Load cell calibration for multiple sensors

[hozone]

Hello,

I'm building a weight scale. Selected load cells are NA1, 40Kg (or 50Kg, yet to decide).

I've not yet selected the load cell amplifier, I'm thinking about the common HX711 or the ADS1231.
Recommended excitation voltage is between 5V and 12V, so both the two IC above will work.
The NA1 sensitivity is 2.0mV/V +-10%, find full specification below.

For my project it will be difficult to calibrate each cell, cause the weight is not equally balanced on that cells, it's some kind of "particular" weight scale.
I don't need that much accuracy. Let's say I need 100gr per cell accuracy.
Supposing to select the HX711 amplifier, a 40Kg cell (sensitivity is 2mV), set the gain at 128x and just keeping 14 bits out of the 24 to remove some noise (see note below), I should get 40000/(2^14*0.128*2) almost 10gr resolution, and that fits my need.
By datasheet rated output (sensitivity) and the other error involved (non-linearity and so on), should not affect the cell reading that much.

So my question is, can I skip calibration for each individual cell or am I missing something? I'm wondering if I can just set the conversion value between the raw IC output (let's suppose the HX711) according to the sensitivity of the cell, and keep this same value for all the difference cells.

Thanks!

Note about the HX711 there's an interesting noise analysis here: https://forum.arduino.cc/t/hx711-noise-free-resolution-tests/351959

[Kean]

Not sure I fully understand the question.  You always have to calibrate, but if you have multiple load cells they should be identical and connected in parallel (or via a summing box).  You want to try balance the weight as much as possible across the load cells, and the platform should sit on pivots (e.g rounded) so that it spreads the load.

I started using the HX711, but moved my design to the AD1232.  As I mention in another topic on HX711:

I've worked with the HX711 in the past, but I found it a bit annoying to work with.  If you don't get the read timing just right you can get spurious readings.

Apart from breakout PCBs and modules (i.e. Sparkfun, M5Stack - both available on Digikey) you can source the ICs from many random sellers on AliExpress.
If you are going to want a commercial relationship, then contact Avia Semiconductor directly. http://en.aviaic.com/detail/730856.html

There is a good Arduino library for the HX711 and lots of helpful info at https://github.com/bogde/HX711/

Personally, I switched to the Texas Instruments ADS1232 (back in 2015) and it has been a great solution for my wildlife monitoring systems.  I use this mostly for dynamic weighing of penguins, but also some other projects.

With a square platform and 4 x 50kg load cells, I can get ENOB of 14 or 15, and sometimes 16 or 17 with the ADS1232.  Some of my PCBs have proven to be a bit noisier than others, but it is also affected by mechanical assembly, installation, and possibly wiring differences.

[Kleinstein]

There is quite some uncertainty in the overall scale factor (10% in the datasheet for the load cell). Withput a calibration one would get this rather large uncertainty.

For the noise it is not so much the noise of the ADC, but more the amplifier noise. It is also not a fixed resolution limit - more averaging (slower readig) gives better resolution.

[lk.dgironi]

@Kean
I have multiple load cells, all the same model/weight.
I can not go in deeper details here due to confidentiality of this project, but I can simplify it.
Let's suppose to have two cells on the same scale mounting, a weight of 50Kg may be distributed as 20Kg on cell #1 and 30Kg on cell #2. I need to measure the weight on each cell to perform some math on the geometry of the project.
Due to the way the scale it's built it will be difficult to calibrate each cell once mounted in place. So I'm wondering if the same conversion factor should be used.
Luckily I don't need that much accuracy.

@Kean
Thank you for the ADS1232 suggestion. I'm planning to use REF195 as reference. Do you have an shareable schematic for your implementation?

@Kleinstein
you are right, 10% are almost 5Kg on full scale, that is too much. If so, maybe calibrating each cell before mounting can be a solution? (please read @Kean above for completeness of project contraints)
Correct me if I'm wrong:
my not calibrated weight conversion formula is
weight = capacity*Vout/(sensitivity*Vexc)
with a capacity of 50Kg, sensitivity of 2mV/V, excitation (Vexc) of 5V my output voltage is Vout
if Vout = 10mV, weight is 50Kg
if Vout = 1mV, weight is 10Kg
so, 10% means 2mV/V can be 2.2mV/V or 1.8mV/V, let's suppose the upper limit
if real-sensitivity is 2.2mV/V but I'm using 2mV/V sensitivity and i'll get 11mV out of the cell, so 55Kg instead of 50Kg (which is the 10% error of the datasheet)

[wraper]

You either need matched load cells or amplify/digitize each load cell separately with individual calibration. In commercially produced scales with 4 load cells I've normally seen 4 matched half-bridge load cells connected together.

