Development of a free Capacity Predicition, need your help

[Sean]

Wolf, did you ever manageto confirm how accurate the capacity tests you record are ?

 

[Wolf]

Wolf, did you ever manageto confirm how accurate the capacity tests you record are ?

Ahh yes the age old question how accurate are our testers anyway.
The answer to this question has many facets.
I will try to condense it as best as I can.
Most sub $70.00 Charger/Testers have a canned program to charge and discharge batteries with some selections available to the users.
These "user" settings are usually allowing a change in mA charging rate and that's it. It assumes that "most" 18650 batteries will test withan acceptable result within those parameters. We cannot change the cutoff mA of the charge cycle which I believe is set at ~80mA. We also cannot change the discharge voltage cutoff most of them have that burned in at ~2.8V.
So what does that mean?
These settings will work for a lot of batteries for reasonable accuracy to at least within ~5% mostly to the higher side.
Cells of different chemistries require different charging and discharging parameters. Hence the spec sheets that battery manufacturers provide that inform us as to how they testedthe cells and how they should be tested to achieve the results that they claim the cells have.
So for a cell (like theUR18650ZTA)that has a max of 4.35V and a cutoff of 3V with a "generic" tester can we get an accurate reading.
NO we can't, but we can get "close".
Another example is theICR18650-30B 4.35V max to 2.75 cut off will the "generic" tester ever show the true capacity of 2950mAh?
Probably not. But again close.
So I got the SKYRC MC3000 as you probably recall. With this unit I can custom tailor the charging and discharging profile according to the manufactures spec sheet down to every detail. I run a spot check on all my testers on every lot I test.
I use 3 testers as you know. 4 Foxnovo, 3 Opus, and 3 Liitokala. So for every 44 batteries that I run through I take 4random batteriesfrom
one of the tester makes,as in 4 from OPUSone time then the next time from Liitokala then from the Foxnovo and run them through the SKYRC to spot check. Usually maximum of~5%deviation to the higher side from the other charger/testers.
The Foxnovo in my findings is the closest the OPUS is slightly higherand the Liitokala is the highestwith capacity results compared to the SKYRC.
So to answer your question I believe my test results after cross checking and spot checking are within a 5% margin of error to the high side.

I am at this point with 2366 cells that meet my criteria for my powerwall build that means 434 more cells to go. With my IR pretest, that has been phonemalysuccessful in weeding out non performing cells, itshould take me another 10 charge and discharge cycles to get to the 2800 mark.

I will then build my first 200p pack add all the cells mAh results up and run a test on the pack with my iCharger. That should give us an indication on how close we are at the pack level.

I hope that answers your question.

Wolf

Oops I forgot to mention the Zanflair way off no good.

 

[Sean]

So to answer your question I believe my test results after cross checking and spot checking are within a 5% margin of error to the high side.

Good answer, thank you.

I was idly pondering the usefulness of the model, given the accuracy of the data being used - which in this instance is reasonably high, because you've taken steps to ensure it is.

Keep up the good work.

 

[drbacke]

@Wolf:
This looks very promising, 5% is not to much. I thought it would be more.

Wolf, did you ever manageto confirm how accurate the capacity tests you record are ?

In the Surrogate Model I builded, wolfs data gets an own correlation function and is used as low-fidelity data. So I can answer the question how accurate the measurement are in a more "surrogate model" way:
Actually the model rates the noise in wolfs data with a Standard Deviation of 166mAh.
In my database (which is much smaller and uses only one tester) the deviation is about 77 mAh. But maybe this value is not correct due to a small amount of data.
This means the model will have a minimum uncertainty of 77mAh SD for my own setup. It will be interesting to test it with other data than wolfs or mine.

And the question how good the quality of the measurements are, will be answered by the model itself. Thats the big advantage of statistical models like Co-Kriging. If you have noisy data, the model will find this fact and the prediction will get a high uncertainty (standard deviation). So if you have low quality data, the model will tell you that with a high predicted uncertainty.


But to answer the question, if the predicitons can help someone else wit another test setup:
Maybe it would be helpfull if someone else than wolf or me gives me testdata and I will post the predictions and we will see how good it performs.I don't know.

 

[Wolf]

drbacke

Interesting results with my data. Even at a LoFi Model the standard deviation is 166mAh and I assume that that is -83 to +83
Which would indicate a pretty good result basically a 3.32%plus or minus on a 2500mAh cell.

Now the question I have is are you taking the whole database of all cells as a source for this or are you filtering by partnumber?

I believe if you filter by partnumber and then run the LoFi Model on just those cellsthe results will be much better? Maybe? Less noise in the model as you are now just working on 1 cell type with a known mAh rating.

