, that provides a standard, digital, and bus-compatible output, would
simplify the connection of sensors to standard electronic signal processors
(microcontrollers, computers, etc.). Currently, the calibration procedure,
required for standardization of the sensor output signal level, contributes
largely to the production costs of accurate sensors. To enable automation
of the calibration procedure, and hence reduce the sensor fabrication costs,
a digital calibrationjunction should be included in the smart sensor.
INTEGRATED SMART SENSORS: Design and Calibration

G. van der Horn et al., Integrated Smart Sensors
© Springer Science+Business Media Dordrecht 1998
Introduction
1.2 Sensors and actuators

In industry many processes are electronically controlled. As depicted in
Fig. 1-1, to be in control of a process one needs to be able to gather ali
necessary information about that process through sensors, then to examine
that information and maybe modify it using a controller, and take the
proper actions through actuators to inftuence the process according to the
desired algorithm [1] . The results of these actions can be measured, which
enables a feedback control [2] .
Digital
Electronic
Controller
Fig. 1-1 Gathering the process information, and controlling the

process, using sensors and actuators.
To gather process information we measure process parameters, such as

pressure and temperature, using sensors. To control the pressure and
temperature of a process we use actuators, such as valves and heaters. Due
to the evolution in microelectronics of the past decades we now have
powerful control and processing tools available which operate in the
electrical signal domain . Therefore, we need sensors which convert signals
from any of the five other signal domains ; mechanical, thermal, chemical,
magnetic, and optical to an electrical signal which can be proces sed
electronically and passed on to an advanced digital computer [7]. For the
control action, we need actuators which can be electrically driven to effect
changes in one or more of the five other signal domains .
2 INTEGRATED SMART SENSORS : Design and Calibration

1.2 Sensors and actuators
Based on a large diversity of underlying physical principles, many sensors

and actuators have been and are being developed . Each of these devices
has its own unique properties and requires its own special electronic
interface. Therefore, the realization of a system as depicted in Fig. 1-1 has
become the terrain of many different technical specialists, thus limiting the
possibilities from a system-Ievel point of view.
From a system-Ievel point of view, a configuration as shown in Fig. 1-2 is

much more practical: sensors and actuators that are able to communicate
with a digital controller using standardized signals on a low-Ievel
communication bus. The digital controller is able to communicate signals
on a high-Ievel bus to a user interface, for example.
low-Ievel bus
communication
Digital
Electronic
high-Ievel bus Controller
communication
User Interface
Fig. 1-2 Bus structure for sensor/actuator communication.
Bus architectures and protocols that have been successfully introduced on

a higher system level (FIELD, PCI, etc.) c\early show the many advantages
with respect to star routing. Besides reduced wiring, advantages of a
low-Ievel bus set-up in data-acquisition systems are also found in the
exchangeability of devices, in multiplexing options, and in easy
maintenance and expansion [3,4,5]. Furthermore, it simplifies division of
a complex system with high-Ievel bus communication, into sub-systems
using low-Ievel bus communication; Fig. 1-3 shows how the system can be
divided into a data-acquisition sub-system to interface the sensors, and
another sub-system for controlling the actuators, each having it's own
INTEGRATED SMART SENSORS: Design and Calibration 3

Introduction
low-Ievel bus and digital controller. It is also possible that some devices
communicate on a high-Ievel [6] and others on a low-Ievel bus [1].
To enable the proposed configurations, many interfacing functions will

have to be integrated with the sensor or actuator device. Such smart
sensors and smart actuators have been only sparsely available up to now.
Whereas most actuators, micro-actuators excluded, are large and power
consuming, the addition of smart electronic circuits to generate driving
signals will enhance performance, for a relatively small increase in price
and energy. Sensors are generally micro-devices which are often
manufactured using special processing. At first sight, the investment
needed to include or add electronic functions with a sensor device seems
relatively large.
low-Ievel bus
commun ication
data-acquisition
sub-system
,-_..L..L_---..
Digital Digital
Electronic Electronic
Controller Controller
high-Ievel bus ~~~~~~~~~~~

communication ~ User Interface
Fig. 1-3 System division into two low-level sub-systems.
A focus on the realization of sensors in standard Integrated Circuit (lC)

technology Of IC-compatible technology will not only help to reduce
senSOf prices in mass production, but will also make it easier to integrate
the required functions in electronic circuits with a sensor device [8,9, IO] .
In the next paragraph we will focus on the functions that should be
integrated on a smart sensor, in order to simplify setting up configurations
as proposed in Fig. 1-2 and Fig. 1-3.
4 INTEGRATED SMART SENSORS: Design and Calibration

1.3 Integrated smart sensors
1.3 Integrated smart sensors

In order to realize a smart sensor which is able to communicate on a
standard bus level as previously proposed, the following functions must be
incJuded in the sensor package [8]:
• sensor(s), preferably silicon sensor(s)

• analog sensor readout and signal conditioning
• analog-to-digital (AD) conversion, to provide a digital
output signal
• bus interface, ta simplify communication to
microcontrollers, PCs, and other devices
• calibration ofthe sensor transfer curve, preferably a
digitally programmable calibration
First of ali, the sensor converts the physical input signal into an electrical
signal. Second, the usually weak analog sensor signal needs to be
amplified and converted into convenient electrical signals (voltage or
current, rather than capacitance, resistance, etc.). Third, an AD-converter
is necessary to allow communication with the digital world of
microcontrollers and computers, often a serial 1-bit output is preferred
here. To simplify setting up systems with multiple sensors and actuators,
each device will also need a bus interface on board to transport the sensor
signal to a digital microcontroller or computer, which further processes the
sensor signal into a digital multi-bit (e.g. bytes) format. Fig. 1-4 shows the
signal processing chain of a smart sensor containing these functions, and
cJarifies the signal conversion realized in each block.
sianal domains:
physical: electrical:
analog; digital: 1-bit serial
input
silic6n analog A-tb-D bus micro-
sens,or interface conv,erter interface controller
sus
Integrated Smart Sensor
Fig. 1-4 Signal processing chain of an integrated smart sensor.

Introduction
Last but not least, on-board calibration electronics will make it possible to
realize smart sensors with standardized output signals, allowing easy
exchange of sensors and easy re-calibration [8, Il]. Calibration is
necessary to 'pinpoint' the transfer from the physical input signal,
temperature for example, to the electrical output signal, so that the
manufacturer can guarantee that a temperature of Odegrees Celsius always
corresponds to a digital output signal O. In the manufacturing of the
sensors this transfer often shows a large device-to-device variation.
Different options to correct this will be discussed in the next paragraph.
1.4 Measurement errors and correction

Before studying measurement errors in sensor systems and ways to correct
such errors, it is important to contemplate what we want to measure and
what we intend to do with the inforrnation on the measurand. It is not
always necessary to relate the sensor signal to a well-defined standard unit
(SI I or SI-derived unit). In a trained (expert) or self-Iearning (neural)
system, a sensor signal is judged subjectively, based on experience, to take
the proper actions; our human bodies are not equipped to measure the
temperature outside in degrees Fahrenheit, Celsius or in Kelvin, yet with
our senses we are perfectly capable to make a healthy decision on whether
to go outside in shorts or to put on a coat. The scientists named have taught
us that we can relate our human temperature sensing to standardized
numbers, which we can communicate in weather reports for example. This
brings us to the advantages of relating the sensor signal to the measurand
in a standardized unit, such as Kelvin for temperature:
• interpretation; we can easily learn how to interpret numeral

information and, moreover, we can let computers do it
• comparison; using numbers we can objectively compare
temperatures at different times and places
• exchangeability of measurement equipment and sensors in
instrumentation and control systems
We have to conc\ude that a sensor signal can only be usefully proces sed in
measurement or control systems, exc\uding neural systems, when it can be
related to a well-defined standard unit. In some control systems the target
1 SI=Systeme International d'unites, Of International System of units.

is to regulate an error signal to zero. In that case it is sufficient to know

only the offset, while non-zero signals do not have to be known in precise
units. For example, when placing a robot arm on top of an object it can be
sufficient to know when the measured distance between the arm and object
has become zero.
The process of relating the measurement or sensor signal to the physical

input signal in precise well-defined units is referred to as calibration.
1.4.1 Process variations and tirne variations

Calibration would be simple if aII sensor devices in a production line had
the same properties. In that case, the relation between electrical output
signal and physical input signal should be traced (measured) for only one
sensor, and then it can be applied to aII sensors of that type. Unfortunately,
many variations in the production process must be tolerated and, therefore,
also the properties of the sensors will vary from device to device.
Particularly in IC-technology we know to expect large absolute variations
(50%) and smaller relative (on-chip) variations (1 %). For example,
micromachined pressure sensors based on piezoresistors show large
variations in offset (due to resistor mismatch, initial stress, etc.), in
sensitivity (position/orientation of resistors, mechanical properties of the
diaphragm, etc.), and also in cross-sensitivity to temperature. An error
band in the output signal range as large as 30% of the full-scale can be
expected [11]. Each sensor device will have to be calibrated individually
in order to guarantee that it exhibits a well-defined and fixed transfer from
input to output. This usually means that each device must be measured
individually, in a controlled environment, using reference signal
conditions.
This individual attention per de vice is very expensive in mass production,

which explains the large price difference between uncalibrated mass-
produced sensors and calibrated measurement equipment. To reduce the
costs of the sensor device calibration, we should look for ways to automate
the calibration of many sensors at a time, in a so-called batch calibration.
The smart sensor configuration proposed in the previous section makes
this possible: many such smart sensors can be put (simuItaneously) in a
controlled environment (such as a temperature oven or a pressure
chamber), via a bus connection each sensor output can be measured
quickly, and the integrated calibration can be programmed to obtain
precisely the desired transfer. Furthermore, the number of necessary

Introduction
reference measurements must be minimized, in order to save time and thus

fabrication costs.
Another problem which can be encountered in the transfer of a sensor is

that it may alter slowly over time; some sensor characteristics depend on
material properties which degrade or "wear off'. Although these changes
are usually small, over a long period (months or years) they may cause a
significant error, known as drift. This drift error has to be reset at a certain
interval by a re-calibration. For this reason, it often makes sense to
integrate at the sensor some kind of actuator, which can generate a
well-defined, additional input signal for the sensor; this introduces the
possibility of self-test and self-calibration. The terms calibration, self-test,
and self-calibration in the context of sensor technologyl will be further
c\arified in the succeeding paragraphs.
1.4.2 Calibration
To define calibration in general we quote Nicholas & White [12]:
"The word 'calibration' has changed its meaning with time and three
distinct meanings can be distinguished ....
The word appears to derive from an Arab word for a mould for casting
metal. This is presumably because calibration originally referred to the
making of guns by casting metal. Calibration referred either to the means
used to determine the calibre (or bore) of the gun Of to the determination of
the range of the gun ....
The second and more common meaning of the word calibration is the
checking and adjusting of the bore of an instrument and the application of a
scale by the manufacturer. . .. That is, a calibration refers to the set of
operations carried out by an instrument manufacturer in order to ensure that
the equipment has a useful measurement scale .... "
It is difficult to objectively define the terms scale, ensuring, and

usefulness, so a more official definition of the term calibration is
unavoidable:
I Unfortunately the terms self-test and auto-calibration are also popular in electronic
circuit technology, especially for sensor interface circuits. The terms could be used
with much more care, by specifymg exactly what is tested or automatlcally adJusted.

"The third meaning for calibration, as used by metrologists, is defined by

the ISO (International Organization for Standardization) as:
CALIBRATION, The set of operations which establish, under special
conditions, the relationship between values indicated by a measuring
instrument or measuring system, or values, represented by a material
measure, and the corresponding known value of a measurand.
NOTES, (1) The result of a calibration permits the estimation of errors of
indication of the measuring instrument ... or the assignment of values to
marks on arbitrary scales. (2) A calibration may also determine other
metrological properties. (3) The result of a calibration may be recorded in a
document, sometimes called a calibration certificate or a calibration report.
(4) The result of a calibration is sometimes expressed as a calibrationfactor,
or as a series of calibration factors in the form of a calibration curve.
This definition is best seen as a pointer to a meaning of calibration
acceptable to modern quality assurance and accreditation practices. It
represents the current practice and if emphasis is put on the NOTES the
definition gives rise to a valuable concept of calibration .... "
The official side of measurement standards and calibration is the concern
of national and international standardization institutes and will not be an
issue in this book.
In practice, a sensor calibration is performed by putting the sensor in a

controlled environment, as indicated by phase J in Fig. 1-5. The
environment is controlled in such a way that physical reference signals can
be generated in accordance with the official standards. A temperature
sensor can be inserted in an oven, in which a heater system can create a
sensor ambience with a temperature of 300, 350 or 400 Kelvin, for
example. Usually an extra (officially calibrated) reference sensor or
instrument is used to guarantee these temperatures. Sometimes, more
(reference) sensors are needed to monitor other ambient parameters, such
as humidity. The measurements of the sensor output signal at several
temperatures can be used to determine the error with respect to the desired
sensor transfer. Depending on the sensor characteristics and the number of
reference measurements the sensor transfer and error curve can be
determined within a certain accuracy. This is not sufficient; the calibration
needs to be completed by adjusting or correcting the sensor transfer with
trimmable (adjustable) elements in such a way that an accurately specified
transfer is obtained, as illustrated in Fig. 1-6.
At a set of input signals [xJ"'xn ], which can be calculated by using the

known transfer of the reference sensor, the sensor outputs [y J... Yn] are
measured. By using a correction table or a correction function, the sensor

Introduction
1. Calibration Phase
, ~ - " environment '-...,.....---r-T"""'"

I ,.....---., -(> (senSOlr(S)]~_ _ _ _~\~in~'orm~a~l;on~_.J
reference signal
correcled
sensor
signal
2. Measurement Phase
I
I
transfer
physicaJ correction calibrated
signai to b t--t----.,~ • • • output
measured ---~ unc!JUbrated signal
- - -s;n-;;" ;n~i~o~;e~t - - sensor signal
Fig. 1-5 Schematic explanation of sensor calibration.
outputs are processed in order to retrieve a signal which relates to the

original physical signal in accordance with the desired transfer curve.
After calibration the correction should be maintained so that the calibrated

sensor can be used to reliably measure the physical signal in an
uncontrolled environment, as indicated by phase 2 in Fig. 1-5. The manner
in which the correction curve is obtained and implemented can be referred
to as the calibration method. A traditional calibration method for pressure
sensors is to insert trimmable resistors (potentiometers or laser-trimmed)
in the Wheatstone bridge and manually adjust them, see also section 4.2. 1.
This book will discuss several possibilities of implementing a more
attractive digitally-programmable transfer correction in the circuitry of a
smart sensor interface.

1.4 Measuremeut errors aud correctiou
reference sensor ca/ibrated output

signa/ slgna/
~",<
<;'1l' calibrated
"'o - + ... output
0"'4' I signa/s
_4o_J. ...
ofC-;--L--+__ physica/ I
I .1 mput slgna/ I ti ca/cu/ated
t
I I
, • f physlca/
-J
sensor Xl I X2 IXI! "1 I I slgna/s
slgna/ I I I -g I I I
I o' c"
-y!: ... -
I '"'''' I -
Y2
-
I
-
I
~ ~ ~
,<>
I ~0<:-~ -
- - - - - ... - r - C:,,0(\.r::P CI) o
_________- Yl 0
,,0''',,~0 "<>
c.Q.~
,(\ G ~CD
7f------- physica/ Q.
input signa/
Fig. 1-6 Determining the sensor transfer and calibration curve.
To summarize, the essential elements of calibration are listed as:
• a calibration phase in a controlled environment

• generation of physical reference signals for sensor
excitation, and measurement of the sensor output signal
• correction of the signal transfer based on those calibration
measurements
Officially, the reference signals used in the calibration measurements
should be traceable 1 to an SI-unit.
When referring to a 'sensor with calibration' we mean that the possibility

of correcting the sensor transfer is integrated with the sensor or sensor
interface.
1.4.3 Self-testing
By installing one or more actuators c10se to the sensor, see Fig. 1-7, it is
possible to influence the physical input signal of the sensor at the
command of an on-chip control circuit. This controller then verifies
whether the sensor has detected the change in signal, if not it generates an
error signal to indicate that the sensor (or actuator) is not functional. This
I Traceability is also c1early defined In Nicholas &White [12]
INTEGRATED SMART SENSORS: Design and Calibration II

Introduction
can avoid incorrect interpretation of the sensor signal inside the system in
which the sensor is used.
sensor
I output signal
self-test
error signal
actuator(s)
multiplicative
I test signal smart sensor I
\ ... _________ ~t~ !e~ '.!~ .. I
sensor environment
Fig. 1-7 Schematic explanation of a sensor with self-test.
For some types of sensors we can realize an additive, for others a

multiplicative test signal. On a temperature sensor, for example, we can
integrate a heater element (i.e. aresistor) which, when turned on, increases
the chip temperature with respect to the ambient. An additive test signal of
a few degrees can be realized . It is important to realize that the self-test
signal has to be detectable under ali conditions, since the sensor
environment is not controlled as in the case of calibration. During the
self-test the sensor output signal should be ignored.
Although from the operation point of view we might have improved

reliability, from the production point of view we have added an extra
element which can fail, namely the actuator. This may affect the yield in
the production and the mean-time-before-failure of the product. It is thus
important to verify that the reliability of the actuator is not much worse
than that of the sensor. For many types of sensors this will not be the case,
or an appropriate actuator which can be integrated with the sensor will not
be available at ali.
Summarizing the essential elements of a sensor with a self-test function:
• an integrated actuator which

• allows a test of the junctionality
• realizes an improved reliability
• operates in an uncontrolled normal sensor environment

1.4.4 Self-calibration
As can be seen in Fig. 1-8, the self-calibration I function is very similar to
the self-test function, but needs the addition of accurate actuator(s) and a
possibility of correcting the transfer, based on sensor measurements which
can be obtained by using different actuator signals.
" -PhY;i~/- - - - - - - - - - - - - - - - '\
L
, inputsignal 1-+..' _ _ calibrated
output signaJ
X
accurate
actuator(s)
precise
multiplication
factor smart sensor with
auto-calibrat ion I
,-------------------~
sensor environment
Fig. 1-8 Schematic explanation of a sensor with selj-calibration.
It is essential for a self-calibration function that the actuators can be used

to improve the properties of the sensor, which means that the transfer of
the actuator has to be more accurate, in some respects, than the transfer of
the sensor. Few examples of sensor/actuator combinations for which this is
true can be found. One example is that of a magnetic Hal\-plate sensor
which may exhibit a relatively large drift in sensitivity and also a cross-
sensitivity to temperature [13]. As a magnetic actuator, a coil can be
realized around the sensor. With an accurate reference current and a
wel\-defined geometry, the coil produces an accurate magnetic signal
which is added to the magnetic input signal. The resulting change in sensor
output can be measured and should always be the same (for a constant
sensitivity). To correct the measured sensitivity errors, the current through
the Hal\ plate can be adjusted. It is also necessary to be able to separate the
self-calibration signal from the physical input signal using a muItiplexing
technique, either in time or frequency.
I Often the term auto-calibration is used, which means the same a~ self-calibration if
the word auto is interpreted as the Greek word for self, as in autodidact, but not as an
abbreviation for automatic.

Introduction
To self-calibrate for offset errors, a multiplicative actuation signal is

needed. The magnetic sensor, for example, could be mechanicaIIy turned
around 180 degrees, or electronicaIIy chopped by inverting the HaII-plate
current [14]. This should resuIt in an exactly opposite sensor output signal
while the offset stays the same. In fact, the signal can be muItiplied by plus
or minus one. The offset is found by taking the average of the two
(inverted and non-inverted) signals. By subtracting the two signals the
offset is eliminated.
Essentials for self-calibration, in addition to those for self-testing, are:
• accurate actuators
• transfer correction
• clear separation of input signal and calibration signal
Since it is very difficult, if not impossible, to realize on-board actuators
which generate independent reference signals traceable to SI-units,
officiaIIy the term self-calibration can be considered contradictory [12].
Self-calibration of sensors does not eliminate the need for 'real'
calibration. It should rather be seen as a method to monitor the parameters
of a sensor and automatically adjust them, so that the time between
(re-)calibrations can be extended. More precisely, we could speak of
auto-zero adjust and auto-gain or auto-scale adjust instead of
self-calibration.
It is useful to extend the potential of self-test and self-calibration by

considering tandem transducers. A tandem sensor obtains the desired
transfer from physical input signal indirectly via another physical signal
into the electrical domain, and vice versa for a tandem actuator [7]. In such
cases we can distinguish self-test or self-calibration options on different
levels. For example, in the case of a thermal flow sensor, which first
converts the flow signal to a thermal signal and then to an electrical signal,
it wiII be difficult to realize a self-test signal for flow, whereas thermaIIy,
using a heater, it can be feasible. In the case of a capacitive accelerometer
it is possible to use a tandem actuation for generating a self-test signal: by
applying a charge on the sensor mass and generating an electrical field, an
electrostatic force and thus an acceleration signal result [15,16].

1.4 Measurement errors am! correction
1.4.5 Redundancy improvement or feedforward testing

In the self-test and self-calibration principles explained above, an actuator
was used to enable feedback to the sensor input signal. In the feedback
loop the sensor and actuator operate in series, which means that the
ensemble can only work properly if both the sensor and the actuator are
operation al. Hence the smart sensor can test itself. If the actuator and
sensor have exactly compensating errors this may remain undetected, but
this is very unlikely.
It is also possible to improve reliability by putting multiple sensors in

paraflel or by using extra sensors to monitor and feedforward the
operating conditions, as suggested in Fig. 1-9. By processing the signals of
, ~
,,--=-----, \
I physical
I input signal al
C
I in
<Il sensor
al
U output signal
o
ă.
I env;ronmenr error signal
condilions
smart sensor with
feedforward test
sensor environment
Fig. 1-9 Schematic explanation of feedforward self-test using

multiple sensors.
multiple sensors in a certain way, single sensor defects can be detected and
failure of the measurement can be prevented . Defects or "out-of-range"
indications can be indicated with an error signal. By using the increased
redundancy provided in the combination of the sensor signals, a more
reliable measurement of the physical input signal is realized . The
reliability is best improved when combining dissimilar sensors with
different failure mechanisms.
It is important to realize that the feedforward principle does not make it

possible to test the operation of the sensor but only to verify the validity of
the output signal.

Introduction
1.4.6 Concluding remarks

For many sensor types self-test or self-calibration is not (yet) feasible
because the essential actuator does not exist or does not fulfil the
requirements. Unless an actuator with a standardized reference signal can
be integrated, self-calibration cannot eliminate the need for calibration l , as
described in section 1.4.2. Only in neural (learning) or fuzzy controIled
systems the relation of the sensor signal to a precisely defined unit is not so
important. In aII other cases, sensor calibration is indispensable. Whenever
the sensor output signal must relate precisely to an agreed standard unit for
the measured physical signal, it is necessary to correct the sensor transfer
based on calibration measurements. This is unavoidable if we want to
make sensors a 'plug-and-play' technology.
1.5 Objective and Organization
1.5.1 Objective
The objective of the research discussed in this book is the incorporation of
a calibration function in the smart sensor concept. The integration of such
a calibration function with the sensor device, next to other smart sensor
functions such as AD-conversion and bus interface, is an attempt to
develop accurate, economicaIly feasible sensors. It is important that the
calibration function can be programmed digitally because this will allow
automation of the manufacturer's calibration procedure which, in
conjunction with batch calibration, can minimize the calibration costs per
sensor. Once reliable smart sensors are available at low cost they will not
only be applied in industry, but they may also enhance consumer products
and improve our personal comfort.
I Nevertheless, the development of self-test and self-calibration techmques should be

further stimulated, as It offers other advantages, especmlly wlth respect to reliability
Improvement of the sensor.

1.5 Objective and Organization
1.5.2 Organization
By recalling the smart sensor signal chain presented in Fig. 1-4 we can
c1arify the organization of this book:
In chapter 2 the 'state-of-the-art' smart sensor concept will be explained,

each of the Junction blocks - sensor, analog interface, AD-converter, and
bus interface - will be discussed briefly and illustrated by implementation
examples.
In chapter 3 the errors which may be encountered in the transfer function

of a sensor will be explained. Different approaches to linearize the sensor
transfer function by means of calibration will be explained from a
theoretical point of view. A special polynomial calibration method will be
presented.
The practical possibilities to correct the overall signal transfer of a smart

sensor for calibration purposes, can roughly be divided in three categories,
according to the three function blocks in which the correction can be
implemented. These categories are issued in the subsequent chapters.
In chapter 4 different options to change the transfer of analog circuits in a

digitally programmable way are examined. The one-dimensional
polynomial calibration method presented in chapter 3 is worked out in the
realization of a programmable analog calibration circuit.
In chapter 5 digital alteration of the transfer of the sigma-delta

AD-converter by use of bitstream signals is discussed. This involves
mixed-mode circuits. It is illustrated by the realization of a digitally
programmable smart temperature sensor.
In chapter 6 the proposed polynomial calibration method is extended to

the two-dimensional case, considering pressure sensors which have a
cross-sensitivity to temperature. The calibration algorithm is implemented
in a simple microcontroller program, which realizes the calibration
correction in the digital signal domain.
1.5.3 Design constraints

Considering the complications of designing modern smart sensor interface
circuits (certainly as single-designer projects), it was decided to make use
of only one IC process, and hence reduce the effort required to set up the

Introdnction
CAD tools for the IC design f1ow. The 3-Ghz bipolar process 'Dimes-Ol'
of the Delft Institute for Microelectronics and Submicron-technology,
DIMES, was available at low cost, and was therefore an obvious choice for
this research project.
Although the choice for the Dimes-Ol process may not be optimal from
the economical point of view for aII the encountered designs, a consistent
use of the same process throughout the research project was favored.
Furthermore, the many silicon sensor developments in the Dimes-Ol
process can now easily profit from the advances in the sensor interface
designs in the same process.

1.5 Objectives and Organization
References
[1] FR Riedijk, lntegrated Smart Sensors with Digital Bus lnterface,
PhD Thesis, Delft University Press, Delft, 1993.
[2] R.C. Dorf, Modern Control Systems, Fourth Edition, Addison-
Wesley Publishing Company, Reading, Massachusetts, 1986.
[3] W.A. Koeleman and P.P.L. Regtien, "Bus-organized
data-acquisition systems", Sensors and Actuators, Vol.5, 1984,
pp.327-333.
[4] J.E. Brignell, "Sensors in distributed instrumentation systems,
Sensors and Actuators, Vol.lO, 1986, pp.239-248.
[5] N. Najafi and K.D. Wise, "An organization and interface for sensor-
driven semiconductor process control systems", IEEE Trans.
Semiconductor Manujacturing, Vol.3, No.4, November 1990,
pp.230-238.
[6] M. Rutka, lntegrated Sensor Bus, PhD Thesis, Delft University
Press, De1ft, 1994.
[7] S. Middelhoek and S.A. Audet, Silicon Sensors, Academic Press,
London, 1989.
[8] J.H. Huijsing, FR Riedijk, and G. v.d. Horn, "Developments in
integrated smart sensors", Sensors and Actuators A, Vol.43, 1994,
pp.276-288.
[9] RF Wolffenbuttel, edit., Silicon Sensors and Circuits: on-chip
compatibility, Chapman & Hali, London, 1996.
[10] RF Wolffenbuttel, "Fabrication compatibility of integrated silicon
smart physical sensors, Sensors and Actuators A, Vol.41-42, 1994,
pp.l1-28
[11] J. Bryzek, K. Petersen, J.R. Mallon, L. Christel, and F Pourahmadi,
Silicon Sensors and Microstructures, NovaSensor, Fremont, USA,
1990.
[12] JV Nicholas and D.R White, Traceable Temperatures: an
introduction to temperature measurement and calibration, John
Wiley &Sons, Chichester/NewYork, 1994.

Introdnction
[13] P.L.C. Simon, P.H.S. de Vries, and S. Middelhoek, "Autocalibration

of silicon Hali devices", Sensors and Actuators A, Vo1.52, 1996,
pp.203-207.
[14] M.J.A.M. van Putten, M.H.P.M. van Putten, and A.F.P. van Putten,
"FuII additive drift elimination in vector sensors using the alternating
direction method (ADM)", Sensors and Actuators A, Vo1.44, 1994,
pp.13-17.
[15] R.P. van Kampen, Bulk-Micromachined Capacitive Servo-
Accelerometer, PhD Thesis, Delft University Press, Delft, 1995.
[16] H.Y. Allen, S.T. Terry, and D.W. de Bruin, "Accelerometer systems
with bui1t-in testing", Sensors and Actuators A, VoI. 2 1-23, 1990,
pp.381-386.

Integrated Smart
Sensor Concept
2
2.1 Introduction
In this chapter we will present a concept for implementing smart sensor
interfaces [1]. Within the scope of this work we cannot discuss many types
of sensors or sensor interfaces, hence we limit ourselves to two common
sensor types, for measuring temperature, and for measuring pressure. This
limitation does not mean that the presented matter cannot be extended to
other types of sensors.
We will explain the basic principle of operation of a monolithic

temperature sensor, and a pressure sensor based on piezoresistors. We will
then show the circuitry which is necessary to read out the sensors and
derive signals suitable for AD-conversion. The preferred type of
AD-converter, a sigma-delta converter, will also be explained, as weB as a
possible serial bus interface.

Integrated Smart Sensor Concept
2.2 Silicon sensors

The temperature sensor and the pressure sensor are both IC-compatible
and thus very suitable examples in the context of smart sensors.
2.2.1 Temperature sensor

The difference in base-emitter voltage of two transistors with different
current densities exhibits a linear relation with respect to the absolute
temperature [2,3]. We take two equal transistors and bias them with
currents that are scaled by a factor of four, as illustrated in Fig.2-1.
1.0, ..
i Voltage
0.6:
+
0.2:
+ +
Temperature (·C) - -
-50 o 50 100 150
Fig. 2-1 Temperature sensor principle based on PTAT voltage.
We can easily derive from the formulas for the collector currents,
'c =, S . e
Ui V BEQ1 )
=4 x 'BIAS
QI QI
(2-1 )
that a (PTAT) voltage proportional to the absolute temperature T results:
!1VBE = VSE
QI
-VSE
Q2
= kT
q
,SQ2 J
_·ln (4.,-
S
(2-2)
QI

2.2 Silicon sensors
In which k represents Boltzmann's constant, q the electron charge and Is

the transistor saturation current. The case of equation (2-2), using a current
density ratio of four, and ISQ1=ISQ2, results in a small positive temperature
coefficient of approximately 0.12 mV per Kelvin. The base-emitter
voltage itself demonstrates a much larger but slightly less linear
temperature dependency which can be expres sed in the form of a Taylor
expansion of the VBE curve at a reference temperature TR :
= VBC T=O + (ll- m )· kT

-
q
R "1
+ I\,T=T.· T
R
(2-3)
1 ( _ m ). kTR.(T - TR)2
--ll ---
2 q TR
In this equation VBC represents the silicon bandgap voltage, Â the linear
temperature coefficient which measures approximately -1.9 mV/K at
TR =300K, and (TI-m) a physical parameter of about 3.5 which is somewhat
process dependent. The voltage signals VBE and VpTAT are displayed in
Fig. 2-1, where the PTAT voltage had to be magnified by a factor of 16 to
obtain a comparable temperature coefficient. Both effects can be used for
measuring temperature. In section 2.3.1 will be shown how to combine
both signals in a fully integrated circuit and to circumvent the need for an
extern al reference voltage.
2.2.2 Pressure sensor

A well-known principle for measuring pressure is by the use of a thin
membrane or diaphragm, because a diaphragm tends to deflect under
pressure. Micromachining techniques make it possible to fabricate thin
mechanical structures on a silicon wafer, even without harming the
integrated circuits processed on the same wafer [4,5,6,7,8,9]. A common
technique, known as bulk-micromachining and also accomplished at
D1MES, is to (KOH) etch the masked backside of the wafer (P-substrate)
and obtain an electrochemical etch stop by applying a voltage to the n-epi
or buried-n layer at the frontside. In that way the wafer thickness
(-500llm) can be locally reduced, under an angle of approximately 55', to
epi-thickness (-3-7Ilm), whereas the horizontal dimensions can be
controlled by the backside mask [10, Il]. In this way, a structure as
iIIustrated in the pictures of Fig. 2-2 can easily be fabricated.

Lx [ 110)
pressure
stress O stress
Fig. 2-2 Pressure sensor principle with Wheatstone bridge.
Now the art is in electrically sensing the pressure induced deftection of the
micromachined diaphragm. A capacitive technique seems attractive but
has the disadvantage that another capacitor plate is required, which has to
be fabricated/mounted as c10se as possible to the diaphragm for maximum
capacitance signal [12,13,14) . This is problematic considering the
packaging and the electrical connections, as well as incompatible with
some of the pressure sensor types indicated in Fig.2-3 because of the
required backside opening (7) . Solutions making use of wafer-to-wafer or
wafer-to-glass bonding are still under development [5, 11,15,15) .
An already welI-developed method makes use of silicon piezo-

resistors [7,16,17). As indicated in the cross-section and top view of the
sensor in Fig.2-2, the deftection of the diaphragm results in a shear stress
at the surface of the diaphragm. Provided that the silicon lattice is
correctly oriented, the piezoresistance effect of silicon can be used to sense
the change in stress [4,17,18) . The topview also gives an impression of the
(absolute) stress profile when under pressure; the darker areas have larger
stress, and the stress-sensitive piezoresistors are thus placed in those areas.
The orientation of the resistors is such that the stress results in an increase
of resistance for R 1 and R3 and a decrease for R2 and R4 . The silicon
resistors (shallow-p diffusion) can be fabricated using standard le

2.2 Silicon sensors
absolute differential sealed gage

illpul pres,~ure pre.'sure I illplll pre.uure
V V V
/ lIb.'wlure VlIClIum"'\..
illPIll pre,<slue pressure2
gage
V ambielll pre.,,~ure
vacuum
I \
I:I I
ambielll pres.mre
il1plll pres.mre
Fig , 2-3 Different types of pressure sensors.
processing (on the same wafer), and are connected in a Wheatstone bridge
configuration, as shown in Fig.2-2. A pressure-related output voltage
signal , V,IUI' proportional to the bridge supply voltage, Vre!' is thus
generated [7] .
2.2.3 Measurement errors

In the fabrication of sensors certain changes in process parameters and
also in material (wafer) properties wiI1 ha ve to be tolerated , and , ergo, a
variation in the properties of the sensors, even when produced in the same
process, must be expected. This usually results in a statistical (Gaussian)
distribution of the errors in the transfer of the sensors with respect to the
average sensor transfer. The manufacturer tries to get the average sensor
transfer as c10se as possible to the desired sensor function, and to
minimize the deviation of the transfer errors , GeneraI1y the spread in the
error is much larger than the desired accuracy, and selection or calibration
is required in order to obtain accurate sensors.
In case of the PTAT sensor the variations in the transfer are determined by
the matching properties of standard electronic components , This wiI1 be
examined in section 2.3,1 ,
INTEGRATED SMART SENSORS : Design and Calibration 25

In the framework of a master thesis' project on pressure sensor self-test

some bulk-micromachined pressure sensors have been deve\oped and
realized in the DIMES-Ol process. We will use those sensors as an example.
Several of these devices have been measured using DRUCK test equipment.
Fig.2-4 shows the transfer, relative bridge output voltage versus pressure,
of two sensors (SensorJ & Sensor2) with a diaphragm of 400 x 400 ~m2,
and two sensors (Sensor3 & Sensor4) with a larger diaphragm of
800 x 800 ~m2. Naturally, the dimensions of the diaphragm (size and
thickness) determine the pressure sensitivity and the pressure range in
which the sensors can be used.
·'CXXJ '000 2CXXJ 3000 4000 5OJO
~5 00
Pressur. (mbar) --..
5aJ
I
~
~'aJ 'aJ '000
>
oS ~5aJ
~
C>
g
o
>
~
C>
~
/
Pressure (mbar) ------.-
Fig. 2-4 Measured transfer of pressure sensors developed in

Dimes-Ol.

