A Multi-Fidelity Convolutional Autoencoder–Transfer Learning Framework for Guided-Wave-Based Damage Diagnosis Using Large Simulated and Limited Experimental Datasets

How can deep learning models for structural health monitoring (SHM) be effectively trained to detect and characterize damage when high-fidelity experimental data is scarce and large-scale simulation is computationally prohibitive? The authors address this by developing a transfer learning framework that leverages lightweight physics-based simulations to pretrain a model, which is then fine-tuned on a small set of experimental measurements.

The researchers utilize a one-dimensional time-domain spectral element (1D-TDSE) model based on zigzag theory to generate a large synthetic dataset of guided wave signals. This model is computationally efficient, allowing for the rapid simulation of piezoelectric transducer interactions and damage scenarios. A convolutional autoencoder (CAE) is first pretrained on this synthetic data to learn a low-dimensional latent representation of the wave signals. A feed-forward neural network (FFNN) is then attached to the encoder to perform damage localization and sizing. Finally, the model undergoes transfer learning, where the encoder is frozen and the decoder and FFNN are fine-tuned using a limited experimental dataset obtained from an aluminum plate with block-mass damage.

The proposed CAE-based transfer learning framework significantly outperforms standard CNN-based approaches in damage localization accuracy. By pretraining on synthetic data and adapting to the experimental domain, the model demonstrates strong generalization capabilities. It successfully predicts damage parameters for scenarios that were not included during the pretraining or fine-tuning phases, achieving R-squared values above 0.93 for localization and 0.99 for sizing. This indicates that the framework is a viable, efficient solution for real-world SHM applications where data acquisition is costly.