Transfer learning for predictive quality in laser-induced plasma micro-machining

Abstract

In laser-induced plasma micro-machining (LIPMM), a focused, ultrashort pulsed laser beam creates a highly localized plasma zone within a transparent liquid dielectric. When the beam intensity is greater than the breakdown threshold in the dielectric media, plasma is formed which is then used to ablate the workpiece. This paper aims to facilitate in-situ process monitoring and quality prediction for LIPMM by developing a deep learning model to (1) understand the relationship between acoustic emission data and quality of micro-machining with LIPMM, (2) transfer such understanding across different process parameters, and (3) predict quality accurately by fine-tuning models with a smaller dataset. Experiments and results show that the relationship learned from one process parameter can be transferred to other parameters, requiring lesser data and lesser computational time for training the model. We investigate the feasibility of transfer learning and compare the performance of various transfer learning models: different input features, different CNN structures, and the same structure with different fine-tuned layers. The findings provide insights into how to design effective transfer learning models for manufacturing applications.