Thermistor NTC tolerance

[bonzer]

Good morning, I would like to understand how exactly NTCs work when it comes to the tolerance. If I have for example an NTC 10k 1% B3977. I know that typically the 1% tolerance is related to the 10k resistance @25°C, and that at higher temperatures like 55°C/130°F it could reach even 4%. Is it general or it depends on the specific technology used for the thermistor and thus it changes based on the types?
But what I really care the most is understanding the following: once I have an NTC sample, let's say SENSOR_1, does that specific sensor keep a certain characteristic that is slightly different from the theoretical but still linear in a limited range small range, like for instance between 55°C and 50°C, thus it keeps the same tolerance error, if for SENSOR_1 it is +3% of theoretical value at 50°C, than we could say that @55°C it is still +3% with respect to 55°C theoretical value? Or from one temperature to the other, it could change completely like +2% or +1% or it could even depend on the previous temperature values because of some hysteresis when cooling or warming up etc? And the tolerance is related to the resistance but when converted to temperature it messes up things much more because over that range a small resistance change results  in a big temperature change. The reason why I'm asking is because I've seen people using them to detect temperature differences of fractions of 1°C (lowest 0.1°C) by performing measurements over the described temperature ranges. To me it looks unreliable but if considering that these are temperature differences measured by the same NTC at the same point in space and time (like at 10 seconds of distance), could it make sense?

[golden_labels]

NTCs designed for temperature measurement will have tolerances given separately both for resistance at 25 °C and for whichever B parameter is used. Some manufacturers, like Vishay, provide worst case value for each temperature in detailed tables.

[bonzer]

Thank you for your answer. I have to deal with a Tasseron sensor. While looking for these tables I have come across a datasheet that shows tables where temperature error is included as well (see attached file). Does it mean that the overall temperature error in the worst case is the one shown (dT) or it is the typical value?

[MrAl]

Good morning, I would like to understand how exactly NTCs work when it comes to the tolerance. If I have for example an NTC 10k 1% B3977. I know that typically the 1% tolerance is related to the 10k resistance @25°C, and that at higher temperatures like 55°C/130°F it could reach even 4%. Is it general or it depends on the specific technology used for the thermistor and thus it changes based on the types?
But what I really care the most is understanding the following: once I have an NTC sample, let's say SENSOR_1, does that specific sensor keep a certain characteristic that is slightly different from the theoretical but still linear in a limited range small range, like for instance between 55°C and 50°C, thus it keeps the same tolerance error, if for SENSOR_1 it is +3% of theoretical value at 50°C, than we could say that @55°C it is still +3% with respect to 55°C theoretical value? Or from one temperature to the other, it could change completely like +2% or +1% or it could even depend on the previous temperature values because of some hysteresis when cooling or warming up etc? And the tolerance is related to the resistance but when converted to temperature it messes up things much more because over that range a small resistance change results  in a big temperature change. The reason why I'm asking is because I've seen people using them to detect temperature differences of fractions of 1°C (lowest 0.1°C) by performing measurements over the described temperature ranges. To me it looks unreliable but if considering that these are temperature differences measured by the same NTC at the same point in space and time (like at 10 seconds of distance), could it make sense?

Hi,

For a long time thermistors were used to temperature compensate various types of meters, so they can't be too bad, and I think they have gotten better since way back then.

One thing about anything thermal that is involved in a measurement is repeatability.  That's when you get a measurement at a fixed temperature and then tomorrow and the next day and next day after that you get the same measurement at that same fixed temperature.  I think thermistors should be good at that because they've been used a lot over time.

What else this means in our circuit is if we measure 2 volts when the temperature is say 20 degrees C and 2.4 volts when the temperature is say 30 degrees, yet that 30 degree measurement was supposed to be 2.45 volts, as long as it measures 2.45 volts whenever the temperature is 30 degrees then we can calibrate the setup to measure over some reasonable temperature range.

