An in-process inspection method integrating deep learning and classical algorithm for automated fiber placement

Abstract

The use of depth cameras, such as a laser profilometer, in defect detection during the automated fiber placement (AFP) process is of great importance as they improve AFP’s layup quality and efficiency. However, the collected point clouds are large and of heterogeneous densities. Several works have projected 3D point clouds onto 2D images, but the method has only led to loss of key geometric details. In this study, we proposed a two-stage segmentation method for collected point clouds during the AFP process, named AFP-Seg, to realize AFP in-process inspection. In the first stage, the point clouds fused with collected laser lines and sampling position data are fed into a semantic segmentation network, and the semantic label of each point can be obtained, and in the second stage, point clouds with specified semantic labels are clustered using the post-process algorithm. Through the AFP-Seg method, information about defects regarding types, sizes, and positions are acquired eventually. Compared with our previous method, the AFP-Seg method can reduce the data processing time by 33–66% and obtain more robust and better inspection results. Furthermore, it can be well integrated into a real-time AFP in-process inspection system and easily adjusted based on actual engineering inspection specifications.