Damage Evaluation in Inconel 718 Alloy by Deep Learning-Assisted Nonlinear Ultrasonic Guided Wave Technique

This article proposes a novel damage evaluation method for Inconel 718 alloys by integrating nonlinear ultrasonic guided wave technique with deep learning (DL). Experimental findings demonstrate that quasi-static components (QSCs) exhibit significantly lower attenuation during propagation within Inconel 718 alloys compared to ultrasonic waves at other frequencies, enabling their potential use in long-distance detection. As a nonlinear response, QSC has also proven effective in accurately capturing variations in material properties, such as thermal damage and stress levels. However, the concurrent presence of thermal damage and stress complicates their mapping relationship with the QSC, making it challenging to invert these damages solely based on QSC amplitude directly. Additionally, thermal damage induces grain coarsening in Inconel 718 alloys, leading to a marked increase in ultrasonic attenuation. To address this challenge, a 2-D convolutional neural network (2D-CNN) model was established to decouple the complex mapping relationships. To enhance the extraction of damage-related features from QSC evaluation signals, the signals were transformed into 2-D images by short-time Fourier transform (STFT) and Gramian angular field (GAF), respectively. Prediction results under controlled laboratory conditions show promising performance of the proposed approach. Notably, STFT outperformed GAF in feature extraction from QSC evaluation signals in both the time and frequency domains. Compared to the predictions obtained from training GAF images, the mean absolute error (MAE) and root mean square error (RMSE) derived from training STFT images were reduced by up to 47.47% and 29.32%, respectively.