Stationary battery capacity testing — only the veteran spots the marginal cell.

The problem

The rated capacity checks out — and the cell fails in six months.

Capacity testing in the formation room filters out the bulk of defective cells: the ones that do not reach nominal at the declared rate go out. But the real problem is the marginal cells, which pass within tolerance while dragging a suspicious curve behind them:

1. A small step in the discharge that the operator never gets to see because they are watching six benches at once.
2. A slow recovery between cycles that only the room's veteran detects — and the day they retire, that sensitivity walks out the door.
3. A temperature asymmetry between cells of the same block that the bench does not flag as a failure but that anticipates premature aging.

Those cells get assembled, pass the final test, reach the customer and fail six months later. The classic system works 99% of the time. That 1% is a claim on a critical backup system — and a UPS bank that does not respond during a power outage is not just cost, it is brand damage that is difficult to recover from.

How it fits the IRIS system

iLEAN does not add a fourth system — it seals the cracks between the bench and the human decision.

The problem is not a lack of data: the formation bench generates them by the thousands. The problem is that nobody can watch every curve in real time. iLEAN acts as the putty that fills the gap between what the bench measures and what the operator can attend to, without asking you to change the bench.

The bench measures. iLEAN Edge sees the curve in real time, against the batch pattern. The doubtful cell is flagged for the person, during the test. The person signs — never the other way around.

The iLEAN pieces applied to stationary battery capacity testing:

• iLEAN Edge — a terminal with a CNN trained on per-cell voltage, current and temperature curves. It compares every cycle against the learned batch pattern and flags the doubtful cell in real time. It works without a network: it records and flags locally; it syncs when the network returns. What is critical does not depend on WiFi.
• Connect — captures the bench data whether it is modern (Ethernet/RS485) or old (reading from an isolated local PC or, as a last resort, vision over the panel). Graduated capture is part of the design: no bench is left out.
• Quality agent — cross-references each cell's curve with the electrolyte batch, the plate manufacturing date, the bench's history and the room's ambient temperature. When a marginal cell shows up, it does not send an email at 10 p.m.: it stops the cycle, flags the cell, and hands the decision to the person in charge. The person signs — the AI does not decide the line.

See the full IRIS architecture →

Before and after

End-of-cycle inspection vs. iLEAN-assisted testing

Aspect	Classic bench inspection	With iLEAN Edge
Marginal-cell detection	At the end of the cycle, operator's eye	During the cycle, against a learned pattern
Voltage/temperature curves	Final chart, many at once	Real-time analysis, pinpoint alert
Capturing the retiring veteran's know-how	Knowledge that leaves with them	Learned pattern preserves the sensitivity
Decision on a doubtful cell	Quick call at shift change	Stitched data, informed quality decision
Rejection traceability	Note in a notebook	Signed record with curve and reason
Customer claim	No origin curve to diagnose with	Formation curve retrievable by serial number

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Mid-sized stationary battery manufacturer (VRLA/OPzS lead-acid or backup LFP lithium), formation room with several charge-discharge benches, manual capacity check at cycle close.
• Edge pilot on one bench (per-cell curve capture + CNN analysis against the batch pattern). First value expected within a few weeks: the first flagged marginal cell the operator would have let through.
• Indicative payback between 4 and 9 months, depending on the room's volume and the average cost of a field capacity claim.
• Expected reduction in marginal cells reaching final assembly ≥ 30%, a defensible floor. The hard lever is the avoided claim on a critical system: it pays for the pilot.

And the quality manager's reasonable doubt

“What if the AI rejects a good cell thinking it is marginal?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI compares a real curve against a learned batch pattern, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN flags the cell and the person decides. The three safety rings exist precisely for this.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.