I cracked the mystery of Chinese precision jewelry scales

[magic]

There is a lot of cheap digital scales out there, typically small and battery powered. Basic ones have 3~4 digits, 0.1g or 0.01g resolution at best and cost peanuts, but for some $40 it's possible to get scales with advertised 1mg resolution and 50~200g capacity, such as this "One Hung Low 8068-series Professional Digital Jewelry Scale" piece of sh!t I had a pleasure of testing today.



Amazingly, this scale collects a lot of positive reviews on auction sites, including people saying they tested it with calibration weights and/or compared against professional lab scales.
Well, I can only hope that they got paid for those reviews because if they did it for free they have to be irredeemable idiots

It's not hard to notice that something is off. The scale ships with a calibration weight included and sure enough, calibrating the scale and weighting the calibration weight gives 100.000g as expected, as many times as you want. But then you take it off, put some random junk and suddenly re-weighting the junk gives results that vary by 2~4mg That's if you are lucky and it's a light junk, because heavy junk (like 160g) may vary by 13mg.

Actually, 0.08% is not a terrible result, but how is the 100g cal weight always coming bang on? I guessed the answer quite fast, and that's what really pissed me off. Yes, they cheat and the results are rounded to nearest full grams. More detailed testing near 1g revealed that results which the scale believes to be ±10mg away are shown as 1.000g, some results further still are displayed as 0998~1.002g, near 1.030g it flips randomly between honest reading and 1.003g and further it works normally. Similar behavior was observed near 2g and 3g and that's where I stopped caring about this turd anymore.

Below is a linearity test I conducted by weighting increasing number of identical metal washers, 50 pieces ~33mg each. Blue is the reading, red is deviation from linear fit multiplied by 100. The washers are not calibration weights, they may not be identical and this may contribute to some of the vertical "noise". We see that measurement-to-measurement noise is 5mg peak-peak (including contribution by washers). There is some waviness at the beginning, which could be nonlinearity or - perhaps more likely - drift. (Drift could be cancelled by re-taring the scale between each measurement, which I didn't do.) And then there is a 20mg "glitch" near 1000mg, because the scale assumed that I am testing it with a 1.000g calibration weight and didn't believe its own sensor, which in itself actually seems far more accurate than 20mg

Can't wait for DMM manufacturers to catch the idea, if they haven't already
(Yes, I know that Zotek hides offset voltage of their meters by replacing low readings with zero, I want more )

[Whales]

Hmm.  I've noticed some of my kitchen electronic scales suspiciously landing on exact .00 numbers quite a few times.  I wonder if it's a common chipset feature -- worth checking scale balance datasheets.

[Ed.Kloonk]

 

[magic]

I forgot to mention that the cheating is somewhat stateful. Sometimes when I weighted 1.005~1.015g in multiple pieces, the scale showed 1.000g and then as I removed the weights, it reported the differences correctly and subtracted them from the reading until it went negative and remained below zero with an empty platform

I also noticed that the reading tends to change by 3 counts more often than by any other number. It occurred to me that 3 is what happens when 200000mg is divided by 65536 counts of a 16 bit ADC. I'm still not sure where the other variation comes from, maybe it's oversampling/interpolation or some drift compensation, or maybe I'm wrong about 16 bits.

I would love to see what's inside, but I'm not sure how to dismantle it safely and I don't feel like dumping $40 down the drain - it's going back to the seller.

[magic]

Hmm.  I've noticed some of my kitchen electronic scales suspiciously landing on exact .00 numbers quite a few times.  I wonder if it's a common chipset feature -- worth checking scale balance datasheets.

There is probably some firmware there. I doubt that any documentation could be found publicly and in English but you can try.

[SeanB]

AS I actually have a set of reference mass pieces, I have done the check and verify step on a number of those small units. Those with a calibration mass will read exactly, because they have been calibrated with that mass, using a 2 point calibration, so setting the zero point, and full scale. Not going to say much for the linearity, as the better units will have at least a 3 point calibration step, with a zero, half scale and full scale calibration point, allowing the internal microcontroller to do a best fit curve, for the conversion from ADC counts to display reading.

Have seen pretty good linearity, at least to within 1%, on them, with just that 2 point calibration, though best is to not rely on that, and instead use a check mass at close to your desired mass.

The same basic load cell, the same signal conditioning blob is used for a lot of them, with full range ranging from 40g to 1kg, but the exact same unit, just the calibration is different between them. Sadly no real way to calibrate them, you can only do the same as the original manufacturer, but you can at least use a certified mass piece to do it, instead of the cheap close to mass one they give.

