Development of a free Capacity Predicition, need your help

I'm currently developing a freely usablecapacity prediction.
The usage will be quite simple:

Input:

• Internal Resistance measured before charging
• Voltage before charging
• Nominal capacity

Output:

• Predicted Capacity + Standard Deviation

I used Wolf's and my ownDatabase to train3 differentsurrogate models. I'm not sure which one will be more suitable, so I did test predictions of ~90 unmeasured cells. I will measure them in the next days and we will see realtime, how good the predictors will perform or not.


Here you can see the test dataand the predictions.I will update the results dayly if measured:
https://docs.google.com/spreadsheets/d/1gfSAhdyL-qo7WtE69umzEC9Mz1LkAM2r13CfJyEt3E8/edit?usp=sharing

I will do this "real time" test is to see ifthe model helps mein a real harvesting process or not.

So this leads me to the next question:
If you would use such a model, how would you prefer to do it?
- Maybe as a website script where I put a table of Pre IR/V and nominal capacities and get the predicted Capacity+Deviation
- Or another option would be to define minimum requirement for the cells. Maybe my personal requirement is 80% of original capacity and min. 2000mAh. Then the Tool would givea probability between 0% and 100% that this cell will fit your requirements or not.
Or maybe you have another idea or suggestion?

When it's done, I'll make the tool available to everyone and hope that it will help you.

Interesting. Look forward to your results.
Are you using AI and NN to run your simulations or something else?

Korishan said:

Interesting. Look forward to your results.
Are you using AI and NN to run your simulations or something else?

Click to expand...

I thought about using NNs, but I tookKriging modelsfor this. Because it's more transparent and robustthan NNs and also the Deviation Prediction is really helpfull.
The biggest problem for the modelswill be to differentiateallthe different kind of cells. At this moment it's only possible through the nominal capacity/(IR Range).

drbacke said:

So this leads me to the next question:
If you would use such a model, how would you prefer to do it?
- Maybe as a website script where I put a table of Pre IR/V and nominal capacities and get the predicted Capacity+Deviation
- Or another option would be to define minimum requirement for the cells. Maybe my personal requirement is 80% of original capacity and min. 2000mAh. Then the Tool would givea probability between 0% and 100% that this cell will fit your requirements or not.
Or maybe you have another idea or suggestion?

When it's done, I'll make the tool available to everyone and hope that it will help you.

Click to expand...

As far as prediction of capacity with a Voltageand an IR measurement I'm not sure if that will work. But with some of the Models you are developing maybe.
I did look at your sheet and any cell there over 50m? will more than likely not produce good results. But you are probably just adding them for control purposes to add diversity and improve the calculation accuracy. The reason I developed my data is to be somewhatsure the cell will have an acceptablemAhoutcome of at least80% of capacity. I did try the prediction albeit not to the extent of building some models quite as extensive as yours but I failed miserably. I was kind of taking a folded paper plane into a wind tunnel at 700kph and wondering if it would fly.

Probability is more what we/you should be shooting for as I am working on the IR chart of as many manufactures and model numbers with at least 50 cells to establish a baseline of good, mediocre, and bad IR. I am doing this with, as you know, actual real life measurements to establish this info.
Can a model predict the probability of a cell being good by those 2 measurements certainly and I hope we will be able to use the tool in the future. But for a model to work very well it needs a lot and I mean a lot of data.
The only thing that I can think of possibly adding to this is the date of manufacture which may have some influence to the quality of the cell.
There are some key chartsthat will tell you how to decipherthe date code on some of the cells.

On a side note,
Here is a chart that will give you the probability of a ICR18650-26A, C, D, F, FU, andH being at ~80% capacity if you keep your IR <60m?. Can you stretch it to 70m? the chart seems to indicate that but your results will vary.
That's a sample of ~480 cells



I will follow your work and see what happens and will assist if you need me to.

Wolf

---

drbacke said:

I thought about using NNs, but I tookKriging modelsfor this. Because it's more transparent and robustthan NNs and also the Deviation Prediction is really helpfull.
The biggest problem for the modelswill be to differentiateallthe different kind of cells. At this moment it's only possible through the nominal capacity/(IR Range).

Click to expand...

Yes that is the problem I ran into as therearedifferent battery chemistries all with different IRs.
Most of the Cells we see are either high drain (low IR)Li-manganese and low drain (higher IR)Li-cobalt.
What I found out was that each and every manufacturer has their own little chemistry lab going on in their cells and the lab changes with capacity and type.

