Berkeley, CA-This Fall, a team of five UC Berkeley students from Data-X partnered with GE Additive in order to explore improvements in metal 3D printing. They did this by extracting features from images from printer data, which they used to train prediction models that predicted an important 3D printing parameter . This can help to make metal 3D printing faster.
Metal 3D printing is a new manufacturing paradigm that enables the production of complex parts in one machine. This new method can also reduce an assembly of parts to a single 3D printed part, simplify supply chains and improve material performance. However, this new technology is complex and has some interesting challenges. GE Additive, one of the leading metal 3D printing companies, brought some of those challenges to the Data-X course to try to find improvements from a data-centric approach.
Team “GE Smart print” worked with data from a GE 3D printing machine, that included images from a 3D printing experiment. The objective from GE was to use this data to generate a predictive algorithm for an important 3D printing parameter.
The first step taken was to extract features from the images (about 80,000 in total) and generate a single csv. The 2nd step that the team took was an exploratory data analysis on the generated CSV files, after which they applied different feature extraction methods from the images, and they ended by training various machine learning models on the generated data. From the model training and testing, it was determined that a linear model gave the lowest RMSE error. The 3rd step that the team implemented was to calculate more features from the images (using Sobel gradients), and exploring new approaches, such as Autoencoders, which is a type of Neural Network.
Enrique Marin: Project Manager
Armyben Patel: Engineering Team
Ziren Lin: Engineering Team
Ada Sun: Engineering Team
Hassan Khawaja: Engineering Team