Author Topic: Gaussian resistors redux.  (Read 13809 times)

0 Members and 1 Guest are viewing this topic.

Offline wilfred

  • Frequent Contributor
  • **
  • Posts: 605
  • Country: au
Gaussian resistors redux.
« on: November 12, 2011, 11:08:56 pm »
 .
« Last Edit: July 28, 2018, 12:35:36 pm by wilfred »
 

Offline amspire

  • Super Contributor
  • ***
  • Posts: 3632
  • Country: au
Re: Gaussian resistors redux.
« Reply #1 on: November 13, 2011, 12:41:51 am »
I would be very surprised if Dave's bench mat had enough resistivity to affect the reading at all. It is not good if an antistatic mat has enough conductivity that a worker could actually get an electric shock through contact with a live mat. It is also not good if the mat has enough conductance to damage a powered circuit board on the mat. It is only meant to dissipate static charges and that is all.

There are lots of reasons that a manufacturer could make a whole batch low, but I think the main one is that they are in the business of making resistors as quickly and as cheaply as they can possibly do it. The spec is 1% and they have to make them for a cheaper price to the 0.1%. So they start a batch, and see that it is making the resistors 0.3% low. Is that within spec?

Yes it is, so they keep going with the batch. Could they take more care and get the average within 0.1%? Yes, but why would they do that if it costs a bit more money? Remember that the Chinese resistors were almost certainly made for a far lower profit margin to the Philips resistors.

Alternatively, another reason is that all resistors can drift with age,  particularly if they are running near maximum power continuously. If they know their resistors over a lifetime drift between 0% and +0.6%, then it would make really good sense to make them 0.3% low. If they don't, then a resistor that started off at +0.5% may be +1.1% after 1000 hours. Thin film resistors are a bit chronic for going high with aging, because the film surface can oxidize at elevated temperatures, and as the film is only about 250 Angstroms thick, any oxidizing makes it increase significantly.

The only way to know for sure is to ask the manufacturer.

Richard
« Last Edit: November 13, 2011, 01:27:29 am by amspire »
 

Offline IanB

  • Super Contributor
  • ***
  • Posts: 9449
  • Country: us
Re: Gaussian resistors redux.
« Reply #2 on: November 13, 2011, 04:47:31 am »
There could possibly be an offset due to the mat.

Those mats are normally constructed with a conductive bottom layer where the ground is attached and a dissipative top layer to work on. The top layer conducts any accumulated charge to the bottom layer where it is dispersed.

In one example of a quality mat I looked up the resistance from top surface to conductive bottom layer was in the order of 1 Megohm, so putting such a resistance in parallel with 1 k may introduce a measurable bias into the readings. Maybe not enough to fully account for the bias Dave observed, but enough perhaps to warrant specifically measuring and eliminating that source of error if necessary.

The conductivity of the mat surface would not be enough to upset the operation of a normal low voltage circuit if a PCB was sitting on it, but it might perhaps upset some high voltage or sensitive circuits.
I'm not an EE--what am I doing here?
 

alm

  • Guest
Re: Gaussian resistors redux.
« Reply #3 on: November 13, 2011, 04:54:53 am »
I don't understand why people would expect the mean of the process to be at exactly the nominal value. Sure, with perfect process control this would be the case, but why bother calibrating your process for 1% resistors to be much better than 1%? As long as you meet the required tolerance plus guardband with a good yield, you're fine. Any better is a waste of money, and competes with your 0.1% resistors.
 

Offline IanB

  • Super Contributor
  • ***
  • Posts: 9449
  • Country: us
Re: Gaussian resistors redux.
« Reply #4 on: November 13, 2011, 05:11:48 am »
I don't understand why people would expect the mean of the process to be at exactly the nominal value.

There are two independent questions:

1) What is the mean value of the resistors?
2) Are there any sources of measurement error in the experiment that might affect the accuracy of the results?

Suggesting that there might be a source of measurement error does not on my part imply any expectations about where the mean value should lie.

