Contact Us

TEL: +86-020-82497630
Mob: +8613694265790
Address: 1F,2F Building 8 Yujing Industry Zone, Dalingshan Road, Tianhe District, Guangzhou, Guangdong, China

Home > Knowledge > Content
LLNL use machine learning to prevent metal parts 3D printing defects in real time
Sep 20, 2018

California's Lawrence livermore national laboratory (LLNL) of engineers and scientists developed a convolutional neural network (CNN), which is a popular algorithm is mainly used for processing images and video, is used to predict the defect of 3D printing parts, and detect whether there is a build in milliseconds will be satisfactory quality.

"This is a revolutionary way to see through video tag data, or even better frame by frame marker," chief LLNL researcher Brian Giera said."The advantage is that you can print out parts of the video collection, and the final conclusion when printing.A lot of people can collect the data, but they don't know how to deal with it, the work is a step in this direction."

Usually, Giera explained that after the build sensor analysis is expensive, and parts quality can only be determined after a long time.To take several days to weeks to print parts, CNN can prove that helps to understand the printing process, faster to understand the quality of the parts, and real-time correction or adjust the build when necessary.

LLNL researchers use about 2000 molten laser track video clips in different conditions (e.g., speed or power) neural network is developed.They used to generate 3D height map tool scanning parts surface, the algorithm USES this information to analyze the various components of a video frame (each region is called the convolution).Giera explained that this process is very difficult and time consuming for the human beings.

Students at the university of California, Berkeley, and researchers at LLNL Bodi Yuan developed can automatically mark each build a height map algorithm, and use the same model to predict the width of the orbit, orbit is destroyed and the width of the standard deviation.Using these algorithms, the researchers were able to shoot video of the ongoing build, and to determine whether the parts showed acceptable quality.As a result, the neural network to test for a continuous parts accuracy of 93%.

Is critical to the success of the "we can learn many useful in the process of training (CNN) video function," yuan said, "we only need to provide a large amount of data to train it, and make sure that it was good at it."

LLNL researchers have spent years collecting laser fusion metal powder bed in the process of 3D printing all kinds of real-time data, including video, optical tomography and acoustic sensors.

"In any case, we are gathering video, so we only join points," said Giera."Just like the human brain using vision and other senses to navigate the world of machine learning algorithm can use all the sensor data to navigate 3D printing process."

Giera said, neural network theory can be used for other 3D printing system.Other researchers should be able to follow the same formula to create components under different conditions, collecting video and using a height map scanning them with standard machine learning techniques to generate the tags used video set.

Giera said, still need to do some work to test parts within the gap, these gaps cannot predict through a height map scanning, but can use the in situ X-ray photographic measurement.

The researchers will also seek to create algorithm, by combining a variety of perception mode except images and video.

"Now, any type of detection is considered a big victory.If we are able to repair it immediately, it's a bigger goal, "said Giera."Given that we are in the collection of machine learning algorithm is designed to deal with large amounts of data, machine learning will be created correctly parts for the first time to play a central role."

The project funded by the laboratory to guide the research and development plan.

Previous: 3D printing of trachea splint recoverable 7-month-old babys breath

Next: The lockheed Martin team develop and deakin university cooperation FORTIS exoskeleton