메뉴 드롭다운
메뉴 드롭다운
KR EN

대학원 소개

연구성과

UNIST 인공지능대학원의 대학원 및 연구성과를 확인하실 수 있습니다.

Bayesian-based uncertainty-aware tool-wear prediction model in end-milling process of titanium alloy (Appl. Soft. Comput.), Prof. Sunghoon Lim

Bayesian-based uncertainty-aware tool-wear prediction model in end-milling process of titanium alloy

Author: Gyeongho Kim  , Sang Min Yang  , Dong Min Kim  , Sinwon Kim  , Jae Gyeong Choi  , Minjoo Ku  , Sunghoon Lim , Hyung Wook Park


Abstract

Tool wear negatively affects machined surfaces and causes surface cracking, therefore increasing manufacturing costs and degrading product quality. Titanium alloys, which are widely used because of their desirable mechanical properties, have problems associated with tool wear due to poor thermal properties, such as specific heat capacity and thermal conductivity. Therefore, the accurate prediction of tool wear is necessary during the titanium alloy end-milling process to improve product quality and ensure reliability for corrective decisions like tool replacement. To this end, uncertainty-aware tool-wear prediction should be performed. In this study, a deep learning-based tool-wear prediction model based on a Bayesian approach was proposed. First, a convolutional neural network (CNN)-based architecture that integrates multiscale information extracted from raw sensor measurement data, termed deep multiscale CNN (DMSCNN), was proposed. It used different-sized convolutional kernels in parallel to enable various receptive field sizes suitable for machining processes. Second, based on a Bayesian learning approach, DMSCNN was transformed into a probabilistic model that produced a predictive distribution for estimated tool wear. In particular, a variational inference was applied to DMSCNN parameters to provide uncertainty awareness. Experiments were conducted with data collected from an actual end-milling process under three different conditions. The results proved the effectiveness of the proposed DMSCNN for tool-wear prediction. Bayesian DMSCNN showed promising results, as it outperformed existing comparative deterministic methods as well as probabilistic methods for tool-wear prediction. The proposed method is expected to be effectively applied in smart manufacturing as well as other machining processes that require data-driven digital decisions.