State Determination of Raw Material and Finished Product Inside a Press Lower Mold Using AI

Abstract

This study presents a You Only Look Once version 8 (YOLOv8) segmentation-based method for determining whether raw materials are correctly inserted into the lower mold of a press, whether a finished product remains in the mold, and whether the raw materials overlap the finished product. An endoscopic camera was fixed at the center of the upper mold to capture the lower mold from a consistent position. Training data were collected in both laboratory and actual press environments and included raw materials, finished products, overlapped states, and material-feeding grippers. The raw material and finished product were polygon-labeled along their visible outer contours, and a YOLOv8 segmentation model was trained. The developed artificial intelligence program used material-feeding gripper detection as an image-acquisition trigger and performed a two-stage state determination using detected classes and the bounding box size of the raw material. All normal-state data were correctly classified, and overlapped cases in which only the raw materials were detected were identified as abnormal through the second-stage bounding box size criterion. The proposed method is therefore applicable to automatic material-state determination inside a press lower mold and may help prevent mold damage and equipment accidents.