Wrapped a 6-week pilot on our die-cast housing line in Juárez using a 12 MP Keyence vision head plus a lightweight anomaly model on an IPC; false rejects dropped 38% and it flagged a porosity drift about 2 hours before SPC, with roughly 80 ms per part latency. If you’ve deployed similar, what guardrails did you set for change control and traceability so QA signs off without slowing the line, and did you push any small design tweaks (we’re testing a 0.5 mm fillet change) to amplify the gains?
Quick example: on our die-cast trim line we pinned shutter/gain/lighting via PLC interlock and hashed the model+config; each inference logs to MES with lot/press/cavity plus the hash so audits line up, @OP. For changes, we do a 24‑hour canary on one station against a frozen golden set and only let operators touch the threshold — everything else needs “change order or rollback.” Small caveat: after any lens wipe or lamp swap we auto-capture a baseline and compare SSIM before going back live so we don’t ‘trust the camera’s new glasses’ blindly.
I’ve seen similar results with our 16 MP vision system, which cut false rejects on a trim line by about 35%. We implemented a change control process where each model update requires a quick review from QA, and we track everything in our MES. That said, keeping the team aligned on potential overfitting risks is crucial; sometimes more data isn’t always the answer. @emma39, any thoughts on balancing that?