Artificial Intelligence in Laser Powder Bed Fusion Procedures – Neural Networks for Live-Detection and Forecasting of Printing Failures.

Abstract

Recently, many disruptive technologies like additive manufacturing were developed and have grown to production grade solutions. While such technologies can reduce the required time and amount of material, they require new methods to address long printing times and error identification in complex process structures. Such systems can be created using artificial intelligence (AI), such that operators can focus on supervising monitoring systems rather than the printing process and can decide upon preprocessed information.For the laser powder bed fusion process (LPBF), a main challenge lies in error detection per layer and in predicting the necessity of near-future operator interaction, to meet quality standards. Yet, only few works aim to automate the task of characterizing build layers.We propose an AI-driven approach to identify and forecast printing errors in LPBF using convolutional neural networks (CNNs) and the Cox proportional hazards (CPH) approach. For training and validation, layer wise greyscale images of 27 print jobs have been recorded. All print jobs’ results were qualitatively rated afterwards by experts. We then perform an autoencoder-based clustering and compare print jobs of high and low quality to create pseudo-labels. We finally train a CNN in a supervised fashion to classify layers as succeeded or faulty, as well as a CPH model to estimate remaining build time without errors. We achieve an overall accuracy of 94.9% for the CNN and statistically significant forecasts (p < 0.005) of the imposed hazard per layer using the CPH model. Our work demonstrates a novel approach of automated build part characterization, that reduces human operator effort and can be used for automated quality assurance and process monitoring.