-
Does anyone have data from long term logging of voltage references that they can share? I'd be especially interested in data with temperature, humidity and barometric pressure also recorded. But even without the environmental data, data from multiple devices over the same time frame would be useful.
I'd like to see if I can construct a mathematical model that accurately predicts future behavior of a voltage reference. As Jason sent me several LM399s, that's the most desirable staring point. But I'll gladly take any data I can get. My intent is to use half the data to predict the next half and then compare to the actual data.
If you have such data please PM me so we can arrange a transfer.
Thanks,
Reg -
Here is some interesting reading from Fluke (attached).
Linked from this thread, this is also good: https://www.eevblog.com/forum/metrology/the-current-cost-of-chasing-ppm's-(fluke-732b-price-list)/
-
Good reading, but the data are inadequate for the sort of methods I have in mind.
That said, it does indicate that even 10 ppm is quite ambitious for a hobby project.
If temperature, humidity and barometric pressure are significant factors, they will be correlated across multiple references, including the measuring device. If one has the environmental data then any effects from those can be removed from the data. However, the accuracy of doing this is dependent upon how many references are involved and how long the time series are. -
Hello,
I have many recorded data. But it will not help you for your individual devices.
Humidity plays only a role if you have certain packages e.g. plastic DIP package
or SMD-ceramic package (like the LT LS8-package) directly soldered to a epoxy board.
Other packages like the TO-46 or TO-99 are nearly insensitive to humidity.
Ok I have one very drifty LM399 device which is also sensitive to humidity.
But I fear that this old and desoldered device is no longer hermetically tight.
If you look at the measurements: (dayly measurements with timescale (x-axis) in days)
Every reference behaves very individual.
There are some common effects: spikes that are related to contact problems during measurement.
Shifts that are related to movement of the ageing box, since the devices are not soldered into the PCB but used with sockets.
On the drift: how do you distingquish the drift of the LM399s from the drift of the measurement system.
Ok: I have other measurements where I measure 2 LTZ1000s and can see how the measurement system drifts against those.
But in the end those LTZs also drift by -1 .. -2 ppm/year normally. (and if I have a accident then they drift several ppms in 1 second).
My strategy would be:
make own measurements with the best instrument available. At least one measurement per day e.g. averaging over 1-5 minutes.
calculate a sliding standard deviation over the last 42 days (1000 hours).
Determine the device with lowest standard deviation and use that as your master reference.
Of course you also can try to measure differences between references.
In this case you have to determine the pair of references with the lowest standard deviation.
But there are more parameters which can fluctuate so this really helps only if your instruments (absolute) stray in 10V range is larger than in 100mV range.
with best regards
Andreas
-
I'd *really* like to get your data in CSV format. I can't do anything with a graph. It would be very useful for characterizing the shape of the aging curves. It won't tell me anything about my devices, but I can only determine that over time.
If you try to fit he wrong model to data you don't get a good result. The only way to determine the correct model is to have a lot of data. Without data someone else has recorded over years I can't do anything for several years. Which means it probably won't happen at all.
In order to determine aging characteristics, the first task is to quantify and remove correlated effects. It's rather like peeling an onion.
I spent 30 years in the oil industry in reflection seismology. Seismic processors routinely take data that has so much noise in it that there is no discernible signal at all. Moreover, no two datasets are the same. And no two answers either, though the important part of the picture is consistent after processing. The number of completely different software packages used for something as complex as seismic processing is huge and there are many undetected bugs because the developers did not implement proper tests when writing the software. Not surprising given that most of it is written by scientists with no real knowledge of software engineering.
I have a long track record of analyzing noisy data that other people said could not be used. I've done it at the scale of several hundred gigabytes over a 180,000 square mile patch of the Gulf of Mexico. The research staff at a super major said it couldn't be done. I not only did it, I did it in 9 months to a hard deadline imposed by the lease bid process. And I wrote all the software for this while I was doing the work.
So please consider giving me a CSV file of your measurements.