Internal Defect Classification for Large‐Section Continuous Casting Round Billets via Multi‐Scale Feature Fusion and Attention Enhancement

Accurate identification of internal defects in large‐section continuously cast round billets is critical for ensuring high‐grade special steel quality. Traditional ultrasonic A‐scan detection methods rely on manual experience and suffer from low signal‐to‐noise ratios and high defect signal similarity, resulting in insufficient identification accuracy. This study proposes a deep learning classification method based on multi‐scale feature fusion and attention enhancement. The method designs a multi‐dimensional feature extraction framework integrating time‐domain statistical features with fractional Fourier transform (FRFT) domain features, utilizing Fisher Score and mutual information mechanisms to select 16 key features from 123 initial dimensions. A three‐branch parallel network structure incorporates adaptive attention mechanisms for dynamic multi‐modal feature fusion. Learnable weighted combination of cross‐entropy, triplet, and contrastive losses addresses class imbalance issues. Experiments on an industrial dataset of 319 S355NL(Q355NE) continuously cast round billet samples demonstrate overall classification accuracy of 89.3% for five defect types: core porosity, shrinkage cavities, cracks, normal conditions, and composite defects, achieving a macro‐average F 1 ‐score of 0.893. Crack defect identification reaches 94.4% accuracy. Compared to baseline methods including MLP, ResNet, and FCN, the proposed method achieves the best trade‐off between classification performance and computational efficiency.