Data collection - noisy signal - how much noise is acceptable?

[refrigerator]

I'm collecting voltage data, and I need a resolution of at least .05mV, which I think I can provide over the range of voltages with my inexpensive A/D board. The problem is, the signal fluctuates maybe 1mV up and down, and so I get a noisy signal. I've tried running a low-pass filter through it and then taking a running average, and I got something somewhat stable. Still, it fluctuates about .03mV. My readings are supposed to be at steady state, so I don't care as long as I get ONE number.

My question is, am I going to be ok just processing my signal until it's "good enough" (and if this is the case, how exactly would I determine what is good enough?), or do I need to start looking for better equipment? My initial thought is that I need to get a signal with fluctuations sufficiently small so that it gives an unambiguous reading, am I right? (So if I need a resolution of .05mV, my signal should fluctuate less than .025mV either way?)

 

[williamshipman]

Hi refrigerator.

0.05mv is amazingly small - do you really need such accuracy?

If you do then you will need to look at your measurement setup to make sure that you don't have any thing that forms an antenna as this will give you noise from the electricity grid, your cell phone, etc.

After that, you will probably need to look at digital signals processing methods. There are a few techniques that might help, I think. Unfortunately, that was never my speciality. Also, what range of voltages are you measuring? If you are measuring 0V to 5V then not even a high end PIC or dsPIC is likely to give you a resolution of 0.05mv.

If possible, try to place an amplifier between your test circuit and A/D. By placing the amplifier very close to your circuit you should end up with most of your high frequency noise from outside sources not going through the amplifier but instead being picked up by the wires connecting your amplifier to the A/D.

If you're amplifier has a large enough gain, then you should be able to achieve your required resolution. For example, a gain of 100 will require your A/D to only have a resolution of 5mv. Rounding off your readings to the nearest 5mv would then negate the 1mv noise. Of course, if the source of the noise is inside your circuit then this won't help at all and you are back to using signals processing methods.

I hope this helps. Cheers.

 

[0xDEADBEEF]

This question is tricky, and I have to deal with this stuff myself. (But my signals are smaller :) The big question is, where the noise comes from. If it is very nice noise, averaging the way you do it will kill it. I assume that you do not want to spend much money, then this is how I would proceed.

1) Make sure that the signal which you want to measure is constant.
2) Take a lot of measurements 1000 or more, make a histogram, check if you can fit a gauss bell. Calculate the skewness and make sure that it is very small.
3) Normally you would stop here, calculate sigma and add that as your error bars.
Now the question to ask is how much you trust your equipment, and how sure are you that the signal really doesn't fluctuate.
4) If you got a nice Gaussian and you are sure that it is due to noise, then you can average over longer time, repeat the process with points averaged for longer and check your histogram again, as long as it is still Gaussian you should be fine.

The main point is, that you have to take your noise seriously, don't just assume that it is good statistical noise, but check it!

What else can you do especially when your noise is not Gaussian?
1) Better cables and good grounding.
2) Shielding of the experiment
3) Twisting the wires
4) Maybe find the source of the noise
5) Amplify the signal close to the source, instrumentation amplifiers are not that expensive
6) differential signal transmission

In theory you can get perfect signal recovery if you are really dealing with noise. In practice there are lots of nasty effects that can give you a dc offset. A way around it is moving your signal to another frequency. It doesn't work on all experiments, and is expensive, but if you can use a lock in amplifier, then you can recover a signal that is 1000 times weaker then the noise.