Off Topic - Algorithm for noise

Doug99 · Joined: 13 May 2004 Posts: 6

Hello ladies and gentlemen... Very Happy

I have a problem figuring out how to go about this.

I am measuring a PWM signal from an radio controlled receiver (variable from 1ms - 2ms) and I would like to create an algorithm to detect an erratic signal from a stable one.

I am measuring the pulses and storing each one as an 8bit value. I can hold about twenty values in a buffer and analyze them in between pulses. I thought I could use statistical calculations like mean, mode, variance, etc. but I am not a math wizard. Is there a better way to sniff out an erratic signal from a stable (but variable) one?

If anyone can help, I'd really appreciate it. If it works, I'll post the code in the code form.

Thank you,

Doug

valemike · Guest

In the past, i've sampled 5 times, put the values in ascending order, then take the average of the 3 middle ones.

PCM programmer · Joined: 06 Sep 2003 Posts: 21708

From the net:
"The median filter is able to remove impulsive noise without smoothing
edges, as a linear filter would do."

Note that MS Excel has a median filter. You can load your data into
Excel and model it. Then write a C function that will do the same thing.
Compare the output of your function to the results from Excel to make
sure you're doing it the same way.

Doug99 · Joined: 13 May 2004 Posts: 6

Thanks for your replies!

I looked up median filters and found tons of info (even found an example in Microchip assembler). I am wondering how I can modifiy or create one to give a boolean output. As in 'yes' there is noise or 'no' there is not. It may not be so easy to give a true or false answer as it would be to simply eliminate the noise. Any ideas?
I apologize if this isn't making any sense.

Thanks again,

Doug

SteveS · Joined: 27 Oct 2003 Posts: 126

Filters, like the median filter, are good at removing noise - which really doesn't sound like what you want. Unless you filter the signal and use the filtered value to test against new values and call anything outside some bounds as 'noise'.

Can you supply more info?

- What is your 'good' signal like? Is it constant, varying, can it jump to a new level and still be considered valid (median filter can handle this well).

- What is the noise like? Does it have a different frequency component than the signal?

- At any instant do you have just good signal or just noise? Or is it noise on top of a good signal? The first would be like signal detection (it's there or not), the second is trying to extract signal from signal + noise.

Try to describe what needs to happen. If you can't it will be hard to build hardware and/or software to do it. If you can decribe it then you may be able to do something.

- SteveS

Doug99 · Joined: 13 May 2004 Posts: 6

Hi Steve, I'll try to answer your questions the best I can.

SherpaDoug · Joined: 07 Sep 2003 Posts: 1640 Location: Cape Cod Mass USA

If all you want is a Good/Bad how about finding the min and max and see how far they are apart? The math is dirt simple and if you just keep the min, max, and sum you don't even need the five bytes it would take to store the values. At the end divide the sum by the number of samples to get an average. Or better still if you take the sum, subtract the min and max, then divide by num-2 you get an average without outlyers. I do the latter all the time.
_________________
The search for better is endless. Instead simply find very good and get the job done.

SteveS · Joined: 27 Oct 2003 Posts: 126

Along the same lines, you can filter the input signal using one of a variety of methods. Decide on a window that separates good signal from noise ( the tricky part). If the new sample is outside the window then it's noise. What you do then I'm not sure. Depends on what you are after.

Do you really want to do something with the noise data or are you just trying to get rid of it?

If you want to get rid of it, then 'just' choose the right filter. The median filter is very good when you get short, large spikes of noise. A low pass filter works well if the noise magnitude is small compared to the signal. If the noise completely masks the signal then you need a lock-in amplifier.

- SteveS

dfnr2 · Joined: 26 Dec 2003 Posts: 2

An autocorrelation function is numerically easy to compute, and its shape will give an idea how much signal vs. random noise is present. I believe excel has the ability to compute and plot autocorrelation functions. You may want to check your library or local tech bookstore for books on engineering communications or on spectral analysis. You could also try google. If you want to differentiate between different signal types in noise, you may want to look into Wiener or Kalman filters.