Infrared Thermal Imaging-Based Turbine Blade Crack Classification Using Deep Learning

Abstract Non-destructive testing is widely applied for the detection and identification of defects in turbine blades of modern aircraft engines. Cracks in turbine blades can affect the turbine performance and pose a risk to safety and service life. For Original Equipment Manufacturers it is, therefore, essential to be able to identify all defects. Heat flow thermography offers, compared to the often used penetrant testing, the potential to improve the detection of defects in turbine blades and is contact-free, reproducible, quick to apply, and can be automated. With induction (heat flow) thermography, it is even possible to detect cracks that lie below the surface and therefore are not externally visible. However, manual inspection of thermography images is very time-consuming. By automating the image classification procedure with a deep learning technique, the speed and accuracy of the classification can be improved over a manually performed classification. The development objective of this AI application is expected to support and assist the highly skilled and experienced inspection specialists in the medium term. Our solution is based on convolutional neural networks. Several challenges of the AI training process, including data imbalance, a small dataset, and extremely small cracks in large images are addressed.