Thermistor used as temperature sensor

[John B]

I want to use an NTC thermistor as a temperature sensor in conjuction with
an 8-bit ADC. The target temperature is 70C and I have chosen a 100k (@25C)
thermistor and a 14k resistor to give me the best resolution around 70C.

I have constructed an Excel spreadsheet to try to determine the transfer
characteristic using the ADC reading as ordinate and the equivalent
temperature as abscissa. This spreadsheet is attached to my post under the
same title in ABSE.

My problem is that I have used the Excel trendline function on sheets 2 & 3
of the spreadsheet and it works correctly on sheet 2 but not on sheet 3.
Sheet 2 shows the scatter diagram with the temperature as ordinate and the
calculated ADC reading as abcissa; the Excel trendline returns an equation
which when plotted is good enough to use. However the trendline returned on
the transposed scatter diagram on sheet 3 is rubbish, as demonstrated by the
plot of the equation returned.

What am I doing wrong?

 

[John Popelish]

I want to use an NTC thermistor as a temperature sensor in conjuction with
an 8-bit ADC. The target temperature is 70C and I have chosen a 100k (@25C)
thermistor and a 14k resistor to give me the best resolution around 70C.

I have constructed an Excel spreadsheet to try to determine the transfer
characteristic using the ADC reading as ordinate and the equivalent
temperature as abscissa. This spreadsheet is attached to my post under the
same title in ABSE.

My problem is that I have used the Excel trendline function on sheets 2 & 3
of the spreadsheet and it works correctly on sheet 2 but not on sheet 3.
Sheet 2 shows the scatter diagram with the temperature as ordinate and the
calculated ADC reading as abcissa; the Excel trendline returns an equation
which when plotted is good enough to use. However the trendline returned on
the transposed scatter diagram on sheet 3 is rubbish, as demonstrated by the
plot of the equation returned.

What am I doing wrong?

I don't know how you got the transposed equation, so I can't verify that.

But I like a different function to fit temperature to divider voltage
that has a natural shape that is about right.

If you treat the A/D full scale value as 1, then you can convert the
fraction of 1 (called f) to temperature (with a maximum error of about
..7 degrees) by this formula:

Temperature= a * (tan(f - b) + c + d*f
Where a=-240.5856, b=0.5515, c=123.0738, d=-102.1026

But if you change the series resistor to 24k, The worst case error is
about half as big (about +-0.3 degrees) with these constants:
a=-196.0598, b=0.5532, c=106.3374, d=-96.91

 

[Spehro Pefhany]

I don't know how you got the transposed equation, so I can't verify that.

But I like a different function to fit temperature to divider voltage
that has a natural shape that is about right.

If you treat the A/D full scale value as 1, then you can convert the
fraction of 1 (called f) to temperature (with a maximum error of about
.7 degrees) by this formula:

Temperature= a * (tan(f - b) + c + d*f
Where a=-240.5856, b=0.5515, c=123.0738, d=-102.1026

But if you change the series resistor to 24k, The worst case error is
about half as big (about +-0.3 degrees) with these constants:
a=-196.0598, b=0.5532, c=106.3374, d=-96.91

tan()? It's only a bleedin' 8-bit ADC, how about a little LUT and be
done with it?


Best regards,
Spehro Pefhany

 

[John B]

..

I don't know how you got the transposed equation, so I can't verify that.

..

That's the equation produced by Excel. However I have just taken another
look and found I had used the wrong columns in the check plot on sheet 3!!
It should have been column D for the ordinate and column E for the abcissa.
It now looks better but is still incorrect. I'll need to do some more work
on it tomorrow.

But I like a different function to fit temperature to divider voltage that
has a natural shape that is about right.

If you treat the A/D full scale value as 1, then you can convert the
fraction of 1 (called f) to temperature (with a maximum error of about .7
degrees) by this formula:

Temperature= a * (tan(f - b) + c + d*f
Where a=-240.5856, b=0.5515, c=123.0738, d=-102.1026

..
I'll have a look at that too, but I'm only using an Atmel 8-bit ATmega32 so
the floating point library will be a bit OTT.

But if you change the series resistor to 24k, The worst case error is
about half as big (about +-0.3 degrees) with these constants:
a=-196.0598, b=0.5532, c=106.3374, d=-96.91

Thanks anyway for your help.

 

[John B]

..
..

tan()? It's only a bleedin' 8-bit ADC, how about a little LUT and be
done with it?

..
..

I have lost count of the number of times I have told other people to go this
way. I got so wrapped up in the challenge that I didn't even consider it.

Thanks.

 

[Spehro Pefhany]

.
.
.
.

I have lost count of the number of times I have told other people to go this
way. I got so wrapped up in the challenge that I didn't even consider it.

Thanks.

;-)


Best regards,
Spehro Pefhany

 

[John Popelish]

On Sun, 06 Nov 2005 17:52:14 -0500, the renowned John Popelish



tan()? It's only a bleedin' 8-bit ADC, how about a little LUT and be
done with it?

That would work just fine. Only 255 values needed.

