Classification of blade’s leading edge based on neural networks in adaptive machining of near-net-shaped blade

The near-net-shaped blade is adopted in the aero engine as it’s material-saving and efficient. However, the leading edge shape’s curvature is sharply changed in its machining process and the deformation trend of each cross section has slight differences. Using the traditional machining method is exhausting and time-consuming. Furthermore, it brings more errors during the whole machining process. Therefore, adaptive machining is imported in the machining of the near-net-shaped blade and the leading edge is to be reconstructed during this process. Besides, it is necessary to know whether the reconstructed leading edge is qualified. To address these two issues, a novel approach is proposed to discriminate and classify leading edges. In this paper, we trained a style transform model of generative adversarial networks with theoretical leading edges and used its discriminator network to evaluate the similarity of reconstructed leading edges we had accomplished in our previous work to establish a standard for the qualified reconstructed leading edge. Then, as the curvature of the near-net-shaped blade changes sharply and has complex features, which requires high accuracy of classification, different DenseNet models were adopted to classify whether these reconstructed images are qualified. We experimented on our LDEG dataset and the highest accuracy on the test set was 88.7%. The experiment results demonstrated that the proposed method is effective in evaluating and classifying leading edges in the machining process.