The usual calibration points are 0 C and 100 C (freezing water and boiling water), but you may need better calibration than that so you would get a calibrated temperature meter to use to calibrate your system.  Once you do that, you should be able to make some reliable measurements.

However, if you are looking for accuracy you probably would want to go with the digital temperature sensors that work using the I2C protocol or even another protocol.  That way you can read it with your microcontroller and get fairly good readings.  You can also calibrate that if you want to get better results, but it depends a lot on why you need to measure temperature.  Usually, 1 degree is well below the tolerance needed for some measurement, except in the medical field and possibly some chemistry stuff and precise lab work.

[bonzer]

So theoretically if they get calibrated at the production line after mounting them over the electronic boards, by using a bench equipped with an accurate thermometer and storing the signal for the two temperature values in the range of interest or for instance taking 0 to 100°C as you said, that could be a good path to go. But from my understanding it would still be unlikely to get any lower than ±0.2°C because of the inherent limits of the technology behind the NTCs.
Therefore a more sophisticated digital sensor could be a better option for this kind of needs.

[Phil1977]

Have you thought about a 4-wire PT100 or PT1000?

The best digital temperature sensors I know have 0.25°C accuracy (ADT7420). A well calibrated PT100 / PT1000 can get you to down to milikelvin resolution and the 4-wire readout significantly enhances the stability of the readout circuit (widely eliminating the influence of cables/plugs etc.)

[bonzer]

PT1000 seems an interesting solution, especially in the 4-wires configuration. I heard about them in the past but because of the low resistance increase per °C, it looked like it's not something that people rush into. Though I didn't expect the accuracy to be so high, that's good. From my understanding, in the 4 wire configuration, I have to generate a current and make it flow through a couple of wires, meanwhile measuring the voltage drop on the PT1000 using the other couple of wires by connecting it to a high input impedance differential amplifier like (for instance instrumentation amplifier?). This way it gets rid of the cabling resistance during the reading.

[Siwastaja]

A well calibrated PT100 / PT1000 can get you to down to milikelvin resolution and the 4-wire readout significantly enhances the stability of the readout circuit (widely eliminating the influence of cables/plugs etc.)

Yeah but that level of accuracy requires some stunning engineering. For example, often you hear stuff like above: "especially in the 4-wires configuration"; sounds like people assume that PT100/1000 is good and 4-wire configuration just makes it even better. In reality if you even try with two wires and some circuit similar to what you would use with NTC, you easily have error of some freaking five degrees C. NTC is much easier because resistance is higher and the curve is steeper, i.e., small change in temperature results in significant change in resistance.

Both PT1000 and PT100 "suffer" from the fact that small temperature changes result in small resistance changes and those are hard to sense, plus the absolute resistance especially on PT100 is very small.

I always laugh when people who talk about millikelvin accuracy then go on with a 3-wire arrangement and say "it's good enough". No, you really need the 4-wire configuration; you really need self-heating prevention (periodic readout), you really need super careful analog front end design.

To get even +/-0.2 degC reliably out of PT100 or PT1000 is a real project. There are helpful ICs and appnotes, but it needs to be taken seriously. Then again, if you are running within some limited temperature range, a fake DS18B20 out of Ebay or cheapest NTC driven from a random 1% 4.7k pull-up into MCU ADC pin, calibrated once at known temperature, might give you the same +/-0.2degC!

TLDR: PT1000 and especially PT100 require care and experience to exceed the accuracy of NTC!

[VinzC]

IMHO the first and foremost question is "do you need that precision at all?" or at least "what precision do you [really] need?"

If, for instance, it's for measuring temperatures that relate to human perception on the skin (like hot water or ambient air) or heating devices, then sub-percent precision makes absolutely no sense at all. I'm not even sure it even does below, say 5 or 10%.

If it's for regulating the temperature for an aquarium with tropical fishes or plants then maybe a 1% precision is not luxury.

