Author Topic: Some old school instruments showing how it's done (HP 3325A and Fluke 8506a)  (Read 59729 times)

0 Members and 1 Guest are viewing this topic.

Offline joeqsmith

  • Super Contributor
  • ***
  • Posts: 12277
  • Country: us
Here I tried not to disrupt the test and let it set at one level for a little over 2.5 hours.   The average is 9.9999929 volts DC.   Histogram is again looking at the two LSDs. 

90 - 1 count
91 - 13 counts
92 - 273 counts
93 - 1439 counts
94 - 0 counts
95 - 86 counts
96 - 2 counts

Without any added noise (in my case added by my changing the loading resistance), I will see buckets without any counts.   

***
Wanted to show the zoomed out histogram. 
***
I have attached the data in ASCII format if you want to have a look.
« Last Edit: January 28, 2021, 09:14:00 pm by joeqsmith »
 

Offline dietert1

  • Super Contributor
  • ***
  • Posts: 2473
  • Country: br
    • CADT Homepage
Since you have about 1800 entries after 9000 seconds, does this mean a Fluke 8506A needs 5 seconds for the 1 uV resolution measurement?
With our 8502As i am currently looking at averages of 20 samples and they need 13.3 seconds to take them. To get a 10 sample average it would take 6.6 seconds, very similar.

Regards, Dieter
 

Offline joeqsmith

  • Super Contributor
  • ***
  • Posts: 12277
  • Country: us
Since you have about 1800 entries after 9000 seconds, does this mean a Fluke 8506A needs 5 seconds for the 1 uV resolution measurement?
With our 8502As i am currently looking at averages of 20 samples and they need 13.3 seconds to take them. To get a 10 sample average it would take 6.6 seconds, very similar.

Regards, Dieter

I haven't looked into it.  I empirically came up with the five seconds.  When I put the software together, if I triggered the meter then request the data too quickly, the meter doesn't have time to respond.  Four must be on the edge as it will miss every so often, so I padded it out to five. 

The manual shows Average selects S10 (H2) or 2^10 or 1024 samples/reading which they show requires 4.3 seconds.   The filter  is set to slow with the timeout disabled.  Not sure if this has any effect.    It would be very easy to measure the time in software. 

One thing I did try was just sticking a cap across the meter, charging it to 10V and then allowing it to discharge.   I would expect if I monitor the LSD, it should be a fairly even spread.   Shown is about 2.6 hours of data. 




Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
Here I tried not to disrupt the test and let it set at one level for a little over 2.5 hours.   The average is 9.9999929 volts DC.   Histogram is again looking at the two LSDs. 

90 - 1 count
91 - 13 counts
92 - 273 counts
93 - 1439 counts
94 - 0 counts
95 - 86 counts
96 - 2 counts
[...]


Looking back at some of my notes,  I studied the section in the manual that explains how the A/D converter works a while back, and decided that it is a 21 bit converter, which is just enough for a 6.5 digit display of a 20V range (positive or negative - the converter can flip polarity and uses 21 bits on either side).

The A/D converter therefore has 2^21 = 2,097,152 discrete steps.

So the converter divides the 7 volt reference over 2^21 steps, meaning each LSB =  3.337860107421875uV.  But the 7V readings have to be scaled up to the 20V range, around a factor three. (Done mathematically/perfectly by the CPU.) Now, each step is pretty much spot on 10uV, which makes sense with the 6.5 digit capability.

The 7.5 digit mode generates an extra digit by taking an average of a number of readings.  It seems intuitive that the number of readings need to be 10, in order to generate an extra digit.

So if the meter is running relatively stable, and the A/D converter is only flipping between two different readings at the 10 uV level occasionally, perhaps it is only natural that the average of those readings end up having missing values?


[Edit:] In other words:  the readings are not noisy enough to give a good average?



« Last Edit: January 29, 2021, 12:44:44 am by SilverSolder »
 

Online bdunham7

  • Super Contributor
  • ***
  • Posts: 8175
  • Country: us
So if that theory is right, I will get my missing numbers if I increase the sample size by enough?
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
So if that theory is right, I will get my missing numbers if I increase the sample size by enough?

If you have everything settled so it moves less than 10uV, the averaging algo won't have anything to work with. 

This is a limitation of the fact that the meter is "better" than its A/D converter can resolve... 

So I suspect running very long sample periods may help, but may not be enough to get the distribution of values to become a clean bell curve.

