Automotive PCB defects — classic AOI does not learn the new component; iLEAN Vision does.

The problem

Classic AOI is excellent with what it knows; blind to what is new.

An SMT line for automotive electronics (BCM, ECU, ADAS systems, infotainment) has run classic AOI for decades — and run it well. The template- and rule-based system works for the defects the process engineer already knows and has coded. The problem is whatever has not been coded yet:

1. New components with no validated recipe. Every NPI (new product introduction) means weeks of tuning the AOI recipe before it stops screaming.
2. Supplier changes. A functionally equivalent MLCC with a different marking fires false positives for a week — and the team learns to ignore them. Until a real one slips through.
3. Subtle defects — marginal cold solder, false solder joints, paste halo, a 5° rotation on an SOIC — that sit in the grey zone of the rules.
4. New defects that appear because of an upstream change (new paste, new oven profile, new PCB batch) and that the old recipe was not looking for.

The classic outcome: the AOI produces a false positive rate that saturates the rework team (and teaches them to ignore it), or it gets tuned down until it goes quiet and then lets the subtle stuff through. In automotive, the subtle stuff reaches the customer's ECU and, one day, comes back as a warranty failure or, worse, as a safety recall.

How it fits the IRIS system

iLEAN does not replace your AOI — it learns what your AOI cannot code and cross-references the root cause.

The problem is not the AOI; the problem is that the reality of the line changes (new component, new supplier, subtle defect) faster than a human can code rules. iLEAN acts as the putty that fills that gap with a CNN which learns from real examples from your plant, without asking you to throw the AOI away.

Classic AOI sees what has been coded. iLEAN Vision sees what is new, what is subtle and what changes. The agent cross-references both, notes the upstream root cause and proposes to the process engineer what to fine-tune.

The three iLEAN pieces applied to automotive PCB defects:

• Edge + Vision — a terminal with a CNN over the line (between the pick-and-place and final inspection, or as a second read after the classic AOI). It learns your process's defects from real flagged boards; retraining in hours, not weeks. It works with no network — inspection continues even if the plant loses WiFi.
• Connect — captures the classic AOI log, the X-ray output (if there is any), the oven recipe, the solder paste batch, the PCB part number and the supplier data from the ERP. And it captures what comes in from outside (an alert from the component supplier, a quality note from the customer) at second zero, with nothing to forward.
• Agents — cross-reference the optical verdict, the classic one and the process data; they correlate today's cold solder spike with yesterday's paste batch change and tell the engineer. They generate a per-panel dossier for a customer complaint. The person validates any change to the recipe — the line does not restart on its own after a hold.

See the full IRIS architecture →

Before and after

Classic AOI alone vs. AOI + iLEAN Vision with root cause correlation

Aspect	Classic AOI alone	AOI + iLEAN Vision + Connect + Agents
New component	Weeks tuning the recipe	Hours retraining with dozens of examples
Supplier change	False positives for days	The CNN learns both variants; the agent traces it to the batch
Cold solder / false solder joint	Grey zone, depends on thresholds	Learned from the real visual signature of your line
New defect (from upstream)	Slips through until a human codes it	The CNN detects it; the agent correlates it with the upstream change
Unified optical + X-ray verdict	Two systems, two sheets	One dossier per panel, cross-referenced
Operation with no network	n/a	Edge keeps inspecting locally

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the immersion diagnostic.

• SMT line with a classic AOI installed, 8-15 active part numbers, frequent NPIs and recurring supplier changes.
• Edge + Vision pilot as a second read after the AOI on the part number with the most false positives or the most rejects. First value expected within a few weeks: false positives down and the subtle defect detected.
• Indicative payback between 4 and 9 months, depending on annual volume, rework cost, component scrap cost and the weight of warranty campaigns with “PCB defect” as root cause.
• Hard lever: a ≥ 30% reduction in false positives and a ≥ 30% reduction in the NPI time needed to get reliable AOI on a new part number (estimate to be validated). A single warranty campaign avoided because of cold solder pays for the pilot.

The standard that measures whoever audits you, and the CAIO's reasonable doubt

Automotive quality operates in the order of 25 PPM (parts per million) of admissible defects ^[1]. For automotive electronics with functional safety, that floor is a ceiling — and classic AOI only reaches it with long NPIs and hand-tuned templates. The CAIO's reasonable doubt (“what if the CNN hallucinates on a new component?”) is defused: AI here invents nothing — it classifies an image against real examples from your line. In tasks anchored to the source, the best models bring error below 1.5% ^[2]. And even so, what is critical is never decided alone: iLEAN holds the suspect panel, the person signs.

^[1] The 25 PPM automotive quality standard — Symestic.
^[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.