Research Symposium

BIO

Gargi is a second-year student at Florida State University pursuing a Bachelor of Science in Computer Science with a collateral minor in Mathematics. She hopes to pursue a Master's degree in Computer Science and continue exploring the intersection of artificial intelligence and engineering. Outside of academics, Gargi enjoys playing cricket and badminton and spending time with her family.

Leveraging Deep Learning Models for Thermal Anomaly Detection in Robot-Assisted Manufacturing

Abstract

Detecting abnormal behavior is critical for robot-assisted manufacturing systems, yet practical deployment is challenged by limited availability of labeled defect data and evolving failure modes. This project investigates how supervised and unsupervised deep learning approaches compare in detecting abnormal manufacturing behavior from a thermal image dataset of an additive manufacturing process. Model performance was evaluated using standard classification metrics, including accuracy and AUC. Supervised models, including a custom Convolutional Neural Network (CNN) and EfficientNetV2, outperformed unsupervised methods when labeled defect data was available. The custom CNN achieved the strongest performance (97% accuracy) compared to EfficientNetV2 (42% accuracy). Unsupervised models included a Convolutional Autoencoder (CAE) and a hybrid CAE + Isolation Forest model, where both the models achieved moderate detection performance (64% and 67% accuracy respectively). This comparative analysis highlights the tradeoff between detection accuracy and labeled data dependence and aims to inform the design of robust anomaly detection frameworks for robot-assisted additive manufacturing.

Keywords: Anomaly Detection, Additive Manufacturing, Deep Learning, Robot-Assisted Manufacturing