99.4% Defect Detection on a High-Mix PCBA Line

Client

Who We Worked With

A contract electronics manufacturer producing high-mix, low-to-medium-volume PCBAs for automotive, industrial, and consumer customers, with frequent product changeovers and tight cycle times.

Problem

The Business Problem

Manual end-of-line inspection was the bottleneck. Inspectors were missing fine-pitch solder defects, tombstoning, and incorrect component placement at unacceptable rates - leading to escapes, customer complaints, and costly rework on already-assembled boards. Existing AOI systems struggled with the variety of board designs.

Baseline

Where They Started

• Manual inspection accuracy estimated at 88 to 91 percent
• Average of 320 boards per inspector per shift
• False-call rates on legacy AOI ranging from 12 to 18 percent
• Escapes to customer measured in hundreds of ppm

Data

Data Sources Used

• Top-down RGB images from line-mounted industrial cameras
• Reference Gerber and pick-and-place data per board variant
• Historic defect-labelled images from past production runs
• MES production-order data for board variant identification

Solution

What We Built

We trained a sequence of computer-vision models - a board-variant classifier, a component-level segmentation model, and defect-type classifiers for solder, placement, and missing-component classes. Models run on edge GPUs at the inspection station with sub-second latency. Confidence scores route ambiguous boards to a human review station, and every decision is logged for continuous retraining.

Integration

Systems We Connected

• Industrial cameras and edge GPU stations at the end-of-line
• MES integration to fetch board variant and routing data
• Defect events written back to MES and traceability database
• Operator HMI for confirm/reject of borderline calls
• Active-learning loop feeds confirmed labels back into retraining

Timeline

Project Timeline

Pilot on a single board variant in 6 weeks. Expansion to top 20 variants representing 80 percent of volume in 10 weeks. Full-line rollout and operator training completed in 4 months.

Governance

Security & Compliance

• Models deployed and retrained inside the client's private network
• Versioned model registry with rollback for every line release
• Daily drift and accuracy monitoring with operator-visible dashboards
• Human override on every reject decision is logged and reviewable
• Client owns the model weights, training data, and labelling pipeline