A novel method based on a convolutional graph neural network for manufacturing cost estimation. (October 2022)

• Record Type: Journal Article
• Authors: Zhang, Hang; Wang, Wenhu; Zhang, Shusheng; Huang, Bo; Zhang, Yajun; Wang, Mingwei; Liang, Jiachen; Wang, Zhen
• Is Part Of: Journal of manufacturing systems. Volume 65(2022)
• Page Start: 837
• Page End: 852
• Publication Date: 2022-10
• Subjects: Deep learning -- Convolutional graph neural network -- Manufacturing cost estimation -- Graph data; 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.2022.10.007
• Languages: English
• ISSNs: 0278-6125

Abstract

Abstract: With the widespread application of mass customization strategy, estimating the manufacturing cost of products to provide suitable references for the quotations of products can assist enterprises to adapt to the competitive market. Moreover, estimating the manufacturing cost of products in the design stage can assist designers in optimizing product designs. With the continuous development of deep learning, studies on part manufacturing cost estimation based on deep learning have started in recent years. However, the existing deep learning-based methods ignore the part's precision information (e.g. roughness, tolerance) that is important to the manufacturing cost, making them less practical. To this end, how to take the precision information of the parts into account in the manufacturing cost estimation to make deep learning-based methods more applicable is a difficulty that needs to be solved. In this context, an innovative convolutional graph neural network (ConvGNN)-based manufacturing cost estimation approach that considers precision information is suggested. Specifically, an attribute graph that is based on machining features and includes precision information is defined to represent the 3D computer-aided design (CAD) model of a part first. Then, a novel ConvGNN framework named Cost Estimation Network (CEN) is constructed that combines spectral-based convolutional layers and spatial-based convolutional layers. The input of CEN is the attribute graph of a part, while the output is the manufacturing cost. After being trained with historical data and an improved loss function based on Root Mean Squared Log Error (RMSLE), CEN can be applied to estimate manufacturing costs. In addition, since the deep learning model is treated as a black box, a modified gradient-weight class activation mapping (Grad-CAM) process is developed to explain the rationale of the manufacturing cost decision and differentiate the degree of influence of various machining features on the cost of parts. In experimental studies, the computer numerical control (CNC) machined rotary parts are used as examples to verify the feasibility and effectiveness of the presented approach. Highlights: Propose a novel deep learning-based part manufacturing cost estimation method that considers the precision information (roughness, tolerance). Propose a novel 3D CAD part representation method that can be learned by graph neural networks. Construct a novel convolutional graph neural network framework named CEN used for regression problems. Propose a modified Grad-CAM process used for explaining the rationale of the manufacturing cost decision. Promote the application of deep learning in the manufacturing industry.