You are on page 1of 12

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/314097187

Interpolation Errors in Thermistor Calibration Equations

Article  in  International Journal of Thermophysics · February 2017


DOI: 10.1007/s10765-017-2194-x

CITATIONS READS

10 1,631

1 author:

David Rodney White


Independent Reseracher
146 PUBLICATIONS   3,517 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

CCT Guides on Thermometry View project

Cryogenic Fixed Points View project

All content following this page was uploaded by David Rodney White on 19 November 2017.

The user has requested enhancement of the downloaded file.


Int J Thermophys (2017) 38:59
DOI 10.1007/s10765-017-2194-x

TEMPMEKO 2016

Interpolation Errors in Thermistor Calibration


Equations

D. R. White1

Received: 24 July 2016 / Accepted: 9 February 2017 / Published online: 27 February 2017
© Springer Science+Business Media New York 2017

Abstract Thermistors are widely used temperature sensors capable of measurement


uncertainties approaching those of standard platinum resistance thermometers. How-
ever, the extreme nonlinearity of thermistors means that complicated calibration
equations are required to minimize the effects of interpolation errors and achieve
low uncertainties. This study investigates the magnitude of interpolation errors as a
function of temperature range and the number of terms in the calibration equation.
Approximation theory is used to derive an expression for the interpolation error and
indicates that the temperature range and the number of terms in the calibration equation
are the key influence variables. Numerical experiments based on published resistance–
temperature data confirm these conclusions and additionally give guidelines on the
maximum and minimum interpolation error likely to occur for a given temperature
range and number of terms in the calibration equation.

Keywords Calibration · Calibration equation · Interpolation error ·


Steinhart–Hart equation · Thermistors

1 Introduction

Negative temperature coefficient (NTC) thermistors are widely used temperature sen-
sors capable of measurement uncertainties as low as 1 mK, approaching that of standard

Selected Papers of the 13th International Symposium on Temperature, Humidity, Moisture and Thermal
Measurements in Industry and Science.

B D. R. White
rod.white@callaghaninnovation.govt.nz

1 Measurement Standards Laboratory, Lower Hutt, New Zealand

123
59 Page 2 of 11 Int J Thermophys (2017) 38:59

platinum resistance thermometers (SPRT). For high-accuracy applications in narrow


temperature ranges near ambient temperatures, their high stability, high resolution, and
low cost makes them very attractive alternatives to SPRTs [1]. However, the extreme
nonlinearity of thermistors means a complicated calibration equation is required to
achieve a close match between the true resistance–temperature behavior of the ther-
mistor and the calibration equation and, hence, to minimize the effects of interpolation
errors and achieve the low uncertainties.
For nonlinear sensors like thermistors, interpolation error is a significant source
of measurement uncertainty. Unfortunately, there is a paucity of good advice on the
selection of calibration equations for thermistors. Many texts e.g., [2–4] and most
manufacturers’ literature recommend the three-term logarithmic polynomial devel-
oped by Steinhart and Hart [5]. Unfortunately, the original recommendations for the
Steinhart–Hart equation were based on numerical errors [6], and there are several
alternative equations that are nearly always better.
To date, the most informative studies on thermistor calibration equations are those
of Bennett [6], Hoge [7] and Chen [8]. Bennett explored the properties of a family
of logarithmic polynomials, which were among those investigated earlier by Stein-
hart and Hart, when applied to a set of twenty thermistors of the same type over a
temperature range of 30 ◦ C. Bennett plotted the peak interpolation error versus tem-
perature range and the number of terms in the calibration equation. A similar plot
is presented in the BIPM Guide on thermistors [1], again based on data for a single
thermistor type. Hoge, using previously published data spanning 200 K to 450 K,
investigated the application of least-squares fitting to thermistor calibration equations
and concluded that several of the equations fitted the data better than the Steinhart–Hart
equation. Most recently, Chen investigated the performance of seven different cali-
bration equations when applied to four different thermistors over a single temperature
range of 70 ◦ C. Chen recommended the four-term version of the logarithmic polyno-
mial, in part because of the low residual errors in the calibration, but also because of
its ease of use. Chen also noted the unsatisfactory performance of the Steinhart–Hart
equation.
The general form of the interpolation error for NTC thermistors is known from
approximation theory [9,10], which predicts a strong dependence on the temperature
range and the number of terms in the equation. However, apart from the two instances
reported by Bennett [6] and White et al. [1], there is very little specific information
about the characteristics of the interpolation errors.
The aim of this paper is to provide firm advice on the selection of thermistor cal-
ibration equations, based in part on a theoretical model of interpolation error and in
part on a numerical evaluation of published resistance–temperature data. Section 2
first reviews some aspects of the mathematics underpinning the family of logarithmic
polynomials used for many of the calibration equations and develops an equation giv-
ing the mathematical form of the interpolation error. Section 3 then presents the results
of numerical estimates of interpolation error based on the resistance–temperature data
reported by a manufacturer of high-precision thermistors. Finally, some conclusions
and recommendations are given.

