Automatic classification of welding defects from ultrasonic signals using an SVM-based RBF neural network approach