John Deere piloted AI and machine vision tech from Intel to improve efficiency in the robotic steel welding process to manufacture heavy equipment and products.
The AI approach could be especially valuable partly because of the volume of work produced with Gas Metal Arc Welding (GMAW) at 52 Deere factories around the globe where hundreds of robotic arms use millions of weld wire a year. Detecting defects in steel welds has traditionally been a manual process relying on skilled technicians, Intel wrote in a blog.
Inspectors commonly look for porosity in the welds where trapped gas bubbles form as a weld cools. The cavities left by the bubbles weaken the weld strength and if not detected require a re-do or possibly a scrapping of steel assemblies, causing delays and expense.
Deere’s approach used AI and machine vision to spot the porosity, relying on hardware and software to detect the flaws in real time at the edge reliably. The approach used a neural network inference engine, which logs a defect quickly and automatically stops the welding work. Deere can correct the issue right away, Intel said.
The work positions Deere to solve other problems in the manufacturing process beyond weld defects, according to Christine Boles, vice president of IoT at Intel.
“The introduction of new technology into manufacturing is opening up new opportunities and changing the way we think about some processes that haven’t changed in years,” said Andy Benko, quality director at John Deere’s construction and forestry division. “The AI solution has the potential to help us produce our high-quality machines more efficiently than before.”
Deere’s weld detection was enabled with Intel Core i7 processors, Intel Movidius Vision Processing Units, the Intel Distribution of OpenVINO toolkit and an industrial grade ADLINK Machine Vision Platform and a MeltTools welding camera.
The pilot of weld porosity detection and control took place at an unidentified production facility using existing industrial PCs, according to Industry Week.
ADLINK has said the use of OpenVINO and other tools provides up to 97% accuracy in detecting porosity defects based on its internal testing data.
To detect porosity defects earlier than human inspectors, Intel and ADLINK have used a camera placed on the welding gun about a foot away from the actual weld. Video frames from cameras are tracked by a 2D classification network, followed by an aggregation process to average responses. The stream video frames are examined the inference engine and when defects are found, the system switches off the welding robot.