Research on Improving Estimation of EDM Precision with Automatic Feature Selection

• Main Authors: YUEH, YEN-CHEN, 岳晏晨
• Other Authors: YANG, HAO-CING
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/6vkr58

碩士 === 國立高雄第一科技大學 === 電機工程研究所碩士班 === 106 === In EDM (Electrical discharge machining), variation of workpiece quality is mostly affected by the relations of voltage and the current while processing. A systsem develoed by previous research is able to collect signal, segment data, extract feature, fit distribution, select key feature, and estimate quality. Due to high correlation of variation of electrode size and workpiece tolerance, it is a challenge how to effectively evaluate electrode size variation and automatically select the key feature set for improving workpiece quality estimation. Based on previous works, this research develops two modules which are the EPS (electrode profile estimation) and AFS (automatic feature selection) modules. The EPS module contains the tool image capturing and tool wear estimation submodules, in which the former can derive tool wear sizes from the off-line captured tool images and the later is capable of on-line estimating tool wear using a regression model with EDM featrures. Based on Glmnet (Lasso and Elastic-Net Regularized Generalized Linear Models) algorithm, the AFS module is used to select key features. Integrating the EPS and AFS modules, the system is improved to estimate workpiece quality. In results, compared with the previous work for estimating accuries of bottom circle of machining 10 mm hole, the mean absolute error of roundness estimation is decreased from 5.09 μm to 4.02 μm, while the 95% error of diameter estimation is reduced from 32.44 μm to 27.89 μm and the max error of Ra estimation is improved from 0.55 μm to 0.5 μm. With module extension and model refresh, the system is promising for estimating workpiece quality while applying in different machines. Keywords: Tool Dimensional Estimation, Feature Selection, EDM, Glmnet.