-
#3820 by frobnicat on 13 Jul, 2016 13:36
-
.../...
@Monomorphic have you made a spreadsheet or raw data file available somewhere ? I must admit I haven't followed your publications with the care they'd deserve, only catching-up here on NSF thread and no time to follow all the links. Motivation : small test in the vertical scale for data acquisition artefacts (too much flattish steps to my taste, I understand you operate the LDS near its limits).
Never mind, a manual check is enough, there is clearly some quantization artefact and we are overinterpreting the meaning of the flat region, or at least the meaning of its extreme flatness. See attached figure where I overlayed (manual fit) some horizontal lines (in green) each time I see a kind of plateauing in the plot : plateaux and "jumps" are aligned (on the vertical scale) all along the plot and moreover they are quite regularly spaced (albeit not perfectly). This is characteristic of a quantization error introduced by the ADC system.
My hypothesis is that three steps are involved
1. at the input of ADC there is a high freq. electronic analog thermal noise on top the slowly varying signal (real position of pendulum)
2. the ADC somehow oversamples compared to the final apparent sampling rate and picks this noise on top of the slowly varying signal. When the base signal is in the middle of a quantization "stripe", the amplitude of noise is not enough to give (discrete) values above or below the thresholds : output numerical value constant. When the base signal is approaching a threshold of quantization, numerical output starts to shoot +1 or -1 with a frequency proportional to transition from one quantized value to the next. The result is still not looking like the plots though...
3. a numerical post-processing occurs (firmware in the ADC ? later on in the acquisition chain ?) and smooths the transitions, probably a low pass of some sort. So that for instance in place of
... 3 3 3 3 3 3 3 4 3 3 3 4 3 4 3 4 4 3 4 4 4 4 3 4 4 4 4 4 4 4 4 4...
we have a kind of semi blurry
... 3 3 3 3 3 3.1 3.2 3.3 3.5 3.6 3.5 3.8 3.9 3.8 3.9 4 4 4 4 4 4 4...
where some plateaux are still present and yet there are kind of continuous transitions between
That or something similar is happening. Conclusion would be that the flat plateau after return loss peak is an artefact, this is just a place of low slope happening to be between 2 thresholds of quantization of ADC. We might wonder why there is little slope here, but not why this is apparently so flat leveled.
Oh, I see you posted the raw data, thanks ! Maybe it can further specify the above hypothesis...
-
#3821 by SeeShells on 13 Jul, 2016 13:37
-
Monomorphic might be correct in the ADC sensor needing to be shielded. He has his resolution set to 3um (a blood cell is 6-9um) and because of the high sensitivity he just might be picking up noise and or RF leaking which could overdrive the input to the ADC sensor.I found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
frobnicat wrote:
...In brief, the flat region of interest is suspiciously too flat given the expected level of noise in said region (given what is to be seen after). Either the explanations are over-fitting the data (flatness is a fluke) or something else is at play like something as improbable as data acquisition perturbed by sudden RF interference when resonance locks, or some mechanical stickiness actually causing the lock in resonance and not the reverse causality, or ferrite chokes ?
...
1) Have you measured whether you have any RF leaking from the cavity?
2) Is there any mechanical stickiness in your setup? Do you have such flat regions of response if the pendulum is given an initial torque, by hand, (free response to an initial angular displacement) and let it come back and oscillate on its own, without power going to the EM Drive?
I would agree with him to see if he could shield the sensor and maybe check the grounding on his beam support structure.
On my tests I'm waiting for a new set of matched HV transformers as I lost one during testing. They were used to start with, so what can I expect.
Shell -
#3822 by Monomorphic on 13 Jul, 2016 14:10
-
Conclusion would be that the flat plateau after return loss peak is an artefact, this is just a place of low slope happening to be between 2 thresholds of quantization of ADC. We might wonder why there is little slope here, but not why this is apparently so flat leveled.
Thanks for the analysis frobnicat. I was noticing the smaller plateaus late last night as well. There are three settings on the LDS, 3um (60ms), 20um (2ms), and 80um (0.15ms). Before I do anymore powered tests, I am going to try and get rid of the artifacts and do a drop weight calibration test. I think 3um may be overkill and want to see what results the 20um setting yields. Once we are all happy that the data is clean, I will stop making changes to the setup and resume tests.
To put things into perspective, the maximum displacement of the beam during the last test was ~0.12volts. If TT and rfmwguy's estimate of 0.216volts/mm is close to mine, then the measurements we are taking occur over a distance of 0.55mm.
-
#3823 by zen-in on 13 Jul, 2016 16:00
-
I found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing.