123
Int J Thermophys (2017) 38:59 Page 3 of 11 59

6000

5000

Beta value, K
4000

3000

2000
-100 -50 0 50 100 150 200 250
Temperature, °C
Fig. 1 Beta values of a range of commercial thermistors plotted against temperature

2 Theory

2.1 Calibration Equations

Thermistors are intrinsic ceramic semiconductors for which the electrical resistance
R varies strongly with temperature T . Negative temperature coefficient (NTC) ther-
mistors are those most commonly used for temperature measurement and have a
resistance–temperature relationship given approximately by [11]
 
Eg
R(T ) = r0 exp , (1)
kT

where r0 is a constant, E g is the bandgap of the semiconductor and k is the Boltzmann


constant. For thermometry, Eq. 1 is usually rewritten in the form
  
1 1
R(T ) = R(T0 ) exp β − , (2)
T T0

where T0 is some convenient reference temperature, often 298.15 K (25 ◦ C). The
parameter β is a characteristic of the thermistor material, with typical values in the
range 2000 K to 6000 K. The ‘beta value’ of a thermistor can be determined from a
thermistor’s resistance–temperature characteristic as

d ln R T 2 dR
β= =− . (3)
d(1/T ) R dT

Figures 1 and 2 show the beta values as a function of temperature and reciprocal
temperature for thirteen different thermistors with R(25 ◦ C) values ranging from
100  to 1 M. The figures show only nine distinct β(T ) curves, suggesting that nine
different thermistor materials are used in the manufacture of this family of thermistors.

123
59 Page 4 of 11 Int J Thermophys (2017) 38:59

6000

5000

Beta value, K
4000

3000

2000
0.0015 0.002 0.0025 0.003 0.0035 0.004 0.0045 0.005 0.0055
Reciprocal temperature, K-1
Fig. 2 Beta values of the thermistors of Fig. 1 plotted against 1/T

Figures 1 and 2 also show that the beta values are not simple constants. Whether
described in terms of T or 1/T , at least a quadratic or cubic equation is required
to describe the beta values well. For this reason, thermistor calibration equations are
usually more complicated than Eq. 2. The general form of the most commonly used
equation expands Eq. 2 as a logarithmic polynomial [1,5–8]:

1
= A0 + A1 ln(R/R0 ) + A2 [ln(R/R0 )]2 + · · · + A N −1 [ln(R/R0 )] N −1 , (4)
T

with the number of terms, N , depending on the temperature range and uncertainty
required. In Eq. 4 the R0 constant may be used in several ways. It may be the units
of measurement (e.g., ohms, kilohms or megohms), or the nominal R(25 ◦ C) value
for the thermistor, or (T0 , R0 ) might be one of the calibration points. In the last case,
fitting one of the parameters in the polynomial expansion may be unnecessary. For
example, Eq. 4 passes through the point (T0 , R0 ) when A = 1/T0 .
Equation (4) is most often simplified to the three-term Steinhart–Hart [5] equation