Wolf

I ran across this article last weekend while I was hunting datasheets, and I thought it was interesting. It regards using X-ray imaging to quantify the remaining capacity and IR of recovered cells.

https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0185922

I plan to read it in more detail when I get some time.

P.S. if anybody out there works in a facility that uses X-ray imaging equipment, this might be worth some experimentation.

Shawndoe said:

I ran across this article last weekend while I was hunting datasheets, and I thought it was interesting. It regards using X-ray imaging to quantify the remaining capacity and IR of recovered cells.

https://journals.plos.org/plosone/articl...ne.0185922

I plan to read it in more detail when I get some time.

P.S. if anybody out there works in a facility that uses X-ray imaging equipment, this might be worth some experimentation.

Click to expand...

Maybe you should start an own thread for this, I think that is to far away from this topic.

Wolf said:

As far as prediction of capacity with a Voltageand an IR measurement I'm not sure if that will work. But with some of the Models you are developing maybe.
I did look at your sheet and any cell there over 50m? will more than likely not produce good results. But you are probably just adding them for control purposes to add diversity and improve the calculation accuracy. The reason I developed my data is to be somewhatsure the cell will have an acceptablemAhoutcome of at least80% of capacity. I did try the prediction albeit not to the extent of building some models quite as extensive as yours but I failed miserably. I was kind of taking a folded paper plane into a wind tunnel at 700kph and wondering if it would fly.

Click to expand...

Yes I use a very wide range to control different locations in the whole space. Also a "bad location which is far away from any known data or is very noisy, should produce a high deviation. So you are able to controll if the prediction is usefull in this area or not.

Wolf said:

Probability is more what we/you should be shooting for as I am working on the IR chart of as many manufactures and model numbers with at least 50 cells to establish a baseline of good, mediocre, and bad IR. I am doing this with, as you know, actual real life measurements to establish this info.
Can a model predict the probability of a cell being good by those 2 measurements certainly and I hope we will be able to use the tool in the future. But for a model to work very well it needs a lot and I mean a lot of data.
The only thing that I can think of possibly adding to this is the date of manufacture which may have some influence to the quality of the cell.
There are some key chartsthat will tell you how to decipherthe date code on some of the cells.

Click to expand...

I also think, that the probability is more usefull and easier to use, because you have several informations in one "handy" value. I dissagree with the amount of data, because this kind of data is really really noisy, due to different kinds of chargers, measurement tools, temperatures etc. and we are only able to measure 2 or 3 useable variables. So about 4000 samples with 2-3 variables are more than enough for a robust kriging model. You won't get much better with more data, the thing you need are more variables like IR,Voltage etc.

Wolf said:

On a side note,
Here is a chart that will give you the probability of a ICR18650-26A, C, D, F, FU, andH being at ~80% capacity if you keep your IR <60m?. Can you stretch it to 70m? the chart seems to indicate that but your results will vary.
That's a sample of ~480 cells
I will follow your work and see what happens and will assist if you need me to.
Wolf

Click to expand...

Thats right and the kind of model I have chosen will model this variation through the predicted deviation. There are limits for doing this (if the deviation function is to complex), but the tests will show how well it actually work. And if the deviation prediction is not suitable we can try a deep gaussian process. This will do it much better, but is not that robust. So I prefer to test the more robust version first.



Of course you can not expect to get perfect results. But I hope to get a tool, which will help you to seperate good cells from bad in a more precise and flexible way as we are doing it till now. But we will see how good it works.
@Wolf: I don't know if such a model will be better than an experienced harvester like you are, but maybe it can be helpfull to other people which don't have that much knowledge and experience like you have. That would be my hope.


------------------------------

I added the first measurements and till now it looks quite well. The Co-Kriging Model seems to be perform better than the other models, which makes sense to me.
At this moment all tested cells have a quite good capacity left, but I'm most interested in the cells which perform bad and the ability to predict those "bad cells" which I want to sort out quite early.

Ok here are the first results.
The X-Axis shows the measured results and the Y-Axis the predicted results with the predicted uncertainty as error bars.
The blue line marks the ideal line, if the model woud be perfect, all values would lie on this line.
Till now it looks quite promining. But the"extreme" cells will be more interesting, especially the low Capacity cells.

@drbacke

Looking good. It will be interesting to see what happens.

I am now using my IR chart exclusively and occasionally I will slip a cell in there just over the marginal m? as a control.
You will see that in my sheet. It also allows me to microtune the IR chart.

