I'm dealing with a signal generated by a rotating machine. I am trying to separate components generated by an anomaly that occurs exactly once per revolution from those generated by an anomaly the occurs exactly twice per revolution. Of course, harmonics of the twice-per-revolution anomaly overlay harmonics of the once-per- revolution anomaly. Seeking suggestions on what analyses to perform to separate the two. My initial guess is that phase will somehow be the discriminating factor. Thanks, Greg
Separating Harmonic Spectra
Started by ●November 30, 2007
Reply by ●November 30, 20072007-11-30
"Greg Berchin" <gberchin@sentientscience.com> wrote in message news:db44a883-dd60-426b-8719-5bb98a602249@s12g2000prg.googlegroups.com...> I'm dealing with a signal generated by a rotating machine. > I am > trying to separate components generated by an anomaly that > occurs > exactly once per revolution from those generated by an > anomaly the > occurs exactly twice per revolution. Of course, harmonics > of the > twice-per-revolution anomaly overlay harmonics of the > once-per- > revolution anomaly. Seeking suggestions on what analyses > to perform > to separate the two. My initial guess is that phase will > somehow be > the discriminating factor. > > Thanks, > GregThe event that happens twice per revolution shows up once by itself and once combined with the event that happens once per revolution. Any chance of phase-locking to the "by itself" event, subtracting it out of the "once-per-revolution event" and then processing the events as separate streams?
Reply by ●November 30, 20072007-11-30
On Nov 30, 8:50 am, Greg Berchin <gberc...@sentientscience.com> wrote:> I'm dealing with a signal generated by a rotating machine. I am > trying to separate components generated by an anomaly that occurs > exactly once per revolution from those generated by an anomaly the > occurs exactly twice per revolution. Of course, harmonics of the > twice-per-revolution anomaly overlay harmonics of the once-per- > revolution anomaly. Seeking suggestions on what analyses to perform > to separate the two. My initial guess is that phase will somehow be > the discriminating factor.If the impulse response of the once per cycle event spans less than an entire revolution, then you could periodically window outside that span in the time/angular domain and see what spectra is left from the other half of the revolution. To find each event, you could try a pair of angular windows 180 degrees apart, and rotate them together until until you find some sort of maxima in the difference between the two spectra. IMHO. YMMV. -- rhn A.T nicholson d.0.t C-o-M
Reply by ●December 1, 20072007-12-01
On Fri, 30 Nov 2007 08:50:15 -0800 (PST), Greg Berchin <gberchin@sentientscience.com> wrote:>I'm dealing with a signal generated by a rotating machine. I am >trying to separate components generated by an anomaly that occurs >exactly once per revolution from those generated by an anomaly the >occurs exactly twice per revolution. Of course, harmonics of the >twice-per-revolution anomaly overlay harmonics of the once-per- >revolution anomaly. Seeking suggestions on what analyses to perform >to separate the two. My initial guess is that phase will somehow be >the discriminating factor.Odd harmonics of the 1/per won't be duplicated in the 2/per. -- John
Reply by ●December 1, 20072007-12-01
On Sat, 01 Dec 2007 17:00:38 -0600, John O'Flaherty <quiasmox@yeeha.com> wrote:>On Fri, 30 Nov 2007 08:50:15 -0800 (PST), Greg Berchin ><gberchin@sentientscience.com> wrote: > >>I'm dealing with a signal generated by a rotating machine. I am >>trying to separate components generated by an anomaly that occurs >>exactly once per revolution from those generated by an anomaly the >>occurs exactly twice per revolution. Of course, harmonics of the >>twice-per-revolution anomaly overlay harmonics of the once-per- >>revolution anomaly. Seeking suggestions on what analyses to perform >>to separate the two. My initial guess is that phase will somehow be >>the discriminating factor. > >Odd harmonics of the 1/per won't be duplicated in the 2/per.Whoops, said it backwards. -- John
Reply by ●December 2, 20072007-12-02
On Sat, 01 Dec 2007 17:01:47 -0600, John O'Flaherty <quiasmox@yeeha.com> wrote:>Odd harmonics of the 1/per won't be duplicated in the 2/per. > >Whoops, said it backwards.Yes; that is one of the properties that I was hoping to exploit. But given a signal that is an unknown combination of both, how do I separate their effects? I tried bispectral analysis, but it's darned near impossible to interpret the results. Thanks. Greg
