Machine Learning-Based Shape Error Estimation Using the Servomotor Current Generated During Micro-Milling of a Micro-Lens Mold

Abstract

The demand for the mass production of micro-lens arrays (MLAs) is increasing. An MLA is fabricated through an injection molding process, and its mold is manufactured by a five-axis high-precision machine tool using a small diameter endmill. A visual exam-ination is not available to judge the quality of the mold while machining. Therefore, an effective process monitoring technology must be developed. A promising approach is to apply a servomotor current to in-process monitoring because as long as the servomotor works well, no external sensors, capital investment, or maintenance processes are required. From this per-spective, a machine learning-based shape error estimation method using only the servomotor current is proposed. To explore the relationship between the motor current generated during micro-milling and the shape error of the mold, the servomotor current in X-, Y-, and Z-axes was recorded, and the corresponding shape error of the MLA mold was measured after ma-chining. Input data were prepared by converting time-domain servomotor current data to frequency-domain data using short-time Fourier transform and reduc-ing the dimensions of the data via principal compo-nent analysis. In terms of a meaningful label for the output data, the average shape error in the machined area corresponding to each window was provided. The input/output relationships were used to train five dif-ferent machine learning models, and the accuracy of shape error estimation using each model was evalu-ated. In addition, the estimation accuracies using the X-, Y-, and Z-axes were compared to find the axis that senses the shape error with the highest accuracy. The results show that the non-linear method using the X-axis servomotor current information closest to the machining point achieved the highest shape error estimation accuracy.