Keithley DMM7510, SMU 2450 & 2460 problems

[niner_007]

Here is an interesting plot of Dr.diesel's data measuring 731B 10V voltage reference using a custom AZ script that I wrote (lownoise.txt). I compare it to the data that I took earlier using a standard 10PLC measurement of a 10V battery. 
https://www.eevblog.com/forum/metrology/dmm-adc-noise-comparison-testing-project/msg1094322/#msg1094322

One can see that in the time range of 10 to 100 sec the custom AZ script removes the noise bump in the Allan variance. At longer times both measurements are affected by the drift of the voltage references in the meters or by the voltages sources themselves.

This confirms that when measuring a finite voltage the autozero problem of DMM7510 is significant and limits the performance on a time scale of 10 to 100 sec.

They have not attempted to fix this problem with the newest firmware release.  To me this problem was obvious within 5 minutes of getting the meter (at first I thought it was defective until I found data from others posted by TiN). Even though the meter meets the specs since there are no specs for longer than 5 PLC, the time domain data look crappy.  I have already spent a lot of time on corroborating and diagnosing this problem, but at this point it's probably better to just stick to Keysight meters.

if it was designed that way deliberatly it's not a bug, and it can't be considered noise; from all the information available, it seems like they optimized it specifically at 5PLC; also, since the script removes the bump I don't see how it is a problem, or just use it at 5PLC; what am I missing? it seems like this whole thing was blown out of proportions (PS sorry in advance for being a noob)

[niner_007]

How does a comparable chart look like from a Keysight meter?

I posted earlier a comparison for Keysight 34465A. For finite voltages it does not look great, but I think its because of the voltage reference. I need to get a 34470A for a fair comparison. Or perhaps someone else here has a 34470A and a decent voltage reference.

I have a 34470A, tell me what to do and I'd love to do it

[Kleinstein]

There are already similar curves for the Keysight meters earlier in this thread. The suitable test would be to collect data for a reasonable long time (e.g. 1-3 hours) under stable conditions. Ideally with a very stable source (e.g. LTZ1000 or better) in 10 V range, but the extra noise is also visible with just a short.

The Alan variance curve is kind of related to the frequency spectrum of the noise. The curve from the custom AZ script is about the normal shape expected. In some metrology applications the minium value reached is an important quality criterion, especially if this is not at too long a time scale. That 10-100 seconds time scale is just the one needed for doing test with dividers or similar. The custom AZ script can do better in this range, but as one can see it is rather noisy on the short time scale, and an internal solution if done right could likely do considerably (about a factor of 1.5-4) better. So the extra script is kind of a work around but not a full solution.

Just using the 5 PLC mode does not solve the problem. Once you average those data taken at 5 PLC to get even lower noise that extra noise comes back and noise even goes up when averaging more than about 100 samples. This might also lead to wrong conclusions if the normal statistics is used to estimate noise. One would normally assume the data are not correlated over more than maybe the next neighbors - but with the Keithley meters (to some extend a similar bump is even visible with the K2002 and may very well effect more meters) there seem to be some correlation up to 100 seconds.

From the analysis done so fat in this thread, it looks like Keithley is using some averaging on the zero readings when doing AZ, even if digital filtering is turned off. This sounds like a good idea at first sight and could reduce the short time scale (high frequency) noise a little. However it comes at the price of not so good suppression of 1/f noise of the ADC/amplifier. So generally this is not a good idea and it would be really nice if Keithley would also offer the normal simple difference AZ mode.

[niner_007]

There are already similar curves for the Keysight meters earlier in this thread. The suitable test would be to collect data for a reasonable long time (e.g. 1-3 hours) under stable conditions. Ideally with a very stable source (e.g. LTZ1000 or better) in 10 V range, but the extra noise is also visible with just a short.

The Alan variance curve is kind of related to the frequency spectrum of the noise. The curve from the custom AZ script is about the normal shape expected. In some metrology applications the minium value reached is an important quality criterion, especially if this is not at too long a time scale. That 10-100 seconds time scale is just the one needed for doing test with dividers or similar. The custom AZ script can do better in this range, but as one can see it is rather noisy on the short time scale, and an internal solution if done right could likely do considerably (about a factor of 1.5-4) better. So the extra script is kind of a work around but not a full solution.

Just using the 5 PLC mode does not solve the problem. Once you average those data taken at 5 PLC to get even lower noise that extra noise comes back and noise even goes up when averaging more than about 100 samples. This might also lead to wrong conclusions if the normal statistics is used to estimate noise. One would normally assume the data are not correlated over more than maybe the next neighbors - but with the Keithley meters (to some extend a similar bump is even visible with the K2002 and may very well effect more meters) there seem to be some correlation up to 100 seconds.

From the analysis done so fat in this thread, it looks like Keithley is using some averaging on the zero readings when doing AZ, even if digital filtering is turned off. This sounds like a good idea at first sight and could reduce the short time scale (high frequency) noise a little. However it comes at the price of not so good suppression of 1/f noise of the ADC/amplifier. So generally this is not a good idea and it would be really nice if Keithley would also offer the normal simple difference AZ mode.

