Standard machine learning algorithms traditionally address isolated tasks and require customization every time for a new kind of dataset. Thus, such a machine learning-based structural health monitoring approach lacks flexibility for any alteration in real-time field application. Most of the published articles have neither demonstrated nor discuss such capabilities. The present study attempts to overcome this by developing an Autoencoder based two-step diagnostic approach; In the first step, a 1D-CNN stacked auto-encoder (1D-CAE) is developed and trained to learn and extract the dominating features present in the Lamb waves. In the second step, the classification module is constructed, which inputs scaled features from the encoder and performs the classification between pristine and damaged conditions. The approach requires relatively minor training datasets than the conventional deep learning counterparts. The 1D-CAE is capable of auto-tuning, offers flexibility with different sizes of the 1D-input dataset. The study utilizes publicly available benchmark datasets collected by “Open guided waves”, which comprises Lamb wave signals recorded for a range of frequencies and artificial defects. The setup consists of arrays of transducers bonded to the 500mm × 500mm and 2mm thin CFRP plate in “Pitch-Catch” configuration, which excites five cycles modulated Hann-filtered sine wave signal with an amplitude of ±100 V. Both the 1D-CAE and the classifier module is trained under the supervision of Adam optimization algorithm with varying learning rate. The fully trained classifier can detect the damage in the thin CFRP plate with 95% accuracy. In the later stage, the performance is evaluated against the unseen samples with defects generated from the experimental setup. The proposed algorithm achieves a high level of generalization.
|