Residual stress regression method based on 1D residual convolutional network for ultrasonic detection

To achieve high-precision, non-destructive prediction of residual stress in metallic materials, this paper proposes a regression model based on a one-dimensional residual convolutional neural network. The model extracts features from ultrasonic time-domain signals and performs stress regression. By incorporating multi-scale convolution, residual connections, and global average pooling, the model enhances deep feature extraction capabilities. Additionally, max pooling is applied in the shallow layers to compress redundant information, followed by fully connected layers for final stress prediction. Experimental results demonstrate that the proposed model outperforms the traditional time-of-flight (TOF) method based on cross-correlation in terms of error metrics, achieving a mean absolute error (MAE) of 3.011 MPa - representing a 52.7% improvement over the TOF-based approach. This method offers an effective solution for non-destructive evaluation of residual stress in materials and shows strong potential for engineering applications.