-
#3824 by SeeShells on 13 Jul, 2016 16:03
-
That's a very fine measurement, equal to about the thickness of 5 sheets of paper. I'm quite impressed. It also shows how small of a shift of beam displacement your measuring and how external vibrations could show up in the measurements. EagleWorks had an air cushioned anti-vibration table they used to mitigate externals.Conclusion would be that the flat plateau after return loss peak is an artefact, this is just a place of low slope happening to be between 2 thresholds of quantization of ADC. We might wonder why there is little slope here, but not why this is apparently so flat leveled.
Thanks for the analysis frobnicat. I was noticing the smaller plateaus late last night as well. There are three settings on the LDS, 3um (60ms), 20um (2ms), and 80um (0.15ms). Before I do anymore powered tests, I am going to try and get rid of the artifacts and do a drop weight calibration test. I think 3um may be overkill and want to see what results the 20um setting yields. Once we are all happy that the data is clean, I will stop making changes to the setup and resume tests.
To put things into perspective, the maximum displacement of the beam during the last test was ~0.12volts. If TT and rfmwguy's estimate of 0.216volts/mm is close to mine, then the measurements we are taking occur over a distance of 0.55mm.
Shell -
#3825 by rfmwguy on 13 Jul, 2016 16:22
-
My old weight cal was 1.34mV/mg and Phil used the LDS chart to convert to mm. Keep in mind my torsion beam is 100+ inches total length with a long resonance cycle of sub hertz, can't recall the exact number and I'm sure its changed with the harness coming off the beam. The advantages are minimal high frequency oscillations due to the dampening and the high isolation of the LDS on the far end of the beam away from the magnetron.Conclusion would be that the flat plateau after return loss peak is an artefact, this is just a place of low slope happening to be between 2 thresholds of quantization of ADC. We might wonder why there is little slope here, but not why this is apparently so flat leveled.
Thanks for the analysis frobnicat. I was noticing the smaller plateaus late last night as well. There are three settings on the LDS, 3um (60ms), 20um (2ms), and 80um (0.15ms). Before I do anymore powered tests, I am going to try and get rid of the artifacts and do a drop weight calibration test. I think 3um may be overkill and want to see what results the 20um setting yields. Once we are all happy that the data is clean, I will stop making changes to the setup and resume tests.
To put things into perspective, the maximum displacement of the beam during the last test was ~0.12volts. If TT and rfmwguy's estimate of 0.216volts/mm is close to mine, then the measurements we are taking occur over a distance of 0.55mm. -
#3826 by jmossman on 13 Jul, 2016 16:41
-
I found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing.
Hi Monomorphic,
Would you happen to have a signal generator and oscilloscope somewhere to help characterize your A/D?
A high frequency signal would help get a feel for potential distortion due to A/D rise/fall mismatches, and a low amplitude signal would help isolate A/D quantization effects (although perhaps a stair-step will be visible on the A/D output regardless of input signal; not sure what your A/D spec looks like). The oscilloscope would provide something to compare the A/D output with (although if the signal generator was "trusted" to be fairly well calibrated/behaved, the oscilloscope wouldn't be as critical).
Also, does your A/D have multiple channels available? Zen-in's theory of a sticky bit might be provable by running the signal generator into multiple channels at the same time (although strictly speaking running the test serially with a load on just 1 channel at a time also has merits). -
#3827 by Monomorphic on 13 Jul, 2016 17:48
-
Would you happen to have a signal generator and oscilloscope somewhere to help characterize your A/D?
I have a signal generator, but no oscilloscope. I'm looking at a couple of USB oscilloscopes. They're not very expensive. -
#3828 by dustinthewind on 13 Jul, 2016 17:57
-
...
When I read this I had the thought to use a hair from my head to push the beam. I am not sure how well it would work or how many hairs you want to lose. There may be another option, like a brush with fine hairs already in it?
As for stickiness, it's something I'm looking into. I will need to redo the 'tap tests' since I switched to high resolution (3um). The previous taps are about 10 times stronger than the anamalous force seen in the latest powered runs. So I have to come up with something that can tap the beam much more lightly. -
#3829 by frobnicat on 13 Jul, 2016 19:32
-
I found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing.
Apart from the least significant bit (or a number of least significant bits altogether), a stuck bit will not result in recurring flat areas. As an example see attached plots on a 7 bits sampling of a full scale ramp up : in blue a normal response, in red the response if middle bit is stuck to 0. There is no recurring flat areas because weaker bits always do their job to "follow" small variations, while the missing more significant bit causes discontinuities. This is more severe than just flat areas (in terms of distortion) and doesn't appear at all like the aspect in monomorphic's plots.
In monomorphic's plot 11 it's more like the (vertical) scale operates on the 3 or 4 least significant bits for the range of input presented to the ADC (with an added layer of low pass filtering after the fact). I wonder what this means that the LDS is operated at "3µm setting", I suspect this is not meaning that the least significant bit corresponds to steps of 3µm, otherwise the 0.55mm range would translate to 183 digitized levels, more like 8 bits resolution than 3 or 4 bits as appears from the number of clearly quantized "flat areas" (vertically).
