Researchers develop collaborative deep learning method for defect detection in additive manufacturing


Taylor Tucker

Chenhui Shao
Associate Professor Chenhui Shao

MechSE doctoral student Manan Mehta and Associate Professor Chenhui Shao recently developed a federated learning-based methodology for pixel-wise defect detection in additive manufacturing (AM). Their work was published in the Journal of Manufacturing Systems and their code is available online.

When performed by a computer, defect detection is typically executed pixel by pixel for each image. Detection algorithms are developed using deep learning. 

“Defect detection is an important area in today’s AM research, as it enables manufacturers to monitor and improve the quality and mechanical properties of AM components and products,” Mehta said. However, obtaining high-quality data to train computer vision algorithms is resource expensive, which can create a bottleneck in AM.

Mehta and Shao sought to address this need by investigating federated learning, which allows multiple manufacturers to collaboratively train a machine learning model under a central server without exposing their own data to one another. “The data never leaves the manufacturer’s location, thus providing a notion of privacy,” Mehta said.

Federated learning model.
Federated learning allows the use of data from multiple sources without compromising privacy.

The team used images of 3D-printed parts as training data for classifying each pixel as powder, part, or defect in a process called semantic segmentation. To simulate a real-world federated learning scenario, they divided the images among eight different sources. Findings indicated that algorithmic performance is significantly improved by federated learning, meaning that manufacturers can use this paradigm to effectively pool their resources without compromising proprietary data.

“This work is one of the first to study federated learning on real-world manufacturing data and opens up exciting opportunities for future work,” Mehta said. “We hope that this work will advance the use of deep learning in data-scarce production environments in AM as well as manufacturing in general.”

Future directions for this research include the development of low-cost, generalizable decision-making methods for manufacturing applications. “Our group is exploring a new interdisciplinary area called ‘cost-effective machine learning for smart manufacturing,’” said Shao, explaining that the field is comprised of federated learning, transfer learning, active learning, data fusion, and physics-formed hierarchical modeling. “We envision that these decision-making approaches will collectively promote the agility, intelligence, and automation of modern manufacturing factories.”