Thank you for explaining it. Is the 10 - 100 seconds here referring to the integration time, or the overall time at which the sample was taken? If the latter (unlikely), couldn't one just wait 100 seconds?

[Kleinstein]

The 10-100 seconds time scale where additional noise appears refers to the timeframe used for comparing data or the length of record to calculate noise. The Allan deviation is averaging data over a given time and than compared adjacent averages. So it is a little like taking 10 second average reading of a voltage of interest and than 10 seconds of zero reading to compensate for an offset that is effected by the noise like the 10 second point of the Allan deviation.
With many normal meters the 10-100 seconds range is about the sweet spot, with the lowest Allan deviation. Thus this time frame is often used in sensitive tests - it also kind of matches what is practical with manual switching and reading. So for such manual operated tests the Allan deviation at some 10 seconds is the relevant noise parameter, not 5 PLC noise.

Waiting more than 100 seconds not really a solution as here noise sources like slow thermal drift will come in. So at 100 seconds it is only back to the 2 seconds value.
The script to force a kind of manual triggered measurement is a workaround, but due to extra waiting times involved it would add noise.

[niner_007]

could one in theory implement better (custom) AZ, by using an external ultra precise reference and custom scripting on the DMM, so when you do the AZ, you switch to the reference (relays), measure, offset (on the data read from the DMM), then continue measuring? without modifications to the DMM, of course, this would all be external

[niner_007]

There are already similar curves for the Keysight meters earlier in this thread. The suitable test would be to collect data for a reasonable long time (e.g. 1-3 hours) under stable conditions. Ideally with a very stable source (e.g. LTZ1000 or better) in 10 V range, but the extra noise is also visible with just a short.

The Alan variance curve is kind of related to the frequency spectrum of the noise. The curve from the custom AZ script is about the normal shape expected. In some metrology applications the minium value reached is an important quality criterion, especially if this is not at too long a time scale. That 10-100 seconds time scale is just the one needed for doing test with dividers or similar. The custom AZ script can do better in this range, but as one can see it is rather noisy on the short time scale, and an internal solution if done right could likely do considerably (about a factor of 1.5-4) better. So the extra script is kind of a work around but not a full solution.

Just using the 5 PLC mode does not solve the problem. Once you average those data taken at 5 PLC to get even lower noise that extra noise comes back and noise even goes up when averaging more than about 100 samples. This might also lead to wrong conclusions if the normal statistics is used to estimate noise. One would normally assume the data are not correlated over more than maybe the next neighbors - but with the Keithley meters (to some extend a similar bump is even visible with the K2002 and may very well effect more meters) there seem to be some correlation up to 100 seconds.

From the analysis done so fat in this thread, it looks like Keithley is using some averaging on the zero readings when doing AZ, even if digital filtering is turned off. This sounds like a good idea at first sight and could reduce the short time scale (high frequency) noise a little. However it comes at the price of not so good suppression of 1/f noise of the ADC/amplifier. So generally this is not a good idea and it would be really nice if Keithley would also offer the normal simple difference AZ mode.

could one use the AZ script as a workaround? what is the problem with using it, beside the noise in the shorter time scale? speed?

[Kleinstein]

The script is a kind of workaround at the costs of higher noise for shorter time scales. In addition the timing is a little different (slightly slower). Depending on the application this can help, though an upgraded ("fixed") software in the DMM would be better (lower noise) and easier to use.

The allan deviation curve looks so odd that the the first though was of a software bug that might find a quick fix. However as a similar effect is found with other meters too  - this "feature" looks more like part of very old software at the heart of the measurement system, that might not be changed so fast.

[niner_007]

The script is a kind of workaround at the costs of higher noise for shorter time scales. In addition the timing is a little different (slightly slower). Depending on the application this can help, though an upgraded ("fixed") software in the DMM would be better (lower noise) and easier to use.

The allan deviation curve looks so odd that the the first though was of a software bug that might find a quick fix. However as a similar effect is found with other meters too  - this "feature" looks more like part of very old software at the heart of the measurement system, that might not be changed so fast.

thank you; and that all in the context of the AZ implementation, correct?

[Kleinstein]

We don't know for sure, but the likely cause of the extra noise for the 10 s time scale is likely due to the AZ implementation. The script for the workaround also supports this.

My suspicion is that they use some kind of filtering on the zero readings used in AZ mode. So instead of simply using the last zero reading the suspicion is that some kind a average zero reading from the past 10-30 seconds is used. I also have a faint memory (was some 20 years ago I used one) of such a description in one of the older Keithley meters (like 196) to do it like this, but I could not find it in the manuals.

This sounds like a good idea at first because it would hide much of the white noise if the zero reading. However the downside is that the 1/f noise in the time frame for averaging is no longer suppressed. Depending on the noise spectrum this might have been OK with the old (lower resolution) meters with high white noise, but for the DMM7510 and most other modern ADCs  this is a really poor choice.