1
= A0 + A1 ln(R/R0 ) + A3 [ln(R/R0 )]3 , (5)
T

which is, unfortunately, recommended by most thermistor manufacturers and in many


texts. The omission of the [ln(R/R0 )]2 term in Eq. 5 causes the equation to give
inconsistent calibrations, and its use should be strongly discouraged. The problem can
be seen by replacing R0 by k R0 , where k is a numerical constant:
       3
1 R R
= A0 + A1 ln − ln(k) + A3 ln − ln(k) , (6)
T R0 R0

so that terms in [ln(R/R0 )]2 emerge from the cubic term and depend on the value
of k. Because Eq. 5 must exclude these terms, it must yield different calibration
equations (and hence different interpolation errors) when the units of measurement

123
Int J Thermophys (2017) 38:59 Page 5 of 11 59

0.1
0
-0.1
Interpolation error, K
-0.2
-0.3 N=3
-0.4 N=2
-0.5
-0.6
-0.7
-0.8
-0.9
0 10 20 30 40 50 60 70 80 90 100
Temperature,°C
Fig. 3 Examples of interpolation error for a calibration of a thermistor. Red line: calibration at two points
(0 ◦ C, 100 ◦ C). Blue line: calibration at three points (0 ◦ C, 50 ◦ C, 100 ◦ C) (Color figure online)

are changed (e.g., from ohms to kilohms), when applied to thermistors of the same beta
characteristic but different R(25 ◦ C) values, or when nominally identical resistors are
connected in series or parallel. To avoid this problem, polynomial calibration equations
should form a complete sequence with no intermediate terms omitted [12]. The use of
a complete sequence is assumed throughout the following analysis. As Bennet noted
[6], Hart and Steinhart’s recommendation for the equation was based on numerical
errors, and as several authors have observed [1,6–8], the Steinhart–Hart equation is
usually not the best three-term equation.

2.2 Interpolation Error

In the context of a thermistor calibration, interpolation error is the difference between


the true resistance–temperature characteristic of a thermistor, and an algebraic approx-
imation determined from exact values of the characteristic at a small number of points.
Figure 3 shows the interpolation error for a two-point and a three-point calibration
of a thermistor. In the two-point calibration, the measurements are used to determine
A0 and A1 in Eq. 4 with the remaining Ai parameters set to zero. In the three-point
calibration, the measurements are used to determine A0 , A1 and A2 . The interpolation
error has zeros at each of the calibration points and a characteristic oscillatory shape
accompanied by changes in sign around each calibration point.
For thermistors, each additional term added to the calibration equation typically
reduces the interpolation error by a factor of 10 or more, and the number of terms
can be increased as required to ensure that the interpolation error makes a negligible
contribution to the total uncertainty.
Further information on the interpolation error can be gained by drawing on approx-
imation theory. Consider a thermistor calibration in which the data from N distinct
calibration points, (Ti , Ri ), with T1 < T2 < · · · < TN , are used to determine the N

123
59 Page 6 of 11 Int J Thermophys (2017) 38:59

parameters in the calibration equation. The values for the measurements are substituted
into the chosen calibration equation to give N equations that can be solved simulta-
neously for the coefficients A0 , . . . , A N . For the following analysis, it is assumed
that the calibration equation R̂(T ) is exact at the calibration points (i.e., there is no
measurement uncertainty). For advice on the treatment of measurement uncertainty
in thermistor calibrations see [1,10].
By taking the logarithm of both sides of Eq. 2 and treating the equation as a function
of 1/T , we can use the expression for interpolation error given by Cheney [9] (see
also [10]) to give the algebraic form of the error:
    
1 1 1 1 1 1
ln R̂(1/T ) − ln R(1/T ) = − − (. . .) −
T T1 T T2 T TN

1 d N ln R(1/T ) 
× , (7)
N ! d(1/T ) N  T =τ

where τ is some unknown value of T with T1 ≤ τ < TN . Equation (7) is based on an


existence theorem that says that a value of τ exists for every value of T . The expression
can be rearranged using Eq. 3 to give

R̂(T ) − R(T ) 1 d N −1 β(1/T ) 
= (T − T ) (T − T ) (. . .) (T − T ) ,
d(1/T ) N −1 T =τ
1 2 N
R(T ) N !T N TmN
(8)

where the (geometric) mean temperature Tm is defined by Ti = TmN . By substituting
the temperature sensitivity of the thermistor,

