Recognition of Residual Cores in Aero-Engine Blade Neutron Images Using Improved Patch SVDD

Abstract

The recognition of residual cores in aero-engine blades is a crucial task in ensuring the safety and reliability of aircraft. Compared to techniques such as borescope and X-ray radiography, neutron radiography, with its strong penetration ability and high sensitivity to light elements, can detect residual cores as thin as 2 mm within complex cavities. This significantly enhances the detection rate of residual cores in aero-engine blades. However, the recognition of residual cores in neutron images currently relies heavily on manual inspection by professionals, which is subjective and inefficient. To address this issue, an improved residual core recognition method based on a patch-level support vector data description (Patch SVDD) algorithm is proposed for neutron images. This study employs an improved gamma transformation to enhance the quality of neutron images and highlight the features of aero-engine blades. A fusion of dilated residual network (DRN) and efficient channel attention (ECA) serves as the feature extraction network in Patch SVDD, improving the capability of feature extraction. Additionally, a residual core grading module is designed to improve core leaching efficiency in production. Neutron images of aero-engine blades were acquired through the reactor-based cold neutron radiography facility (CNRF) to construct a dataset. The results demonstrate that this improved method achieves areas under the receiver operating characteristic curves (AUCs) of 94.8% at the image level and 95.6% at the pixel level, indicating its favorable recognition efficacy. This study provides an intelligent method for quality monitoring in aero-engine blades.