Recognition of incorrect assembly of internal components by X-ray CT and deep learning

Abstract

It is important to make sure that all components of a complex product are assembled correctly. Because in many cases, some components are enclosed in an opaque shell, X-ray imaging is currently used to extract their characteristics and match prior-known ones. However, X-ray imaging is not very robust in recognition of incorrect assembly of internal components, because some of them may overlap. To solve this problem, we propose a new method to detect internal component assembly fault, by X-ray computed tomography (CT) and convolutional neural network (CNN). Multi-view imaging is implemented by mechanical rotation of a product in respect with an X-ray CT machine to capture multiple projection information on each internal component, and then the component can be recognized by making use of deep learning. A CNN model is trained to classify the internal components and give the coordinates of each component. Based on the CNN recognition results and the CT projection sinogram, a projection corresponding to a reference in a projection data set of a standard product can be found. By comparing and matching the locations of each component, transposition or dislocation can be recognized. Both simulation and experiment show that this new method can effectively identify incorrect assembly, missing assembly, transposition, and other problems, improving the product quality.