Skip to main content

4.2 AOI Fundamentals

Lighting/angle, rule-based vs ML libraries, and keeping false calls low without letting escapes slip through.

Automated optical inspection (AOI) is the production line’s set of trained eyes, combining cameras, controlled lighting, and software to spot visual risks before boards move downstream. It excels at confirming part presence, orientation, position, and visible solder quality, but it’s not a substitute for electrical tests. Lighting choice and camera angles are the real performance levers—ring light for edges and markings, low-angle light for fillet contours, coaxial for reflective surfaces, and side or 3D views for height and heel visibility. Inspection logic can be purely rule-based, using geometry and contrast, or machine learning–assisted for subtle, variable features like fillet smoothness; most stable systems blend the two. False calls waste touch-up time, while escapes let real defects slip through, so thresholds are tuned by defect risk rather than aesthetics. By keeping programs tied to a “Golden Board” reference, versioning lighting profiles, and actively linking AOI results to downstream feedback, the system stays consistent even as materials and builds change.

4.2.1 What AOI is (and what it isn’t)

AOI is a fast, consistent camera + lighting + software check that answers three questions:

  1. Is the right thing there? (presence, polarity, value markings)
  2. Is it where it should be? (X/Y/θ offset, lift/tilt)
  3. Does it look soldered? (bridges, insufficient/fillet shape, wetting clues)

It’s not a lie detector for every electrical fault. AOI catches visual risks early; ICT/FCT catch electrical or parametric faults later. Use both.



4.2.2 Where to put AOI (pre- vs post-reflow)

  • Pre-reflow AOI (after placement): perfect for orientation/polarity, presence, wrong footprint, and big offsets before you bake mistakes in.
  • Post-reflow AOI: adds solder joint judgement (bridges, opens, tombstones, fillet quality).
     Most lines run post-reflow as the main gate and use targeted pre-reflow checks for risky builds (e.g., lots of polarized parts).



4.2.3 Lighting & angles—your biggest quality knobs

Think like a photographer: the wrong light makes good joints look bad, and vice-versa.

A) Lighting modes (pick the fewest that work)

Mode

What it highlights

Use it for

Ring / bright-field

General edges, silks, text

Presence/offset, OCV/OCR, polarity marks

Low-angle / dark-field

Tiny height changes, fillet edges

Bridges, lifted leads, tombstones

Coaxial (on-axis)

Flat, reflective surfaces

BGA masks, code reading on shiny parts

Structured/3D (fringe)

Height map of pads/parts

Lifted QFN corners, bent leads, solder mound height

B) Wavelength & color

  • White RGB covers most; blue can pop solder fillets; red tames mask glare; UV helps fluorescing conformal coats (post-coat lines).
  • Mask color/finish matters: OSP matte vs ENIG shiny will want different gains/angles. Save a lighting profile per product.

C) Camera geometry

  • Top camera (2D/3D) does the bulk of work.
  • Side cameras (oblique ~25–45°) see heel fillets on gull-wings, lifted QFN edges, and hidden bridges along tall parts.
  • If you have 3D AOI, use height to demote nuisance calls (e.g., silkscreen glare that 2D thinks is a bridge).



4.2.4 Rule-based vs ML libraries (and when to use which)

Approach

How it thinks

Strengths

Watch-outs

Rule-based (thresholds, geometry, color)

“Edges here, contrast there, pad X has area Y.”

Transparent, fast to tune; great for presence/offset/polarity and clear defects.

Brittle across finish/mask changes; many tiny rules to maintain.

ML-assisted (template/CNN classifiers)

Learns “this is OK vs NG” from examples.

Better at subtle fillet/joint quality, varied lighting, component cosmetics.

Needs curated training images; beware overfitting and hidden bias.

Practical blend: use rules for the must-haves (polarity, offset, bridges), and sprinkle ML on the subjective bits (fillet quality, cosmetic scratches). Keep ML outputs explainable (scores + example images).



4.2.5 Building a stable AOI program (simple recipe)

  1. Start from FA: teach on your Golden Board photos, not a random sample.
  2. Teach minimal features per part:
    • Presence box, pin-1/polarity locator, and pad windows for solder checks.
    • Avoid teaching glossy logos or changing lot codes as truth.
  4. Light sanity: lock exposure/gain per product. If you need more than two lighting modes on most parts, improve lighting before adding rules.
  5. Guard-band smartly: tighter on fine-pitch and polarity-critical parts; looser on cheap resistors that AOI will feed to process SPC anyway.



4.2.6 False calls vs escapes (how to balance)

  • False call = AOI flags a good feature → wastes touch time.
  • Escape = AOI misses a real defect → hurts yield/field.

Targets that keep lines calm (tune to your product):

  • False calls: ≤ 0.5–1.0 per board average, with a cap per panel.
  • Escapes: 0 on critical classes (polarity, bridges on high-risk nets); very low on others, verified by audit sampling.

Levers to pull

  • Use 3D/side-view to reduce nuisance calls on fillets.
  • Create risk classes: Class A (polarity/bridges) = strict; Class B (cosmetics) = tolerant.
  • Route uncertain calls to a review queue with zoomed crops; don’t slow the belt for a beauty contest.



4.2.7 Controlling drift (why yesterday’s good board fails today)

  • Mask/finish changes between lots change reflectivity → keep lighting profiles versioned by product and finish.
  • Lens/cover contamination raises false calls → clean optics on a schedule.
  • Board warp changes apparent fillet shape → fix supports upstream (Ch. 8.4) and let 3D judge height, not glare.
  • ECNs → treat AOI like code: rev the program when land patterns, silks, or part numbers change; attach the change note.



4.2.8 Metrics that matter (and ones to ignore)

Track per product, per side:

  • False calls/board (and by top 10 refdes)
  • Escapes found at ICT/FCT/rework (with AOI image back-link)
  • Top 5 AOI defect categories (bridges, tombstones, polarity, opens, cosmetic)
  • Review rate and auto-pass rate
  • Time to clear a board (don’t let AOI be the bottleneck)

Ignore “total defects counted” without context—it’s often just lighting noise.



4.2.9 Fast troubleshooting (AOI says NG—what now?)

  • Bridge calls spiking? Check stencil cleanliness and separation speed (7.5), then revisit lighting angle; don’t just widen thresholds.
  • Polarity NGs on LEDs/diodes? Verify silks & pin-1 marks are visible and consistent; switch to coaxial or add a simple OCR/OCV on the mark.
  • Fillet insufficient calls on gull-wings? Add side-view or 3D height; tune reflow TAL/peak slightly (9.2/9.4) before rewriting half the library.



4.2.10 Release checklist (stick this near the AOI)

  • Lighting profile saved (mode, angle, gains) for this product/finish
  • Program taught from Golden Board; pin-1/polarity checks on all polarized parts
  • Rule vs ML split documented; ML thresholds show scores + example images
  • False call and escape targets set; Class A defects strict, Class B tolerant
  • Optics clean, calibration check OK; side/3D cameras enabled where needed
  • Feedback loop active: escapes at ICT/FCT link back to AOI images/program rev



Bottom line: AOI is only as good as its light and teaching. Use the simplest lighting that reveals the truth, blend rule-based checks with ML where judgement is fuzzy, and tune thresholds by risk class—not vibes. Keep programs tied to the Golden Board, watch the two numbers that matter (false calls, escapes), and AOI becomes a quiet guardian instead of a noisy critic.


Back to top