Sensor Signal Processing Using Machine Learning for Reliable Surface Quality in Direct Laser Interference Patterning

Vorschau

27 Min.

Laser surface microtexturing processes, such as Direct Laser Interference Patterning (DLIP), are governed by complex physical effects including thermal accumulation and plasma generation, which directly affect surface quality [1,2]. Real-time monitoring with off-axis photodiode sensors enables in situ observation of these interactions by capturing backscattered emission signals during laser irradiation [3]. In this study, a picosecond DLIP system (ELYPsis®, SurFunction GmbH) was used to fabricate periodic surface patterns on additively manufactured Ti-6Al-4V substrates, while plume emissions were recorded using a broadband photodiode sensor (Plasmo GmbH). Surface quality was assessed through spectral entropy analysis of White Light Interferometry (WLI) images, extracting spatial frequency disorder as a scalar metric and applying K-means clustering to classify structures as “OK” or “NOK.” These unsupervised labels were assigned to time-resolved photodiode signals—cropped to the first 6 milliseconds—and paired with corresponding laser parameters (fluence, overlap, repetition rate). Rather than relying on time–frequency transformations, raw photodiode signals were directly processed by a dual-input 1D Convolutional Neural Network (1D-CNN), which performed early fusion with the scalar laser parameters to learn joint signal–context representations [4]. The trained model achieved a classification accuracy of 86%, demonstrating its ability to detect structural deviations using minimal signal preprocessing. These results confirm the viability of entropy-guided, sensor-driven deep learning for real-time quality monitoring in DLIP and provide a scalable foundation for closed-loop control in high-throughput laser microstructuring applications.

Abstract

Erwerben Sie einen Zugang, um dieses Dokument anzusehen.

Dr. Marcelo Sallese