YOLOv8 Performance for Leakage Detection Using Pneumatic Leak Test Images

Abstract

Leakage defects in automotive components can compromise product quality and reliability, making accurate inspection essential in quality control processes. At PT. AMI, leakage inspection of Cap Rear Sub Assembly components is performed using a pneumatic leak test by visually observing air bubbles in water, making the process susceptible to human error and inconsistent inspection results. This study evaluates the performance of the YOLOv8 object detection model for automated leakage detection using pneumatic leak test images. A dataset consisting of 1,000 images (500 OK and 500 No Good) was collected and annotated using Label Studio, with 862 air bubble objects labeled as leakage indicators. Three experimental configurations involving YOLOv8s and YOLOv8m models with different input image resolutions were evaluated using Precision, Recall, mean Average Precision at IoU 0.5 (mAP@0.5), and mean Average Precision at IoU thresholds from 0.5 to 0.95 (mAP@0.5:0.95). The results showed that the YOLOv8m model achieved the best overall performance, with a Precision of 83.82%, Recall of 73.45%, mAP@0.5 of 79.66%, and mAP@0.5:0.95 of 28.91%. The best-performing model was successfully deployed in a Flask-based web application capable of automatically detecting leakage by identifying air bubbles and classifying inspected components as OK or No Good. These findings demonstrate that the proposed YOLOv8-based approach can improve the consistency, efficiency, and reliability of leakage inspection, providing a practical solution for automated quality control in automotive manufacturing.