DC-accurate Low-pass-filter

[iMo]

..fyi..

[EC8010]

Do you mind posting the 1 mHz circuit? This looks like some fun projects to build into a tin.

Here it is. As you can see, I shifted everything down by one decade, so it's now a 100Hz low-pass rather than 1kHz. It needed quite a large (Marks & Spencer toffee) tin. The choke was a commercial one found at a junk shop in Toronto. Be warned that large chokes are usually made for power supply smoothing and are magnetically leaky, so they easily pick up hum. If you have the choice between an EI or toroidal choke (of the same value), the toroid will pick up less hum. Of course, you could just fit your larger capacitor in the 1kHz filter in place of the 14uF - it won't change the other required values.

The reason I reduced the low-pass frequency by a decade was that if you genuinely want to look down to 1mHz, you need a sweep speed of 10,000s because the resulting FFT will go down to 0.1mHz but you need to discard the first ten frequency bins because they're unreliable. If you then note that you need to average 100 FFTs together to get a smooth noise spectrum, that's 100 x 10,000s = 11.6 days. That's an awful lot of data, so you need to be able to reduce sample rate, hence reducing the filter high frequency response to 100Hz. You also need to compensate in the spreadsheet (into which you imported the FFT data) for the LF loss of the filter. Thus, my 1mHz filter is actually only used down to 10mHz, requiring a mere 28 hour collection time and no equalisation is needed.

[coppercone2]

also, what frequency did you measure those parasitics at? and is it the same one for both filters?

I was thinking of not using a RM5 former and instead winding a torroid. its alot of work but so is making a good box so you might as well. Do you have a material suggestion for it?

The 1mHz filter is interesting but I think its getting to the realm of 'annoying' to measure based on how you described it, its really a lot. 28 hour data log.. someone would need to pay me to try that one. I don't have the patience. 

[EC8010]

Also, what frequency did you measure those parasitics at? And is it the same one for both filters?

Parasitics were not measured at a single frequency but by measuring impedance at 68 different frequencies between 20Hz and 200kHz, then least-squares fitting a model, so those values are valid for the entire working frequency range of the filters.

Winding a toroid will indeed be a lot of work but you should get a very good result.

The requirement for the box is to be electrically conductive with no gaps. Anything that is capable of keeping food fresh is good for gaps. We're coming up to the time of year when shortbread and chocolates appear in fancy tins. They make very good enclosures for electronic testing. Look for tins having a large overlap on their lids. Tobacco tins are good, too. Use BNCs fitted to the tin to make external connections via screened cables. You can improve screening by adding self-adhesive copper tape (as sold to gardeners for dealing with slugs) over the join between lid and tin. A glass fibre pencil will take paint off the outside of tin/lid allowing solder to make a definite connection between tape and tin. Beware that modern tins have an epoxy coating inside, so the lid is insulated and will cause hum. Glass fibre pencil inside and an earth strap between lid and tin solves that problem. The tin is too thin to provide electromagnetic screening <200kHz; you just have to experiment to find a position that suffers least hum.

Little patience is need for a 28 hour noise log - you just set it running and come back much later. What is a problem is ensuring that it doesn't get disturbed.

[Overspeed]

Hello

Why not use a analog filter as in use on the Adret 104 ?

Regards
OS

[iMo]

The above filter from EC8010 is an AC filter, not a "DC-accurate LP filter"..
If you try with the above "Adret 104" patented filter - see below 1Hz cutoff for say 10k i/o impedance.
You may scale it easily - for 0.1Hz cutoff simply multiply both L and C values 10x, and so on..

[coppercone2]

Also, what frequency did you measure those parasitics at? And is it the same one for both filters?

Parasitics were not measured at a single frequency but by measuring impedance at 68 different frequencies between 20Hz and 200kHz, then least-squares fitting a model, so those values are valid for the entire working frequency range of the filters.

Winding a toroid will indeed be a lot of work but you should get a very good result.

