Session: 03-01-07: Annual Conference-Wide Symposium on Additive Manufacturing

Paper Number: 147155

147155 - Machine Learning-Enabled Process Monitoring and Error Detection in Material Extrusion-Based Additive Manufacturing 

Additive manufacturing (AM), also known as 3D printing, has revolutionized the landscape of advanced manufacturing. Among various AM techniques, material extrusion (MEX) stands out for its exceptional design freedom, unmatched potential for customization, and significantly reduced product development times. Thermoplastic materials are deposited layer-by-layer in a meticulously controlled manner, enabling the creation of intricate geometries and functional prototypes with remarkable ease. This process simplicity, coupled with the vast array of available materials, has cemented MEX's position as the most widely adopted AM technique. However, the success of MEX hinges on a critical factor: the precise extrusion and subsequent deposition of material onto the printing surface. Inconsistencies in this process can lead to a multitude of defects, compromising the final product's structural strength, aesthetics, and functionality. Unfortunately, current limitations exist in real-time monitoring and error detection within MEX processes. Traditional methods often rely on manual inspections, which are time-consuming, prone to human error, and incapable of immediate corrective actions. This research addresses this critical gap by proposing a novel approach that leverages the power of machine learning and cost-effective vision systems. A customized camera module was installed on the MEX extruder, providing a clear view of the localized material deposition area. This strategic placement allowed the camera to capture high-resolution images under various process parameters and layer heights. Flow rate and z-height platform offset were first identified as the primary factors influencing MEX's material extrusion quality. The total of 299,411 captured images, from different combinations of the flow rates and z-height offsets, formed the foundation for training an advanced convolutional neural network (CNN) algorithm with multiclass classification capabilities. By meticulously analyzing the vast dataset of captured images, the CNN algorithm learned to identify subtle variations and patterns associated with different material deposition scenarios. This empowered the algorithm to predict multiple process parameter values simultaneously. The trained CNN algorithm demonstrated remarkable performance in real-time prediction of the critical process parameters. Utilizing a separate test dataset of 31,400 images, the algorithm achieved an impressive accuracy rate of 91.4% for flow rate prediction and 82.5% for z-height offset prediction. These exceptional results showcased the immense potential of this novel approach in ensuring consistent and high-quality material deposition within MEX processes. This research presents an effective approach that harnesses the combined power of machine learning and vision systems to revolutionize MEX processes. The ability to predict process parameters with high accuracy paves the way for the implementation of automated error detection and correction systems. Such advancements will not only minimize process inconsistencies and improve product quality but also unlock the potential for increased automation and intelligence within MEX.

Presenting Author: Dian-Ru Li National Taiwan University

Presenting Author Biography: Dr. Dian-Ru Li is currently an Assistant Professor in the Department of Mechanical Engineering at National Taiwan University. Dr. Li’s research interests lie in the field of design and manufacturing with the focused areas on biomedical engineering and advanced & smart manufacturing. She is experienced with applying engineering methodologies to reveal human-tissue interaction mechanics and enable innovative medical device development. Beyond this, her current research themes also focus on implementing advanced computer science techniques into manufacturing processes to improve the product quality, and achieve process intelligence and sustainability. Dr. Li has 11 peer-reviewed journal articles, 11 conference proceedings, and 4 patent applications (with 1 USA patent awarded). She is also a Member of American Society of Mechanical Engineers.

Authors:

Tsan-Hung Fu National Taiwan University
Ting-Yuan Huang National Taiwan University
Dian-Ru Li National Taiwan University