Time to brush up on current affairs. Because we're predicting Li-ion batt lifetimes using impedance and AI
- Reference: 1586241611
- News link: https://www.theregister.co.uk/2020/04/07/ai_battery_cycle_lifetime/
- Source link:
The constant cycle of discharging and charging Li-ion batteries gradually knackers their maximum capacity, though the degradation process is difficult to estimate. Having a system that can automatically figure out the number of charging cycles left in a battery before its maximum capacity drops too far, and the component is thus on the way out, would be nice.
And so, boffins at the University of Cambridge and Newcastle University in the UK have [1]developed software capable of forecasting the “remaining useful lifetime” of Li-ion batteries using AI and a method known as electrochemical impedance spectroscopy ( [2]EIS ).
EIS in this instance works by briefly applying an oscillating voltage across the battery, and measuring the current response. These readings – which reveal the battery's impedance characteristics, or its resistance to an alternating current – are fed into a trained model, which uses them to predict the remaining lifetime of the battery under test.
“Our system sends in an oscillatory signal into the battery, and measures its response,” Alpha Lee, first author of the [3]paper and a research fellow at the University of Cambridge, told The Register . on Monday. “This impedance spectrum provides information about the different electrochemical processes that are happening within the battery.”
The approach relies on [4]Gaussian process regression , a statistical algorithm that can be trained to learn what properties are most indicative of degradation in order to predict how many charging cycles are left before a battery’s capacity drops to 80 per cent of its initial capacity.
The model was trained using 20,000 EIS measurements from batteries spanning a range of health, to learn what impedance characteristics a battery is most likely to exhibit when it’s about to drop to an unacceptable level of capacity. After training, when shown EIS readings of an arbitrary battery, the model can predict the number of cycles that battery has left before its performance falls below that acceptable level.
“Our model accurately predicts the remaining useful life, even without complete knowledge of past operating conditions of the battery,” the team's paper stated. The academics claimed that their model [5]is more accurate than conventional methods.
If it's Goodenough for me, it's Goodenough for you: Canuck utility biz goes all in on solid-state glass battery boffinry [6]READ MORE
Lee hopes that the algorithm will be used in commercial Li-ion batteries in the future so as to generate a warning system that tells users when their batteries need to be replaced.
“This is particularly important for electric vehicles because battery failure between service stations could be a major inconvenience, and some failure modes could pose safely concerns," he said.
"The model is also important for battery recycling because it can rapidly assess how ‘healthy’ a battery is, informing the decision of whether to re-use it for less demanding applications or recycle it as scrap metal.
“A key advantage is that our method does not require any modifications to the battery chemistry, so it is a simple turnkey solution. We are looking to work with electric vehicle manufacturers, original equipment manufacturers in the consumer electronics space, as well as battery manufacturers.” ®
Sponsored: [7]Forrester Build a Digital Experience Portfolio
[1] https://github.com/YunweiZhang/ML-identify-battery-degradation
[2] http://lacey.se/science/eis/eis-principles/
[3] https://www.nature.com/articles/s41467-020-15235-7#Sec2
[4] https://towardsdatascience.com/quick-start-to-gaussian-process-regression-36d838810319
[5] https://www.nature.com/articles/s41467-020-15235-7/figures/2
[6] https://www.theregister.co.uk/2020/03/02/canadian_firm_to_develop_goodenoughs_new_glass_battery/
[7] https://go.theregister.co.uk/tl/1936/-8554/forrester-build-a-digital-experience-portfolio?td=wptl1936
"It's not an algorithm, its a black box"
It's linked to from the article - you can download it and run it in Matlab.
"if they can tell me why specific voltages returning specific impedence values means a specific number of charge cycles, fine, but I bet they can't."
It's in the paper. You train a model to take these variables - frequency, temperature, impedance, etc - and match them to battery lifetime. So that when you show it arbitrary EIS values, it predicts the lifetime.
C.
It's not an algorithm, its a black box
You don't know what's happening inside a black box, by definition.
However, when you generate an algorithm using machine learning, as in this case, you can see exactly what it's doing i.e. you can see how it works, you just don't have a clue why it works.
Battery analysis
Having once spent a year (at least it felt that way) analysing the charge state of NiCd batteries, I can say that impulse response (a form of ac excitation which does indeed measure the impedance) is a pretty accurate measurement method for rechargeable batteries.
The problem is that battery construction (there are a lot of ways to construct a battery in a given technology depending on the end use) requires each type of construction to have its own training data which I had to do by hand (this was 30 years ago and I tested literally thousands of batteries from multiple manufacturers and batches) although the 'typical' construction for a given battery yielded type consistent results across manufacturers (the manufacturers said it was impossible to know the charge state but they would - they were in the business of selling new batteries).
There are a lot of gotchas in this field but with sufficient training data the analysis can be quite accurate.
I ended up building an analyser with a microcontroller and the results were tested by actually discharging the battery and confirming that the test had reported the correct charge state - it was always within 5% of the reported state of charge and typically within 2% so it is possible.
Re: Battery analysis
You speak from a time when it was fashionable to think. Now we say "It's all too complicated, so give it to a neural network and we'll move on to something else". Oh, and invent some pretentious term to impress the natives. We measure the battery impedance at two or three different frequencies, so call it "electrochemical impedance spectroscopy".
Sounds nice
Call me when you sell it in a handheld format, where I put the battery I want to test in, press a button and get a cycle number on a LED screen.
And, obviously, it has to be able to test all three common battery formats (AAs and AAAs especially).
Re: Sounds nice
... and it has a USB-C connector allowing testing of phones and laptops.
They hope the algorithm....
It's not an algorithm, its a black box, if they can tell me why specific voltages returning specific impedence values means a specific number of charge cycles, fine, but I bet they can't.
This type of thing will reduce our advancements in the future as it will reduce our understanding of how and why things happen. Taking the easy route will make things harder in the future.