3-D Inspection Method for Industrial Product Assembly Based on Single X-Ray Projections

Abstract

Industrial product assembly inspection is an indispensable part of industrial production, directly affecting product performance and reliability. Due to the capsulation of many products, most of the current inspection methods are based on X-ray technologies, such as digital radiography (DR) and computed tomography (CT). The inspection methods based on DR are fast but cannot obtain the spatial distributions or interrelationships of the assembled components and are easily affected by the occlusion of internal components. The detection methods based on CT can reconstruct the 3-D volume of a product and have high detection accuracy but cannot achieve real-time inspection because of the reconstruction speed. To address the above problems, we propose an inspection method based on a single-projection reconstruction network for industrial product assembly. An improved autoencoder network is trained to learn the shared structural features in 3-D volumes from 2-D projections, realizing the reconstruction of the 3-D volume at a specific viewing angle from a single projection from any viewing angle. The 3-D volumes output by the network are used to detect the position and posture of the components in the product to judge whether the product has been assembled correctly. The feasibility of the proposed method is verified with the data of a certain fuse under different assembly conditions. The experimental results show that the proposed method can improve the inspection efficiency, reduce the inspection costs, and simplify the imaging system.