Deep learning-based borescope image processing for aero-engine blade in-situ damage detection

Abstract

As an important part of aero-engines, blades are used to compress the air entering the engine and produce a lot of thrusts. Catastrophic and sudden accidents caused by blade failure seriously threaten aero-engine operation safety, hence it is necessary to check blades regularly to ensure their health and reliability. As a common non-destructive testing technology, borescope inspection is widely used in health monitoring and maintenance of aero-engine blades. However, traditional borescope inspection mainly relies on artificial vision and it is a time-consuming and experience-dependent process. In this paper, a deep learning-based blade damage detection method is proposed to endow borescope inspection with intelligence. The proposed method pays more attention to texture information, which reflects the types of damages. It is applicable to the situation requiring higher localization accuracy due to the balance between coarse-grained and fine-grained localization. In this method, the enhanced Mask R-CNN network with three functions of damage mode classification, damage localization, and damage area segmentation is constructed. Moreover, a texture-focus multi-scale feature fusion network is used to give more attention to the shallow texture information which reflects the shape of damage. Balanced L1 loss is introduced to balance coarse-grained and fine-grained localization by adjusting the gradient and loss of easy samples. We also propose practical evaluation metrics for blade damage detection and make detailed evaluations and discussions. Extensive experiments are conducted on simulated and real aero-engine damaged blade datasets to verify the effectiveness and progressiveness of our method, and the results show the method has great potential for intelligent detection of aero-engine in-situ blade damage.