Recognizing "sinusoidal" signal

[Pallern]

I am working with a pattern recognition problem, using 1-dimensional movement data, and some small movement patterns i'm looking at are characterized by looking like small sinuses (nice rounded curves), they appear sometimes here and there, with different duration (sometimes 1/2 period, sometimes 2 periods etc), and the amplitude and frequency is changing dynamically.

Can someone tip me of a "filter"/"pattern recognizer"/set of mathematic properties that will recognize this pattern from other types of signals? I have done some testing using neural networks (backprop) to discriminate the sinus-like patterns from everything else, but making the training set robust enough is a teedious job... Any input appreciated

 

[cah2oeng]

This certainly won't solve your problem entirely, but to shed more light on the situation you could try performing a spectral analysis on the signal (i.e. take the DFT of the signal). That would at least tell you how much power is alloceted at each frequency.

 

[GregLocock]

DFT won't help with such short signals. I'd stick with time domain filtering - you might have some luck with IIR (?) filters.

Could you post some data?

Cheers

Greg Locock

 

[Pallern]

Thanks for the response! I have tried using fft, but as the frequency and amplitude are varying in time, the information gets a bit "cluttered" in frequency domain. I guess what I'm asking for is an algorithm to recognize a signal that is smooth, looks like sinuses of varying amplitudes/frequency, as opposed to signal that contains straight lines/spikes -> not as nice-looking.

- how do I post my data here, is it possible to "attach" files to messages?

 

[cah2oeng]

You can take the fourier transform of any signal, regardless of length, so the short nature of this data wouldn't preclude its use. However, it sounds like you've tried that route.

What about using a correlation algorithm? In my line of work, we use these to track a particular pattern as it moves in space from one sequential image to another. Maybe this idea could be adapted to search for the sinusoidal patterns you expect in your 1-D data.