It's almost like the A/D converter's own noise needs to be 5uV in order for the LSB to randomly "trip" to either side of the real value, if the real value is exactly between two 10uV steps, if you see what I mean.

In other words, if your measured voltage is stably parked between two steps of the A/D converter, so it is 'stuck' on one value, no amount of averaging can rescue that situation...

Adding noise to the voltage would cause the A/D to trip to both sides of the measured voltage - and it will trip more often on the side it is closer to, so averaging can find the value we are looking for.

Obviously there is going to be some noise in the circuit that will lead to this effect naturally...  but there may not be enough noise, is the problem!
 

Offline dietert1

  • Super Contributor
  • ***
  • Posts: 2473
  • Country: br
    • CADT Homepage
Yesterday night i had some time to fix the guard problem in one of our 8502As. My mistake - while replacing lots of caps i forgot to cut short the wires of one of them and it touched the guard inside the A/D module. Lucky enough it was the autozero memory, so nothing else happened.
Now i have two diagrams to show how i am using the DVMs. First one shows a log of this night with a drift of about 1.5 ppm. The reference and the repaired DVM got turned on at about 23:00. Vertical raster is 1 ppm. Each DVM has "sticky" levels even after averaging, so effective p2p noise is less than 1 ppm. The levels aren't the same for both instruments, because i am recalibrating the measurements in software, such that both DVMs report similar values. When i average both instruments, the result looks like a 7 digit DVM, with a noise level at about 1 or 2 uV at those time scales.
After the 8:00 bump the measurement will continue during the day without further drift or bumps. During working hours the lab will be at roughly constant temperature.

Regards, Dieter
 

Offline joeqsmith

  • Super Contributor
  • ***
  • Posts: 12277
  • Country: us
When I was reading the manual, I thought they stated that the fast R^2 converter was 24 bits including the sign. 

With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.   

Online bdunham7

  • Super Contributor
  • ***
  • Posts: 8175
  • Country: us
When I was reading the manual, I thought they stated that the fast R^2 converter was 24 bits including the sign. 

With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.

I also remember believing it was 23 bits + sign, although I can't find it at the moment.  That would still give 8,388,608 possible values to cover a display that shows 20,000,000 counts.  That is a little less than 2.4 displayable counts per possible binary output of the A2D converter.  Some of those displayable values will be close to a binary value and some will not.  If you add in a slight misalignment here or there and a relatively small sample size being averaged, it seems plausible that some values are more likely than others--especially if, as SilverSolder has proposed, the meter settles at one or two binary values.

If this is how it works, then you would expect to see a more even distribution over a range like you are testing with the capacitor.  No one particular displayable LSD is more likely over the whole range, even if specific full counts are less likely or impossible.  I set my sample size to 212 and let it run, but the reading has drifted down 7uV overnight so I'll have to just keep watching.
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
When I was reading the manual, I thought they stated that the fast R^2 converter was 24 bits including the sign. 

With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.

You would expect the bins to look balanced through a sweep -  if not, there is definitely an alignment issue! -  this test is really quite good, I'll try it and see what happens here...

I believe the A/D converter works like this:   The converter has 5 real bits which are used several times to "nibble away" at the voltage being measured, in 5 repeated measurements, where the remainder between each of the five readings is stored as an analog voltage on a capacitor.  This is the "recirculating remainder" that they are talking about.

You would think 5 readings of 5 bits add up to 25 bits, but in fact only the first reading is a 5 bit reading -  the next 4 readings all overlap the previous 5 bit "nibble" by 1 bit, so they are effectively 4 bit readings.  This is done right at the analog level, by the amplification chosen for the op amps that handle the remainder and feed it back.   The reason they do that is to reduce noise, (EDIT: according to the manual actually I can't remember where I read that, or if I just thought that must be the reason).  So, effectively, we get 5 bits plus 4x4 bit readings = 21 "real" bits for a complete cycle.

The sign is handled by a separate circuit that flips the voltage reference to the opposite polarity.  I suppose we can call that one extra bit, but the sign is a completely separate business from the recirculating remainder processing.  The processing is equally happy working with positive or negative voltages, as long as the reference is flipped!

When you look at how this thing actually works, you can't help being just a little bit awestruck that they actually managed to get this principle to work at all, let alone be accurate, since it is actually all being done by analog electronic circuits that are just "guided" by the CPU!  :D

[Edit:] Yes, the manual does say the result is "assembled into a 24 bit word describing the polarity and magnitude" - but that doesn't actually say how many of the bits are significant!  :D
« Last Edit: January 29, 2021, 04:36:34 pm by SilverSolder »
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
Yesterday night i had some time to fix the guard problem in one of our 8502As. My mistake - while replacing lots of caps i forgot to cut short the wires of one of them and it touched the guard inside the A/D module. Lucky enough it was the autozero memory, so nothing else happened.

