Comparative Evaluation of Task-specific Machine Learning and Foundation Models for Melt Pool Anomaly Detection in Laser Powder Bed Fusion

Abstract

Foundation models enable low-cost development and operator-friendly deployment of online quality inspections in manufacturing. However, whether foundation models pre-trained can effectively adapted for specialized industrial process monitoring remains insufficiently understood, particularly regarding fault detection effectiveness, adaptation efforts, and computational trade-offs compared to task-specific machine learning models. This paper presents a systematic comparative benchmark study using a reference framework that models learning and adaptation processes. Both task-specific machine learning and foundation model-based approaches are aligned along common spectrums of data availability and adaptation mechanisms and evaluated with uniform metrics. Through a case study of melt pool monitoring in laser powder bed fusion additive manufacturing, we compare six learning strategies spanning unsupervised, semi-supervised, supervised, zero-shot, few-shot, and fine-tuned foundation model-based approaches. The analysis compares performance trade-offs based on data availability for adaptation, highlighting when adapted foundation models become competitive against task-specific learnings. In addition, this work evaluates industrial anomaly detection approaches for computational constraints and resource requirements. The systematic methodology with an example can guide the industry to select the most appropriated process anomaly detection methods for their applications, which balances accuracy, data efficiency, and computational feasibility in resource-constrained manufacturing environments.