spread-spectrum switcher

On Oct 3, 1:04=A0am, John Larkin

Like it or leave it. Build your own - if you can.

Of course you can't speak any other language.

At lowish frequencies, you can use one of those fancy digital storing scopes to record the digitized stuff to a file and then FFT the file. I have done this to check the spectrum of a class D circuit that ran at 500KHz to 3MHz and back.

Most of our scopes now do FFTs internally. It's nowhere as good as a spectrum analyzer, but fine for quick looks.

John

Nice. Hopefully 300kHz spread will be enough for you. Where's that

If you do this a lot get a real analyzer. Since you guys have office space galore maybe "the" baseband analyzer of all times, the HP3585. That's the one I recommend to all my ultrasound clients. But please don't buy all of them like you did with sampling scopes :-)

Looks English to me. You should see some of the error messages on my Instek scope when a cable isn't plugged in or something. "DSO not connect" and so on. But the thing works and that's most important. And has a dozen selectable languages.

I think it's other spurious lines being smeared sideways. Neat shape. It even has sort-of-Gibbs' ears at the corners.

The Rset resistor on this chip is 150K for 2 MHz, and the node is very noise sensitive. It's mildly spread-spectrum all by itself, without help from me, with a broad-ish spectral line and visible jitter looking at the PWM on a scope.

One of my guys researched SAs and picked the Aeroflex, based on specs and features, so we bought it. It's OK, but has a lot of bugs and quirks. It turns out to be Korean, and the local guys don't know much about it and can't get anything, like the firmware, fixed. All the menus and messages are in kringlish.

It says "Wait Until Calibration!" and worse.

John

The internal one is usually limited to power of two lengths. The one I use on the PC can take any length where the biggest prime factor is less than some number above 100. This means that I can try various length FFTs to remove the skirts on the peaks caused by the length not being an integer number of cycles.

The scope I have is a Tek-um-err-um well there are 4 digits in it. It has a 32K memory.

Some run with ridiculously low set currents. This is one reason why I often roll the complete switcher myself. Because too many commercial chips are too noisy for sync'd stuff like ultrasound.

On my current design I slammed the chip into permanent foldback so I can use a set current about five times higher than normal. I could sense some disgust in the face of the app engineer but he said that's fine. I guess to him it looked like mounting Rubicon-style tires on a Jaguar.

Mine's Chinglish but can be switched to Korean :-)

Ouch ...

I remember one serious translation goof-up. English manual of ultrasound machine said to freeze image when turning machine off. German version only said something like "Bitte einfrieren", meaning please put into freezer. So, someone dutifully put the transducer into the freezer. Tsk ... pop ... roughly $10k worth of damage.

This looks like a strange way to do things; why would they do that? The standard way to reduce the artifacts of FFT is windowing. If you want the frequency of interest exactly on a FFT bin, you can interpolate either in time or in frequency domain.

Vladimir Vassilevsky DSP and Mixed Signal Design Consultant

It's particularly odd since the spectral artifacts due to truncating the data length are still there regardless of what length you pick, unless the data are actually periodic in that length.

Cheers

Phil Hobbs

Windowing only reduces the skirts. If the FFT contains an exact number of cycles of a frequency that frequency will have no spread on it at all.

Interpolation in time introduces its own artifacts, because you have to assume things about what the signal does between samples. It is very common for sample rates to be nice even multiples of 10 and 1,2 or 5. The same can often be true of the frequency you are looking for. Having a FFT that does a length 1 million FFT can be very handy in this case.

In my case I wanted to make sure that there were no strong peaks at unwanted sub harmonics of the crystal things ran on near the frequency that was supposed to be there. ie: I needed to see things near a strong carrier and small set of side bands at frequencies I knew.

The artifacts are there for anything that isn't an integer number of cycles but not there for things that are.

In an unrelated example: Consider the case where you sample at 1000Hz and can't avoid having some 60Hz and its harmonics in the data. If you can make the data length an integer number of seconds, the 60Hz etc will be confined to one bin but the general noise will be spread out.

not

The=20

=20

Spot on. That is also why many good (FTT using) instruments provide a variety of windowing (temporal data weighting) functions to allay various artifacts of sampling windows. Each windowing weighting method has its good an bad properties. Sometimes you want to just capture raw and try several different windowing functions just to see the differences in impact on the sample.

All of the windowing functions necessarily trade some resolution to control the edge effects inherent in the implicit periodicity of the FFT. The result of multiplying in real space is a convolution in frequency space and vice-versa. Some practitioners use the mirror boundary condition FFT rather than translation. This eliminates any edge discontinuity but can introduce other artefacts.

Though if you are really interested in weak detail at frequencies close to a powerful harmonic you can usually do better with maximum entropy methods. Their ability to model sharp features accurately in the frequency domain gives much better results than windowing. The artefacts are different so you have to be a little bit careful. Some basic details of using MEM for power spectrum estimation are in Numerical Recipes.

Driving the noise floor down by averaging multiple independently measured power spectra is also beneficial.

Regards, Martin Brown

If you have multiple measurements, then shift-and-add with about 75% overlap of the data sets will give good results. It also wastes less data, since (apart from the beginning of the first data set and the end of the last one) every sample has the opportunity to be in the middle of the window, and so contributes about equally.

MEM has problems when there are actually discrete sine waves in the data. In that case you're usually better off subtracting the sine wave before doing either MEM or DFT analysis.

Cheers

Phil Hobbs

as a

one

is

various

not

=20

you=20

=20

edge=20

close=20

artefacts=20

details=20

=20

This relates to an interest of mine. Just how computationally intensive is this? My interest relates to segregating phase relative consistent harmonic series from a complex signal.

In practice I agree. Although it does introduce some marginal bias into the results if there is any overlap. I suppose it depends how much data you have to play with.

A decent MEM algorithm should be able to handle a huge dynamic range even with a pure sine wave present. That isn't always true in practice.

The only snag is that you can get peak splitting due to noise if you attempt to overfit the data (which a lot of users tend to do). Having said that you are right that if there is one obvious carrier wave you may as well fit it by other means and then look at the spectrum of the residuals. No point in sacrificing dynamic range if you don't have to.

Answer is it depends. The general MEM algorithms tend to be between 100 and 200 times slower than a classic linear inverse. And with modern kit this isn't prohibitive any more.

But there are a handful of closed form analytical results for 1-d time series that are considerably faster. They only work for certain problems but one of them is specttal analysis. For an intro to it have a look at the descriptions in Numerical Recipes (but don't trust their code for anything critical).

Sections 13.6 & 13.7 using the Burg entropy definition S=log(f) and a closed form solution that is relatively fast to compute. The other MEM variant uses the Gull & Daniell entropy definition S= flog(f) and is generally more computationally expensive non-linear optimisation problem. See the original paper in Nature (ppv or subscriber):

A better early MEM algorithm is described in detail in Skillings paper in MNRAS.

Although a closed form flogf MEM solution is known for 1-D spectral estimation problems it is not often used because of numerical instability.

Regards, Martin Brown

Interesting, thanks. One more for the toolkit.

Cheers

Phil Hobbs