I am also going straight to the charger/testerswith anything ?2.00V. Anything <2.00V gets charged at 4.2/50mA for however long it takes.
Also anytime I use the SKYRC tester I test the cell at the manufactures standard charge/discharge specifications with the charge cut off mA if know otherwise the default of 80mA is used. By the way the charge cut off mA is incredibly critical in determining the "true" capacity of the cells.
Cells # 5784 to 5827 are froma new batch ofcells I just harvested

and they are being tested as we speak. I would say 50 percent of the harvest hasn't made it this far. I cringe when I put the rejectedcells aside but I know they wont produce the numbers I want/need so away they go.


Wolf

Hi,

I had similar idea couple of months back, I will measure temperature, intial IR and voltage while charging @1A and record the data for the first 20-30minutes. Than using simple ML to predict the data, my initial data shows that it is possible. I am also discussing it with some HW colleagues to make some prototypes. I believe that for such a device that can estimate capacity in 30minutes people might be interested...

If someone is interested in helping me out with this, you are welcome. Big help would be with collecting the data - everyone that has a camera, temperature sensor and YR1030 is greatly welcome.
I need:
data for the first 30minut of charging (temperature, voltage, IR) with 1 minute between each measurement, measurement start from completely discharged cell. Also important is to finish with capacity measurement to be able to test the model.

When we have enough data, I can build the model and say for sure if it works. But as I say, my initial measurements say it is possible...

Satiriasis said:

When we have enough data, I can build the model and say for sure if it works. But as I say, my initial measurements say it is possible...

Click to expand...

@Satiriasis

That isinterestingcombining these 3 parameters into a ML for prediction. I will watch to see how it progresses.
I am more of trying to get the probability of a cell being good and within 80% of rated capacityrather than theprediction of capacity which I tried with several different methods.None with any AI,NN or ML though. I was not successful. Hopefully some of you are.

I do not have the time to do the 30 minute watching and recording these parametersonce every minuteof the cells though.
I have 11 charger/testers going pretty much 24hrs a day so logisticly that would not be possible without extensive temp, and voltage sensors going to a database. I do have a SKYRC though that records and graphs this info.
There are some questions I havewith this though.
1 IR should not be an issue as all measurements are made with the YR1030 (that is good)
2 Temperature can be certainly an indicator very quickly if a cell is bad but we would have hopefully reduced our chances of having a hot cell with our initial IR reading. I find that most good cell while charging at 1A will reach ~30C to 35C within an hour level off and when the amps drop the temp will follow the charging curve. This cell was first discharged at 550mA and thencharged at 2A. Here is the graph.




3 Still on temperature. If using an off the shelf charger such as Liitokala, Opus,Foxnovo, and god forbid a Zanflair there is some heat that is created by the charger/tester itself. The Opus is the only one of the reasonably priced charger/testers that I am aware of (other than the expensive SKYRC) that try to manage heat production with a small fan albeit unsuccessfully. Hence all the mod pages for Opus.
So how would that be taken into account to the Model.
I have a 20p charging board that I use to charge <2V cells at 50mA to see if they are revivable but only if they pass my initial IR test.
The cells that increase in temp very quickly <10minare no good despite the IR can you spot them? We are only looking at a 2C to 4C difference.



Compared to a board with all good cells with the same charging parameters.


\

If I have time I will run some control cells through my SKYRC and publish the results on my google drive.
If you give me some cell specs as far as manufacturer, model and IR range that you want me to test I will be more than happy to do that.
I have a lot of cells mostly Sanyo, Samsung, Panasonic and LG. So if you give me the parameters oflets say 20 cellsto begin with in this formatICR18650-30B with~3.76V and an IR of 48m? to 52m?. I will see if I can find suchcells within those parameters (should be no problem) and run them on my SKYRC and chart the results.

I look forward to your results and may the IR be with you.

Wolf

Isn't manufacture date also supposed to be relevant to cell capacity over time?
I suspect too many things will influence change in capacity from when it was new to easily predict capacity of a cell taken out of a pack.

Oz18650 said:

Isn't manufacture date also supposed to be relevant to cell capacity over time?
I suspect too many things will influence change in capacity from when it was new to easily predict capacity of a cell taken out of a pack.

Click to expand...

Not really. Number of cycles is probably the biggest factor. Degradation over time is minimal by comparison. Other variables include storage conditions - what sate of charge was the cell stored at, and what temperature was it stored at.

Measuring internal resistance is indeed the only valid method for predicting capacity.

