Adaptive receptive field based on multi-size convolution kernel for micro-defect detection of turbine blades

Abstract

Because turbine blades are essential parts of aircraft engines, micro-defects on the surface of turbine blades induce accidents in aircraft crashes. The detection of micro-defects in aerospace turbine blades is achieved within the realm of non-destructive evaluation. Given that most of the defects are tiny and spread randomly on a curved surface, which cannot be completely detected by current target detection methods, it is feasible for micro-defects to be effectively detected by the fluorescent magnetic particle method. In this paper, we propose a Defect Classification (DCLS)-Deformable Detection Transformer (DETR) model to inspect micro-defects in an automatic fluorescent magnetic particle inspection system. In the DCLS-Deformable DETR model, an adaptive receptive field (ARF) attention module is designed, which considers the interdependencies between the channel features of different feature maps. The weights between the channels of each feature map were also considered, while adaptively adjusting the receptive field according to the detection target size. The DCLS-Deformable DETR model with ARF increased the AP from 63.4% to 64% and AP50 from 95.2% to 97.2%, compared to the Deformable DETR. Turbine blades include three typical defects: linear cracks, reticular cracks, and pitting corrosion. Compared with the original model, the proposed model enhances the AP of three defect types by 1.8%, 2% and 4.7% respectively. The DCLS-Deformable DETR model considers the position, level information, and channel information of the input samples, which can capture micro-defects around large defects.