Development of Product Picking-Quality Inspection System through Convolution Filter and Ensemble-based Deep Learning Analysis for Distance Deviation Measurement and 2D Image

With increased customized personal production, vacuum grippers have been increasingly used to handle various products in a single manufacturing system. Vacuum grippers are less affected by the shape or material of the target part. However, regardless of the quality of the part, it is sometimes impossible for the gripper to successfully pick a part owing to a tiny curvature on the target surface. Therefore, we propose a non-destructive picking-quality inspection system. In this system, the distance deviation was measured in the form of a 2D grid, and an image of the gripping surface was captured. The collected measurements and images were applied to the convolution filter-based autoencoder and neural network, respectively, to evaluate the grip possibility. Consequently, the picking-quality score was calculated through the soft-voting-based ensemble method. The effectiveness of the proposed method was evaluated using 80 box data with 5-fold cross-validation. The proposed method demonstrated better performance with an accuracy of 99.8%.