Noise measurement with spectrum analyzer

[Mrt12]

Hi guys
I am not quite sure whether this is the right section of the forum, but I think my question fits here best.

I just had a discussion with a colleague. We tried to measure some noise spectra of oscillators with a spectrum analyzer and had a debate about which detector needs to be used.
I do remember slightly that I once read an application note from Dieter Scherer at HP, that for the measurement of noise related stuff, the "sample" detector should be used. So I used that and my measurements look reasonably good.
However my colleague thinks the "RMS" detector should be used. But I am not sure.

Can someone please explain to me which detector I should use and why?
My spectrum analyzer lets me choose between "pos peak", "neg peak", "RMS", "sample" and "Quasi peak". I know the latter is only used for EMC stuff, so I don't use that.

[jlake]

Are you trying to measure phase noise?

If so, check that your SA doesn't have a built in routine for measuring such things. Lots of modern SAs do. Also, make sure that your SA's LOs have a better phase noise than your DUT (>10dB).

I would say that both sample and 'normal' detectors would give accurate results if the BW is configured accordingly. I'd probably use normal detection though.

Sample is just going to give you the frequency in the middle of each frequency bucket (span/(points-1)). It's therefore quite good at providing noise information. Sample can miss CW signals though, because it only takes into account one point per bucket!

Normal detection is a bit more complicated and uses the Rise and Fall algorithm to present the data. It's typically used when CW signals and noise need to be measured at the same time. It first measures whether the signal is rising or falling or both within a bucket. If it's going up and down within the bucket, it's noise. If it's just going up or just going down, it's a signal. It's this distinction and subsequent processing that allows it to represent signals and noise ostensibly accurately.

RMS is going to add up all of the (data)² in your bucket and then divide it by the points per bucket and then sqrt it. This is going to give you a power.

[David Hess]

Any should work assuming that the spectrum analyzer's noise marker function correctly compensates for the measurement type, but RMS is the most appropriate one.  If you do not have a noise marker, then you will have to do the compensation to find RBW (resolution bandwidth) manually and RMS should be used.

[jlake]

Any should work assuming that the spectrum analyzer's noise marker function correctly compensates for the measurement type, but RMS is the most appropriate one.  If you do not have a noise marker, then you will have to do the compensation to find RBW (resolution bandwidth) manually and RMS should be used.

While this is a good shout (just using a noise marker---if your SA has one), this doesn't make any attempt to explain why RMS is the most appropriate detector. Why would RMS be advantageous over Rosenfall or sample?

[David Hess]

Any should work assuming that the spectrum analyzer's noise marker function correctly compensates for the measurement type, but RMS is the most appropriate one.  If you do not have a noise marker, then you will have to do the compensation to find RBW (resolution bandwidth) manually and RMS should be used.

While this is a good shout (just using a noise marker---if your SA has one), this doesn't make any attempt to explain why RMS is the most appropriate detector. Why would RMS be advantageous over Rosenfall or sample?

It is straightforward to combine the RMS measurement and RBW to get the noise density at any one point while other measurements require a correction factor which may not be available.

This is also what makes noise measurements using the FFT on a DSO difficult; you have to do the correction for RBW and FFT window yourself.

[SilverSolder]

Any should work assuming that the spectrum analyzer's noise marker function correctly compensates for the measurement type, but RMS is the most appropriate one.  If you do not have a noise marker, then you will have to do the compensation to find RBW (resolution bandwidth) manually and RMS should be used.

While this is a good shout (just using a noise marker---if your SA has one), this doesn't make any attempt to explain why RMS is the most appropriate detector. Why would RMS be advantageous over Rosenfall or sample?

It is straightforward to combine the RMS measurement and RBW to get the noise density at any one point while other measurements require a correction factor which may not be available.

This is also what makes noise measurements using the FFT on a DSO difficult; you have to do the correction for RBW and FFT window yourself.

Is it possible to sample noise with a SA by taking a sequence of point measurements so the RBW "borders" touch each other, over the frequency band of interest?

[cncjerry]

Depending on your analyzer and if it is supported, try John's PN.exe.

http://www.ke5fx.com/gpib/pn.htm

The problem with most analyzers is the noise floor is no where near low enough to measure PN of anything but possibly RF synthesizers, poorly performing DDS units, etc.  I've been trying to measure PN with many analyzers and gave up.  I found a relatively low-cost way to measure PN, low cost in that it is under $1200 or so.  The noise floor is below -180dBc which is relatively good for such a low cost setup.

This setup is based on a method of using a reference quadrature double balanced mixed signal to translate the DUT oscillator signal to baseband removing the carrier in the process with the noise remaining.  The remaining noise is then amplified above the noise level of a baseband analyzer for display. I use an HP 3562a but there are others including low noise PC audio cards.  I've not had a lot of success with this method as I don't seem to have the low noise amplification needed.  But before I solved that I found the FPGA method and decided to invest in it.

Jerry

[Mrt12]

Dear colleagues

I'm sorry for my delayed response. For some reason I don't get an email notification!

Indeed I am trying to measure phase noise - for a cheapo synthesizer at 2 GHz with a cheap XTAL (not even an OCXO!).
So the phase noise is so bad I can directly measure it with the S/A. I am lucky because my spectrum analyzer (R&S FSP30) does have an FFT option which easily allows me to use 1 Hz RBW filters without waiting for ages.

I used the spectrum analyzer's REF out as reference for the synthesizer and measured the spectrum at different offsets from the carrier with 1 Hz RBW. I used the sample detector and trace averaging.

But still I don't understand yet what the difference between "SAMPLE" and "RMS" detectors is.

