Real-Time Defect Monitoring of Laser Micro-drilling Using Reflective Light and Machine Learning Models

Abstract

Laser micro-drilling is a significant manufacturing method used to drill precise microscopic holes into metals. Quality inspection of micro-holes is costly and redrilling defective holes can lead to imperfection owing to the misalignment in re-aligning the removed specimens. Thus this paper proposes an in-situ, automatic inspection method using photodiode data and machine learning models to detect defects in real-time during the fabrication of SK5 steel plates with 1064 nm Nd:YAG Laser machines to reduce the workload and increase the quality of products. Further, it explores the possibility of generalizing the models to 51 different scenarios of fabrication by classifying unseen data into 51 classes. A dataset of around 1,500,000 time series data points was generated using an optical probe while drilling over 56,000 holes into test specimens. 15 different combinations of thickness and diameter were drilled using suggested parameters. An additional 12 potential defect-prone conditions were designed to obtain data during conditional drilling. Hole quality was measured for each hole using OGP 3D profile microscope measuring machine. Results showed high accuracy in specialized defect detection within each scenario and showed a possibility of classifying photodiode data patterns, offering opportunities to improve the practicality of the proposed solution.

An Experiment on Anomaly Detection for Fault Vibration Signals Using Autoencoder-Based N-Segmentation Algorithm

“…In the existing AD using AE, the maximum reconstruction error value (Kang, Kim, Kang, & Gwak, 2021;Wei, Jang-Jaccard, Xu, Sabrina, Camtepe, & Boulic, 2023) or the 3sigma value (Lee, Lee & Kim, 2024;Panza, Pota, & Esposito, 2023) among the reconstruction errors of the training data has been set as a threshold. In this case, the threshold has to be only one.…”

“…Applying supervised learning to AD requires data for all types of anomalies. Since gathering anomaly data for enough training is practically impossible, unsupervised learning using only the normal-condition data of the facility is more suitable, and the Autoencoder (AE) is representative of unsupervised learning (Lee, Lee, & Kim, 2024;Wei, Jang-Jaccard, Xu, Sabrina, Camtepe, & Boulic, 2023). AE is an AI model that learns how to produce output data as close as possible to the input data without data labels.…”