As far as anyone having a database such as I have worked on I don't know I haven't seen anyone post any results that they have.
Another 103 cells and I will have 5000 cells recorded.When I started this on 12/13/2018 and entered my first cell I never in my wildest dreams thought I would ever get to this number. Now I can't wait to get to the 10000 mark. Although it will slow down some as I do need to get my powerwall built rather than enter data all the time.
I also have some more testing projects I want to finish as I am backlogged on those.

Wolf

 

[Wolf]

And the question how good the quality of the measurements are, will be answered by the model itself. Thats the big advantage of statistical models like Co-Kriging. If you have noisy data, the model will find this fact and the prediction will get a high uncertainty (standard deviation). So if you have low quality data, the model will tell you that with a high predicted uncertainty.

Drbacke,

Any news to share with us? Anything I can help with?

Wolf

 

[completelycharged]

This might be worth having a look at, as I think if you can get the data this approach may give you a lot more of an easier approach and more fitting to NN data for cell differentiation.
https://secondlifestorage.com/t-Charge-and-discharge-cycle-data-help

Ths would limit the cell evaluation to a longer test cycle, but the end result would be a lot more accurate and with additional information.

If you take a charge and discharge cycle and then pulse the charge or discharge at different currents at 10 different capacity / voltage points through the cycle this will give a lot more information and and idea on the net chemical balance of the parts inside.

If a cell has been held at full charge for a couple of years it will chemically deteriaorate in a different way to a cell that has been held at a low state of charge but a high temperature. Both of which may still be usefull to operate at a certain voltage range where the deteriaorated chemicals do not matter as much (different elements in the electrolyte mix). That is just my thoery at the moment.

The thought is that if a cell is still useable for 500 cycles, juts at a different voltage range and these cells can be identified and grouped together in volume they are then useable. In isolation they are just another item to go to the dump/recycling.

My belief is that voltage, IR and cell are a rough guide, but not going to be reliable enough and then requires that all cells are able to be identified accurately AND they dont differentiate rejection cells from a good production lines that are then sold on the chinese market.....


The approach would be to take number of a given well known cells that have known varying capacities and the cells also have reliable test data for cycle life (hardly any do out of the 100's or cells on the market at the moment). Then cycle test each (pulse charge/discharge for varing IR evaluation under differeing currents) one for the capacity, map this to the remaining known cycle life, then add these values into the NN without the cycle test capacity figure.

Just some thoughts and ideas that may or may not help....

 

[drbacke]

And the question how good the quality of the measurements are, will be answered by the model itself. Thats the big advantage of statistical models like Co-Kriging. If you have noisy data, the model will find this fact and the prediction will get a high uncertainty (standard deviation). So if you have low quality data, the model will tell you that with a high predicted uncertainty.

Drbacke,

Any news to share with us? Anything I can help with?

Wolf

Hey Wolf, sorry I'm currently working on a cheap Raspi based battery monitoring system. So I forgot to post the new data, here it is. I think the results are good enuough to be helpfull:






However, the most interesting resultwere these guys here (Cell Name, Nominal mAh, pre Voltage, pre IR):
Sanyo UR18650FJ21000,08575,7
Sanyo UR18650FJ21000,07970,8
Sanyo UR18650FJ21000,07671,6
Sanyo UR18650FJ2100 0,07870,5


Here the predictions (the lowest of all predictions you can find in the whole table):
1402 mAh
1456[size=small]mAh
1448mAh
1460mAh[/size]

And now the really cool result:
It turns out that these cells were all heaters!

So the model seems to "know" that these cells are really bad.
Maybe it was just a coincidence, but I don't think so.
They had all a very low pre Voltage and for this type of cell a very high IR, so this makes sense.

 

[Wolf]

@drbacke
Interesting predictions and not to far off.
Good work on the model.
The predictions are close to what my tests show. Good stuff!!!



My cut off on those cells(UR18650FJ) is Optimal ?68m? and Marginal up to 76m?
There is only 1 thing I do different with any <1V cells that are within my acceptablem? readings.
I charge them in parallel at a CC/CV of 4.2V at 50mA per cell.
This usually takes a couple of days if not 3 or 4 before they come up to voltage.
If a cell gets remotely warm during the 50mA charge it is taken out.
These happen to be UR18650FMs and not a heater in the bunch "yet". all cells<58m?.



Once they hit 4.19 they will rest a bit and I will retake the IR and then into the charger/ testers they go.
I label them with a yellow font in my spreadsheet that way we know they were low voltage cells to begin with.
Once in the charger/testers charging the rest of the way to 4.2 at 500mA that is where we will see if there are any heaters.
Most of the time they behave very well. We will see.
I have now started to change the font colorof the "known" heaters that make it through the preliminary IR tests in the tested capacity to red.


As you can see I only have 4so far. All the other ones never made it this far. I do still have a boxload of ~300 ofthem of all various makes and models but mostly Sanyo.