1 dR β
= − 2, (9)
R dT T

into the left-hand side, the interpolation error can be expressed in terms of the equiv-
alent temperature error:

1
T = T̂ − T = − (T1 − T ) (T2 − T ) (. . .) (TN − T )
N !T N −2 TmN

1 d N −1 β(1/T ) 

× . (10)
β d(1/T ) N −1 T =τ

With the calibration points equally spaced, the interpolation error is similar to a
Tchebycheff
 polynomial of the same order. Tchebycheff polynomials have the form
(T − Ti ) with the points Ti chosen so that all of the peaks in the function have the
same amplitude. For a given temperature range Tmin to Tmax , Tchebycheff polynomials
have a maximum amplitude

  (Tmax − Tmin ) N
 
 (T − Ti ) = , (11)
max 22N −1

123
Int J Thermophys (2017) 38:59 Page 7 of 11 59

so that the peak interpolation error for the thermistor calibration can be estimated as

(Tmax − Tmin ) N 1 d N −1 β(1/T ) 
|T |max ≈ . (12)
N !22N −1 T N −2 TmN β d(1/T ) N −1 T =τ

This equation shows that the interpolation error scales in proportion to the temperature
range, Tmax − Tmin , raised to the N th power (where N is the number of terms in the
calibration equation), and that it falls with increasing mean temperature. Although an
important parameter in Eq. 12, the mean temperature, Tm , is usually determined by
the application and is not a free parameter that can be adjusted to minimize the error.
Note too, if the (N − 1)th derivative of the beta value is zero, then the interpolation
error is zero, and the calibration equation is exact.
Although the derivation above applies to a calibration in which the number of
measured points equals the number of parameters in the calibration equation, the
conclusion also applies to calibration equations that are over determined, i.e., where
the number of calibration points exceeds the number of parameters in the equation
and the equation is fitted by least squares. With least-squares fits, the N values of
Ti in Eq. 12 correspond to the N zeros that always occur in the residual error of an
N th-order fit [9].
To use Eq. 12 directly as a guide to selection of calibration equations for thermistors
requires advance knowledge of the β(1/T ) function, which is not provided explicitly
by any manufacturer and, therefore, not known in advance of a calibration. This situa-
tion is made more difficult by the lack of standardisation of thermistors characteristics
and the wide range of different thermistor materials available, as illustrated in Figs. 1
and 2.

3 Numerical Evaluation

To provide general guidance on interpolation errors in thermistors, numerical exper-


iments were carried out using the high-quality resistance–temperature data supplied
in some thermistor specifications. The availability of the data is a consequence of
manufacturers’ publishing resistance–temperature data for thermistors that are inter-
changeable at the level of 0.1 ◦ C or better.
While it should be possible in principle to apply the analysis below to any published
resistance–temperature data, it was found that the utility of all published data was
limited by the resolution of the resistance values at high temperatures, which are often
reported only to the nearest 1  or 0.1 . The resulting round-off errors limited the
resolution in the analysis of interpolation error, in some cases to no better than 10 mK.
To overcome this limitation, (1) only the data covering a temperature range of 200 ◦ C
or more were used, (2) the temperature range of the analysis was limited to 40 ◦ C less
than the upper temperature limit of the thermistor and (3) the data were pre-treated as
follows. Firstly, the full range of the published data for each thermistor, which in some
cases ranged from −80 ◦ C to 250 ◦ C, was fitted to a sixth- or seventh-order polynomial
of the form of Eq. 4. This equation was then used to generate a set of high-resolution
resistance–temperature data that were free of round-off errors. It is assumed that the

