A supervised community detection method for automatic machining region construction in structural parts NC machining. (January 2022)

• Record Type: Journal Article
• Authors: Liu, Xu; Li, Yingguang; Deng, Tianchi; Wang, Pengcheng; Lu, Kai; Chen, Jiarui; Yang, Dingye
• Is Part Of: Journal of manufacturing systems. Volume 62(2022)
• Page Start: 367
• Page End: 376
• Publication Date: 2022-01
• Subjects: NC machining -- Structural part -- Machining region construction -- Supervised community detection -- Graph neural network; Manufacturing processes -- Periodicals; Production engineering -- Data processing -- Periodicals; Robots, Industrial -- Periodicals; Production, Technique de la -- Informatique -- Périodiques; Robots industriels -- Périodiques; Electronic journals; 670.42
• DOI: 10.1016/j.jmsy.2021.12.005
• Languages: English
• ISSNs: 0278-6125

Abstract

Highlights: A data-driven method for automatic machining region construction in structural part NC machining is proposed. The task is converted to detecting machining region communities (MRC) from the attributed graph of part model. A new concept named affiliation similarity (AS) is defined, with which a new MRC detection algorithm is established. Graph neural networks for AS prediction are designed and trained to support supervised MRC detection. Abstract: In structural part NC machining, part faces need to be grouped to form various machining regions to support machining sequencing, parameter planning and toolpath generation. However, existing automatic machining region construction methods are usually only acceptable for parts in specific fields due to the heavy dependence on domain-specific expert defined rules or strategies. This research proposes a data-driven method that learns the knowledge required by machining region construction directly from the historical data. The task is first converted to machining region community (MRC) detection from the attributed graph of the 3D part model. As existing supervised community detection methods in network science need to know the total MRC number in advance which is unpractical in this task, a new concept named affiliation similarity (AS) is defined to describe the overlap of the MRC affiliations of multiple nodes, then with which a new MRC detection algorithm that does not require any information of the MRC number is established. The maps from the face nodes to the AS information is modeled using graph neural networks (GNN) which are trained using the data samples constructed based on the historical process files. A case study using the data of aircraft structural part NC machining from real industry is carried out and the result shows the proposed method is practical in automatic machining region construction. More importantly, it is believed that the proposed method is convenient to be extended to parts of other domains by modifying the node attributes, redesigning the GNN structures and retraining the GNN models with corresponding dataset.