LOL that's almost the same problem I made for myself by using a too-long screw to put the bottom chassis cover on.  Lesson(s) learned!  :D

Now i have two diagrams to show how i am using the DVMs. First one shows a log of this night with a drift of about 1.5 ppm. The reference and the repaired DVM got turned on at about 23:00. Vertical raster is 1 ppm. Each DVM has "sticky" levels even after averaging, so effective p2p noise is less than 1 ppm. The levels aren't the same for both instruments, because i am recalibrating the measurements in software, such that both DVMs report similar values. When i average both instruments, the result looks like a 7 digit DVM, with a noise level at about 1 or 2 uV at those time scales.
After the 8:00 bump the measurement will continue during the day without further drift or bumps. During working hours the lab will be at roughly constant temperature.

Regards, Dieter

That is very neat!

I have attempted a similar idea with 4 meters, but found that the difference between the "best" and "worst" performer was too great - the worst performing unit didn't really add any value to the best ones! 

So, if nothing else, the exercise helped me identify which meter to work on next!  :D

 

Offline dietert1

  • Super Contributor
  • ***
  • Posts: 2473
  • Country: br
    • CADT Homepage
..
When you look at how this thing actually works, you can't help being just a little bit awestruck that they actually managed to get this principle to work at all, let alone be accurate, since it is actually all being done by analog electronic circuits that are just "guided" by the CPU!  :D
...

Yes a true miracle. Those high resolution DVMs from HP mainly used time division to realize the DAC, with even less bits. But i remember that for high resolution digital audio there are multi-bit sigma-delta DACs, that should be a mixture of both methods. So it's still around. Also the MSP430 MCUs that we use for our oximeters include segmented DACs: One 7 bit ladder and another 5 bit ladder to be muxed similar to a Kelvin-Varley Divider. The MCU has two of those and they can be software tuned for 19 bits accuracy. A stone remains a stone, independent of what marketing names it.

Regards, Dieter

 
The following users thanked this post: SilverSolder

Offline guenthert

  • Frequent Contributor
  • **
  • Posts: 771
  • Country: de
[..]
With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.

     I haven't read the specifications of the 8505/6A, but generally those long(ish) scale DMMs specify an input resistance of >10GOhm (in some ranges, which should include 10VDC).  That's only a nominal U/I ratio, not to be understood as plain resistor.  There will be some bias current of the FETs at the input leaking out and the AZ circuit's capacitance will want to be fed, etc. .   At 10V you should expect an average current of less than 1nA (it might be less than 100pA, but peak values might very well be considerably higher), hence the nominal >10GOhm.

    The bias current means, that a low-leakage capacitance on the input will actually get charged (not necessarily towards +10V), which can be used to determine the average current.
« Last Edit: January 29, 2021, 06:10:28 pm by guenthert »
 

Online bdunham7

  • Super Contributor
  • ***
  • Posts: 8175
  • Country: us
    The bias current means, that a low-leakage capacitance on the input will actually get charged (not necessarily towards +10V), which can be used to determine the average current.

All true, but the numbers on the 8505/06A are lower.  If you tweak it carefully, you can get bias currents below 1pA, specifications are <5pA adjusted, <30pA @ 1 year.  So the effect on a capacitor of anything more than a few hundred pF will be pretty minimal on a properly adjusted meter of this type.  I've no idea whether or how the U/I ratio changes in operation during measuring or over time, that would be another interesting question.
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
[..]
With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.

     I haven't read the specifications of the 8505/6A, but generally those long(ish) scale DMMs specify an input resistance of >10GOhm (in some ranges, which should include 10VDC).  That's only a nominal U/I ratio, not to be understood as plain resistor.  There will be some bias current of the FETs at the input leaking out and the AZ circuit's capacitance will want to be fed, etc. .   At 10V you should expect an average current of less than 1nA (it might be less than 100pA, but peak values might very well be considerably higher), hence the nominal >10GOhm.

    The bias current means, that a low-leakage capacitance on the input will actually get charged (not necessarily towards +10V), which can be used to determine the average current.

The specification is max 5pA bias current.  I measured mine to around 3pA with an electrometer.  I've tried to look at the circuit diagram to figure out how they get it so low, but I'm not understanding it...  analog black magic overload with stacked dual FETs etc.