-
#3830 by Rodal on 13 Jul, 2016 19:45
-
I found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing.
Apart from the least significant bit (or a number of least significant bits altogether), a stuck bit will not result in recurring flat areas. As an example see attached plots on a 7 bits sampling of a full scale ramp up : in blue a normal response, in red the response if middle bit is stuck to 0. There is no recurring flat areas because weaker bits always do their job to "follow" small variations, while the missing more significant bit causes discontinuities. This is more severe than just flat areas (in terms of distortion) and doesn't appear at all like the aspect in monomorphic's plots.
In monomorphic's plot 11 it's more like the (vertical) scale operates on the 3 or 4 least significant bits for the range of input presented to the ADC (with an added layer of low pass filtering after the fact). I wonder what this means that the LDS is operated at "3µm setting", I suspect this is not meaning that the least significant bit corresponds to steps of 3µm, otherwise the 0.55mm range would translate to 183 digitized levels, more like 8 bits resolution than 3 or 4 bits as appears from the number of clearly quantized "flat areas" (vertically).
This is really great analysis, frobnicat, showing once again the value of open technical forums like NSF where the Internet can bring expertise from a large number of people with expertise in a wide number of areas.
Showing the value of technical criticism in finding out the truth: that is how it the "peer-review" process works in science, everything has to be questioned. -
#3831 by jmossman on 13 Jul, 2016 20:52
-
Would you happen to have a signal generator and oscilloscope somewhere to help characterize your A/D?
I have a signal generator, but no oscilloscope. I'm looking at a couple of USB oscilloscopes. They're not very expensive.
I've personally not had any luck with USB oscilloscopes in the < $400 range due to bad/buggy software and very poor A/D resolution.
Maybe the landscape of products has changed over the past 1-2 years.
In the meantime, I'd recommend oversampling your inputs as much as possible. If the LDS really was generating samples at ~16Hz, then your sampling rate should be at least double that (i.e. 32Hz or faster) to satisfy the Nyquist rate. Since storage is pretty cheap, using a kHz+ sampling rate should help demonstrate any noise in the system (implementing a low-pass filter in Excel isn't difficult).
Also, using your existing signal generator and your existing A/D will provide the abilities of a crude low-bandwidth oscilloscope.... unless the generator can't generate frequencies low enough for your existing A/D to accurately sample.
Best of luck with your experiments! With each additional piece of characterization data, evaluating the potential error bars in the experimental results gets a little bit easier.
P.S. Any chance you could share the Make/Model information for the LDS and A/D that you're using? -
#3832 by rfmwguy on 13 Jul, 2016 21:32
-
Same as I have from my Test Report of last October: https://forum.nasaspaceflight.com/index.php?action=dlattach;topic=38577.0;attach=1072184;sess=47785Would you happen to have a signal generator and oscilloscope somewhere to help characterize your A/D?
I have a signal generator, but no oscilloscope. I'm looking at a couple of USB oscilloscopes. They're not very expensive.
I've personally not had any luck with USB oscilloscopes in the < $400 range due to bad/buggy software and very poor A/D resolution. Maybe the landscape of products has changed over the past 1-2 years.
In the meantime, I'd recommend oversampling your inputs as much as possible. If the LDS really was generating samples at ~16Hz, then your sampling rate should be at least double that (i.e. 32Hz or faster) to satisfy the Nyquist rate. Since storage is pretty cheap, using a kHz+ sampling rate should help demonstrate any noise in the system (implementing a low-pass filter in Excel isn't difficult).
Also, using your existing signal generator and your existing A/D will provide the abilities of a crude low-bandwidth oscilloscope.... unless the generator can't generate frequencies low enough for your existing A/D to accurately sample.
Best of luck with your experiments! With each additional piece of characterization data, evaluating the potential error bars in the experimental results gets a little bit easier.
P.S. Any chance you could share the Make/Model information for the LDS and A/D that you're using? -
#3833 by rfmwguy on 13 Jul, 2016 21:46
-
Zen, you should know that I took your advise and used the 30dB attenuator in-line with the USB spec an. It did what you said it would do, drop the sidebands out of existence and provide a sharp spike on the fundamental...albeit 30dB lower. This allows me to better focus on the center by removing the clutter even tho the center is closer to the noise floor...ThanksI found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing. -
#3834 by zen-in on 13 Jul, 2016 22:45
-
Zen, you should know that I took your advise and used the 30dB attenuator in-line with the USB spec an. It did what you said it would do, drop the sidebands out of existence and provide a sharp spike on the fundamental...albeit 30dB lower. This allows me to better focus on the center by removing the clutter even tho the center is closer to the noise floor...ThanksI found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing.
