Effective pre-stress identification in steel strand based on ultrasonic guided wave and 1-dimensional convolutional neural network

The accurate assessment of the effective pre-stress in steel strands is a challenging task, and ultrasonic guided wave (UGW) technique has shown certain application prospects in this field. However, the existing UGW-based approaches require manual parameter extraction from signals in time domain or frequency domain, which is a cumbersome and time-consuming process, and pre-stress identification based on individual parameters may not be reasonable. This study proposes a framework for identifying effective pre-stress in steel strands based on UGW and one-dimensional convolutional neural network (1D-CNN), which does not require any parameter extraction operation and achieves high identification accuracy. The output features of various convolutional layers in 1D-CNN were downscaled and visualized, and the prediction results of 1D-CNN were compared with those of a support vector regression (SVR) model. Results show that with the deepening of the network, the correlation between output features of the convolutional layers and pre-stress values increases significantly, indicating that the 1D-CNN model is able to automatically extract features related to the variation of pre-stress. The pre-stress prediction accuracy using 1D-CNN is significantly higher than that using SVR, and the prediction error is within 3%. The proposed 1D-CNN model exhibits excellent noise-robustness, with the prediction error remaining within 10% even at the SNR level of −5 dB. Even after removing half of conditions in the training set, the proposed 1D-CNN model is still able to achieve accurate identification of effective pre-stress.