Keeping Battery Pack EOL Test Meaningful When Production Scales

Keep the mass-production end-of-line test equivalent to or stricter than the pilot program, and write three things into the transition plan before the first production build: the exact build event where 100% EOL testing starts, the number of consecutive passing units required before full release, and the record fields the results land in. The fixture, the handler, the cycle time, and the test software are all allowed to change on the way to scale. The measurement basis is not. The moment production results can no longer be compared against the pilot results, the only honest baseline you had for what a good pack looks like is gone, and no amount of yield reporting brings it back.

I usually end up at the new end-of-line station in the last week before the production-intent builds, with a laptop open beside the supplier's monitor. At pilot, EOL was a bench against the wall: a programmable DC supply, an electronic load, a sequencer script on a PC, and one technician who clipped the pack in by hand, waited out the soak, typed the readings into a spreadsheet, and saved it. At volume it is a transfer handler with a pneumatic clamp, a barcode scanner, and a screen that shows a green box for about forty seconds before the pack indexes away. The new station is genuinely better. It runs four hundred packs a shift instead of twelve.

That is also the problem. The screen in front of me shows voltage, DC internal resistance, a communication pass flag, and a serial number. The spreadsheet on my laptop shows voltage, AC impedance at 1 kHz, a saved CAN trace per unit, insulation resistance with a thirty-second dwell, and a hand-recorded torque value off the busbar fasteners. Two of those five have a rough equivalent on the new station. One has something that looks like an equivalent but is a different measurement entirely. Two have nothing at all.

I do not learn that from the summary report. I learn it by asking the supplier's test engineer for the sequencer configuration and reading it next to the pilot procedure, column by column.

Equivalence is a program-to-program comparison, not a spec-to-spec one

The datasheet will tell you the pack is tested to the same specification at both stages. That is usually true and almost never sufficient. What I want side by side is the pilot test procedure, which is often a written document with technician instructions in it, and the production sequencer, which is a configuration file with a list of steps, conditions, limits, and record fields. Put them next to each other and the gaps show up fast.

Typical findings, in the order I usually hit them:

• A manual step with judgment in it becomes a pass-fail flag, or disappears. The technician at pilot who looked at the CAN trace and noticed a node dropping out is not replaced by anything at volume.
• A dwell time shrinks. Thirty seconds of insulation resistance at pilot becomes five seconds at production because the station has a cycle-time budget and thirty seconds blows it.
• A measurement changes basis. AC impedance at 1 kHz on a bench meter and DC internal resistance from a ten-second pulse are both real numbers and are not interchangeable. If the pilot limits get copied onto the new station anyway, you get false rejects on day one and someone quietly widens the limit on day three.
• A record field changes name, unit, or disappears. This one is invisible until you try to pool the two data sets, which is usually six weeks too late.

Coverage that is lost without anyone deciding to lose it

Most coverage reduction I have seen was not a decision anyone made. It was a consequence of the station's cycle time and the supplier's process engineer doing the reasonable thing, which is making the automated sequence fit the takt time. Nobody signs a document that says "we dropped the leak test." It just is not in the new sequence, and the new sequence is what the operator runs.

So I ask the test engineer directly, not the account manager: what did you cut to hit the cycle time, and who approved it. If the answer includes a 100% safety-related measurement moving to audit sampling, that is a buyer-side decision, not a supplier-side one, and it needs to come back to you with a rationale attached.

Pinning down where 100% testing actually starts

"From start of production" sounds precise and is not. Between the pilot build and steady state there are usually pre-series builds, soft-tooling builds, and a ramp where some units are saleable and some are not. Units from that window have a way of ending up in the field.

What I want written down: the build event that triggers 100% EOL, the number of consecutive units that must pass cleanly before the line is released to full rate, and what happens to a failure inside that window. Hold and tear down, not rework and continue. If a unit fails and gets reworked and then passes and nobody looks at why, you have converted your release criterion into a formality in about four days.

I also want the transitional builds identified. If a pre-series unit ships without the full sequence, its label and its record should say so.

Build the bridge while you still can

The pilot population is a perishable asset. Once the bench is decommissioned and the technician moves to another program, you cannot recreate it. So while the two stations both exist, get units through both.

The clean version is running pilot-built packs through the production station, so the same physical units produce readings on both systems and you can compute the offset directly. Mechanical differences sometimes block that, in which case the fallback is running the first thirty or so production packs on the pilot bench in parallel, or at minimum keeping a matched pair of golden units, one known-good and one known-marginal, and running them at the start of every shift on the new station. The golden units do double duty: they catch a fixture going out of calibration before it produces a shift's worth of scrap.

Fixtures change the measurement, so the limits need a review

This is the part that gets skipped most often. A new handler means new leads, new probe geometry, new contact resistance, a different thermocouple location, and possibly a different ambient condition at the station. Those things move a resistance reading by a few percent without a single pack being different. On a good day that is noise. On a bad day it eats your margin and the line starts rejecting good product for a reason that lives in the fixture.

Re-derive the limits from the production station's own distribution with a guard band, and be honest with yourself about the difference between recalculating a limit and relaxing one to improve yield. The first is engineering. The second is how a real failure mode gets waved through for three months. A simple repeatability study helps here: run one pack ten times on the new station and look at the spread. If the spread is wider than the tolerance you are trying to enforce, the station is not capable and no limit adjustment fixes that.

Operators and stations have to be qualified before the first shipment

The pilot had one technician on days. Production has three shifts, and the readings from third shift are always the interesting ones. Operator qualification on the new station, plus a documented station capability check, belongs on the checklist ahead of the first shipment rather than after the first complaint. It is a boring item that takes half a day and saves a quarter.

The record fields are the thing that quietly breaks

When the supplier changes test software, even to a newer version from the same vendor, the export changes. Field names change. Units change. A timestamp format changes. The serial number the station writes may not match the serial on the label if the label printing and the MES are not in step. I ask for one raw result file per unit, not a PDF summary and not a roll-up, and I diff the header row against the pilot export before the first production run. If the software changed mid-program, I want the field mapping in writing.

That is the whole reason the pilot data is worth keeping. Not as an archive. As the reference set that tells you whether the process moved.

If coverage is reduced, document the risk and name the control

Sometimes the reduction is correct. A slow test with a strong upstream process control behind it genuinely can move to sampling. When that happens, I want the risk written down, the compensating control named, and a signature on it. Not for the paperwork. So that eighteen months later, when a field return shows up with a failure mode that the dropped test would have caught, there is a record of who decided what and on what basis.

Comparing the first month of production data

The first weeks of production data are not a yield number. They are a comparison exercise. Voltage mean and spread against pilot. Impedance or resistance mean and spread against pilot, on whatever basis the new station uses. Communication failure rate, which at pilot was near zero partly because someone was watching the traces. If the means shift and the spreads hold, that is usually the fixture or the measurement basis. If the spread widens, that is the process.

Either way, you find out in week two instead of at the first field failure, and only if the two data sets were kept comparable on purpose.