On the other hand, I would expect the manufacturing system to be continuously measuring the mean value in production and adjusting the process in a feedback loop to reduce the net offset to zero. But then maybe you do not find such process sophistication in "Wun Hung Lo" brands...
I'm not an EE--what am I doing here?
 

alm

  • Guest
Re: Gaussian resistors redux.
« Reply #5 on: November 13, 2011, 06:04:15 am »
Sure, measurement errors are possible, but the expectation that the error results in a lower value does make assumptions about the value of the resistors. I would expect them to adjust the process to have a mean that's close to nominal, but I wouldn't expect them to use a precision reference for this in a 'Wun Hung Lo' factory, plus I expect this process to have some hysteresis. As long as the spread is low enough, there's no advantage to adjusting them very close to nominal.
« Last Edit: November 13, 2011, 08:23:36 am by alm »
 

Offline grummund

  • Newbie
  • Posts: 3
Re: Gaussian resistors redux.
« Reply #6 on: November 13, 2011, 07:21:37 am »
I'm just wondering Dave, if you checked a sample of the resistors using the HP bench meter to confirm none were above 1k?  Well I'm sure you did  ;) but what were the results?

If the results are confirmed with the same bench meter as used in the previous test then my money would be on the factory test kit being out of cal.  Well, perhaps not strictly out of cal. as such, but at least of lower spec. than the Agilent used for this test.
 

Offline IanB

  • Super Contributor
  • ***
  • Posts: 9449
  • Country: us
Re: Gaussian resistors redux.
« Reply #7 on: November 13, 2011, 08:08:35 am »
Let's ignore the manufacturing process for a moment and just consider the work surface. In the interests of good experimental technique I would want to place the two pronged probey thing against the blue surface to measure the resistance between the probe tips and then report what it is. Such as "infinity", "10 Meg", "1 Meg", or whatever. If it's infinite that's all well and good, but at least it has been measured and confirmed, rather than assumed and neglected. Doing good science involves controlling all the variables, and that includes verifying things that are assumed to have no impact on your results. Stating the result to camera assures the viewers that you considered this and avoids them asking or wondering later.
I'm not an EE--what am I doing here?
 

Offline Kiriakos-GR

  • Super Contributor
  • ***
  • Banned!
  • Posts: 3525
  • Country: gr
  • User is banned.
    • Honda AX-1 rebuild
Re: Gaussian resistors redux.
« Reply #8 on: November 13, 2011, 08:15:49 am »
Just watched the video, and I have one tip for Dave, so to cut down the extra zeros in the Agilent GUI , you can use the Truncate option in the menu.
This will cut down some resolution, but this is what this function does !! 
Try it.  ;)
 

Offline amspire

  • Super Contributor
  • ***
  • Posts: 3632
  • Country: au
Re: Gaussian resistors redux.
« Reply #9 on: November 13, 2011, 09:28:48 am »
It is clear from the video and from previous videos that Dave is very rigorous in his attention to measurement accuracy. I just think he has allowed a "mental bias" to creep in to his expectations of a resistor branded "OneHungLow". He could be so accustomed to lumping all inferior Chinese goods into the Onehunglow bin that it has subconsciously affected his thinking.

If  you go back to the first video you can easily form the impression Dave expected to see a perfect Gaussian distribution centered around the mean. He got a slight negative offset (I think that was his words) and was a little surprised by it. But not this time. It was a big offset and not outside his preconceived ideas about low quality Chinese goods.

I don't see this as true is any way. I think Dave didn't know what to expect from the Philips resistors. I don't think he knew what to expect from the Chinese resistors.

All he did was report what he saw.

0.3% is not a big offset if you are making 1% resistors - it is a perfectly acceptable offset.

And don't get too caught of on this Gaussian thing.  Gaussian curves usually show the net result of maybe 20 factors, and typically when you average those 20 factors out, you tend towards randomness. It is common for each individual factor to be nothing like a Gaussian curve. Say the resistor machine had a sticking cog. Each time it stuck, the resulting resistor was about 0.2% higher then it normally would be.  The curve will end up with 2 peaks - one about 0.2% higher in resistance then the other.

