0
\$\begingroup\$

I'd like to measure low-level noise in the 1 Hz to 100 kHz range, for example at the output of a LDO. For higher frequencies a spectrum analyzer would do the job, but most only go down to the MHz range. I have access to an Agilent E7495A which goes down to 500kHz (kinda), and I have seen a few go down to 9kHz (Agilent E4411B), but I've never seen a 1 Hz analyzer.

TI for example has a whole video that describes how to measure LDO noise: https://training.ti.com/engineer-it-how-measure-ldo-noise-and-psrr They show using some kind of analyzer, but what is it?

What is the usual type of instrument used for this? Spectrum analyzer connected in a special way, audio analyzer, something else?

\$\endgroup\$
  • 1
    \$\begingroup\$ An audio analyser will go down to 2 HZ and lower, but it is expecting to analyse a signal (clean sine-wave) that it generates in order to give you THD, etc. A DSO set to a low sweep rate maybe best. \$\endgroup\$ – Sparky256 Jul 30 '18 at 3:35
  • 1
    \$\begingroup\$ Any oscilloscope with built-in capability of spectral analysis ( or just a properly arranged FFT) will do the job as long as you acquire sufficiently long periodograms (and proper sampling rate to cover your 100 kHz point). \$\endgroup\$ – Ale..chenski Jul 30 '18 at 3:46
  • \$\begingroup\$ You can use a Spectrum Analyzer, the thing to keep in mind is that depending on the LDO you might or might not need a 'Post Amplification' stage, take a look at ti.com/lit/slyy076. Do you have an LDO in specific in mind? Often times manufacturers can might be able to tell you this information. \$\endgroup\$ – Victor S Aug 1 '18 at 3:15
3
\$\begingroup\$

A Windows PC can do this for free with Audacity. You still need the coax cables, Faraday shield box if necessary and an AC coupled 50 ohm load. View into a scope and amplify as needed from 0.1 V to 1 Vpp

Then measure for 1 minute for a smooth response as shown below.

Then feed the noise into your audio input and calibrate with a sine wave.

enter image description here

Some PC's have a DC-coupled mic input, with low offset and others have some HPF. You would choose a 1 Hz HPF.

\$\endgroup\$
3
\$\begingroup\$

You need something with low noise that can do FFT's, and can sample up to ~2x your measurement frequency.

What you need depends on how much noise there is. A switchmode or a poor linear regulator is relatively easy. A low noise regulator is likely to need a low noise instrument. Below 100Hz, and especially below 10Hz, the cmos inputs of many devices will be worse than the DUT.

If its very noisy, a digital scope might work. LDO's and random LM317, 7805 etc, often have very nasty noise peaks in the 10's of kHz caused by their feedback. These can be easy to see with anything that has FFT.

A good 14 bit digital oscilloscope can do it.

A low noise preamp with low pass filter, or using several simple analog filters to break the noise into 3 decade bands will also help, with a low resolution system.

Good soundcards can cover part of the range. Some used to sample at up to 192kHz (i.e. you can measure to 96kHz), and down to 20Hz. Below 20Hz the ADC often has a digital highpass function built in that you can't get access to turn off.

Below 10Hz you may be able to use a good digital multimeter.


Note that using the FFT results is quite confusing, when you want to make some absolute value measurements.

When you have a single frequency signal, the fft produces a correct value, and the value does not change with samplerate/fft-bin-width.

When you look at the broadband noise, the the amount of energy in one fft bin, depends on how wide that bin is. Change the samplerate, and the values all change. You must normalise the value to power-spectral-density (psd) i.e divide the power by the bin width.

However when you have both single bin spikes, and broadband noise, you have to decide which is which, and normalise them differently.

You will also notice in TonyEE's graph, that there is not a nice smooth accurate line, but rather a great hairy mess, which you can't read the value of. This is because the FFT contains an enormous number of tiny frequency bins, which have wildly different values in adjacent bins (it's noise after all).

TonyEE's graph

To clean this up, you need to "thin" the spectrum, summing multiple adjacent bins, so that you have far fewer bins - say 20 bins per decade, with logarithmiclly equal widths.

