A deep learning model for stress estimation from full time-domain ultrasonic signals

Residual stress is a critical factor which influences the service life and structural integrity of many metallic components. To address the limitations of conventional ultrasonic stress measurement methods, this study proposes a uniaxial stress estimation approach based on a hybrid deep learning architecture using full time-domain ultrasonic waveforms. The proposed 1D CNN-TGA model integrates a one-dimensional convolutional neural network (1D CNN), a temporal convolutional network (TCN), a gated recurrent unit (GRU) and an attention mechanism to capture both local waveform characteristics and long-range temporal dependencies. Ablation experiments demonstrate that the proposed 1D CNN-TGA hybrid model achieves high accuracy in ultrasonic stress estimation. Grad-CAM is used to visualise the temporal regions of the ultrasonic signal that contributed most to the model’s stress estimation. Grad-CAM visualisation reveals that the model mainly focuses on signal regions containing amplitude-energy variations and time-shift characteristics, providing physical interpretability for the learned features. Experimental results show that the model achieves high estimation accuracy for stress estimation with R2 = 0.9995, MAE = 1.3098 MPa and RMSE = 1.5157 MPa. Compared with traditional time-of-flight (TOF) methods, the proposed approach demonstrates improved robustness and generalisation for ultrasonic stress estimation.