The requirement for the box is to be electrically conductive with no gaps. Anything that is capable of keeping food fresh is good for gaps. We're coming up to the time of year when shortbread and chocolates appear in fancy tins. They make very good enclosures for electronic testing. Look for tins having a large overlap on their lids. Tobacco tins are good, too. Use BNCs fitted to the tin to make external connections via screened cables. You can improve screening by adding self-adhesive copper tape (as sold to gardeners for dealing with slugs) over the join between lid and tin. A glass fibre pencil will take paint off the outside of tin/lid allowing solder to make a definite connection between tape and tin. Beware that modern tins have an epoxy coating inside, so the lid is insulated and will cause hum. Glass fibre pencil inside and an earth strap between lid and tin solves that problem. The tin is too thin to provide electromagnetic screening <200kHz; you just have to experiment to find a position that suffers least hum.

Little patience is need for a 28 hour noise log - you just set it running and come back much later. What is a problem is ensuring that it doesn't get disturbed.

Oh no I meant magnetic material.

And IMO, those food tins are almost not good enough for connectors because they are too thin. They buckle when you plug stuff in. For isolated connectors, you can put a backer plate behind it, but if it needs to be RF grounded, that one is a bit harder. I guess you can still put a backing plate with a big hole for the back nut of the connector.

I am good with sheet metal chassis stuff, not so good at magnetic cores


If you wanna take it to the next level, copper plate it, then kool-amp silver plate the interior, then varnish it after masking the ground points.


For altoids tin projects, that have isolated connector (i.e. battery power supply), I glue aluminum in there and then drill it out, its WAY stronger when you plug it in. I caught my BNC connector on a larger shallow tin 'gapping' after awhile because the thin metal got distorted! When the cable was plugged in, sometimes it could get enough force to open a gap. I did not notice anything electrically, but to me it means it needs reinforcement. The wall buckled in and there was a gap between the lid and the interior

[EC8010]

3H1 A100.

Yes, tin plate tins are bendy and a bit marginal on strength, but they screen well. A quick glance around spotted twelve tin plate tins ranging from shortbread to 1 oz. baccy tins fitted with BNCs and small switches; I haven't had problems. I put a large washer behind BNCs and switches on the inside of the tin to spread the stress. I mark the BNC holes with dividers and scriber, drill a 6mm hole (about as big as I dare in something that thin) then file the D-shaped hole with needle files to be a close fit. Having a close fit is key, otherwise the connector can move and cause damage. Altoids and 1 oz. baccy tins are more robust than large cake tins. If you fit XLRs, you definitely need reinforcement behind the connector. I made a spanner that locates on BNC lugs to allow BNCs to be done up tightly using appropriate spanner on nut. If you fold your own box, self-adhesive copper tape over the folds will obscure apertures that let hum and RF in; essential if an electrometer is inside.

[Kleinstein]

An LC type filter could work, but one should not have a resistive load at the output for a DC accurate filter. The damping part would be with a resistor parallel to the inductor and / or an RC series element at the output. For the start and comparison with the other filter circuits one should use a lower order. Already just 1 inductor gives 2nd order.

A problem with an LC fitler is that one needs rather large inductance and capacitance. The kH range would be somewhing like a higher voltage transfromer winding. The magnetic core can also add noise (Barkhausen noise) and pick up hum.
One may consider a supercapacitor, though these have a rather limit voltage range. In combination with a relative low resistance inductor the leakage current may not be that bad. However the loading to the source could be an issue.

[EC8010]

It's not a case of could work; they do work. And very well; those component values are practical. The circuits shown are full circuits including measured parasitics of what was actually built and is used. The original poster wanted to investigate noise in voltage references, and that's exactly what these filters were made for. You don't actually want to pass DC because that limits the gain you can apply to look for noise. Obviously, the polypropylene coupling capacitor sets a low frequency limit, but so does your measurement time - something nobody has explicitly discussed.

Barkhausen noise in an inductor is quite difficult to observe. Conversely, DC references are noisy; 100nV/root Hz is not unusual. Agreed, hum pick-up is always an issue with wound components, but if you capture a spectrum, it's not a problem to remove it in the spreadsheet. With care, the hum is only 20dB or so above the noise floor - at least, that's what I see.

What I have posted has been built, tested, and found to be fit for purpose; it is not a theoretical musing living only in SPICE.

