When parts inspectors at the aviation manufacturer Safran received a toilet lid that the company just manufactured, they used to spend 20 to 30 minutes examining it.
On airplanes, lavatories need to be perfect, so the inspectors search for consistent colors and smooth surfaces free of paint bubbles and scratches, among other defects.
Some inspection professionals have been performing these quality-assurance roles at Safran for more than 10 years and can detect up to 90% of the production defects in cabin galleys, passenger seats, and other parts, said Bindioa Ouali, the senior director and head of digital production, robotics, and automation at Safran.
Still, fatigue can set in after hours of inspections, causing detection rates to vary. AI can help streamline these intensive quality assurance processes — but only if companies have enough data to train the technology, said Zheng Yao, a principal research scientist at Lehigh University’s Energy Research Center, in an interview with Business Insider. When there’s not enough, they can use synthetic data, or artificial information that mimics real data points, he said.
Enter software companies like Loopr AI, Spectron, and Akridata that are creating synthetic data from computer-generated images, 3D models, and other simulation tools to address data scarcity and help prevent the costly scrapping and reworking of faulty parts.
Solving for data scarcity
Safran began pilot projects two years ago with Loopr AI to see whether artificial intelligence could improve defect identification while reducing inspection and documentation time, Ouali said.
In situations like Safran’s, where the company produces a wide variety of parts but not a high volume of each, “it’s extremely hard to have thousands of data sets very early on,” Priyansha Bagaria, Loopr’s founder and CEO, told Business Insider.
In cases like this, Loopr synthetically generates data by creating new training samples from existing images, altering them to increase the number of samples from which algorithms can learn. This allows Safran to simulate various production scenarios while still using real-world data to validate the results, said Bagaria.
The analysis software works well at detecting cosmetic defects in cabin assemblies, comparing them against the production drawings and the listed parts to use, Bagaria said. She added, “AI is really good at removing the subjectivity and retaining the knowledge of your operators. You can automate the repetitive tasks and standardize your processes.”
With paint analysis, for example, AI learns the visual characteristics of the defect, as well as the specific part on which it could appear, like a table or a toilet seat, Bagaria said. The model also uses data from each specific manufacturing component and from lighting conditions that mimic those in the areas where the parts will ultimately be installed.
With Safran’s toilet seats, inspection now takes five to 10 minutes and includes an automatically created defect and inspection sheet before the parts move to the next stage, Ouali said. This documentation was previously created by the inspector.
Additionally, Loopr’s technology achieved a 90.9% recall — the ability to identify all defects, such as scratches, paint discoloration, and paint bubbles — on the project, Ouali said.
One AI software program, two inspection methods
Depending on the setup, the Loopr AI inspection process can use a more automated or hybrid approach.
For Safran’s paint inspection program, the company already had visualization hardware in place, said Bagaria. The program was automated in that the image capture and AI analysis were automated, but the platform allows a human to review, confirm, or correct the AI’s findings, she said. This feedback also helps improve the model over time.
Loopr is hardware-agnostic and can run on various camera and tablet types a company already uses. For this project, a robotic arm photographed parts prior to and after paint application, then provided them to Loopr for analysis.
With a hybrid approach, like for a cabin assembly that includes a lavatory or galley, the inspectors manually take photos at different angles and upload them to Safran’s computer system. The Loopr algorithm accesses the photos and identifies potential issues, creating a defect and inspection sheet.
With a faster inspection process, “we’re able to produce 10% to 15% more than before. The throughput increased to send parts to the final assembly,” Ouali said.
Bagaria said that with successful pilots completed, Safran is now introducing Loopr’s technology to its Marysville, Washington, and Santa Maria, California facilities.