\$\endgroup\$
  • \$\begingroup\$ The deviation of adjacent bins from random noise depends on seconds of samples / sample rate or in other words number of log (samples) affects the noise standard deviation between bins and bin resolution. \$\endgroup\$ – Sunnyskyguy EE75 Jul 31 '18 at 22:51
  • \$\begingroup\$ @TonyEErocketscientist Doing overlapping FFT's (I think thats what you are saying) would reduce STDEV of the bins. However to get the same smoothing at 10kHz, as you would get from summing adjacent bins would take ~1000 samples, i.e. 1000 secs of acquisition. The point is that when you do an FFT down to 1Hz, you have a stupidly large numbers of tiny (1Hz), meaninglessly noisy, bins at 10kHz. If you collate the bins you get a meaningful, low stdev line. (I think doing averages of separate acquire/ffts, is worse, as it is uncorrelated averaging i.e. sqrt(N)) \$\endgroup\$ – Henry Crun Jul 31 '18 at 23:19
  • \$\begingroup\$ With a log-log FFT the bins are not equal AFAIK so the ripples are similar in each decade. My example uses 64k bins in a 1 minute sample. Audacity has the option to use less. I chose maximum on 2k resolution screen which is overkill. \$\endgroup\$ – Sunnyskyguy EE75 Jul 31 '18 at 23:29
  • \$\begingroup\$ An FFT is always linear, plotted onto a log scale. They may do multiple FFT's , one for each decade. (The old instruments only had enough memory to do the fft decade by decade). Still within the decade, the bins are linear, not logspace. \$\endgroup\$ – Henry Crun Jul 31 '18 at 23:36
  • \$\begingroup\$ You might be able to hit the "export" button, and if it can export a datafile which has the frequencies, we could see what Audacity is actually doing rather than conjecture. It would be interesting to know. \$\endgroup\$ – Henry Crun Jul 31 '18 at 23:43
2
\$\begingroup\$

This can be done by using the audio frequency spectrum analyzers (e.g. SR785 or SR1 or venerable HP3562A). They have input noise of ~10-20 nV/rHz which is well below the noise level of things like the LM7805/7905.

Of course, better would be something built in this century. An instrumentation amp (either 3 opamp or the AD620) with some AC coupling can be used to amplify ahead of a commercial digitizer like LabJack or even the microphone input on your laptop.

\$\endgroup\$
1
\$\begingroup\$

Some LDOs operate near 1uA Idd, controlling the gate of a huge PMOS device. With various internal circuits and stages that each need 100nanoAmps to operate, and needing an onchip feedback voltage divider, that voltage divider may be your noise-floor defining circuit.

Assuming 100nA (0.1uA) through the divider and 5 volts across the divider, the total resistance is 50MegOHm. Assume the divider is formed of 2 resistors, each of value 25MegOhm.

The noise floor of the divider will be 25Mohm || 25Mohm = 12.5MegOhm.

The noise density will be 4 nanoVolts/rtHz * sqrt(12.5Mohm / 1Kohm)

or 4nV * sqr(12,500) = 4nV * 110 = 440 nanoVolts/rtHz.

In 10KHz bandwidth, this scales up to 440nV * sqrt(10,000 Hz)

or total integrated noise of 440nV * 100 = 44 microVolts RMS.

Because of this (large) noise floor, people designing high-end vinyl-playback systems --- particularly the preamplifers for Moving Coil cartridges, design their own discrete high-idling-current power regulators. Often the topology of choice is a shunt-regulator.

You can find discussion of this in "diyAudio", under "simplistic njfet riaa".

Notice you are trying to measure 44 microVolts RMS, if you are using a 1uA LDO. For even lower noise LDOs, with some having 1uV total integrated output but quite HIGH idling current because in general the path to low-noise is to use LARGE transistors. And those large transistors have lots of parasitic capacitance that require high currents to achieve adequately high bandwidth in the feedback servo regulator amplifier loop.

Thus you need a low-noise preamplifier, AC coupled, between the LDO output and your modern scope with FFT or your spectrum analyzer.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.