AOI false calls and PCB defects: an AI second-look guide

What AOI false calls are, and the PCB defects behind the calls

Automated optical inspection (AOI) is the camera-based check that surface-mount lines run after placement or after reflow (solder paste is usually checked earlier by a separate SPI machine). The machine images the board, compares each component and solder joint with its inspection program, and flags every location that falls outside the program's limits. An operator then verifies each flagged location — the call — at a review station: real defects go to repair, the rest are false calls and the board moves on. A false call costs the operator's time; a real defect that gets through costs far more, as an escape found at test, at the customer or in the field.

Why AOI produces false calls

• Thresholds tuned tight. Nobody wants an escape, so inspection limits are set on the safe side. A tight limit catches more real defects and also flags more joints and parts that are within acceptance.
• Lighting and reflection. Solder is shiny. A glossy fillet, a reflective component body or a shadow from a tall neighbor can make an acceptable joint look different from what the program expects.
• Component variation. The same part from a second source can have a different body color, marking or lead finish. A program built on one supplier's part can flag the other's as missing, shifted or wrong.
• Rule-based libraries. The inspection program describes each part with fixed rules — windows, color ranges, dimensions. Rules don't know what the part is for, or what a human inspector would accept; they only know whether a measurement is inside its limits.

The defects that matter

Behind the false calls are the defects the line exists to catch. The common acceptability standard for electronic assemblies, IPC-A-610 (Acceptability of Electronic Assemblies), describes what acceptable and defective joints and placements look like. On a surface-mount line, the defect types a second look is asked to read include:

• Solder bridges — solder joining two pads or pins that should be separate, a risk on fine-pitch ICs.
• Insufficient or excess solder — a joint without enough fillet to be reliable, or so much solder that it hides the joint or risks a bridge.
• Tombstoning — a small chip component standing up on one end, soldered on one pad only.
• Missing, shifted or rotated parts — a part not placed, placed off its pads, or turned on its pads.
• Polarity — a diode, electrolytic capacitor or IC placed the wrong way round.
• Foreign material — solder balls, debris or other material on the board where it doesn't belong.

Why traditional methods miss real PCB defects

AOI review relies on people: every call goes to an operator who looks at it and decides. That works while the queue is short. As it grows, the method itself becomes the risk.

• Fatigue. Verifying call after call is repetitive work. Attention drops over a shift, and the decision becomes a habit rather than a look.
• Real defects waved through. When the operator expects a false call, a real bridge or a reversed part can be accepted along with the false ones. The defect was flagged — and still escaped.
• Loosening thresholds trades one problem for another. Opening the inspection limits reduces false calls and lets more real defects pass the machine unflagged. Tightening them sends the queue back up.
• Program tuning never ends. Each new product, part source or process change can bring a new wave of false calls, and the library is tuned one rule at a time.
• Sampling and end-of-line checks come late. A sample audit or a functional test can find what AOI review missed, but by then more boards with the same defect may have been built.

What these methods lack is the judgment an experienced inspector brings to every call: what is this part, what does the call claim is wrong, and would this joint be accepted?

How vision-language models make it solvable

Primarch builds on and enhances vision-language models (VLMs): models that read images and language together. That changes what a second look can be. Instead of measuring a joint against another fixed window, the model is designed to read the flagged location the way an experienced inspector would — and to do it for every call, on every shift.

• It reads the call, not just the pixels. The model sees the part, its pads, its joints and its neighbors together, and reads the location in the light of what the AOI flagged it for.
• It brings domain knowledge to the image. What a solder bridge looks like, what makes a fillet acceptable, which way a polarized part must face: this knowledge is put on top of the model, so the reading follows the rules an inspector works to rather than a raw measurement.
• It copes with variation in context. A reflective fillet or a part from a second source is read for what it is, rather than for how far it is from one stored image. Reading in context is designed to address the cause of false calls described above: limits that measure without knowing what they are looking at.
• It explains its verdict in plain language. Each result is a sentence an operator or engineer can read — what was flagged, what the second look sees, why it counts as a false call or a defect — next to the image it was read from.
• It remembers. Results are kept, so repeats can be grouped by line and shift, and a process engineer can see which defects keep returning.

The useful question is not "is this measurement inside the window?" It is "would an experienced inspector accept this joint?" The second look is built to answer the second question for every call.

A second look can only read what the image shows. Joints hidden under a component, such as the balls under a BGA, are not visible in an optical image of the board.

What to look for in a solution

Whichever vendor you talk to, these are fair questions to ask about a system meant to take a second look at AOI calls:

• Works with the AOI you already run. How does it receive the calls and the images — and does it depend on one AOI brand?
• Uses your acceptance criteria. Can your workmanship standard and your customers' requirements be built into how it decides, or does it apply its own idea of a good joint?
• Explains every verdict. Each call should come back with a reason a person can check, not just a pass or a fail.
• Handles uncertainty openly. Ask what happens when it isn't sure, how calls it reads as false are handled — shown, sampled or logged — and who decides that policy.
• Is proven on your boards. A pilot on your own products and your own calls, judged against measurable criteria agreed before it starts — not a demo set.
• Shows patterns, not just verdicts. Can defects be grouped by line and shift, so engineers see what keeps coming back?
• Keeps your data with you. Board images and designs are sensitive. Is there an on-premise option where the data never leaves your facility?
• Adapts without retraining from scratch. When a new product, part source or defect type comes along, how much work is it to update?

