Abstract

The presence of anomalies in components generated using laser-based Additive Manufacturing (AM) processes greatly affects mechanical properties, making anomaly detection and prevention significant for the improvement of AM. Melt pool thermal imaging of laser-based AM processes such as Direct Energy Deposition (DED) provide rich data from which anomalies can be detected. Although through feature-based learning methods (e.g., support vector machine and k-nearest neighbor) or end-to-end deep learning models (e.g., self organizing map) the anomalies can be identified, the sparsity of defective instances and lack of interpretable analysis, together with high dimensionality and velocity of melt pool images further increase the difficulty of anomaly detection. We propose implementing Neural Vector Quantized Variational Autoencoder (Neural-VAE) to analyze thermal images from a DED process to determine where and when anomalies occur, indicating a flaw in the generated component. A Vector Quantized VAE is designed to represent high-dimensional images as low-dimensional vectors. Then, hyperdimensional computing is introduced to leverage the low-dimensional representation of image data and determine defective and non-defective images. Our proposed methodology can detect anomalies with an accuracy, sensitivity, specificity, and f-score of 99.2%, 99.6%, 98.8%, and 99.2%, respectively, while successfully generating anomalous images.

This content is only available via PDF.
You do not currently have access to this content.