Thermal Image Classification Using Convolutional Neural Network (CNN) For Thermal Stress Prediction In Metal 3D Printing

Aini Zuhra Abdul Kadir; Yee Seng Hoo; Nurul Husna Mohd Yusoff; Shao Wei Koh; Abdul Hamid Ahmad; Mohd Azlan Suhaimi

Authors

Aini Zuhra Abdul Kadir Faculty of Mechanical Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor, Malaysia
Yee Seng Hoo Faculty of Mechanical Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor, Malaysia
Nurul Husna Mohd Yusoff Faculty of Mechanical Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor, Malaysia
Shao Wei Koh Faculty of Mechanical Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor, Malaysia
Abdul Hamid Ahmad Faculty of Mechanical Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor, Malaysia
Mohd Azlan Suhaimi Faculty of Mechanical Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor, Malaysia

Keywords:

Thermal Stress, Additive Manufacturing, CNN, MobileNetV2, DLMD

Abstract

In this paper, a neural network strategy consisting of Convolutional Neural Networks (CNNs) is introduced to classify thermal stress in metal additive manufacturing. Past thermal images of a Directed Laser Metal Deposition (DLMD) process were pre-processed and labelled depending on the determined values of thermal stress. Three CNNs, DenseNet201, MobileNetV2, and ResNet50, were tested in two scenarios: as feature extractors, combined with Support Vector Machine (SVM) classifier and end-to-end training with ADAM optimizer. The experimental performance indicated that MobileNetV2 fared the best at attaining the highest accuracy (96.36 %) since it has a less resource-hungry framework and a quicker convergence. The model also performed high generalization when validated on unseen data through both individual and batch validation. This paper shows how it is possible to combine real-time thermal inspection and a CNN to automatically determine defects in metal 3D printing.