Thank you. I don't think it was just I who suggested that. Network analyzers always have a 10 dB or 14 dB attenuator on the input. The general Radio 1713 network analyzer used 5x attenuators. This photo has bad color balance because the 874 connector attenuator should be green.
-
#3835 by rfmwguy on 13 Jul, 2016 23:11
-
If you only knew the grief GR connectors used to cause me back in the day. Also, bet I'd seen thousands of 14dB attend before I realized they're part of an FAA cal procedure.
Zen, you should know that I took your advise and used the 30dB attenuator in-line with the USB spec an. It did what you said it would do, drop the sidebands out of existence and provide a sharp spike on the fundamental...albeit 30dB lower. This allows me to better focus on the center by removing the clutter even tho the center is closer to the noise floor...ThanksI found the Return Loss scan I did with HDPE tuned to where it is now. I've marked it in the image. Dr. Rodal, notice the same 'flat' regions? One at RF on and one at RL peak. Damn interesting!
It looks to me like your A/D is non-monotonic. The flat areas occur at the same level throughout the graph. It could be that one bit of the A/D is stuck at one level. That will result in recurring flat areas. To test the A/D drive it with a triangle wave, if that is possible. You can also analyze the data near each flat spot to see what codes (DN values) are missing, by dividing the LSB value into each data point. You can do this with a spreadsheet and then plot a histogram of DN values. It looks like several DN values are missing.
Thank you. I don't think it was just I who suggested that. Network analyzers always have a 10 dB or 14 dB attenuator on the input. The general Radio 1713 network analyzer used 5x attenuators. This photo has bad color balance because the 874 connector attenuator should be green. -
#3836 by frobnicat on 13 Jul, 2016 23:31
-
.../...
To put things into perspective, the maximum displacement of the beam during the last test was ~0.12volts. If TT and rfmwguy's estimate of 0.216volts/mm is close to mine, then the measurements we are taking occur over a distance of 0.55mm.
How much of the input range of your ADC is 0.12V ? What is the width of the output of the ADC, 8 bits, 10 bits, 12 bits, 16 bits ? If the ADC takes in 0-5V full scale and is 8 bits output then that could explain that your experiment is recorded on the 3 or 4 least significant bits only (between 8 or 16 discrete levels). Sorry if this is documented in some previously linked data.
-
#3837 by SeeShells on 13 Jul, 2016 23:48
-
It's correct., the currents have to flow up and around in the slotted which prevents a strong counter EMF.
I was researching electromagnetic braking a month or so ago and when I saw the "flats" I thought of this video.
Shell
Very interesting video, SeeShells...
However, it makes me wonder what would happen if it had horizontal cuts instead of drilled holes, as obviously, the vertical cuts fails to generate a counter-magnetic field.
The only difference I see between the drilled holes and the vertical cuts, is the lack of electromagnetic induction current flow (hence lack of counter magnetic field) in the last one. Am I'm right in assuming that? -
#3838 by Monomorphic on 14 Jul, 2016 00:54
-
.../...
To put things into perspective, the maximum displacement of the beam during the last test was ~0.12volts. If TT and rfmwguy's estimate of 0.216volts/mm is close to mine, then the measurements we are taking occur over a distance of 0.55mm.
How much of the input range of your ADC is 0.12V ? What is the width of the output of the ADC, 8 bits, 10 bits, 12 bits, 16 bits ? If the ADC takes in 0-5V full scale and is 8 bits output then that could explain that your experiment is recorded on the 3 or 4 least significant bits only (between 8 or 16 discrete levels). Sorry if this is documented in some previously linked data.
It's the cheapest model Dataq available. Here it is: http://www.dataq.com/products/di-145/
I would be willing to purchase a better one if it made a difference and didn't break the bank.
-
#3839 by WarpTech on 14 Jul, 2016 01:13
-
.../...
To put things into perspective, the maximum displacement of the beam during the last test was ~0.12volts. If TT and rfmwguy's estimate of 0.216volts/mm is close to mine, then the measurements we are taking occur over a distance of 0.55mm.
How much of the input range of your ADC is 0.12V ? What is the width of the output of the ADC, 8 bits, 10 bits, 12 bits, 16 bits ? If the ADC takes in 0-5V full scale and is 8 bits output then that could explain that your experiment is recorded on the 3 or 4 least significant bits only (between 8 or 16 discrete levels). Sorry if this is documented in some previously linked data.
It's the cheapest model Dataq available. Here it is: http://www.dataq.com/products/di-145/
I would be willing to purchase a better one if it made a difference and didn't break the bank.
What are you using for input conditioning? 10 bits is fine if the input signal fills the range, but if the signal is only using 2 or 3 bits, then it requires a front end OpAmp to boost the signal.
...
Maybe the landscape of products has changed over the past 1-2 years.