Let's hope they get this one fixed when they bring out an DMM85xx - it at least would make it much easier to get 8 digit resolution, when you get about a factor of 2-4 improvement from a simple software change.

[maxwell3e10]

I have slight hope that the autozero problem is fixed in DMM6500, but wouldn't hold by breath. If it has as low noise as DMM7500 and the only difference is in the voltage reference, it would be a great meter.

In general, I think they are putting more effort into a nice user interface than into the intricacies of analog design. And it makes market sense, it's both easier to hire engineers who can design apps and easier to sell meters based on shiny interface than on obscure specs. Even on eevblog  there are more posts about "noise" as in fan noise instead of actual electrical noise. So, as someone else had said here, its not obvious that a meter rivaling HP3458 will ever be build.

[niner_007]

We don't know for sure, but the likely cause of the extra noise for the 10 s time scale is likely due to the AZ implementation. The script for the workaround also supports this.

silly question, what if you turn AZ off and don't use it at all?

[maxwell3e10]

silly question, what if you turn AZ off and don't use it at all?

Look at reply #3 in this thread. Basically, the autozero algorithm behaves as if it's not even used until about 30 sec, while the 1/f noise of the meter is such that it would benefit from autozero after 1 sec.

[niner_007]

silly question, what if you turn AZ off and don't use it at all?

Look at reply #3 in this thread. Basically, the autozero algorithm behaves as if it's not even used until about 30 sec, while the 1/f noise of the meter is such that it would benefit from autozero after 1 sec.

is the noise because of the drift? if you have a super stable environment, would it still happen? so after AZ is ran, it starts to drift until you run it again? basically constantly drifting?

[kj7e]

AutoZero corrects offset voltage drift from the input gain and A/D stage.  These drifts happen to all DMM's including the 7510, 34470, or the 3458.  The AZ function switches from the input signal to an internal short or known zero to calibrate the zero offset.  The issue with the 7510 seems to be how and when the AZ happens.  Its not a major issue (IMHO), just appears not to work as well as it should and it shows up as some extra front end noise.  It is a concern, especially if it could be better optimized in software.  So far Keithley has not made any public comments on this, so its possible it's not an easy task.

It would not matter how stable your environment is, you cant stop the internal voltage offset drifts from happening.  See this video which helps explain some of this behavior;

[Kleinstein]

Besides drift the auto-zero function also suppresses 1/f noise of the ADC and input amplifier (the part included in AZ function).
For some reason this does not work well with the DMM7510 until some 0.02 Hz.  Especially with a higher grade DMM were the ADC is not adding much white noise this giving away some possible performance.  For many applications it is not a problem,  more like a negative surprise as the performance to something like 100 PLC or with external averaging is way lower than what is expected from the 5 PLC specs. Its kind of the step from barely 7.5 digit resolution to nearly 8 digits that is given away due to some odd software.

I don't have much hope to get it fixed with the DMM6500 - the problem is more like some old tradition carried on for to long and not need to change this for 6 digit performance.  I have more hope if they want to bring out an 8 digit version that they might realize it would help to suppress 1/f noise.

[PA4TIM]

new 7510 , first hour of use a BSD.
Firmware rev 01.06.4C,
faultcode: PC=0012D6F8

I was in the continuity function, pressed MENU and the screen got blue.

OT: Great meter, I really like it. For me the speed, the sampling part (even better as I hoped), the dry circuit ohms, the nOhm-Gohm , pA resolution and the easy user interface are the winners but I do not know if I like the continuity and diode functions.
2 functions that are very important for me.  (I really like the continuity from my 2000)
Continuity is fast enough but it latches much to long and the beep sounds like a mix of bird with a cold and a choked mouse.
Diode test has two levels and a beep (the same choked mouse), what is great on its own but not for my use. (component level repair work) My Agilent beeps once for a diode and stays beeping on a short so I do not have to look to the screen. The 14V is also great but not always for use in situ.

[PA4TIM]

Today the continuity stalled the meter. I changed the limit to 1.5V . I could change to the function screen but when I pressed another function the round in progress thingy showed up and stayed. I had to re-power it.

I still like the meter, it is so comfortable in use. I normally have a handful of meters powered up so I do not need to switch functions, but it is a bit sick a meter in this class after 3 years still has bugs. My Siglent scopemeter has some too but not in a way it stalls and need a repower cycle (and is probably the cheapest isolated scopemeter on the market)

[maxwell3e10]

Yes, I have seen it hang up with various less frequently used features, like temperature in my case. I think they managed to get the bugs out for the most common voltage and current measurements, but other may still not be tested thoroughly. There is a Tek forum devoted to this topic: https://forum.tek.com/viewforum.php?f=363
If you post there, they will try to reproduce the problem and maybe it will get fixed in the next FW update. Not sure why we have to do their work for them, though.