[SilverSolder]

Depending on your analyzer and if it is supported, try John's PN.exe.

http://www.ke5fx.com/gpib/pn.htm

The problem with most analyzers is the noise floor is no where near low enough to measure PN of anything but possibly RF synthesizers, poorly performing DDS units, etc.  I've been trying to measure PN with many analyzers and gave up.  I found a relatively low-cost way to measure PN, low cost in that it is under $1200 or so.  The noise floor is below -180dBc which is relatively good for such a low cost setup.

This setup is based on a method of using a reference quadrature double balanced mixed signal to translate the DUT oscillator signal to baseband removing the carrier in the process with the noise remaining.  The remaining noise is then amplified above the noise level of a baseband analyzer for display. I use an HP 3562a but there are others including low noise PC audio cards.  I've not had a lot of success with this method as I don't seem to have the low noise amplification needed.  But before I solved that I found the FPGA method and decided to invest in it.

Jerry

Is that the same basic idea as notching out the fundamental with a twin-T filter when measuring audio distortion?

[Mrt12]

Is that the same basic idea as notching out the fundamental with a twin-T filter when measuring audio distortion?

not quite. You need a second oscillator which has much better performance than the DUT and a PLL. Then, the DUT is mixed (heterodyned...) with this auxiliary oscillator. The PLL enforces the mixer DC output voltage to zero such that the two oscillators are in quadrature.
The mixer DC output voltage contains the phase noise spectrum of both oscillators. It can be amplified and measured with a spectrum analyzer.

[SilverSolder]

Is that the same basic idea as notching out the fundamental with a twin-T filter when measuring audio distortion?

not quite. You need a second oscillator which has much better performance than the DUT and a PLL. Then, the DUT is mixed (heterodyned...) with this auxiliary oscillator. The PLL enforces the mixer DC output voltage to zero such that the two oscillators are in quadrature.
The mixer DC output voltage contains the phase noise spectrum of both oscillators. It can be amplified and measured with a spectrum analyzer.

So a kind of very precisely controlled double balanced mixer that suppresses both the DUT and local oscillator, leaving just the "garbage" for analysis?   Sounds like a cool idea!

[David Hess]

It is straightforward to combine the RMS measurement and RBW to get the noise density at any one point while other measurements require a correction factor which may not be available.

This is also what makes noise measurements using the FFT on a DSO difficult; you have to do the correction for RBW and FFT window yourself.

Is it possible to sample noise with a SA by taking a sequence of point measurements so the RBW "borders" touch each other, over the frequency band of interest?

I do not know why you would do that or that it would have any advantage.  Once you know the RBW, which depends on the filter's bandwidth and shape factor or FFT bin width and FFT window, and have the RMS measurement, then that is all you need to calculate the noise at one point.

[SilverSolder]

It is straightforward to combine the RMS measurement and RBW to get the noise density at any one point while other measurements require a correction factor which may not be available.

This is also what makes noise measurements using the FFT on a DSO difficult; you have to do the correction for RBW and FFT window yourself.

Is it possible to sample noise with a SA by taking a sequence of point measurements so the RBW "borders" touch each other, over the frequency band of interest?

I do not know why you would do that or that it would have any advantage.  Once you know the RBW, which depends on the filter's bandwidth and shape factor or FFT bin width and FFT window, and have the RMS measurement, then that is all you need to calculate the noise at one point.

Sometimes you want to known the noise in a specific bandwidth?  - can the RBW be set for a specific range, e.g. 20Hz - 20KHz to measure an audio amp?  I was thinking you might have to use e.g. 1KHz RBW repeatedly until the range was covered...

[cncjerry]

I think John`s PN.exe measures the 1hz noise at selected points, preventing the long, full plots from 100hz to 1Mhz.  Again, it works but is limited by the noise floor of most analyzers I can afford.

[SilverSolder]

I think John`s PN.exe measures the 1hz noise at selected points, preventing the long, full plots from 100hz to 1Mhz.  Again, it works but is limited by the noise floor of most analyzers I can afford.

It also assumes the noise is evenly distributed?

[David Hess]

I do not know why you would do that or that it would have any advantage.  Once you know the RBW, which depends on the filter's bandwidth and shape factor or FFT bin width and FFT window, and have the RMS measurement, then that is all you need to calculate the noise at one point.

Sometimes you want to known the noise in a specific bandwidth?  - can the RBW be set for a specific range, e.g. 20Hz - 20KHz to measure an audio amp?  I was thinking you might have to use e.g. 1KHz RBW repeatedly until the range was covered...

That is called a spot noise measurement.  You can either integrate the noise density over the bandwidth of interest to calculate the spot noise, or create a filter of the appropriate bandwidth and shape factor to measure the spot noise directly.  The results will be the same and you can roughly convert between the two.

Back before DSOs with FFT capability were common, the options were to use a spectrum analyzer to measure noise density or to make a spot noise measurement with a filter and RMS detector.  The later method is actually very easy to do on an analog oscilloscope if you can filter the input signal.  Some analog oscilloscopes supported variable low and high pass filters exactly for this sort of application.

[David Hess]

I think John`s PN.exe measures the 1hz noise at selected points, preventing the long, full plots from 100hz to 1Mhz.  Again, it works but is limited by the noise floor of most analyzers I can afford.

It also assumes the noise is evenly distributed?

No, it means that the noise is measured over a 1 HZ bandwidth at the frequency of interest, which results in the noise spectral density at that frequency of either Volts^2 / Hz or Volts / SqrtHz.  If you were making a spot noise measurement which encompasses a much wider bandwidth, then it does assume the noise is evenly distributed.