How can I improve ADC accuracy?

[gpsmikey]

If you are interested, a couple of books I have found quite helpful and keep on my shelf are "The Circuit Designers Companion" (Tim Williams) and "The Art of Electronics" (Horowitz and Hill). Not cheap, but both excellent books and very handy as everyday reference.
http://www.amazon.com/Circuit-Designers-Companion-Second-Engineers/dp/0750663707
http://www.amazon.com/Art-Electronics-Paul-Horowitz/dp/0521370957
The "Art of Electronics" is a bit dated, but still an excellent reference. See if your local technical library has copies of them you could snoop through.

mikey
_________________
mikey
-- you can't have too many gadgets or too much disk space !
old engineering saying: 1+1 = 3 for sufficiently large values of 1 or small values of 3

[hawking1122]

Thank you guys for invaluable help, I'll do the hardware job and feedback anything I explore.

[hawking1122]

I've added some bypass caps and the result is now much more stable. Thank you all!
However, something I want to see clearer:
1. My Vref is generated using LM317L, the schematic is below:

The 3 caps are chosen based on datasheet. As you can see, the resistor and pot are quite small, so the current flowing through them is not small, so I do not have to worry about instability of LM317L even though current flowing into Vref pin of microcontroller is small?

2.

[hawking1122]

3. The amount of time for 1 sample is 22usec, but the maximum number of samples I can get from my Pic18f2550 is just 16 samples/1second (my computer software need to display analog value every second, I'm averaging 16 samples to get more accurate value). I've tried averaging 64 samples but it take about 4 second to receive all 64 10bit values.
The phenomenom above does not appear (I can average up to 300 samples each time) if I used value=read_adc() instead of
read_adc(ADC_START_ONLY);
sleep();
value=read_adc(ADC_READ_ONLY);
Does it mean sleep() takes too much time to complete?

[hawking1122]

I've double checked my code and see there are some delay functions which slow down the rate of transmiting/receiving usb data, but after I've deleted them all, the maximum rate of sampling is 40 samples/second. It seems not as short as 22usec mentioned in datasheet?

[Ttelmah]

On the 'why use 8,16,32 for averaging'. This is down to division. Dividing an integer by eight for example, is just a 3bit rotation. Takes a handful of machine cycles. Dividing by 10 instead, takes perhaps 50* as much work.....
A good averaging algorithm, is:

[SherpaDoug]

With simple averaging, the more samples the better, but the answer only improves by the square root of the number of samples, and that is assuming the noise is gaussian.

If the data is generally good, with occasional bad readings, I often use what I call "Olympic Scoring." I take a number of samples, often 10, and keep track of the maximum, the minimum, and the sum. Then I subtract out the max and the min from the sum and average the rest. 10 samples minus the two extremes leaves you 8. Dividing by 8 is a simple bit shift. You don't even have to store all 10 samples, just 3 numbers, min, max, and sum. And a sum of 10 samples at 12 bits each still fits in an INT16. So the math and RAM consumption is minimal and you eliminate one outlier either high, low, or both. It works a lot like a median filter but with a little smoothing and a lot less CPU resources.

Search the forum for "olympic" to get more info on the method.
_________________
The search for better is endless. Instead simply find very good and get the job done.