A framework for advancing machine vision in wire arc directed energy deposition manufacturing system

Deep learning-driven machine vision technologies are revolutionizing numerous industries, including metal additive manufacturing (AM). Wire arc directed energy deposition (WA-DED), a key branch of metal AM, has increasingly adopted these technologies to enhance process monitoring and defect detection. However, the development of accurate and reliable deep learning models for WA-DED machine vision applications is challenged by the extensive manual effort required for labelling large image datasets, which becomes even more labour-intensive when changes in deposition processes or metal feedstock necessitate new model training. To address this challenge and promote machine vision application in metal AM industries, this study proposes a novel framework that integrates a semi-automatic labelling method based on unsupervised learning (UL) techniques to improve model training efficiency while maintaining high accuracy. The framework comprises a semi-automatic labelling module, which leverages t-SNE for efficient dataset annotation, and a classifier construction module, which utilizes a ResNet-based architecture for image classification. The effectiveness of the proposed framework is validated through classification accuracy and efficiency assessments, achieving 94.65 % accuracy across nine molten pool image categories, along with a 94.3 % improvement in data labelling efficiency and 30.5 % improvement in the entire classifier construction efficiency. Additionally, the classifier is further evaluated through defect detection in a real-world WA-DED road barrier part. The detection results demonstrate that the types and locations of multiple observable defects, as well as a variety of other subtle defects, have been accurately identified. Overall, these outcomes confirm the effectiveness of the proposed framework in promoting machine vision applications in WA-DED manufacturing systems.