Quality evaluation modeling of a DED-processed metallic deposition based on ResNet-50 with few training data

Abstract

The additive-manufacturing (AM) field necessitates a robust process-monitoring system for quality assurance and control. To meet this industrial requirement, quality-evaluation models have emerged as powerful tools for providing quality feedback. Recently, convolutional-neural-network- (CNN)-based classification models have gained popularity in quality evaluation using image data. However, such models require sufficient image data for training, a requirement that is challenging to fulfill in the context of metallic AM due to the complexity of decomposition and analysis. This challenge is particularly pronounced in start-up or medium-sized metallic-AM enterprises. Moreover, many countries around the world have faced a decline in population and a shortage of labor in the engineering field. This growing shortage of workers to collect datasets exacerbates this challenge. In this study, experiments of directed-energy-deposition (DED) processes for single-line and single-track metallic deposition using AISI 316 L stainless-steel powders were conducted with two experimenters. After the process, a minimal amount of cross-sectional surface image data of the metallic deposition was binary-processed and analyzed across three quality states: normal state, surface burrs, and internal defects. To compensate for the lack of training data, multiple strategies are proposed, including image preprocessing and ResNet transfer learning. The selection of an optimization solver and layer depth for maximizing classification performance was discussed. The potential performance of ResNet dealing with binary images and performance standards with few training data was also identified by comparing with other higher-level architectures (Inception and Xcepition).