2.3 Analog Interface circuits
Despite the fact that the sensors come from the same wafer, a significant
difference in offset and sensitivity between two sensors is found, for both
cases. An even larger difference can be expected when also taking into
account the variation in temperature coefficients of offset and sensitivity.
The large offset errors in the shown examples are mainly due to
imperfections in the first resistor layout, which can easily be improved in a
next version. Nevertheless, error sources such as initial stress, geometrical
variations, doping concentration profiles, and resistor mismatch will result
in a large error variations. The sensor construction does not aIlow to put
resistors c10se together to improve matching. The source of the signal
errors also explains why there is not a typical sensor-characteristic
nonlinearity error but a large device-to-device variation. An overview of
error contributions in piezoresistive pressure sensors can be found
in [7:chapter 4] and [19]. A total error band of 30% of the fuIl-scale can
be a reasonable expectation.
Notwithstanding, one of the largest sensor markets is that for pressure

transducers. Thus a strong push from that market is expected for
stimulation of integrated calibration techniques. We have to conc1ude that
in order to realize production of a pressure sensors with standard transfer
from pressure range to output range, the calibration will have to
compensate large error variations.
2.3.1 Temperature sensor readout

To enable the read-out of a temperature signal, as presented in
section 2.2.1, by an AD-converter, another signal has to be generated
which can be used as a reference. Hence, we will first describe the
construction of a so-called bandgap reference. The temperature signal and
the reference signal must be combined in such a way that a ratiometric
signal is obtained which makes optimal use of the AD-converter's signal
range. This will be explained next.
By summing a VBE signal and a signal proportional to VPTAT the first-order

(linear) temperature coefficients can be cancelled, and a reference signal
results, which is referred to as bandgap reference. As can be seen from
equation (2-3) it relates direct1y to the bandgap voltage VBG , although a

small offset error and second-order error are present. Bandgap references
and PTAT current sources are well-known circuits in le design as they are
often used for biasing purposes.
The left half of the circuit displayed in Fig.2-5 shows a low-voltage PTAT
source [20], which makes use of the ratio-ed transistors Ql and Q2. These
transistors are inserted in a controlloop that establishes equal bias currents
for both transistors. Because of the ratio in current density the difference
between the base-emitter voltages is proportional to the absolute
temperature, as already shown in equation (2-2). The value of the bias
current is determined by the PTAT voltage at the emitter of Q2 and the
value of resistor R 1. The current /VPTAT inherits the proportional-to-
absolute-temperature property of the voitage. The current mirror on the
left-hand side only serves to generate a small start-up current to prevent a
zero bias solution at start-up.
~ O O
t..::
~
IVPTAT ,:--1VPTAT
IVbe
r--
H - 1VRe
*
I
r.:02 I
,~
01 == 03
1x ~ 4x ~+ RO
R) + I R2 V
be
6k V pTAT 96k I-
I
Fig. 2-5 Bandgap reference and PTAT current generator.
The right-hand si de of the circuit, around Q3, generates a current which is

determined by the base-emitter voltage of Q3 and the value of resistor R2 .
The current /Vbe thus inherits the Vbe temperature characteristic, including
the negative temperature coefficient.
The two current signals can easily be summed at the output, resulting in
/V8G=/Vbe+/PTAT. When inserting that current in a resistor R, it is converted
back into a voitage signal. The voltage will then be proportional to the
weighed sum of the PTAT voitage VPTAToRlR 1 and the base-emitter voltage
Vbe oRlR 2 , which can be equalized to the silicon bandgap voltage.

A simulation of the circuit, see Fig. 2-6, shows the PTAT signal with a
positive temperature coefficient, the VBE signal with a negative
temperature coefficient, and the constant sum signal. By correctly
dimensioning the ratio of the resistor values of R[ and R2 , in this case a
ratio of 16, the temperature coefficients of both current signals can be
matched so that the sum of the currents, referred to as [VBG' remains
constant as a function of temperature.
t 15!·················· ........................................................................................~
Current (JlA)
i !
:
1VBG- VBANDGAp/R
, ,
i Temperature ce) ---+-

o............................................................................................................... ,:
-50 o 50 100 150
Fig. 2-6 Current signals as function of temperature, using ideal
resistors.
The simulation shown in Fig.2-6, was done for ideal resistors, meaning
that the resistance was considered independent of temperature. In reality
the temperature coefficients of the resistors directly affects the curves of
the current signals. However, when using the same kind of resistors for R,
R [, & R2 , the bandgap voltage signal will not be affected because it
depends on the resistor ratios. Similarly, the effect of the temperature
coefficients will cancel if we use the ratio of two current signals.
Temperature sensor circuit

For use in a smart temperature sensor with AD-converter we like to realize
a circuit which generates two current signals of which the ratio varies
approximately from 10% to 90% for the intended temperature range.
Within the operating temperature range of le technology, the ratio of the
PTAT and the bandgap reference signal varies from 35% to 65%. Therefore,
we prefer not to use that ratio directly, but to modify the current signals.

This can be done by combining the PTAT and bandgap currents, for
example in the way shown in Fig.2-7. A fraction ofthe bandgap current is
copied and subtracted from the PTAT current to generate a new temperature
current signal f.r. A down-scaled copy of the bandgap current is used as a
new reference current signallREF'
Fig. 2-7 Temperature" sensor" circuit generating f.r and 1REF;

IVBG and IVPTAT come from the circuit of Fig. 2-5.
The simulation displayed in the upper graph of Fig.2-8, shows the two
currents as a function of temperature, now taking into account the
temperature coefficients of the resistors in the le process. It clearly shows
5 r-40'C 12S'C
..·........·····......··........····....·· ..···· ....······· .. ··········..............................
current
(IlA)
t o :..........................................................................................1...........
1 ! ......................................................................... ···· .... ······ ....··T.. ······ '!
0.9
current
ratlO
t ".....::::::..--------------+--i 0.1
o •...
-
u ~ ~
••••••••••••••••••••• . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
-50 o 50 100 150

Temperature ce)
Fig. 2-8 Ratio of f.r and 1REF used as measurement signal.

that the absolute current IREF is certainly not constant, and that IT is not so
linear. However, the ratiometric signaIIT/IREF, displayed in the lower half
of Fig.2-8, is good and is not affected by the temperature dependence of
the resistors. The sensor circuit is dimensioned in such a way that the
ratiometric signal range of [0.1-0.9] maps on the temperature range of
-40 ·e to + 125 ·e. The current mirrors in the sensor circuit of Fig. 2-7 can
be dimensioned differently to obtain a different temperature range, within
the operation limits of Ies of course.
To study the effect of process variations on the signal, a simulation has

been do ne for a worst-case relative mismatch of ± 0.5% and an absolute
variation of±25%. Fig. 2-9 shows the results on the ratiometric signal and
the error with respect to the desired linear curve.
:f32j-
t
7
......
j
. ,
1 0 ! b ·.. ··
error
(%)
t,:~_ ~
l
i
. _ ...
-10.. . . • . ... ... ... . ....

-50 o 50 100 150 -50 o 50 100 .. 156
Temperature CC) _ Temperature CC) _
Fig. 2-9 Effect on transfer and error curves resulting from

worst-case relative (1 %) and absolute (50%) variations.
First of alI, it demonstrates that the effect of relative errors is much larger.
Second, it can be conc1uded that, in fabrication of such temperature
sensors, a spread in offset and gain errors of several percents, say ± 5%,
should be expected. The linearity error will be relatively- smalI, in the order
of 1%, for the specified temperature range. To realize a silicon temperature
sensor with exact1y the desired transfer from temperature to ratiometric
output, these errors will have to be calibrated. Furthermore, it has to be
noticed that the nonlinearity curve has a shape that is characteristic to this
type of sensor. The shape doesn't vary much from device-to-device. It is
determined by the nonlinearity in the VBE curve as expressed in
equation (2-3). The temperature coefficient of the resistors is a second

source for nonlinearity but results only in a very small device-to-device

variation of the nonlinearity error.
2.3.2 Pressure sensor readout

The piezoresistive pressure sensors based on a Wheatstone bridge
configuration inherently make use of a ratiometric conversion; as
explained in section 2.2.2 a pressure signal will result in an output voltage
signal related to the bridge reference voltage. The AD-conversion which
follows the analog readout circuitry has to realize a conversion of this
ratiometric signal, by using both the bridge output voltage and the bridge
reference voltage. As seen in section 2.2.3, pressure sensors may display
an offset which is in the same order of magnitude as the full-scale signal
range. This offset can be eliminated by using a chopping technique. As an
example, Fig. 2-1 O gives an overview of a bridge interface circuit which
fulfils these requirements [21]. A combination of standard analog cells of
the DIMES-Ol library and custom cells has been used for the realization of
this circuit.
resistive
divider
dlfferent,al
reference
current
LLl AD-
converter
dlfferent,al
Input current
+
chopped vp
bridge
chop
Fig. 2-10 Pressure bridge readout circuit.
The chopping is implemented by using saturated bipolar transistor

switches which invert the polarity of the bridge reference voltage.
Changing the polarity of the reference voltage, inverts the polarity of the
bridge output voltage signal but not that of the bridge offset voltage. By
subtracting the inverted and non-inverted signal the bridge offset is
eliminated. This is preferably done at the end of the signal chain, as shown
in the figure, so that aII other offset contributions are removed as well.

2.4 Analog-to-Digital conversion
As seen in section 2.2.3, the ratiometric bridge signal usuaIly does not get
much larger than 1-2 %. To optimaIly use the dynamic range of the
AD-converter it is necessary to scale down the bridge reference signaI. In
Fig. 2-10 this is realized with, accurately scaled, and carefuIly laid-out
resistors. Both the down-scaled reference voltage, and the bridge output
voltage, are converted to currents, using VI-converters, and pas sed on to
the sigma-delta AD-converter for a ratiometric conversion.
For the pressure sensor as well as for the temperature sensor we have
shown how to obtain a current ratio proportional to the sensor signaI. The
same can be done for many types of sensors. The obtained current ratio
can be accurately converted into a digital (bitstream) signal, using the
continuous-time sigma-delta converter explained in the next section.

Much effort has been, and is being put in comparing resolution, speed, and
power consumption of several AD-conversion principles and
architectures, however it does not fit in the scope of this work to discuss aII
the investigated AD-converter types. Information can be found
in [22,23,24], for example. We wiIl contine to summarizing the
advantages of the preferred AD-converter type, the delta-sigma or
sigma-delta converter 1. Advantages of sigma-delta converters are:
• serial one-bit digital clock-synchronous output

• simple and compact architecture which can be implemented
using a small number of components
• a low-power consumption, especially at low sample speeds
• no accurate components or component matching required
• exchange of sample speed and resolution possible
• high resolution possible (J 6-bit)
Disadvantageous can be the fact that the digital bitstream output needs
digital post-processing in the form of an averaging, or noise-shape fiIter, in
order to convert the one-bit output signal to multi-bit samples, which can
I Both names, delta-sigma and sigma-delta, are common, and refer to the same type of
converter.

be used in a digital processor. Such a filter might be implemented in

hardware on a separate (CMOS) chip which can be connected to the sensor
bus, or in software on a microcontroller Of PC bus master. This enables a
shared use of the filter for aII the smart sensors that produce a bitstream
output, see also section 2.5.
We will first proceed by explaining the sigma-delta conversion principle.
2.4.1 Sigma-delta converter

The sigma-delta principle is based on oversampling, and on feedback and
filtering (noise-shaping) of the quantization error. The operation principle
in the time domain will be explained using Fig.2-ll and Fig.2-12.
The basic sigma-delta converter consists of an integrator, often

implemented in a feedback configuration using an OpAmp, and a clocked
comparator, mostly implemented as a master-slave ftipftop.
The fact that we largely oversample the input signal,
f clock » fNyqUist=2xbandwidth (2-4)
permits the assumption that the input signal is considered constant for
many c10ck periods.
The first figure iIlustrates the situation that the presumably constant input
current is integrated on the integration capacitor, and results at the
integrator output in a ramping voltage signal, with an upward slope
proportional to the input signal. The c10cked comparator compares this
voltage signal to the comparator's reference level Vcmp ' and outputs a
digital decision, low or high, only once each c10ck period, for example at
the positive clock edge. In this situation the output remains low until the
first c10ck edge after the integrated signal exceeds the comparator's
reference level. Then the comparator switches the output signal to high,
and turns the switch of the reference current, as illustrated in the next
figure. Provided that I ref> lin' an opposite integrator input current of
I ref -Im results now in a proportional negative slope at the integrator
output. The comparator output is high until the integration signal has
dropped below the reference level, and a c10ck edge occurs. The output
will become low again, and the sigma-delta converter switches back to the
situation of Fig. 2-11, and the procedure repeats itself.

I'~ ;"~ -,~id:~- ------:

- - - - - -
ooor o.
t
dock I
mput
Lurrent r..
> ___..L..-~bitstream
output
integrator
Fig. 2-11 Sigma-delta converter principle, output low (A).
1
mput
current t n.
bitstream
> ___ ..L..-~ output
clocked
integrator comparator
Fig.2-12 Sigma-delta converter principle, output high (A).
When considering a larger number of c10ck periods we see that the

integrator signal osci\lates around the reference level, and the digital
output produces a series, or pattern of ones and zeroes at the beat of the
c10ck signal, as depicted in Fig.2-13 for an input signal of I m =O.19.lre!"
The input signal range is usually limited to the signal range from
O.l.Jrel to O.9.Jrel , to avoid very low slopes in the integrator signal.
Through simple reasoning it can be found that a large input signal results
in steeper upward slopes, and less steep downward slopes, and thus a large
number of high outputs. The reverse app1ies for small input signa1s; the
number of high outputs will be small. Hence the output pulse density
seems proportional to the input signal; in fact the value of the input signal

O . 5-r------~-------------,-------__,
-0.5
-1 .o ··,.-r-r-r... - ..'··rT-T·-,-T'·t·..,· "t-T · ...

.. •• _ _ 64··_~··
72
_____ J... __ ~·_·· ·' 1'60 ee ....____ _ _..96
_ _ _ _ _ _ •_____
_____ • _ _ _ _ 104
_. ____ ....... ______ ''''ro--~'
112 120
_, ,,·,·' ...._H ...·.•.. _.•H'·.·· _____ 12e . .
~·.·.~.·_
1.S :fb . - - - - - - - - - - - - - .b NI'

1.0 I m m+ :
~5- ,
0.0 , ,
~ r
64 72 80 8S 96 104 112 120 128
Fig. 2-13 Sigma-delta integratar and bitstream signal.
is encoded in the digital output bitstream. To obtain a digital number

representing the input signal, the output bits wiIl have to be counted (or
otherwise processed) over a certain time interval, interpreting high outputs
as + 1 and low outputs as O. The resolution is determined by the number of
c\ock periods taken for the averaging, in other words by the aversampling
ratia, which is defined as:
fc/o ck
OSR = (2-5)
fNyqulSl
A very high resolution AD-conversion can be obtained provided that

integrator "Ieakage", c\ock jitter, delay, and hysteresis are
minimized [24, ... ,29]. To fulfill those requirements accurate components
are not required, as opposed to the requirements for other AD-converter
types. Besides the OSR and the noise-shape order, the performance IS
determined by the circuit design and the invested bias power.
By realizing the sigma-delta converter in a fully differential topology, as

depicted in Fig.2-14, common-mode and supply interference, important
error sources in instrumentation can be strongly suppressed. This
configuration aIlows a differential input signallimited in signal range from
-O.9.lret to +O.9.lre /" High outputs should now be interpreted as + 1, and
low outputs as -1. As often seen in practical realizations, the two displayed
reference sources can be replaced by a single reference current of 2.Jret ,
provided that a common-mode (CM) control loop at the input of the

bU.Hream
>....I.r-JII oulpu[
Fig. 2-14 Differential sigma-delta converter.
integrator takes care of fixing the input CM-Ievel. This is necessary to

obtain a symmetrical differential integration signal. The clocked
comparator in this configuration simply decides the polarity of the
differential signal, it compares the differential signal with the intrinsic
reference of O. Fig.2-15 further demonstrates the operation of the
(differential) sigma-delta converter. A slowly varying sine-wave input
signal clearly results in a pulse-density modulated bitstream signal. In this
example it is attempted to reconstruct the original signal by taking the
ave rage of the bitstream signal in clusters of 16 clock periods.
Fig. 2-15 Sine-wave input. bitstream. and decimated output signal.

Essentially the principle of sigma-delta conversion is based on simple

(one-bit) quantization of the (filtered) input signal, whereas the
quantization error is fed back to the input, and averaged (filtered) to zero
(for DC) by signal integration. The output signal is composed of the
digitized input signal at low frequencies, and the quantization (noise-
shaped) error at high frequencies. A more elaborate explanation of the
sigma-delta converter, using frequency domain analysis, will be given in
chapter 5.
Another approach is to look at the charge balance which is obtained: on

the long term the signal at the output of the integrator averages to zero (i.e.
the remaining charge error can be neglected). This means that over a long
interval the integral of the sum of aII integrator input signals is zero, or in
the discrete time domain:
m+N m+N
L b n '/ ref = c.~v "" O (2-6)
n=m+l n=m+l T'/k
in which N is the number of clock periods in the interval, and b n represent

the corresponding series of output bits, referring back to Fig.2-13. The
remaining charge, determined by the difference ~ V between the integrator
voltage, at the start and at the end of the interval, and the integrator
capacitor C, can be neglected if the time interval NxT,'/k is large with
respect to the c\ock period Tclk : N» 1. The maximum voltage error
results when the minimum/maximum input signal is just not detected by
the comparator, thus ~Vmax=/refxT,/k/C. Which means the maximum
error on the right-hand side of equation (2-6) corresponds to 1x/ref'
whereas the maximum on the left-hand side of the equation is N times
larger.
Provided the input signal remains constant during the interval, we can
derive for the average of the bitstream output:
1 m+N
- '"
N L. b n
(2-7)
n=m+l
which confirms that the output bitstream signal re\ates to the ratio of the
input current and the reference current.

2.5 Digital bus or microcontroller interface
2.5.1 Microcontroller compatible interfaces

The bitstream output of a smart sensor with a sigma-delta converter is
microcontroller compatible in the sense that it can easily be read in, and
interpreted by a microcontroller or PC, as it is a serial, one-bit, digital, and
cIock-synchronous signal [30]. The appropriate sigma-delta cIock signal
can be generated by the microcontroller, and is sent to the sigma-delta.
The digital bitstream signal is sent from the sigma-delta to the
microcontroller. The microcontroller processes the bitstream in order to
obtain multi-bit samples with the desired resolution. In order to read out
multiple smart sensors with bitstream outputs the microcontroller needs to
be equipped with a multiplexer which enables the selection of the different
sensors; the clock line might be shared by several smart sensor devices.
This situation is pictured on the left side of Fig. 2-16.
microcontroller microcontroller
.------1 or PC or PC
multiplexer bus interface
smart sensors with digital outputs smart sehsors with bus interface
Fig. 2-16 STAR-coupled versus BUS architecture.
It is cIear that the with the number of sensors also the wiring complexity
increases. Other disadvantages of this star-coupled architecture are that
the number of sensors is limited by the multiplexing capabilities of the
microcontroller (usually 8 to 16), that the position of the sensor determines
the multiplex address, and thus sensors cannot be swapped, and that only
direct sensor-to-master communication is possible.
The bus architecture shown on the right-hand side of Fig. 2-16, overcomes
aII these disadvantages. The sensors and the microcontrollerlPC are
equipped with a bus interface, and connected by common lines, a
clock line and a data line for example. A bus-slave sensor will put its data
on the bus data line only after it has been correctly addressed by the

microcontrollerlPC bus master according to a bus protocol. The sensor

data can be read by the master but also by other devices on the bus, a
digital bitstream filter for example. Bidirectional communication is also
possible, for example to program the calibration factors of a sensor before
reading it out. Other advanced features which become available in a bus
architecture are sensor identification and alarm interrupts.
2.5.2 Smart Sensor bus interfaces

Because it often is not possible to integrate complex digital circuitry with
the sensor device, the first requirements for a smart sensor bus is minimum
complexity at the bus-slave side. Another requirement is to have a
minimum number of wires, bec au se for use in sensor systems each
connection introduces extra fabrication costs. A sensor bus which fulfills
these requirements is the Inter-Integrated Circuit (12C) bus, which has
been developed by Philips as internal bus for example for consumer
products with microprocessor control. The bus has two supply wires and
two wires for communication, using a simple addressing and handshake
protocol on open collector (OC) lines. A flow and temperature sensor with
an 12C bus interface has been realized in a BiCMOS process [31].
Unfortunately the 12C bus protocol allows only serial byte
communications, which means that for sensors with a serial one-bit
(sigma-delta or duty-cyc\e) signal an on-chip conversion to 8-bit or 16-bit
numbers is needed which requires a considerable amount of digital
circuitry.
For this reason the Integrated Smart Sensor (l S2) bus has been
invented [32]. The IS 2 bus is very similar to the 12e: it also has two supply
lines and two serial communication li nes, but it has an open data
transmission format [33,34,35]. In addition to the open collector
technique Manchester encoding is used, so that in combination with the
clock four different data states can be distinguished, as iIlustrated in
Fig.2-17.
The clock and the data line are both pulled up by resistors, which are
connected to the positive supply line. Using a strong open collector driver
(R clk "",500Q), the master puts a highly stable c\ock signal on the c\ock
line, which is indispensable for the accuracy of the sensors'
(time-continuous) sigma-delta converters. The data line can be pulled
down by any device on the bus, using a less powerful open collector

open-collector
Manchester Code clock and data line
pull-up
resistors
clock~
data
idle----
zero \.JL
one~
stop~
Fig. 2-17 Using a Manchester encoding in combination with

open-collector lines.
transistor (Rdata"'SkQ). Both the clock and the data line are read by the
,bus devices using high-ohmic input buffers.
When not a single bus device is active the data line remains high, and the
bus is in idle state. After at least two idle states, a bus master can address a
(sensor) bus slave by serially transmitting a start bit, eight address bits, and
awaiting an acknowledge bit from the addressed bus slave, as explained in
the bus protocol shown in Fig. 2-18. The slave can then put its data on the
data line.
Digital clock-synchronous data is transmitted according to the Manchester

code described in Fig. 2-17, a zero means pulling down the data line during
the high phase of the c\ock, and a one means pulling down the data line
during the low phase of the clock. When a one and a zero collide, because
two sensors accidental1y have responded to the same address, the line
remains low during the whole clock period which corresponds to a stop
signal which wi\l be detected by the active bus devices, and reset the bus
communication. This is also the way the master or an alarm sensor can
send an interrupt. The length of a digital transmission is controlled by the
master, so that it is possible to communicate bytes as well as any arbitrary
number of bits, for example a long series of sigma-delta output pulses.
As mentioned before, it is also possible to put other kind of data on the

bus. In that case the master keeps the clock line sti\l (high) to prevent
cross-talk distortion, and the slave can put an asynchronous digital signal
(pulse-width or pulse-frequency modulated) or an analog voltage on the
data line. It is up to the master to correctly interpret the data. Interruption

Transmission acknowledge
Protocol ; 2-i~ address ~F--d-:-a-:'ta--~t~p=
- - I I I I I
c/ock line .§
I I - (II
clock-synchronous } .~.!Il
'1' digital data I _21 ~
da/afine
~ n 'Olij
U1L- .t;
I I
'1 I
: ~}c»'*
aSynChronous : r - v - L
r:s::z1
digital ~ata
..Q .!!!
~~
.
analog data
1,
-
I
.::"'
Fig. 2-18 Explanation of the IS 2 bus communication protocol.
is now only possible by (shortly) pulling the c10ck line down which should
"wake up" the master, and get the c10ck running again .
It is also possible to make bus slaves receive data, for example to program
a sensor's input range . At the side of the master the bus addresses will ha ve
to be organized in such a way that the master knows what kind of data to
expect from or to send to the devices on the bus. This could be do ne by
setting up a table with information about the sensors attached to the bus,
containing address, type of data, number of samples, etc. or by splitting up
the address range in sub-ranges for analog sensors, for sensors with
bitstream outputs, etc. With eight address bits the maximum number of
addresses is 256.
The hardware at the sensor (bus-slave) side is very simple. The scheme is
shown in Fig. 2-19. A Manchester decoder is needed to detect the status of
the data line. After a start condition has been detected the decoded data
bits are shifted in a register. A digital comparator is needed to compare the
received address to the slave's own address. After positive comparison the
slave immediately returns an acknowledge bit, and than puts its data on the
bus . For digital data a Manchester encoder and OC driver are needed, and

2.6 Integrating a calibration function in the smart sensor concept
for other data an analog switch and a voItage driver. The slave set-up is
thus very simple, and the necessary digital circuitry is so Iimited
(approximately 20 gates and 12 ftipftops) that it can be integrated in
CMOS on a very small surface, but also in bipolar technology using for
example Emitter-Coupled Logic (ECL) as wiIl be demonstrated in
chapter 4.
c/ock line
Manchester
Decoder
Error
21dles
dii:i ral alla/"i:

Sellsor Data
Fig. 2-19 Configuration of an IS2 bus-inteiface, for a bus slave.
2.6 Integrating a calibration function in the smart sensor

concept
The favored smart sensor configuration, for bipolar implementations at
least, is reviewed in Fig. 2-20. Besides the (silicon) sensor it consists of an
analog interface which realizes a sensor signal current and a reference
current that are converted by the sigma delta converter into a digital
bitstream signal. Next, the bitstream signal can easily be pas sed on by the
IS 2 bus interface to the system's microcontroller or another bus master for
example a PC. For many sensor types this scheme is probably the simplest
way to realize a high-resolution AD-conversion and BUS interface
2.6.1 Error sources in the smart sensor signal path

Besides errors in the transfer of the sensor, from the physical input to the
electric al output signal , we can also expect errors in the electrical transfer

Jntegrated Smart Sensor sus
. I
mlcro- I
controller /
Fig. 2-20 Signal chain ofthe proposed integrated smart sensor.
of the other function blocks. Especially the analog interface may

contribute offset and gain errors. However in the electrical signal domain
we have methods available to reduce offset errors (chopping, auto-zero)
and gain errors (dynamic element matching) to values which are much
better than most sensor specifications [36,37]. Such advanced methods are
not necessary in a smart sensor interface when we intend to calibrate the
sensor anyway: we might as well calibrate the total error in the smart
sensor signal chain. An exception has to be made when we expect
considerable drift errors in the electronic circuit: in that case dynamic
chopping or element matching methods will be useful.
2.6.2 Conventional sensor calibration techniques

Despite the smart sensor concept, most sensors are still fabricated (and
distributed) separately from the electronic interface. Therefore,
conventional calibration techniques concentrated on correction of the
sensor transfer itself, during or just after fabrication. An example is the use
of laser-trimmed resistors in a pressure sensor bridge, see section 3.2.1. A
disadvantage of this approach is that it is difficult to automate the
calibration, and that it is thus an expensive procedure in a fabrication line.
Furthermore, it means that the electronic interface must match the
specifications of the calibrated sensor, which could require a separate
calibration or a more advanced circuit technique.
On the system level, sensor calibrations are usually performed on the

systems computer. Data-acquisition programs, such as LabView, offer
many possibilities for an advanced digital calibration. Disadvantage of this
approach is that each new or replaced sensor in the system first has to be
calibrated by the user, rather than by the manufacturer, which means the

user needs (expensive) reference equipment. Furthermore, the calibration

requires computational power of the system, and the calibration
information is not attached to the sensor but stored in the system's
computer. Altogether it will not stimulate the use of sensors in consumer
products if such a calibration technique is needed.
2.6.3 Correction of the sensor signal transfer

As the objective is to find a way to integrate a calibration function in the
smart sensor concept, we will start by looking for different possibilities to
change the signal transfer of one of the different function blocks that have
been presented. In the smart sensor signal chain shown in Fig. 2-20 we can
choose to alter the transfer in:
• the analog signal domain

• the AD-converter
• the digital signal domain
In aII cases it is important that the calibration is made digitally
programmable. This will make it possible to fully automate the calibration
in a fabrication process. It is also important that the errors of the whole
signal chain are calibrated, even when we apply the correction at the front
of the chain. This is done by taking the final (digital) output signal when
measuring the sensor transfer during the calibration phase.
The three options will be further elaborated in the succeeding chapters.

The sensor signal can be pas sed through an analog circuit with
programmable offset, gain, etc. To realize a digital programmability for
analog signals Digital-to-Analog Converters (DACs) are needed. This will
become c1ear in chapter 4. In the sigma-delta AD-converter we have a
mixed signal circuit which offers different possibilities for a
programmable correction as will be c1arified in chapter 5. Concerning the
digital signal domain, it is often not possible to integrate advanced digital
circuitry on the sensor chip. Alternatively it is possible to hybridly
integrate a microcontroller die in the sensor package, or to use the
microcontroller bus master for a digital calibration. Therefore, chapter 6
concentrates on the realization of an advanced calibration function in the
form of microcontroller software.
Before looking at those different practical implementation forms, the

important aspects of calibration, and different linearization methods will

be explained in chapter 3. Also an attractive polynomial calibration

technique will be proposed.
2.6.4 Calibration memory

An issue that has not yet been addressed is the storage of the carrection
factors after calibration of the sensor. Obviously it is desirable that the
correction is stared in a non-volatile memory integrated with the sensor or
interface chip. Possible forms of non volatile memories, such as (E)PROM,
zener or metal fuses, are mostly not available in standard IC processes.
The development of such a memory in the used process could easily turn
into a separate research project, and has not been given any attention
within this project. The work in this book concentrates on the development
of the circuitry needed to execute the correction. For the storage of the
calibration coefficients simple volatile RAM is used, in the form of digital
ftipftop cells. The advantage of developing digital calibration circuitry is
that the memory can easily be replaced with whatever digital memary is ar
becomes available in the used IC process.

References
[1] J.H. Huijsing, FR Riedijk, and G. v.d. Horn, "Developments in
integrated smart sensors", Sensors and Actuators A, VoI. 43, 1994,
pp.276-288.
[2] G.C.M. Meijer, "Thermal sensors based on transistors", Sensors and
Actuators, VoUO, 1986, pp.103-125.
[3] G.C.M. Meijer and A.W. van Herwaarden, Thermal Sensors,
Institute of Physics Publishing, Bristol/Philadelphia, 1994.
[4] S. Middelhoek and S.A. Audet, Silicon Sensors, Academic Press,
London, 1989.
[5] R.F Wolffenbuttel, "Fabrication compatibility of integrated silicon
smart physical sensors, Sensors and Actuators A, VoI.41-42, 1994,
pp.II-28.
[6] K.D. Wise and N. Najafi, "The coming opportunities in microsensor
systems", Proceedings Transducers 1991, San Fransisco, USA,
pp.I-7.
[7] J. Bryzek, K. Petersen, J.R Mallon, L. Christel, F Pourahmadi,
1990.
[8] J.H. Huijsing, "Signal conditioning on the sensor chip", Sensors and
Actuators, VoUO, 1986, pp.219-237.
[9] R. Frank, "Pressure sensors merge micromachining and
microelectronics", Sensors and Actuators A, Vo1.28, 1991,
pp.93-103.
[10] L.K. Nanver, EJ.G. Goudena, and H.W. van Zeijl, "Dimes-Ol, A
base-line bifet process for smart sensor experiments", Sensors and
Actuators A, Vo1.36, 1993, pp.139-147.
[II] RP. van Kampen, Bulk-Micromachined Capacitive Servo-
Accelerometer, PhD Thesis, Delft University Press, Delft, 1995.
[12] F Schnatz et al., "Smart CMOS capacitive pressure transducer with
on-chip calibration capability", Sensors and Actuators A, Vo1.34,
1992, pp.77-83.

[13] C. Sander, J.w. Knutti, and J. Meindl, "A monolithic capacitive

pressure transducer with pulse-period output", IEEE Trans. Electron
Devices, VoI. ED-29, 1980, pp.42-48.
[14] L.S. Lee and K.D. Wise, "A batch-fabricated silicon capacitive
pressure transducer with low temperature sensitivity", IEEE Trans.
Electron Devices, VoI. ED-29, 1982, pp.42-48.
[15] T. Rogers and J. Kowal, "Selection of glass, anodic bonding
conditions and material compatibility for silicon-glass capacitive
sensors", Sensors and Actuators A, VoI.46-47, 1995, pp.113-120.
[16] S. Ansermet, D.Otter, R.w. Craddock and J.L. Dancester,
"Cooperative development of a piezoresistive pressure sensor with
integrated signal conditioning for automotive and industrial
applications", Sensors and Actuators A, VoI.21-23, 1990, pp.79-83.
[17] D. Tandeske, Pressure Sensors: selection and application, Marcel
Dekker Inc., New York, 1991.
[18] Y. Kanda, "piezoresistance effect of silicon", Sensors and Actuators
A, Vol.28, 1991, pp.83-91.
[19] S. Kim and K.D. Wise, "Temperature sensitivity in silicon
piezoresistive pressure transducers", IEEE Trans. Electron Devices,
VoI. ED-38, 1983, pp.802-81 O.
[20] M. Gunawan, G.C.M. Meijer, J. Fonderie, and J.H. Huijsing, "A
curvature-corrected low-voltage bandgap reference", IEEE Journal
of Solid-State Circuits, Vol.28, No.6, 1993, pp.667-670.
[21] L.G. Fasoli, ER. Riedijk, and J.H. Huijsing, "A general circuit for
resistive bridge sensors with bus compatible output", Proceedings
IMTC'96, June 1996, pp.1205-121O.
[22] R. v.d. Plassche, lntegrated Analog-to-Digital and Digital-to-Analog
Converters, Kluwer Academic Publishers, Boston/Dordrechtl
London, 1994.
[23] P.R. Gray, B.A. Wooley, and R.W. Brodersen, edits., Analog MOS
Jntegrated Circuits Il, IEEE press, 1988.
[24] J.c. Candy and G.C. Temes, Oversampled Delta-Sigma Data
Converters, IEEE Press, New York, 1992.
[25] ER. Riedijk, lntegrated Smart Sensors with Digital Bus lnterjace,

[26] B.E. Boser, Design and implementation of oversampled analog-to-

digital converters, PhD Thesis, Stanford University, 1989.
[27] PJ.A. Naus and E.C. Dijkmans, "Low signa1-1eve1 distortion in
sigma-delta modulators", Proc. Audio Eng. Soc., March 1988.
[28] O. Feely and L.O. Chua, "The effect of integrator leak in sigma-delta
Modulation", IEEE Trans. Circuits and Systems, Vo1.38, No.11,
November 1991, pp.1293-1306.
[29] RW. Adams, "Design and implementation of an audio 18-bit
analog-to-digital converter using oversampling techniques", J. Audio
Eng. Soc., Vo1.34, No.3, March 1986, pp.153-166.
[30] A.W. van Herwaarden and RF Wolffenbuttel, "Introduction to
sensors compatible with microprocessors", Microprocessors and
Microsystems, Vo1.14, No.2, 1990, pp.72-82.
[31] EJ. Hogenbirk, HJ. Verhoeven, FR Riedijk, and J.H. Huijsing, "an
integrated smart sensor for flow and temperature with i2 c bus
interface based on thermal sigma-delta modulation", Proceedings
Transducers 1993. Yokohama, Japan, pp.792-795.
[32] J.H. Huijsing, RF Tuk, FR Riedijk, M. Bredius, and G. v.d. Horn,
Mixed AnaloglDigital Two-Line Bus System, Dutch Patent
Application, Nr.9320\ 595-1, June 1993.
[33] FR Riedijk RF Tuk, M. Bredius, and J.H. Huijsing, "A bus-
controlled interface environment for an accurate read-out and
flexible interconnection of integrated smart sensors in local
systems", Proceedings Transducers 1993, Yokohama, Japan,
pp.327-330.
[34] M. Bredius, FR Riedijk, G. v.d. Horn, and J.H. Huijsing, "The
Integrated Smart Sensor (l S2) bus", Proceedings Sensor Expo West,
Philadelphia, USA, 1993, pp.243-247 ..
[35] FR Riedijk and J.H. Huijsing, "Sensor interface environment based
on sigma-delta conversion and serial bus interface", SENSORS,
journal of applied sensing technology, Sensors Expo issue, April
1996.
[36] FM.L. van der Goes and G.C.M. Meijer, "A novel low-cost and
accurate multi-purpose sensor interface with continuous auto-
calibration", Proceedings 1MTC'96, June 1996, pp.782-786.

