Automated quality inspection across six assembly lines

Context

An electronics assembly plant in Bac Ninh running six lines across three shifts a day. Visual inspection was performed by twelve people per shift, by eye.

The challenge

Miss rates climbed noticeably towards the end of each shift, and results were inconsistent between inspectors. Customer complaints typically arrived two to three weeks later, once the batch had already shipped.

What BigAI delivered

A two-day on-site assessment showed the biggest constraint was not the algorithm but the lighting: the lines used overhead fixtures that cast shadows, making captured images inconsistent. The first work item was redesigning localised lighting for each camera position.

We then collected and labelled roughly 400 images per defect class, trained a detection model, and deployed it to edge devices sitting directly on the line. Inspection results feed the MES and a real-time quality dashboard.

Results

After five months, defects reaching customers fell from 2.3% to 0.85%. Inspection staff moved to supervisory roles handling the cases the model flags as uncertain.

Results

Swipe to see the full table

We assumed the problem was the model. The site assessment showed the first thing to fix was the lighting on the line.

Technology used