Machine Learning and Metal 3D Printing Combine for Real-Time Process Monitoring Algorithm

A finished metal part (damaged region in red).

The day is slowly coming when metal 3D printing will be widely considered as a reliable industrial manufacturing method, but there are still some issues to deal with before we get there. A lot of research goes into figuring out the underlying cause of flaws in metal 3D printing, which can result in weaknesses such as spatter and micro cracking in the final 3D printed parts – unacceptable when you’re dealing with high-risk applications like aviation components.

But two researchers from the College of Engineering at Carnegie Mellon University (CMU) have figured out how to combine 3D printing and machine learning for real-time process monitoring, a practice which can detect anomalies inside a part while it’s being 3D printed. Their research could one day lead to self-correcting 3D printers.

Luke Scime, an alumnus of CMU’s Department of Mechanical Engineering (MechE), partnered up with NextManufacturing Center director Jack Beuth to create a machine learning algorithm that practices process monitoring with laser powder bed fusion technology, which is prone to errors caused by the powder layer being spread incorrectly.

Other researchers are using methods like acoustic technologies, spectroscopy, and temperature monitoring to learn what’s going on inside the structural level of a build. But while there have been some limited types of monitoring available on the market, they don’t typically have the capacity for automated analysis, and can only provide a reading that a machine operator has to interpret. But Scime and Beuth’s work goes in a different direction: a computer vision algorithm.

Scime said, “One of the biggest hurdles between just making a part that looks good and actually putting it on an aircraft is making sure that the part you’re producing doesn’t have flaws in it.

“Computer vision is a term for using data analysis techniques to understand what’s happening in an image.”

The algorithm analyzing a single layer of the powder bed.

Scime’s innovative algorithm takes images of the powder bed and extracts features, which are then grouped and compared over various levels of analysis until a fingerprint of the image is created. The machine has learned how to recognize different flaws because the researchers fed the algorithm hundreds of pre-labeled training images; now, it can compare the fingerprint of new images it receives to the ones it already knows to isolate various anomalies.

A 3D reconstruction of the part from several thousands of image analyses.

In a paper Scime and Beuth published in the Additive Manufacturing journal, titled “Anomaly detection and classification in a laser powder bed additive manufacturing process using a trained computer vision algorithm,” they demonstrate how the algorithm is able to detect flaws in powder spreading in the millimeter scale while they’re developing. The algorithm can determine what the flaw is and where it’s happening, which can help increase process stability (the ability to print).

For the purposes of the paper, six types of anomalies were used.

The abstract reads, “This work presents an approach for in-situ monitoring and analysis of powder bed images with the potential to become a component of a real-time control system in an LPBF machine. Specifically, a computer vision algorithm is used to automatically detect and classify anomalies that occur during the powder spreading stage of the process. Anomaly detection and classification are implemented using an unsupervised machine learning algorithm, operating on a moderately-sized training database of image patches. The performance of the final algorithm is evaluated, and its usefulness as a standalone software package is demonstrated with several case studies.”

This work is a big step forward in making metal 3D printing a reliable, safe method for industrial manufacturing.

Representative examples of the six powder bed anomaly classes chosen by the authors: (a) Recoater hopping, (b) Recoater streaking, (c) Debris, (d), Super-elevation, (e) Part failure, and (f) Incomplete spreading.

Scime said, “The holy grail is to deploy this in a real-time environment where you’re automatically analyzing data, doing something about it, and then moving on.

“What it really comes down to is, can we detect it, understand that it’s an issue, and then design what we call processing parameters to do something different than we were doing in order to reduce the amount of warpage?”

Scime explained that self-correcting automation could end up working in a few different ways, the most basic of which is a 3D printer sending an alert to an operator once an anomaly is detected so the issue can be addressed early on. Then, you move on to teaching a 3D printer to recognize critical flaws and automatically perform simple fixes, like clearing the blade the spreads the powder bed or stopping 3D printing one part while allowing others to keep going.

However, the crowning achievement in automated self-correction is fighting superelevation. This anomaly, responsible for most part damage, occurs when part of the build starts to warp or curl up out of the powder. While it may be a while before this level of automation is reached, the CMU machine learning algorithm is already able to accurately identify several anomalies and is ready for application in the real world. However, Scime hopes to examine how additional sensor data can be added to its analysis, and also improve its accuracy.

Discuss this and other 3D printing topics at 3DPrintBoard.com or share your thoughts below.