[37] P.e. de Jong and G.e.M. Meijer, "Absolute voltage amplification

using dynamic feedback control", Proceedings IMTC'96, June 1996,
pp.67-71.

Calibration and
Linearization
3
Techniques
3.1 Introduction
In the first part of this chapter we will identify and distinguish the different
types of errors which affect the transfer of the sensor, and explain which
errors can be corrected by calibration. In the following section we will
explain different Iinearization techniques which can be used to calibrate
the offset, gain, and Iinearity errors in the sensor transfer. In the last part
we will propose and explain a polynomial calibration method which can
be used to calibrate and Iinearize the sensor transfer in a step-by-step
approach. It will be shown how the method can be expanded to a
two-dimensional polynomial calibration to be used for calibration of a
cross-sensitivity error.
3.1.1 Errors in the sensor transfer curve

In sensor production it is desired to produce large batches of sensors at a
time, and to have ali of the sensors respond according to the same
well-defined transfer within a certain accuracy. To verify this the sensors
could be measured individually or by sample. When measuring the

Calibration and Linearization Techniques
transfer of a sensor we may be confronted with one or more of the

following error types [1,2]:
• offset: when a zero (or minimum) physical input signal is

applied, the measured output signal is not zero (or minimum
of output range) but shows an 'ojfset' value.
• gain, range, or full-scale error: the sensitivity ofthe sensor
is not as intended, the maximum physical input signal does
not match with the maximum electrical output signal.
• nonlinearity: the sensor output does not change linearly
with the physical input signal.
• cross-sensitivity: the transfer curve changes when measured
at different ambient conditions (temperature, for example),
thus the sensor is not only sensitive to the input signal but
also to other parameters.
• hysteresis: the sensor transfer curve is different for
decreasing physical signals thanfor increasing signals,
once the signal passes a certain level.
• drift: the sensor transfer curve changes (slowly) in time.
The errors are explained graphically in Fig. 3-1. Generally, some of these
errors are larger than the desired accuracy allows. To correct these errors,
ali sensors should be measured and calibrated individually to guarantee the
accuracy.
Gain, offset, and also linearity and cross-sensltlvlty errors may be

observed for practically ali types of sensors. Hysteresis and drift errors
need to be examined for some types of sensors, but can be neglected for
other sensors. Although it may be somewhat complicated, hysteresis errors
could be calibrated. However, other sensor-specific solutions should be
considered first as they will often be less complicated or provide better
results. Drift errors are difficult, if not impossible, to compensate or
calibrate as it would require prediction of the future errors (extrapolation),
and time measurement. Sometimes it is possible to reduce drift of the
sensor by means of self-calibration, using an additional transducer (sensor
or actuator) with a lower drift, see section 1.4.4, Fig. 1-8, and Fig. 1-9. For
many sensors it is acknowledged that in a certain time (-n years) the
sensor transfer curve may drift out of the specified range of accuracy, and
recalibrations will be necessary. The smaller the drift errors, the longer the
time span that can be allowed between recalibrations.

3.1 Introdnction
output
input
output T=250K output t=18 mon/hs

T=300K 1=12 months
T=350K t=6 mOllthv
T=400K 1=0
input input
Fig. 3-1 Possible errors in sensor transfer curve.
The treatise presented in this book is limited to calibration of the offset,

gain, linearity, and cross-sensitivity errors.
3.1.2 Systematic and random errors

Some errors may be characteristic for the type of sensor that is considered.
It can be described in a model of the sensor. A thermal flow sensor, for
example, shows a typical nonlinear rather than a linear transfer from flow
velocity to sensor output signal. The source of such systematic errors
should be traced in the sensor construction or read-out electronics. Other
errors show a statistical distribution, due to random variations in the
production process. For example, the offset of a resistive sensor bridge
might show a Gaussian distribution because of uncontrolled and
unpredictable differences in the values of the four resistors in the bridge.
When the errors in the transfer are larger than the desired sensor accuracy,
the sensors in production need to be calibrated (or selected on measured
specifications). Both systematic errors and random errors could be

calibrated for each individual sensor. However, when the systematic error
is much larger than the deviation of the random errors, this approach is not
optimal. It is better to trace the source of the systematic error and correct
it, or introduce a systematic error compensation for the whole sensor batch
or production. After removing the systematic error, the remaining random
errors can be calibrated per individual sensor, so that ultimately the error
distribution fits within the accuracy limits. Fig.3-2A may represent the
statistical distribution of the offset for a large batch of sensors, showing an
average (~) of 45mV and a standard deviation (cr) of 5mV For the
example, we suppose that the desired accuracy allows a maximum offset
of±2mV
1 probabtlity
J.l=45mV
o 20
(A)
60
--
offset (mV)
1
probabt hty
-20 -10 O 10 20 --
offset (mV)
I
(B)
probabthty
±2mV
-20 -10 O
(e)
10 20 --
offset (mV)
Fig. 3-2 Statistical offset distribution, before (A), and after

systematic correction (B), and calibration (C).

3.1 Introduction
PracticalIy aII sensors (99.76%) have an offset within the 11 ± 30" range,
and so, initialIy, none of them has an offset smaller than 2mY. To calibrate
the offset error, a correction of maximum 60mV with a 4mV accuracy
would be required. By applying a systematic offset correction of -45mV
first, the shifted distribution of Fig. 3-2 B is obtained, and approximately
30% of the sensors have an offset according to the specification. Next, an
offset calibration with a correction range of ± 15 mV and a resolution of
4mV (-3-bit) is sufficient to put almost aII sensors in the ±2mV offset
range. The offset error distribution of the sensors, after such a calibration
is shown in Fig.3-2C. The calibration accuracy can be increased to further
reduce the offset error and the width of the distribution.
3.1.3 Linear and nonlinear errors

We can distinguish between linear errors and nonlinear errors. Offset and
gain errors and their first-order temperature coefficients are linear errors.
In order to correct linear errors only linear operations are required, which
are easy to implement in circuit technology. In a straightforward manner a
programmable offset and programmable gain can be realized. Some
examples will be given at the beginning of chapter 4. The calibration or
correction of the nonlinearity, referred to as linearization, is more
complicated. We need to counteract the sensor's nonlinearity and somehow
realize an inversely nonlinear electronic transfer. Many different methods
can be found in literature. The basic linearization principles are explained
in section 3.2. In section 3.3 we propose and explain the polynomial
linearization method which wilI be used in the realizations presented in
subsequent chapters.
For linearity errors we should distinguish also between systematic and

random errors. A systematic nonlinearity, which can be explained from the
sensor's physical model or reproduced consistently from measurements,
should also be corrected systematicalIy. An example of a systematic
linearity compensation wilI be given in section 3.2.1. Preferably,
calibration is used only to compensate random errors resulting from
device-to-device variations. A certain number of measurements need to be
taken in order to determine and correct a sensor's nonlinearity. The number
of measurements necessary to obtain a certain reduction of the linearity
error, strongly depends on the linearization method used. To reduce the
costs of calibration it is important to minimize the number of
measurements, because measurements take (production) time. This is an

important criterium in the selection of an appropriate Iinearization method

for sensor calibration.
3.1.4 Normalized sensor transfer functions

One-dimensional sensor transfer function
To simplify the mathematical descriptions of the linearization methods we
wiII work with normalized sensor functions. The electrical output signal
eoul of a sensor can normally be considered as a function S of one physical
input signal <Pin: e oul = S ( <i' in) .
Through normalization, the range of interest [<Pmtn' <PmruJ for the physical
input signal is mapped on the range [-1,1] for a normalized input signal.
Similarly, the desired output range [emin,e max ] of the electrical sensor
output must be mapped on the [-1,1] range for a normalized output signal.
We can derive the normalized input signal and normalized output signal of
the sensor respectively as:
2<p - <i'max - <i'min 2e - e max - e mln

x y (3-1 )
<i'max - <i'mln emax -e mln
The normalized sensor output y can be described by the normalized

transfer function, which is expressed as a one-dimensional function of the
normalized input signal x: y = f(x) . Furthermore, we assume that the
desired normalized transfer is a linear function of x with unity gain:
y=g(x) =x.
A normalized transfer function example, f(x), and the ideal transfer

function g(x) are shown in Fig. 3-3. We wiII use the example curve for the
explanation of the Iinearization methods in the next section. The shown
function f(x)=-1.2 o ln(1.7-x) is taken as an arbitrary example, and not
because it would characterize a typical sensor. Assuming that for a real
case sensor the systematic errors have been compensated first, then only
arbitrary errors would remain. Thus we do not know beforehand how to
describe the transfer functionf(x). Instead, we have to take several, say N,
calibration measurements f(x n ) [n=1 to N] which can be used to compose
an interpolation functionj'(x) which resembles the real sensor transfer. On
the basis of the interpolation function, we can linearize the sensor transfer
so that it approaches the ideal sensor curve g(x). The Iinearized transfer

3.1 Introduction
1.0
t normallzed
output
~em~or
"gnal
g(x)=x
0.5
normahzed
."Iensor I nput
f---i--+--+-r-+---f---:A---j "gnal
-1.0 -o 5 1.0 - - x
-1.0
Fig. 3-3 Normalized one-dimensional sensor transfer function,

and ideal transfer function.
function y=h(x) will thus be composed of the sensor output signal f(x), the
calibration measurementsf(xn ), and the corresponding ideal output values
g(xn):
(3-2)
The obtained linearization can best be examined by looking at the error

function, defined as:
f (x) = h (x) - g (x) (3-3)
Because of the normalization, the error curve can also be directly

interpreted as the relative error (with respect to one half of the symmetrical
scale).
Two-dimensional sensor transfer function
If we want to examine sensors with cross-sensitivity to another physical

input signal, we must consider the sensor transfer as a two-dimensional
function. We may express the electrical sensor output eoul as a function S
of two physical input signals <P,n and 'I',n: e oul = S (<I'in' 'Vin)' Again, we
may transform the electrical output signal to a normalized output y, and the

physical input variables to normalized input signals x and z. We can then

examine the normalized two-dimensional transfer function y = f(x, z) .
Ideally, the normalized output y should be linearly dependent on the input
variable x with unity gain, but independent of the other input z (zero-gain).
Thus the ideal normalized function is given by: y = g (x, z) = x. The
two-dimensional transfer functions can be presented graphically as a
surface in a three-dimensional (30) picture. Examples are shown in
Fig.3-4. In this case, the goal of the calibration is to process the sensor
...... '; ' "
"mbiCnl "mbienl
tcmpcrl.:Hure prc,"ure i npUI lempermure prcs. . urc input
Fig. 3-4 Normalized two-dimensional sensor transfer function,

and ideal sensor transfer function.
transfer function f(x,z) in such a way that a function h(x,z) results, which
resembles the linear function g(x,z). This will require calibration
measurements f(xn,zm) for various input conditions defined by
x n [n= l to N] and zm [m= I to M]. Referring to the number of
measurements this is called an NxM calibration. Similarly to
equation (3-3), we can define the two-dimensional error surface
E (x, z) = h (x, z) - g (x , z) .
3.2 Linearization
Some types of sensors show a characteristic nonlinear transfer, which can
be explained from a physical model , or which can be reproduced
consistently. In that case, the required calibration can be simplified by first
applying a systematic linearization, as proposed in section 3.2 .1. The

3.2 Linearization
remaining random variations of the nonlinearity can then be calibrated by

using a general linearization method. Such methods are applied directly
for sensors with a more or less linear transfer, showing only linearity
errors that vary randomly from device-to-device. The following sections
will explain several such linearization methods. All those methods are
based on the use of calibration measurements of the sensor output signal.
Besides for linearizing the signal transfer, the calibration measurements
are also used for correcting the offset and gain error.
3.2.1 Sensor characteristic linearization

Certain kinds of sensors always show a typical non linear transfer curve.
The shape of the nonlinearity is reproducible and can often be traced in the
model of the sensor. Such a systematic error should preferably be
corrected by a systematic linearization, rather than by means of
interpolation of multiple calibration measurements. A sensor may, for
example, typically show a logarithmic transfer which could be written in
the form of equation (3-4) with a, b, c, and d as sensor-dependent
constants:
es = f ( c.p) = a + b . In (c + d . q» (3-4)
Such a logarithmic transfer curve can easily be linearized by using an

electronic circuit with an exponential transfer function. We could use the
exponential relation between base-emitter voltage and collector current of
a bipolar transistor [3], see equation (2-1). The idea is illustrated in
Fig. 3-5. The sensor output signal es is applied at the base-emitter voltage.
The collector current I( is converted into a output voltage ~)ut by an
IV -converter. For the overall transfer of sensor and circuit, we achieve:
h ( q» = R . le = R· e T be
K V
= R. e K T {a + b In (c + d . q» } (3-5)
= R·e aKT· e bKT. (c+d'q»
which can be written in a simplified form as:
h (x) = A +B .x (3-6)

Exponenttal
CIrcuit IV Converter
Sensor
~)ut
output
vottage
physlcal mputvottage physlcal

mputsignal mputslgnal
Fig. 3-5 Linearization of a typicallogarithmic transfer function.
Due to the inverse transfer curve of the transistor with respect to the
transfer of the sensor, the output is now linearly proportional to the
physical input signal «>, and only the gain and offset errors remain to be
calibrated. Attention should be paid to the temperature sensitivity
introduced by K Ţ = q/kT. This temperature sensitivity could be
cancelled easily by multiplying the sensor signal es with a PTAT
(proportional to 1) signal, before applying it to the transistor [3].
It may not always be possible to find a way to integrate the inverse

function of the sensor curve in an electronic circuit. Furthermore, the
compensation applied may not sufficiently reduce the nonlinearity. In
those cases, additional linearization can be improved, using one of the
approaches presented in the next sections.
3.2.2 Linearization based on a look-up table

A simple approach to correct any type of transfer is to store the complete
inverse function of the sensor transfer in a table and to look up the
corrected output value corresponding to the sensor output signal
measured [4,5,6,7]. An Analog-to-Digital converter (ADC) is used to
quantize the sensor output signal into a digital number, which can be used
as an address of the memory containing the look-up table. At the memory
address, the corrected digital value for the desired transfer curve can be
read out. The digital output can be converted into an analog signal by
means of a Digital-to-Analog converter (DAC). A graphical explanation is
shown in Fig. 3-6.

3.2 Linearization
/
look-up table
MEMORY i
~ 1-_.diit8i,.Ia.''.IU..
I".UI. calibta/ed ou/put
slgnal
S<Il
senso,
oufpuf address atlalo}! OU/jJlll
signal
c,,><:''O ! (x ) inpuf signal

4"
",'O'"
inpul signal
Fig. 3-6 Look-up table calibration method.
During the calibration phase, the correct output signals must be stored in
the memory at the correct addresses. This can be done by (slowly)
scanning through the range of the physical input signal of the sensor, and
storing the desired digital output values in the memory, at each single
address change. The correct digital output values could be obtained from a
reference sensor with an ADC [6]. Filling the complete look-up table
requires a large series or 'scan' of calibration measurements. Another
approach is to limit the calibration to a few measurements, and use an
advanced interpolation algorithm on a computer to obtain the required
table values.
The quantization of the uncorrected nonlinear sensor output signal implies

a very uneconomical use of the ADC's dynamic range. First of ali , when
large offset and gain errors (distributions) are taken into account, it is
possible that some sensors will use only a small fraction of the ADC's
input range. Furthermore, a nonlinear transfer may locally show a low
sensitivity, which means less ADC quantization levels will be available for
a certain range of input signals. This effect will become c1ear in the
example, shown in Fig. 3-7.
A 6-bit quantization of the normalized input signal over the [-2,2] range
was used, and applied to the example sensor function f(x) . The quantized
signal f w "{x) is shown in Fig.3-7. Each quantization level on the y-axis
can be used to obtain a corrected output value, which results in the
calibrated signal transfer function hw.,{x). It is c1ear that for input signals x
in the range [-1,0], there are less quantization levels available than in the
[0,1] range due to the nonlinearity off(x). The corresponding error curve,

10
t~utput
0.5 !(x)
_x
-10 -0.5 1 O Input
10% tE
error
+ 1/2LSB
mput
-1 I -X
_1/2LS B
-10%
Fig. 3-7 Look-up table linearization of the transfer curve, and

corresponding errar curve.
also shown in Fig.3-7, demonstrates that the quantization error will not
remain within the ±'/2 LSB limits.
The disadvantages of using a look-up table can be listed as:
- requirement of a large memory: for an 8-bit quantization

256 bytes are needed,for J6-bit J35 kilobytes are needed
- many calibration measurements needed or a good
interpolation to obtain ali values for the look-up table
- quantization ofthe (nonlinear) sensor output signal is
needed

3.2 Linearization
Advantages of the method can be summarized as:
+ once the look-up table is filled, a very fast correction of the

sensor signal is possible, as no signal processing is needed
+ the look-up table calibration can be rea/ized in a simple
straight-forward hărdware or software implementation
+ the calibration includes the errors ofthe sensor interface
circuit and the AD-converter
The system configuration shown in Fig.3-6 can easily be expanded to
achieve calibration of a two-dimensional sensor transfer curve. In case of
cross-sensitivity to temperature, we can add a temperature sensor with an
ADe. The digital temperature signal is then used in combination with the
original sensor signal to compose an address for the look-up table.
Naturally, memory requirements and the number of calibration
measurements will increase accordingly.
3.2.3 Linearization based on piecewise-linear interpolation

Instead of storing the complete sensor transfer curve in a large memory,
we could use a small memory and store only the coefficients of a model
which describes the sensor transfer. We can measure the sensor output
signals for a small set of known input signals. These calibration
measurements represent 'knots' on the sensor transfer curve. Between all
adjacent knots we can draw straight lines, described by y = an + b nX .
These lines compose the piecewise-linear (PWL) interpolationfpwix) of the
sensor transfer curve. Based on these linear 'portions', the sensor output
signal can be proces sed in a simple way to obtain a linearized transfer
curve. In each sub-range an 'offset' and a 'gain' correction will be applied.
Only the values indicating the sub-ranges and the required offset and gain
corrections for each sub-range need to be stored in a memory.
The method is explained graphically in Fig. 3-8. After the piecewise-linear

model of the sensor transfer has been established in the calibration phase,
a range detection is sufficient to determine what offset and gain correction
must be applied to the sensor signal. Although a digital implementation is
the most obvious, the signal processing can also be done in the analog
signal domain. Hence the dotted print used for the ADC and the DAC in
the picture of Fig. 3-8.

MEMORY
1:7
senSor
ou/put signal
signal
&<;10 fix )
0<
1:'<;
~0
input signal
otTset & gain input signal
correctlon
Fig. 3-8 Piece-wise-linear methodfor linearity calibration.
A piecewise-Iinear interpolation of the example transfer curve is shown in

Fig.3-9. It is based on five calibration measurements represented by the
knots (x /,y /) to (x5,Y5). The effect of quantization is not taken into account
as it is not essential for the piecewise Iinearization to quantize the sensor
output signal.
The linear interpolation between the first two knots can be found as:
(3-7)
Which can be proved simply by verifications with x=x/ and x=x2 .
With N knots the complete PWL curve can be described mathematically by

the sum ofthe N-1 linear componentsfn(x):
(3-8)
f pwl (x) (3-9)

3.2 Linearization
10
t~utput
0.5
__ x
-1.0 -o 5 10 input
-1.0
x/,y/
4% t:rror
2%
-1.0 -0.5 0.5 1.0
'<_---+--j;:--+-=-,k-=--+----,r----+----j __ x
input
-4%
Fig. 3-9 Piecewise-linear interpolation, and corresponding

linearization and errar curve,
We are not really interested in the interpolation curve of the sensor

transfer, but rather obtain an expression for the linearization function of
the sensor transfer, Similar to equation (3-8) we can describe the
components hn(x) which transform the output signalj(x) as follows:
f(x) E [Yn'Yn+I)-j
g(xn+l) -g(x n)
hn(x) = g ( x n) + - - - - - - . {f(x) - Yn} (3-10)
Yn+l-Y n

based on the desired output values g(xn) and the measured output signals:
(3-11 )
By substitution it can be proven that the values of components hn(x) match

the desired output values in two calibration points, hnCxn)=g(xn) and
hn(xn+l)=g(xn+l)' The linearization function hpwdx) is then defined as the
sum of the sub-range components:
N-l
hpwt (x) = L hn (x) , (3-12)
n=l
and thus intersects the desired transfer function g(x) at aII input values x n
used for the calibration.
The resulting calibrated transfer curve and the error curve for the example
case with five calibration measurements are shown in Fig.3-9. The
Iinearity peak error is reduced to 5%, while the uncalibrated nonlinearity
was about 30%. The error reduction could be improved somewhat by
optimizing the division of the sub-ranges, instead of taking equidistant
calibration points xn- This would only be possible for specific transfer
curves of which the shape is known beforehand. The Iinearization can be
further improved by increasing the number of interpolation knots, at the
expense of more calibration measurements and a larger memory
requirement. Alternatively, a better interpolation between the knots of the
transfer curve can be used to improve the linearization, as will be
explained in the next section.
The advantages and the disadvantages of the PWL calibration and

linearization method can be summarized as:
+ low memory requirements

+ a reasonably smal! number of calibration measurements is
required
+ fast and simple signal correction; no advanced signal
pracessing is required, only gain and offset correction
- reduction of the linearity errar is limited; very nonlinear
transfer curves stil! require a large number of calibration
steps

3.2 Linearization
- separate calculation of ofJset and gain correction is

required for each subsection of the signal range
- large discontinuities in the derivative ofthe linearized
signal
3.2.4 Linearization based on piecewise-polynomial or spline

interpolation
Rather than using linear interpolating between two knots on the sensor
transfer curve, we can use a polynomial interpolation between three or
more consecutive knots. These interpolations can also be expressed
mathematically as splines [8,9]. A series of such interpolations or splines
can be used to compose a 'piecewise-polynomial' or spline function which
closely resembles the original sensor transfer function over the complete
input range [10, 11,12].
The second-order polynomial interpolation through three knots can be

described by:
with the coefficients an and bn based on the three measurements Yn=f(xn),

Yn+J=f(xn+J)' and Yn+2=f(xn+2) of the transfer function:
Yn + l - Yn
(3-14)
b = Yn +2- Yn Yn + l - Y n
n (x n + 2 -xn ) (x n + 2 -xn + 1)
We can divide the input signal range into sub-ranges, based on the
measurement points xn' In each sub-range there are two polynomial
interpolations available, except for the first and the last sub-ranges where
just one interpolation is valid. Therefore, we define the linearization

function fsPL(x) for N interpolation knots as the following combination of

the interpolation components or splinesfn(x):
(3-15)
Thus, in the sub-ranges where there are more interpolation functions

available they are averaged.
Equations (3-13) to (3-15) describe the interpolation function of the sensor

transfer curve f(x) and not the linearization function. We can apply the
same approach as was used for the piecewise-linear case to find the
demanded linearization function hSPL(x).
First, the sub-range components are described by:
h n (x) = g (x n) + An {f(x) - Yn} + Bn {f(x) - Yn} {f(x) - Yn + l}

(3-16)
in which the coefficients An and Bn can be based on the calibration

measurements of the transfer function, indicated by Y n = f(x n ), as
follows:
Yn+l-Y n (3-17)
g(x + ) -g(x )
n 2 n
B = ....,.----""---=----""----
n (Y n +2- Yn) (Y n +2- Yn+l)
The Iinearization function hSPL(x) is then composed of the components

hix) in the same way as the interpolation function in equation (3-13).
Fig. 3-1 O shows the interpolation and linearization function for the
example case based on five calibration points (Xl>Y 1) to (x5,Y5), and also
the resulting error curve.

3.2 Linearization
1.0 t~utput
05
_x
-10 -0.5 1.0 input
X/>YI
t:rror
10%
0.5%
-10 -0.5 0%
O
-0.5%
-1.0%
Fig. 3-10 Spline interpolation and linearization functions, and
Clearly, the linearity erroT is reduced significantly from about 30% for the
uncalibrated transfer function to below 1% for the linearized function.
Here again, we see that the peak error is in the x=O.S to x= 1 range and
could have been reduced by a better choice of the calibration points xn- For
finding the optimal choice based on statistical information about the sensor
function, Crary even developed an advanced computer program [13,14].
Again, it only helps for transfer curves with a c\early characteristic
nonlinearity, and a limited deviation from this curve for a batch of sensors.
Advantages and disadvantages of the spline interpolation technique are:
+ good linearization can be obtained with a low number of

calibration measurements

+ systematic approach is used, which can be set up with

recursive equations [9J very suitable for software
implementations
- ali calibration measurements need to be executed and stored
before correction curves are computed
- for each sub-range a different correction function is
computed, using different coeJficients
- advanced computations are needed, preferably implemented
in computer or microcontroller software
Further improvement in the linearization can be expected when even more
knots are used for each interpolation, or, in other words, when higher
degree splines are used [9]. When alI measurements are used in one single
interpolation function we obtain an expression of a curve that 'fits' the
entire sensor function on the full input range. Similarly, we can obtain an
inverse interpolation curve which linearizes the sensor transfer function.
This case will be described in the next sections.
3.2.5 Linearization based on curve fitting

As explained in section 3.2.1 it is possible to counteract the nonlinear
transfer of the sensor with an inversely nonlinear transfer when the
expression for the sensor transfer curve is known. It is also possible when
the sensor transfer function is not precisely known but only characterized
by some calibration measurements [1,2,4,5,15,16,17]. Curve-fitting
techniques make it possible to compute a mathematical function which
intersects the actual sensor function in the measurement points. This is
usualIy done by defining a weighted sum of expressions based on the
sensor output y=f(x). The weight fac tors are caIculated on the basis of the
measurements f(x n) of the sensor transfer function. One may then try to
compute the exact inverse function of the found curve-fit function. A
better approach is to compose a curve-fit function H {f(x)} for the inverse
sensor transfer right away. The curve fitting then refers to matching the
resuIting transfer function h(x)=H {f(x)} to the desired linear transfer
function g(x). The linearization function can usualIy be expres sed in the
folIowing form:
h (x) = L c n . En {f(x) } (3-18)

n=O

3.2 Linearization
The expression En depends on the specific curve-fit technique used, and

could be, for example, an exponential, sinusoidal, or polynomial function
of increasing order (n) . The weight factors c n ' also referred to as
calibration coefficients, are caJculated, using the caJibration measurements
f(x n ) and the corresponding values for the desired transfer g(xn ) .
The computations required can be split in two groups. First, the caJculation
of the correct values for calibration coefficients during or after the
caJibration phase. This may be done externally of the sensor (on a
computer). Second, caJculation of the Jinearized output signal based on the
sensor output and the calibration coefficients. This correction must be
integrated with the smart sensor, either in analog or digital hardware, or in
software in a microcontroller interfacing the sensor. Memory must be
present to store the calibration coefficients. The procedure is explained
graphically in Fig. 3-11.
sensOl MEMORY callbrared OUIPUI

ourpul $'gnal
calibrat ion cocflicicnl'
slţJnal g(x)
(,~,..e fix)
.0~~O\ __
curve fii
COM PUTATIO
",pUI Signai IfIpul sJ!}nal
Fig. 3-11 Inverse transfer function method for linearity

calibration.
A well-known curve-fitting technique makes use of polynomial functions

and Lagrange interpolation. For N calibration measurements, the
linearization function can be described by :
And the N coefficients can be caJculated by taking the complete set of

calibration measurements:
(3-20)
which will not be worked out here. A modified polynomial calibration

technique will be elaborated in section 3.3.

