AI Can Spot Defects in Aircraft Rivets by Looking at Video
Article Source: MDPI
Rivets play a crucial role in constricting joints on aircraft fuselages and wings, and this has been the case since the early times of aviation. Simply put, aircraft manufacturers cannot use welding for bonding panels together because aluminum isn’t tolerant to heat, riveted joints are stronger, and finally, they are easier to inspect. Compared to screws, rivets are again much better because they can withstand vibrations, whereas screws would come loose after a while.
On the matter of inspection, rivets require many work hours to perform the task, and this time translates to significant amounts of money for the global aviation industry. And even then, some defective rivets may be missed, as modern civil and military aircraft contain anything from 10,000 to 30,000 rivets. This may result in degraded aerodynamic performance when flying and, in extreme cases, may even lead to panel detachments.
A team of Russian academic researchers has developed a deep neural network-based solution to this problem, which can inspect rivets by looking at video or images of aircraft joints and spot the defective ones or at least those that require further inspection by a technician. This system would save service crews hundreds of hours each month and cut the maintenance costs and times by a lot.
The researchers trained their neural network with a rich dataset and created a detailed mathematical model for evaluating rivet health and structural status by looking at its head. The head can have notches, undercuts, irregularities, cracks, deformations, and other signs of potential trouble. The system learned to recognize these signs and classify the rivets into six different categories, two non-defective and four defective.
After several training rounds, the system reached a precision of 98%, which means that for an average of 20,000 rivets, only 400 were misclassified. As such, these systems can be used in combination with human inspections to help speed up the process considerably.