 

[John Popelish]

It's not even too bad with 10 bits in many cases these days.




Which value do you propose omitting?


Best regards,
Spehro Pefhany

Either or both extreme ones should probably be replaced with
thermistor open and thermistor shorted cases. ;-)

 

[Spehro Pefhany]

That would work just fine.

It's not even too bad with 10 bits in many cases these days.

Only 255 values needed.

Which value do you propose omitting?


Best regards,
Spehro Pefhany

 

[John Popelish]

.
I'll have a look at that too, but I'm only using an Atmel 8-bit ATmega32 so
the floating point library will be a bit OTT.

I'm afraid these constants won't work at all. They are the solution
to a different, but related function.

If you really want to plug these in, just for fun, the first set
should have been:
a=-797.4246, b=0.551, c=-316.7359, d=696.1691

and the case with the 24k series resistor:
a=-653.0378, b=0.5539, c=-255.8475, d=556.9077

And for an even better fit, change the resistor to 30k and plug these
constants into temperature=a* tan(d*(f-b)) + c
Where:
a=-43.504, b=0.5526, c=47.3743, d=2.1974

 

[John Popelish]

Either or both extreme ones should probably be replaced with thermistor
open and thermistor shorted cases. ;-)

Actually, the curve is not fit to temperature for A/D values below 62
or above 240, so the table could be 179 values and cover the entire
fitted range from 10 to 105 C. This allows for half degree resolution
with 8 bit values, or 1/3rd degree if you waste just a bit of the range.

 

[Spehro Pefhany]

Either or both extreme ones should probably be replaced with
thermistor open and thermistor shorted cases. ;-)

Indeed. Probably a few near each end, so a comparison would be more
efficient.


Best regards,
Spehro Pefhany

 

[Pooh Bear]

I want to use an NTC thermistor as a temperature sensor

Poor choice IMHO.

Graham

 

[Mike Harrison]

Poor choice IMHO.

Without knowledge of the application requirements, this is a ridiculous statement.

NTCs are cheap and easy to interface. They are not especially accurate but that may not be a
problem.

 

[Spehro Pefhany]

Without knowledge of the application requirements, this is a ridiculous statement.

NTCs are cheap and easy to interface. They are not especially accurate but that may not be a
problem.

Many, many millions are made with *significantly* better than +/-0.5°C
accuracy near RT. That's extremely good unadjusted accuracy--
unmatched by any other inexpensive sensor. Also, they have high output
and do not require a voltage reference, only a cheap reference
resistance.

There are a couple market segments in which they dominate because of
these unique characteristics-- and many more in which they are dubious
or totally unsuitable, of course.


Best regards,
Spehro Pefhany

 

[Fred Bloggs]

I don't know how you got the transposed equation, so I can't verify that.

But I like a different function to fit temperature to divider voltage
that has a natural shape that is about right.

If you treat the A/D full scale value as 1, then you can convert the
fraction of 1 (called f) to temperature (with a maximum error of about
.7 degrees) by this formula:

Temperature= a * (tan(f - b) + c + d*f
Where a=-240.5856, b=0.5515, c=123.0738, d=-102.1026

But if you change the series resistor to 24k, The worst case error is
about half as big (about +-0.3 degrees) with these constants:
a=-196.0598, b=0.5532, c=106.3374, d=-96.91

Isn't that precision relatively absurd considering the manufacturing
tolerance on all those things is something like 25%?

 

[[email protected]]

Isn't that precision relatively absurd considering the manufacturing
tolerance on all those things is something like 25%?

Almost all of that tolerance is in the absolute resistance of the
thermistor - a good single point calibration ( in a well stirred ice
bath where the ice has been made from distilled/de-ionised water) can
get you to accuracies of some millidegrees.

You can buy trimmed "interchangeable" thermistors which offer
unadjustaed accuracies of +/-0.2C (about +/-0.8% tolerance on
resistance). Farnell offered Betatherm parts, and Newark has parts from
YSI - including at least one offering +/-0.05C accuracy. Thermometrics
and Fenwall are also in the business.

The Steenhart-Hart fitting function is usually cllaimed to be good
enough to get millidegree accuracy over a respectable temperature
range.

 

[Winfield Hill]

[email protected] wrote...

The Steenhart-Hart fitting function is usually cllaimed to be good
enough to get millidegree accuracy over a respectable temperature
range.

Steinhart-Hart

 

[Rene Tschaggelar]

Fred Bloggs wrote:

Isn't that precision relatively absurd considering the manufacturing
tolerance on all those things is something like 25%?

Fred,
you're somewhat outdated, there are standard
products which offer 0.5%

Rene

 

[John Popelish]

Isn't that precision relatively absurd considering the manufacturing
tolerance on all those things is something like 25%?

Quite possibly. But if the O.P. needs a simple function to produce a
lookup table of temperatures fore A/D states, having a function that
adds little additional error to that of the parts, is useful.

My point was, that just because Excel produces polynomial fits for any
curve, that form may not be the best natural fit to the curve.