Machine Learning Based Approach to Process Characterization for Smart Devices in 3D Industrial Manufacturing

Abstract

A key differentiator between the additive manufacturing and the traditional injection molding is the precision-manufacturing. An error-free print job significantly guarantees good part quality with minimum wastage of the material and energy. In practice, however, achieving error-free production is quite challenging and this emphasizes the need to learn what behavior of the machine leads to an erroneous job. Knowing the health of the printer or the print job helps quantify print job performance, as well as build system alerts to take reactive actions. Intending to run the model dynamically while the machine is printing, time-series based deep learning models like LSTM are most suitable. This paper presents a machine learning based anomaly detection approach to discover patterns in sensors' measurements in the streaming mode in MultiJet Fusion 3d printer developed at HP Inc. A hybrid architecture of LSTM and Auto-encoder has been proposed to learn the printer behavior generate an alarm in the event of an anomaly. The results of both LSTM and LSTM-Autoencoder models have also been discussed by taking real-life examples of 3D printing jobs.