A Dual-regularization Mechanism used for Ultrasound Signal Classification by Acoustic Velocity-guided Dropout and Squeeze-and-excitation Attention

In intelligent nondestructive evaluation (NDE), overfitting on small datasets poses a significant limitation to the generalization of ultrasound classification models across different materials. Consequently, the development of effective regularization techniques is crucial for designing robust multi-material NDE systems. In this paper, we propose a versatile and lightweight dual-regularization module comprising two sub-modules: Acoustic Velocity-Guided Dropout (AVGD) and Squeeze-and-Excitation (SE) attention. The AVGD integrates physical domain knowledge with conventional regularization methods by dynamically adjusting the dropout rate of feature channels based on acoustic velocity information. Meanwhile, the SE attention mechanism enhances critical features in conjunction with dropout, thereby improving the model’s learning capacity. Both sub-modules are encapsulated into a dropout layer and an SE block, respectively, and seamlessly integrated into a classical neural network architecture. The proposed method is evaluated on a collected ultrasound signal dataset and compared against standard regularization mechanisms. Experimental results demonstrate that the dual-regularization mechanism significantly enhances the generalization capability of the baseline model.