Bought a set of those from China, and took them, figured out how they adjusted it (undo the top cap and drill down to lower, and add lead to raise) so I could get them to agree closer to the recently calibrated mass meter, accurate to 0.5mg. Then sent them in to be certified. Cost more than the masspieces cost originally, but well worth it, as I then knew, down to the milligram, the exact mass, and traceable via 3 hops to the standard kilogram.

[HighVoltage]

What great idea!
To make it more real, they should have the least significant digit change a "little"

To be accurate in 1 mg is not that easy.
I have repaired lots of mg scales and it is amazing what effort the real manufacturer put in scales to make them reliable.
My most precision scale goes down to 10 ug and it is mind twisting what is needed to make that happen.

What surprises me is that so many fall for these cheap Chinese scales.

[gamalot]

Multimeter manufacturers have done something similar a long time ago, the famous Agilent/Keysight multimeters will show zero at very low AC currents. I also know of some scale manufacturers that track zero drift and "cancel" it. Of course, none of them are as smart as this jewelry scale maker! 

[Gyro]

What great idea!
To make it more real, they should have the least significant digit change a "little"

To be accurate in 1 mg is not that easy.
I have repaired lots of mg scales and it is amazing what effort the real manufacturer put in scales to make them reliable.
My most precision scale goes down to 10 ug and it is mind twisting what is needed to make that happen.

What surprises me is that so many fall for these cheap Chinese scales.

My 1ug resolution [Edit: non-Chinese] scales. 

[HighVoltage]

My 1ug resolution [Edit: non-Chinese] scales. 

WOW!
How old is that one ?
And who is the manufacturer?
That must be a very special 1ug scale !

My 10ug scale is about 10 years old from Mettler Toledo, Made in Switzerland.

[free_electron]

just say no and buy a  sartorius or mettler scale

[Gyro]

My 1ug resolution [Edit: non-Chinese] scales. 

WOW!
How old is that one ?
And who is the manufacturer?
That must be a very special 1ug scale !

My 10ug scale is about 10 years old from Mettler Toledo, Made in Switzerland.

It's a Stanton Instruments MC1A  http://stanton-instruments.co.uk/page19.html

It pre-dates the 1960 catalogue, but irrc, it's 1950s (there's a datasheet in the Catalogues section of the above web page). I still have the original calibration/periodic inspection card somewhere, it originated from the Atomic Energy Research Authority Establishment at Harwell, but I picked it up in a 'treasures' shop around 40 years ago. I don't like to think what it was used to weigh, but I went over it very thoroughly with a Geiger counter.

I should probably clarify that the page says that the early models were only good to 2-3ug resolution in practice, probably also depending on the remaining sharpness of the Agate knife edges and synthetic Sapphire planes.


P.S. That site is an interesting resource, I only found it today!

[AVGresponding]

What great idea!
To make it more real, they should have the least significant digit change a "little"

To be accurate in 1 mg is not that easy.
I have repaired lots of mg scales and it is amazing what effort the real manufacturer put in scales to make them reliable.
My most precision scale goes down to 10 ug and it is mind twisting what is needed to make that happen.

What surprises me is that so many fall for these cheap Chinese scales.

Oooh, do you have any repair threads relating to these? I have a very poorly A&D GX-600, it would be useful to read something like that!

[Bud]

There you go, a look inside a no-shit precision scale 

https://youtu.be/eKsEmukmByI

Edit: The video even includes the mandatory Lab cat, though with a twist 

[magic]

Those with a calibration mass will read exactly, because they have been calibrated with that mass, using a 2 point calibration, so setting the zero point, and full scale.

Doubly so if they also round 99.990g to 100.000g

Not going to say much for the linearity, as the better units will have at least a 3 point calibration step, with a zero, half scale and full scale calibration point, allowing the internal microcontroller to do a best fit curve, for the conversion from ADC counts to display reading.

I have seen them too and wondered if it's because they are better or because they are worse and need it more. Maybe I should construct some ~50g test weights and check the midpoint INL of mine, since I haven't returned it yet.

That being said, it's not very important to me. I'm currently looking for good linearity below 10g, for ratiometric measurements on low quantities. I picked the 200g version figuring it may come in handy in the future.

Similar behavior was observed near 2g and 3g and that's where I stopped caring about this turd anymore.

