NTC temperature sensor

[jmarcelino]

@tobru
There is some known non-linearity in the ESP32 ADC which is not (yet) being compensated on the ESP-IDF (and thus neither in micropython)

https://twitter.com/ba0sh1/status/815759185816129536/

For example up to 0.1v it registers 0.0. So at the least you need to account for that.

You may be able to compensate these in software. Personally I'd just use a much more accurate dedicated external ADC - connected over I2C for example - for NTC measurements. For example a ADS1115 breakout can be had for less than $2.

[robert-hh]

I'm still thinking how you have connected the thermistor a) between Vcc and ADC (like sketched below) or b) between ADC and GND. From your code, it looḱs like b)
And yes, there were quite a few notes about the ADC's poor performance. If you need more precise values, you should compensate at least for the actual range.

[tobru]

@robert-hh said in NTC temperature sensor:

I'm still thinking how you have connected the thermistor a) between Vcc and ADC (like sketched below) or b) between ADC and GND. From your code, it looḱs like b)

It's currently connected like this:

3V3
 +
 |
 |
+-+
|-|
|-| 10k resistor (instead of NTC)
+-+
 |
 |
 +--------------+ P16
 |
 |
+-+
|-|
|-| 10k resistor
+-+
 |
 |
 +
GND

Do you think the code doesn't fit this scheme?

@jmarcelino said in NTC temperature sensor:

use a much more accurate dedicated external ADC - connected over I2C for example

I've ordered a Velleman MM112 external ADC with I2C interface, I can then compare the values.

[robert-hh]

I think it should. If you NTC is connected between Vcc and ADC, this line:

resistance = SERIESRESISTOR / ((4095 / average) - 1)

should be:

resistance = SERIESRESISTOR * ((4095 / average) - 1)

For instance:
at 50°C, the value of the NTC is 3588 Ohm. That should give an ADC reading of:

4095 * (10000/(3588 + 10000)) = 3014

Put that into your formula, and you get 27881 Ohm
Put that into the other formula, and you get 3587 Ohm.
Regards, Robert

[tobru]

@robert-hh said in NTC temperature sensor:

resistance = SERIESRESISTOR * ((4095 / average) - 1)

Yes, that looks better, I use the following test code now:

import machine
import time

THERMISTORPIN = 'P16'
NUMSAMPLES = 5
SERIESRESISTOR = 10000

adc = machine.ADC(0)
adcread = adc.channel(attn=3, pin=THERMISTORPIN)
samples = [0.0]*NUMSAMPLES
for i in range(NUMSAMPLES):
    samples[i] = adcread()
    print("sample %u: %u" %(i, samples[i]))
    time.sleep(1)
  
average = 0
for i in range(NUMSAMPLES):
    average += samples[i]

average /= NUMSAMPLES

print("average sample: %u" %average)

resistance = SERIESRESISTOR * ((4095 / average) - 1)

print("resistance: %u" %resistance)

Vin = 3.2
buffer = average * Vin
Vout = buffer / 4095
buffer = (Vin / Vout) - 1
R2 = 10000 * buffer

print("resistance2: %u" %R2)

and the output is:

sample 0: 1850
sample 1: 1836
sample 2: 1837
sample 3: 1845
sample 4: 1829
average sample: 1839
resistance: 12262
resistance2: 12262

Nevertheless, the ADC accuracy seems to be very bad so measuring the temperature this way seems useless. I'll try it with an external ADC as soon as it arrived.

[robert-hh]

I think you could stil use the built-in ADC, if you requirement for precision is not extreme (like with 1% error). But you have to compensate the errors of the ADC.
a) the zero point. look at which input voltage the output is different to zero. According to a comment below, this is somewhat like 0.1 V
b) the scaling: Look at which input voltage the output just reaches the maximum value.
c) nonlinearity. You could make a compensation table, or ignore it, if it is small enough and outside the range of interest.
d) Noise
a) and b) is easily compensated for. d) is reduced with taking an average of many samples, like you did. You do not have to wait a second between the samples. A few ms is sufficient, if at all, and a small capacitor like 10 nF with short leads directly at the ADC helps.

Another observation: You connect the NTC to 3.2 V (are you sure it's not 3.3V?). The ADC max range is at 3.6 V, the reading at 3.2 V would be 3631. Instead of 4095, you should therefore use 3631 as the virtual maximum value. Then, the resistor value is 9793 (10412 for 3.3 V), which is 2% off the nominal value, and depending of the precision class of the resistor that could be the real value.