Example use cases

The walk-throughs below are illustrative examples of how a second-look result is built. They are not customer cases and contain no performance figures.

Example 1: a shiny fillet flagged as insufficient solder

1. 01

   Moment

   A surface-mount line, after reflow, on a control board with a row of chip capacitors near a tall connector.

2. 02

   What the AOI flagged

   Insufficient solder on one end of a chip capacitor next to the connector.

3. 03

   What the second look reads

   The fillet is wetted along the termination and has an acceptable shape. It looks darker on one side because the tall connector beside it shadows it.

4. 04

   The finding + image

   "Likely false call: the solder joint on the right-hand end of the capacitor next to the connector is well wetted with an acceptable fillet; the flagged difference comes from the connector's shadow." The image of the location is attached.

5. 05

   Who sees it

   The review station operator, who sees the call with the reason next to it.

6. 06

   What the pattern shows

   Asked in plain language — "show false calls on this board type this week" — the memory returns calls on capacitors next to the same connector, which the AOI programmer can use to review that part of the inspection program.

Example 2: a real solder bridge among the calls

1. 01

   Moment

   Line 2, early in the morning shift, on a board with a fine-pitch IC.

2. 02

   What the AOI flagged

   A possible bridge between two pins of the IC.

3. 03

   What the second look reads

   Solder joins the two adjacent pins along the toe of the leads; the gap between them is closed.

4. 04

   The finding + image

   "Defect: solder bridge between two adjacent pins of the fine-pitch IC. Send to repair." The image of the location is attached.

5. 05

   Who sees it

   The review station operator, with the call marked as a defect, and the repair station.

6. 06

   What the pattern shows

   Bridges the second look read on this IC fall at the start of the morning shift on line 2. The process engineer checks the paste printing at shift start.

Example 3: a second-source part flagged as missing

1. 01

   Moment

   A new batch of boards built with a resistor from a second source.

2. 02

   What the AOI flagged

   Missing component at several resistor positions on every board.

3. 03

   What the second look reads

   A resistor is present, placed on its pads and soldered at both ends. Its body is a different color from the part the program was built on.

4. 04

   The finding + image

   "Likely false call: the resistor is present and soldered; its body color differs from the one the AOI program expects." The image of the location is attached.

5. 05

   Who sees it

   The review station operator; the false calls also appear in the shift digest for the AOI programmer.

6. 06

   What the pattern shows

   The calls start with the new batch and repeat at the same positions on every board, which tells the programmer to add the second-source part to the inspection library.

Example 4: a reversed electrolytic capacitor

1. 01

   Moment

   A power board after reflow, with several electrolytic capacitors.

2. 02

   What the AOI flagged

   A possible polarity error on one capacitor.

3. 03

   What the second look reads

   The capacitor's polarity marking faces the opposite way from the polarity mark printed on the board.

4. 04

   The finding + image

   "Defect: electrolytic capacitor placed with reversed polarity; its marking is opposite the board's polarity mark. Send to repair." The image of the location is attached.

5. 05

   Who sees it

   The review station operator and the repair station; the defect also appears in the quality engineer's digest.

6. 06

   What the pattern shows

   No other boards in the run show the same call, so it stays a single event in the record rather than a pattern.

How Primarch's Custom Experts help

This section is about our product. Custom Experts put your domain knowledge on top of the Primarch platform core, for problems that can be seen but are not in a catalog. We don't train models from scratch; what makes the expert custom is the knowledge. For AOI calls and PCB defects:

• Takes a second look at AOI calls. A detailed PCB inspection runs as a post-filter on the locations the AOI machine flagged, reading each call and separating likely false calls from real defects.
• Reads PCB defects such as a solder bridge. Visual quality control in electronics, with defect detection and classification on the line. Which defect types the expert reads is defined with you in discovery and tested in the pilot.
• Writes each finding in plain language, next to the image it was read from, so the operator and the engineer see the same thing.
• Keeps a queryable memory. Findings are written into a memory you can ask in plain language, and repeats can be chained by line and shift.
• Reads more than line cameras. Photo archives, periodic captures and microscope output can also feed the expert.
• Works from the images the AOI machine already saves for each call, so no extra camera is needed at the review station.
• Uses your acceptance criteria. Your acceptance criteria and your customers' requirements are part of the domain knowledge the expert is configured with.

Putting it to work

1. 01

   Discovery

   We define the problem and the success criteria together: which boards and calls, which defect types, which decision the second look supports.

2. 02

   Configuration

   An expert profile is built from a domain knowledge base, rule templates and thresholds; targeted fine-tuning if needed.

3. 03

   Pilot

   Validation on real images from your own line, tested together against measurable criteria. Depending on scope, discovery to pilot typically takes a few weeks.

4. 04

   Deployment

   On-premise or cloud rollout, team training and a continuous improvement loop.

• Alerts, digests and dashboards. Instant notifications, periodic digests and role-based dashboards, so each finding reaches the operator, engineer or manager who acts on it.
• Your data stays with you. With on-premise deployment, development and validation can happen inside your facility; your board images and data never leave.