Perpustakaan Digital ITB

Carbon-fiber reinforced polymer (CFRP) composites are extensively used in aerospace for their high specific strength, yet they are susceptible to internal failures such as delamination and cracks. Traditional Structural Health Monitoring (SHM) using ultrasonic Lamb waves often requires extensive baseline signals and dense sensor arrays. This research proposes a baseline-free deep learning framework, specifically a Bayesian Neural Network (BNN), trained on FEM-generated dataset to automate damage localization. The methodology involves using 3D finite element method (FEM) simulations to generate training data by varying damage coordinates on CFRP thin plates. The model was implemented using the Bayes-by-Backprop variational inference framework, in which each network weight is represented by an independent Gaussian variational posterior optimized via the reparameterization trick, trained by minimizing a mini-batch Evidence Lower Bound objective combining mean squared prediction error with a KL-divergence regularization term. Results demonstrate that the model can predict damage locations on a 500×500 mm plate with a mean euclidean distance (MDE) of 66.34 mm and RMSE of 65.168 mm using only one actuator-sensor pair. This performance is 4.25% higher compared to performance of model trained on experiment dataset. This framework eliminates the need for manual feature extraction or healthy reference signals and enables development of wider and more complex damage detection model.