Reply by ●December 2, 20072007-12-02
Greg Berchin wrote:> On Sat, 01 Dec 2007 17:01:47 -0600, John O'Flaherty <quiasmox@yeeha.com> > wrote: > >> Odd harmonics of the 1/per won't be duplicated in the 2/per. >> >> Whoops, said it backwards. > > Yes; that is one of the properties that I was hoping to exploit. But > given a signal that is an unknown combination of both, how do I separate > their effects? I tried bispectral analysis, but it's darned near > impossible to interpret the results.Can we assume that the twice-per-revolution anomaly is actually a once-around for a part geared 2:1? If not, does it actually consist of dub dub dub dub 180 degrees apart, or might it be a syncopated lub-dub ... lub-dub ...? (I hope that's not too cryptic!) Will the harmonics of the anomalies have zero phase when the fundamental is also zero? There's a good chance of that if the acoustic delay is small enough. Then you might be able to sort the harmonics into groups based on their phases. Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
Reply by ●December 2, 20072007-12-02
On Sun, 02 Dec 2007 07:46:19 -0500, Greg Berchin <gberchin@comicast.net> wrote:>On Sat, 01 Dec 2007 17:01:47 -0600, John O'Flaherty <quiasmox@yeeha.com> >wrote: > >>Odd harmonics of the 1/per won't be duplicated in the 2/per. >> >>Whoops, said it backwards. > >Yes; that is one of the properties that I was hoping to exploit. But >given a signal that is an unknown combination of both, how do I separate >their effects? I tried bispectral analysis, but it's darned near >impossible to interpret the results.Just exploring this, how do you know the 2/per isn't just a harmonic of the 1/per? If they're separate anomalies, can you eliminate either temporarily, to find out what the harmonic structure of the other is? Or, alternatively, can you temporarily increase the effect of the 1/per without affecting the 2/per? If you could once measure the amplitude and phase of the 1/per spectrum, and it was stable, then you might subsequently be able to infer its complete spectrum from the measured strength of a non-duplicated odd harmonic, and get the 2/per by subtraction. -- John
Reply by ●December 2, 20072007-12-02
On Sun, 02 Dec 2007 12:09:07 -0500, Jerry Avins <jya@ieee.org> wrote:>Can we assume that the twice-per-revolution anomaly is actually a >once-around for a part geared 2:1?No. The "once-per" is a physical situation that occurs at only one rotation angle; the "twice-per" is a physical situation that occurs at two different rotation angles.>If not, does it actually consist of >dub dub dub dub 180 degrees apart, or might it be a syncopated lub-dub >... lub-dub ...? (I hope that's not too cryptic!)Nominally 180�, but this signal is indicative of severe wear, so there will be plenty of room for slop.>Will the harmonics of the anomalies have zero phase when the fundamental >is also zero?There is a fixed phase relationship between the shaft angle, the angle at which the "once-per" occurs, and the angles at which the "twice-per" occur. They will vary from device to device, but within a given device they will be fixed.>Then you might be able to sort the harmonics into groups >based on their phases.That's my working theory. I liked the idea from Ron N about selectively windowing in the time domain, but I just haven't yet figured out exactly how to implement it to the greatest effect. I know almost nothing about the waveforms a priori, except that one occurs once per rotation and the other occurs twice per rotation. Greg
Reply by ●December 2, 20072007-12-02
Greg Berchin wrote:> That's my working theory. I liked the idea from Ron N about selectively > windowing in the time domain, but I just haven't yet figured out exactly > how to implement it to the greatest effect. I know almost nothing about > the waveforms a priori, except that one occurs once per rotation and the > other occurs twice per rotation.Greg, What if you integrate the signal in the two dimensions per revolution and per half revolution? Let's say you have N samples per revolution. For every sample: I[n] = x[n] + x[n + N] + x[n + 2N] .... ^^^^^^^^^^^^^^ This accumulates single and double events. J[n] = x[n] + x[n + N/2] + x[x + N] + x[n + N + N/2].... ^^^^^^^^^^^^ This accumulates 2xdouble events, 1xsingle events and noise which is hopefully zero mean. So: J[n] - I[n] = double events + noise which is averaged out. The two dimensional FFT also comes to the mind. Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