I meant that you won't know things that have gone on previously, such as number of cycles.
I do believe that IR will be a good indicator, but you will need a good sample size of each type of cell to get meaningful data - as wolf has done.
I am hoping to do similar predicting as my cells are all the same type, so I will have a good sample size.
I also know manufacture date of the cell packs, so I am hoping to see how that affects performance of the cells.
I was planning to have started testing already, but I got more cells so I have continued extracting instead.

I like the new ideas that started in this thread.
But using some kind of Metamodel and put some data in it and hope that it will create good results doesn't work in reallife.
The development of such models is my dayly work and the models I show here are also complete developed by myself over years.
So you will not get that special kind of models in scipy or numpy etc..

And sample size is not everything.If you think so, you don't understand that kind of models and maybe you don't understand data modelling itself.

So next data, looks really promising in my opinion:

Please explain the picture for us?
I think I get it, but want to make sure.

Oz18650 said:

Please explain the picture for us?
I think I get it, but want to make sure.

Click to expand...

drbacke said:

Ok here are the first results.
The X-Axis shows the measured results and the Y-Axis the predicted results with the predicted uncertainty as error bars.
The blue line marks the ideal line, if the model woud be perfect, all values would lie on this line.
.....

Click to expand...

---

This diagram showsrather what we are interested in:
On the X-Axis is the measured Capacity and on the Y-Axis the predicted probability that the cell will have more than 2000mAh. So this looks quite good, most of the cells which have a probability>50% are actually over 2000mAh.
Of course you can use any capacity instead of 2000mAh, it's just an example.

Wolf said:

Satiriasis said:

When we have enough data, I can build the model and say for sure if it works. But as I say, my initial measurements say it is possible...

Click to expand...

@Satiriasis

That isinterestingcombining these 3 parameters into a ML for prediction. I will watch to see how it progresses.
I am more of trying to get the probability of a cell being good and within 80% of rated capacityrather than theprediction of capacity which I tried with several different methods.None with any AI,NN or ML though. I was not successful. Hopefully some of you are.

I look forward to your results and may the IR be with you.

Wolf

Click to expand...

Hi Wolf,

thanks for your reply, yes I also do not have much time, currently I have just one Opus where I have setup the environment, my way is to record the video for the complete charging/discharging process and than manually extract the data. So it will take ages, I need hundreds of batteries + different chemistry and vendors.

Currently, I am building the model that is able to predict it from complete discharge, later I believe, I can build the model also from completely charged batteries and watch characteristic while discharging them (again about 30min from the fully charged state). The final HW will have ability to decide - if you harvested fully charged battery, it will start from fully charged one, if almost discharged, it will discharge it and start testing for 30min.

The problem you have mentioned with the noise of the temperature is not a problem. I am not interested in the overall accuratetemperature, rather than changes within the time (deltas), but I will not be saying more at the moment...

@drbacke

Quite interesting to view yourCapacity Prediction Test sheet on google drive and look at your capacity results using IR as a guide and theymatch my own results pretty close. Pretty much within a 5% deviation. Not bad.

Wolf

Wolf said:

@drbacke

Quite interesting to view yourCapacity Prediction Test sheet on google drive and look at your capacity results using IR as a guide and theymatch my own results pretty close. Pretty much within a 5% deviation. Not bad.

Wolf

Click to expand...

Which results do you exactly mean?






Next results:
There are now some cells with low capacity and itseems to work quite well also.
But there is one cell, which was predicted quite bad in my opinion (in the probability plot ~30% prob and 2500mAh). It's aSanyo UR18650NSX with 12.5 mOhm.
Maybe in this case its really a problem of lacking data, because the predicted standard deviation is at maximum, which means there is no data in this place.
But also you can see, that the deviation is quite high, what tells you: "I don't know anything here"

drbacke said:

Which results do you exactly mean?

Click to expand...

Your tested capacity results in your sheet.
I just sort my data by cell Part number and then sort by IR selecting the IR you recorded as close as possible and eliminating the rest.

Certainly there is some deviation but most of them match up pretty close.

Next results:
There are now some cells with low capacity and itseems to work quite well also.
But there is one cell, which was predicted quite bad in my opinion (in the probability plot ~30% prob and 2500mAh). It's aSanyo UR18650NSX with 12.5 mOhm.
Maybe in this case its really a problem of lacking data, because the predicted standard deviation is at maximum, which means there is no data in this place.
But also you can see, that the deviation is quite high, what tells you: "I don't know anything here"

Click to expand...

The Sanyo UR18650NSX is a high drain manganesebased cell which will have a very low IR compared to acobalt based low drain cell.
You really can't put those cell into the same category as it will throw your model way off.

Wolf