Post-bond defects like sagging, shift, and missing wires are subtle in scale; manual inspection is slow and fatigue-prone to escapes, leaving capacity and quality at constant odds.
After wire bonding, this packaging plant must check each metal wire's loop height, routing, and connection. Defects like sagging wires, wire shift, and missed bonds are at the tens-of-microns scale. Previously inspectors checked unit by unit under a microscope, limited by human eyesight, with fatigue-induced escapes over long shifts.
Once an escaped defect flows into downstream molding, it is nearly impossible to rework, scrapping the finished unit and reaching the customer. To suppress escapes the line could only slow its pace and raise sampling ratios, keeping quality and capacity perpetually at odds.
DaoAI Solution
DaoAI deployed the trained wire-defect model onto the edge of the AI-AOI inspection device for inline per-unit inference: the device handles image capture, detection, and classification locally, with no server round-trip. The model was iterated on field samples for specific morphologies such as sagging wires, shift, and missing wires; rare defects were filled in with APDT positive-sample learning, avoiding the need to manufacture defects for negative samples.
- Defect model pushed to the AI-AOI edge, per-unit inline inference without round-trip
- Covers typical defects: sagging wires, wire shift, missed bonds
- Low-latency edge inference matched to bonder takt
- Results drive real-time sorting so defects never reach molding
With the model at the edge, detection decouples from takt, ending the quality-versus-capacity tradeoff.
After deployment, detection accuracy at this step reached 98%, edge inline inference roughly doubled inspection throughput versus manual visual check, and critical-defect escapes approached zero. Inspectors shifted from full-inspection posts to equipment maintenance and anomaly confirmation, letting the line speed up while holding quality.