ChargeWisp. Battery health, read from ordinary charging

Knowing a pack’s health is not the same as knowing what to do with it.

ChargeWisp reads the shape of a battery’s charging curve — data your docks and swap stations already record — and returns both a health number and the decision that follows from it: keep in service, monitor closely, reassign, repair, give a second life, or recycle.

No hardware, no downtime

Health is classically measured with a full controlled discharge on lab equipment: hours per battery, specialised hardware, the pack out of service. That is impossible at fleet scale and absurd to repeat monthly.

Charging is controlled by the charger, so it looks the same in the field as it does in a lab. That makes every ordinary charge a measurement — including partial ones, because riders rarely charge from empty to full.

Validated on 793 batteries

Across 9 independent public battery-aging datasets, spanning every mainstream chemistry. Every number below comes from batteries held out of training entirely — never seen by the model — with each battery weighted equally, so no single well-behaved cell can flatter the result.

793Batteries validated on, across 9 independent datasets
4 ptsMean error across all 160 never-seen test batteries
2.5 ptsMean error on the packs the model certifies
Estimated versus lab-measured state of health, shown for all validation batteries and for the certified subset. Points concentrate along the diagonal, more tightly in the certified panel.
All 160 test batteriesmean error 4.4 points · 73% within ±5
The 98 it certifiesmean error 2.5 points · 89% within ±5
How to read it: each dot compares the model’s estimate (vertical) against the lab-measured truth (horizontal), on batteries it had never seen. On the diagonal is perfect. The right panel is the same model after it has set aside the packs it judges unscoreable — the cloud tightens onto the diagonal.

It knows when it doesn’t know

Some packs behave too erratically to score honestly. The model identifies those itself — before seeing the true answer — and flags them for inspection instead of publishing a number it cannot stand behind. Of the 160 test batteries, 32 were flagged rather than scored.

This matters more than headline accuracy. A wrong “recycle” scraps a pack with years left in it; a wrong “keep” becomes a roadside failure. Certified estimates additionally carry a mathematically guaranteed error bar, which is the format insurers and second-life buyers ask for.

The same charging data also yields remaining useful life — typically within about 19% of the truth on unseen packs, and biased conservative early on, which is the safe direction to err for fleet planning and warranties.

How working together actually goes

  1. You send charge logs. An export of what your stations or BMS already record — voltage, current, temperature, time. No integration project, nothing to install on your side.
  2. I run the model and send back results. Per-pack health, the lifecycle decision, a confidence flag, and an honest list of the packs the model declined to score.
  3. You check it against what you already know. If you have ground truth on some of those packs — bench measurements, packs you already retired — that is the cheapest possible test of whether this works on your batteries, on your own data, before anyone commits to anything.

On your data. What you send is used to produce your evaluation — not passed to third parties, not resold, and deleted whenever you ask. What I keep from it are the aggregate accuracy numbers, never your raw logs. Calibrating the model to your batteries specifically is a separate conversation, and only happens if you want it. Happy to put all of that in writing, or to work from an anonymised or partial export if that is easier to approve.

Retiring a pack too early is expensive twice over

Once on the balance sheet: a pack pulled at 80% health when it had two good seasons left is capital thrown away, and one pulled too late is downtime, a warranty claim, or a safety event. Made by hand, that call costs technician time on every single pack.

And once in the ground. Every battery retired before its time is another one manufactured before its time — more lithium, cobalt and nickel mined, more energy spent in cell production, and a pack sent to a shredder while it could still have served a lighter duty cycle. Accurate end-of-life decisions are the cheapest form of circularity available to a fleet: no new material, no new hardware, just not throwing away what still works. Sorting packs correctly into repair and second-life streams instead of defaulting them all to recycling is the same lever pulled again.

Where it fits