 

Online bdunham7

  • Super Contributor
  • ***
  • Posts: 8175
  • Country: us
So since my bench is now occupied with large Fluke meters anyway, I went out and retrieved the old 8500A I had purchased for parts in case I needed any for the 8505A.  Of course almost no parts are compatible and I didn't need any anyway--and the 8500A worked perfectly, so I just left it alone.  At one point I adjusted Ibias and input zero so that I could use it as a null in place of an electrometer that I don't have.

So after a warmup, it seems to work just fine and since we were discussing input impedance and capacitors, I set up a test.  I put a 0.47uF mylar cap on the inputs and then put a 10V source on it, then removed the source.  Two hours later, the measured voltage has just dropped below 8.8V, giving an apparent input impedance of about 150G. 
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
How do you actually adjust the input zero and bias - is there actually an adjustment, or does it mean changing fixed components?

The 8502A, 8505A, and 8506A seem to share a number of compatible parts.
« Last Edit: January 29, 2021, 11:03:10 pm by SilverSolder »
 

Online bdunham7

  • Super Contributor
  • ***
  • Posts: 8175
  • Country: us
How do you actually adjust the input zero and bias - is there actually an adjustment, or does it mean changing fixed components?

The 8502A, 8505A, and 8506A seem to share a number of compatible parts.

Under the cover of the DC Signal Conditioner module there are two trimpots.  You set it to 100mV range and connect a shorting plug and adjust the zero. Then you put a 1M resistor + small cap (I think I'm using 2.2nF) across it and adjust the Ibias trimpot to zero.  At 1M, 1uV implies 1pA bias current.  It's quite fiddly, especially the bias.  If you sneeze it will display HHH.HHHH H.

Quote
The 8502A, 8505A, and 8506A seem to share a number of compatible parts.

AFAIK, the 8505A and 8506A are essentially the same, with the mainframe and most of the modules sharing almost all components except some special internal connections for AC and of course, the front panel lettering and indications.  I haven't actually seen an 8506A, they seem to be valued because of the novel thermal converter.  Personally I'm glad to have the 8505A because it can do AC amps, you don't have to choose between ohms and current converters and the AC TRMS converter actually works pretty well--accurate down to near zero % of range and reasonably accurate to 1MHz.

The 8502A I'm not sure about.  The 8500A has almost no compatible large parts.  Of course, I'm sure many of the individual board components are the same, but the most of the modules are different p/n, the power supply is not the same and won't fit, the module retention clips (the ones that break 100% of the time) aren't used and even the case top is not ventilated.  The 8500A actually has a very nice dual-bobbin transformer that I was hoping to use to quiet my humming 8505A, but that's also a no-go.  The 8502A looks a lot more like an 8500A with the additional buttons than anything else.
« Last Edit: January 30, 2021, 12:29:25 am by bdunham7 »
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline joeqsmith

  • Super Contributor
  • ***
  • Posts: 12277
  • Country: us
After 24 hours, the capacitor discharged to 6.8V.  LSD is basically flat.   


[..]
With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.

     I haven't read the specifications of the 8505/6A, but generally those long(ish) scale DMMs specify an input resistance of >10GOhm (in some ranges, which should include 10VDC).  That's only a nominal U/I ratio, not to be understood as plain resistor.  There will be some bias current of the FETs at the input leaking out and the AZ circuit's capacitance will want to be fed, etc. .   At 10V you should expect an average current of less than 1nA (it might be less than 100pA, but peak values might very well be considerably higher), hence the nominal >10GOhm.

    The bias current means, that a low-leakage capacitance on the input will actually get charged (not necessarily towards +10V), which can be used to determine the average current.

The specification is max 5pA bias current.  I measured mine to around 3pA with an electrometer.  I've tried to look at the circuit diagram to figure out how they get it so low, but I'm not understanding it...  analog black magic overload with stacked dual FETs etc.

I used a new 9V transistor battery in series to try and get the noise low enough to make some sort of ballparkish measurement.   With my old electrometer placed in the lowest range (20pA), standing back and let it settle looks like around 1.5pA.    Any sudden movements, over 2pA.  Walk away and don't twitch, 1.33pA.   I can believe it's <5pA. 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
[...]
The 8502A I'm not sure about.  The 8500A has almost no compatible large parts.  Of course, I'm sure many of the individual board components are the same, but the most of the modules are different p/n, the power supply is not the same and won't fit, the module retention clips (the ones that break 100% of the time) aren't used and even the case top is not ventilated.  The 8500A actually has a very nice dual-bobbin transformer that I was hoping to use to quiet my humming 8505A, but that's also a no-go.  The 8502A looks a lot more like an 8500A with the additional buttons than anything else.

