# Can Autocorrelation be used to Differentiate Signal Quality (High, Medium, Low, Very Low) for Periodic Signals?

I'm working on a project to classify the quality of periodic signals into four categories: high, medium, low, and very low (noisy). I was initially exploring autocorrelation as a potential feature for this classification task. However, upon investigation, I observed that the autocorrelations of signals with different qualities appear very similar.

Is it theoretically possible to use autocorrelation to differentiate between the quality (high, medium, low, very low) of periodic signals? If so, what characteristics of the autocorrelation function might be indicative of different signal qualities for periodic signals?

Thank you in advance for any insights or suggestions.

So "quality" means less noise?

And the signal is periodic? Would the period be an integer number of samples?

Some form of autocorrelation can be used for pitch detection. And, once you know the pitch (by that I mean you know the period of your periodic signal), then you can tune a comb filters using precision delay to separate the periodic signal from everything else (that I presume is the "noise" that you want to measure). You can make the teeth of that comb filter really sharp, but how sharp depends on how periodic your signal is.

• Thanks! Your explanation clarifies the role of periodicity and autocorrelation in my case. Yes, by "quality" I mean less noise. Ideally, a high-quality signal would have a strong periodic component with minimal noise and also the signal is periodic. I appreciate the suggestion about using autocorrelation for pitch detection and comb filters for noise separation. Commented May 22 at 17:28

An alternative to the comb filters is to determine the period, chop up the signal into individual cycles, determine the average cycle and than simply subtract it out. Whatever is left is the noise.

As RBJ has already pointed out things are a lot tricker if the period length isn't an integer amount of samples. In this case, it may be the easiest to resample the whole signal by a factor very close to 1.

Once it's properly aligned you can alternatively just do an DFT. That will give you the energy of the fundamental, the energy of the harmonics and the residual noise.

Which method is the "best" depends on how much drift you have in fundamental frequency, phase & amplitude of the harmonics, etc.