1/f flicker noise in regulators

[Teleno]

Hi all.

I've built the following oscillator that I'm trying as a pinpointer:



Instead of the battery in the schematic I'm using an LM317 regulator (noisy!) as the power source because I want to control the amplitude.

After passing the probe's voltage through a peak detector I measured a lot of flicker noise. The drift at 1Hz to 10Hz is in the order of tens of millivolts.

The problem is that a small DC change in V1 is multiplied by 10x in the oscillator (I measured that), and so does the 1/f noise.

As a remedy I can think of various possibilities:

1. Use a regulator with low flicker noise (any recommendations for 4V-6V, 200mA?).

It's the obvious solution, but then again it'll get multiplied by 10x.

2. Use an active noise cancelling circuit. Example: http://www.wenzel.com/documents/finesse.html

Increased complexity and component count.

3. Use a DC-DC boost converter.

Question: Are boost converters a source of flicker noise just like series regulators? The switching frequencies about 1MHz is far above the bandwith of the pinpointer, no problem there.

4. Use a battery without a regulator.

Question: How noisy is a battery, if at all?

Thanks in advance for your comments and insights!

[Davor]

A battery is the only voltage source/reference with negligible 1/f noise. For that reason there are some new regulators with voltage reference input aimed precisely for using batteries as a noiseless reference.

As for this circuit... it lacks stability even with a regulator. It makes it a good candidate for pinpointing by virtue of wild amplitude variation.
So you are out of luck on this one. You'll not be able to tame it without a fixed voltage source.

You may build a regulator with a red LED (or 2, or more) as a decent reference, instead of a zener, but make sure it receives enough current. Also zeners below about 5.1V tend to be considerably quieter than those with higher voltage.

[6666]

Also zeners below about 5.1V tend to be considerably quieter than those with higher voltage.

Thats interesting , so would that mean we would be better off with a string of smaller voltage zeners rather than just one ?

[Davor]

Yes.
Also using Si diodes, or LEDs, except the blue or white ones (those are basically the same, except the white ones have phosphors added on top).
But the best of the best is designing your equipment without regulators. Just bare batteries.

Guess some sort of polarity and overvoltage protection, combined with proper design that works well in a wider range of supply voltages, would be the best choice.

[6666]

Thanks Davor

[Teleno]

Thanks for the insights, Davor.

The oscillator is very sensitive to tiny changes in its supply voltage. If I settle for the battery - the less noisy option - then I'll need a separate for the rest of the system: MCU, beeper etc. because any change in the current drawn by these devices will be felt by the oscillator.

A solution would be to draw a constant current from the battery to feed the rest of the electronics, a current high enough to satisfy any demand (worst case). This way there will be no variation in the voltage. The drawback is a higher current consumption.

The simulation shows the comparative effect of feeding the rest of the electronics from the battery using the above solution. In the first example (top), a 1mV change in the supply voltage translates in 10mV amplitude change at the oscillator. In the second example the constant current draw totally prevents the passage of glitches to the oscillator.

[Teleno]

While investigating 1/f noise I came across Brownian noise (random walk noise) that has a 1/f² power spectrum and happens to be the integral of white noise.

I'm using a classical control loop to keep the amplitude of the oscillator at a fixed value, like this one:




I suspect the white noise from the oscillator and the envelope detector becomes brownian 1/f² noise at the integrator's output (my signal). The extreme DC drift you can see in the 10 minutes of action recorded below look a lot like brownian motion:

[Davor]

Actually, oscillators containing LC elements are performing integration on their own. It depends on the level of excitation, but you may observe nearly linear rise in envelope of an oscillator building up its oscillation. Once they reach rail voltage, their amplitude gets limited. For that reason an oscillator is in effect a P-I regulator of the oscillation envelope.
So when you add one more instance of integration, it becomes unstable. Such oscillators will overshoot on turn on, and will be very touchy with just about everything. Therefore, adding integration in regulation path is a bad idea - it is already there.
Your best bet regarding amplitude stability is to establish some tight relationship between rail voltage and oscillation amplitude, and let the oscillator reach it on its own. Make sure the rail is quiet enough in 1/f region, and that's it. You can't erase its integration nature, so instead you simply go with it, but forget about trying to fix it with more integration.

[Teleno]

I've built the circuit and connected it to an Arduino Pro Micro for signal processing.

Q3 and Q4 form a fast reference current source (Iref) to subtract a fixed voltage from the oscillator output (R3 x Iref). This configuration works as a dead-band amplifier, the peaks of the sine wave are level shifted from around 40V down to 3V so the changes in amplitude can be measured without loss of gain.

Q7 and U3 form a simple peak detector. U1 and U2 form an amplitude control loop that tries to maintain the output of the peak detector at 3V (Vref). The control voltage at the output of U2 is the signal we're interested in. The gain is approximately R7/R4 = 47.




As I said, there are sever issues with 1/f noise in the control loop, that's why I'm using digital signal processing: (1) differentiation to block DC, (2) lossy integration of the difference and separation of the target response patterns from the background noise.

This is the raw signal dominated by 1/f noise and 1/f2 random walk:




And this is the processed signal, the green peaks indicate the detection of a target.




The nugget is detectable at a 5cm. distance, a somewhat larger one at 10cm.

Here are the PCB rig and the coil (the small PCB inside the coil is the oscillator).







It's like a supercharged pinpointer that wants to become a metal detector.

[Davor]

You'll need some black sands together with a nugget, and see if you can detect it. There are good reasons for using GB in nowadays detectors.

[Teleno]

You'll need some black sands together with a nugget, and see if you can detect it. There are good reasons for using GB in nowadays detectors.

Ferrite very easily drives the detector far off the working point causing the feedback loop to saturate. The cause is the increased inductance of the LC tank.
I've found out that amplitude changes due to inductance follow a linear law dA/dL = k. I will try to have the MCU measure k based on the frequency and the (constant) value of C3. Then this factor can possibly be subtracted from the output signal for GB.




The ferrite signal is ignored by the detection algorithm.

[Teleno]

There are good reasons for using GB in nowadays detectors.

For this kind of detector I propose the following GB scheme: A PLL (CD4046) tracks the LC tank, outputs a voltage proportional to the resonance frequency (Vf) which is used to adjust the reference current of the current mirror Q3/Q4 (Iref) so that the working point (R3 x Iref) does not change with frequency.

This would at least prevent large swings of the working point, the MCU can then find the actual relationship between Vf and Vout for a pure magnetic target and subtract this coponent from the received signal.

[Davor]

It is not my intention to burst your bubble, but this scheme assumes ground influence will remain the same regardless of proximity. Guess you are in for a surprise.