End-of-line test automation: what a pass actually proves about the product

Coverage is bought with cycle time

Every extra measurement costs seconds, contacts and something that can go wrong. Deciding what to test therefore means deciding what not to test, and that decision should follow from failure modes that genuinely occur: what has escaped to customers, what assembly finds, what the design engineers consider marginal. Testing everything the product can do is neither affordable nor necessary; testing what the process is capable of getting wrong is. Write down what each measurement exists to detect, because a station with no documented purpose accumulates tests nobody dares delete and loses the ones that mattered.

The fixture is the commonest source of bad verdicts

Most arguments about test results turn out to be fixture problems. Spring contacts wear and lose force, contamination builds on probe tips, connectors have a finite number of insertion cycles, a locating feature wears until the unit sits differently, and cabling degrades where it flexes every cycle. Each produces failures indistinguishable from genuine product faults, and the usual symptom is a climbing failure rate on one station while the others hold steady. Fixtures need scheduled maintenance, a cycle count, and consumable items treated as consumables. A fixture check with a known-good unit at shift start catches most of it.

Limits should come from evidence, not from comfort

Test limits often start as the design tolerance and then get tightened after every escape, until yield falls and nobody can explain how the current values were arrived at. A sounder basis is the measured distribution from a capable process combined with the values that correlate with real field failure. Where a limit is tighter than the process can hold you are scrapping good product; where it is looser than the failure threshold you are shipping bad. Record who set each limit and on what evidence, then demand the same standard from anybody proposing to change it.

Retest culture destroys the data you would need

Once a failing unit is retested until it passes, the test has stopped meaning anything. Some retesting is legitimate, since fixtures do produce genuine false failures, but it needs rules: how many attempts are permitted, what must happen between them, who authorises it, and every attempt recorded including the failures. Without that record you lose the ability to see a station degrading or a batch drifting, because only eventual passes survive in the data. Units that pass on retest deserve investigation rather than shipment, particularly where the same unit keeps reappearing.

Keeping stations agreeing with each other and with themselves

Stations disagree. Two nominally identical testers give different readings because of cable lengths, fixture wear, instrument calibration and ambient temperature, and the difference surfaces as a product failing on one line and passing on another. Manage it with reference units circulated between stations on a schedule, traceable calibration for the instruments, and a defined action when a station moves. Decide in advance what happens to product tested by a station later found out of calibration, since that decision is far harder to make under pressure with a shipment waiting and no agreed rule to point at.

Frequently asked questions

Should a failed unit be retested?

Only under a documented rule, with every attempt recorded. Genuine false failures do happen, mostly from fixture contact, so a controlled retest with a stated limit on attempts is reasonable. What is not reasonable is retesting until a pass appears and reporting only that pass, since it conceals both product problems and station degradation. Units passing on retest should be reviewed rather than shipped without thought, especially where the same unit or the same measurement keeps recurring.

How do we decide what to test at end of line?

Work backwards from failure evidence. List what has escaped to customers, what assembly catches, what warranty returns show and what the design team regards as marginal, then design measurements that would have caught those. Weigh each against cycle time and fixture complexity. Record the purpose of every test so the station can be reviewed sensibly later, and revisit the set periodically, because failure modes shift as the design matures and the process settles down.

Why do two identical test stations give different results?

Because they are not identical. Cable lengths and routing, fixture wear, instrument calibration state, firmware revision, ambient temperature and even the supply all move readings. The differences are usually small and matter only where a limit sits close to the spread. Circulate reference units between stations on a schedule, keep instrument calibration traceable, and record station identity against every result so a disagreement can be investigated rather than argued about across a table.