Pengembangan Sistem Inspeksi Kualitas Otomatis Berbasis YOLOv8n untuk Deteksi Cacat Pin pada PCB Controller

Abstract

Quality inspection of pins on PCB Controllers at an Electronics Manufacturing Company in Batam currently relies on a template matching-based system using a Keyence CV-X camera, which produces a high False Reject Rate (FRR) due to excessive sensitivity to minor image variations. This study proposes a deep learning-based verification layer using the YOLOv8n model as Stage 2 within a dual-stage verification architecture to reduce FRR in pinbent inspection. The YOLOv8n model was trained through a transfer learning approach on a dataset of PCB Controller pin images, exported to ONNX format, quantized to INT8, and integrated into a VB.NET desktop application via ONNX Runtime. Testing was conducted in Standalone Inspection Mode on 524 PCB Controller units from actual production data, using an inference pipeline identical to the production mode encompassing pin-area cropping, letterbox resizing, ONNX inference, and hierarchical decision logic. Evaluation results show that the model achieved an [email protected] of 96.68% and a recall of 99.82% on pin-level validation data. At the system level, the YOLOv8n model increased overall accuracy from 58.78% to 85.50% and reduced FRR from 39.54% to 13.87% a relative reduction of 64.9%, which was statistically significant based on McNemar's test (p < 0.001). These findings demonstrate that a YOLOv8n-based deep learning approach effectively addresses the limitations of template matching systems for pinbent inspection in manufacturing environments.