Nonlinear ultrasonic testing of composite porosity based on CWT and KAN-ResNet50 under simulation data enhancement

Linear ultrasonic techniques often exhibit insufficient sensitivity to small defects in composite materials. Furthermore, current deep learning approaches for porosity detection rely heavily on hand-crafted features and suffer from limited generalization capabilities when trained on small datasets. To address these challenges, this paper proposes a novel nonlinear ultrasonic detection method for composite porosity based on Continuous Wavelet Transform (CWT) and a KAN-ResNet50 architecture, enhanced by simulated data augmentation. First, a random pore model with varying porosity levels is constructed based on random medium theory. COMSOL Multiphysics is then utilized to simulate and extract the corresponding nonlinear ultrasonic response signals. Subsequently, CWT is applied to both simulated and experimental ultrasonic signals, converting them into time-frequency images to construct a comprehensive dataset for porosity classification. An ImageNet-pretrained ResNet50 model serves as the backbone network, with its top-level residual blocks fine-tuned. Notably, a Kolmogorov-Arnold Network (KAN) replaces the traditional fully connected layer as the classification head to form the final porosity evaluation model. Experimental results demonstrate that the proposed method achieves an average classification accuracy of 98.15% across four types of porosity samples on the test set, verifying its effectiveness. This study provides a highly feasible, automated, and intelligent approach for the non-destructive evaluation of composite porosity, significantly reducing the reliance on large-scale experimental samples.