Automotive: Plastic Part Defect Inspection — Field Practice

01/9Challenge

The hard part is freezing imaging and criteria—not “having a model”

Ribs, holes and edges occlude each other; different defects need different viewing directions. When ambient light shifts, specular highlights move with it—the same notch looks bright one day and dark the next. A classic “collect thousands of images and train a dedicated appearance model” path breaks on mould changes, batch changes and surface changes: long onsite training, slow replication.

So the path changes: freeze optics and pose first, cover irregular short shots and burrs with Transformer few-shot segmentation, and attach area, length and protrusion height to configurable thresholds. Inference stays on the edge GPU—protecting both cycle time and data locality.

02/9Defects

Five checks: three morphology, two metrology

The inspection scope converges to the five items below, numbered without gaps. All five completed segmentation validation in the coaxial trial; thresholds and accuracy are frozen on official samples in a fixed fixture.

Defect 02 is about structural integrity, not surface texture. Defects 04 / 05 ask “how tall”, not “is there a burr”—changing a tolerance only edits the millimetre threshold.

03/9Optics

Coaxial lighting: turn defects into a stable contrast cue

The part is black, curved and locally specular. Oblique light makes highlights drift with pose and ambient light. Coaxial lighting aligns illumination with the viewing axis: a good surface returns light to the lens (brighter); notches and burr sidewalls deflect light away (darker). A light shield then kills ambient light so the same defect looks the same across shifts.

Camera, lens and coaxial light follow the trial selection; exposure, WD, station count and structure freeze after official samples, mechanical space and cycle time are checked. Light parameters are managed in recipes per station.

04/9Tensor Flow

Three station branches into one workflow

The vision agent runs on the edge: configuration, inference, HMI and outputs share one software stack. Zero-code canvas wiring—drag operators, connect data ports. This project is not a single serial chain: three station branches run in parallel (capture → preprocess → segmentation → event detection), then merge into a state machine, save images and talk to the PLC over serial.

Double-click a node for the sub-editor: segmentation → annotate tool, event detection → GPU event composer, state machine → state-graph editor. A product change is mostly recipes and samples, not a rewrite.

05/9Segmentation

Pixel masks: lock a new defect with 1–10 samples

Short-shot boundaries are irregular and burr contours are fine; boxes fit poorly. The plan uses pixel-level semantic segmentation: each class is defined in the annotate tool from a small set of line images—product guidance is 1–10 sample images and minutes of setup. Masks then drive area, length or protrusion vs. a baseline. Training and inference stay on the local edge GPU; station frames do not leave the plant.

Across moulds and batches, similar morphologies often need only a few extra labels to transfer—the field value of a few-shot base versus “one long train per defect”: shorter launch, less permanent onsite training staff.

06/9Trial Evidence

All five items, each with its own detection result

Whether the project is worth starting rests on one question: under coaxial light, do the target defects actually form learnable, segmentable features? That cannot be argued—it has to be shot. The trial went through all five items one by one; below is a representative detection result for each: human labels on the left, model output on the right.

07/9Events

Turn masks into “hook short shot” and “burr over height”

Segmentation sees pixels; the business needs decidable events. Events are listed for the state machine; logic is built in the GPU event composer: presence, width measurement and other primitives get a class and ROI, then AND / OR / NOT. Burrs use width measurement for max perpendicular distance—fit a clean reference edge, measure contour-to-baseline distance; the millimetre limit lives in the recipe.

08/9State Machine

Arrive, inspect, merge five results into one OK / NG

The state machine turns discrete events into a line flow: wait for part → item checks → pass or NG; NG parts leave for review and can return after rework. All five must pass to release and count; any hit saves NG images and alarms. Results join by part ID across stations so you never get “each station OK, whole part missed”.

09Deployment

Independent views, one conclusion per part

Station count can merge with mechanical space; viewing directions stay independent:

• Station A · hook & bracket: fixed local views for defects 01 and 02 (short shot, break, structural integrity).
• Station B · top snap: fixed lighting and pose for defect 03 morphology.
• Station C · edge burr: side profile, baseline fit, heights for defects 04 and 05.

Fixtures limit pose drift; shields and light recipes are per station. The edge unit runs multi-station inference, HMI and archive; arrival I/O / PLC triggers capture and receives pass signals. The front end covers live view, NG alarm, manual review and search by time / part ID / defect; recipes manage cameras, model versions, thresholds and station enable—changeovers edit configuration, not a whole line software rewrite.

Equipment list

The table fixes categories and purpose only; exact models freeze after production samples, mechanical space and cycle time are reviewed.

Delivery follows freeze requirements → station trials → mass-production samples → model & repeatability → onsite acceptance; hardware, lights, models and thresholds freeze only after validation. Fit: plastic appearance QC with frequent part/defect change and multi-view local checks. Deep cavities or burrs far off the imaging plane need process help or an orthogonal view.