An Interpretable Machine Learning based Predictive Control Method for Manufacturing Processes with Turning as A Case Study

Abstract

Modern digital manufacturing machines can record massive process signals through embedded internal controllers and instrumented external sensors. This provides a good base for the development of machine learning-based control methods for modeling and control of manufacturing processes. This study develops an interpretable machine learning based predictive control method, and that can learn effective ordinary differential equations as process models to achieve reliable and robust control performance. Verification and validation results on a synthetic example of Lorenz 3 system with control show the proposed method can achieve superior performance on model discovery from highly noisy data, and both set point and trajectory controls. In addition, a case study on the turning process using experimentally validated time domain simulator is conducted. Compared with existing methods, the proposed method shows a two order-of-magnitude anti-noise capability improvement for model discovery, and that enables stable process parameter selection with maximum material removal rate. The proposed method has the potential to be used for process model discovery, and model-based monitoring and control for machining and other manufacturing processes.