Traditional physical model-based nondestructive evaluation (NDE) and damage detection methods are often unreliable due to the complex dependence of model parameters on minor differences in material properties (e.g., thickness, temperature, or loading effects). While classic data-driven approaches appear to eliminate model complexity, their performance highly depends on feature extraction, for which domain-expertise-based data preprocessing is required. Wavefield analysis is a promising alternative for non-contact NDE but suffers from the problem of slow data acquisition. As a result, effective structural health monitoring (SHM) based on wavefield analysis of guided waves in large-scale systems, such as mechanical, civil, or aerospace structures, has remained challenging. To address these challenges, we present a deep convolutional neural network (DCNN)-based transfer learning approach to interpret ultrasonic guided waves with small training data sets, thereby achieving rapid, effective, and automated SHM. Specifically, the proposed learning framework includes a pre-trained DCNN for automated feature extraction from the raw inputs (i.e., wavelet-transformed time-frequency images) and a fully connected classification stage that is trained with partial wavefield scans. Experiments on full wavefield scans of various thin metal plates demonstrate the effectiveness and efficiency of the proposed approach: >95% classification accuracy is obtained with only 5% training data, thus enabling fast scanning and fully automated damage detection of large-scale structures.
|