Session: 07-01: NDE for Additive Manufacturing / 03-01: Electromagnetic NDE Techniques

Paper Number: 147665

147665 - Novel Deep Learning Method for Real-Time Diagnostics and Correction for Additive Manufacturing 

Abstract: 

This study presents a novel approach to overcome process reliability challenges in Material Extrusion (ME), one of the prominent additive manufacturing (AM) techniques. Despite ME's advantages in cost, versatility, and rapid prototyping, it faces significant difficulties in commercial scale production, primarily due to quality issues such as overextrusion and underextrusion, which compromise final part performance. Traditional manual monitoring methods fail to detect these defects efficiently. Researchers have proposed some in-situ monitoring techniques such as optical imaging, acoustic sensing, accelerometer-based displacement analysis, etc., which are semi-automatic, yet the overall performance of these techniques is still non-reliable. Among these techniques, optical imaging is the most widely implemented method of detecting defects in material extrusion process where the printing nozzle stops after printing each layer of the part to capture the image of the just printed layer. This not only hinders the printing process, but it is also a non-viable option for commercial implementations. Therefore, it is necessary to develop an efficient and real-time monitoring solution. Considering these challenges, an innovative and field deployable infra-red thermography based in-situ real-time defect detection and feedback control system is proposed in this study. A novel experimental setup has been developed to integrate IR camera which is capable of capturing images in real-time. A robust data acquisition framework is developed and automated to collect spatial and temporal thermal information during the extrusion process. The study elaborates on the preprocessing of this data for deep learning model training, including sub-segmentation and temporal trimming. Two deep learning architectures, Convolutional Neural Network (CNN) and hybrid CNN and Long-Short Term Memory (LSTM), are trained and evaluated for their ability to detect defects in real-time. Based on these algorithms, classification models are developed and analyzed for performance based on accuracy and required time to detect a predefined under or over extrusions. In this study, CNN+LSTM model is found highly accurate (> 98%) and readily implementable online for detecting defects within three seconds. To eliminate the propagation of under or overextrusions, a feedback control system is designed in MATLAB based on the outcome from the selected deep learning model. The proposed closed-loop control system demonstrates the dynamic adjustment of relative flow rate and ensures consistent production of high-quality parts. This approach not only ensures the production of defect-free parts within tolerable ranges but also sets a foundation for future research in automated monitoring systems for additive manufacturing. The findings and methodologies detailed in this study contribute to the advancement of AM and offer a path toward overcoming one of its major barriers to commercial and industrial application.

Presenting Author: Sourav Banerjee University of South Carolina

Presenting Author Biography: Sourav Banerjee, is a Professor in the Department of Mechanical Engineering at the University of South Carolina

Authors:

Asef Ishraq Sadaf Georgia Southern University
Hossain Ahmed Georgia Southern University
Sourav Banerjee University of South Carolina