As an example of the curve-fitting technique, Fig.3-12 shows the

linearized transfer curve h(x) based on a fourth-order polynomial of the
sensor function f(x). The calculation of the coefficients then requires five
calibration measurements. The corresponding error curve shows how well
the method reduces the nonlinearity.
lOt ~utput
0.5
~~~~--~--~~-+--~~~--~ --.x
-1.0 -0.5 1 O input
-1.0
0.5% tEerror
0%
~==c=~~~==~~~~~~~--~--.x
-1.0 -O 5 O 0.5 10 input
-05%
Fig. 3-12 Fourth-order polynomial linearization curve, and
Advantages and disadvantages of the (inverse) curve-fitting technique can

be listed as:
+ low number of calibration measurements results in good

linearization
+ low memory requirements because of a low number of
coefficients
+ one single correctionformulafor the complete signal range

3.2 Linearization
- higher order polynomials might require high-accuracy

computations (jloating-point operations)
- rather complex calculation of calibration coefficients may
be needed
This curve-titting technique based on polynomial interpolation can be set
up in a such a way that is compatible with the spline function method. The
main difference is that the spline method uses multiple polynomial
interpolations over sub-ranges with a subset of the measurement points,
while the curve-tit polynomial uses one interpolation over the full range,
using aII measurement points.
3.2.6 Linearization based on error minimization

Another interpolation technique, much different from the previous ones, is
based on constructing and tun ing a curved function which does not exactIy
go through aII the measured data points, but which still approaches the
measured curve cIosely. Using a mathematical algorithm, the coefticients
of the interpolation curve are tuned in such a way that the errors at the
measured data points (difference between measured y-value and y-value
on the interpolation curve) are minimized. Usually the mean square
error is taken as the error function that should be minimized. Such
methods can be applied provided that there is some sort of redundancy in
the calibration data, - we have a-priori knowledge of the measured curve
(a model) and we have more than the minimum number of calibration
measurements required to compose a corresponding interpolation curve.
Suppose, for example, that we know that the measured transfer should
obey a quadratic function and that we have four measurement points.
Then, it is better to tune the three coefticients of a second order polynomial
(quadratic curve) and minimize the errors in the four points, rather than to
calculate a third order polynomial and zero the errors. The reason for this
is that the last option might introduce a much larger error in between the
measured data points, as we knew before-hand that the introduced third
order polynomial term should be zero. However, based on the a-priori
knowledge we could also concI ude that have taken one unnecessary
calibration measurement, as three measurement points would have been
sufticient to compose a second order interpolation curve intersecting the
measured curve.

The technique for interpolation explained above can also be applied for
linearization. While the linearization techniques presented in the previous
sections required straight-forward calculations only, the optimization of
several (or many) coefficients for the minimization of the error function
requires a different class of mathematics [13,14]. How to determine the
optimum order for the linearization curve? Where to start tuning the
calibration coefficients? How to determine a good error function and how
to minimize it and assure a global (rather than a local) minimum? These
are complex issues, especially for higher order linearization and two-
dimensional calibrations. Solutions of such optimization problems are
usually found by using iterative algorithms (e.g. simulated annealing)
implemented in software [13,14]. Considering the amount of complex
calculations required, the approach is not very suitable for realization in
hardware integrated on a smart sensor l , nor for an elaborate treatise fitting
within the scope of this book.
We confine ourselves to summarizing the pro's and con's of linearization

techniques based on error minimization, but further ignore the option:
+ ideally suited for using large numbers of calibration

measurements, or
+ effectively using redundant information or a-priori
knowledge about the sensor transfer function
+ good linearization with a very low ermr can be obtained,
provided the correct choice of a linearization function based
on a known or assumed sensor model
+ good alternative approachfor when increasing the order of
a (polynomial) curvefitting does not impmve the error.
- storage ofthe (many) calibration measurements is needed,
because the are used in the ermr function optimization
- complex iterative calculations requiredfor finding the
optimum choice of calibration coefficients
- requires redundant information, and thus more than the
minimum number of calibration coefficients
- unsuitablefor hardware implementation
1 However. it is not unth10kable that a sImple Iinearization/calibration function is

integrated 10 the sensor hardware. where as the calculallon of the optimal calibration
coefficlents is do ne externally in computer software.

3.2 Linearization
3.2.7 Conclusion
Different linearization techniques were presented, which can be used to
calibrate the sensor transfer function. Sensor calibration techniques found
in literature are practicaIly aII based on one of these techniques or on a
combination of them [18,19]. Hardware and software implementations of
the look-up table and the piecewise-linear techniques can be found,
whereas the more advanced techniques are always implemented in
software. Application of sensor characteristic linearity compensation in
analog electronic circuitry can be found for thermal sensors
especially [20,21].
With respect to the complexity of the hardware that needs to be integrated

with the sensor, the look-up table and piecewise-linear methods are
favored; the software implementations can be (hybridly) integrated in the
form of a microcontroller. However, using a look-up table, it is required to
vary the physical input signal step-wise through the whole signal range.
This will be time-consuming for physical signals, such as temperature and
pressure, having a slow response or large settling time (which depends on
the actuator used). Dynamic scanning in combination with a reference
sensor may sometimes be used to speed up the look-up table
calibration [6].
Furthermore, it is made clear that, with respect to the number of

calibration measurements, the best linearization can be achieved with the
spline interpolation and the curve-fitting techniques.
In aII cases, a set of calibration measurements has to be completed and

stored first. These will then be used to compute the coefficients needed in
the correction of the sensor signal. A disadvantage of that approach is that
we have to choose the number of calibration measurements beforehand. If
the calibration appears to reduce linearity errors insufficiently, we have to
execute additional calibration measurements or a new and larger set of
measurements, and completely reda the calibration ca1culations. Using the
polynomial (curve-fit) calibration technique that will be explained in the
next section, such reca1culations can be avoided.
INTEGRATED SMART SENSORS: Design and Calibratian 75

3.3 Progressive polynomial calibration method

The polynomial calibration and linearization method we propose operates
on the principle that each calibration measurement can be used directly to
calculate one programmable coefficient in the correction function. The
correction is then immediately applied to modify the sensor output. The
next calibration step makes use of this corrected sensor signal. Each
succeeding correction step is applied in such a way that the previous
calibrations remain undisturbed.
The first measurement is done to calibrate the offset which means the
transfer function is shifted or translated, as shown in the first picture of
Fig.3-l3 .
y y
t
Offset
t
)',,.- - -- - -- -
--"x
y y
2nd order 3 rd order
t linearity
t linearity
....- - - - - ---"x
il . U:l(x)-Y.lI·tf2(x)-Yz l·l!,(x)-y ,1
)" -- y ,~~___4_______
Y=,.~
r· Y
~r=-
el
/ --"x
Fig. 3-13 Four steps of the progressive polynomial calibration/

linearization method.

A following calibration measurement is used to correct the gain error,

without affecting the offset calibration. This is achieved by rotating the
function around the previous calibration point, as shown in the second
picture of Fig. 3-13. In this way the previous offset correction at y=y 1 will
not be affected by the gain correction. The cross-dependence between
offset and gain setting, found in most other calibration methods, is
eliminated in the calibration method proposed here.
The next calibration measurements are used to correct linearity, which is

achieved by 'bending' the function in such a way that the previous
calibration points stay fixed. The third picture in Fig.3-13 shows the
second order linearity correction, and the last picture of Fig. 3-13 shows
the third order linearization. More calibrations can be done in the same
manner to further linearize the sensor transfer function. This explains an
incremental way of constructing a polynomial correction curve to tit the
desired linear transfer function. In fact, at each calibration step, the
calibrated transfer function progresses towards the desired transfer, hence
the labei progressive polynomial calibration.
We will now describe the mathematical aspects of this calibration method.
3.3.1 One-dimensional functions

In accordance with section 3.2, the (physical) input variable is indicated as
x, the (electrical) sensor output is indicated by y; the uncalibrated sensor
response is denoted by the sensor transfer function y=f(x) ; the desired
transfer function is denoted as y=g(x) and is assumed to be a linear
function of the input signal g(x)=Kx. The purpose of the calibration is to
obtain a calibrated sensor response function which closely resembles the
desired transfer function. This is done by taking calibration measurements
for a set of well-known input signals x n ' and comparing the measured
sensor output value f(x n ) to the desired output value Yn =g(xn ). The
proposed approach is to calculate a corrected sensor transfer curve hn(x)
after each calibration measurement. We make use of the previously
corrected transfer curves hdx) to hn_dx), the previous calibrated output
values Yl to Yn-l' and the n-th calibration measurement f(x n ). The
calibration is thus expressed in a series ofnestedformulas:
(3-21 )

in which an is referred to as the n-th calibration coefficient. At each

calibration step the calibration coefficient an has to be calculated in order
to obtain that the corrected sensor output becomes equal to the desired
output: hn(xn)=g(xn)=yw
We will clarify the procedure by showing the first four steps.
• The first calibration function hdx) is given by:
(3-22)
A (constant) reference signal Yref is needed only to give the correction

factor the same unit as the functions, so that al can be a dimensionless
number l . Based on the first calibration measurementf(xl) and the desired
output value YI=g(xl)' the calibration coefficient al can be calculated:
(3-23)
Substitution of equation (3-23) into equation (3-22) proves that the desired
equalization hdxl)=Yl is obtained.
• The second calibration function h 2(x) uses the previously corrected

function hdx) as follows:
(3-24)
The additive correction term is zero in the first calibration point (xl,yI),
which keeps the previous equalization still valid for the new function
hlxl)=hdxl)=YI' independent of a2' The second equalization h 2(x2)=Y2
has to be obtained by finding the correct value for calibration coefficient
a2' This requires a second calibration measurement f(x2), from which
I The reference signal is introduced for compatibility with the ImplementatlOn in

chapter 4. Normalization of the input and output signals, see page 56, can still be
applied, but is not necessary.

hJ{X2) can be ca\culated, and can then be compared with the desired output
value Y2=g(x2). The equalization is obtained when:
Y2 - h, (x 2)
(3-25)
{h,(x 2) - Y,}
• The third calibration function should be constructed in such a way that

the previous two equalizations remain valid, while a second-order
polynomial correction is obtained. A correction term that is zero in the
previous two calibration points (x"y,) and (x2,y2) is added in the third
calibration function:
{h,(x)-y,}· {hix)-Y2}
= h 2(x) + a 3 . - - - - - - - - - - (3-26)
y ref
where a reference Yref is only inserted to preserve the unit of the correction
term. A third calibration measurement !(x3) is done and immediately
corrected by ca\culating hdx3) and h2(x3), which are used to find the
desired value for the third calibration coefficient a3:
(3-27)
Substitution into equation (3-26) proves that the a third equalization of the
calibrated sensor transfer h3(x) to the desired transfer g(x) has been
achieved, namely h 3(x3)=Y3 .
• Similarly a third order polynomial correction can be realized in the

fourth calibration function, which is formulated as:
{h,(x) - y,}. {hix ) - Y2}· {h (x) - Y3}

= h 3(x) + a 4 . -:....----.::....----.::....----=-----.:----=-
3
. Y ref· Y ref (3-28)
Here too, the reference Yref is used to preserve the unit of the correction
term. Again a new calibration measurement !(x4) is needed to obtain

hdX4), h 2(X4) , and h 3(X4) , which enables the ca1culation of the fourth
calibration coefficient a4:
(3-29)
After four calibration steps, a calibration function y=h 4(x) is obtained

which intersects the desired transfer function y=g(x) in the four chosen
calibration points (x\,y\), (x2,y2), (x3'Y3)' and (X4,y4). In fact, we have been
building up a Lagrange interpolation polynomial for the sensor transfer
function. Equations 3-22, 3-24, 3-26, and 3-28 can be worked out to obtain
the standard polynomial form:
This would result in expressions for ca to c3 which prove that each

coefficient then depends on alI calibration measurements.
Additional calibration steps can be built up in the same manner as

described, increasing the order of the polynomial correction term. The
equations can be generalized, and the expression for the n-th calibration
function is given by:
(3-31 )
The calibration coefficient an is ca1culated based on the n-th calibration

measurement, giving hlxn) to hn_/(xn ), expres sed as:
(3-32)
Four calibration steps are illustrated in Fig.3-13 on page 76. Clearly, the
method achieves that the calibrated transfer function progresses the
desired function in a step by step manner. The number of steps and the
measurement points xn are not necessarily fixed beforehand.

The reduction of the error signal, the difference between the calibrated
output and the desired output, E(x)=hn(x)-g(x), depends on the number of
steps and the choice of the calibration points on the signal range:
1. When taking only one calibration step, the calibration

measurement is best taken with a reference input signal xl at
the middle of the input range.
2. When taking two calibration steps, the calibration
measurements should be done for input signals at about
25% and 75% of the input range.
3. For three or more calibration steps, generally a good
linearization is obtained when the calibration points are
selected in the following sequence: the first point at one end
of the sensor range of operation, the second point at the
other end ofthe range, andfurther calibration points
halfway between two previously selected points.
The method has been successfully tested on many different fictive sensor
transfer functions. An example is shown in Fig.3-14, where the sensor
curve was expres sed as y=flx)=-1.2·ln(1.7-x), and the desired sensor
curve as y=g(x)=x. The input signal range from -1 to + 1 was considered.
The calibration measurements were taken respectively at xI =-1, x2= 1,
x3=0, X4=+'/2, and x4= -'/2' The figure shows the calibrated transfer
function after one step h\ex), after two steps h2(x), after three steps h 3(x),
after four steps h4(x), and hsex).
The corresponding error curves tn(x)=hnex)-g(x)=hn(x)-x, are shown in

the pictures of Fig. 3-15. The initial and the first 3 calibrated error curves,
then, the third and fourth error curve, and then, the fourth and fifth error
curve are displayed, using different scales. We see that after calibration of
the offset and gain, a linearity error of approximately 30% (peak-to-peak
band) remains for this case, which can be reduced to approximately 6% by
the third calibration step, and to less than 0.5% by the fourth calibration
step, and even down to 0.2% by the fifth step.
The reduction of the linearity error that can be obtained depends, of

course, on the initial nonlinearity of the sensor transfer function. When the
sensor transfer function has a well-known typical nonlinearity, the
selection of calibration measurements can be optimized accordingly. But a
sensor-characteristic linearization would then be a better approach than the

10
t~utput
05
Xl x5 Xz
~--~--F---+-~~--+r~~~+-~ --.X
-1.0 -0.5 I O mput
-1.0
Fig. 3-14 Example of a sensor transfer function, and the

succeeding calibrated transfer functions after J, 2, 3, 4,
and 5 steps.
polynomial method. The proposed polynomial calibration method IS

valuable for sensors which may show an arbitrary linearity error.
Advantages of the proposed method are:
• a good linearization is obtained through a minimum number

of calihration measurements
• a low number of calibration coefficients, thus only a small
memory is required
• the required calibration coefficients can be obtained
directly, without mathematical iterations
• a systematic step-by-step calibration, using a repetitive
algorithm
Signal Flow Diagram

The last advantage becomes very c1ear when the signal processing is
displayed in a ftow diagram as shown in Fig.3-16. The sensor output
signal y, exhibiting a transfer function f(x) with respect to the physical
sensor input signal x, is proces sed through a network of adders,

-10 -o 5
20%
t;rror E3(xl
10 __ x
O 05 in put
stepl step3
step2
E2(x)
E1(x)
E(xl
-80%
4%
2%
__ x
-10 05 10 in put
-2%
-4%
10%
t;rror
05%
E4(x) stepS
O __ x
-10 Es(x) -O 5 0% 05 10 mput
-05%
-10%
Fig. 3-15 Error curve of the example transfer function, before and
after 1, 2, 3, 4, and 5 calibration steps.
multipliers, and trimmable calibration factors a 1 to an- First, ali calibration

factors are set to zero so that the sensor signal y=f(x), left in the diagram,
goes straight through to the output signal h4 (x) at the right side of the
diagram. Then a first calibration measurement is done by providing
reference inputf(x1)' and adjusting a1 to obtain an offset calibrated signal
h 4(x)=h 1(x)=f(x)+a1'Yref' which should be equalized to the desired
value Y1' A second calibration is done at the next reference input x2 by
adjusting a2, which will add a factor a2·{h 1(x)-Y1) to h 1(x) so that a gain-
calibrated function h 4 (x)=h 2(x) is obtained. The third part adds the term

3
a {h,-r)){hz-\'2){h l- r tl
4 )-'rcl "rei
Yref
Fig. 3-16 Graphical diagram of the polynomial calibration

method.
a:dhl(x)-YI }·{h 2(x)-Y2}/Yret to the calibration function, to realize a

second-order polynomial linearity correction. From there the flow diagram
becomes a repetitive chain as the encircled part in Fig.3-16 is repeated to
increase the order of the polynomial correction. Because of the repetitive
character, the calibration algorithm can be very well implemented in a
(microcontroller) software routine. But it is also possible to implement the
flow diagram in hardware, digital as well as analog.
3.3.2 Two-dimensional functions

The output of a pressure sensor is determined not only by the pressure
applied to the sensor but it is also affected to a certain extent by the
operating temperature of the sensor, in such a way that the errors, offset,
gain, and nonlinearity are temperaturc dependent [22,23,24 J. Such a
sensor has to be calibrated for both pressure and temperature, hence the
term two-dimensional calibration. The calibration method explained in the
previous section can be extended to handle two-dimensional
calibration [25,26]. The basic principle is to select lines or 'axes' in one
dimension by fixing the value for one input variable, and then proceed
along each selected axis with different values for the other variable, as in
the one-dimensional calibration. The values for the first variable must be
selected according to the same procedure as for the one-dimensional
calibration.
To explain this principle mathematically, we will use normalized input

variable x for the desired sensitivity, and z for the cross-sensitivity
variable. The uncalibrated sensor transfer function is given by y=f(x,z). To
calibrate the cross-sensitivity for signal z, we assume we can make use of

an additional sensor which senses z independent of other variables,

z'=k(z). The sensor transfer function wilI be calibrated by the use of:
1. NxM calibration measurements ofthe sensor output at

different input 'vectors', denoted by f(xmz m ), with n=1 to N
andm=1 toM
2. M simultaneous measurements of the output of the
additional z-sensor, denoted by z'm= k(zm)' m= 1 to M
3. the N desired values for the sensor output signal, according
to the ideal linear transfer curve Yn=g(xn), n=1 to N
After each calibration measurement, a corrected transfer curve hnm(x,z) is
constructed, based on the preceding calibrated transfer functions and a
calibration coefficient anm . Each coefficient can be ca1culated on the basis
of the additional values obtained in the list above. Here, we will only show
the calibration functions, but not the equations for ca1culation of the
calibration coefficients.
We start at a fixed value xI for input variable x. We calibrate offset and the
cross-sensitivity of the offset to variable z, by applying M different
conditions zI to ZM, and ca1culating the calibrated transfer functions
respectively as:
hll(x,z)=!(x,z) + aii (3-33)
M-I
hIM(x, z) = hl,M_I(x, z) + a lM · n (z' -z'm)
m= I
Each coefficient al m can be ca1culated by using one calibration

measurementf(xI,zm)' and the desired output value for hlm(xl,Zm) which is
g(xl,Z)=YI in alI z-points.

The set of equations in (3-33) can be compared with equation (3-22) in

which the reference is left out and the coefficient al is gi ven an M-point
polynomial calibration with respect to variable z.
We have now achieved that, at x=xl, the calibrated transfer function

hIM(xl,Z) equals to Yl more or less independent of z. In the next set of
calibration steps, we will rotate the transfer function around the line
(x=xl,Y=Yl) which runs along the z-axis, see Fig. 3-4 on page 58. This way
we adjust the gain of the transfer function, without disturbing the previous
calibration steps. To calibrate the gain we apply a second reference for the
input variable x=x2' The cross-sensitivity of the gain for variable z IS
calibrated by applying M different conditions ZI to ZM and computing:
n
M-I
h 2M (x, z) = h 2,M_I(x, z) + {l2M' {hIM(x, z) -YI} (z' -z',,)
m=:.: 1
(3-34)
The coefficients a2m can be calculated by using the calibration

measurements f(x2,Zm) and the desired output value in ali points
g(x2,z)=h
Again, the set of equations (3-34) can be compared with the one-
dimensional gain calibration of equation (3-24) in which the calibration
coefficient a2 is given an M-point polynomial cross-sensitivity
compensation for variable z.
Next, a linearity calibration is obtained by adding a second-order

correction term which is zero along the two previous calibration lines

(X=Xl'Y=Yl) and (x=x2'Y=Y2)' and by measuring at a third series of input

conditions X=Xl and z=z 1 to zM:
M-I
h 1M (x, z)
-
= h",-, M-I (x, z) + a 1M
-
. {h l M(x, z) - YI} {h 2M (x, z) - Y2} TI (z' -z'm)
m = 1
(3-35)
In the same manner, the procedure can be continued to further calibrate the
nonlinearity. Similar to the one-dimensional case, new input conditions for
the input variable x need to be used, x=x4 to x=xN in combination with the
z-conditions described before. The Nth and last calibration can then be
described by the following equations:
N-I
hNl(x,z)=hN_I.M(x,Z) +a NI TI {hnM(x,z)-y n }
n= 1
N-I
h N2 (x,z)=h NI (x,z) + (lN2(z'-Z'I) TI {hnM(x,z)-y n }
n=1
N-I
h N3 (x, z) = h N2 (x, z) + (IN3 (z' - Z' 1) (z' - z'2) TI {hnM(x, z) -Y n}

n=1
M-I N-I
hNM(x, z) = h N, M-I (x, z) + (INM TI (z' - z'm) TI {hnM(x, z) - Yn}

m =1 n =1
(3-36)

After aIl the coefficients aii to aNM have been caJculated, the final
calibration function, h(x,z)=hNdx,z), has been given an (N-1)th-order
polynomial linearization for input variable x, and an (M-l )th_order
compensation for cross-sensitivity to variable z. If the linearization has to
be improved, the polynomial order for input x can be increased with an
additional set of (M) equations and measurements as expressed in
equation (3-36). Only the new calibration coefficients need to be
caJculated, the previous ones remain valid. On the other hand, if the order
of the cross-sensitivity polynomial needs to be increased, then each set of
equations (3-33) to (3-36) gets one additional formula. We need N
additional calibration measurements for inputs xI to XN at a new condition
ZM+ l' It can be shown that, despite the nested (recursive) use of the
functions hnm(x,z), the additional calibration coefficients can be caJculated
without having to recompute the previous calibration functions.
Practical implications and results of the two-dimensional calibration will

be presented in chapter 6.
3.3.3 Multi-dimensional functions

In the same way as the polynomial calibration has been extended from one
to two dimensions, it can be extended to a third dimension. AII the
calibration coefficients a nm in equations (3-33) to (3-36) then need to be
calibrated for sensitivity to the third variable in a 'nested' series of
measurements. It is also possible to put 2 two-dimensional calibrations 'in
series', first calibrating for one cross-sensitivity as described above, then
calibrate the resulting transfer curve for the other cross-sensitivity. Besides
the mathematical complications, the practical complication of the NxMxL
calibration measurements with three controlled physical input signals will
make such a three-dimensional calibration very unattractive. Using vector
and matri x expressions it is possible to expand the polynomial calibration
to even more dimensions, but it would be more expedient to concentrate
on elementary ways to reduce the sensor's cross-sensitivity to so many
parameters.

3.4 Conclusion
3.4 Conclusion
The discrepancies that can be found in the sensor transfer have been
distinguished to clarify what errors can be corrected by means of
calibration. Generally, the aim of the calibration is to correct the sensor
transfer for linear errors, like offset and gain (scale), as well as
nonlinearity and cross-sensitivity.
Several linearization methods have been explained and applied to an

sensor transfer function example. Without prior knowledge about the
sensor characteristic, a general curve-fitting method offers the possibility
to linearize the sensor transfer, using a low number of calibration
measurements. The polynomial curve-fitting technique has been adapted
in such a way that the sensor curve can be corrected right after each
calibration step, without affecting previous calibration steps. The
corrected sensor transfer function then progresses towards the desired
(linear) transfer in a step-by-step manner. The proposed method can be
extended to two-dimensional transfer functions, which enables a
systematic calibration of linearity and cross-sensitivity.

References
[1] J.W. Gardner, Microsensors: principles and applications, John
Wiley & Sons, New York, 1994.
[2] W. Gopel, J. Hesse, and J. Zemel, "Sensors: A Comprehensive Study,
Vol.l: Fundamentals and General Aspects", T. Grandke & W.H. Ko
(Vol.eds.), VCH, Weinheim, 1989.
[3] J.H. Huijsing and J.A. van Steenwijk, "A monolithic analog
exponential converter", IEEE Journal of Solid-State Circuits, VoI. 15,
No.2, 1980, pp.162-168.
[4] J.E. Brignell, "Digital compensation of sensors", Journal of Physics,
VoI. E: Scientific Instrumentation, 1987, pp. 1097-1 102.
[5] J.E. Brignell, "Software techniques for sensor compensation",
Sensors and Actuators A, VoI.25-27, 1991, pp.29-35.
[6] P.P.L. Regtien and P.J. Trimp, "Dynamic calibration of sensors using
EEPROMs", Sensors and Actuators A, Vo1.21-23, 1990, pp.615-618.
[7] PN. Mahana and F.N. Trofimenkoff, "Transducer output signal
processing using an eight-bit microcomputer", IEEE Trans.
Instrumentation and Measurement, Vol.IM-35, No.2, June 1986,
pp.182-186.
[8] C. de Boor, "A practical guide to splines", Applied Mathematical
Sciences, Vol.27, Springer, New York, 1978.
[9] P. Dierckx, Curve and surfacefitting with splines, Oxford University
Press, Oxford, 1993.
[10] S.B. Crary, w.G. Baer, J.c. Cowles, and K.D. Wise, "Digital
compensation of high performance silicon pressure transducers",
Sensors and Actuators A, VoI. 2 1-23, 1990, pp.70-72.
[Il] S. Huang, R.Z. Morawski, and A. Barwicz, "Static calibration based
on superposition of splines in one variable", Proceedings IMTC'96,
June 1996, pp.49-53.
[12] C. Berthoud, M. Ansorge, and F. Pellandini, "Effective static
response compensation suitable for low-power asic implementation
with an application to pressure sensors", Proceedings IMTC'96, June
1996,pp.1168-1173.

3.4 Conclusion
[13] M.E. Snow and S.B. Crary, "The use of simulated annealing in the
I-optimal design of experiments", Michigan Academician XXIV,
1992, pp.343-354.
[14] S.B. Crary, L. Hoo, and M. Tennenhouse, "I-Optimality algorithm
and implementation", Computational Statistics, Y. Dodge and
J. Whittaker (eds.), Proceedings of the 10th Symposium on
Computational Statistics, Vol.2, Neuchâtel, Switzerland, 1992,
pp.209-213.
[15] S.R. Ashley, M. Muggeridge, and J. Lucas, "An inexpensive digital
Iinearizer for nonlinear transducers", Journal of Physics E:
Sci. Instrum., VoI. 11, 1978, pp.576-580.
[16] w.T. Bolk, "A general digital linearising method for transducers",
Journal of Physics, VoI. E: Scientific Instrumentation, 1985,
pp.61-64.
[17] D. Patranabis and D. Gosh, "A noveI software-based transducer
Iinearizer", IEEE Trans. Instrumentation and Measurement, Vo1.36,
No.6, December 1989.
[18] P. Malcovati, c.A. Leme, P. O'Leary, F. Maloberti, and H. Baltes,
"Smart sensor interface with AID conversion and programmable
calibration", IEEE Journal of Solid-State Circuits, Vo1.29, No.8,
August 1994, pp.963-966.
[19] M. Yamada and K. Watanabe, "A capacitive pressure sensor
interface using oversampling ~-L demodulation techniques", IEEE
Trans. Instrumentation and Measurement, Vo1.46, No.l, February
1997, pp.3-7.
[20] M. Gunawan, G.C.M. Meijer, J. Fonderie, and J.H. Huijsing, "A
curvature-corrected low-voltage bandgap reference", IEEE Journal
of Solid-State Circuit~, Vol.28, No.6, 1993, pp.667-670.
[21] S. Kaliyugavaradan, P. Sankaran, and VG.K. Murti, "A new
compensation scheme for thermistors and its implementation for
response linearization over a wide temperature range", IEEE Trans.
Instrumentation and Measurement, Vo1.42, No.5, 1993, pp.952-956.
[22] S. Kim and K.D. Wise, "Temperature sensitivity in silicon
piezoresistive pressure transducers", IEEE Trans. Electron Devices,
VoI. ED-38, 1983, pp.802-81 o.

[23] H.-J. Kress, F. Bantien, J. Marek, and M. Willmann, "Silicon

pressure sensor with integrated CMOS signal-conditioning circuit
and compensation of temperature coefficient", Sensors and
Actuators A, VoI.25-27, 1991, pp.21-26.
[24] M. Akbar and M.A. Shanblatt, "Temperature compensation of
piezoresistive pressure sensors", Sensors and Actuators A, Vo1.33,
1992, pp.155-162.
[25] K.F. Lyahou, G. v.d. Horn, and J.H. Huijsing, "A non-iterative
polynomial 2-dimensional calibration method implemented in a
microcontroller", Proceedings IMTC'96, June 1996, pp.62-67.
[26] G. v.d. Horn and J.H. Huijsing, "Integrated smart sensor
calibration", J. lntegrated Analog Circuits and Signal Processing,
Vo1.I4, No.3, November 1997.

Calibration
using Analog
4
Signal Processing
4.1 Introduction
In this chapter we will focus on calibration techniques based on analog
signal processing. Almost all sensors require some kind of analog signal
conditioning or amplification before the sensor signal can be converted
and proces sed digitally. Basically, sensors produce an analog output signal
only. But even for modern smart sensors with digital output, a
standardized analog output signal is often desired as well. Therefore, we
will first consider the possibilities of adjusting the transfer of an analog
circuit for the purpose of calibrCltion.
Adjustable elements, like potentiometers or laser-trimmed resistors, can be

used. An example will be explained in section 4.2. As mentioned in
previous chapters, it is preferable to realize a digitally programmable
calibration. This would allow batch calibration and in-situ calibration,
and thus simplify automation of the (factory) calibration. It would also
enable calibration of the sensors after packaging and allow (customer)
re-calibration, and thus improve sensor performance. The laser-trimmed
resistors in the conventional calibration circuit could be replaced by
resistor arrays with switches, but instead we prefer to make use of Digital-

Calibration using Analog Signal Processing
to-Analog converters, as will be explained in section 4.3. Pulse-width and

pulse-density modulation are very promising techniques for calibration, as
will be shown in section 4.4. The examples in those sections will be
limited to calibration for offset, gain, and the linear temperature
coefficients of offset and gain. In section 4.5, it will be demonstrated that
the polynomial linearization explained in the previous chapter can also be
implemented in a programmable analog circuit.
4.2 Conventional sensor calibration

Before calibrating the transfer of a sensor electronically, other options to
improve the sensor characteristics should be considered first; the less the
size and number of errors to be calibrated the better the final performance
will be. If cross-sensitivity to another physical parameter can be
eliminated systematically in the sensor construction, this is preferable to
cross-sensitivity compensation that requires calibration. Nevertheless
some errors wilI always remain, and require calibration. Often the sensor
output can only be related to the physical input signal through a
calibration. The transfer may then be corrected by processing the electrical
output signal, or for some sensors (modulating sensors), by modifying the
electrical reference signal.
4.2.1 Resistor trimming

In discrete electronics, most passive elements are also available in
mechanicaIly adjustable implementations: potentiometers, variable
capacitors, etc. An advantage of such elements is that they have inherent
analog memory: potentiometers stay in their last position. Comparable
adjustments in Integrated Circuit (IC) technology make use of laser
trimming. A laser beam can be used to modify the geometrical properties
of a resistor on the chip surface, in order to adjust the resistance value. An
obvious disadvantage of the laser-trimming technique is that it al ters the
resistors permanently.
For example, with sensors based on a resistive bridge configuration it

makes sense to use trimmable resistors. Fig.4-l shows a Wheatstone
bridge in which some extra resistors, R, to R6 , have been inserted [1,2,3].
The resistor combination R" R 2 makes it possible to adjust the bridge
balance to calibrate offset. Resistor R3 is a temperature-dependent resistor

4.2 Conventional sensor calibration
fu II-scule Te
bridge + R6
supply instrumentation
voltage - fu II-scule amplifier
normuflzution
Fig. 4-1 Wheatstone bridge with resistor compensation.
which may be adjusted to compensate for the offset temperature

coefficient. The voltage di vider Rs, R6 realizes an adjustment of the bridge
reference voltage and its temperature gradient, so that gain errors can be
compensated and the full-scale output (FSO) can be normalized. The
voltage divider can only reduce the bridge voltage and thus decrease the
signal. With normalization of the FSO, aII sensors will thus be adapted to
the one with the lowest sensitivity. It is preferable to obtain a gain-
adjustment without reducing the bridge reference. This could be done with
a laser-trimmed feedback resistor in an instrumentation amplifier as shown
in Fig.4-1.
Disadvantages of laser-trimming calibration techniques are:
• the adjustments are not independent of each other (the offset

resistors also affect the gain)
• resistors can often only be changed one way, usually
resistance is increased
• the trimming is irreversible, so no re-calibration is possible
• in-situ calibration (under test) is difficult or impossible
• automation in afabrication process is difficult
• it is a time-consuming expert operation, thus expensive
An advantage of laser-trimming, and also a reason why it is stil! common
practice, is that, besides the resistors, no other elements need to be
integrated with the sensor. If enough research is done to make le
technology and sensor technology compatible, the integration of active
elements with the sensor may not be a problem in the future. Smart sensor

electronics wiII then provide more attractive solutions than laser-trimming

techniques.
4.2.2 Programmable resistor-arrays

A direct approach to overcome some of the disadvantages of laser-
trimming calibration is to replace the trimmable resistors with
programmable resistor arrays. A requirement, then, is that switches, such
as MOS transistors, can be integrated with the sensor (interface), and their
position/state can be stored in an on-chip (permanent) memory. An
example of an instrumentation amplifier with programmable gain is shown
in Fig.4-2. The feedback is built-up with an array of binary-scaled
Fig. 4-2 lnstrumentation amplifier with programmable feedback

resistor.
resistors that can be switched in paraIleI. The result is that the voitage gain
of the amplifier can be programmed digitaIly, in this case with a 4-bit
resolution. Likewise, other trimmable resistors in conventional
configurations could be replaced with programmable resistors. However, if
IC-technology is available, it is better to investigate more advanced circuit
techniques rather than to adapt conventional sensor calibration techniques
and inherit many of their disadvantages.
4.3 Analog calibration circuits

When an analog signal has to be inftuenced in a programmable way the
use of some sort of Digital-to-Analog (DA) conversion is unavoidable'.
Despite the many DA-conversion techniques that exist and are being

developed [4], we will concentrate on the use of a conventional

multiplying DA-converter (DAC) configuration.
The principle of this DA-conversion is based on the creation of binary-

scaled portions of an analog reference signal, which can then individually
be switched on or off by the bits of the digital input. The parts that are
switched on need to be summed together in the analog signal domain
somehow. Current signals can be summed easily on a common low-
impedance node, whereas voltage signals have to be made time-discrete
(sampled) before they can be summed in a switched-capacitor circuit.
When sampling of the analog signal has to be avoided, it is logical to
create binary-scaled currents and use current switches. Fig.4-3 shows a
simple 4-bit example of a multiplying DAC in both CMOS
implementation and bipolar implementation. Instead of connecting an
analog reference, we apply an input signal. The operation will be
explained in short, so that the application of the DACs in the next section
will be clear.
do
Jt---t--- d 1
~--+--+-- d2
~--+--+-+-- d3
R R R
Fig. 4-3 Simple 4-bit multiplying DAC examples, using CMOS

voltage switches (left), or bipolar current switches
(right).
I unless rehable analog electric al memories become available

The CMOS circuit on the left uses a resistor ladder network to obtain
binary-scaled currents. Depending upon the digital bit values, the vertical
