Enhancing defect detection in additive manufacturing using a conditional autoencoder with skip connections and in situ infrared sensing

Developing reliable and efficient defect detection strategies for additive manufacturing (AM) is critical to advancing its adoption in industrial environments. To this end, studies have focused on constructing defect detection frameworks combining unsupervised anomaly detection algorithms and sensor...

Celý popis

Uloženo v:

Podrobná bibliografie
Vydáno v:	Journal of manufacturing processes Ročník 156; s. 268 - 283
Hlavní autoři:	Segura Ibarra, Diego, Hossain, Rifat-E-Nur, Lamonte, Riley C., Moore, Arden L., Chen, Jinyuan
Médium:	Journal Article
Jazyk:	angličtina
Vydáno:	Elsevier Ltd 26.12.2025
Témata:	Anomaly detection Deep learning Fused deposition modeling Melt-pool temperature sensing Process monitoring Self-supervised learning Process monitoring Deep learning Melt-pool temperature sensing Fused deposition modeling Anomaly detection Self-supervised learning
ISSN:	1526-6125
On-line přístup:	Získat plný text
Tagy:	Přidat tag Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!

Popis
Shrnutí:	Developing reliable and efficient defect detection strategies for additive manufacturing (AM) is critical to advancing its adoption in industrial environments. To this end, studies have focused on constructing defect detection frameworks combining unsupervised anomaly detection algorithms and sensor-based process monitoring systems. However, some of these approaches rely on traditional neural network architectures with inherent limitations that hinder their anomaly detection performance, while others implement algorithms that are not well-suited for estimating defect severity. To overcome these challenges, this work presents a defect detection framework that employs a novel reconstruction-based time series anomaly detection algorithm, which enhances sensitivity and precision through a conditional autoencoder with skip connections and a specialized self-supervised learning scheme. This algorithm produces an anomaly score that is proportional to the intensity of an anomaly, enabling the estimation of defect severity from this score. In the defect detection framework, the deep learning scheme is used to detect anomalous infrared (IR) sensor signals, which monitor the melt-pool region temperature in real time. Compared to IR camera-based systems that demand significant computational resources and acoustic emission sensors that require extensive denoising, the IR sensor offered a more efficient and practical solution for in situ real-time process monitoring. Experiments conducted on a commercial fused deposition modeling (FDM) machine demonstrated the framework’s ability to reliably detect void defects of varying sizes in both a simple L-shaped test geometry and a more complex industry-standard test article. Compared to conventional autoencoder-based strategies, the proposed algorithm improves the F1 score by 34% on average, highlighting the significance of the more advanced approach. Although validated on FDM, the framework’s adaptable design enables its use across diverse AM processes, defects, and sensor types.
ISSN:	1526-6125
DOI:	10.1016/j.jmapro.2025.10.099