[coppercone2]

I have a family of low-pass filters I use for measuring noise of voltage references. All are simple LC. 600 turns of 0.2mm enamelled copper wire on an RM5 former gets you 36mH when the cores are added; that's the series element. 470nF shunt capacitor determines high frequency cut-off, and its resonance is damped by the 430R series resistor. Series 14uF polypropylene capacitor blocks DC and sets 11mHz high-pass, allowing lots of gain to be added. All the other components are measured parasitics of the real components. You don't have to use the 14uF capacitor, but you can't use lots of gain if you don't. I use the filter in conjunction with a 1M input resistance LNA feeding an oscilloscope set to FFT. The 1kHz high frequency limit allows me to clearly distinguish between white and 1/f noise. Another variant uses 120uF series capacitor (polypropylene, but quite large) for a 1mHz high-pass filter. Both variants have grounding switches after the input coupling capacitor to reduce the wait for stable operation.

what do you think of nanocrystaline as a core material for this application?

[Roehrenonkel]

Hi EC8010,

....... The original poster wanted to investigate noise in voltage references, and that's exactly what these filters were made for.
.......

No, exactly the opposite. I want a DC-accurate Low-pass-filter so i have a clean +10V without the noise.

Ciao4now

[EC8010]

I want a DC-accurate low-pass filter so I have a clean +10V without the noise.

Whoops! My mistake. You can filter high frequency noise away from a voltage source/reference but (by definition) you can't filter the lowest frequencies; you need a source that is itself low noise. I've used my filters in the search for quiet voltage references.

No idea what nanocrystalline core material would be like, I'm afraid. But I had some 36mH inductors wound on RM5 cores, so I used them. The 100Hz low-pass filter used some large inductors made by UTC. My measurement filter really isn't critical, so almost anything will do the job.

[coppercone2]

Well I experimented how to get 36 mH on my nanocrystaline core.


It looks like about 18 turns of wire is needed on a 60uH Al core to get 36mH at 120Hz and 1KHz. Drops at 10KHz and you get 4mH at 100KHz. It's not what I expected but I made sure the measurement is repeatable and verified with similar value manufactured components (I got a large amount of different shielded axial through hole chokes for a pretty penny).

I think that is alot nicer then trying to do 600 turns ! Especially because its a torroidal core.

I am thinking to follow up with a different filter at say 100KHz to supplement the NC core that drops off there.



Also, with fine gauge manganin wire, it looks like that winding should give me between 400 to 600 ohms of resistance if wound on the same coil, making it possible to combine the resistor and the inductor on the torroidal core, but working with the sub 1/20th of an mm manganin is a little frustrating (first time), it is easy to kink that tiny wire. My plan is to use dots of UV adhesive to kinda attach it to the core while working on the torroid.

Comes out to like 0.7 meters of wire to get the correct turn count. I used the DE5000 meter, but I can follow up on a HP 1KHz mohm meter possibly, it does measure L, but not C.

The manganin should be +-10ppm, making it comparable to a pretty good resistor.

What do you think?

And any thoughts on a second filter stage to follow up the first?


For the construction I plan to wind the wire on the core using dots of UV adhesive, then mount it to a PCB by lashing the core to some holes using nylon lace over a pool of epoxy that is then drizzled with more epoxy to make a robust connection between the plastic core and a garolite board, and terminate the windings to turret terminals (2 stage).

[Kleinstein]

I would prefer the inductor to be low resistance and thus use copper wire. This way one can adjust the damping if needed and also consider a parallel resistor in stead of the series one. For DC accuracy low series resistance would help and allow more leaky capacitors, possibly even electrolytic.

I would not worry so much about the higher frequencies like 10 KHz or 100KHz. There the filter would still attenuate enough from the capacitor even if the inductance goes down somewhat.

For a low pass filter to suppress ref. noise one may want a really low cross over and may want more than only 36 mH. AFAIR some 100 mH could be available as ready made CM chokes and using both coils in series could give an extra factor of 4. They are often just ferrite, but could still be wirth a first try.

[coppercone2]

Lol my first attempt at thinking it was linear the NC torroid got 130mH from like 25 turns or something

these cores can give you absurd inductance with not too many turns. I imagine getting it into the Henry range would not be very time consuming

I imagine what would happen if you really fill up one of the larger torroids, it might become a 'magnetic blocker' more then anything else lol

[coppercone2]

I really want the built in resistor though. If its too much you can put a splice to make a hybrid coil with two wires, if its too small then there seems to be no good way to get rid of the resistor.

[EC8010]

It looks like about 18 turns of wire is needed on a 60uH Al core to get 36mH at 120Hz and 1KHz. What do you think?

And any thoughts on a second filter stage to follow up the first?