[Kean]

@Kean
I have multiple load cells, all the same model/weight.
I can not go in deeper details here due to confidentiality of this project, but I can simplify it.
Let's suppose to have two cells on the same scale mounting, a weight of 50Kg may be distributed as 20Kg on cell #1 and 30Kg on cell #2. I need to measure the weight on each cell to perform some math on the geometry of the project.
Due to the way the scale it's built it will be difficult to calibrate each cell once mounted in place. So I'm wondering if the same conversion factor should be used.
Luckily I don't need that much accuracy.

As mentioned, if they are decent quality matched load cells you can just wire the load cells in parallel.  They still form a bridge circuit, and this will just increase the excitation current draw.  The sense signal voltages will be summed proportionally.

You can add a variable resistor into the excitation legs if you need to trim the cells for mismatch or imbalance.  Most summing boxes include this.  One problem I've had with this is in outdoor use, using poor quality potentiometers can lead to significant thermal instability.

Edit to add:  Also make sure you have a way to isolate the load cells for diagnostics.  Summing boxes with terminal blocks help here.

@Kean
Thank you for the ADS1232 suggestion. I'm planning to use REF195 as reference. Do you have an shareable schematic for your implementation?

I mostly used the ADS1232 reference design plus an ATmega328 and RS232 interface.  Power supply via a 12V to 6.5V buck converter, then two 5V LDO regulators - one for analog and one for digital.

@Kleinstein
you are right, 10% are almost 5Kg on full scale, that is too much. If so, maybe calibrating each cell before mounting can be a solution? (please read @Kean above for completeness of project contraints)
Correct me if I'm wrong:
my not calibrated weight conversion formula is
weight = capacity*Vout/(sensitivity*Vexc)
with a capacity of 50Kg, sensitivity of 2mV/V, excitation (Vexc) of 5V my output voltage is Vout
if Vout = 10mV, weight is 50Kg
if Vout = 1mV, weight is 10Kg
so, 10% means 2mV/V can be 2.2mV/V or 1.8mV/V, let's suppose the upper limit
if real-sensitivity is 2.2mV/V but I'm using 2mV/V sensitivity and i'll get 11mV out of the cell, so 55Kg instead of 50Kg (which is the 10% error of the datasheet)

It is definitely worth doing this matchmath as a check of the fundamentals, but just be clear that you cannot use those calculations for getting the final weight.  You probably know this, but I'll just make it clear...

You must zero/tare the platform with all the additional mounting hardware and platform, i.e. measure the offset.  Then you need a calibrated weight to place on the platform and calculate the conversion factor/multiplier.

You can then run checks with the weight placed in various positions, to check for imbalance errors.  Finally you can take a series of rapid sequential measurements to check for peak/RMS noise and calculate ENOB.  The ADS1232 reference design provides some guidance on this.

[Kleinstein]

Having a load shared by multiple load cells can indeed be an issue for the calibration. Ideally one would create a loading (possibly from the other side) that only effects a single load cell.
With a little more math one should still be able to get get at least a reasonable good calibration (maybe a factor 2 or 3 less accurate than separate cal.) from using somerhing like the same weight, just positioned differently. One gets the result as a conbination of multiple raw readings with the unknown cal factors. This would give a system of linear equations to solve, which is still relatively simple math. It just needs enough load cases (e.g. as many as load cells - more would allow an addttional check), that are also significant different enough.
This assues a seprate digital measurement and numerical summation. It gets tricky with an analog combination (e.g. split half bridges over to load parts).

The reference for should not be that important as the measurements should be done ratiometric. With secondary effects, like sensor heating it can still help to have a reasonble stable reference (maybe a 7805 could be good enough).

[hozone]

Thanks all.
Reading your comments makes me think the only solution will be calibrating each cell one by one.

Note on price constraints, availability optimization and accuracy. I would like to change the current load cells due to price availability mostly, but even accuracy. The actual project is using TE FX29 with i2c output, which is suppose to be factory calibrated. The full scale span vary is (on digital world) +- 700 counts, which is almost 5% (2.5Kg out of 50Kg ~500N one), that is a lot (if I've understand the datasheet)

You either need matched load cells or amplify/digitize each load cell separately with individual calibration...

Problem is that I need 3 cell for my geometry, and I've also price constraints, matched one cost a little more.

As mentioned, if they are decent quality matched load cells you can just wire the load cells in parallel. ...

They will not be matched, I've some price constraints too. For this reason I'll going to wire that as a full bridge stand alone.

It is definitely worth doing this match as a check of the fundamentals, ...

