This study aims to detect seeded porosity during metal additive manufacturing by employing convolutional neural networks (CNN). The study demonstrates the application of machine learning (ML) in in-process monitoring. Laser powder bed fusion (LPBF) is a selective laser melting technique used to build complex 3D parts. The current monitoring system in LPBF is inadequate to produce safety-critical parts due to the lack of automated processing of collected data. To assess the efficacy of applying ML to defect detection in LPBF by in-process images, a range of synthetic defects have been designed into cylindrical artefacts to mimic porosity occurring in different locations, shapes, and sizes. Empirical analysis has revealed the importance of accurate labelling strategies required for data-driven solutions. We formulated two labelling strategies based on the computer-aided design (CAD) file and X-ray computed tomography (XCT) scan data. A novel CNN was trained from scratch and optimised by selecting the best values of an extensive range of hyper-parameters by employing a Hyperband tuner. The model’s accuracy was 90% when trained using CAD-assisted labelling and 97% when using XCT-assisted labelling. The model successfully spotted pores as small as 0.2mm. Experiments revealed that balancing the data set improved the model’s precision from 89% to 97% and recall from 85% to 97% compared to training on an imbalanced data set. We firmly believe that the proposed model would significantly reduce post-processing costs and provide a better base model network for transfer learning of future ML models aimed at LPBF micro-defects detection.
Surface deformation is a multi-factor, laser powder-bed fusion (LPBF) defect that cannot be avoided entirely using current monitoring systems. Distortion and warping, if left unchecked, can compromise the mechanical and physical properties resulting in a build with an undesired geometry. Increasing dwell time, pre-heating the substrate, and selecting appropriate values for the printing parameters are common ways to combat surface deformation. However, the absence of real-time detection and correction of surface deformation is a crucial LPBF problem. In this work, we propose a novel approach to identifying surface deformation problems from powder-bed images in real time by employing a convolutional neural network-based solution. Identifying surface deformation from powder-bed images is a significant step toward real-time monitoring of LPBF. Thirteen bars, with overhangs, were printed to simulate surface deformation defects naturally. The carefully chosen geometric design overcomes problems relating to unlabelled data by providing both normal and defective examples for the model to train. To improve the quality and robustness of the model, we employed several deep learning techniques such as data augmentation and various model evaluation criteria. Our model is 99% accurate in identifying the surface distortion from powder-bed images.
The advent of Additive manufacturing (AM) of 3D printed objects is revolutionising the manufacturing industry. Despite its promise, the extensive post-processing requirements of 3D objects remains a significant barrier to AM’s wider adoption. Identifying defects in real-time from powder bed images taken before and after the laser melting of the current layer presents a promising strategy to reduce post-processing efforts. Traditional methods focusing on the top-layer images fall short in identifying multi-layer defects such as keyhole porosity, balling, and lack of fusion, which are critical to the integrity of 3D printed objects. Addressing this challenge, our study introduces an innovative multi-layer technique for the detection of keyhole porosity using high-quality X-ray Computed Tomography (XCT) images, leveraging the capabilities of the cutting-edge YOLO (You Only Look Once) object detection algorithm. Our findings reveal that this approach achieves a remarkable mean average precision (mAP) score of 92.585%, underscoring the efficacy of deep learning models in accurately identifying keyhole porosity across XCT images. This research not only demonstrates the potential for improving the quality and reliability of AM processes but also paves the way for reducing the dependency on labour-intensive post-processing steps.
6 March 2023 in American Journal of Cardiovascular Drugs
This study focuses on the detection of seeded porosity during metal additive manufacturing by employing convolutional neural networks (CNN). The aim of the study is to demonstrate the application of Machine Learning (ML) in in-process monitoring. Laser Powder Bed Fusion (LPBF) is a selective laser melting technique used to build complex 3D parts. The current monitoring system in LPBF is inadequate to produce safety-critical parts due to the lack of automated processing of collected data. To assess the efficacy of applying ML to defect detection in LPBF by in-process images, a range of synthetic defects have been designed into cylindrical artefacts to mimic porosity occurring in different locations, shapes, and sizes. Empirical analysis has revealed insights into the importance of accurate labelling strategies required for data-driven solutions. Two labelling strategies based on the computer aided design (CAD) file and X-ray computed tomography (XCT) scan data was formulated. A novel CNN was trained from scratch and optimised by selecting the best values of an extensive range of hyper-parameters by employing Hyperband tuner. The accuracy of the model was 90% when trained using a CAD-assisted labelling, and 97% when using XCT-assisted labelling. The model successfully spotted pores as small as 0.2mm. Experiments revealed that balancing the data set improved the model's precision from 89% to 97% and recall from 85% to 97% when compared to training on an imbalanced data set. We strongly believed that the proposed model would significantly reduce post-processing cost and provide a better base model network for transfer learning of future ML models aimed at LPBF micro-defects detection.
Deep learning is transforming breast cancer screening by automating mammogram analysis. However, the rapid evolution of model architectures necessitates a critical evaluation of the field's trajectory and its readiness for clinical translation. This systematic literature review, conducted in accordance with PRISMA guidelines, critically analyses a core set of 37 peer-reviewed studies published between 2022 and 2024, a period marked by the emergence of Vision Transformers and large-scale self-supervised learning. In total, our synthesis draws on 110 references to provide a comprehensive analysis of this evolving field. While Convolutional Neural Networks (CNNs) remain the dominant architecture, our analysis reveals a significant shift towards more complex, data-intensive models. This review synthesises findings across the deep learning pipeline, from data acquisition to model evaluation, to identify systemic challenges. We argue that the field faces two critical, intertwined challenges: a significant reproducibility issue stemming from a pervasive reliance on large, inaccessible private datasets that impedes independent validation of state-of-the-art claims, and a critical gap in algorithmic fairness. This gap is evidenced by a profound lack of performance reporting across key patient demographics (such as ethnicity and breast density), which poses a significant risk to health equity. Promising innovations in self-supervised learning and domain-specific pre-training offer pathways to mitigate data scarcity and improve model robustness. However, translating algorithmic potential into tangible clinical benefit will require a concerted effort to foster open science, implement rigorous evaluation of generalisability and fairness, and develop models that are not only accurate but also equitable and trustworthy.