Battery Pack Failed Cycle Life Test Early. What Next?

Quarantine the pack, stop the test clock, and freeze the data before anyone opens a housing. An early cycle-life failure is an evidence problem before it is a hardware problem, and the records that decide whether you are holding one bad pack or a lot-wide defect get destroyed by the first engineer who takes a quick look inside, or the first BMS that gets power-cycled and rolls its buffer.

The last time this landed on me, the number was 412. We had warranted 1,200 cycles to 80 percent of beginning-of-life capacity, and the sample pack crossed that line in a chamber on a Tuesday night with nobody in the building. I found out from a graph. Flat, boring capacity fade for three hundred cycles, then a knee, then a cliff, and the test software dutifully logging it all into a file nobody had opened in six weeks.

That file was the most valuable thing we owned for the next month. Not the pack. The pack was a brick. The file was the case.

The window where you still have evidence

Before anyone touches anything, you want the raw channel data out of the cycler, not a screenshot and not a summary PDF. Current, voltage, and temperature at whatever interval the equipment records, plus the calibration certificate for the channel that ran this test. I have been burned by a cycler channel that read current about four percent high — not enough to notice on a dashboard, enough to quietly shorten a life test by a couple hundred cycles.

Then the BMS logs. Pull them first, before the pack is powered down for shipping or moved to a bench, because a lot of automotive-grade BMS boards run small circular buffers that overwrite themselves in days. If you lose the cell-group voltage history and the cutoff events, you have lost the ability to tell a cell failure from a protection setting.

Then the environment. Chamber setpoint file, any excursions, the ambient log, and a plain answer to a plain question: did this pack sit at high state of charge in a warm room over a holiday shutdown while the test was paused? I have seen a three-week unplanned rest at 95 percent SOC and 35 degrees do more damage than a thousand cycles would have.

Photograph the pack in place, label it, bag it, and put it on a shelf with a sign on it. Do not let anyone disassemble it. The moment your team opens the housing, you own the pack and the supplier owns the explanation. Whatever you were going to learn from the internal geometry, you can learn later, from a pack the supplier opens with you watching.

The uncomfortable step that comes before you call the supplier

Check your own test against your own warranty language.

I make this a rule now. If the purchase specification warranted 1,200 cycles at 0.5C charge, 1C discharge, 25 degrees C ambient, 80 percent depth of discharge, and the validation lab ran it at 1C charge, 1C discharge, and 35 degrees, you have not found a defect. You have found thermodynamics, and you will look bad on the call.

So I sit down with the profile file and the spec side by side and walk every line: C-rate in each direction, cutoff voltages, rest periods, temperature, and whether the pack was charged and discharged through the BMS or through a bench supply that bypassed it. That last one matters more than people expect. A pack cycled through an external charger with no BMS conversation never sees the same float and taper behavior the field will produce.

If the test drifted from the spec, the right move is not to bury it. It is to say so early, re-run at spec, and treat the original data as an out-of-spec observation. Suppliers remember which buyers are straight with them, and a re-run you can defend is worth more than a claim you cannot.

One failure is an anecdote. Two is a signature

Assuming the test matched the spec, the next move is a control pack from the same build lot, run under conditions as close to identical as you can manufacture — different cycler channel, same chamber if possible, same profile file, same thermocouple placement.

What I am buying with that test is a definition. If the control pack runs to a thousand cycles without a knee, I have an isolated failure, and the conversation is about that one unit. If it knees at roughly the same cycle count, I have a build characteristic, and the conversation changes completely: it is about the lot, the line, and every pack already on a robot.

Roughly the same is doing real work in that sentence. Cycle counts scatter. I do not need the second pack to fail at 412. I need it to fail in the same neighborhood, with the same shape of curve, and ideally with the same cell group showing the spread at end of charge. The shape matters more than the number.

Splitting cell-level from pack-level before you cut anything open

Most buyers jump straight to the cell. Cells are where the energy is, so cells are where the blame goes. But a meaningful share of early cycle-life failures I have worked through were pack-level, and a few were not failures at all in the way the buyer meant.

