AI visual inspection is automated optical inspection that uses learned models running on edge hardware to find defects on a production line. A complete system is cameras and optics, controlled illumination, encoder-locked triggering, inference hardware, a PLC handshake and reject actuation — and the part that decides whether it works is almost never the model.
Scratches, dents, inclusions, porosity, contamination, coating and plating faults.
Gauging, edge and hole position, gap and flush, profile against tolerance with a stated uncertainty.
Components fitted, fasteners present, seals seated, contents correct, count verified.
Orientation, sequence, mating, connector engagement and correct variant.
Print quality, label placement and skew, legibility, wrinkle and code grading.
Date, lot and serial reading on print, laser mark and engraving, verified against expected content.
Shade matching, gloss, weave, grain and finish consistency.
Bead geometry, continuity, spatter, undercut and porosity.
Resolution derived from the smallest feature that must be resolved, field of view from the mechanical envelope, depth of field from part variation. Line scan where the part moves continuously, area scan where it indexes.
Geometry chosen for the defect physics — low-angle for surface relief, dome for specular and curved, backlight for dimension, coaxial for flat specular, structured for three-dimensional form.
Capture locked to part position rather than to a timer, with strobe timing and exposure short enough that motion blur does not consume the resolution the optics bought.
Classical vision where it is deterministic and sufficient, learned models where appearance varies beyond what a rule can express. Running on edge hardware sized to the cycle time.
Trigger in, result out, within the window the line allows. Digital I/O or fieldbus, with an agreed protocol for timeout and fault conditions.
Air blast, diverter, gate or robot pick, fired against the part the verdict belongs to — which means part tracking, not just detection.
Result attached to a batch or serial record, images retained per the traceability policy, available for audit.
Uncertain and rejected cases captured for review, feeding the next model revision.
The cycle-time budget. A line running 120 parts per minute allows 500 ms per part end to end, of which capture, transfer, inference and decision must fit within whatever remains after the mechanical motion. That number constrains camera, resolution, model and inference hardware simultaneously, and it is fixed before anything else is chosen.
Part tracking. A verdict is useless unless it reaches the right part. On an indexed station that is straightforward. On a continuous conveyor with a reject station three metres downstream it requires encoder-based tracking and a shift register that survives line stops and restarts.
The fault protocol. What the line does when the vision system does not answer in time, when the camera fails, when the illumination degrades — agreed with the controls engineer at design time, not discovered at commissioning.
Traceability. In regulated production, which image produced which verdict under which model version has to be answerable years later. That is a data-retention design decision, not a logging afterthought.
A defined protocol: which parts, which conditions, which defect classes, at what rate, with what metadata.
A written labelling standard, inter-annotator agreement measured, and a review process that keeps the dataset trustworthy as it grows.
Architecture chosen against the latency budget. Rare classes handled by augmentation, synthetic data or an anomaly-detection framing rather than pretending the imbalance is not there.
Against a golden sample set held out from training, with confusion analysis weighted by the real cost of each error rather than by count.
Quantised, compiled and measured on the actual inference hardware — worst-case latency, not average.
Input statistics, output distribution and reject rate tracked per station, because a fleet average hides a single failing cell.
Triggered by monitoring rather than by calendar, using data the review queue has already labelled.
Every inspection system has these. Naming them early is what separates a deployment from a pilot.
| Constraint | Consequence | How it is handled |
|---|---|---|
| Lighting geometry | A defect with no contrast under the installed illumination cannot be detected | Imaging trial on real defective parts before hardware selection |
| Occlusion | Features hidden by fixturing or part orientation are not inspectable | Multi-view capture, or a fixturing change agreed with the line owner |
| Rare defect classes | Too few examples to train a classifier | Anomaly detection trained on good parts, augmentation, or synthetic generation |
| False-reject cost | Over-sensitive systems destroy throughput and get bypassed | Threshold set against the measured cost of each error type, not against an accuracy target |
| Cycle time | Inference must complete within the window the line allows | Budgeted first; camera, resolution and model chosen to fit it |
| Part variation | Legitimate variation misread as defect | Training distribution that spans the real variation, including suppliers and seasons |
Real defective parts photographed under several illumination geometries. Output is a written answer on whether the defect is detectable, and what would be needed if it is not.
Camera, optics and lighting selected and proven on the bench against the actual part set.
One station fully integrated, running in parallel with existing inspection long enough to compare against it honestly.
PLC handshake, reject actuation, MES connection, operator interface and fault handling commissioned.
Documentation, retraining pipeline, monitoring dashboards and the training that lets the site run it without Faststream.
Indicative ranges. A specific cell is specified against its own part set, line speed and defect classes.
| Parameter | Typical range | Driven by |
|---|---|---|
| Camera resolution | 2–20 MP area scan; 2k–16k line scan | Smallest feature to resolve across the field of view |
| Line speed | Static indexing to several metres per second | Process; determines exposure and blur budget |
| Inference latency | 5–50 ms typical at the edge | Cycle time remaining after mechanical motion |
| Decision-to-actuation | Within one station pitch | Reject mechanism position and conveyor speed |
| Environmental rating | IP54 to IP67 enclosure typical | Washdown, dust, coolant and vibration at the station |
| Image retention | Rejects only, to full retention | Traceability and regulatory policy |
AI visual inspection is automated optical inspection that uses learned models, running on edge hardware, to classify or locate defects in images captured on a production line. A complete system comprises cameras and optics, controlled illumination, encoder-locked triggering, inference hardware, a PLC handshake and reject actuation.
The PLC issues a trigger when a part is in position and expects a pass or fail result within an agreed window. The result is returned over digital I/O or fieldbus, and reject actuation is fired against the tracked part. A fault protocol defines what the line does if the vision system does not answer in time.
It depends on the cycle-time budget rather than on the vision system alone. Edge inference typically completes in 5 to 50 milliseconds; whether that fits depends on how much of the station cycle remains after mechanical motion. The budget is established before cameras are chosen.
Not by supervised classification, which needs examples. That case is handled by anomaly detection trained on good parts, by augmentation and synthetic generation, or by reframing the acceptance criterion. Which of those applies is decided in the feasibility study.
By monitoring rather than by hoping. Input statistics, output distribution and reject rate are tracked per station; low-confidence cases are routed to a review queue that generates labelled data; and retraining is triggered by the monitoring rather than by a calendar. Hardware drift — lamp ageing, lens contamination — is checked separately, because retraining will not fix a dirty lens.
No. Dimensional gauging, presence and absence checks and code reading are usually better served by classical vision, which is deterministic, explainable and does not drift. Learned models earn their place where defect appearance varies in ways a rule cannot capture.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.