Real-Time Defect Detection in Additive Manufacturing Using Vision-Based Machine Learning

Tuesday, October 21, 2025: 8:00 AM
335 (Huntington Place)
Dr. Navin Manjooran, Ph.D., MBA, CEng. FASM, FACerS, FIIM, FIE, FIMMM, FIIE, FAEM, HoF-VTAEE , Solve, Windermere, FL
Shahjahan Hossain , University of Central Florida, Orlando, FL
Pranta Sarkar , University of Central Florida, Orlando, FL
Ranajay Ghosh , University of Central Florida, Orlando, FL
Dr. Ramesh Subramanian , Siemens Energy Inc., Orlando, FL
Dr. Gary R. Pickrell , Virginia Polytechnic Institute and State University, Blacksburg, VA
Ensuring consistent print quality in additive manufacturing (AM) requires robust, real-time monitoring systems capable of detecting process anomalies as they occur. This work presents a machine learning–driven framework that utilizes computer vision techniques to monitor AM processes and identify defects dynamically. The approach leverages simulated AM datasets to train convolutional neural networks (CNNs) for accurate defect classification based on real-time visual input.

The system captures video streams of the printing process, extracting key visual features such as layer alignment, filament deposition uniformity, and surface texture. These features are analyzed frame-by-frame using a trained CNN to determine the presence or absence of defects. To develop a reliable training dataset, a variety of simulated defects—such as layer shifting, over-extrusion, under-extrusion, and inconsistent filament flow—are introduced. The dataset is further diversified with variations in lighting and camera angles to ensure robustness across a range of real-world printing conditions.

Once deployed, the system continuously evaluates visual features during printing. Upon detecting a potential defect, it can trigger real-time alerts, recommend parameter adjustments, or pause the process for intervention. This capability enables proactive quality control, minimizes print failures, and reduces material waste.

By combining machine learning with real-time visual monitoring, this approach offers a scalable, non-invasive solution for improving reliability in AM processes. It supports the broader goal of autonomous, self-correcting manufacturing systems, where quality assurance is embedded directly into the production workflow.