resistors are either short-circuited to ground by the PMOS transistor (bit
low) or to a virtual ground node by the NMOS transistor (bit high). The
effective resistance of the ladder network from input to ground equals R.
The input voltage signal results in an input current IIN=VIN/R which scales
down by a factor of two at each successive node in the ladder. The input of
an IV-converter can be used as a virtual ground and summing node for the
binary parts.
In the bipolar version, a cascode with scaled transistors is used to create

the binary portions of the current. If resistor matching is better than
transistor matching, the emitter resistor network as shown will improve
the accuracy of the binary division. The currents are switched by
differential pairs which are controlled by digital signals with limited
amplitude.
The formula in Fig.4-3 shows that the analog input current is multiplied by
the digital number applied at the switches. The next section will show how
such multiplying DACs can be used for calibration of the transfer of an
analog sensor read-out circuit.
4.3.1 Classical offset and full-scale calibration

The multiplying DACs can only reduce the signal. Therefore, it is
necessary to first amplify small sensor signals to a signal level which is
high enough to leave some 'headroom' for processing. In other words, the
dynamic range between signal amplitude and noise+distortion must be
sufficient before applying the signal to the multiplying DACs.
Fig.4-4 shows a straightforward application of the DACs to make the gain

and offset of an analog read-out circuit programmable. The sensor output
signal V~ell.w,., presumably a voltage l , is amplified or buffered at the input
of the circuit and the voltage signal is then converted by the DAC into a
current. The magnitude of the current is controlled by the switches of the
DAC. The position of the switches can be stored in a very small and
simple digital memory (register), which is not shown in the figure. The
memory should be programmable and preferably non-volatile.
J Other kmd of sensor signals could be converted to a voltage signal, mstead of Just
amphfied

r------~---,
I R J Ro I() I
VOUT
instrumentation
amplifiers
Fig. 4-4 Analog calibration circuit, using multiplying DACs for a

programmable gain and offset control.
The output current of the DAC is converted back into a voltage by the
IV -converter. The digital number applied at the switches of the DAC
controls the gain of the transfer. When there are no other signal paths, the
gain can be varied over the full range, from zero to the maximum. Often
the gain needs only a small adjustment relative to the full range. This can
be achieved by connecting a resistor Ro across the DAC, as illustrated with
the dashed wires. Now an additional current 10 proportional to the sensor
signal will contribute to the output voltage. Depending on the ratio of R[
and Ro, only a fraction of the full-scale signal (gain) can be varied.
Correction of offset of the transfer requires a reference which is

independent of the sensor signal. The reference voltage VREF is treated in
the same manner as the input voltage which results in a programmable
current h The current is added to the other signals at the low-impedance
input node of the IV -converter, and thus realizes a shift of the output
voltage level. The offset correction range depends upon the value of the
reference voltage and the ratio of R2 to R[. Here, too, a resistor could be
placed across the DAC in order to create an additional prefixed offset
correction (not shown in figure).
4.3.2 Classical cross-sensitivity calibration

Clearly the circuit construction of Fig.4-4 allows the addition of many
correction signals at the input of the IV -converter. For example, this can be
done to realize a correction for cross-sensitivity to temperature, provided

that a signal is available which depends linearly on the operating

temperature of the sensor device only, and not on the sensor signal itself.
Such a signal can be generated by using a PTAT source or a resistor with a
linear temperature coefficient 1. As pictured in Fig.4-5, the temperature
signal VŢ can also be controlled by a programmable DAC and added to the
other signals. This resuIts in a temperature effect for the offset of the
circuit's signal transfer, which can be used to compensate for the offset
temperature coefficient (TC) of the sensor. The correction range for the
offset TC is set by the amplification A4 and the choice of R 4 . For the
correction of the TC of the gain, a correction signal is used which depends
on the input signal and is Iinearly proportional to the temperature. This
correction signal is obtained by muItiplication of the temperature signal VŢ
and the sensor signal V~ensllr in an analog multiplier circuit. The multiplier
output voItage signal is also converted to a current, controlled by a DAC,
and added to the other signals.
IV -converter
I\um VOUT
Fig. 4-5 Analog circuit with multiplying DACs for offset, gain,
gain TC, and offset TC control.
I For plezoresistive bridges with a constant current excitation, sometJmes the bndge
voltage can be measured and used as a temperature signal

In principle, the four DACs control the four different errors, offset, offset
TC, gain, and gain TC independently. But the different correction signals
may also have transfer errors which need to be measured. Therefore, in the
calibration procedure, each of the four signal contributions has to be
measured separately at two different physical input signals and at two
temperatures (2x2 conditions). This can be do ne by respectively setting
each DAC on maximum and the other three on zero. These 2x2 x4 output
measurements can be used to characterize ali four signal contributions.
Then the required values for the DACs can be calculated and programmed
in the DACs, so that the desired calibration is obtained.
The schematic of Fig.4-5 can be worked out in a bipolar implementation.

In that case, we prefer to make use of current multiplying DACs, see the
beginning of section 4.3. The input amplifiers must then be replaced by
VI-converters, the current signal of a PTAT source can be used as
temperature signal, and a (Gilbert) multiplier with current output can be
applied for multiplication. The idea is i\1ustrated in Fig.4-6.
Isensor
VI-converter current DAC
VREF~-+--1 VI-converter 1-.....---1 currentDAC
Gilbertmultiplier
current DAC
Fig. 4-6 Bipolar implementation using current-mode DACs.
In the schematics of Fig. 4-4 to Fig.4-6 we have ignored the fact that the
errors may have a different polarity than expected. For example, the TC of

the offset can be either positive or negative. By realizing the calibration

circuit in a dijferential implementation, it becomes possible to give the
correction signals a positive or negative sign. The input amplifiers are
replaced by differential instrumentation amplifiers, and the output
IV-converter is replaced by a differential equivalent, see also section 4.5.
The programmable output signals of the DACs can then be switched to
either the positive or the negative input of the IV-converter to control the
polarity of the correction. A differential circuit also offers an improved
performance with respect to the common-mode and supply rejection
ratios.
4.3.3 Improved analog calibration circuit

In the analog calibration circuits proposed in the previous section, the
calibration settings are not independent of each other. This means that the
offset has to be re-adjusted after a gain calibration, and the other way
round. Unless the individual signal contributions are measured separately,
and the required calibration factors are ca\culated externally, it turns into
an iterative calibration.
According to the progressive polynomial calibration method introduced in

chapter 3, this problem can be overcome. Using the equations for the
transfer functions in a 2x2 calibration:
h ll (x,z) =j(x,z) +a ll
hlz(x,z) =hll(x,z) +a lz ' (Z'-Z'I)

(4-1)
h ZI (x,z) = hlz(x,z) +a zl ' {hlz(x,z) -YI}
hzz(x,z) =hzl(x,z)+a zz ' {hlz(x,z)-YI}' (z'-Z'I)
with x as the sensor input signal, and z' the measured temperature signal,
we can construct the desired transfer of the calibration circuit.
For simplicity, we choose the first sensor calibration point at the input
level SI where we want the output signal to be zero, so YI=O. Suppose,
furthermore, that we can make use of a voltage signal VT-T1 which is zero
at a known temperature TI' Next, the functions in equation (4-1) are

replaced by voltage signals. Then, we can derive that the following

transfer is desired:
This transfer can be implemented in an analog calibration circuit as

illustrated in Fig.4-7. It is, in fact, a slight modification of the original
circuit shown in Fig.4-5.
R
IV -conver!er
11111 IV -conver!er
112 122
2R
2R
Fig. 4-7 Alternative analog circuit for independent calibration of

affset, gain, and temperature coefficients.
The calibration can now be executed in the manner of the progressive

calibration method. First, aII the 'DAC coefficients' are set to zero. At
temperature TI' and a zero physical input signal 5 I for the sensor, DAC
coefficient a II is adjusted in such a way that the desired zero output
voltage is obtained. Second, at another temperature T2 , al2 is adjusted to
compensate the temperature gradient in the offset of the sensor. Now the
signals 1/ 2 and V/2 are zero for a zero sensor input signal. At the next
calibration step another sensor input 5 2 is provided, and at temperature TI
the gain can be adjusted with a21' At the other temperature T 2 , the
temperature coefficient (TC) of the gain can be compensated with a22' In
that way, the calibration settings can be controlled separately. This is
achieved because, independent of the DAC settings, the signal

contributions for the gain corrections are zero when the sensor input signal
is equal to the calibration point SI, and the signal contributions for Te
compensation are zero at temperature TI.
4.4 Classical pulse-modulated calibration

The performance of the multiplying DACs depends on the matching of the
transistors and resistors used. In particular, it is the monotonicity of output
versus increasing digital code that is affected by mismatches [4]. Without
taking special measures the highest resolution for which monotonicity can
be guaranteed is about 7-8 bits. With special transistor and resistor layouts
to improve matching, the resolution may be improved to 10 bit. Further
improvements require more advanced techniques such as dynamic element
matching.
As long as the maximum correction signals are only a small fraction of the
full-scale signal, the limited resolution of the calibration is not such a
problem. If high resolutions are desired, we need another type of
DA-conversion, based on scaling in the time (or frequency) domain.
Instead of creating binary divisions of the signal, we create (binary) scaled
time slots in which the signal is turned on or off. Using techniques known
as pulse-width modulation and pulse-density modulation we multiply the
analog signal with a digitally controlled high-frequency pulse signal [5].
The low-frequency average of the pulsed analog signal can thus be
accurately controlled by the digital signal. The resolution is now
determined by the time accuracy of the pulses and the maximum pulse
width. These properties relate respectively to the stability of the digital
c10ck signal (jitter) and the required analog signal bandwidth.
The analog calibration circuit of Fig.4-5 can be very much simplified by

replacing each multiplying DAC by a single resistor or VI-converter and a
(current) switch. This is displayed in Fig.4-8. Each switch is controlled
separately by a digital pulse train or stream. The four pulse trains regulate
the average contribution of the current signals IJ to 14 to the input of the
IV-converter. The averaging function can be realized by placing a
capacitor across the feedback resistor of the IV -converter. The low-pass
filtering achieved suppresses the high-frequency contents of the pulse
trains. Although the analog circuit is simplified, the complexity is now
moved to the digital circuit which has to generate the four pulse train
signals. There are several options for the implementation of a digital pulse

4.4 Classical pulse-modulated calibration
R)
v,\en.\or
~ 1,
_nnnJl_ gain
low-pass filter
R2 IV-converter
VREF 12
~
_JULJlIl_ offset
Vour
... --, VTxV\'en.ltJr
I,um
VREF R3
~ 13
voltage
multiplier _1U1JUUL gain TC
Vr R4
~ 14
_nnrL_ offset TC
Fig. 4-8 Analog calibration circuit using switched-current DACs.
generator which creates a digital l-bit pulse signal of an n-bit digital

number. One option is to use the digital equivalent of a sigma-delta
converter, which outputs a so-called bitstream signal. This technique will
be elaborated in chapter 5 and combined with a sigma-delta AD-converter
in place ofthe IV-converter.
The circuit of Fig.4-8 should preferably be implemented in a differential

version for the same reasons mentioned in the previous section. But also to
improve resistance against the digital interference, which is expected to
increase considerably.
Furthermore, it must be noted that there is no longer a time-continuous

analog signal path; the analog signals are interrupted by the switches. To
prevent interference and aliasing effects, the bandwidth of the analog
signals must be limited in such a way that it does not overlap the frequency
band of the pulse-train signals.

Calibration llsing Analog Signal Processing
4.5 Analog implementation of a polynomial calibration

The circuit examples discussed so far have shown the possibilities of
correcting linear errors in the transfer of an analog circuit in a very
straightforward way [6]. As discussed in chapter 3, there are several
methods to correct the nonlinearity of the sensor transfer. A very effective
linearization could be obtained with the presented progressive polynomial
calibration method. In this section we will describe a complete analog
implementation and realization of a four-step linearity calibration. For
clarity, the signal flow diagram explained in section 3.3 is repeated below
in Fig.4-9.
3
li (h 3'\3){hn2){h 1,\J1
4 )'rc! Yrcl
Yref
Fig. 4-9 Graphical diagram for four steps of the polynomial

calibration method.
4.5.1 System configuration

To demonstrate the feasibility of an analog hardware implementation, the
flow diagram of Fig.4-9 has been implemented as a programmable analog
circuit in a low-cost bipolar process [7,8]. Fig.4-1O shows the block
diagram of the analog part. When compared carefully, it shows great
similarity with the diagram of Fig.4-9. AII signals are represented by
differential currents which are superimposed on common-mode bias
currents. Current signals in(V1N ) in Fig.4-1O correspond to calibration
functions hn(x) in Fig.4-9; currents In correspond to calibration points Yn'
and current I REF to Yref'
The input voltage VIN, coming from the sensor output, the reference signal
Vas, and the set-points for the calibration VI, V2 , V3 are aII converted to
differential current signals, using Voltage-to-Current (VI) converters. At

4.5 Analog implementation of a polynomial caIibration
currenl dupltcalors 1 10 5
R
'lli "
"'
"
2
VOI'!
A,
V, V,
R R
Fig. 4- \O Programmable analog signal processor, implementation

ofFig.4-9.
the output, a Current-to-Voltage (IV) converter turns the processed current

back into a differential voltage signal. The signal processing is
implemented in current-mode circuitry. The required processing functions
are listed below.
• The adding and subtracting is done by correctly connecting

the differential current signals (Kirchoff's law).
• The multiplication ofthe analog current signals is
implemented in a modified Gilbert multiplier, a well-known
translinear circuit, using only a few transistors [9].
• The calibration coefficients a J to a4 are implemented as
multiplying Digital-to-Analog Converters (DAC), according
to the principle explained in section 4.2.2. The DACs are
combined with programmable digital registers which are
connected to the digital inputs of the DA Cs, indicated by A J
to A 4. The digital registers are not shown in Fig. 4-1 O.
• As can be seen in theflow diagram ofFig.4-9, a duplication
ofthe signal is necessary at the nodes numbered 1 to 5. This
is implemented in current duplicator blocks. lnside each of
INTEGRATED SMART SENSORS: Design and Calibratiofl \07

these blocks an IV-converter turns the current into a voltage

which is then used to generate two differential current
signals, using two Vl-converters, as explained in Fig.4-IO.
In the next sections the implementation of these function blocks wiIl be
elaborated. The operation of the complete system wiIl be further explained
in section 4.5.6.
4.5.2 Analog current multiplier

In bipolar technology, it is possible to make use of the weIl-known
exponential relation between collector current and base-emitter voltage in
so-called translinear circuits:
(4-3)
A special cIass of circuits is based on the principle that when the sum of
base-emitter voltages of a set of transistors is constant, this also applies for
the product of the collector currents [9,10]. An advantage of this principle
is that it is valid for a very large range of values, because the exponential
relation is a physical property of the bipolar transistor. The best known
example of a translinear circuit is probably the current multiplier.
A variant of the translinear multiplier for the multiplication of two

differential current signals is shown in Fig.4-11. It consists of two
transistor mazes, one with Q 1, Q2, Q3, Q4 and another with Q 1, Q2, Q5, Q6.
The transistors are connected in such a way that the sum of base-emitter
voItages on both sides of a maze are equal, thus V be3 + V be2 equals
V be /+V/Je4 and V/JeS+Vbe2 equals V/Jel+V/Je6. Using equation(4-3), it
results in the foIlowing translinear relation:
At aII four input terminals at the bottom of the circuit, currents are pulled
down to the negative supply or ground. The bias voltage at the base of Q l
and Q2 determines the voItage level at the current inputs.

+
V b1as
Ilnl+
X
Ilnl- Iln2- 11112+
Vhta,1 02 01
03 04
05
Fig. 4-11 Four-quadrant current-mode multiplier.
Because of the differential set-up, each input current consists of two parts,
lin+ and Itn_. Each part is made up of a common-mode component, lin' and
a differential component, M,n' When neglecting the base currents, we see
for inputJ that ItnJ+=Ie3+le4""1(]+lc4 and linJ-=leS+le6""lcS+lc6 and for
input2 that Itn2+=/in2+f'l.I,n2""/d and lin2-=ltnrf'l.ltn2""lc2' which can be
worked out in combination with equation (4-4) into:
~1,n2)
1 ( 1+
=_. I14-21 ( 1 -~/m2)
I
c3 2 I
- - ·1
ml+
-_o
I- - ·1 ml+
m2 m2
I(6 =!.
2
( 1 - ~1,n2)
I
. Iml- I1'5 1 (
=_. ~/m2)
2 1 +I - - ·1 In 1-
m2 m2
(4-5)
Furthermore, we see in Fig.4-11, for the two parts of the output current
that lout+=Ic3 +lc6 and lout-=lc4 +lcS ' When substituting linJ+=linJ+f'l.lmJ
and ItnJ-=ltnrf'l.ltnJ into equation (4-5) we find for the output currents:
~I 2
Illut+ = I mi + ~I mi . __
m_
I (4-6)
m2
~I
I out- = I mI - M mI
. ---.!.!!.3
I
m2

With Mout=/oUl+ -/our- it simply folIows for the differential output current
that:
2. I'l./ml . l'l./m2
(4-7)
/m2
So, using a small six-transistor circuit, we have obtained a differential

output current which is proportional to the product of the two differential
input currents scaled by the common-mode component of the second input
current. From the circuit it can be found that the common-mode
component of the output current is practicalIy equal to the common-mode
of the first input current.
The multiplication property is tested by means of PSPICE simulations.

Fig.4-12 shows the simulation results. The first picture shows the
40:'
-20; to
out
=
M
(IlA) /
.
L'l./1I/2=1-16.-12.-8.-4.0.4, 8, 12.1611lA
-40. . . . -40. .
5
"
-16 -8 O MInI (JlA)--+ 16 O Tlme (ms) --+ 10
Fig. 4-12 Simulation results showing the operation of the

multiplier.
differential output current as a function of one differential input current for

different values of the other differential input current. The second picture
shows the two differential input currents sweeping between plus and
minus 1611A, and the resulting output current. Clearly, the multiplication
works in alI four quadrants, i.e. for negative and positive values of both
differential input currents. The common-mode input currents were biased
at 20I1A.

4.5.3 Programmable current divider DAC

The programmable calibration coefficients are implemented in the form of
so-called multiplying DACs. As explained in section 4.2.2, a DAC can be
used to obtain a multiplication of an analog signal with a digital number.
By connecting the DAC's digital input to a programmable digital storage
register, the desired function is realized. Fig.4-13 shows the 8-bit
DA-converter designed for differential current signals.
I I
0[7-IIJ
I I I I I I I I I I I I I
Vhwll
07~ ~ ~~ ~ ~~ J~ ~ ~~ ~oo OO~ ~~ ~~ ~~ ~~ ~~ ~~ ~~ ~I07
h;+~ 11;+. 11;+. I~~++ I~~++ I~~++ ~+ ~~;+ ~~;+ .~ 256 .~
256 .~ 128 .~ .~ +II;~ I~- n·
."
•
64 • .!!!!.:
32 16 T
V!)/(J\2
8x 4x 2x 1x 1x 1x 1x 2x 4x 8x
Vhwd
8x 4x 2x 1x 1x 1x 1x 2x 4x 8
2R Q2R 2R Q2R 2R 2R 2R Q 2R 2R 2R
R R R R R R
Fig. 4-13 8-bit differential current-mode DAC
The currents on the positive input I tn + and on the negative input l/fl- are
divided into 8 binary-scaled parts, using cascode dividers [4]. Each binary
part is switched to either the positive or the negative output by using a
differential pair. The circuit is configured in such a way that the
differential input current M,"=(I111+ -I,"J is multiplied with the two's
complement value of the 8-bit digital signal D[7-0J applied at the switches.
This is obtained by reversing the connections of the differential pair for the
most significant bit (MSB) D 7 , and by connecting the remainder of the
current division ...!... M in opposite phase to the output. When the base
256 In
currents of the transistors are neglected, the following transfer for the
differential current signal can easily be derived:
M oul
M III
(4-8)
INTEGRATED SMART SENSORS: Design and Calibration III

in which the digital bits Dn can have a value of +1 (high) or -1 (low). The
differential pair switches in the DAC in Fig.4-13 are aII biased with one
base to a reference voltage V blas 1. The other side of each switch either gets
a lower or a higher voltage. As a differential voltage of 200m V is
sufficient, the switches can be controlled by the output of a digital ECL
(Emitter-Coupled Logic) circuit. The other bias voltages, V bIas2 and
V bIas3' need to be set at such a level that saturation of the transistors is
prevented. V bIas3 can be used to control the voltage level at the current
inputs.
To allow a normal interpretation of the digital bits with values of 1 and O,

equation (4-7) can be rewritten as:
M()Uf _ _ 1 + D +! D +! D +! 1 D +~ 1 1
~- 7 2 6 4 5 8 D 4 + T6 3 32 D 2 +6;ţ D I + 128 Do
In
(4-9)
The range of the differential output current spans from -1 M In to
+ 128
127 M
In
. The digital signals D[7-01 can now be interpreted as a digital
two's complement encoded 8-bit number with value d. Which makes it
possible to simply write the transfer as:
(4-10)
The digital number can have positive as well as negative polarity. This,
naturally, also applies for the differential input current, thus, again, a
four-quadrant multiplication is obtained. It is cJear from the schematic that
the common-mode output current is practically equal to the common-
mode input current.
Because of transistor and resistor mismatch, the scaling in the binary

dividers of the DAC is not perfect. This Iimits the resolution for which
monotonicity can be guaranteed [4]. For this reason a low resolution of
8 bits was chosen for the experimental realization. Higher resolutions, 10
to 12 bit, would be possible when taking special measures for improved
component matching. More advanced circuit techniques (e.g. dynamic
element matching) are necessary when resolutions in the order of 12-16 bit
are desired. This is not elaborated because it would complicate the design
of the complete system considerably.

4.5.4 VI-converters and IV-converters

For the VI-conversion that is needed at the voltage inputs of the system, a
standard cell VI-converter was available in the analog library. The
schematic is shown in Fig.4-14. It originates from [4]. Transistors Ql and
Fig. 4-14 V/-converter used for converting the input voltages.
Q2 deliver the output currents, while Q7+Q3 and Q8+Q4 perform a

voltage following action from the input voltage to the resistor R 1. The
local feedback loops with Q 1,Q3,Q5 and Q2,Q4,Q6 are constructed in
such a way that the currents through Q3,Q7 and Q4,Q8 are, in first
approximation, independent of the signal. The differential input voltage is
copied exactly (besides ofIset) to the emitters of Q 1 and Q2, across the
feedback resistor RI. The differential output current, b./out=/out+ -/out-, is
twice as large as the current through the resistor, which is simply
determined by the input voltage difference and the resistor value,
neglecting the base currents of Q 1 and Q2:
2. ilV
__ In
(4-11 )
R
Because of the constant current through the voltage following transistors,

the linearity of the circuit is considerably improved with respect to a
normal degenerated differential pair. Another advantage of the circuit is
that the input common-mode voltage can be biased on "ground-Ievel"; one
voltage input can be connected to the negative raiI (ground), while the
voltage input signal is applied at the other input.
INTEGRATED SMART SENSORS: Design and Calibratiofl 113

For the current duplication blocks, see section 4.5.1, a combination of one
IV-converter and two VI-converters is used. In fact, this combination
forms a sort of current mirror, but then only for the differential current
signal. The IV-converter turns the differential input current into a voItage
which is then used by two VI-converters to generate two differential
current signals.
Fig.4-15 shows both converters. The VI-converter, an the right, is just a

differential pair with Darlington transistors and a degeneration resistor.
) 11 12)
20u 20u
Fig. 4-15 IV-converter (left) and VI-converter (right), as used in

the current duplicators.
The transfer of the VI-converter is not linear over the full signal range, but
this problem is tackled by building the IV-converter around the same kind
of differential pair. As shown on the left side of Fig.4-15, the same
differential pair is used as a feedback from the output voltage to the input
current. The loop is c10sed by the cascode transistors Q5 and Q6. Thus, the
IV-converter has exactly the reciprocal transfer of the VI-converter. A
common-mode control loop is made by splitting the resistor in two, and
inserting Q7 and Q8. This is necessary to fix the common-mode output
voltage level, independent of the common-mode input current.
The simulation result displayed in Fig.4-16 shows the IV -VI-transfer from

differential current signal to a differential voltage, and then back to a
differential current for the case that the IV-converter output is connected to
the VI-converter input. The feedback resistors in the IV-converter
114 INTEGRATED SMART SENSORS: Design and Calibratioll

t :: ...........................................................................................................]
Error: t. .
(%) i mearltyerrar
t 4~1·····n . . n. n. n. . . . . . n.nnnn . . . . n.
.2 1...........................................................................................................
n~
Mout~
(ILA) i Mout(VI-converter) ii
-40· ............................................................................................................. ;
t ' ,
1 2 r····················································· ....................................................... 1
, ,,
,, ,
oi i
~Vout i
(Voit) i ~ v"lUt(IV-converter) ,
--12' ............, ............ ,............. , ............. ,............. , ............., ............ , ............. ;
-40 -20 O M (IIA) 20 40
ltl ,.. ----.
Fig. 4-16 Total transfer of I-to- V and V-to-I conversion.
(RI+R2) and in the VI-converter (RI) were given equal values (50 kQ).
The upper graph shows the relative error between differential input and
output current, before and after a gain correction. We see that the linearity
error in the current duplication wiIl be small enough ( «±0.2%= 1;2 LSB
for 8 bit) when the current range is limited to about [-30IlA, 30IlA].
In the current duplicator blocks used in Fig.4-lO, two VI-converters are

connected to the outputs of the IV-converter, so that two 'copies' of the
differential current are obtained. By choosing different values for the
feedback resistors, the output currents can be scaled up or down. This
makes it possible to set the error-correction-range of each of the
calibration steps.
The common-mode output current of the duplicator blocks can be adjusted

independently of the input current. The common-mode (CM) voltage level
at the current inputs is determined by the bias voltage at V bws in Fig.4-15.
In aII described function blocks, the CM-voltage level at the outputs has to
be higher than the CM-voltage level at the inputs, with the exception of the
IV -converter shown in Fig.4-15. Here, the CM-Ievel can go down. This is
necessary to avoid a high 'stack' of function blocks and the need for a
large supply voltage.
INTEGRATED SMART SENSORS: Design and Calibratioll 115

4.5.5 System implementation

To further cJarify the system implementation, the schematic of Fig.4-1 O is
recalled in the figure below. The function blocks are implemented as
described in the previous sections. Only the IV -converter at the output has
not been described. It makes use of a simple differential amplifier which
will also be used in chapter 5 and will be explained there in detail.
IV-( onverfer
R R R
Fig. 4-17 Programmable analog signal processor for progressive

polynomial calibration.
The system of Fig.4-17 is completed with the circuitry for the biasing of
ali the different analog blocks, and a digital circuit in Emitter-Coupled
Logic (ECL), as presented in Fig.4-18.
The digital circuit consists of four 8-bit registers and a small serial
interface which makes it possible to program each register separate\y. The
digital outputs of the registers are connected directly, without level-shift,
to the 8-bit digital inputs of the DACs. These digital signals are on
ECL level which is equal to V" for high bits and V,e-VEcL for low bits.
Ve, is the positive supply voltage and VECL is the small differential voltage
used in the ECL gates, approximately 300 to 400 mV As a reference for
the switches in the DACs, a voltage halfway high and low, V,,_1/2 VElL ' is
derived in the digital part and buffered to the analog part of the circuit.
The other bias voltages needed in the analog blocks are generated as
shown in the figure. The bias voltages are chosen in such a way that

ECL bJa'
CMOS
I V« - I~
/2 VU { ~ to V"""I of DAC,
voltage
to
ECL r'T-r-r-r-r-r-r, A J buffer
buffer; ~CrJ::CrJ::~~17-01
I toDACI
clock
serial ~CrJ::CI:J:~A~217-01
I to DAC2
mput
selea{ '"'::lCID:~CI~A~317-01
I to DAC3 V" -1 4V to V"",,3 of DAC,
to V""" of Mulupher;
~CrJ::CI:J::~A~417-01
I to DAC4
~------------------~---------------- --------~,,~--------/
digital (ECL) analog
Fig. 4-18 Additional circuit blocks for the analog signal processor.
saturation of transistors is avoided when the blocks are connected as

shown in Fig.4-17, and the supply voltage is 4.5 VoIt or higher.
The digital inputs of the serial interface are connected to small buffer
circuits which convert digital input signals on CMOS-Ievel [0-5 VoIt] to
signals on ECL-Ievel [4.7 -5Volt]. This makes it easier to program and test
the Integrated Circuit, using standard digital signals.
4.5.6 System operation

The circuit of Fig.4-17 has a straightforward path through which the main
signal ftows: at the input the voltage signal is converted into a differential
current, which then goes through the current duplicator blocks numbered 1
to 3, and at the output it is converted back into a differential voltage signal.
The output currents of the DACs are added at several places in the forward
path. When the calibration starts, alI DACs are programmed to give a zero
output current. As a result, the differential output voltage is directly
proportional to the differential input voltage. The gain is determined by the
ratio of the resistors in the input VI-converter and the output IV -converter,
and the current scaI ing in the duplicators 1 to 3:
V OUT R IV
=G 2·_·G
R ·G ·G
123
(4-12)
VIN VI

The resistors R in Fig. 4-17 are all equal, and the current duplicators l to 3
have scale factors one, thus the overall gain G is 2. However, by choosing
different resistor ratios it is possible to obtain a larger gain.
Calibration Procedure
At the voltage inputs Y 1, Y 2' and Y3' reference signals have to be applied
to fix the rotation set-points of the calibration. Using a stable reference
signal which relates to the minimum/maximum of the input voltage range
(sensor signal), the voltage inputs of the calibration circuit can be
connected as shown in Fig.4-19. This results in calibration points for the
input signal at respectively Vm=V}=-Yreferenw Vm=V2=+Yreference'
Vl1l=V3= OYolt. Through equation (4-12) the corresponding calibration
points on the output voltage range can be found.
+~+
-,ell.wJ/__ ~: referell( /
Fig. 4-19 Connection of the voltage inputs of the calibration

circuit to input and reference voltage sources.
To follow the procedure, suppose that a sensor with a voltage output is

connected to the input of the calibration circuit, and also a reference
voltage is connected as shown in Fig.4-19. The combination of sensor and
calibration circuit is then inserted in a test environment to calibrate the
overall transfer.
The calibration starts by providing a first sensor reference signal (<Pmln)'

which leads to an input voltage VIN at the circuit. DAC l is then
programmed in such a way that the output voltage of the circuit obtains the
desired value for the first measurement (Vour=-C·Yreference). The
required digital value can be ca\culated simply by taking two
measurements of the output signal for two different values stored in the
DAC (for example, O and +1/2 or -Ih), and interpolating the two
measured output signals.
Then, a second input reference (<Pmax) can be provided to the sensor and
DAC2 can be programmed to obtain the desired output voltage for the
second measurement (Vour=+C·Yreference). Because of the system set-up,
it will not be necessary to reprogram DAC 1.

Then, a third calibration step is done to program DAC3, and a fourth step
for DAC4, according to the calibration principle explained in section 3.3.
After calibration, the combination of sensor and circuit has a signal
transfer which resembles the desired transfer, and matches it exactly in the
calibration points.
Simulation
By means of a PSPICE simulation, the circuit is tested for an input voltage
which is swept through the input range. The input range is set to [-250 mV,
+250 mV], Vos to 250 mV, and the successive calibration points are set by
voltages V I= -250 mV, V2= +250 mV, and V 3= O mY.
With R=25kQ, the corresponding signal range for the differential current
is approximately [-20 /lA, +20 /lA] and the full scale is 40/lA. The
function blocks in the calibration circuit are biased accordingly for
operation in c1ass-A mode. Fig.4-20 shows the differential current signals
at the inputs of the DACs as a function of the input voltage VIN, when ali
DACs are set to zero.
maxImum oft.\et (orre( fum
maxImum gam torre( (ton

IOu'"
maximum 2ud·order (orreefum
maxImum 3,cI. order Lorrec fum

(DAC 4)
-IOu ~
Ideal (urrent .\lgnal
-20u .... ---- ---------,.-------------------,-------------------..,--------------------r-------------------,.--------------

-200mV -lOOmV OV lOOmV 200mV
Fig. 4-20 Simulation result showing the input current signals of

the DA Cs as a junction of the input voltage.

The current signals in Fig.4-20 represent the maximum correction signals

that can be applied for, respectively, offset, gain, and second-order and
third-order linearity calibrations. In this case the maximum offset
correction is almost ±50% of the full scale. This might be adjusted by
changing the voltage applied at the Vos input. The output currents of
current duplicators 1 and 4 are scaled in such a way that the maximum
gain correction signal is ±50% of the full-scale signal. The scaling in
duplicators 2, 3, 4, and 5 is such that the second-order correction is
maximum ±8%, and the third-order approximately ±2% of the full-sca)e
output. All the corrections can be done with an 8-bit resolution.
Calibration Errors
The limited resolution in the calibration of the offset leaves a remaining
offset error in the calibration point: in the worst case, it could be plus or
minus a half LSB (Least Significant Bit) of the offset calibration range.
The gain correction at the next calibration step makes use of the error
signal, which is now not exactly zero in the previous calibration point.
This can have a negative effect on the remaining offset error, as illustrated
in Fig.4-21. The transfer will not rotate around the intended point in the
origin but around a different point.
output
mpUf
infended
rotatlOn poinf
Fig. 4-21 Effect of an error in the jirst rotation point.
When the offset calibration error was calibrated down to 1/2 LSB, and,
next, the gain is increased by 50%, then the offset error will be magnified
to 3f4 LSB. The same effect also applies for the succeeding steps, but with
less severity. In order to limit the effect of error magnification, the
correction ranges of succeeding calibration steps should decrease, or the
resolution of the first steps should be higher.

Fig.4-22 shows the error curve after calibration for a worst-case example.
Four calibration steps were do ne at calibration points corresponding to
-250mV, +250mV, OmV, and -125mV on the input range. The considered
1.0%
i Error (% of full scale)
05%
Fig. 4-22 Error after four calibration steps with 8-bit resolution.
example curve required almost maximum gain and linearity corrections.

Initially, at the first calibration step, the offset error at V/N=-250mV was
reduced to J12 LSB, or 0.2% (= 1/2 xT 8x I 00%) of the full-scale signal. But
after the gain and linearity correction it increased to approximately 0.35%.
The gain calibration itself left an error at V/N=250m V of J12 LSB, or
-0.2%. Because of the much smaller signal range for the linearity
corrections, the same 8-bit resolution leads to a much smaller error in the
third and fourth point. The resuIting integral error over the input range is
relatively smalI, despite the larger error values at the edges of the range.
To keep the error within the ±J12 LSB band, the resolution of the DACs
should be adapted to the correction ranges for each calibration step. For
practical reasons the use of the same 8-bit DAC and 8-bit registers for alI
four steps was preferred in the presented realization.
4.5.7 Realization
For the digital part of the circuit, the developed ECL standard cell library
for the DIMES-O l process was used. The layout of the digital circuit has
been generated automaticalIy by the computer layout program L-Edit.
Correspondingly, the presented analog function blocks have been laid out
as analog standard cells, so that they could also be used in the automatic
'Place and Route' routine of the same layout program. The standard celIs

on transistor level were laid out manually. Where necessary, transistor

pairs were placed in a cross-coupled double pair configuration (quad) to
mini mize the offset. Where the ratio of feedback resistors was important,
the resistors have been put close together in a standard cell to improve
matching. On the system level the connections were made automatically.
The analog and the digital part also had to be connected by hand. Fig.4-23
shows a photograph of the resulting chip, which measures approximately
5x5.5mm 2. This seems large, but it has to be considered relative to the
minimum dimensions of the process. The DIMES-OI process is not very