In fact you use divergence from the Gaussian curve to help track down specific problems, because it indicates something is happening that is not the result of random noise.
Quote

I think 400 and 1000 data sample is sufficient to get a good Gaussian distribution centered around the mean  if you are measuring a truly random sample. If not then it begs the question of what advantage does it offer a manufacturer of 1% resistors to make them all at one end of the tolerance curve. I just don't see any.

Regards
Wilfred

I did try and answer that above. There are several perfectly sensible reasons.

Richard
 

Offline Rufus

  • Super Contributor
  • ***
  • Posts: 2094
Re: Gaussian resistors redux.
« Reply #10 on: November 13, 2011, 12:20:06 pm »
Dave can knock my theory off in 10 minutes if he wanted too.

He could knock it off in about 2 seconds by probing the bench.

With the obvious care over not putting the resistance of his hand in circuit, the contact repeatability tests, and checking the bandolier tape had no effect in the previous vblog I would be amazed if he didn't do that.

As for offset away from the mean it is all speculation. We don't know how they are manufactured. For all we know a few thousand more resistors and the offset might be the other side, we don't know another batch of philips resistors wouldn't be offset as much as the onehunglows.

The missing peak in the Gaussian distribution could indicate that two (or more) resistor manufacturing machines feed a single bandoliering machine.

The only dodgy thing I thought I saw was the -1% outliers in the distribution plot didn't seem to appear in the scatter plot.
 

Offline amspire

  • Super Contributor
  • ***
  • Posts: 3632
  • Country: au
Re: Gaussian resistors redux.
« Reply #11 on: November 13, 2011, 12:49:01 pm »
Dave can knock my theory off in 10 minutes if he wanted too.

He could knock it off in about 2 seconds by probing the bench.

I will predict that the meter will show OL for resistance.
Quote

With the obvious care over not putting the resistance of his hand in circuit, the contact repeatability tests, and checking the bandolier tape had no effect in the previous vblog I would be amazed if he didn't do that.

As for offset away from the mean it is all speculation. We don't know how they are manufactured. For all we know a few thousand more resistors and the offset might be the other side, we don't know another batch of philips resistors wouldn't be offset as much as the onehunglows.

The missing peak in the Gaussian distribution could indicate that two (or more) resistor manufacturing machines feed a single bandoliering machine.
That is a good theory. It would explain the results exactly. Like it! Also if that were the case then both machines would have been set a little low.  Perhaps they are deliberately making the values low for the reason I mentioned before - namely that metal film resistors characteristically drift high when run at elevated temperatures. It is better to make resistors that are low, then to make resistors that are high.
Quote

The only dodgy thing I thought I saw was the -1% outliers in the distribution plot didn't seem to appear in the scatter plot.

That is the only issue of concern in the whole test in regards to quality of the Chinese resistors.  If you were comparing Philips to the Chinese resistors in terms of quality, you would have to give the nod to Philips as they had no resistors near the limit. This might be nothing to do with manufacturing quality, but simply a question of quality control policy. Philips may reject any part near the limits as a matter of policy. The Chinese company may accept any part within spec as a matter of policy. Both would be valid policies.

Richard
 

Online EEVblog

  • Administrator
  • *****
  • Posts: 28915
  • Country: au
    • EEVblog
Re: Gaussian resistors redux.
« Reply #12 on: November 13, 2011, 09:44:15 pm »
Anyway, like I said I don't want to start an argument but I do think the skew does need to be explained. A quick run along the old  resistors with the new test setup or with the old leads on the new resistors would shed some light. You wouldn't need to plot the results or do very many measurements to squash my theory. Or confirm it. What do you have to lose?

I can assure you, it makes absolutely no difference.
The HP meter reads the same, with the resistor in or out of the bandoleer, on or off the anti-static mat. The results are correct and real.
FYI these anti-static mats are static dissipative to the order of many gigaohms at those distances (yes, I've measured it), and I did actually try to piece through it with my jig and was not able to penetrate through to the conductive backing of course, these are very tough mats.
Your theory is squashed, sorry. I should have shot video of these tests in the first place to avoid any confusion, but I thought I had mentioned or tested the bench mat in a previous video somewhere?
I included the 0.01% test resistor to show that the meter was spot on absolute.

