Discrete signal recognition from continuous waveform

Question

Hi Folks,
I'm hoping some of you can give some general pointers on how to better approach this problem. This is the first application I will be actively performing waveform analysis so forgive the amateurishness.
Anyway I'm acquiring a physiological signal at 50Hz and attempting to recognize 'breaths' from this source. The problem is that the frequency, amplitude, and offset will vary depending on the rate and depth of breath (among other things). Below are the ranges I'm working with (in units of "WC pressure):
Freq: 10mHz-500mHz
Amp: 0.005 to 0.145 "WC
Offset: -0.25 to -1.0"WC
In general, I'm trying to reliably get a measure on each breath's amplitude (and offset, but that's easy) and frequency. I have tried two methods: to isolate each breath and analyze, or to buffer say 10seconds worth of signal and analyze. The problem mostly comes in when I try to use the "Pulse Measurements" and "Amplitude and Levels" vi's as part of the analysis package. The way in which they perform their calculation relies on set crossing thresholds - if the signal jumps in amplitude or quickly changes frequency (not likely) and I'm using the buffer approach, I lose calculation on the buffer. Obviously using a fixed buffer size (can) give different results depending on the frequency of the signal.
It's amazing how we (as humans... right?) can discern the breaths from this signal, how we can easily identify problematic shapes, but to do this programatically, it's a completely different story.
I'm just unsure how to better approach this - should i be trying to fit an ideal curve to every group of samples? Attached is a screenshot of ideal (at two distinct rates) and actual.
Message Edited by 8bitbanger on 07-10-2009 02:19 PM
v2009 devel. w/RT
Attachments:
inWC_actual.PNG ‏44 KB
inWC_ideal.PNG ‏43 KB

Answer 1

First, I would filter out the small quick changes is the waveform (unless they are significant to what you are doing or you consider them to be a  breath).
In the past, I have tried to heavily filter respiration down to a sinusoid signal and then take the Hilbert transform to create an analytic signal (f(t) + jH{f(t)}).  Take a look at the derivative of the phase of this signal which is equivalent to its instantaneous frequenc.
Take a look at attached.   I saved data into the Data In control for you to use as an example.  This data is monkey respiration, so it is a little faster than human. Play around with the filters to see what is best for your data.  Hopefully, this will be of some help.
Randall Pursley
Attachments:
Resp Rate.vi ‏269 KB