dense; an NPN transistor measures 40x561lm2, and a meta12-metall via is
12x 121lm2. In a modern high-density bipolar process, the square chip size
would scale down by a factor of four at least.
Fig.4-23 Chip photograph ofthe realized calibration circuit.
The functionality of the integrated circuit has been tested in a

measurement set-up as iIIustrated in Fig.4-24. Only the supply of 5Volt is
not shown. A constant voltage signal from a DC source is applied at the
reference inputs in such a way that V,=-250m V and Vas= V2=250m V
The third input is short-circuited so that V3=OmV Using a signal
generator, a low-frequency triangle-wave signal with an amplitude of
approximately 300mV is applied at the input VIN' The CM-Ievel of the
input voltage was set at the ground raii (OV). From a PC using the

data-acquisition program LabView, the digital registers which control the

DACs could be programmed via the four digital inputs of the circuit. The
resulting differential output signal was observed on an HP digital storage
oscilloscope. Because of a gain of two, the output signal is twice as large
as the input signal. However, the output voltage is Iimited by the circuit's
signal range at approximately +600mV and -600mY.
\!le -_
Labvlew
E> _
\!l-""
C · /Il
se/eu
calibration circuit
~-+-...L..J4++--"1II--+-, +
250mV
Fig. 4-24 Measurement set-up.
The triangular signal observed on the oscilloscope shows the transfer of

the circuit for the considered input range. The four registers of the
calibration circuit were programmed with different values, and the output
traces were stored on the scope. This made it possible to display several
traces simultaneously on the oscilloscope, which is shown in the scope
plots of Fig.4-2S. The first picture shows the traces for different values
programmed in the first register for the offset. The second picture shows
the traces for different values stored in the second register for the gain.
Idem ditto for the third and the fourth picture.
The pictures c1early show the rotation points in the output range at
-SOOmV, +SOOmV, and OY. The first three calibration points thus have to
be chosen at those output values. The fourth and last calibration point can
be chosen freely, but the best choice for a third-order Iinearity correction is
around plus or minus 2S0mY.
Because the analog signal goes through a cascade of local-feedback

amplifiers (IV-converters and VI-converters) with a relatively large
bandwidth, the bandwidth of the complete circuit can be large as well. So,

. : : :
.. ...... i.·"· ... ·;.-.. ,, ..·, . .. .. . " .
.. +. i "1'" "., .. -!. " ';'
1'" ", • • • .. 1 - •• ;.0 -'"
.. :.... ~ .......; ..... {.

. i. • -, ...... ,,, ..•. • .• - ..
.".......! ... ".-., .......... .
Fig. 4-25 Measured output traces on the oscilloscope

(x-axes: 5ms per division, y-axes: 200mV per division).
after the sensor transfer is calibrated, using low-frequency or static

reference signals, the correction also works for analog signals of high
frequencies. Depending on the power invested in the bias currents of the
different circuit blocks, bandwidths in the order of MHz can be achieved.
To reduce output noi se, the bandwidth of the realized circuit has been
limited on purpose with a low-pass filter to approximately IOkHz. The
specifications are summarized in the table below.
4.6 Conclusion
Several programmable analog calibration circuits, based on the use of
Digital-to-Analog Converters, have been presented. For uncomplicated
sensor calibrations requiring a limited resolution for the correction factors
(8-10 bit), these circuits can be advantageous. With respect to
conventional sensor trimming techniques, the advantage lies in the
possibility to digitaIly program the calibration and hence facilitate
automation. Furthermore, digitizing of the sensor signal can be avoided,

Table 4-1 Specijications of the analog calibration circuit.
Description Value Unit
Supply Voltage Veea 4.5-12 v

Current Consumption
mA
A=Analog, D=Digital
Input CM-range (forVee=5V) -0.3-0.6 v
Maximum differential
±450 mV
input signal
Nominal Gain 2 VIV
Max. Offset Correetion b ±50% of FS, 8 bit
Max. Gain Correetion ±50% of FS, 8 bit
Max. 2nd -order Linearity Correetion ± 10% of FS, 8 bit
Max. 3rd -order Linearity Correetion ±2% of FS, 8 bit
Max. Error after ealibration
±0.4% ofFS c
(peak value)
Conditions: Vcc=5 Voit, Vos=-V I=V 2=250mV, V 3=OV, FS ln =O.5V, FS out = I V
a Provided that the digital inputs can be scaled up accordingly

b. Adjustable wtth Vos
c. Provlded that the nonlinearity of the original slgnal fits wtthtn the
correction range.
and thus the analog calibration option offers the possibility to maintain the
bandwidth and dynamie range obtainable with compact analog circuitry.
The polynomiallinearization method presented in the previous chapter can

be worked out systematically in a programmable analog calibration and
linearization circuit. The implementation of such a circuit was elaborated
and demonstrated by a bipolar chip realization. However, the
implementation in such analog circuitry is feasible only for a small
number of calibration steps or for sensors requiring a high bandwidth.

References
[1] D. Tandeske, Pressure Sensors: selection and application, Marcel
Dekker Inc., New York, 1991.
[2] J. Bryzek, K. Petersen, J.R. Mallon, L. Christel, and F. Pourahmadi,
1990.
[3] P. Ciureanu and S. Middelhoek, Thin Film Resistive Sensors,
Institute of Physics Publishing, Bristol, 1992.
[4] R. v.d. Plassche, lntegrated Analog-to-Digital and Digital-to-Analog
Converters, Kluwer Academic Publishers, Boston/Dordrechtl
London, 1994.
[5] A.B. Carlson, Communication Systems: an lntroduction to Signals
and Noise in Electrical Communication, McGraw-Hill Book Co.,
Singapore, 1986.
[6] J.H. Huijsing, F.R. Riedijk, and G. van der Horn, "Deve\opments in
integrated smart sensors", Sensors and Actuators A, VoI. 43, 1994,
pp.276-288.
VoI. 14, No.3, November 1997 (to be published).
[8] G. v.d. Horn and J.H. Huijsing, "Programmable analog signal
processor for polynomial calibration", Proceedings ESSClRC' 96 ,
Neuchate\, Switzerland, 1996, ppAOO-403.
[9] B. Gilbert, "A precision four-quadrant multiplier with sub-
nanosecond response", lEEE Journal of Solid-State Circuits,
VoI. SC-3, 1968, pp.365-373.
[IO] C. Toumazou, F.J. Lidgey, and D.G. Haigh, edits., Analogue lC
design: the current-mode approach, Peter Peregrinus Ltd., London,
1990.

Calibration
using Sigma-Delta
5
Analog-to-Digital
Conve rs ion
5.1 Introduction
In section 4.4 the use of pulse modulation techniques to control the signal
transfer of a smart sensor was already suggested. The pulsestream or
bitstream modulation can be combined nicely with sigma-delta bitstream
techniques. A digital sigma-delta (noi se shaper) may be used to generate
programmable pulsestreams or bitstreams which modulate the analog
signa\, and a sigma-delta AD-converter is used to convert the modulated
analog sensor signal into a digital bitstream.
Before presenting the application of this technique for smart sensor

calibration, the explanation of the sigma-delta converters of section 2.4
will be extended to the frequency domain in section 5.2, In section 5.3 we
will then explain the principle of the bitstream modulation, and the effects
on the transfer and the performance of the AD-conversion.
Implementation of a first-order sigma-delta converter in bipolar circuits
will be presented in section 5.4. In section 5.5, we will present a smart
temperature sensor with programmable calibration as an example for the
application of the technique. Realizations of such smart temperature
sensors, incIuding one with a bus-interface, will be shown in section 5.5.

Calibration using Sigma-Delta Analog-to-Digital Conversion
5.2 Sigma-delta AD-converters

For c\ ari ty, we recal1 the differential configuration of a first-order
continuous-time sigma-delta AD-converter in Fig.5-1. Instead of using
two switched reference currents, as in Fig.2-14, they have been assembled
in one current source of twice the reference value. This is possible
provided that the differential amplifier has an input common-mode (CM)
control which takes care of the cM-component of the input current. The
converter further consists of an integrator, which can be regarded as a first-
order low-pass filter, and a clocked comparator, which can be viewed as a
combination of aI-bit quantizer and a c\ocked gate.
Integrator Quantizer+Gate
Output
bttstream
Reference
Fig. 5-1 Differential sigma-delta converter configuration.
Despite the quantization accuracy of only l-bit, a high resolution

conversion can yet be obtained on the basis of two wel1-known principles:
oversampling and noi se shaping. The oversampling is achieved by running
the c\ock at a frequency much higher than the Nyquist sample frequency.
The noi se shaping is obtained by the low-pass filter function in the loop.
A high resolution multi-bit digital output can be obtained by digital1y

processing a long series of output bits. This is referred to as decimation.
These digital output samples become available at a sample rate much
lower than the original sample frequency, hence the often-used terms
downsampling and decimation filter. Similar to the oversampling ratio, see
equation (2-5), we can define a downsampling ratio DSR as the ratio of the
original sample frequency fp and the decimated sample frequency fds:
(5-1 )

Usually, the maximum down-sample ratio DSR=OSR is chosen, and the

decimated digital output samples are available at the Nyquist rate again.
Fig.5-2 explains the complete signal processing schematically. As the

I-bit serial output of the sigma-delta converter is a very convenient sensor
output signal, the digital filter is usually not incorporated in the smart
sensor design. Instead, it may be implemented in software on the system's
computer or microcontroller. Or, it can be implemented in digital hardware
on a separate chip which is, for example, connected to the sensor bus.
Sigma-delta Converter Digital Filter

~_------------PL~---------- ~ __ r-----~A---f,----\
fI doi
1-bit ,-----+--.1-:,......., N-bit
output
output ........
I---r--+I ~ DS
L-_---'"_ _....
Decimation l Down
Reference : Sampling
I
Signal Domains: analog I 1-bit digital I N-bit digital
Freguencies: baseband ~ (over)sample rate I Nyquist rate
Fig. 5-2 Signal processing chain of a sigma-delta AD-converter.
The three principles involved in the sigma-delta type of AD-conversion

will be explained shortly and intuitively. More elaborate discussions can
be found in [1,2,3,4].
5.2.1 Principle of oversampling

The effect of quantization is often studied under the practical assumption
that the probability of the input signal is distributed uniformly between the
quantization levels. The quantization error is then distributed uniformly
between J /2 LSB and +//2 LSB, with the Least significant Bit representing
the smallest quantization step qJ= 1 LSB. By integration of the square error
over that range, the average power of the quantization error can be
estimated as [5] :
(5-2)

Calibration using Sigma-Delta Analog-to-Digital Con vers ion
The error introduced by the quantization is then considered as an

additional noise source. It is further assumed to be uncorrelated to the
sampled input signal. The noise will thus be spread evenly over a
frequency band equal to half the sample frequency, as shown in Fig.5-3 .
quanliUllion
noLe
,,1,
:;'\.
bCL'eoolld \ -2-
LSB2 Nyquist-rate sampling
frcqucncy
quanti7A1lion
noise
LSB2 Oversampling
baJebcJ/I~ TIOsR ;-!ff
?
\
j,= OSR-!Nyquisl
Fig. 5-3 Effect of oversampling on the quantization noise in the
baseband.
When sampling at the Nyquist rate, ali the quantization noise falls in the
baseband. However, when oversampling with a certain ratio OSR, the
quantization error is spread over a much larger bandwidth [2,6,7,8]. The
part outside the baseband can be removed by a filter and does not affect the
attainable signal-to-noise ratio. The baseband noise power is thus reduced
by a factor equal to the OSR. Accordingly, the signal-to-noise ratio is
improved with 3dB for each doubling of the sampIe frequency . The
improvement is given by:
Fa dB = -20IogJOSR (5-3)
5.2.2 Principle of noise shaping

To explain the principle of noi se shaping a discrete-time model of the
sigma-delta converter is most often used [1,2] . Even when using a
continuous integrator rather than the inherently time-discrete switched-
capacitor integrator. The model remains valid because the comparator
processes the integrator output at discrete time steps only. For a linear

integrator, the output voltage at discrete time steps, t=nTclk , is determined

by the previous output and the input current:
(5-4)
A general model of a first-order converter, also called noise-shape coder, is

shown in Fig.5-4. The input signal X(z) is normalized with respect to the
reference of the converter, so that the reference signal could be omitted
from the model.
Input Output
---t~
X(z) Y(z)
Fig. 5-4 Discrete-time model of a first-order sigma-delta

converter.
The quantizer is modeled by the noise source N(z) and the gate by a unit
delay Z-I. As we will see, the transfer of the integrator is essential for the
noise shape. It is given by:
H (z) (5-5)
1- Z-I
Using the figure, the transfer from the signal input X(z) and the noise input
N(z) to the quantized output Y(z) can easily be found, and expressed as:
Y(z) =z-I·X(z)+(l-z-I)·N(z) (5-6)
Whereas the input signal is only delayed (phase-shift), the quantization

noise is shaped by the factor (I-z-I). The effect on the noise magnitude
characteristic is obtained by ca1culating II-z-I 12 with substitution of

z=e-JCiJt , ro=2rcf, and 't=Tclk=lIf~. As a result, we find the following

function for the shape of the quantization noise power:
(5-7)
The function shows that the quantization noise is suppressed at low

frequencies, in the baseband, and elevated at higher frequencies, around
half the sample frequency. The noise-shape function is shown in Fig.5-6.
By integration over the baseband, f=0 to f=fb='/2f!osR, we can find a

noise reduction factor with respect to the normal oversampled uniform
noise [2]:
2-2 sin (ni OSR) (5-8)

(nIOSR)
For sufficient1y large oversampling ratios, this can be worked out, using a
Taylor expansion for sin(1fIoSR). Expressed as an improvement factor for
the signal-to-noise ratio in dB, it is given by:
FI, dB = -20log J3 ( 1t
OSR) (5-9)
For each doubling of the sample frequency, a 6dB improvement in the

signal-to-noise ratio in the baseband may be obtained by the noise
shaping, in addition to the 3dB improvement obtained by the
oversampling action (equation (5-3) ).
Quantization
Noise N(z)
Integrator2
Input Output
--+{,
X(z) Y(z)
Quantizer
Fig. 5-5 Possible model for a second-order sigma-delta

con.figuration.

Further improvement can be obtained with higher order noise-shape

functions. With higher order noise shapers, the number of possible
contigurations also increases, and the stability becomes an issue for
concern [9, ... ,14]. We will restrict ourselves to a second-order sigma-delta
modulator as modeled in Fig.5-5.
With integrator functions H 1(z) and H 2(z) as given by equation (5-5), a

transfer of the quantization noi se N(z) to the output Y(z) proportional to
(l_z-l)2 is obtained. The corresponding amplitude characteristic gives the
following noise-shape function:
2nf 4nf
6 - 8cos- + 2cos- (5-10)
f~ f~
Both the tirst-order and second-order noise-shape function are shown in

Fig.5-6. Clearly, the second-order function suppresses the quantization
noise in the baseband even further.
By integration of the equation (5- 10) over the baseband, we tind the noise
reduction factor, again with respect to the uniform case, expres sed as a
function of the oversampling ratio [2]:
6 _ 8 sin (ni OSR) + sin (ni OSR) (5-11 )

(nIOSR) (nIOSR)
By means of Taylor expansions this can also be simplitied. The

improvement in the signal-to-noise ratio can then be expressed in dB by:
FZ,dB n
= -20log ",5~ (OSR)2 (5-12)
Thus a 12dB improvement of the signal-to-noise ratio for each doubling of

the sampling frequency can be achieved with the second-order noise
shaping, in addition to the 3dB for the oversampling action.

1 second-order /
---------~ 1,",·"",
................ . ......... ... .. .l. . . ........ . .... .. . .. ................ ........ .. . ... .......... ...... .
\ uniform
\
ba~eband noise \ fh
\ \
& - - - - - - - \ .- 0.5!,
0.25 1.00 400 16.00
(normalized)/reqllency flfh •
Fig. 5-6 Noise-shape functions of a first-order and second-order

sigma-delta modulator for OSR=16.
The above described method can be applied for a generalized n-bit noise
shaper (sigma-delta) of order L, to tind a signal-to-(quantization)noise
ratio that can be estimated by [3] :
S
N q/jf2 O
n
- = q s 2 /2j2 . F . F = 2n 31t L + - . (OSR)L
L 2
-
1t
+~ J ( 1) (5-13)
s
A maximum sine-wave signal with a peak-to-peak value of 2n 'qs was

assumed. For low resolution (Iow n) noise-shaping coders, however,
van de Plassche proposes a correction factor for the maximum signal
amplitude [2] . For a 'one-bitter', the maximum peak-to-peak value of the
sine wave signal would be 4q/1t instead of 2q~.
In practice, the quantization noise of sigma-delta converters does not

follow the explained noise-shape function smoothly. The tirst-order
modulator in particular shows very discrete spectrum components, which
relate directly to the input and the clock signal. These spectrum
components correspond to so-called limit cycles which occur when a
bitstream pattern of limited length repeats itself continuously. For De input
signals the behavior is completely deterministic.

Fig.5-7 shows the spectrum of the bitstream output of a first-order

sigma-delta converter with a -3dB sine-wave input signal atf=fb' and one
hundred times oversampling. It was obtained by means of a simulation in
the data-acquisition program, LabView. The figure also indicates the
theoretical noise-shape function.
1
/ ba cbnnd . ignal
-20
-60
-BO
-1
1000
(normalized) frequency f/fh •

Fig. 5-7 Frequency spectrum of a first-order sigma-delta
bitstream output, for OSR= 100.
O
1/
1
ba.<ebnnd signnl
-20
ii;'
~ -40 1,
~
~
E
-60 ,1
<:S
o:
II
a i!
-BO .(
~ !
S
:.c:;
,.. Ihcorclical . hape
-100
010 100 10.00 10000
Ib o.sl.
Fig. 5-8 Frequency spectrum of a second-order sigma-delta
bitstream output, for OSR= 100.

Fig.5-8 shows the output spectrum for a second-order modulator, under

certain conditions, and the theoretical noise-shape function. The spectrum
c1early resembles the theoretical noise-shape function much better than in
case of the first-order modulator. The quantization error is decorrelated
more from the signal, and, hence, the assumption of white quantization
noise is more applicable.
Despite the better performance of the second-order sigma-delta converter,

the first-order converter is sometimes preferred in smart sensor
configurations because of the simple implementations and a guaranteed
stability. The typicalIy low bandwidth signals of sensors « l kHz) alIows
a large oversampling ratio (> 1000) so that sufficient1y high resolutions
can be obtained, even with a first-order sigma-delta converter [4].
5.2.3 Principle of decimation

To retrieve the baseband input signal from the high-frequency (HF) digital
one-bit output signal, some kind of low-pass filter is needed. Suppression
of the HF signals is important in order to reduce the noise that folds back
into the baseband when downsampling the signal. Digital signal
processing offers many advanced filtering techniques [15,16, 17], but we
restrict ourselves to the simplest technique used for bitstream decimation.
Each decimated output sample is obtained by processing the bitstream

output for a certain number of c10ck periods, equal to the DSR. The
considered series of bitstream pulses, as seen through a 'window', are
proces sed by applying weighting factors and summing the weighted
pulses. This is explained schematicalIy in Fig. 5-9. The values of the
weighting factors depend on their position in the window, and are defined
by a weighting function, or so-called window function. The resuIting
decimated output value must be interpreted in relation to the minimum (alI
pulses low) and the maximum filter output (alI pulses high). The
ca1culation of the decimated output occurs each time a window is filled
with fresh output pulses. This leads to a downsampling of the digital signal
by a factor DSR equal to the window size.
A uniform window function corresponds to eounting the high pulses (o nes)

in the bitstream output during a certain number of c10ck periods. The
Fourier transform of the uniform time-domain window is a sinc-shaped
low-pass function, as shown in Fig. 5-1 o. Hence the name sine filter is used
for this simple first-order decimation filter.

; • L .; signals:
_~~Ilstream
~
I
I I window
I function , , I
I __ __
I I
I welghted
I I bitstream
I I
-•------------r
liN-bit
output
Ilme_
Fig. 5-9 Application of window functions for decimation of the

sigma-delta bitstream.
10
1 08
04
02
O~{2-
5------~--~--1~OrO-------~--~~~
4~OO~--'~~~~~~1600

Fig. 5-10 Frequency transfer functions of decimation filters with
uniform and triangular window (ofthe same size).
A triangular window function, as explained in Fig.5-9, corresponds to a

low-pass filter function with a sinc-square characteristic. The transfer of
such a second-order or sinc 2 filter is also shown in Fig.5-1O. For both
cases the same window size was used . Clearly the suppression of the
higher frequencies by the sinc 2-filter is stronger than for the sinc-filter. 1
Steeper filter shapes can be obtained with more advanced window

functions, such as Hanning, Hamming, Blackman, etc. [18] . Theoretically,

a sinc-shaped time-window (with infinite window length) would realize

the ideal uniform baseband filter in the frequency domain. In contrast to
the advanced windows or filters, both the uniform and the triangular filter
are very easy to implement in hardware, using digital counters and
adders [15], as will be shown in section 5.4. The configuration is then
somewhat different as the FIR (finite impulse response) structure explained
in Fig.5-9, in order to reduce the amount of hardware (imagine the delay
line for large OSR).
If the order of the decimation filter is one higher than the order of the noise
shaper, it provides enough digital resolution for the obtainable signal-
to-noise ratio. The spectrum of the digital output signal of a first-order
sigma-delta converter with a sine-wave input signal, after decimation with
a triangular window function is shown in Fig. 5-11. The filter shape, also
shown in the figure, c1early suppresses the HF quantization noi se. The
peaks appearing in the spectrum are the higher harmonics of the
square-wave digital output signal representing the sampled sine-wave.
oo 10 00
(normalized) frequency flfi,
Fig. 5-11 Output spectrum of decimated first-order sigma-delta

output with OSR= 100 and a sinc 2 filter with DSR= 100.
J One may recognJze that when a rectangular wmdow functlOn IS used m a movmg-
average filler (the ave rage is computed every c10ck period), thls corresponds 10 a
Iriangular-shaped we lghting funcllon for a wmdow thal IS IWlce as large.

5.2.4 Transfer from analog input to digital output

Now that we have explained aII the processing required for the sigma-delta
analog-to-digital conversion, we can show the input-output transfer.
Fig.5-12 shows the transfer of a differential first-order sigma-delta
converter in combination with a decimation filter based on a triangular
window, using an oversampling ratio and window size of 128.
10
t 0.5
~
~
:::s
o
~ 00-
.~
"1:::l
"ţ:s
~
.~ -O 5
~
::::
~
::
-1.0
-1 O -O 5 0.0 05 1.0
t
....
e.... 000
~
-001
-0.02
-1.0 -O 5 0.0 0.5 1.0
normalized inpUl signal (li,l'ref) ~
Fig. 5-12 Transfer and ermr of a first-order sigma-delta converter

using a sinc 2 filter for normalized signals and oSR=J28.
The input signal is normalized with respect to the sigma-delta reference,

and the output signal is normalized with respect to maximum and
minimum digital output.

In most parts of the input range the error is quite small, in accordance with
expectations based on equation (5-13) (s/N::=2000). However, for input
signals just around basic fractions of the reference, such as -0.5, -0.33, O,
0.33, and 0.5, the error can get much larger. Around zero the error is
approximately as large as 1/0SR, which is due to long limit cycIes
occurring for small non-zero signals. These result in very low frequency
components in the bitstream output, which can not be filtered out by the
decimation filter. The error must be reduced sufficiently by choosing a
large oversampling ratio.
The peak errors can be reduced by applying a technique known as

dithering [8,19,20]. By adding some (white or pseudo-random) noise at
the quantizer input in the sigma-delta loop the bitstream behavior becomes
more chaotic. The limit cycIe lengths will be randomized so that the
discrete peaks in the frequency spectrum are smoothened and the spectrum
approaches the theoretical noise shape.
5.3 Alteration of the sigma-delta transfer

We will explain in this section how we can modify the transfer of a
sigma-delta AD-converter for the purpose of calibration of a smart sensor.
We will multiply analog signals by programmable digital l-bit signals.
These l-bit signals are generated by a so-called bitstream generator, which
will be explained first.
5.3.1 Digital sigma-delta bitstream generator

The discrete-time model of the first-order sigma-delta converter, shown in
Fig. 5-4, can be implemented quite literally in digital circuitry. Such a
circuit would convert a multi-bit digital number or LF signal into al-bit HF
digital bitstream with a noise-shaped spectrum 1. Different terms are
presently in use for such a converter, incIuding noise-shaping coder,
sigma-delta (de)modulator, sigma-delta DA-converter, and fractional-N
generator. As for the sigma-delta AD-converter, there are also many
I The same technique is applied for inexpensive high-accuracy DA-converters, for

example the well-known 'one-bitters' for CD players. The one-bit digital-to-analog
converter is easy to Implement and mherently linear, and the subsequent analog filter
can be very Simple because most quantization nOi se IS shlfted to hlgh frequencies

different implementations and several orders of noi se shaping

possible [2,21,22,23].
We wiIl settle for a simple first-order configuration as shown in Fig.5-13,

and refer to it as a bitstream generator, in accordance with its function .
The c10cked register and the full adder c1early realize the discrete-time
integrator function as shown in Fig.5-4. A shift register provides the
digital input signal to the other input of the full adder. In that way, the
input value can easily be programmed by a serial interface.
Seriul Proli fW'l/I/lIblre _ _ _ _-,

illPl1l - -- - . t
N-bi/ illpu/ vulue
I -bi/ oU/Pili (bi mreum )

N-bil I,;c=(/=
rry= ==:::::;-__...
FuI! Adder t-
+
Clod -....1..----1
N-bi, iII/eliTa/i im vulue
Fig. 5-13 Con.figuration of a first-order sigma-delta bitstream

generator.
Because the N-bit full adder outputs a carry signal each time the digital
output exceeds the maximum value (2 N _I), it thus functions as a digital
comparator with a built-in reference level. The full adder continues its
count at a shifted output value as it loses the bits above the most significant
bit (MSB). Hence the subtraction of a reference, needed in a sigma-delta
converter, is also inherently fulfilled in the digital full adder. The carry
signal, which corresponds to a comparator output, is used as the bitstream
output signal, proportional to the programmed N-bit input signal. Matters
are c1arified in Fig.5-14, by showing the signals of a simple 4-bit
implementation of the bitstream generator.
One may also notice the difference in the bitstream signal with a pulse-
density or pulse-width modulated signal [24]. In fact, the configuration

Calibration llsing Sigma-Delta Analog-to-DigitaI Conversion
bitstream
output
I I I I
r,
16 +---r+-I---.. . . . . ----r--r----,-+-+----'--'---T"""I- maximum
12
I integration
I value
4
Fig. 5-14 Digital signal values of the sigma-delta bitstream

generator.
operates as a first-order sigma-delta converter with a (pre-)quantized input

signal. Because only quantized input levels are possible, corresponding to
a multiple of the fraction 2"N, the length of the produced limit cyc\es is
always limited to 2N . Assuming a constant digital input signal, the lowest
frequency component which may occur in the noise-shaped bitstream
output spectrum is the cIock frequency divided by 2N. When filtering the
bitstream output at half that frequency or lower, the original digital input
signal can be recovered with the same N-bit resolution.
As with the differential sigma-delta AD-converter, we can introduce a

polarity to the bitstream signal by assigning a value of -1 (instead of O) to a
low output state. The N-bit input value can be 2's complement coded by
inverting the MSB. The maximum symmetrical signal range that can be
used goes from _(2N -1_1)I2N -1 to +(2 N-1_1)/2N -1, avoiding the use of a
constant (not-switching) bitstream signal. For large N the range practically
equals ±1.
In the following sections we will explain that the generated bitstream

signal can be used to modulate the analog input signals of a (normal)
sigma-delta AD-converter in order to realize a programmable
multiplication factor.

5.3.2 Modulating the input signal for gain calibration

Just like the reference current, the input current may also be switched
between the positive and negative input of the integrator, as illustrated in
Fig. 5-15 . The switch is controlled by a bitstream generator as described in
section 5.3. 1.
:>---r--1Ii ou I pu I
bil.l'lreWI1
Fig. 5-15 Sigma-delta converter with modulated input current.
Imagine, for example, that the input current Itn is being switched to the
negative side just as many times as to the positive side, the average
contribution will then be zero. In fact, the analog input current is simply
modulated by the programmable bitstream signal. After averaging or
low-pass filtering by a decimation filter, the sigma-delta output will be
proportional to the multiplication of the current ratio Itdlref and the input
value of the generator, indicated as gain in the figure.
The average of the generated bitstream, and thus the gain, can be
programmed with an accuracy that is determined by the resolution of the
bitstream generator. The resolution of this fully digital circuit can be made
very high.
In the configuration of Fig. 5-15, the gain may be controlled almost from
-1 to + 1. Such a large correction range may not be desired. Fig.5-16
presents an alternative configuration in which a down-scaled copy of the
input current is being modulated and added at the integrator inputs.
Provided that the input signal can be duplicated and scaled by a factor a
(a< 1), a gain ranging from l-a to 1+a is now possible. The next figure,
Fig. 5-17, shows the transfer of the complete sigma-delta converter (with
sinc 2 decimation) for different values of the gain. The resolution of the
bitstream was 6-bit, <X was 0.25, and the window size of the decimation

l in >..,..~ OUlPUI
+L---hr---h-~"/"--"""'T-I biwream
Fig. 5-16 Sigma-delta converter with modulated additional input

signal.
1.0..,-----;------:-----:------,--..,..",...,..-r--::r- - ,
05
00
-o 5 --/---------
-10;--~~~LJ-----,------r_--~----~._--_1
!,
-15 -10 -0 .5 00 05 10 15
normalized input signal (i ilii re;)
Fig. 5-17 Signal transfer of the gain-programmable sigma-delta

converter, for different gain values.
filter was 128. For this test case, the digital 2's complement numbers O,
plus and minus 13/32 , and 26/32 were programmed in the bitstream
generator, corresponding to gain values of approximately 0.8, 0.9, 1.0, 1.1,
and 1.2.
It is c1ear that we can map a larger or smaller range of input signals,

compared to the usual ±Iref' onto the normalized output range ± 1, so that

the correction can be used for a programmable gain calibration of a sensor

signal.
To understand the effect on the performance of the AD-conversion and

under which conditions the modulation can be used, we examine the
behavior in the frequency domain. The first picture in Fig. 5-18 shows the
frequency spectrum of the bitstream output of a sigma-delta with a normal
unmodulated sine-wave input signal, and an oversampling ratio of 100.
The dotted lines in the next picture are the frequency components of the
digitally generated bitstream used for the gain correction. Because of the
fixed (DC), quantized (digital) input, and the first-order generator, there are
only a few discrete spectrum components. The DC-component, not visible
because of the logarithmic frequency axis, represents the average
bitstream value corresponding to the programmed digital input. The
spectrum is shaped in the same way as that of a first-order sigma-delta
converter.
The sine-wave input signal that is being modulated by the bitstream results
in sideband signals next to each bitstream component, as indicated in the
same figure with solid Iines. Also, at the signal frequency, a component
results which represents the gain correction. Because the amplitude of
these sideband components is proportional to that of the modulating
frequency (of the bitstream), also the modulated correction signal has a
shaped noise spectrum.
The modulated correction signal is added at the input of the sigma-delta

converter, and thus some additional shaped 'noise' is introduced. The level
of the noise can be lowered by reducing al, which controls the amplitude
of the correction signal. Furthermore, the presence of this 'noise' effects a
dithering of the integrator signal, which results in generation of a much
denser noi se spectrum. The last picture of Fig. 5-18 shows the spectrum of
the bitstream generated by the sigma-delta AD-converter with modulated
input signal. To demonstrate the effect, a large factor a= l was chosen. It
can be seen that extra noise peaks appear, that the shape of the
quantization noise has been preserved, and that the amplitude of the input
signal has been reduced by the gain calibration.
The effect of the modulation is thus twofold. On one side the performance
of the AD-converter is affected in a negative way: an extra noise source is
I acan be seen as the moduiaruJIl deprh of the mput signal

t / ba.~eband signal
-80
0.20 100 5.00 2500 100.00
-
o
1
- gain com:clion .1
gain bit.lrcam componcnls ': . "
·20
sidcband signals ~ : ,! ~ II! i j I
~ :' ~ : ~
·40-
l i' "Ii
~I
-60
li: il
1' '1
I
"
-80 il
0.20 100 5.00 2500 10000
o
t 1/ rcduccd baseband signal
·20 !
-40
!
1,
-60-
-80-
j
0.20 100
b (Ilormalized) !reqllellcy !!fh 0.5 ,
..
Fig. 5-18 Spectra of, respectively, the unmodulated bitstream
output, the correction signal, and the modulated output.

introduced at the input. The (frequency-)shape of that noise is determined

by the digital sigma-delta, and the amplitude depends on the factor <1,
which also controls the range of the gain correction. If <1 is small enough,
suppose a 25% gain correction range is sufficient, or if the noise shaping of
the correction signal is better than that of the sigma-delta AD-converter,
the introduced noise can be neglected.
On the other side, the behavior of the AD-converter is improved due to a

dithering effect on the input signal. The noise-shaped output spectrum,
which for a first-order usually consists of many discrete peaks, is
smoothened to approach the theoretical shape more c1osely. This
improvement especially applies for Dc-input signals.
5.3.3 Modulating an additional signal for offset calibration

In the same way as for the gain correction, we may add a modulated
reference current at the input of the sigma-delta converter, as shown in
Fig. 5-19. This allows a programmable offset correction of the transfer of
the sigma-delta converter. Aiso here the correction range can be set by a
scale factor <1, for the copy of the reference current.
>~-111 OII/PUI
biHlre/ll1/
oi{"! 1correCI ;OIl
Fig. 5-19 Sigma-delta converter with modulated additional

reference current.
The effect on the transfer is shown in Fig.5-20, using the same conditions
as for the gain correction: DSR=128, <1=0.25 , and a 6-bit bitstream
generator programmed with the digital numbers 0, plus and minus 13/:n ,
26
an d /32'

/tjL ~ ____._ .
1.0-r-----;----- - - - - ; - - - - - - - - - r - : -.,.,.-::,...,...----,
f
,,' / 1
I i~, /
0.5 ----.-----ţ _________ ~
0.0- · - __ ~L _. _ _.__.-+.- - -
/; ~
-0.5
//
,d~//~:,~ : i
.. -
1/'/ :
-1.0+---/c.....<':....r/_· r.. . J,= ---.-___1_ _ _ _• _----r----/
-1.5 -1.0 -0.5 0 .0 0.5 1.0 1.5
tlormaLized input signal ( / itl/ ref)
Fig. 5-20 Signal transfer of the offset-programmable sigma-delta

converter. for different offset values.
5.3.4 Sigma-delta calibration configurations

The offset and gain corrections, as presented in the previous sections can
simply be combined in one sigma-delta AD-converter. Even more signals
can be added in the same manner, to introduce a programmable correction
of the cross-sensitivity or the nonlinearity [25,26], provided that the
required analog (current) signals are available or can be generated easily.
The analog correction ~ignals must be modulated by separate bitstream
generators 1, which can be made programmable through a common serial
interface.
As an example, Fig.5-20 shows a sigma-delta configuration with a

modulated current a/lin for gain control, a modulated current a2/ret for
offset control, and, additionally, a temperature dependent current IT for
control of the temperature-coefficient in the offset of the transfer. To avoid
confusion, it must be mentioned again that the aim of the circuit is to
calibrate the overall transfer of a smart sensor, not that of an AD-converter.
I Multlplc usc of a smgle oltstrcalll gCllcratOi through IllUltlpleXlI1g IS also posslole.

out sOlllcwhat compltcatcd. oecause for thc mtegratloll value must also oe preserved
Another optlOn IS to multlplex the generator over vcry long tllne-slots (ovcrsulllpltng
factor). Then the wmdow size of toc declmatlon must be mcreased accordmgly
148 INTEGRATED SMART SENSORS: Design and Calihration

The input current should be deducted from the sensor signal, the
temperature current can be obtained by use of a PTAT source Of another
temperature sensor.
OUIPUI
bilSlream
('/ock
diJ:ilul
biWrellm
J:enerulor
diJ:illll
biwrelll1l
xel!eralOr
oJj.~el
dixillli
biwream
Te
J:eneralor
. 2l reI
Fig. 5-21 Sigma-delta configuration with programmable ojfset,

ojfset temperature-coefficient, and gain.
To realize a correction possibility for the temperature-coefficient in the

gain of the transfer, a signal proportional to the product of IiI! and IT is
needed. One approach may be to use an analog multiplier, as suggested in
section 4.4. However, the multiplication can also be achieved with an
additional modulated sigma-delta AD-converter for the temperature
signal, as illustrated in Fig. 5-20. First, a bitstream which is proportional to
the temperature and a programmable calibration factor is generated. This
bitstream is used to modulate an additional current source proportional to
the input signal which realizes the multiplication.
Even more configurations become possible when it is conceived that the

bitstream signals can also be added in the digital signal domain, as
suggested in [28,31]. This would require a separate sigma-delta converter
for each analog signal, and a noise-shaping digital adder at the output, in
which the bitstreams are weighted and added I .
I The author believes there stilllies a research field here to be explored, but already
mitiated by [28 , ... ,34)

clod
I'm~rallll/l.ill~ illf'1I1
Fig. 5-22 Sigma-delta configuration with bitstream calibration of

offset. gain. atţd temperature coefficients of both.
Two things become very apparent in the presented configurations when

looking at the many (modulated) input signals that are summed at the input
of the integrator.
First, the integrator is loaded with a much larger input current than usual,
because at some point in time ali the currents may get summed. To avoid
clipping of the integrator output, either the integration capacitors must be