Dave.
« Last Edit: November 13, 2011, 09:49:11 pm by EEVblog »
 

Online EEVblog

  • Administrator
  • *****
  • Posts: 28915
  • Country: au
    • EEVblog
Re: Gaussian resistors redux.
« Reply #13 on: November 13, 2011, 10:25:38 pm »
This plotting data lark is actually pretty interesting though. The results are not at all what I would have expected. Not with the OHL brand anyway.

I was expecting maybe a small offset and a much wider spread, out toward the limits.
Occasional glaces at the 5% data in fact show very similar results to the 1% ones! With the 5% ones I was hoping to see the mythical "double hump" where are the 1% ones have been picked out!  ;D

Quote
If you get the chance to find out more about how resistors are made I think that would be quite an interesting topic to present. Not much resistor manufacture in Australia i expect. None probably.

Correct, zero.
Interesting, but probably quite different between manufacturers.

Quote
Can you tell me what type of mat you use. I think I ought to get one.

http://www.chemtools.com.au/showroom/index.php/component/content/article/133-bench-mat-rolls-a-kits/163-glass-and-enclosure-cleaner
http://www.oritech.com.au/productDetail.aspx?productID=20508

Not cheap stuff, usually you have to buy it by the roll.

Dave.
 

Offline ivan747

  • Super Contributor
  • ***
  • Posts: 2010
  • Country: do
Re: Gaussian resistors redux.
« Reply #14 on: November 14, 2011, 01:37:16 pm »
I included the 0.01% test resistor to show that the meter was spot on absolute.

Dave.

A VERY expensive 0.01% resistor:
http://search.digikey.com/us/en/products/Y07851K00000T9L/Y0785-1.0KA-ND/2609889
Possibly one of the most expensive electronic components in terms of price/weight and size ratio, I think. The winner of that category one is probably some custom chip or something ridiculously precise.

Store that resistor quickly before it is swept away with the dust!
« Last Edit: November 14, 2011, 01:43:11 pm by ivan747 »
Nothing like the smell of rosin core solder in the morning.
"Could you not use some of that crowdfunded $1.5 million to hire a graphic designer who understands perspective?" -Delta
"A soldering station I bought once had a sticker on it that said, I shit you not, 'QENUINE'." -c4757p
 

Online EEVblog

  • Administrator
  • *****
  • Posts: 28915
  • Country: au
    • EEVblog
Re: Gaussian resistors redux.
« Reply #15 on: November 14, 2011, 01:44:53 pm »
I included the 0.01% test resistor to show that the meter was spot on absolute.

Dave.

A VERY expensive 0.01% resistor:
http://search.digikey.com/us/en/products/Y07851K00000T9L/Y0785-1.0KA-ND/2609889
Store that quickly before it is swept with all the dust!

I have an even more expensive 0.005% one too. I put both in a box with a nice label and binding posts.
I now have a neat spot reference checker.

Dave.
 

Offline grummund

  • Newbie
  • Posts: 3
Re: Gaussian resistors redux.
« Reply #16 on: November 14, 2011, 10:06:06 pm »
The HP meter reads the same
Cool.   8)

Thanks, that was a really interesting and instructive test.  Paralleling up say 10 x 1% resistors to achieve an average higher tolerance as you mentioned in the blog is not something I've ever attempted but it's always been an idea in the back of my mind - now I'm not so sure!

One thing that I'm curious about though: in the initial setup trialling of the Agilent the meter readings looked to me roughly centred around 1k, the highest I noticed was 1.0042k.  That's a 4.2 ohms offset (assuming the particular resistor in question was one at the top end of the distribution) which would make for a whopping error on ohms range if as you say the delta measurement is held across auto-ranging.
 

Online EEVblog

  • Administrator
  • *****
  • Posts: 28915
  • Country: au
    • EEVblog
Re: Gaussian resistors redux.
« Reply #17 on: November 14, 2011, 10:23:09 pm »
Thanks, that was a really interesting and instructive test.  Paralleling up say 10 x 1% resistors to achieve an average higher tolerance as you mentioned in the blog is not something I've ever attempted but it's always been an idea in the back of my mind - now I'm not so sure!