How did you charge the UR18650FJ cells?

Wolf

 

[100kwh-hunter]

@wolf
As you know i keep a very close eye on your work.
But i can not help myself to ask the following question, its playing for a very long time in my mind:
Why did you never tried to make your own charger/discharger to get optimum results?
With your knowhow it must be a peace of a cake!
There...... finally got the guts to ask

By no means i want to offend people and i am in doubt if this is way off topic.
Sinds we must trust charger/dischargers?

Very good read!
Best

 

[Wolf]

@wolf
As you know i keep a very close eye on your work.

So that is the eye in the sky I keep seeing when I turn around.



I'm going to turn all my cameras off.

But i can not help myself to ask the following question, its playing for a very long time in my mind:
Why did you never tried to make your own charger/discharger to get optimum results?
With your knowhow it must be a peace of a cake!
There...... finally got the guts to ask

Had I known 8 months ago what I know now I probably would have. But alas I jumped in without forethought and bought a bunch of charger/testers. Little did I know what I was going to get involved in. Now that I am about 10 cells away from my 2800 cell goal I will devote more attention to putting something together like that. This weekend I put together a nice ZHIYU ZB206 board so I am working on more testing equipment.

By no means i want to offend people and i am in doubt if this is way off topic.
Sinds we must trust charger/dischargers?

Very good read!
Best

On my Vacation to Austria from the 15th of June to the 3rd of July to visit my mother I will immerse myself into Arduino code and RPi and play around with that. Notwithstanding the mountain climbing trips and general relaxing.
We will see what becomes of it.

Wolf

 

[Cherry67]

@drbacke
Interesting predictions and not to far off.
Good work on the model.
The predictions are close to what my tests show. Good stuff!!!



My cut off on those cells(UR18650FJ) is Optimal ?68m? and Marginal up to 76m?
There is only 1 thing I do different with any <1V cells that are within my acceptablem? readings.
I charge them in parallel at a CC/CV of 4.2V at 50mA per cell.
This usually takes a couple of days if not 3 or 4 before they come up to voltage.
If a cell gets remotely warm during the 50mA charge it is taken out.
These happen to be UR18650FMs and not a heater in the bunch "yet". all cells<58m?.



Once they hit 4.19 they will rest a bit and I will retake the IR and then into the charger/ testers they go.
I label them with a yellow font in my spreadsheet that way we know they were low voltage cells to begin with.
Once in the charger/testers charging the rest of the way to 4.2 at 500mA that is where we will see if there are any heaters.
Most of the time they behave very well. We will see.
I have now started to change the font colorof the "known" heaters that make it through the preliminary IR tests in the tested capacity to red.


As you can see I only have 4so far. All the other ones never made it this far. I do still have a boxload of ~300 ofthem of all various makes and models but mostly Sanyo.

How did you charge the UR18650FJ cells?

Wolf

It may be the wrong thread to start a discussion about it, but you have so woderfully and thoroughly descriped your approach of reviving low-U cells.
I did already in the past offer "sceptical interest" because your process links the cells together and do not treat them "separately", but however in the end the success beats the science.

I come back to this because it didnt leave my mind to try to understand what this process does in effect, and what makes it so obvious successful in your case.

For once i have new "strange" positive results in even lowering the starting current (see my lucky thread), and reading your nice desription again i found a new insight which help unterstanding better what tricky thing you do.

What i found is, that (in the end) good cells come up from even 0.6 V per cell in an amazing dU/dt , taking some 3 mAh (only) to come up to some 2... Volt. Only then they start to slow down in dU/dt taking up charge and storing it properly in the intended chemical processes.

My thinking (which is yet to be "proved", it is only hypothesis) came to the idea, that "future" "bad" cells may have already damages in the separators, or "dearranged" chemicals which will not anymore take part in the storage process or even distort that (like copper dendrites).
These damages or dearrangements may take charge without essentially storing them, so being losses. That would slow down the dU/dt value, being an indicator for good or worse cells.

Now coming back to your method, in essence it (perhaps) can be looked like you indirectly compare the dU/dT of twenty cells, by distributing the current to the cells having equal dU/dt, so giving more current to the bad ones, giving them more losses, which can be identified als heat.....
(The Genie Idea of Wolf is just comparing good to bad cells, instead of scientifically determining what is good or bad)

Now, your 50 mA is quite high compared to the 2 mA i used. There is lots of rooms for more experiments.
A last thought may be a very low cells may have a very slow chemical speed to take even the low currents, and any "excess" does "damages".

I look forward to what you may say to this considerations, and if Wolfs obvious succesfull "sorting" arrangement is only a sorting or even more....
(what means probably there is a limit of current where more is itself damaging the Low-U cell more)

One thing more - perhaps it would be a good idea to transfer this and many other things to the Low Voltage thread which is just newly created.