Neural network put on guard of aircraft parts quality in MAI
Scientists of Moscow Aviation Institute developed software to check the quality of aircraft parts. The engineers created the program on the basis of neural networks R-CNN and ImageNet. They were trained on thousands of images of defects and pores in metal products printed on a 3D printer. The output turned out to be a multi-layered neural network that is able to recognize defects on CT scans with high accuracy.
- The algorithm resembles of how people recognize objects. Imagine that you are looking at a fruit basket. Your brain is able to distinguish an apple from an orange or identify a banana, even if only the half of it is visible, because the rest is behind the grape. Our software does something similar, but with computed tomography images, revealing defects and pores in metal parts, - said Konstantin Korobov, the project developer, MAI engineer.
Jet engine components or aircraft structural parts are still checked manually. The development will speed up the verification by 1.5-3 times, automating the process. The method will increase the reliability of quality control, since due to automation the possibility of human error is reduced.
- Detecting and eliminating defects early in the production process can prevent failures in a complex process chain. Thus, the introduction of this technology at enterprises might lead to significant cost saving,” said Konstantin Korobov.
In addition to the aircraft manufacturing industry, the program is of interest to the doctors. It may be used to test components of medical devices, such as surgical instruments or prosthetics.
Work on the project is carried out on the basis of departments No. 903 “Advanced materials and technologies for aerospace applications”, No. 904 “Engineering graphics” and No. 910B “Mechanics of nanostructured materials and systems”. At the current stage, the product is going through the testing stage. At the same time, developers continue to improve the performance of the algorithms. A fully functional version of the software is planned to be released in early 2024. At the same time, MAI engineers are ready to sign agreements in order to introduce their product into production.
Main photo: Unsplash.