If it's for a bakery oven, then most probably 5-10% precision is enough.

These are only examples that come to my mind but you get the picture.

[Phil1977]

@Siwastaja: Is it still that difficult to buy dedicated ADCs for Pt100/Pt1000-sensors?

I just remember using a Keithley 2000 for that purpose a while ago. I think it uses 0.1mA test current in the 10k-range, dissipating around 20uW at 2kOhm sensor resistance (somewhere >200°C). This dissipation creates a deviation of 6mK with a pessimistic thermal resistance of 300K/W of the sensor to the DUT.

The Keithley 2000 has a resolution of 10mOhm in this range, giving a temperature resolution of around 3mK. The one-year stability is 110ppm or 1.1Ohm, which is around 0.3K.

We also had some specialised meter from Omega that was specified better but broke during usage and wasn't available for quite some time.

I don't say it´s easy to measure with this accuracy. But if you want a stability <0.1K then I think the 4wire-Pt1000 is still one of the best options.

[Kleinstein]

With PT1000 it is no longer that difficult to find a suitable ADC. There are quite some SD ADCs with differential input and ref. connection and some internal "gain". So the PT1000 and ref. resistor are in series and the ref resistor provides the reference to the ADC. It usually still needs a good ref. resistor to get OK accuracy.

There are a few ADCs with PT1000 in mind (e.g. include a current source ), not really a dedicated version needed. It can get a bit more tricky when protection is needed and leakage currents are possible. The leakage currents are one reason why for high temperatures PT100 is preferred over PT1000 and sometimes even lower resistance (e.g. PT25) is used.

Most of the PT1000 sensors are the thin film type with not that great stability and accuracy - not really better than the NTC made for measurement. There are good, wire wound PT100 / PT1000, but these can get quite expensive.

[MrAl]

So theoretically if they get calibrated at the production line after mounting them over the electronic boards, by using a bench equipped with an accurate thermometer and storing the signal for the two temperature values in the range of interest or for instance taking 0 to 100°C as you said, that could be a good path to go. But from my understanding it would still be unlikely to get any lower than ±0.2°C because of the inherent limits of the technology behind the NTCs.
Therefore a more sophisticated digital sensor could be a better option for this kind of needs.

I would think you would want to look into the digital sensors.  Thermistors are only so good but they do seem to be repeatable just like a lot of other thermal products.  You could do some simple tests to see what you can get out of them, and maybe the application would need swapping out different units for different ranges or something like that.
I would think the range over a short interval is somewhat linear, but at least monotonic, and you could test the repeatability over very small ranges, but you need a really good reference meter.

If you are looking for perfect temperature readings then you have to look a bit farther.  It's hard to get that from anything really.  Even voltmeters are not that exact unless you put out the money.  Probably the digital sensor is the simplest way to go with this but you have to decide that.

[Kleinstein]

The choice of a suitable sensor depends also on the temperature range and what points are really important. With all sensor types there are different grades with accuracy and stability. One often has to make compromises at one end or the other. One would need more information and there may be multiple working options. Mounting the sensor is often also an important point.

The repeatablity/hysteresis can actually be one of the weak points of digital sensors - especially if they come in a plastic case. Already soldering them to a board can reduce the accuracy from the soldering temperarture and PCB stress.
The digital/electronic sensors are usually still good with the initial accuracy and relatively easy to use on a PCB.

The NTCs get good resolution/low noise and can get away with 2 wire connection much better than PT1000. They can also be quite small and the read out is relatively easy, though nonlinear. For moderate temperature they can be quite stable and accurate. With the higher resistace leakage current (at high temperature or with condensation at low temperature) can be an issue.

PT100 could get very good stability with expensive wire wound sensors. The thin film ones are still OK, but not as accurate. One may need 4 wire conection even for PT1000 and the readout still needs more care. Good sensors can get quite expensive and also tend to be a bit larger.