123
59 Page 8 of 11 Int J Thermophys (2017) 38:59

round-off errors in the manufacturers’ data were random and that the least-squares fit
averaged their influence.
Unfortunately, only the data from one manufacturer were found to be usable in
this way. This is the data on which Figs. 1 and 2 are based. Efforts were made to
apply the analysis to thermistors from other manufacturers, but in most cases, the
temperature range over which the data were reported was insufficient to reliably fit to
high-order and hence to determine the fourth-order interpolation errors. In one case,
the manufacturer had already smoothed the data using separate low-order equations
in three abutting temperature ranges, so that the first derivative of the data was clearly
discontinuous and did not accurately represent the thermistor behavior at the level
required.
Once the high-resolution data had been generated, least-squares fits of the data were
performed for each thermistor data set with 2- , 3- , and 4-parameter fits of Eq. 4 over
a variety of temperature ranges. The temperature ranges extended from 0 ◦ C upwards
in temperature to a maximum of 200 ◦ C. The decision to make Tmin = 0 ◦ C for all
the fits was based in part on the temperature ranges for the available data, and in part
on the likely practical need to include the ice point or water triple point within the
calibration range of a thermometer.
All of the equations were fitted by performing a linear fit of data in the form of
Eq. 4, i.e., with the data in the form (ln(Ri ), 1/Ti ). Nonlinear fits with Eq. 4 rewritten
in the form

1
T = , (13)
A + B ln(R/R0 ) + C [ln(R/R0 )]2 + D [ln(R/R0 )]3 + · · ·

were tried, but the fourth-order fits all failed to converge well. The rms error in the fit,
which was taken to be a measure of the interpolation error, was then scaled according
to the uncertainty relation

u(T ) = Tm2 u(1/T ), (14)

to give the rms interpolation error in the temperature measurement.


Figure 4 plots the calculated rms interpolation error for the thirteen different ther-
mistors, over the different temperature ranges and for the three different least-squares
fits. The figure shows several interesting features. Firstly, the interpolation error fol-
lows quite closely the expected (Tmax − Tmin ) N behavior suggested by Eq. 12. For
comparison, the lines in the figure enclosing the bands of results for the respective fits
follow the (Tmax − Tmin ) N behavior exactly. The small difference in slope between
the results and these lines is possibly accounted for by the increasing value of mean
temperature, Tm , as the temperature range increases (see Eq. 12). Another possible
influence is that the value of the corresponding derivative of β(1/T )(see Eq. 12)
changes over the temperature range.
The breadth of the bands enclosing the results is quite wide, there being a factor
of about 6 between the highest and lowest rms error for the N = 2, a factor of 4.5
for N = 3 errors and a factor of 10 for the N = 4 errors. Table 1 gives equations
approximating the lower and upper bounds of the bands of results. For all thermistors,

123
Int J Thermophys (2017) 38:59 Page 9 of 11 59

10

1
rms Interpolation error, K

0.1

0.01

0.001

0.0001
10 100 300
Temperature range , K
Fig. 4 rms interpolation error for a range of different thermistors versus temperature range and order of
the interpolating equation. Blue squares: N = 2; red dots: N = 3; green triangles: N = 4. The solid lines
indicate approximate boundaries (see text) (Color figure online)

Table 1 Approximate upper and lower bounds of rms interpolation error for different order fits and tem-
perature ranges, T = Tmax − Tmin

Number of parameters Lower bound Upper bound

N =2 u = 1.2 × 10−5 T 2 u = 7.0 × 10−5 T 2


N =3 u = 0.8 × 10−8 T 3 u = 3.5 × 10−8 T 3
N =4 u = 1.1 × 10−12 T 4 u = 1.1 × 10−11 T 4

the addition of an extra term in the equation decreased the interpolation error by at
least a factor of 10.
Although not shown in Fig. 4, there is a trend for the thermistors with the lowest
R(25 ◦ C) values to have the lowest second-order (N = 2) interpolation error. There
did not seem to be the same correlation with higher-order errors or with beta values.
This observation contrasts with the error for linearized thermistor resistance networks
[13] where the beta value has a strong influence on the error.
With all the sets of N = 2 and N = 3 interpolation errors, the points in Fig. 4 for
each thermistor followed the T N trend well. However, some of the N = 4 fits showed
evidence of additional effects, for which there are at least three possible causes. Firstly,
the additional effects were also characteristic of fits based on low-resolution data or
data drawn from a narrow temperature range, so the variations are possibly caused by
the round-off errors and perhaps give an indication of the reliability of the analysis.
The greater breadth of the envelope of the N = 4 errors may also be a consequence
of these effects. Secondly, Eq. 10 shows that β(1/T )(N −1) must be constant for the
points for a single thermistors to fall on a straight line. It may be with some thermistors

123
59 Page 10 of 11 Int J Thermophys (2017) 38:59

that β(1/T ) is more complicated. Another possibility is that the interpolation errors
for some of the high-order fits might reflect the differences in T − T90 , where T90
is temperature according to the International Temperature Scale (ITS-90) and T is
the thermodynamic temperature. The thermistors should be expected to follow laws
determined by the thermodynamic temperature, but the resistance–temperature data
are reported in terms of T90 .

