Skip to content

Latest commit

 

History

History
110 lines (80 loc) · 5.28 KB

Internal ADC.md

File metadata and controls

110 lines (80 loc) · 5.28 KB

Poor ADC ?

The ESP32 internal ADC has a poor reputation. It suffers from non-linearity and is known to be noisy. But is it really that bad?

The linearity can be fixed with calibration and we keep a margin to the ADC voltage range, like Espressif suggests (e.g. at 6dB PGA attenuation 150 mV ~ 1750 mV, see suggested range) For simple curve fit see here. This can be improved with a look-up-table (LUT).

About the noise issue, we should take a closer look how we take measurements. A simple and naive way is to set up a periodic timer and take a single-shot measurement with Arduino API analogRead(), expecting a 12-bit precision. But what about conversion time? The shorter the AD conversion time, the higher the noise. We didn't tell the ADC for how long to measure. If we measure the time analogRead takes until return, its only a couple 10 microseconds. Digging through the code, we find that it calls adc_oneshot_read, adc_oneshot_hal_convert. The ADC characteristics in the ESP32-S3 data sheet show 100ksps (for the ESP32 even 2Msps) sampling rate with the DIG controller. So we can assume that measurements are taken with sub-µs time, which explains the noise issue quite well.

The more time the ADC spends for conversion, the better the results get. Let's assume a measurement that takes 100µs and we take it every 100ms. This will give us an effective ADC duty cycle of 0.1% (1/1000). So 99.9% of the time the ADC doesn't do anything.

To get most of the ADC performance we start a continuous measurements at maximum sampling rate (SOC_ADC_SAMPLE_FREQ_THRES_HIGH) and average the samples so we get the desired sampling rate. Averaging should be done with the raw adc int16 readings for best performance.

For the ESP32-S3, a good configuration appears to be 83.3kHz SR and 32 samples averaging. With 3 channesl this gives an per-channel rate of ~ 868 sps (=83333 / 32 / 3).

We cannot use the single-shot measurement API when the ADC is in continuous mode. So when adding another channel that has low requirements (e.g. an NTC), we need to add it the continous reading pattern, which will reduce bandwidth of all channels to 651sps (83333/4/32). We can improve this by duplicating the pattern without the NTC channel.

For example ch0,ch1,ch2,ch3,ch0,ch1,ch2, where ch3 is the NTC channel. This increases sampling rate of ch0,ch1 and ch2 to 744sps (83333/4/32*8/7). ch3 is sampled with 372sps ( 83333/7/32).

Wiring

This applies to the original fugu PCB design with either ESP32 or ESP32-S3

To use the internal ADC1 of the ESP32/ESP32-S3, just omit ADS1015 or ADS1115 (U10) and the ALERT 10k pull-up resistor (R31).

The firmware will automatically switch to internal ADC1 if it doesn't find the ADS1015 or ADS1115 on the I2C bus.

Remove the ADS chip and the pull-up resistor next to it. Then use thin cables or isolated copper wire to connect the analog signals directly with the ESP32 chip.

ESP32-S3-WROOM-1

A2 (solar current) ---> GPIO4 (ADC1_CHANNEL_3)
A3 (solar voltage) ---> GPIO5 (ADC1_CHANNEL_4)
A4 (bat voltage)   ---> GPIO6 (ADC1_CHANNEL_5) (orignally ADC-ALERT)

ESP32-WROOM-32

A2 (solar current) ---> GPIO36 / SENSOR_VP (ADC1_CHANNEL_0)
A3 (solar voltage) ---> GPIO34 (ADC1_CHANNEL_6) (orignally ADC-ALERT)
A4 (bat voltage)   ---> GPIO39 / SENSOR_VN (ADC1_CHANNEL_3)

Sampling Rate

ESP32-S3 max ADC sampling rate according to the datasheet is 100 kHz, however only 83.3 kHz is possible with esp-idf (SOC_ADC_SAMPLE_FREQ_THRES_HIGH in soc_caps.h). ESP32 can do up to 2 MHz.

We configure the ADC1 in continuous mode with a sampling rate of 83333 Hz (adc_continuous_config). This means that the ADC does 83.3k conversions per second. Reading more than one channel reduces the sampling rate per channel. We apply averaging of 32 samples to reduce the quite noisy readings. With 4 channels this gives 868 samples per second per channel (83333 / 3 / 32).

On the ESP32-S3 we can achieve this theoretic rate, the ESP32 appears to be a bit slower (511 / 625 @ 80 kHz, 639 / 781 @ 100khz ) and always at ~80% of the calculated rate. This ratio is constant across tested sampling rates 80k, 100k, 125k, 128k, 156.25k, 160k, 200k, 250k, 312.5k, 320k and 400kHz.

At 640 kHz the ESP32 works at 70% of expected sampling rate and with significant variance. This might be due to a slow control loop.

Latency can be configured through conv_frame_size. with frame size 128 bytes, 4 bytes/sample, SR=83.3kHz, latency is expected to be 384µs (128 / 4 / 83333).

looks like esp32 is the only one with 2Msps

Other Notes

  • Waveshare ESP32-S3 Mini (ESP32-S3-Zero) does have no floating pin noise
  • ESP32-S3-WROOM DevBoard has peak noise level of 100~120