Unified Transfer Learning Framework for Structural Health Monitoring of Plate-Like Structures

One of the most promising candidates for a real-time SHM system for thin aircraft structures is the Lamb wave-based inspection method. The Lamb wave-based SHM systems have a great potential as an integrated structural monitoring system thanks to the ongoing machine learning revolution. Despite the fact that the current ML algorithms have demonstrated robust classification capabilities, it is undeniable that the success of such models depends heavily on the input data. The fundamental reason for this is that these ML models are often created to perform standalone tasks, which makes them highly data-dependent. In order to address these problems, the current study suggests a transfer learning strategy. The transfer learning model imports pre-trained layers from earlier ML models to help it learn new data. The proposed unified TL model utilizes pre-trained layers of a ResNet-autoencoder to assist a 1D-CNN classifier in learning from new data. The model is constructed in two phases, where in the first phase, the autoencoder is trained on noisy baseline differential signals from “Open Guided Waves” publicly available benchmark datasets. In the later stage, the trained layers from the autoencoder are then merged with a 1D-CNN classifier to diagnose new data. The proposed model identified both damaged and undamaged cases from the set of 144 unseen samples with an impressive 82% accuracy. The study will assess strategy’s generalization ability in both composite and metallic plate-like structures.