From Production to Testing
Testing is not a separate department that grabs a few cases at random. Every check on the list exists because some production drift can defeat it, and the sampling plan is written around that pairing. Structure gets one acceptance level, appearance gets another, because the two fail in different ways and at different costs. Batches are coded in three segments so a failure can walk back to a resin lot, a mould and a shift, and a retained sample is what makes that walk possible six months later, when nobody remembers the run.
01Start from the failure list, then pick the samples
Sampling is not grabbing a few cases and dropping them. By the time a batch ships, the process already knows how it fails: thin walls crack first, a misaligned parting line leaks first, low foam density lets shock through first. The test list is built backwards from that list. Write down which check covers which drift and sampling stops being a ritual.
The pairings hold fairly steady. Walls 0.3 mm under the lower limit crack at a corner on impact. A parting line off by 0.2 mm lets moisture track in after immersion. Latch force 15% below spec pops open under vibration. Foam 10% light pushes peak acceleration up at the same drop height. Any failure on a report should trace back to a named process parameter.
02AQL 1.0 for structure, 2.5 for appearance
One batch does not run on a single AQL. Structural attributes — wall thickness, latch force, sealing, drop — are accepted at 1.0. Appearance — colour shift, flow lines, light scuffs — runs at 2.5. The logic is simple enough: a cosmetic flaw is visible and harmless, a structural one is invisible and ends the product. They should not share a tolerance.
Sample size comes off a table by lot size, and for cases this size it usually lands at 8–32 units per batch. Small lots are cheaper to screen fully. Large lots do not scale linearly; the table steps. The decision rule is Ac/Re. Reach the rejection number and the whole lot goes to 100% inspection, rather than the bad units being pulled and the rest waved through.
03Three-level lot code: resin, mould, shift
The lot code carries three segments: resin lot, mould number, shift. Drop any one and traceability breaks in the middle. A leaking case with only a finished-goods code tells you the day it was made. Resin lot reaches melt index, mould number reaches the shot counter, shift reaches that night's temperature log. Stacked together they scope the risk in under 2 hours.
Retained samples follow the lot code and stay 6–12 months. They are not a collection; they are the physical reference for a retest. When a complaint arrives, dropping the retained unit tells you whether the design is wrong or that batch was an outlier. Sort the cabinet by lot, clear it quarterly, and scrap expired samples nobody ever asked about.
04CPK 1.33: in tolerance is not the same as stable
A batch can sit entirely inside tolerance and still be one shift away from crossing. CPK measures how many standard deviations fit between the process mean and the nearest limit. Key characteristics — wall thickness, latch force, critical internal dimensions — should hold 1.33 or better. Below 1.0 the process cannot hold the tolerance and full inspection only hides it. Rework the tool or the window.
CPK needs a run of data, not three parts. Five units per shift across twenty shifts gives about 100 points worth using. If the result lands under 1.33, check mould temperature swing first, then resin lot changes, then ejection distortion, before deciding between adjusting the machine and tightening the drawing. Do that and sampling can relax from 32 units to 8.
05Retest every 3–6 months and feed failures back
A released product is not tested once. Retesting runs on a 3–6 month cycle, and any change of resin, tool or shift triggers one early. Only the high-risk items from the original 8–14 need repeating. The question is drift: after six months on a shelf, how much latch force has faded, how much gasket rebound is left. That is what shelf life actually rests on.
A retest failure goes back into the first-off sheet. If an item comes in low twice running, it becomes a mandatory check at the start of every batch. Close that loop and the trail is complete: lot code to shot counter, counter to service record, record to the measured deviation. The whole walk stays under 2 hours, and none of it depends on anyone's memory.
| AQL, structural attributes | 1.0 |
|---|---|
| AQL, appearance attributes | 2.5 |
| Sample size per batch | 8–32 units |
| Retained-sample storage | 6–12 months |
| Lot-code levels | 3 levels (resin / mould / shift) |
| CPK on key characteristics | ≥1.33 |
| Test items per batch | 8–14 items |
| Full trace time | ≤2 h |
FAQ
The lot is only a few dozen units. Is AQL sampling still needed?
No. At that size, testing everything is faster than working out a plan, and it catches isolated defects that sampling would miss. The AQL tables earn their keep on large lots, not below roughly 50 units.
Can a sample kept for 12 months still be retested?
Yes, but compare like with like. A retained unit shows a trend, not an absolute figure. Drop it and a fresh unit from the same height, and the difference is what shelf time cost you.
Is CPK 1.33 a hard requirement?
For key characteristics, yes. General features can run at 1.0. Below 1.0 the process cannot hold the print, and more inspection only sorts defects the next batch will repeat. Change the tool or widen the tolerance.
Is tracing to resin, mould and shift enough?
Usually. The three segments cover material, tool condition and that shift's settings. If a batch ran across two machines, add a machine digit to the code, or the trail stops at the shift and scoping takes far longer.