A Vision-Based Neural Networks Model for Turbine Trench-Filler Diagnosis

Vision-based neural networks as artificial intelligence models have been critical in many manufacturing industries, including automotive, food, and aerospace. Machine vision and deep learning have provided practical, promising, and accessible innovations to address many problems during manufacturing and diagnostics. Recent studies offer beneficial results in implementing industrial artificial intelligence systems that require determining, comparing, and evaluating optimal technological solutions. The aerospace industry has to deal with the problem of diagnosing critical components that continually need the expertise of a trained human being. On the other hand, automatic diagnostics are becoming a critical technology to deal with this problem. However, these systems require a particular configuration of computer vision algorithms which, in the case of turbine trench-filling components, have yet to be added to the literature. Considering the above, we report in this paper a new methodology that uses a pre-trained deep neural network framework available online to automatically diagnose geometric nonconformities in aeronautical components that protect the turbines from vibrations and reduce noise from the engines to the aircraft deck. The method is based on the following stages: a computer vision system with a monochrome camera to acquire images, training of deep neural networks with transfer learning, and a stage to analyze nonconformities automatically. In addition, a benchmark of several pre-trained deep neural networks is presented to address the problem. According to the experimental results, the YOLO deep neural network topology significantly contributes to automatic diagnosis with high precision and exact rate. Finally, we show that integrating deep neural networks with control graphs is a promising strategy to optimize online diagnostics in the production lines of the aeronautical industry.