One thing that I'm curious about though: in the initial setup trialling of the Agilent the meter readings looked to me roughly centred around 1k, the highest I noticed was 1.0042k.  That's a 4.2 ohms offset (assuming the particular resistor in question was one at the top end of the distribution) which would make for a whopping error on ohms range if as you say the delta measurement is held across auto-ranging.

I used the Philips resistors from the previous blog in the initial setup video.

Dave.
 

Offline samgab

  • Frequent Contributor
  • **
  • Posts: 423
  • Country: nz
Re: Gaussian resistors redux.
« Reply #18 on: November 14, 2011, 11:26:30 pm »
Have you had any joy sussing out that data upload issue, Dave?
 

Offline Zad

  • Super Contributor
  • ***
  • Posts: 1013
  • Country: gb
    • Digital Wizardry, Analogue Alchemy, Software Sorcery
Re: Gaussian resistors redux.
« Reply #19 on: November 15, 2011, 06:03:25 am »
Having become the proud owner of a Tektronix DMM4020, I thought I would try a few tests of my own over the weekend. Not as extensive and structured as Dave's but nevertheless I got some interesting data. Unfortunately, all my 1% and better resistors are SMD and a real pain to test, which is why I only sampled 50 or so of them.

Without exception, every one read below the nominal value, all within half the stated tolerance. The 1 remaining 49.9 Ohm 0.5% 50ppm resistor I have, was just 0.005 Ohms low. It may be that, in-circuit with a good soldered contact, it was even closer than that. Of the 100 Ohm, 0.1% resistors I have (10) all were within 0.02%. Now I am the first to admit that a small sample size of random resistors is not massively thorough but, combined with Dave's data so far, I have some useful provisional conclusions.

  • It seems that new resistors are manufactured to err on the side of low resistance .
  • I had also thought that resistors would average around the marked value. They don't.
  • It seems from the first video that resistors age to higher values.
  • It also seems that resistors that are higher than the nominal value are either scrapped or trimmed to the next value.
  • I had thought that resistors made in the same batch would have very similar values. They don't.
  • A good proportion of nominally poor tolerance resistors can have very close tolerance indeed.


Offline BravoV

  • Super Contributor
  • ***
  • Posts: 5855
  • Country: 00
Re: Gaussian resistors redux.
« Reply #20 on: November 15, 2011, 06:13:05 am »
..<snip>..

  • It seems that new resistors are manufactured to err on the side of low resistance .



Any chance you pick few of them, and run current at their rated wattage or half of it, and leave it for a while, maybe a day or two and see if they're drifted "after" the burn in ? Really curious if they're going to drift higher after that ?

Offline robrenz

  • Super Contributor
  • ***
  • Posts: 3035
  • Country: us
  • Real Machinist, Wannabe EE
Re: Gaussian resistors redux.
« Reply #21 on: November 15, 2011, 06:43:08 am »
Zad,  My following comments are meant to be helpful not critical.  I am not bashing your meter, I have a new 8846A which is the brother of your meter with 1 more digit.  I am just amazed at how much less accurate on resistance these meters are than you would be led to think.  If you are taking measurements at the lower end of a scale you need to do the math to see what the real accuracy is.  Your example of the 49.9 Ohm resistor will have a possible one year error of +/-  0.023 Ohm which is +/- .046%.  It gets worse as you get lower in the range.  A one Ohm measurement  will have a possible one year error of +/-  0.0083 Ohm which is +/- .83%.  A .010 Ohm measurement  will have a possible one year error of +/-  0.00803 Ohm which is +/- 80.0%.  The reason for this is the +/-% of range that gets added to the error and this gets ugly as you get to the bottom of a scale.  like mine, your meter is probably way more accurate than the guaranteed specs as shown in my calibration data that came with the meter. This info from fluke explains why  http://support.fluke.com/calibration-sales/Download/Asset/2547797_6200_ENG_A_W.PDF

alm

  • Guest
Re: Gaussian resistors redux.
« Reply #22 on: November 15, 2011, 10:05:05 am »
Yes, accuracy sucks at the low end of the range, which is why you usually pick the lowest possible range. Picking resistors that are near the top end of the range (eg. 100 ohm) would improve the accuracy.
 