increased, or the input currents must be scaled down. The first requires a
larger bias current of the integrator output stage and demands a larger
dynamic range for the comparator input. The latter affects the dynamic
range requirements at the integrator input. In both cases more bias current
is required to maintain the (analog) signal-to-noise ratio.
Second, the multiple randomized input signal improves the dithering

effect, because ali the bitstream generators add a sort of pseudo-random
noise on top of the De average correction signals. This especially
manifests itself when the bitstream generators produce wide, but shaped
noise spectra. Thi s happens for irregular bitstream patterns generated by
high-resolution input values (123/512 , for example) or with higher-order
modulators.

The positive effect on the noise shaping of the AD-converter's bitstream

output can be seen in the spectrum shown in Fig.5-23 . It resulted from a
sigma-delta converter with an arbitrary offset and gain calibration, using
modulation depths al=az=O.25. The spectrum shows a good resemblance
to the theoretical first-order noise shape.
1
/ ba. ebnnd signal
~
~
.~
~
i5
§
~
!::
:.c;
00
(normalized)jrequency f/fh
Fig. 5-23 The bitstream spectrum of a first-order sigma-delta

converter with modulated input signals, and OSR= 100.
0.02-
1 0.01-
~
~
1:: ~
000-
<Il
-001-
-002-
-1.0 -o 5 00 0.5 1.0
tlormaliloed illpUi sigllal ('iri' rei ..
Fig. 5-24 Errar versus input-level for a first-order curve sigma-
delta converter with modulated input signals. using a
second-order decimat ion with a window size of 128.

The positive effect on the error for De input signals is demonstrated by the
error function shown in Fig. 5-24. It was obtained by simulating a
sigma-delta modulator with a small offset and a small gain calibration, and
by using a second-order decimation with a window size of 128. When
comparing Fig.5-24 with the original error curve in Fig.5-12, we see that
the error has been smoothened, but that the average has gone up slightly.
To summarize, we list the disadvantages and advantages of the presented

technique for smart sensor calibration:
- the progressive polynomiallinearization method proposed

in chapter 3 cannot easily be applied here, since the method
is based on summation of correction signals
- preferably, a small correction range should be used, say
below 50% ofthefull-scale
- the demands on the dynamic range ofthe analog
components will be more severe
+ accurate digital calibration, resolutions up to 16-bit or
higher are feasible
+ simple, robust, and accurate analog-to-digital conversion
which does not rely on device matching
+ one-bit serial correction signals are used, thus requiring a
minimum number of routing wires from the digital to the
analog part of the circuit
As a conc\uding remark, it must be mentioned that the presented bitstream
modulation technique can also be used in combination with higher-order
modulators. The order of the digital sigma-delta modulator (bitstream
generator) as well as the order of the sigma-delta AD-converter can be
increased. Because of the noise spectrum added to the input signal, the
order of the digital sigma-delta should be at least as high as the order of the
AD-converter. In that case, the small increase of quantization noise can
easily be compensated with a small increase of the oversampling ratio.
5.4 Implementation of the sigma-delta AD-converter

The main analog building blocks of the sigma-delta converter are the
differential amplifier and the c\ocked comparator. The implementation of
those blocks will be addressed in the subsequent sections 5.4.1 and 5.4.2.

The current switches have been implemented simply by differential pairs,

preceded by emitter followers to obtain a level shift. The circuitry for the
generation of the current signals depends on the sensor application. For
exampIe, the PTAT and reference current source of section 2.3.1 and the
VI-converters of section 4.5.4 can be used. The operation of the presented
blocks is demonstrated by a simulation shown in section 5.4.3.
The digital blocks were designed in such a way that implementation is

possible, using standard digital CMOS cells or the bipolar ECL cells,
developed for the DimesO 1 process. The latter option, although not very
compact, allows integration with the bipolar analog part on one chip.
Therefore, the ECL logic was used for the bitstream generators.
A second order decimation filter was realized as a test-chip in DiMOSOI, a

1.6~ CMOS process which only recently became operational in DIMES. It
will be issued in section 5.4.4. Besides the hardware implementation ofthe
bitstream decimator, a first-order decimation (counter) in the read-out
software on a PC was also available for testing the sigma-delta converter.
5.4.1 Differential integrator amplifier

The differential amplifier is used in the integrator of the sigma-delta
converter. The integrator amplifier fixes the voltage level at the input, and
reduces negative influence of the Early effects in the currents sources and
switching transistors, while providing the integrated large-swing voltage
signal at the output.
The amplifier used is shown in Fig.5-25. It is a very simple 2-stage

amplifier design, built up with a differential pair input stage QI-Q4 (quad),
a folded cascode, Q7 and Q8, and two voltage followers, Q Il and Q12, for
providing the output current. The voltage amplification is realized by the
input stage cascode combination only. The dominant poIe is at the bases of
the output stages. No compensation capacitors were needed so that the
amplifier has no internal slew-rate Iimit. A voltage gain of 66dB (2000)
and unity-gain bandwidth of 20MHz were obtained. The current gain was
comparable (-1200). These figures are sufficient to neglect integrator
leakage for sigma-delta resolutions up to 10-12 bit [4,35].
Furthermore, the amplifier contains two common-mode (CM) control

loops, one at the input and one at the output. Transistors Q5 and Q6, in
combination with the level-shifted connection through R5 to the common
emitter node of the input stage, fix the CM input level to a few hundred

Vcc
05
r~---+
R5
32K
+--+---+--...... out+
out-
In- .....--""'1I---+-,r-----+----' +---+--f:: 011
+----+---""'11-----«: 012
Vee
~-~-+--~----~-~--+--+-~--
Fig_ 5-25 Dijferential design ofthe integrator amplifier.
millivolts under the supply. This a\1ows to drive the current switches, that
are connected to the integrator input, by digital voltage signals on
ECL level (see section 4.5.5) without saturating them.
The second CM-control does not rea\1y fix the cM-Ievel of the output
voltage, but regulates the maximum of both voltages. The degenerated
differential pair, Q9 and Ql O, detects which of the two output voltages is
higher, and the common node (with R9 and RIO connected) fo\1ows that
voltage. A feedback through the cascode transistors Q7 and Q8 combined
with emitter resistors R7 and R8 prevents the output voltage from rising
too far. The cM-control 'gent1y' clips the voltage at the rising output node
just below the supply level, while the voltage the other output can go
down. The asymmetry of this control with respect to the differential signal,
is compensated by the differential feedback configuration in which the
amplifier is used. The scope plot shown in Fig.5-26 demonstrates the
operation of the control loop. While the individual output voltages are
distorted, the differential output voltage amplifies the input correct1y.
An advantage of such a control is that when the differential voltage is

sma\1, and the connected comparator has to make a precise decision, the
absolute voltage levels are close to the supply voltage. This means the
(NPN) input stage of the comparator has to be operational only in a sma\1
range around the positive supply (not, as usual, halfway the supply rails)l.

'..t. .... ... :. . 0, 0 •• , • • •••
i-
.
T
r
·.·· ·· ·· ·· --.,. · ··· ···· f· ····· ···, ·····
Fig. 5-26 Input and output signals of the differential amplifier,

used in afeedback configuration.
The presented amplifier design has been completed with the bias circuitry
for the current sources and was implemented in a very compact standard
layout ceH for the analog library. It is shown in Fig.5-28.
5.4.2 Clocked comparator

The c10cked comparator uses the familiar master-slave configuration and
is shown in Fig. 5-27 . The amplifying stage, QI and Q2, and the latching
stage, Q3 and Q4, are turned on and off alternately, controlled by the c10ck
signal. The other combination, Q5, Q6 and Q7, Q8, is switched in the
opposite way. On positive c10ck edges, the input voltage amplified onto
R1 and R2, is latched by Q3, Q4, and pas sed on by Q5, Q6 to the output. At
the negati ve c10ck edge the second latch Q7, Q8 holds the output state.
The input pair is biased with an additional current source II to avoid that it
is switched off completely. The emitter followers, Q13 and Q14, buffer the
c10ck signal onto the c10ck switches, so that the digital c10ck signal can
also be provided on ECL level. Saturation of the transistors will not occur,
provided that the input signal is provided at a high CM level, as discussed
at the end of the previous section.
I Especmlly 10 low-voltage apphcatlOns the proposed asymmetrical cm-control will

thus be ve ry practlcal.

Calibration using Sigma-Delta Analog-to-Digital Con vers ion
Vcc
Rl R2 RS R6
l2K l2K l2K l2K
out-
+--t-_out+
clk+
clk- --t-"""""J...
Fig. 5-27 Differential design ofthe clocked comparator.
Also, the comparator has been completed with the transistors for the
current sources, and laid out in a compact standard cell . It was added to the
analog cell Iibrary, which was also used for the design in chapter 4.
Fig.5-28 gives an impression of the layout cells of the two function blocks
discussed here .
Differenttal Amphfier Clocked Comparator
Fig. 5-28 Standard layout cells ofthe differential amplifier and the
clocked comparator.

5.4.3 Sigma-delta converter

The functionality of the blocks has been tested by means of a pspice
simulation of the differential sigma-delta configuration. Depending on the
clock frequency the value of the input currents and the supply voltage,
smaller or larger integration capacitors must be used. Fig.5-29 shows the
result of a simulation for a clock frequency of 200kHz, a reference current
of 12/-lA, and integration capacitors of 40pP. The supply vOltage was
3 VoIt. After start-up the integrator signal clearly shows the desired
behavior for a first-order sigma-delta modulator.
clod
r···········tiiiiiJuiiiiiiireiiiii .................................... _... __ .... .

v" =3 i
,
t :\ f\ !\ {\J\ f\ !\ f\f\ :
(V) O~\j~~YVV~
·1 j dtfterential In/elira /or ou/put i
·2! ............. _............. -............................................ _.................................... j
O 20 40 60 80 100 120 140
time !l.iS) ..
Fig. 5-29 Integrator output signal and bitstream signal of the

differential first-order sigma-delta converter; (pspice
simulation).
The next figure, Fig.5-30, shows the simulation result for the same
sigma-delta converter, but then with a modulated input current. The
modulating signal is shown on top. The effect on the slopes of the
integrator signal is clear: there are now 4 possible input conditions. The
maximum voltage-swing also increases.
The computation effort for the pspice simulations is too much to simulate
enough clock periods for determination of the resolution on the basis of
the bitstream output signa!.
Using the laid-out analog library cells presented in the previous sections,
completed with layout cells for a PTAT bias source, for the current switches

clock
Nain-
modulatinx
bllstream
v" =3 [ ---~;uipui -bit:~i;-eaiil- ----------------------------------------------------------------------------------
!
t o;
(V)
-1 j
-21------------- ______________________________ . ______________ . ____________________________ .______________ . _____ oi
o 20 40 60 80 100 120 140
time HH) ..
Fig. 5-30 Integrator output signal and bitstream signal of the

first-order sigma-delta converter with modulated input
signal.
and for the capacitors. the layout for different sigma-delta configurations
could be generated automatically. It can be combined with automatically
generated digital designs for the bitstream generators. as will be shown in
section 5.5. Fig.5-31 shows one of the sigma-delta converters on a chip
photograph.
Fig. 5-31 Chip photograph of a sigma-delta converter with two

60pF (6 x 20pF) integration capacitors.

5.4.4 Digital decimation fiUer implementation

As a test chip for the new DiMOS process, a bitstream decimation filter
with smart sensor bus interface has been developed. Fig.5-32 shows the
configuration.
After reset, an 8-bit up/down counter generates the triangular shaped

weighting function for a window size of 512 c10ck periods. Each clock
period, the weighting factor from the up/down counter wiIl be added to the
memorized sum if the bitstream input is high. When the count returns to
zero, at the 512th c10ck period, the total sum wiIl be stored in the output
register. The maximum output equals 512x256/2=65386, hence the
16-bit resolution for the fuIl-adder and the register in the loop. The
decimation result can be shifted out of the output register seriaIly, which
makes it easy to pass it through to the bus interface. It also makes it
possible to process the bitstream and read out the converted signal at
different c10ck frequencies.
BUS
111111'111 VlIItU' (IIIbl' . eli!)
received bitstream
serial
loud outp"1
/6·bil /6·bil
Fu 11 shifl
Addcr regislcr
11111(/
L - _ - - j -- - - - ' wi/ldoll"jilll (5 /2 dbJ
Fig.5-32 Configuration ofthe 16-bit sinc 2 decimationfilter.
The bus interface is connected in such a way that any bitstream put on the
bus by a sensor wiIl be decimated by the filter. The bus master must first
address a sensor which outputs a bitstream, then, after at least 512 c10ck
periods, it can address the filter to read out the decimated value.
A digital standard cell library was adapted to the DiMOS design rules, so
that the layout could be generated by the place & route routine of the

layout program L-Edit. Fig.5-33 shows a photograph of the resulting

CMOS chip, which measures approximately 3x2 mm 2 .
The chip can be used to connect any sigma-delta output to the integrated
smart sensor bus, for easy read-out and testing on a PC-controlled
measurement set-up. It has been tested, using the integrated smart sensor
(IS 2 ) bus and several smart temperature sensors, as presented in the next
section.
Fig. 5-33 Chip photograph of the CMOS realization of the /6-bit

decimation filter.
5.5 Smart temperature sensor realizations

The techniques presented in section 2.3.1 and section 5.3 can be combined
in the design of a smart temperature sensor with programmable offset and
gain calibration. The goal of the calibration is to overcome the device-to-
device variations by means of calibration, so that ali sensors can then be
delivered with a fixed transfer from temperature to output. It is also
possible to allow a customer programmable (variable) temperature range.
These possibilities have been worked out in the next sections .

5.5 Smart tempera ture sensor realizations
5.5.1 Smart temperature sensor with calibration and IS 2 bus

interface
Fig.5-34 shows the configuration of a smart temperature sensor based on
the combination of (l) a PTAT and bandgap reference circuit, (2) a sigma-
delta AD-converter, (3) an IS 2 bus interface, and (4) two programmable
bitstream generators. The combination of a PTAT temperature sensor and a
sigma-delta AD-converter has been reported earlier [36].
oulPUI biwreall/
clod
received
byle
hil.H refll/l
OUII'UI
~ ~
. 4-bit
bitstream
(!fJ"Sel
adjuJI
.;;0 Sigma-delta convener
2 '"
generator
Xaill
.::: 4-bit adjuJI
;: ·5 bitstrcam
"" generator
<.r.
Analog tcmperntu rc scnsor ci rcuit
Fig. 5-34 Configuration of the smart temperature sensor with bus

interface and programmable calibration.
The analog circuit for the temperature sensor is constructed in such a way
that a temperature signal lr and a reference signal Iret are generated, as
well as copies of both currents which are scaled down by a factor of 16. It
is aimed to obtain a ratiometric signal IylIret which varies from O to l for
the temperature range of -30°C to +70°C. The ratiometric conversion is
obtained by the sigma-delta converter. Relatively large integration
capacitors 60pF are used, because the maximum sigma-delta clock
frequency that currently can be obtained on the smart sensor bus with a PC
bus master is low, about 70kHz.
The signal 1ret/l6 is used to correct the offset error in the transfer. It can be
added or subtracted from the normal input signal with a 4-bit precision,
using the first bitstream generator. In the same way, the signal Iyl16 is

modulated by the output of the second bitstream modulator, and added at

the sigma-delta input for correction of gain errors. The maximum
correction for offset and gain is thus approximately ±6.25% of the full
range. This is sufficient to compensate for the errors that can be expected
to result from process variations, provided that the process is characterized
well so that the actual nominal (average) sensor transfer is c10se to the
desired transfer (see section 3.1.2).
The two 4-bit digital values for the offset and gain can be programmed
through the bus interface. At one of the two sensor addresses the sensor
can receive this calibration byte programmed by the bus master. At the
other bus address of the sensor, the bus interface sends out the bitstream of
the sigma-delta converter which can be received by the bus master.
Fig. 5-35 Chip photograph of the bipolar realization of the smart

temperature sensor.
Fig.5-35 shows a photograph of the chip which was realized in the

Dimes-Ol bipolar process of the research institute DIMES. It measures
approximately 4.4x5 .3mm 2 . A commercially more feasible chip size may
be obtained by integration in a high-density bipolar or in a CMOS process
(CMOS temperature sensors have been reported in [37,38], for example)
which would reduce the size of the digital part especially. In addition, the

5.5 Smart tempera ture sensor realizations
size of the analog part may be reduced by shrinking the integration

capacitors, and running the sigma-delta converter at a higher c10ck
frequency. In aII, the chip size can be reduced to below 1 xl mm 2 .
Through the use of the smart sensor bus, several sensors can be measured
at the same time, including a reference temperature sensor for calibration
purposes. In addition, the decimation filter presented in section 5.4.4 can
be connected to the bus for decimation of the bitstream output of each
addressed temperature sensor. An impression of the measurement set -up is
given in Fig.5-36. Multiple temperature sensors can be connected to the
bus and inserted together in a temperature chamber (oven) for collective
calibration.
decimatÎO" fllter
IS2 refere"ce se"sor
Sus
Master
se"sor1
IS2 bus lill es
Fig. 5-36 Impression of the measurement set-up using the

integrated smart sensor bus for communication.
Fig.5-37 presents the transfer and error of two measured temperature

sensors before and after calibration of offset and gain.
Clearly, the curves can be calibrated onto the desired transfer curve. After
calibration, the error remains almost within the ±0.5% range,
corresponding to an 8-bit resolution. The calibration step size (LSB) was
0.8%, however a noise error was present as well. The parabolic error
curve, remaining after offset and gain calibration, can be recognized with
difficulty.
The table below summarizes the specifications of the chip. Because of the
ECL realization, the digital part consumes a considerable amount of
supply current which explains the self-heating of about 0.5 ce.

t D1Rltai Output
08
desired transfer
06
-20 o
t 20
Ermr (%)
40 60
-
Temperature (OC)
80
.4
------------- - _ _. _
-40 -20 o 20
-
Temperature (OC)
80
t DiRital Output
08
06
-20 o
t 20
Ermr(%)
40 60
-
Temperature (0C)
80
-2
Fig. 5-37 Measured temperature transfer and relative errar oftwo

smart temperature sensors, before and after calibration.

Table 5-1 Speci,fications of the smart temperature sensor with gain

and ofJset calibration, and bus inteiface.
Description Value Unit
Temperature range -30 to +70 cC
Calibration aeeuraey (LSB) 0.8 % ofFS

Calibration range, offset & gain, 4-bit ±6.25 %ofFS
Useful output resolution 7-8 bit
Maximum error after ealibration ±0.8 % ofFS
Cloek frequeney 50-200 kHz
Supply Voltage, Vee 3-6 V
Supply eurrent, analog part 04 mA
Supply eurrent, digital part 16 mA
Self-heating -0.5 cC
Conditions' Vee=5, DILl6 paekage
5.5.2 Temperature sensor with linearity calibration

For a more accurate temperature sensor, a correction of the nonlinearity is
needed. When using the ratio of a PTAT signal and a bandgap reference, the
error typically shows a parabolic curve [36, ... ,39]. It can be shown that
this is in fact a logarithmic error, of which the Taylor polynomial has a
dominant quadratic component [39]. This can be compensated with a
signal based on the product of a PTAT signal with a positive temperature
coefficient, and a Vbe signal which has a negative temperature coefficient.
The multiplication can be obtained with a simple analog current multiplier,
as shown in Fig.5-38.
The PTAT, V BE, and reference signals are derived from a temperature
sensor signal as presented in section 2.3.1.
The temperature signals can be combined with the bitstream calibration

technique, as proposed in the eonfiguration of Fig.5-39. This smart
temperature sensor has been presented in [27].

Vcc ..... -.................. -".-.

i
4 ~ '-t------------'-'------"----- ··· · · -
In/ (~)
I
31
S'lvbe'lptat
Iret
Q1
,
Ij
0L. ........... ,. ....................................................... J

-50 O 50 100 150
Vee'______~------~------___
Temperature (0C) -
Fig. 5-38 Current multiplier for generation of a parabolic

compensation signal, and output current Inl versus
temperature.
clock
bit.ltream
output
,...--'----. Jiain
IO-bit adjust
bitstream
generator oJJ.let
lO-bit adjuJt
bitstream t--I~-+---\
Sigma-delta conve l1er
generator linearity
8-bit adjllsi
bitstream t---~-t---+--\
generator
Analog temperature sensor ci rcuit
Fig. 5-39 Configuration of the second, programmable smart

temperature sensor.
Using high-resolution bitstream generators the three different current

signals Irel , IT> and Inl can be accurately weighted with a digital value, and
summed at the input of the sigma-delta converter. The sum is converted
with respect to the sigma-delta reference current, into a digital output. The
three ratiometric signals 2lrellrel' Irllrel' and Inll rel have respectively

zero-order (constant), first-order (linear), and second-order (quadratic)

dependency on the temperature, as shown in Fig. 5-40. The supposedly
linear signal, however, has a small quadratic error. The contribution of
each of these signals can be varied from -1 to +1, so we have fuIl control
of the overall temperature transfer. Through a serial interface the bitstream
generators can be programmed in such a way that the temperature can be
measured accurately over a certain range.
3.0 r-r······ . ·......·. ·........·. ·. ·. · ·....· ·....·......· ....·. ·. ····....................... ·.... ···1
j Current RatlOs !
: :
2.0: 2/ret :
1 ~ ~
1.01 /
' ~
.! 7;;f
0.01.. ............................................................................................................. !
-50 O 50 100 150
Temperature (ac) _
Fig. 5-40 The three ratiometric temperature signals used in the

second smart temperature sensor.
A calibration shou1d proceed as foIlows. At three different temperatures,

preferably 'spaced' as far apart as possible, each of three programmable
signals must be measured separately. This can be done by consecutively
programming the bitstream generator of each signal with a fixed value,
and the other two with zero. The three signals can then be characterized
precisely, using second-order polynomials for the two
temperature-dependent signals. For a desired linear temperature transfer it
can then be caIculated what the contributions of each of the three signals
should be. The required values can then be programmed in the shift
registers of the chip. The sigma-delta output range can be mapped on a
temperature range of at least 75°C, and in between the operation limits of
the IC, from -40°C to 130°C.
Fig.5-40 shows how the parabolic linearity error, remaining after offset
and gain calibration, can be reduced to a much smaller third-order linearity
error, using the addition of a quadratic temperature signaL

Calibration llsing Sigma-Delta Analog-to-Digital Conversion
Error with and without Non-linearity Calrbration Qn % of FS)
05
03
OI
1-
-25 -10 20 35 50 80 95 110 125
-O 1
-O 3
-O 5
T (Celclus)
Fig. 5-41 Error in the temperature transfer. with and without

linearity calibration. for a temperature range of
-40 to 125°C.
5.6 Conclusion
The construction of the sigma-delta Analog-to-Digital converter allows
modification of the transfer by summing several analog input signals
which are modulated by digitally generated bitstream signals. The
sigma-delta (noise-shape) techniques can also be applied in the digital
bitstream generation. which makes it possible to calibrate very accurately.
However, the suggested configurations only allow for a low order of
calibration.
The bitstream technique is appropriate for integration in the smart sensor

concept, as demonstrated with the implementation of a smart temperature
sensor with bus-interface and programmable calibration in a bipolar chip.
For high accuracies, the required digital circuitry can only be integrated in
a feasible manner, using a more compact CMOS technology.

5.6 Conclusion
References
[1] J.c. Candy and G.c. Temes, Oversampled Delta-Sigma Data
Converters, IEEE Press, New York, 1992.
[2] R v.d. Plassche, Integrated Analog-to-Digital and Digital-to-Analog
Converters, Kluwer Academic Publishers, Boston/Dordrecht/
London, 1994.
[3] B.E. Boser, Design and implementation of oversampled analog-to-
digital converters, PhD Thesis, Stanford University, 1989.
[4] F.R Riedijk, Integrated Smart Sensors with Digital Bus Interface,
[5] W. Bennet, "Spectra of quantized signals", BeII Syst. Tech. J.,
VoI.BSTJ-27, July 1948, pp.446-472.
[6] J.c. Candy and 0.1. Benjamin, "The structure of quantization noise
from sigma-delta modulation", IEEE Trans. Communication, Vo1.29,
No.9, September 1981, pp.1316-1323.
[7] R.M. Gray, "Oversampled sigma-delta modulation", IEEE Trans.
Communication, Vol.35, 1987, pp.48 1-489 .
[8] R.M. Gray, "Quantization noise spectra", IEEE Transactions on
Information Theory, Vo1.36, No.6, 1990, pp.1220-1244.
[9] J.c. Candy, "A use of double integration in sigma-delta modulators",
IEEE Trans. Communication, VoI. COM-33, March 1985,
pp.249-258.
[10] RW. Adams, "Design and implementation of an audio 18-bit
Analog-to-Digital converter using oversampling techniques",
J. Audio Eng. Soc., Vo1.34, No.3, March 1986, pp.153-166.
[Il] J. Robert, G.C. Temes, V. Valencic, R. Dessoulavy, and P. Deval, "A
16-bit low-voltage CMOS AlD converter, IEEE Journal of
Solid-State Circuits, VoI. SC-22, April1987, pp.157-163.
[12] J. Robert and P. Deval, "A second-order high-resolution incremental
AlD converter with offset and charge-injection compensation, IEEE
Journal of Solid-State Circuits, Vo1.23, No.3, June1988, pp.736-741.
[13] Y. Matsuya K. Uchimura, A. Iwata, T. Kobayashi, M. Ishikawa, and
T. Yoshitome, "A 16-bit oversampling A-to-D conversion

technology using triple integration noise-shaping", IEEE Journal of

Solid-State Circuits, VoI. SC-22, 1987, pp.921-929.
[14] B.E. Boser and B.A. Wooley, "The design of sigma-delta modulation
Analog-to-Digital converters", IEEE Journal of Solid-State Circuits,
Vo1.23, 1988, pp.1298-1308.
[15] J.c. Candy, yc. Ching, and D.S. Alexander, "Using triangularly
weighted interpolation to get 13-bit PCM from a sigma-delta
Modulator", IEEE Trans. Communication, Vo1.24, 1976,
pp.1268-1275.
[16] R.E. Crochiere and L.R. Rabiner, "Interpolation and decimation of
digital signals - A tutorial review", Proceedings ofthe /EEE, Vo1.69,
1981, pp.300-331.
[17] E. Dijkstra, M. Degrauwe, J. Rijmenants, and O. Nys, "A design
methodology for decimation filters in sigma-delta AID converters",
Proceedings ISCAS'87, 1987, pp.479-482.
[18] A.V Oppenheim and R.W. Schafer, Discrete-time Signal Proces ing ,
Prentice-Hall, New Jersey, 1989.
[19] L.R. Carley, "An oversampling Analog-to-Digital converter
topology for high-resolution signal acquisition systems", IEEE
Trans. Circuits and Systems, Vo1.34, 1987, pp.83-90.
[20] Pc. de Jong, G.C.M. Meijer, and A.H.M. van Roermund, "A new
dithering method for sigma-delta modulators", 1. Integrated Analog
Circuits and Signal Processing, VoI. 1O, August 1996, pp.193-204.
[21] J.c. Candy and A. Huynh, "Double interpolation for
Digital-to-Analog conversion", IEEE Trans. Communication,
Vo1.34, 1986, pp.77-81.
[22] L.R. Carley and J. Kenney, "A 16-bit 4'th order noise-shaping DIA
converter", Proceedings of the Custom Integrated Circuits
Conference (C/CC), 1988, pp.2I.7.1-2I.7.4.
[23] PJ.A. Naus, E.C. Dijkmans, E.F. Stikvoort, AJ. McKNight,
D.J. Holland, and W. Bradinal, "A CMOS stereo 16-bit DIA
converter for digital audio", /EEE Journal of Solid-State Circuits,
Vo1.22, 1987, pp.390-395.
[24] A.B. Carlson, Communication Systems: an Introduction to Signals
and Noise in Electrical Communication, McGraw-Hill Book Co.,
Singapore, 1986.

5.6 Conclusion
[25] G. v.d. Horn and J.H. Huijsing, "Sensor calibration using

sigma-delta bitstream modulation", Proceedings EuroSensors VIII
Conference, Toulouse, France, 1994, pp.34.
[26] G. v.d. Horn and J.H. Huijsing, "Sensor calibration using
sigma-delta bitstream modulation", Proceedings National Sensor
Technology Conference, J994, pp.183-186.
[27] G. v.d. Horn and IH. Huijsing, "Smart temperature sensor with
integrated bitstream calibration", Proceedings ESSCJRC'95, LiIle,
France, 1995, pp.234-237.
[28] N. Kouvaras, "Delta-modulation/P.C.M con verter" , Electronics
Letters, Vol.14, No.20, 1978, pp.660-662.
[29] N. Kouvaras, "Some novel elements for delta-modulated signal
processing", The Radio and Electronic Engineer, Vol.51 , No.5,
1981, pp.241-249.
[30] M. Freedman and D.G. Zrilic, "Nonlinear arithmic operations on the
delta-sigma pul se stream", Signal Processing, Elsevier, Vol.21,
1990, pp.25-35.
[31] P. O'Leary and F. Maloberti, "A bitstream adder for oversampling
coded data", Electronics Letters, Vol.26, No.20, 1990,
pp.1708-1709.
[32] F. Maloberti, "Non conventional signal processing by the use of
sigma delta technique: a tutorial introduction", Proc. IEEE
International Symposium on Circuits and Systems, San Diego, USA,
1992, pp.2645-2648.
[33] F. Op 't Eynde, "A power metering ASIC with a sigma-delta based
multiplying ADe", Digest of Techn. Papers ISSCC, 1994,
pp.186-187.
[34] Y.F. Dias, "Signal processing in the sigma-delta domain",
Microelectronics Journal, Vol.26, No.6, 1995, pp.543-562.
[35] O. Feely and L.O. Chua, "The effect of integrator leak in
Sigma-Delta Modulation", IEEE Trans. Circuits and Systems,
Vol.38, No.II, November 1991, pp.1293-1306.
[36] F.R. Riedijk and J.H. Huijsing, "An integrated absolute temperature
sensor with sigma-delta A-D conversion", Sensors and Actuators A,
Vol.24, 1992, pp.249-256.

[37] P. Krummenacher and H. Oguey, "Smart temperature sensor in

CMOS technology", Sensors and Actuators A, VoI.21-23, 1990,
pp.636-638.
[38] A Bakker and J.H. Huijsing, "Micropower CMOS temperature
sensor with digital output", IEEE Journal of Solid-State Circuits,
Vo1.31, No.7, July 1996, pp.933-937.
[39] G.C.M. Meijer and AW. van Herwaarden, Thermal Sensors,
Institute of Physics Publishing, Bristol/Philadelphia, 1994.

Calibration
using Digital 6
Signal Processing
6.1 Introduction
It is not the aim of this chapter to investigate or discuss aII the possible
digital forms of sensor calibration techniques, but rather to demonstrate
the feasibility of the progressive polynomial calibration method proposed
in chapter 3.
It is acknowledged that digital compensation is the most attractive

technique for high-resolution sensor calibration. The implementation in
software for a digital processor enables an advanced, complex but also
ftexible correction of sensor signals. Hence, for most linearization
techniques explained in chapter 3, examples of computer software
implementations can be found in literature [1,2,3,4]. For the simpler
linearization techniques, implementation in microcontroller programs has
also been attempted [5,6].
On the other side, a hardware implementation allows to optlmlze for

speed, chip area, or power consumption, but lacks ftexibility [7,8]. In the
next section we will first shortly consider the possibility to implement the
proposed polynomial calibration method in digital hardware.

Calibration using Digital Signal Processing
Using modern computers, it obviously is possible to realize an advanced

digital sensor calibration with high orders of linearization and cross-
sensitivity compensation. Such might be interesting for instrumentation
systems, but not for the integration on a smart sensor. We must recall our
purpose: to produce smart sensors of which the output signal is
standardized with respect to a reference signal, and to achieve this at
minimum production cost.
When a digital (either dedicated or standard) processor can be (hybridly)

integrated at an acceptable price, still another cost aspect must be
considered. Namely the time required per calibration per sensor. This time
can be minimized not only by using an effective calibration method, but
also by minimizing the data communication needed between the sensor(s)
and the calibration controller (computer). Such can be obtained if the
sensor calibration circuit or program is capable of calculating it's own
calibration coefficients. In that case, the calibration controller only has to
send a signal to indicate which reference signal is applied during the
calibration phase, see section 1.4.2. The signal can be sent to many sensors
at a time, which then process their own output signal and calculate their
own calibration coefficient, alI in paralIel. It is cIear this may save time
with respect to the situation in which a central computer has to read each
sensor output, calculate each calibration coefficient, and store it in each
sensor, one at a time.
For the proposed polynomial calibration method the calculation of the

calibration coefficients is not complicated, and may very well be
incorporated in the digital implementation.
6.2 Hardware implementation

For cIarity the equations for the one-dimensional polynomial calibration
procedure, explained in section 3.3.1, are summarized in Table 6-1. Since
we now deal with digital numbers and not with physical signals, the
reference signal Yret has been left out. It means a different 'unit' is
incorporated in each calibration coefficient.
It was already cIear from the cascaded implementation shown in Fig. 3-16
on page 82, that the procedure has a repetitive character. The same
algorithm is repeated for each calibration step, using the previous output
hn_dx) and the polynomial product IT {hlx)-Yd as inputs. While in analog

Table 6-1 Overview of equations for the one-dimensional

progressive polynomial calibration technique.
calibrated function calibration coefficient
initial y = f(x) --
step 1 fix) +a l al = YI - fi x l )
h/(x)
{Y2 - h l (x 2)}
step 2 a2 =
hl(x) + a 2 · {hl(x) - YI}
h2(x) {h l (x 2) - YI}
{Y3 - hix 3)}

step3 hix) + a 3 · {hl(x) - YI}' {h 2(x) - Y2} a3 =
hlx) {hl(x)-YI}' {h 2(x)-Y2}
... ...
n-I Yn - hn_I(xn)
stepn
hn_I(x) + an' IT {h,(x)-y,} a n = n-I
h,,(x)
,= 1 TI
{h,(xn)-y,}
,= 1
hardware the multiplication and summing actions can be realized very
compactly, in digital hardware such arithmetic functions require complex
circuitry. Therefore, it makes more sense to include the algorithm in a loop
as proposed in Fig.6-1, and so reduce the number of multipliers and
adders. The highlighted algorithm is the same as used in Fig.3-16, besides
the absence ofYret. The signals indicated in Fig.6-1 are for the n-th and last
calibration step, supposing the calibration points YI to Yn are known, and
the coefficients al to an have already been caIculated and stored.
The output and correction term of a calibration step are stored in 'sample
& hold' registers, so that they can be used as inputs in the next calibration
step. For each calibration step i the values of Y,_1 and the required
coefficient ai are multiplexed from a memory, and the loop is executed
once. The first calibration step, for offset, may be simplified and executed
first, outside the loop as shown in the figure. At the second calibration step
the inputs are connected in such a way that the first multiplier outputs the
signal (hdx)-YI) which is used for the gain correction. At succeeding

...---------l Sample & Hold 1+----,

Registers
H-.L....t-. hll(x)
n
11 - 1
, =1
{",(x) - Y,}
Memory
Fig. 6-1 Algorithmic loop implementation of the one-dimensional

polynomial calibration method, at the n-th correction
step of the measurement phase.
calibration steps the inputs are switched to the previous output values
stored in the SH registers , and so the product term for the polynomial
corrections is built-up.
As mentioned in the introduction, it is desirable that the calculation of the

coefficients, during the calibration phase, is incorporated in the
implementation itself, and not externally. Knowledge about the desired
output values YI to Yn can be built in or received from the calibration
controller. By expanding the block diagram as illustrated in Fig.6-2, the
calibration coefficient can then be calculated internally. During calibration
the controller sends a signal to indicate which input reference is being
applied to the input of the sensor.
At the n-th calibration step, the reference signal XII is applied by the
calibration controller, corresponding to the desired output Yn- The input
f(x n ) is sampled, and proces sed through n-I correction steps. The n-th step
is then executed with an=O, and h n. 1(xn) is obtained at the output. The
output signal is subtracted from the desired output Yn' and the result
{yn-hn-I(xn)} is used to obtain the correct value for al/" This can be
achieved in two ways . As shown in the figure , the signal {Yn-hn-I (x n)} can
be divided by the value of the available correction term n {hlx)-Yl}, in

r-- - -- ----i Sample & Hold 1+------,

Registers n {h ;(x) - Y; }
II -
; =I
I
YII
(1) read all=O

r -- L--....J.-_-,
Fig. 6-2 Algorithmic loop implementation with calculation of the

calibration coefficient, at the n-th step of the calibration
phase.
accordance with the formula's in Table 6-1 . The other way is to compute
aII algorithmically. As in a Successive-Approximation converter, the bits of
aII are determined, and programmed one by one from MSB to LSB, based on
the polarity of the signal {Y1l-h ll _1(XII)}. Bit by bit that signal is reduced to
zero, the output is equalized to the desired value, and the resulting value
for aII is stored in memory.
When considering implementation of the block diagrams in Fig.6-1 and

Fig.6-2 in digital hardware, the following objection comes to mind. The
digital implementation will turn out to be very similar to a standard
microprocessor. Adders and multipliers are needed; these functions are
usually efficiently implemented in the Arithmetic Logic Unit (ALU) of a
microprocessor. Furthermore, the necessary memory addressing is also
present in any microprocessor, and the sample & hold registers are
available for example as 'stack' memory. The question rises whether the
development of an custom processor for digital calibration would be worth
the investment. The only advantages of a dedicated digital processor for
the digital calibration would be the possibility to optimize for speed,
power consumption or chip area, whereas flexibility and debugging

options have to be missed. Hence the option to implement the progressive

polynomial calibration method in a small software program for a standard
microcontrolJer is preferred for experiments. It is presented in the next
section.
6.3 Software implementation
