Latent Representation and Characterization of Scanning Strategy on Laser Powder Bed Fusion Additive Manufacturing

Despite the transformative capability of laser powder bed fusion (LPBF) additive manufacturing to create components with intricate geometry, the large-scale adoption remains a barrier owing to the process complexity and significant build quality concerns. In-process melt pool imaging offers an unparalleled capability to tackle the problems by evaluating the impact of prominent process parameters (e.g., laser power, laser velocity, and hatch spacing) on build quality. However, the current investigations overlook the effect of other influential factors such as scan strategies. Because of the multitude and high-dimensionality in melt pool images, the extraction of manual features to characterize and intertwine diverse scan strategies (e.g., orthogonal serpentine, pre-scanned boarder, and clockwise spiral) is cumbersome or inefficient. While end-to-end deep neural networks realize automated feature extraction from melt pool images, they are limited in providing meaningful signatures for the characterization of various scan strategies. This paper presents a systematic image-guided analysis based on variational autoencoder (VAE) that enables the semantic representation of image data on low-dimensional latent space to characterize similarities between scan strategies. Further, hyperdimensional computing as a cognitive solution is integrated to differentiate various scan strategies according to latent features. Experimental results on the real-world case study based on 30,000 in-situ melt pool images show that VAE is significantly effective in interpretable characterization associated with 12 different scan strategies. In addition, the cognitive model differentiates scan strategies using the latent representation with an accuracy of 81.20 ± 0.8%.