Session: Research Posters

Paper Number: 119720

119720 - A Comparative Study of Feature-Based and Image-Based Clustering Techniques for Laser Powder Bed Fusion Process Monitoring 

When it comes to producing complex parts, Additive Manufacturing (AM) is quickly becoming a viable option. Laser Powder Bed Fusion (L-PBF) is one of the most popular AM technologies that can melt and fuse powder material into solid, layered parts based on intricate 3-D patterns. However, the L-PBF process is complex, involving powder spreading, heating, melting and solidification and the quality of the finished parts is highly dependent on the characteristics of the AM feedstock material, the process parameters, and the performance of the machine. Any deviations from the specifications may cause process anomalies and result in part defects. Understanding the interplay of these factors remains an ongoing task. Melt pool monitoring (MPM), a common method for measuring thermal conditions, can also be used to detect process irregularities in real time, and thus predict the part quality. 

The purpose of this research is to improve melt pool-based process monitoring and anomaly detection for real-time control through a comparative study. Various approaches have been reported for melt-pool based anomaly detection, including traditional Machine Learning (ML) algorithms using melt pool features and deep learning based techniques. In this research, we conducted a thorough comparison of methods for process anomaly detection using melt pool images. In particular, the investigation delves into both Deep Learning (DL) and feature-based ML paradigms, accentuating their utilization in anomaly detection via clustering techniques. We explore how various melt pool features and clustering algorithms such as Kmeans, SVM classifier, spectral clustering and hierarchical clustering affect the anomaly detection performance. For the feature-based study, the features extracted from the melt pool images include melt pool morphological characteristics, intensity distribution and gradient. To identify the most relevant features for process anomaly detection, we applied Formal Concept Analysis (FCA) to down-select the features, extract and highlight the most crucial features necessary for the analysis of the melt pool. FCA is a multi-relational data mining technique that maps data and applies mathematical theory to derive association rules from multiple parameters. These association rules can later be reduced for dimension reduction and then decrease and select the input features for the clustering algorithms. In the image-based method, a meaningful representation is directly extracted from the image without supervision using a Deep Convolutional AutoEncoder (DCAE). The extracted representation is used as input for clustering algorithms. The ML and DM methods will be compared in terms of precision using MPM labeled images, and computational efficiency to demonstrate their real-world applicability. The study employs data from the NIST Additive Manufacturing Metrology Testbed (AMMT). The analyzed part has a total number of 40 layers. The examination will narrow down to 21 layers with each layer containing around 5000 images. The preliminary results show that the process of clustering is significantly sensitive to the selection of features and algorithms. An additional observation underscores that image-based analysis, if used exclusively, might not provide all-encompassing insights. The study also illustrates that minor groups representing MPM locations in defect regions can still be identified without data labeling.

Presenting Author: Yande Ndiaye NIST

Presenting Author Biography: Yande Ndiaye is a Ph.D. student at the University of Lorraine, France. She has a master's degree in IT from a prestigious school in France. Her research focuses on using Formal Context Analysis to extract Knowledge in the Additive manufacturing area.

Authors:

Yande Ndiaye NIST
Jaehyuk Kim National Institute of Standards and Technology (NIST)
Zhuo Yang National Institute of Standards and Technology (NIST)
Yan Lu National Institute of Standards and Technology (NIST)
Mario Lezoche CRAN, Université de Lorraine