That's very few turns, so the effective permeability of the core must be enormous. Of course, a toroid isn't gapped, so that will help. However... The nice thing about an air gap is that it stabilises the inductance. I would expect the (gapped) RM5 core to have a much more stable value of inductance than an ungapped toroid. 600 turns isn't nearly as much of a problem as you'd think provided you have a coil winder with a counter. Such things are now cheaply available. What might surprise you is that the winder's crucial quality is smooth rotation - it's jerkiness that breaks fine wire. My cheap and cheerful winder had an awful rotating plastic sleeve as a handle, so I made a closer fitting wooden rotating sleeve that allows smoother rotation.

I've never felt any need for a second stage of analogue filtering. When measuring noise, an accuracy of +/-1dB is perfectly acceptable, so provided your sample frequency is 5x the filter's cut-off, its 12dB/octave slope is entirely adequate to keep aliases under control.

[coppercone2]

you could get a gapped core that will probobly have a large inductance too. the distributed gapped version (not with a cut, looks solid)

I think I saw NC Distributed gap cores somewhere. The gap is leakage I think, so it also has a downside

Or the powder core is kind of  gapped. I see a few exotic powder cores.


Anyway, its been years since I put anything into the thermal chamber, maybe I can put this in there.


But, are you sure that it still needs a gap, because the core I believe is wound, so I think it ends up being kind of gapped?

https://www.transmart.net/a-news-effect-of-temperature-on-nanocrystalline-core-performance




Nanocrystalline cores are known for their good thermal stability compared to conventional ferromagnetic materials. This is primarily due to their fine grain structure, which limits the movement of domain walls and reduces the overall eddy current and hysteresis losses. However, even nanocrystalline cores have a limit to their thermal stability, beyond which their magnetic properties can significantly degrade. It is important for designers and engineers to consider the thermal stability of nanocrystalline cores when selecting materials for specific applications to ensure reliable and efficient performance over the expected operating temperature range.

[coppercone2]

I have another option too, a 39mH shielded axial inductor. Its not a torroid but I don't have to wind anything. Perhaps its good enough. This one is ferrite.

I wonder how it compares to your R5 core. And I could give it a wrap in iron, hy-mu or NC foil.

From my not so junk box

16uF polyproyplene foil axial 10% 16.1
470 ohm ceramic composition 10% 477.5
39mH shielded axial ferrite inductor 10% 40.1
470nF C0G ceramic 5% - measures 46539

I wonder how it will do. It looks to be over dampened because the inductor is 120 ohm and the resistor is like 460.



The RC HP calculation is 9.88mHz

[coppercone2]

What exactly will the result of having a over dampened filter be ?

In a scope your rising edge would be like slow. If you look at the average under the wave, I guess you lose data, for a square wave.

how does the error look like in measurement for noise in real life?


If you put a giant value like 5k, the bode plot looks like it basically is a much lower pass filter

Did you ever quantify the amount of error you get from over/under dampened filters?


I never thought about the actual amount of error you can get from a over or under dampened RLC filter in a RMS measurement before.

THe other part of it seems also to be a voltage divider between Rseries and the 1M

Would you bother tuning that series resistor? I kind of wonder how it would be if you put a potentiometer there. Like a 1K pot

[Kleinstein]

An underdampend filter will show a resonance and thus peaking at the transition. An overdampend filter will be more smooth, a bit more like 2 x 1st order filters and not a proper 2nd order filter. There is anyway no perfect value for the damping part. There are a few classic values (Tchebychev, Bessel, Butterworth and Linkwitz) for the damping depending on what aspect one is aiming for. Damping values in between are possible too.
One may want to tune the damping, if one needs a special phase response, especially linear phase, which acts like a delay. Other uses are likely less critical and a resistor from the E24 series would be good enough.

For the LF noise measurement (the classic 0.1 to 10 Hz band) the more logical way to make the low pass filter for the 10 Hz limit is anyway some form of digial filtering. An active 10 Hz RC based filter is also not problem and the filter in not so noise critical as it can be behind quite some amplification. The question may be if one could implement the input AC coupling (high pass) as an LC filter to reduce the noise there. This however would be at 0.1 Hz or lower, so not so easy. There is also the good alternative to have the AC coupling at the input at a frequency lower than 0.1 Hz and have a separate high pass filter for the 0.1 Hz later - possibly also in the digital domain. This avoids the RC filter noise in the transition region from the often relatively large resistor.  For a LF noise measurement there is little use for an LC based fitler.