Conclusions

Figure 4 and Table 1 give guidelines on the maximum and minimum rms interpolation
error versus temperature range and the number of terms in the calibration equation
for a range of thermistors including at least nine different thermistor materials with
different beta values. This information allows users to select the appropriate interpo-
lating equation for a given temperature range and uncertainty. The analysis suggests
that a four-term equation is sufficient to ensure interpolation errors are below 1 mK
for an 80 ◦ C range, a three-term equation sufficient for 10 mK over a 60 ◦ C range and
the two-term equation is sufficient to ensure the error is below 100 mK over a 35 ◦ C
range.
This information is especially important where the number of calibration points
used to calibrate the thermistor is the same as the number of unknown parameters
in the equation, as the user may have no other indication of the magnitude of the
errors. In contrast, where the number of calibration points exceeds the number of
fitted parameters and the equation is fitted using least squares, the standard deviation
of the residual errors in the fit will include the effects of the interpolation errors.
As suggested by the theoretical expression for interpolation error, Eq. 12, the inter-
polation errors scale approximately as (Tmax − Tmin ) N , where N is the number of
parameters in the calibration equation. The interpolation errors also fall rapidly with
the increasing number of terms in the calibration equation. For all the thermistors in
the study, and for all temperature ranges below 200 ◦ C, the addition of a single extra
term in the equation reduced the interpolation error by more than a factor of 10.
Although the data on which the study is based include at least nine distinctly
different types of thermistors, the data were from a single manufacturer, so it is difficult
to know how representative the results are of all high-quality thermistors. Further
research on thermistors from other manufacturers is required.

References
1. D.R. White, K. Hill, D. del Campo, C. Garcia Izquierdo, Guide on Secondary Thermome-
try: Thermistor Thermometry (BIPM, Paris, 2014). http://www.bipm.org/utils/common/pdf/ITS-90/
Guide-SecTh-Thermistor-thermometry.pdf
2. P.R.N. Childs, Practical Temperature Measurement (Butterworth-Heinemann, Boston, 2001)
3. J. Fraden, Handbook of Modern Sensors: Physics, Designs, and Applications (Springer, New York,
2010)
4. T.W. Kerlin, R.L. Shepard, Industrial Temperature Measurement (Research Triangle Park NC, ISA,
1982)
5. J.S. Steinhart, S.R. Hart, Calibration curves for thermistors. Deep Sea Res. 15, 497–503 (1968)

123
Int J Thermophys (2017) 38:59 Page 11 of 11 59

6. A.S. Bennett, The calibration of thermistors over the range 0 ◦ C–30 ◦ C. Deep Sea Res. 19, 157–163
(1971)
7. H.J. Hoge, Useful procedure in least squares, and tests of some equations for thermistors. Rev. Sci.
Inst. 59, 975–979 (1988)
8. C.-C. Chen, Evaluation of resistance–temperature calibration equations for NTC thermistors. Mea-
surement 42, 1103–1111 (2009)
9. E.W. Cheney, Introduction to Approximation Theory (American Mathematical Society, Providence,
1982)
10. D.R. White, P. Saunders, Propagation of uncertainty with calibration equations. Meas. Sci. Technol.
18, 2157–2169 (2007)
11. S.M. Sze, Physics of Semiconductor Devices, 2nd edn. (Wiley, New York, 1981)
12. D.R. White, Some mathematical properties of the ITS-90 scale, in Temperature: Its Measurement and
Control in Science and Industry, vol. 8, ed. by Christopher W. Meyer (AIP, Melville, 2013), pp. 81–88
13. D.R. White, Errors in linearizing resistance networks for thermistors. Int. J. Thermophys. 36, 3404–
3420 (2015)

123
View publication stats

You might also like