Abstract: Internal defects in high-quality large-section continuous casting billets are a primary cause of quality disputes in special steel production. To address the issues of strong subjectivity and susceptibility to missed detections and misclassifications in handheld ultrasonic flaw detection, this paper introduces machine learning methods to achieve automatic classification and identification of internal defects in large-section 42CrMoA continuous casting billets through time-domain feature extraction and modeling of ultrasonic A-scan signals. A total of 13 time-domain features are extracted, and classifiers including fine-grained decision tree and support vector machine (SVM) as well as a feedforward neural network are constructed. The neural network is trained using three algorithms: Levenberg-Marquardt (LM), Bayesian regularization (BR), and scaled conjugate gradient (SCG). The results show that SVM achieves the highest classification accuracy of 99.8%. Among the three neural network algorithms, BR exhibits the best fitting accuracy, while LM and SCG show faster convergence, with SCG striking a balance between convergence speed and generalization performance. Regression analysis indicates that the correlation coefficient between the output and target values of the SCG model exceeds 0.98, validating its effectiveness. The study demonstrates that machine learning methods can effectively replace manual judgment for intelligent identification of internal defects in billets. Based on the above comparative analysis, this paper proposes a cascading strategy of “preliminary screening with traditional methods followed by refined inspection with neural networks”, providing a feasible technical pathway for intelligent quality assessment of continuous casting billets in special steel enterprises.