Print defect detection during laser powder bed fusion by one-dimensional convolutional autoencoder

Abstract Metal Additive Manufacturing (AM), particularly Laser Powder Bed Fusion (L-PBF), is increasingly used in high-precision industries such as aerospace and medical device manufacturing due to its ability to produce detailed and complex parts. Despite these advantages, L-PBF faces challenges related to build faults, which can degrade part quality, increase material waste, and raise computational costs due to extensive post-process inspection. To address these issues, integrating in-situ process monitoring with machine learning enables early fault detection and reduces reliance on post-process analysis. However, most existing approaches rely on supervised learning, often using high-resolution image data that are computationally expensive to process or sensor-based signals that still require substantial preprocessing, making real-time deployment difficult in industrial AM environments. This study presents a One-Dimensional Convolutional Autoencoder (ODCAE) as an semi-supervised model for detecting laser power anomalies during the L-PBF process. The proposed model is trained on real-time photodiode signals collected using the Infini AM process monitoring system, which captures optical emissions generated during the build process. It learns the periodic patterns of these signals and detects anomalies by identifying deviations from nominal behaviour through reconstruction-error analysis. Since it operates directly on signals, the method avoids complex preprocessing steps and reduces computational cost. ODCAE demonstrates that high detection accuracy can be achieved alongside low-latency, energy-efficient inference, enabling a practical pathway toward real-time, sustainable in-situ monitoring in L-PBF. Consequently, by enabling earlier detection of abnormal process conditions, the proposed model has the potential to reduce failed builds, material waste, and associated production costs. The proposed model was evaluated on samples with engineered anomalies, including prints with different geometries and laser power settings, as well as data acquired from a second L-PBF machine, covering multiple datasets and experimental conditions. The results demonstrate that the proposed method consistently outperformed the baseline deep autoencoder across all evaluated experiments.