Plastic injection quality control with AI: catching short-shot, flash and colour on the line.

Why do short-shots, flash and colour drift still escape a moulding cell?

Injection moulding runs stable until it does not. A blocked gate, a cold barrel, a worn tool or a bad colour batch turns a run of good parts into short-shots, flash, sink marks and off-colour shots, often for a few cycles before an operator notices. On a fast tool running many cavities a minute, a manual spot-check at a sampling table sees a fraction of the shots and misses the intermittent fault entirely, so bad parts reach the good bin.

The defects themselves are the reason ordinary checks fall short. A short-shot is an incomplete fill where the melt froze before the cavity filled. Flash is excess material squeezed out at the parting line. A sink mark or warp is a surface and geometry deviation left by thick sections cooling unevenly. Colour non-uniformity comes from masterbatch mixing or regrind. Each is a subtle, appearance-level departure from a good shot, exactly the judgement that a fixed brightness or edge threshold handles badly. Write a rule tight enough to catch flash and it false-rejects a good part; loosen it and the short-shot slips through.

How does anomaly detection trained on good shots only fit injection moulding?

Anomaly detection learns what a good shot looks like and flags anything that deviates, which fits a stable moulding process well. A qualified tool running in control produces a consistent good part, so the model can be trained on those good shots alone and then flag the short-shot, the flash, the sink or the colour drift as a departure from that norm, including a defect no one wrote a rule for. Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and anomaly on good parts only is one of its core modes.

Training on good parts only is also what keeps the data requirement small. You capture about 20 good reference images from a qualified run, and the model trains under 48 hours, so bringing a new moulded part, a new cavity layout or a new colour onto inspection is a two-day task, not a defect-cataloguing project that stalls a launch. The unit runs Anomaly, Defect, Counting and Quality modes and combines classical computer vision with AI inference, so a measurement stays auditable rather than a black box. Anomaly mode carries the first-seen cosmetic faults, defect mode carries the known repeat offenders like parting-line flash, and quality mode carries the colour and finish comparison against the good reference. For the method behind training on 20 good images, see how to inspect a part with only 20 images.

Can one unit check cosmetic defects and dimensions in millimetres?

Cosmetic and dimensional checks run on the same part in the same cycle. The anomaly and defect modes catch the visual faults, short-shot, flash, sink and colour, while classical computer vision measures feature position and size in millimetres for the dimensions that matter, a boss location, a mounting hole, a trim edge. On a delivered geometry-and-position case the unit reports feature coordinates and dimensional measurement per part in millimetres, the same measurement path a moulded housing, connector or closure needs for traceability.

Pairing the two is what makes moulding inspection worth wiring in. A part can fill fully, pass a cosmetic check and still sit out of tolerance because the tool crept. Reading the cosmetic and the dimensional result together, per shot, means the cell catches both the obvious short-shot and the slow dimensional drift that a sampling gauge only finds hours later.

Which inspection mode catches which moulding defect?

One unit maps each common moulding defect to an inspection mode, and every result leaves as a signal the cell controller already reads.

Does it keep up with the moulding cycle?

Inspection is locked to the cycle, not to a timer. The unit triggers off an encoder pulse, a photoelectric sensor or a fixed interval, so it captures each shot as the part ejects or indexes past the camera. On a live cap-inspection line, a moulded plastic closure, the measured field result is about 30 ms per part, and the catalog bound is 0.5 s per part at 100+ parts per minute.

Treat those figures as the envelope, not a promise for your tool. The per-part time you can commit to for your own part and colour needs an application-specific measurement on your parts and lighting. For most injection cycles, which run seconds per shot, a decision in tens of milliseconds means inspection is never the bottleneck, and a bad shot is flagged and diverted before the next cycle stacks it into a good bin.

How does it wire into a moulding cell without a vision engineer?

The unit drops into an existing moulding cell without a PC, a cloud account or a vision engineer on staff. It carries pass/fail over PROFINET, EtherNet/IP, Modbus TCP, EtherCAT or OPC UA, matching whatever the cell controller or take-out robot already speaks, and it has 4 inputs and 4 outputs at 24V for triggering and reject actuation onto a diverter, a chute gate or a robot pick. The AV-S100 standard unit (IP54, 0-45 C) suits a dry moulding cell producing caps, housings, trims and medical consumables. It mounts in about 30 minutes by one person because it weighs under 9 kg, and inference runs on a fanless Jetson-class board inside the enclosure. Every image is processed on the unit, so part imagery stays on the line, and models update by USB stick for an air-gapped cell.

This post is a spoke of the pillar guide on AI visual inspection; see where it maps across real applications.