Scalable and Generalizable AI Solutions - Applied to Vision-Based Quality Monitoring for Directed Energy Deposition

Abstract

This dissertation focuses on enhancing the scalability and generalizability of machine learning (ML) methods for predicting process outcomes in industrial automation, with a particular emphasis on additive manufacturing (AM) via directed energy deposition (DED). The research introduces several novel frameworks to address key challenges in this domain, including data annotation, domain generalization, and real-time adaptability.

First, a novel active learning method is proposed to reduce the need for extensive data annotation. This method optimizes model performance by balancing uncertainty with random sampling, allowing for a transition from exploration to exploitation throughout the learning process. It achieves a notable reduction in annotation requirements, and is designed to be broadly applicable to any ML task involving prediction uncertainty.

Second, the dissertation presents a new validation framework to address domain generalization issues in AM. Traditional ML models often rely on limited and non-diversified datasets, which can lead to an overestimation of performance under real-world conditions. This research introduces a diverse DED dataset designed to simulate various environmental changes typically encountered in real-world scenarios. The framework employs a categorization and validation simulation protocol to assess model performance across different environments. Evaluation of various ML architectures on this dataset provides insights into cross-environment performance and inference latency.

Finally, a training-free domain adaptation framework is developed to tackle real-time process measurement challenges in AM. This framework adapts to diverse operational contexts with minimal annotations through human-in-the-loop interactive annotation guided by the architecture itself. It demonstrates exceptional real-time processing capabilities, and segmentation performance of melt-pool related signatures.

Together, these contributions advance the industrial adoption of AM and enhance its competitive edge over conventional manufacturing processes by improving the scalability, generalizability, and real-time adaptability of ML methods.