Model Uncertainty Estimation for Thermal Error Compensation in Machine Tools

Abstract

This paper introduces a method to compensate for thermal errors in machine tools (MT) using LSTM neural networks, with a focus on addressing prediction uncertainties. It presents the application of Monte Carlo Dropout (MC-Dropout) to estimate the uncertainty of LSTM predictions using data generated in a simulated MT environment. MC-Dropout offers a practical, computationally efficient method to allow for effective thermal error compensation without repeated on-machine measurements. Incorporating uncertainty estimates can enhance decision-making, thus allowing more autonomous machine operations and improve the selection of training data for machine learning models, leading to greater overall prediction accuracy.