Here is roughly the decision tree I use, and none of it requires destroying the pack.

If capacity fades at a normal rate but DCIR climbs much faster than the capacity curve predicts, I look at interconnects: weld quality, busbar geometry, contact resistance between cell terminals and the current collector. A resistive joint heats under load, the heat spreads into the adjacent cells, and the whole assembly ages unevenly. You can often see this in the temperature channels before you see it in the capacity number.

If the pack's capacity falls off a cliff but individual cell groups look fine in the BMS log, I look at the protection layer. A cutoff set at 2.9 volts per cell under a 1C load will terminate discharge early on a cold pack, and the cycler will happily record the shortened discharge as capacity fade. The cells did not fail. The setting was wrong for the operating window. I have seen a pack lose fifteen percent of apparent capacity to a cutoff voltage that was never re-validated against the actual duty cycle.

If one cell group consistently hits end-of-charge voltage well before the others, and consistently bottoms out first, that group is the weak one, and the question becomes whether the weakness came from the cell, the weld, or the sense line. A sense-line offset can fake a weak group for months.

If BMS-reported capacity and the cycler's coulomb count disagree by more than a percent or two, suspect the measurement chain before you suspect the chemistry.

And if the failure is genuinely cell-level — one group plating out, self-discharge climbing, a group that will not hold a charge after a rest — then the question is lot variation, formation, or the cell maker's aging data, and that is a question the pack assembler has to answer on the cell maker's behalf.

What you ask the supplier for, and what to reject

An 8D, or something structurally equivalent. Not a root cause statement in an email. The discipline is the point.

What I expect to see: a containment section that tells me what happens to remaining inventory and to any packs already in transit or in my fleet, with dates. A root cause that names the mechanism and also names the escape point — how the defect got past their outgoing inspection, because a defect with no escape point means their process found it and shipped anyway. A corrective action with an owner and a completion date. And a verification-of-effectiveness section that is specific: what test, what sample size, what pass criterion, what date it will be run.

The last one is where most reports turn to fog. "Process improved" is not verification. "Twenty packs from the next build, full cycle test at spec, threshold set at X, results due the 14th" is verification. If they will not commit to a number and a date, the report is a correspondence exercise.

Alongside the 8D, I ask three traceability questions in writing. What is the cell lot code and date code for the failed pack. Do you have formation and aging data for that lot. And is that lot present in any other pack delivered to us, or to anyone else, and can you tell me which serial numbers.

That third question is the one suppliers are slowest to answer, and it is the one that matters most to a fleet. Cell substitution inside an approved part number is common, sometimes disclosed and sometimes not, and the only defense a buyer has is a paper trail that ties serial numbers to lots.

The finding most buyers skip entirely

If a sample pack failed your gate and then failed in the field, the gate failed too. That is a second finding, and it belongs in the same review.

So I go back to the sample qualification and ask what it actually proved. How many cycles did it run? At what temperature? How many units? Were they built on the production line or were they hand-assembled engineering samples with a note that production would be "equivalent"? Was the profile run to the warranty conditions or to something gentler because the chamber was busy?

Repeated early failures are a signal to tighten the gate before the next batch ships, not to accept a replacement pallet and move on. A replacement shipment that arrives through an unchanged gate just relocates the failure into your fleet, on robots that are harder to reach than a test bench.

When you cannot get to the bottom of it

Sometimes you will not. Cell makers do not open formation data to strangers. Small assemblers genuinely do not know which lot went into which pack. You can spend six weeks and end up with a plausible mechanism and no proof.

When that happens, I stop chasing certainty and start pricing risk. What fraction of the delivered lot has actually been cycled. What does the failure distribution look like — clustered or scattered. What does it cost to cycle-test the remaining inventory versus what a mid-fleet replacement costs in downtime. Then I write the decision down with the uncertainty still in it, so that whoever reads it in two years can see what we knew and what we assumed.

That memo is the part nobody enjoys writing. It is also the part that holds up.