Yes, I know this. I've experimented with HX711 and cheap weight cells years ago. And I always had to calibrate each load cells, tare... and so on. Anyway thank you for the ADS1232 reference design tip!

Having a load shared by multiple load cells can indeed be an issue for the calibration...

I was thinking about having the math so solve the geometry problem for calibrating cells using just a single weight. But maybe it will be better and accurate to modify my hardware and make it capable of calibration for one single cells, building some "calibration" hardware tool.

[Kleinstein]

Loading single load cells (e.g. turn over the whole system and load from the other side) would likely be more accurate than the math for separating data from different load positoins. Another advantage is that one may be able to do an analog trim and than combine the load cells before the ADC.  For the trim it is only about the scale factor, not yet the offset. The offset part can handled with alread all combined.

[hozone]

As an alternative, maybe I can put a know weight in the center of gravity of my geometry, and then consider it equally distributed on the 3 load cell I have, is that right?
To make you understand better, let's image the device I'm building it's something like the wii balance board.

As for the TE FX29 accuracy of my previous post, are my though right? I mean, do I understand the "full scale span" of the datasheet well?

Thanks again

[Kleinstein]

With 3 points and knowing the center point well the assumption of equal load is probably OK. There is still some uncertainty in the positions.

A point to look at may also be lateral forces. The load cells may also react to latera forces to some degree - ideally they should not, but nobody is perfect.

I think the understanding the ful scale span is correct. It is normal to have some uncertainty in the sensitivity factor, though 10% is quite a bit. Chances are parts from the same batch are quite a bit closer together, even from not specially matched sets.

[hozone]

Thank you for all the clarification!

[wraper]

With 3 points and knowing the center point well the assumption of equal load is probably OK. There is still some uncertainty in the positions.

Only if all 3 of them are decently vertically aligned with a very flat plate bolted on them. Or if coupling to the load cells in not hard with sort of pivot points that would eradicate problems due to alignment.

[mendip_discovery]

Given the information you have been able to share, I think it would be best to remove the load cells and stack them on top of each other. Then apply known loads to them so that way you can get them to all be similar at the same time as accurate enough for the task.

I have used a calibrated load cell and a g-clamp to calibrate load cells which are a little bit of a challenge to balance weights onto.

Whenever it comes to scales/weighing systems I always recommend you get a test weight for regular checking, or at least a "gold" standard aka a scrap part that you can run through to check that it is all ok.

[David Hess]

Scales which use multiple load cells have a summing circuit with a separate gain adjustment for each load cell.  A test weight is moved around on the scale and the gain adjustments are trimmed so that the output is constant.

[hozone]

Thank you all for clarification.
I'm thinking about trying a design to calibrate each single cell by disassembling the top cover.

[Mori]

Not sure I fully understand the question.  You always have to calibrate, but if you have multiple load cells they should be identical and connected in parallel (or via a summing box).  You want to try balance the weight as much as possible across the load cells, and the platform should sit on pivots (e.g rounded) so that it spreads the load.

I started using the HX711, but moved my design to the AD1232.  As I mention in another topic on HX711:

I've worked with the HX711 in the past, but I found it a bit annoying to work with.  If you don't get the read timing just right you can get spurious readings.

Apart from breakout PCBs and modules (i.e. Sparkfun, M5Stack - both available on Digikey) you can source the ICs from many random sellers on AliExpress.
If you are going to want a commercial relationship, then contact Avia Semiconductor directly. http://en.aviaic.com/detail/730856.html

There is a good Arduino library for the HX711 and lots of helpful info at https://github.com/bogde/HX711/

Personally, I switched to the Texas Instruments ADS1232 (back in 2015) and it has been a great solution for my wildlife monitoring systems.  I use this mostly for dynamic weighing of penguins, but also some other projects.

With a square platform and 4 x 50kg load cells, I can get ENOB of 14 or 15, and sometimes 16 or 17 with the ADS1232.  Some of my PCBs have proven to be a bit noisier than others, but it is also affected by mechanical assembly, installation, and possibly wiring differences.

Do you consider that this would be environmentally stable and good. Please review its manufacturing process.

[David Hess]

With a square platform and 4 x 50kg load cells, I can get ENOB of 14 or 15, and sometimes 16 or 17 with the ADS1232.  Some of my PCBs have proven to be a bit noisier than others, but it is also affected by mechanical assembly, installation, and possibly wiring differences.

Do you consider that this would be environmentally stable and good. Please review its manufacturing process.

The best load cells are only good to about 0.02% so a resolution of 14 bits is often sufficient, and was not difficult to achieve even 20 years ago.  Even 16 or 18 bit was easy to achieve 20 years ago.