Online EEVblog

  • Administrator
  • *****
  • Posts: 28915
  • Country: au
    • EEVblog
Re: Gaussian resistors redux.
« Reply #23 on: November 15, 2011, 01:09:35 pm »
Having become the proud owner of a Tektronix DMM4020, I thought I would try a few tests of my own over the weekend. Not as extensive and structured as Dave's but nevertheless I got some interesting data. Unfortunately, all my 1% and better resistors are SMD and a real pain to test, which is why I only sampled 50 or so of them.

Without exception, every one read below the nominal value, all within half the stated tolerance. The 1 remaining 49.9 Ohm 0.5% 50ppm resistor I have, was just 0.005 Ohms low. It may be that, in-circuit with a good soldered contact, it was even closer than that. Of the 100 Ohm, 0.1% resistors I have (10) all were within 0.02%. Now I am the first to admit that a small sample size of random resistors is not massively thorough but, combined with Dave's data so far, I have some useful provisional conclusions.

  • It seems that new resistors are manufactured to err on the side of low resistance .
  • I had also thought that resistors would average around the marked value. They don't.
  • It seems from the first video that resistors age to higher values.
  • It also seems that resistors that are higher than the nominal value are either scrapped or trimmed to the next value.
  • I had thought that resistors made in the same batch would have very similar values. They don't.
  • A good proportion of nominally poor tolerance resistors can have very close tolerance indeed.

Hmm, interesting confirmation of my results.
Still doesn't make much sense to me though!

Dave.
 

Offline amspire

  • Super Contributor
  • ***
  • Posts: 3632
  • Country: au
Re: Gaussian resistors redux.
« Reply #24 on: November 15, 2011, 01:49:33 pm »
Having become the proud owner of a Tektronix DMM4020, I thought I would try a few tests of my own over the weekend. Not as extensive and structured as Dave's but nevertheless I got some interesting data. Unfortunately, all my 1% and better resistors are SMD and a real pain to test, which is why I only sampled 50 or so of them.

Without exception, every one read below the nominal value, all within half the stated tolerance. The 1 remaining 49.9 Ohm 0.5% 50ppm resistor I have, was just 0.005 Ohms low. It may be that, in-circuit with a good soldered contact, it was even closer than that. Of the 100 Ohm, 0.1% resistors I have (10) all were within 0.02%. Now I am the first to admit that a small sample size of random resistors is not massively thorough but, combined with Dave's data so far, I have some useful provisional conclusions.

  • It seems that new resistors are manufactured to err on the side of low resistance .
  • I had also thought that resistors would average around the marked value. They don't.
  • It seems from the first video that resistors age to higher values.
  • It also seems that resistors that are higher than the nominal value are either scrapped or trimmed to the next value.
  • I had thought that resistors made in the same batch would have very similar values. They don't.
  • A good proportion of nominally poor tolerance resistors can have very close tolerance indeed.

Hmm, interesting confirmation of my results.
Still doesn't make much sense to me though!

Dave.

As I mentioned previously, a chronic problem with thin film resistors like the metal film resistors is the film surface can oxidize, particularly when run hot. This always causes an increase in resistance as the film is only a few hundred Angstroms thick anyway.

I think metal film resistors vary rarely decrease in resistance with time, so making them a bit low in value is a good tactic. Has anyone ever seen metal film resistors decrease as they get older?

Thick film resistors also tend to rise in resistance over time, but for a different reason - thermal cycling degrades the connections between particles in the film.

Metal foil and wire don't have these chronic problems so they are potentially more stable. Resistance wire does have its own issues - avoiding stresses in the wire, the temperature coefficient of the resistor former, and the quality of the welding at the ends. Stresses in resistance wire can either result in crystallization in the metal increasing resistance, or the metal relaxes over time and the resistance decreases. Oxidation is also an issue with wire and foil, but you are usually talking parts per million or less.

Richard
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf