A deep learning based automatic surface segmentation algorithm for painting large-size aircraft with 6-DOF robot. Issue 2 (20th December 2019)

• Record Type: Journal Article
• Authors: Wang, Guolei; Hua, Xiaotong; Xu, Jing; Song, Libin; Chen, Ken
• Is Part Of: Assembly automation. Volume 40:Issue 2(2020)
• Page Start: 199
• Page End: 210
• Publication Date: 2019-12-20
• Subjects: Deep learning -- Automatic surface segmentation -- Robot painting; Automation -- Periodicals; Automatic machinery -- Periodicals; Assembly-line methods -- Periodicals; Industrial engineering -- Periodicals; 670.42705
• DOI: 10.1108/AA-03-2019-0037
• Languages: English
• ISSNs: 0144-5154

Abstract

Abstract : Purpose: This paper aims to achieve automatically surface segmentation for painting different kinds of aircraft efficiently considering the demands of painting robot. Design/methodology/approach: This project creatively proposed one method that accepts point cloud, outputs several blocks, each of which can be handled by ABB IRB 5500 in one station. Parallel PointNet (PPN) is proposed in this paper for better handling six dimensional aircraft data including every point normal. Through semantic segmentation of PPN, each surface has its own identity information indicating which part this surface belongs to. Then clustering considering constraints is applied to complete surface segmentation with identity information. To guarantee segmentation paintable and improve painting efficiency, different dexterous workspaces of IRB 5500 corresponding to different postures have been analyzed carefully. Findings: The experiments confirm the effectiveness of the proposed surface segmentation method for painting different types of aircraft by IRB 5500. For semantic segmentation on aircraft data with point normal, PPN has higher precision than PointNet. In addition, the whole algorithm can efficiently segment one complex aircraft into qualified blocks, each of which has its own identity information, can be painted by IRB 5500 in one station and has fewer edges with other blocks. Research limitations/implications: As the provided experiments indicate, the proposed method can segment one aircraft into qualified blocks automatically, which highly improves the efficiency in aircraft painting compared with traditional approaches. Moreover, the proposed method is able to provide identity information of each block, which is necessary for application of different paint parameters and different paint materials. In addition, final segmentation results by the proposed method behaves better than k-means cluster on variance of normal vector distance. Originality/value: Inspired by semantic segmentation of 3 D point cloud, some improvements based on PointNet have been proposed for better handling segmentation of 6 D point cloud. By introducing normal vectors, semantic segmentation could be accomplished precisely for close points with opposite normal, such as wing upper and lower surfaces. Combining deep learning skills with traditional methods, the proposed method is proved to behave much better for surface segmentation task in aircraft painting.