Never say never

New linearity measurement.
The extra 1980mg is a container which enabled easy taring between measurements.
Each measurement was repeated 3 times and the median was used.
Noise is similar but drift is gone, only some unexplained issue at 2~5 and 11.
Good linearity except near 2000mg and 3000mg, as expected

[Zucca]

Bastards! They will burn in hell for eternity.

[SiliconWizard]

Oh well. Cheating for calibration or testing purposes... who does that...
Remember VW?

[Circlotron]

I've got an old Sabtronics frequency counter that does something similar. After running both it and the new-at-the-time Rigol DG1022Z function generator (1 ppm accuracy) for 24 hours I put 10 megs into the counter. It displayed 10000000.0 Hz. Dang! That's good I thought. Wasn't drifting either. Then I dropped it to 9 megs and it showed maybe 7 or 8 Hz low. Similar with 8 megs. Basically the error was  a fixed proportion of the measured frequency, as you might expect. But not at 10 megs...

Dieselgate.
Partygate.
Countergate.

https://www.radioexperimenter.us/rm-1981-05/to-shield-1.html

Edit -> Oh yeah. SiliconWizard beat me to it.

[Bud]

Drug dealers are not going to be happy reading this thread 

[HighVoltage]

What great idea!
To make it more real, they should have the least significant digit change a "little"

To be accurate in 1 mg is not that easy.
I have repaired lots of mg scales and it is amazing what effort the real manufacturer put in scales to make them reliable.
My most precision scale goes down to 10 ug and it is mind twisting what is needed to make that happen.

What surprises me is that so many fall for these cheap Chinese scales.

Oooh, do you have any repair threads relating to these? I have a very poorly A&D GX-600, it would be useful to read something like that!

Sorry to say, that I did not post any scale repair threads
I probably will repair more in the future and then post it on eevblog

[Bicurico]

When I started to read OP, I immediatly thought that the included sample/caliber had a magnet inside...

perhaps I should sell that idea to chinese manufacturer.

[SilverSolder]

Wow, that's lame!!   

I guess you could get around this problem by adding a small, accurate mass to the scale if measuring near a round number....

[AVGresponding]

When I started to read OP, I immediatly thought that the included sample/caliber had a magnet inside...

perhaps I should sell that idea to chinese manufacturer.

  Me too!

Though for extra calibration accuracy, the manufacturer could make several different reference weights, each with a different rfid tag in...   

[SiliconWizard]

Drug dealers are not going to be happy reading this thread 

Maybe the "hobbyist" ones. 

The "serious" ones probably make way enough money to buy decent scales, not cheap crap off Aliexpress.

[magic]

This damn scale is either very good at cheating or it is actually very good except for the cheating

I wanted to do the 0-100g INL test. I found four AA cells which weight about 24g each, numbered them 0,2,4,9 for certain reason and weighed all four of them and then divided them in different sets of 1, 2 or 3 cells and weighed the sets alternately.

This was to hopefully confuse the scale should it try to remember previous results and assume that it is being tested for linearity and lie
Yes, I'm paranoid. Not sure if sufficiently?

The results are below. Each measurement was repeated five times allowing the scale to go back to zero in between. Lines starting with -> are median results. Repeatability was fairly good today and I found that linearity error is virtually nonexistent. One of the 50+50=100 comparisons is spot on, the other is off by 2mg. The 25+25+25+25=100 is off by 5mg, or we could say that each cell was undermeasured by 1.25mg because 100g is supposed to be calibrated. It could easily be quantization error. The 25+25=50 and 25+25+25=75 sums are off by similar margins. All errors are less than 0.01%.
 

Code: [Select]

0249
95.448
95.448
95.451
95.450
95.451
-> 95.450

09/24
47.787
47.662
47.790
47.661
47.787
47.662
47.788
47.661
47.790
47.662
-> 47.788 / 47.662

024/9
71.564
23.886
71.561
23.887
71.563
23.886
71.563
23.890
71.562
23.886
-> 71.563 / 23.887

049/2
71.629
23.820
71.630
23.818
71.631
23.816
71.630
23.819
71.630
23.818
-> 72.630 / 23.818

0/4
23.897
23.843
23.899
23.841
23.898
23.842
23.899
23.841
23.896
23.843
-> 23.898 / 23.842

04/29
47.739
47.707
47.741
47.707
47.742
47.708
47.741
47.707
47.742
47.710
-> 47.741 / 47.707