As a ready made high high value inductor with relatively low resistance one could consider a CM choke:  100 mH are available of the shelf and the 2 coils in series would give 4 x the inductance. A closed core is a big plus, as it gives high inductance (though early saturation) and it also keeps the hum pick-up small.
For a DC accurate LC filter I would avoid intentional resistive loading to ground. So only the capacitor leakage to compete with the coil resistance.

[coppercone2]

but what does it end up looking like for RMS measurement? Usually noise is RMS

[EC8010]

It all depends on how you measure your noise. The traditional method used quite steep high-pass and low-pass filters to define bandwidth, then used a meter preceded by an RMS rectifier to measure the total noise. The implicit assumption being made was that the noise was white (unchanging voltage with frequency).

But we can do much better now. A much more useful measurement is made via the FFT of an oscilloscope. The FFT breaks the frequency range into bins, each defined by a mathematical high-pass and low-pass filter and assumes that the noise between the filters is white. If the bins are narrow (perhaps only 1Hz wide), then the assumption of unchanging amplitude with frequency within the measurement bandwidth remains appropriate even for 1/f noise (-10dB/decade slope). Thus, the FFT is the ideal way to characterise noise.

But! Oscilloscopes rarely contain an anti-alias filter. Worse, with 20MHz of minimum bandwidth, they see a lot of noise, whereas we are often only concerned with 10kHz and below. So my filters are designed to remove the high frequency noise we're not interested in so that we can wind in oscilloscope gain and use the FFT to measure what we are interested in without overloading the front end.

As pointed out, although an LC filter has a slope of 40dB/decade, it is has a resonance before entering that final slope that must be correctly damped. We want Q = 0.7 for a maximally flat response. And that's the only critical thing, but it's easily set by adding the right resistance to the inductor. That's why my filters have an additional series resistance.

What isn't immediately obvious is that the inductor has shunt capacitance and your deliberate capacitor has series inductance. Both of those parasitic components mean that the 40dB/decade slope does not continue for ever. Fortunately, you have to try quite hard to make a 36mH inductor that has enough shunt capacitance to be a problem. Likewise, you'd need extremely long wires on your deliberate capacitor to cause a problem. If you do a SPICE analysis of my circuits (which include the measured parasitic compoonents) you'll see that they're (deliberately) quite well behaved.

Once we've added our input filter, we no longer need to worry about overloading the oscilloscope. But oscilloscopes sell by bandwidth and small low capacitance devices have plenty of 1/f noise. So we need a low noise pre-amplifier having at least x100 gain to ensure that any 1/f noise we measure is actually from the device we're trying to measure, not the oscilloscope itself. I use an OPA1641 configured as a non-inverting amplifier having A = 100 after the LC filter. That does for most work, but it is perfectly possible to do better using discrete FETs connected in parallel and driving an NE5532A to get noise down to 1nV/root Hz.

Finally, once we've got an FFT on the oscilloscope, it is ragged. We need to average FFTs together to smooth the noise spectrum. Averaging FFTs together seems to be reserved for the more expensive oscilloscopes. If you average 100 FFTs together, the deviations from a fitted curve will be about 0.6dB, which is why I earlier commented about 1dB for noise. Having averaged 100 FFTs together, we export the FFT data as a csv file to a spreadsheet and plot it again (logarithmic frequency axis). Having the raw data available, we can now fit a line to the white noise and characterise that. We can fit a -10dB/decade slope to the 1/f noise and characterise that. Usually, I find that there's also -20dB/decade noise due to thermal dependencies, so that's also needed. It all sounds complicated (and it is), but having fitted in this way and defined what the noise is, you can try reducing it and see whether you have actually made a difference. The -20dB/decade thermal dependency noise is the easiest to tackle and is usually reduced by adding thermal insulation around semiconductors and shielding them from draughts (just like the voltage reference data sheets tell you to do).

As Lord Kelvin observed, once you have some proper numbers, you have proper knowledge and you can advance. The old-fashioned single figure noise measurements aren't really much use. Here's an example of what I measured via one of my filters:

[EC8010]

@coppercone2: Your junk box bits should do just fine, just set Q = 0.7.