A Pore Classification System for the Detection of Additive Manufacturing Defects Combining Machine Learning and Numerical Image Analysis

Abstract

This study aims to enhance additive manufacturing (AM) quality control. AM builds 3D objects layer by layer, potentially causing defects. High-resolution micrograph data capture internal material defects, e.g., pores, which are vital for evaluating material properties, but image acquisition and analysis are time-consuming. This study introduces a hybrid machine learning (ML) approach that combines model-based image processing and data-driven supervised ML to detect and classify different pore types in AM micrograph data. Pixel-based features are extracted using, e.g., Sobel and Gaussian filters on the input micrograph image. Standard image processing algorithms detect pore defects, generating labels based on different features, e.g., area, convexity, aspect ratio, and circularity, and providing an automated feature labeling for training. This approach achieves sufficient accuracy by training a Random Forest as a hybrid-model data-driven classifier, compared with a pure data-driven model such as a CNN.

Keywords

• additive manufacturing
• pore classification
• machine learning
• numerical image analysis