6.3.1 Sensor system configuration

The purpose of the implementation of the two-dimensional calibration
procedure in a microcontroller program [9,10], is to eventually apply it in
a smart sensor system set-up as proposed in Fig.6-3.
Smart Sensor Package

~--------------------, v++ computer
clock interface
data
gnd
, Cii Q)
ot
«Q)
u
cn <ti
> :::J't:
<le:: .o Q) micro-
W o ~
u controller
L ____________________ ~
memory
Smart Sensor Inteiface Circuit
1s2 bus for

multiple sensors
Fig. 6-3 Pressure sensor system conjiguration for including a

calibrationJunction in microcontroller software.
Using a smart sensor interface chip with IS 2 bus interface the

microcontroller can read out the pressure sensor and an integrated PTAT
temperature sensor [Il]. A multiplexer controlled by the bus interface
makes it possible to read out both sensors using a single sigma-delta
AD-converter. The output of the pressure sensor is determined by the
applied pressure and, because of cross-sensitivity, to the operating
temperature of the sensor. The sensor interface with PTAT sensor should
thus be integrated close to the pressure sensor so that the measured
temperature is related closely to that of the pressure sensor. Different
levels of integration are possible, depending on which components can be

integrated on the same chip, multi-chip module, or package. It is also

possible to keep the microcontroller external, so it can be used for multiple
sensors in a larger bus-wired system.
For the development of the microcontroller program it was assumed that

the microcontroller can invoke samples of the pressure and temperature
sensor 1, and that it must autonomously compute the corrected output
pressure based on stored calibration data. Furthermore, it is desirable that
the microcontroller program can autonomously compute the required
calibration coefficients, when the sensor system is inserted in a calibration
set-up. Through a computer interface the microcontroller must then be
signalled about the reference inputs being applied and the desired output
value.
The calibration program and the calibration coefficients must be stored in

the memory of the microcontroller, preferably a non-volatile memory such
as PROM.
6.3.2 Calibration algorithm

The two-dimensional polynomial calibration method has been explained
shortly in section 3.3.2. The equations for an NxM calibration are adapted
for a pressure sensor calibration, and summarized in Table 6-2. The
computation of the calibration coefficients, using NxM calibration
measurementsf(Pn' T m) is presented in Table 6-3.
The pressure sensor output signal is represented by y=f(p, T), whereas

hnm(p, T) represent the subsequent calibrated sensor functions. The
calibrated output signal is given by the last calibration function hNM(P, T).
For the calibration of the sensor, N different pressure values have been
used, corresponding to desired output values y I to YN At each pressure
value, M different temperatures have been used for calibration. The signal
T represents the output of the temperature sensor rather than the actual
temperature. Also for the temperature calibratioll" points T m the
temperature sensor outputs, measured in the calibration phase, must be
used. In that way calibration of the errors of the temperature sensor will be
inc\uded automatically.
1 For this purpose the ffilcrocontroller needs to be equipped with an IS 2 bus interface,
which can be realized either in hardware or in software. These developments have
not yet been completed.

("î
00 e:.
O s:..,
Table 6-2 Overview of calibration functions for the two-dimensional progressive polynomial calibration method. Ql
....
o'
:::
...... m
Z ~
-l n 1 ", M S'
(JQ
1:'5
;:<l O
~ ~'
tTl M-I
CI [
1 hll(p, T) = f(p, T) + aII ... hIM(P,T)= h1,M_I(P,T) + a lM · n(T-Tm)
CIl lJl
~ m=I ~'
~
V :::
~ e:.
CIl k V M-I ~
tTl
Z
(Il 2 h 21 (P, T) = h1M(P, T)+ a 21 ·{h 1M (P, T) -YI} ... h 2M (P,T) = h 2,M_l(P,T) + a1M·{hIM(P,T)-YI}· n(T-Tm)
O ~
<Il
;:<l m=I <Il
(Il
:-- S'
(JQ
tl
(\) J..-..--'"
~
0<;'
'" .. , ... ...
;::! V
>:>
;::! ~
\::l.. ~
N-I -- M-l N-I

12: N hN1(P,T)= hN_I,M(P,T) +a N1 TI {hnM(p,T)-y n } ... hNM(P, T) = hN,M_I(P, T) + a NM TI (T-Tm) TI {hnM(p, T) -Y n }
&: n=I m =1 n =1
i:l......
O'
;::!
Table 6-3 Overview of calibration coefficients for the two-dimensional progressive polynomial calibration
method.
Z
-l m 1
- n
M
2:1l T=T 1 T=TM
::o
~
tI]
o YI - hl.M-1(PI' T M)
CIl 1 a lM = M-j
s:;J> al! = Y1 - f(pl' TI)
P=PI TI (TM-Tm )
q m=l
CIl
tI] ------+----
Z ~
CI'l
o::o Y2 - h IM (P2' TI) Y2 - h 2,M-I(P2' T M)
CI'l
2 a 21 Q2M = M-I
P=P2 {h IM (P2' T1)-YI} {h IM (P2,TM)-YI}' TI (TM-Tm )
tl
~
c., m=l
0;' ~
;:: 0'1
i::l 1...1
;:: ~
i::l... --" 'J1
o
-------r--- ~
~
c:;: A:-- ~
~ YN-hN-I,~PN' TI) YN- hN,M-I(PN' T M) ~
.....
(5' N aNI = N-I a NM = N-I M-I 3"
;:::
P=PN TI {hnM(P n, TI) -Y n } TI {hnM(P n, TM ) -Y n }' TI (TM-Tm) '"O
ii'
n =1 n =1 m=l 3ro
=
S"
00 ....
o"
=
The arrows in Table 6-2 and Table 6-3 indicate the direction of execution .
Along the rows of the tables the temperature dependency is calibrated,
whereas the columns represent the calibration steps on the pressure axes. It
is important that at each pressure calibration step the sensor curve is first
corrected for temperature before proceeding to the next pressure
calibration steps. Otherwise the temperature dependency gets multiplied
by the pressure corrections, and gets harder to rectify. The example may
c1arifies this . It is assumed that generally the sensor signal depends more
on pressure than on temperature.
The algorithm for the two-dimensional polynomial calibration has been

tested, using the mathematical computer program MATLAB. This gives the
obvious advantage that we can study the behavior of the calibrated sensor
response, where in practice this would require an impractical number of
measurements . For that purpose, several fictive sensor responses were
generated, using mathematical descriptions of two-dimensional functions
with different kinds of nonlinearity and cross-sensitivity. Fig.6-4 shows
one such a uncalibrated sensor response used for testing. It c1early shows a
slightly nonlinear pressure dependency and a small cross-sensitivity to
temperature. The frame indicates the ideal transfer function.
Output
(normalized)
....
'
'
.' .
Temperature
(normalized) Pressure
-, -, (normalized)
Fig. 6-4 Sensor transfer function (example), output versus

normalized pressure and temperature.

The calibration is now started by measuring the sensor output at one

pressure P 1, for different temperatures. After correction at for example
three temperature points TI' T2, T 3, the output is more or less constant
along the axis p=p 1, y=y l' Fig.6-5 shows the sensor transfer in that
situation.
Output
(normalized)
'. '
"
:.
1
Temperature O. 5 ~
(normalized)
·1 ·1
Fig. 6-5 Sensor transfer curve, after calibration at one pressure

point and three points along the temperature axis.
It is now possible to rotate the response surface around that axis, using
{h(p,T)-YI} for the gain correction . Because the error at P=PI is already
made very small , it will not get affected much by the gain correction . At
the next pressure calibration point P2 the gain correction is determined.
And again the temperature-dependency is reduced first by calibration at
temperature points TI' T 2, T 3, before proceeding to the next calibration
step.
After 5x3 correction steps the transfer curve seems practically identical to
the desired transfer curve, see Fig. 6-6.

Out put
(normalized)
..
1
Temperature O.5 ~
(normalized)
·1 ·1
Fig. 6-6 Corrected sensor transfer curve, using a 5x3 polynomial

calibration.
The progress in the calibration can be examined better by looking at the

error surface. Fig. 6-7 shows the error between the actual sensor output and
the desired output as a function of pressure and temperature, without
calibration.
After correction based on a 5x3 calibration, the error surface is ftattened as

shown in Fig .6-8. The error is made zero in aII the used calibration points
(P1,T 1) to (Ps,T,,). Along the temperature and the pressure axes the
third-order and fifth-order polynomial may be identified. The labeled
points on the "ground" surface indicate the calibration order. After
calibration the peak in the error surface is reduced to about I percent of the
full-scale outpul.
Further reduction of the error may be obtained by using more calibration

points. Fig.6-5 shows the error surface in case the correction was based on
a 7x5 polynomial calibration . The remaining error peak was reduced to a
value smaller than 0 .2%.

Error
Temperature
(normalized)
·1 ·1
Fig, 6-7 Initial errar surface,
. ~ '.: ..
Error
.. ' . ...
~
. .'
" ,
. ' ",
"
" ,
Temperature
(normalized)
·1 ·1
Fig, 6-8 Errar surface after correctÎon based on a 5x3

polynomial calibration,

Error
t
-3
X 10
. . ... . .•...
.. .
.- ...
•..
' .
: .•.... : ,- ..~
Temperature
(normalized)
·1 ·1
Fig. 6-9 Error surface after correction based on a 7x5

polynomial calibration.
The reduction of the error has been examined for several different sensor
transfers, and different numbers of calibration steps.
Calibration coefficients
The calculation of the calibration coefficients suggested in Table 6-3,
implies a certain order of the calibration measurements. The calibration
coefficients in the table are calculated row by row. In case of a pressure
sensor the order is disadvantageous, namely at each pressure calibration
point Pn' the output must be measured for different temperatures TI to TM '
This means that a lot of temperature sweeps, NxM, would have to be
made. Because of the large settling time of a temperature chamber (oven),
in which pressure sensors are calibrated, the calibration would take a long
time and become expensive. For most pressure sensors, the test pressures
for the calibration can be generated much quicker. Hence it is preferred to
minimize the number of temperature steps, and perform multiple pressure
measurements at each temperature.

Fortunately the calculation of the calibration can easily be modified to

enable a different arrangement of the calibration measurements. Looking
at the equations of the calibration functions, see Tables 6-2, we can
observe that for T=T I :
Thus we don't have to compute the first row completely to be able to

calculate the calibration coefficient a21 on the second row, as suggested in
Tables 6-3. Because for aII equations in the first column:
the calculation of the calibration coefficients can be modified as shown in

Table 6-4. Now the coefficients can be calculated column by column as
indicated with the arrows. Accordingly, the calibration measurements are
done in that order as well. In that case the number of temperature sweeps
is minimized to M, whereas the pressure needs to be changed MxN times.
In fact the calibration coefficients can be calculated in any order working

through a similar table but always from left to right or downwards. In that
way the calibration can also be expanded afterward, using extra columns
and/or rows of ca1ibration measurements.

n
00 e:.
00 o:
Table 6-4 Overview of calibration coefficients for the two-dimensional progressive polynomial calibration ~
method. s'
-=
......
Z CI)
-l
C!l
fi 1 M =
S'
O n T=T I T=TM ~
;:o 1 ", 1 O
~ tiC'
C!l
O YI - hl,M_I(PI' T M) [
VJ 1
3:: ali = YI-!(PI,T I ) .,.
I a lM = M-I 'Jl
>- P=PI
I IT (TM-Tm) tiC'
~ m=1
=
e:.
VJ I "'CI
C!l
Z
(/)
Y2 - h 2,M-I(P2' T M)
g
('!)
O Y2- h ll(P2' TI) CI)
;:o 2 CI)
(/) a 21 = a 2M = M-I
P=P2
5'
i::l {h ll (P2' TI)-YI} {h IM (P2,TM )-YI}' IT (TM-Tm) ~
m=1
00'
'"'"
;:,:
>:>
;:,: .. , ... ...
>:>..
II II
~ ~I
&: yN-h N _ I , I(PN' TI) I YN- hN,M-I(PN' T M)
~ N aNI = N-I a NM = N-I M-I
~. ...
;:,: P=PN IT { \ I (Pn' TI) -Yn} IT {h nM (p n' T)
M - Yn } . IT (TM - T m
)
n= I n=I m=1
6.3.3 Calibration program

The microcontroller we used for evaluation of the calibration method is
the 80c552 which is an inexpensive 8-bit microcontroller driven by a
12 MHz c1ock. The system software consists of a program written in C,
which is compiled, assembled, and then downloaded into the
microcontroller memory through an RS232 interface. And the program
could be tested on microcontroller evaluation board, using PC-
communication.
The program consists of two parts. First the calibration part, to calculate
the calibration coefficients when applying reference signals to the sensors.
At each calibration step the output signals are measured, and proces sed
according to the equations in Table 6-2, to compute the calibration
coefficients as described in Table 6-4. Second the measurement or
correction part, to calculate the corrected values of the pressure output,
according to equations in Table 6-2, using the sensor outputs and the
values of calibration coefficients already calculated during the calibration
cyc1e. From the implementations of the one-dimensional polynomial
calibration it was already c1ear that both the calibration and the correction
can be implemented algorithmicaIly. In programming this means that the
calculations are based on the repetitive use of the same procedure. For the
two dimensional calibration a nested loop is needed. Altogether it results
in a very compact program. The complete calibration program, together
with the calibration coefficients and reference data for a 5x5 calibration,
required only 6.6 kilobytes memory of the microcontroIler.
InitiaIly the program is developed as an interactive program which during

the calibration process inquires the user about the reference data and the
desired output signal, at each stage of the calibration process. Then the
microcontroller proceeds with the necessary calculations, which involves
calculating a new calibration coefficient and checking the error limit
supplied by the user. Once this limit has been reached, entering more
calibration data can stop, and the measurement cyc1e can start. Once the
optimal number of calibration points and reference signals has been
chosen, the system can be made autonomous by storing the program and
the reference data in the microcontroller PROM, and by using hardware
interrupts to indicate that certain reference signals (pressure and
temperature) have been applied for calibration. During batch calibration,
many of such pressure sensor systems can then be connected to the same
interrupt signals, and each system independent1y calculates and stores its
own calibration coefficients.

The microcontroller program was operated and tested, using the fictive
sensor data already used in the simulations. As expected, the actual results
from the microcontroller program were found to coincide with the
simulation results. With the microcontroller type already mentioned, and
32-bit precision ftoating point arithmetic, it was found that the time
required to calculate one corrected value during the measurement cycle
was approximately 30 ms based on 5x5 calibration steps. This is fast
enough for sensors with a signal bandwidth lower than 15 Hz. Of course
speed and accuracy can be exchanged. To increase speed one could for
example reduce calibration accuracy to 5x3 or 3x3 steps, chose a lower
arithmetical resolution, or chose a faster 16-bit (more expensive)
microcontroller. Maximum speed can then be in the order of 100-1000
samples per second.
Finally, Fig.6-1O gives an overview of the remaining peak errors after

several calibration steps, for six different sensor curves that were used for
testing of the two-dimensional calibration method.
en r ca e cos.. (6)
Fig. 6-10 Overview of the rematntng peak errors for different

sensor cases, when increasing the number of the
calibration steps.

6.4 Conclusion
For the six test cases, different pressure and temperature dependencies
have been combined, using polynomial, root, exponential, and logarithmic
functions. The large bars represent the initial linearity error, since with a
2x2 calibration only the linear errors are corrected. The figure shows how
the linearity error reduces when increasing the number of calibration
measurements to respectively 3x3, 5x3, 5x5, and 7x5. One has to keep in
mind that the bars indicate the maxima of the error surfaces
6.3.4 Future work

Still some work remains to be done before the calibration program can
actually be applied in the intended smart sensor configuration of Fig.6-3
and tested with real sensor data. The developed pressure sensor interface
described in section 2.3.2 needs to be combined on one chip with a smart
temperature sensor as described in section 5.5.1. The microcontroller
program must be expanded with a bus-interface for sensor read-out. Also
the packaging for the combination of the sensor, the interface le, and the
microcontroller must be investigated. Further investments are required to
set up a (computer-) controlled measurement configuration for pressure
and temperature, which is needed to calibrate and test such smart sensor.
6.4 Conclusion
In the consideration of digital implementations of calibration techniques
we concentrated on the proposed polynomial calibration method. The
step-by-step correction of the sensor transfer curve can be set up in an
algorithmic way. For each calibration step the same calculation procedure
is then used. This applies for the correction of the sensor transfer function
as well as for the computation of the calibration coefficients. It allows a
compact realization in hardware in the form of an algorithmic loop
implementation. When implemented in software, the required calibration
program can also be made very compact. The latter option can be
successfully applied in future microcontroller-based smart sensors. With
respect to the implementations of the previous two chapters, it offers the
advantage of an advanced two-dimensional calibration, at the costs of
(hybrid) integration of a microcontroller, and some sacrifice in sensor
signal bandwidth.

References
[1] J.E. Brignell, "Software techniques for sensor compensation",
[2] D. Patranabis and D. Gosh, "A novel software-based transducer
linearizer", lEEE Trans. lnstrumentation and Measurement, Vo1.36,
No.6, December 1989.
[3] S.B. Crary, w.G. Baer, J.c. Cowles, and K.D. Wise, "Digital
compensation of high performance silicon pressure transducers",
[4] S. Huang, R.Z. Morawski, and A. Barwicz, "Static calibration based
on superposition of splines in one variable", Proceedings lMTC'96,
June 1996, pp.49-53.
[5] w.T. Bolk, "A general digital linearising method for transducers",
Journal of Physics, VoI. E: Scientific Instrumentation, 1985,
pp.61-64.
[6] PN. Mahana and F.N. Trofimenkoff, "Transducer output signal
processing using an eight-bit microcomputer", lEEE Trans.
Instrumentation and Measurement, Vol.IM-35, No.2, June 1986,
pp.182-186.
[7] S.R. Ashley, M. Muggeridge, and 1. Lucas, "An inexpensive digital
linearizer for nonlinear transducers", Journal of Physics E:
Sci. Instrum., Vol.l1, 1978, pp.576-580.
[8] C. Berthoud, M. Ansorge, and F. Pellandini, "Effective static
response compensation suitable for low-power ASIC
implementation with an application to pressure sensors",
Proceedings lMTC'96, June 1996, pp.1168-1173.
[9] K.F. Lyahou, G. v.d. Horn, and J.H. Huijsing, "A non-iterative
polynomial 2-dimensional calibration method implemented in a
microcontroller", Proceedings IMTC'96, June 1996, pp.62-67.
[10] G. v.d. Horn, K.F. Lyahou, and J.H. Huijsing, "Calibration and
linearization method for microcontroller-based sensor systems",
Journal on Measurement + Control, Vo1.29, No.9, November 1996.
Vo1.14, No.3, November 1997.

Summaryand
Conclusions
7
Smart Sensors
Electronic sensors are applied in industrial machines, as welI as in
consumer products, in order to enhance functionality, or to improve
performance and reliability. Currently, the application of sensors requires a
high-Ievel experti se, due to the large diversity in sensors and,
correspondingly, in sensor output formats, and in the required electronic
sensor interfaces. The application of electronic sensors will be greatly
stimulated if the sensors are equipped with smart electronics, which
provide conversion of the sensor signal to a standardized digital output and
bus interface. Using modern silicon IC-technology, such smart sensors can
be produced in high volume and at low cost. This will make them
available for application in more products, and may also lead to new
products. A smart sensor configuration, composed of silicon sensor(s),
analog front-end circuit, one-bit sigma-delta AD-converter, and a serial
bus interface, forms a suitable concept for complete integration.
Calibration
For standardization, a fixed transfer from sensor input to electrical output
is desired for each sensor (type). Due to variations in the production, most
sensors exhibit some unpredictable errors with respect to the desired
transfer. Each sensor needs to be calibrated individualIy to accomplish the

Summary and Conclusions
desired fixed transfer. This has to be realized in an inexpensive way, which

means that the calibration procedure has to be automated in the sensor
production line. Currently, sensor calibration requires an expensive (time
consuming) laser-trimming technique. Automated calibration could
become feasible if electronics are included in the smart sensor, to enable
correction of the sensor transfer function in a digitalIy programmable
(computer-controlIed) way. Furthermore, the costs of calibration in a
sensor production line can be reduced by executing a minimum number of
calibration measurements, and by measuring a maximum number of
sensors at a time. The latter is facilitated if the smart sensors are equipped
with a bus interface. The first must be obtained by choosing an effective
linearization method.
Linearization
Different linearization methods have been examined. GeneralIy, the use of
a single (fulI-range) interpolation function, or a set of (sub-range)
interpolation functions (splines), is the best choice for linearizing the
sensor transfer curve, using a minimum number of measurement data.
ClassicalIy, the sensor transfer curve is first measured completely (alI
calibration points), and then alI the required calibration coefficients are
calculated. The progressive polynomial calibration method is proposed,
which linearizes the sensor transfer curve in a step-by-step manner. At
each step, one calibration measurement is done, and one calibration
coefficient is caIculated and applied immediately. Every succeeding
calibration step is done in such a way that the previous calibration(s)
remain undisturbed. At each step, the calibrated transfer curve progresses
towards the desired transfer curve. In addition to the linearization of a
one-dimensional sensor transfer function, the progressive method is also
worked out for the two-dimensional polynomial calibration of sensors,
showing a cross-sensitivity error.
Implementation
At several places in the signal processing chain of a smart sensor, the
overall transfer function can be modified. TraditionalIy, the sensor transfer
function is adjusted in the analog (physical) signal domain, by means of
trimmable components, or in the digital domain, by using a digital
memory with a look-up table, or by using a data-acquisition computer.
These three options are problematic concerning the desired features for the
calibration function, respectively, programmability, a low number of
calibration points, and integration with the sensor.
194 INTEGRATED SMART SENSORS: Design and Calibratiofl

It was shown that an advanced two-dimensional calibration procedure,
based on the progressive polynomial linearization method, can be
implemented in a simple microcontroller program. The calibration
program has been tested with computer-generated fictive sensor data, for
different numbers of calibration steps, up to 7x7 measurement points. This
(software) calibration feature can be added to one sensor excJusively, by
hybrid integration of a microcontroller die and a sensor die in one
package. After the packaging, an automated sensor calibration procedure
can be inserted in the production line in order to manufacture calibrated
smart sensors. At the current technology level, it is often more cost-
effective to use one microcontroller, in a complete (bus-wired)
data-acquistion system, for readout and calibration of multiple sensors.
However, this implies that the sensors will still have to be calibrated by the
customer who applies the sensors in the data-acquistion system.
Alternative options have therefore been examined.
In the analog signal path, the sensor transfer function can be altered in a
digitally programmable way, by using Digital-to-Analog (DA) converters.
Calibration of offset, gain, and cross-sensitivity errors can be implemented
in a cJassical approach. These implementations can be modified in
accordance with the progressive calibration method, to eliminate influence
between successive settings (offset, gain, etc.), and to circumvent iteration
of the adjustments. A programmable analog circuit is presented which
implements four steps of the progressive polynomial calibration method.
Programmable analog circuits can be used for sensor calibration with
limited resolution (8-10 bit) and limited correction of higher-order effects
(nonlinearity/cross-sensitivity). As an advantage, a large bandwidth can be
maintained for the (analog) transfer function, as it is not limited by any
digital sampling or processing speed.
As an alternative for the DA-conversions, digital pulse modulation can be

used to modify the transfer of an analog signal. This technique has been
elaborated in combination with sigma-delta AD-conversion. The
technique multiplies analog input signals with digitally-controlled
bitstream (pulsed) signals. This offers the possibility of a high-resolution
correction. Because the signals are aII summed at one point, the technique
should be applied only to calibration of (Iower-order) non-idealities that
can be compensated with additive correction signals. The use of the
bitstream modulation technique is demonstrated by the realization of two
smart temperature sensor configurations, providing calibration of offset,
gain, and second-order linearity errors.
INTEGRATED SMART SENSORS: Design and Calibratioll 195

Summary and Conclusions
To summarize, the Table 7 -1 rates the important features of the examined

implementations of a programmable calibration function.
Table 7-1 Ratings for the features of the differellt

implementations of a programmable sensor
calibration function.
Analog Bitstream Digital

Calibration Signal Modulation, Signal
Features Processing AD-conversion Processing
Offset, Gain Cal. +++ +++ +++

Linearity Calibration + + +++
Cross-sensitivity Cal. + + +++
Progressive Poly-
+ +++
nomial Calibration
Resolution ++ ++
Bandwidth ++ +
Integration with
++ +++
Sensor (interface)
Flexibility +++
---=impossible/very low, ... ,+++=very suitable/very high
For low-resolution calibration of a small number of error corrections, the

use of an analog circuit with programmable DA-converters is feasible. It is
suitable for integration with the sensor or sensor interface, provided that a
high-density (VLSI) integration technique, compatible with the sensor
technology, is available. If the advantages of the analog circuit technique,
namely, a compact integration (relatively low number of transistors) and
large 'analog' bandwidth, are not really needed, then a digitally oriented
calibration technique is favored.
For a low number of high-resolution corrections, the bitstream modulation

technique is the best option. It is also a suitable technique for integration
of the calibration circuit in the proposed smart sensor concept. Because of
the amount of digital circuitry required for the generation of the bitstream
signals, a CMOS implementation would be more compact and thus
preferable.
196 INTEGRATED SMART SENSORS: Design and Calibratioll

For high-resolution calibration with an advanced correction method (such
as for higher-order polynomial Iinearization and cross-sensitivity
compensation), the use of a microcontroller or a dedicated digital signal
processor is the best option. Advances will still have to be made to make
the integration of such digital signal processors with sensors economically
feasible.
Several techniques have been presented which make it possible to include

a digitally programmable calibration function in a sensor interface circuit.
To become useful, these techniques must be combined with (existing)
non-volatile memory techniques for storage of the digital calibration
coefficients. A calibration circuit can be integrated with each sensor at an
increase in the production costs, which is related to the silicon surface of
the total calibration circuit. However, by making use of this enhanced
functionality of the sensor for automation of the calibration procedure, the
production costs of accurate (factory-calibrated) sensors can be reduced.
Some costs per calibration measurement will remain (heating a
temperature chamber costs time). For some kinds of sensors the netto
effect will be a reduction of production costs per sensor. Considering the
rapidly increasing integration densities and decreasing price per transistor,
an integrated smart sensor with calibration feature will become more and
more viable in the fu ture. Furthermore, the calibration function that can be
integrated with a sensor will become more advanced (faster, more
accurate, or a higher-order Iinearization). As usual in circuit technology,
and also for the calibration circuits, accuracy, speed, and circuit
complexity can be exchanged.
Finally, it should be recognized that calibration techniques can only

improve some aspects of a sensor, and at a relatively high price
(calibration measurements). If a sensor transfer function can be improved
in another way, for example by using an improved construction or material
or an auto-calibration, this is given preference. Research on such sensor
developments deserves to be stimulated.

Index
A c
actuators ........................ 2 calibration ..................... 5, 7
AD conversion ................ 5, 33 analog implementation . .106
auto-calibration . 13
AD converter ................... 33 batch calibration . . 7
sigma-delta converter . 127 definition . . . . . . 8
analog calibration errors . . . . . . .120
cross-sensiti vity · 99 calibration algorithm ............ 179
full-scale or gain · 98
calibration coefficient ........ 78, 186
IC realization . 121
implementation \06 calibration memory .............. 46
independent . . \03 calibration procedure .... 118, 167, 189
offset . . . . . . · 98
pulse-modulation 104 calibration program ............. 189
temperature coefficients 101 cascode dividers ................ III
analog signal processing .......... 93 charge balance .................. 38
B comparator ..................... 34
implementation . . . . . . . . .. 155
bandgap reference ............... 27
controller ....................... 3
bandgap voltage ................. 23
cross-sensitivity ................. 57
batch calibration ................ 93
curve-fitting .................... 70
bitstream ................... 36, 104
bitstream generator ............. 140 o
bitstream spectrum ............. 151 DA conversion .................. 96
~s~~~~re ................. ~ DA converter
implementation . . . . . . . . . . III
bus interface ..................... 5
multiplying DAC . . . . . . . 97, III
IS2 interface . . . . . . . . . . .. 43
DAC coefficient ................ 103
bus master ..................... 40
data-acquisition .................. 3
bus protocol .................... 40
IS2 protocol . . . . . . . . . ... 42 decimation ................ 136, 138
bus slave ....................... 39 decimation filter ................ 136
implementation . . . . . . . . . . 159

Index
digital calibration
hardware implementation . 174 in-situ calibration ................ 93
one-dimensional . . . . . 174
progressive polynomial . 175 integrator ....................... 34
software implementation 178 implementation . . 153
two-dimensional. . . . . 179 transfer function . . . . . . . . . 131
discrete-time domain ............. 38 interpolation
Lagrange . . . . 71,80
dithering ............. 140,145,150 piece-wise linear .. 63
downsampling . . .......... 128, 136 splines . . . . . . . . 69
downsampling ratio (DSR) ....... 128 interpolation curve ............... 65
drift . . . . . .. . ................... 8 IV converter . . . . . . . . . . . . . .. . .. 114
E L
Emitter-Coupled Logic .......... 116 Lagrange interpolation ........ 71, 80
error compensation .............. 54 laser trimming .. . .............. 94
error minimization ............... 73 limit cycles ... . . . . .. ... . ..... 142
error surface lineanzation .................... 55
two dimensional . 184 curve fitting . . . 70
look-up table . . 60
errors
piece-wise linear 63
cross-sensitivity .52
splines . . 67
distribution. . . .54
drift . . . . . . . .52 linearizatlon methods . . ...... 56,75
error distributlOn .25 look-up table. . . . . . . . . . . . . . . . .. .60
full-scale . .52
gain . . . . . . . . .52 M
hysteresis .52
linear versus non linear .55 Manchester coding . . . . . . .. . .... .41
measurement errors .6,25 microcontroller . . . . . . . . . . . .. . .... 5
nonlinearity . . . . . .52 interface . . . . . . . . . . " . 39
offset. . . . . . . . .52 programs ...... . . 173
sensor characteristic . .53
micromachimng . . . . . . . . . . . .. . .. 23
systematic versus random .53
mismatch effect ................. 31
F multiplier
feedback testing .. . ............. 15 Gilbert multiplier . . ... 101
implementation .. . 108, 165
feedforward testing .............. 15
full adder ...................... 141

Index
N quantizer ...................... 131

noise shaping .......... 128, 130, 136
R
first and second order . . . . . . 133
spectrum . . . . . . . . . . . . . 135 ratiometric conversion ............ 32
normalized input signal ........... 56 resistor arrays ................... 96
normalized output signal .......... 56
s
normalized transfer function ...... 56
self-calibration .................. 13
Nyquist rate sampling ........... 130
self-test ........................ II
o sensor .......................... 2
oversampling .............. 128- 129 pressure sensor . . . . 23
temperature sensor . . . . . . . . 22
oversampling ratio (OSR) ..... 36, 130
sensor transfer function
p normalized transfer function . . . 56
one-dimensional . . . . . . . . . 56
piece-wise polynomial interpolation 67 two-dimensional . . . . . . .58, 183
piecewise-linear interpolation ..... 63 sigma-delta ..................... 33
piezo-resistors .................. 24 bitstream generator . . 141
principle. . . . . . . . . . . . . . 34
pressure sensor .................. 23
signal multiplication . . . . . . . 149
readout circuit . . . . . . . . . .. 32
sigma-delta AD converter ........ 128
progressive polynomial calibration .77
differential . . . . . . .. 128
algorithmic loop. . . 176
discrete-time model . . . 131
one-dimensional. . . . . . 77
implementation . . . 153,157
signal f10w diagram . . . . 84 input-output transfer .139
software implementation 178
output spectrum. . . . .. 135
successive approximation 177
signal processing chain . . . 129
two-dimensional . . 84 signal-to-noise ratio 130, 132- 134
pseudo-random noise ........... 150 sigma-delta calibration
PTAT voltage ................... 23 cross-sensiti vity .148
pulse stream ................... 104 gain . . . . . . . . .143
modulated spectra .147
pulse-density modulation ........ 104 offset . . . . . . . .147
pulse-width modulation ......... 104 sigma-delta DA-converter ....... 140
Q
signal conditioning ............... 5
signal f10w diagram ............. 106
quantization ................... 129
quantization level ............... 61 signal transfer
correction 45
quantization noise .............. 131 errors . . . 52

Index
signal transfer function ........... 56

inverse function . . 70
logarithmic. . . . . 59
one-dimensional . . 56
two-dimensional . .57
sinc filter ...................... 136
sinc2 filter ..................... 137
smart sensor ..................... 4
calibrated tempo sensor . 161, 165
configuration. . . . o . o 43
integrated smart sensor o 05
interface o . o o 21
mterface chip . o . . . 178
~2~soo.o.o.o. o~
microcontroller configuration 178
signal chain . . . . o o . . o. . 44
smart temperature sensor . .. . 29
splines o.. oo... ooo.. o. oo.... ooo067
standard unit .... o.. o.. o.. oo. o. oo.6
T
temperature sensor ... oo.. o. oo.... 22
implementation .. o o ... 160, 165
translinear Circuit ......... o.. o.. 108
trimming . o. o.. ooo.... oooo. o.. o094
two's complement. oo. o.. o. o112,142
u
user interface .. o ..... o..... oo... 3
v
VI con verter o.... o. . . .. o.. o.. o113
w
Wheatstone bndge ..... o.. o... o. o25
window function .. o. oooo.... oo. 136
triangular window o 137
uniform window o .. o .. o . . 136

Integrated Smart Sensors

Uploaded by

Copyright:

Available Formats

Integrated Smart Sensors

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Integrated Smart Sensors

Uploaded by

Copyright:

Available Formats

INTEGRATED SMART SENSORS

INTEGRATED SMART SENSORS

Gert van der Horn

SPRINGER-SCIENCE+BUSINESS MEDIA, B.V.

ISBN 978-1-4419-5016-1 ISBN 978-1-4757-2890-3 (eBook)

Printed on acid-free paper

AII Rights Reserved

2 lntegrated Smart Sensor Concept 21

INTEGRATED SMART SENSORS: Design and Calibration v

2.3 Analog Interface circuits ......................................... 27

3 Calibration and Iinearimtion

VI INTEGRATED SMART SENSORS: Design and Calibration

3.3 Progressive polynomial calibration method ........... 76

INTEGRATED SMART SENSORS: Design and Calibration Vll

5 Calibration using Sigma-Delta

viii INTEGRATED SMART SENSORS: Design and Calibration

6 Calibration using Digital

7 Summary and Conclusions 193

INTEGRATED SMART SENSORS: Design and Calibration IX

INTEGRATED SMART SENSORS: Design and Calibration

1.2 Sensors and actuators

Fig. 1-1 Gathering the process information, and controlling the

To gather process information we measure process parameters, such as

2 INTEGRATED SMART SENSORS : Design and Calibration

Based on a large diversity of underlying physical principles, many sensors

From a system-Ievel point of view, a configuration as shown in Fig. 1-2 is

Fig. 1-2 Bus structure for sensor/actuator communication.

Bus architectures and protocols that have been successfully introduced on

INTEGRATED SMART SENSORS: Design and Calibration 3

To enable the proposed configurations, many interfacing functions will

high-Ievel bus ~~~~~~~~~~~

Fig. 1-3 System division into two low-level sub-systems.

A focus on the realization of sensors in standard Integrated Circuit (lC)

4 INTEGRATED SMART SENSORS: Design and Calibration

1.3 Integrated smart sensors

• sensor(s), preferably silicon sensor(s)

Fig. 1-4 Signal processing chain of an integrated smart sensor.

INTEGRATED SMART SENSORS: Design and Calibration 5

1.4 Measurement errors and correction

• interpretation; we can easily learn how to interpret numeral

1 SI=Systeme International d'unites, Of International System of units.

6 INTEGRATED SMART SENSORS: Design and Calibration

is to regulate an error signal to zero. In that case it is sufficient to know

The process of relating the measurement or sensor signal to the physical

1.4.1 Process variations and tirne variations

This individual attention per de vice is very expensive in mass production,

INTEGRATED SMART SENSORS: Design and Calibration 7

reference measurements must be minimized, in order to save time and thus

Another problem which can be encountered in the transfer of a sensor is

It is difficult to objectively define the terms scale, ensuring, and

8 INTEGRATED SMART SENSORS: Design and Calibration

"The third meaning for calibration, as used by metrologists, is defined by

In practice, a sensor calibration is performed by putting the sensor in a

At a set of input signals [xJ"'xn ], which can be calculated by using the

INTEGRATED SMART SENSORS: Design and Calibration 9

, ~ - " environment '-...,.....---r-T"""'"

Fig. 1-5 Schematic explanation of sensor calibration.

outputs are processed in order to retrieve a signal which relates to the

After calibration the correction should be maintained so that the calibrated