Self-Supervised Deep Learning and Statistical Feature-Based Approaches for Melt Pool Anomaly Detection

In this work, we explore the efficacy of two distinct approaches for classifying melt pool images generated during laser powder bed fusion (LPBF), an additive manufacturing process. Our first approach utilizes a self-supervised deep learning model based on Bootstrap Your Own Latent (BYOL). The model is initially pre-trained on an extensive set of unlabeled melt pool images, capturing the intrinsic features of the images. Subsequently, it is fine-tuned on a smaller labeled dataset to improve its classification accuracy on melt pool anomalies. In parallel, we define a set of statistical features to quantitatively describe each image and employ these features in a conventional supervised classification model. The aim is to identify the optimal approach between self-supervised learning and statistical feature-based classification for detecting anomalies in LPBF melt pools. Following comparative evaluation, the more effective model will be leveraged in further analyses to deepen our understanding of melt pool dynamics and improve the LPBF process. This research contributes to advancements in LPBF quality assurance by addressing the challenge of anomaly detection in complex manufacturing environments with minimal labeled data.