I have successfully swapped modules between 8502A, 8505A, and 8506A.   Of course, just because the modules "work" doesn't mean the parts can't be made to different tolerances etc. -  I haven't tested them that thoroughly (yet!... now that I had the idea...)  -  but some cards, like the GPIB card etc., should be a no-brainer.

The transformers hum loudly in every single unit I've ever laid hands on - it is safe to say this is a "feature", not a bug...   I guess it might be fixable by rubber mounting it - I have a theory that it is perhaps  too stiffly coupled with the chassis, mechanically? - but the hum is hardly the worst noise offender in the old test gear collection...  :D
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
After 24 hours, the capacitor discharged to 6.8V.  LSD is basically flat.   


[..]
With that 10Gohm input impedance my little cap is not discharging very fast.  After 15 hours its down to 7.4 volts.  The LSD bins continue to look fairly even.

     I haven't read the specifications of the 8505/6A, but generally those long(ish) scale DMMs specify an input resistance of >10GOhm (in some ranges, which should include 10VDC).  That's only a nominal U/I ratio, not to be understood as plain resistor.  There will be some bias current of the FETs at the input leaking out and the AZ circuit's capacitance will want to be fed, etc. .   At 10V you should expect an average current of less than 1nA (it might be less than 100pA, but peak values might very well be considerably higher), hence the nominal >10GOhm.

    The bias current means, that a low-leakage capacitance on the input will actually get charged (not necessarily towards +10V), which can be used to determine the average current.

The specification is max 5pA bias current.  I measured mine to around 3pA with an electrometer.  I've tried to look at the circuit diagram to figure out how they get it so low, but I'm not understanding it...  analog black magic overload with stacked dual FETs etc.

I used a new 9V transistor battery in series to try and get the noise low enough to make some sort of ballparkish measurement.   With my old electrometer placed in the lowest range (20pA), standing back and let it settle looks like around 1.5pA.    Any sudden movements, over 2pA.  Walk away and don't twitch, 1.33pA.   I can believe it's <5pA.

I also used a 614 electrometer, and had the same user experience of having to look at it from across the room to avoid disturbing the measurement!  :D

What is the principle behind putting a battery in the circuit to help noise?
 

Offline joeqsmith

  • Super Contributor
  • ***
  • Posts: 12277
  • Country: us
I was just trying to get a somewhat close measurement.  A battery is a fairly low noise source and I could place it inside the metal enclosure to minimize the lead lengths.   

Offline guenthert

  • Frequent Contributor
  • **
  • Posts: 771
  • Country: de
    Don't know about the 614, but the 617 would be too slow to reliably catch the peaks of the input bias current of a DMM.  The 617 has an analog output (amplifier output, not just a synthesized analog signal, as e.g. in the Keithley 181 nanovoltmeter) which could be used to connect a oscilloscope.
 

Online bdunham7

  • Super Contributor
  • ***
  • Posts: 8175
  • Country: us
    Don't know about the 614, but the 617 would be too slow to reliably catch the peaks of the input bias current of a DMM.  The 617 has an analog output (amplifier output, not just a synthesized analog signal, as e.g. in the Keithley 181 nanovoltmeter) which could be used to connect a oscilloscope.

If it doesn't use autozero, why would there be peaks?  The only meters I've seen that have issues in that area (and can cause issues when connected in parallel with other meters) are the ones with continuous autozero.  My picoammeter is also too slow to catch anything but the slowest of peaks, so if there is some other mechanism that caused them I'd be interested to know what it is.
A 3.5 digit 4.5 digit 5 digit 5.5 digit 6.5 digit 7.5 digit DMM is good enough for most people.
 

Offline SilverSolderTopic starter

  • Super Contributor
  • ***
  • Posts: 6126
  • Country: 00
    Don't know about the 614, but the 617 would be too slow to reliably catch the peaks of the input bias current of a DMM.  The 617 has an analog output (amplifier output, not just a synthesized analog signal, as e.g. in the Keithley 181 nanovoltmeter) which could be used to connect a oscilloscope.

614 does have a "real" analog output as well, but I didn't think to trace it on the scope to look for peaks.  Another interesting project...

 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf