Ultrasonic high-precision stress prediction algorithm using multi-feature-fusion Kolmogorov–Arnold network

Accurately characterizing the stresses in components poses a significant challenge in industry. Ultrasound offers a quick and convenient option for stress characterization. However, low resolution limits its application. To address this issue, this study proposed a high-precision stress prediction algorithm based on a Kolmogorov–Arnold network (KAN) method with multi-feature fusion. First, feature indicators from different domains, including the time and frequency domains, were extracted. Then, by combining the XGBoost method with feature correlation analysis, the contribution rates and importance values of the extracted feature indicators for stress characterization were computed for feature selection. Subsequently, a deep KAN was constructed, a physical consistency constraint loss function was designed, incorporating adaptive activation functions to enhance the stress characterization accuracy. Finally, single-feature and multi-feature models were compared against various machine learning regression models, and the mean absolute error of the proposed model was 1.7319. The results showed that the multi-feature-fusion KAN algorithm offers higher prediction accuracy for stress characterization.