[tobru]

@robert-hh said in NTC temperature sensor:

if you requirement for precision is not extreme (like with 1% error)

There is no exterme precision requirement. 1% would be completely fine.

a) the zero point. look at which input voltage the output is different to zero. According to a comment below, this is somewhat like 0.1 V
b) the scaling: Look at which input voltage the output just reaches the maximum value.
c) nonlinearity. You could make a compensation table, or ignore it, if it is small enough and outside the range of interest.
d) Noise
a) and b) is easily compensated for. d) is reduced with taking an average of many samples, like you did. You do not have to wait a second in between, and a small capacitor like 10 nF with short leads directly at the ADC helps.

Thanks for the inputs!

Another observation: You connect the NTC to 3.2 V (are you sure it's not 3.3V?).

It is connected to 3V3, so it is of course 3.3V. I've chosen 3.2V for Vin in the second equation because of the parameter attn=3. I've read that in this threads: https://forum.pycom.io/topic/226/lopy-adcs and here https://forum.pycom.io/topic/602/ntc-temperature-sensor/6

The ADC max range is at 3.6 V, the reading at 3.2 V would be 3631. Instead of 4095, you should therefore use 3631 as the virtual maximum value. Then, the resistor value is 9793 (10412 for 3.3 V), which is 2% off the nominal value, and depending of the precision class of the resistor that could be the real value.

Interesting. I must have missed this information in the documentation, I thought 3.3V is the maximum.

[robert-hh]

As always, the information given by espressif is poor. Nothing compared to the manuals issued for instance by Atmel, STM or Freescale. It is unclear, how much the attenuation level really is, some say 11db, some say 12 db. 11db would be a factor 3.55, 12 db a factor 3,98 (10**(atten/20)) of the base range, which is ?. In any case, better give it a try. If you have a voltage meter, get the value at the input and scale it by:

Vrange = Vinput * (4095/reading)

[robert-hh]

So I gave it a try, connected a power source to the input and read the average of 100 samples each. The results are not convincing.
a) Even at an average of 100 samples, there is a lot of noise..
b) it looks pretty nonlinear., at least at attn 3
results:

attn = 3
value 0 until 125 mV
107	mV -> 0
200	mV -> 95
300	mV -> 215
500	mV -> 455
700	mV -> 700
1000 mV -> 1070
1500 mV -> 1700
1600 mV -> 1815
2000 mV -> 2310
2500 mV -> 2930
2900 mV -> 3540
3000 mV -> 3740
3100 mV -> 3955
3170 mV -> 4095


attn = 0
Value 0 until 57 mV
100 mV -> 127
200 mV -> 550
300 mV -> 974
500 mV -> 1820
700 mV -> 2650
900 mV -> 3510
1000 mV -> 3933
1044 mV -> 4095

If you look at that in a graph, at attn 3 the bends up beyond 2.5 V, while the attn=0 graph looks more linear, besides the 0-Offset.
That's worse than I expected. The best one can get is to work with attn = 0 and use an external voltage divider.
Edit: I found a link that shows an experience, which pretty much matches mine:
http://www.esp32.com/viewtopic.php?t=1045
The Figures I got are pretty much the same. So 6db attn would be an option too.

[jmarcelino]

@robert-hh
this is an open issue on ESP-IDF and also the Arduino core

https://github.com/espressif/esp-idf/issues/164
https://github.com/espressif/arduino-esp32/issues/92

The official response from Espressif is:

The ADC frontend indeed has a non-linear response. Once characterization data is ready, we will take it into account inside ADC APIs, and update the datasheet with correct values of INL/DNL and effective resolution.

So any efforts to compensate it at this stage will likely be rendered useless in a future update once the characterisation is complete (at least at ESP-IDF/MicroPython level)

That's also why I suggested an external ADC for the time being, at least that won't change.

[robert-hh]

@jmarcelino: Yes, Ive read that. But you would still have to compensate for the non-zero range (at least in your circuit). So at the moment and with limited requirements for precision, using the lower attenuation variants 0 and 6db and simply adding an offset would be sufficient as compensation.
If you need precise values, yes, then an external ADC is anyhow recommended.
B.T.W. Is there still the option to use the ADC in 10 bit mode, like by an option res = xx in the channel set-up?