Deep transfer learning for delamination damage in CFRP composite materials

Abstract

This article presents the health monitoring of carbon fiber-reinforced plastic (CFRP) structures using a data-driven deep transfer learning approach to facilitate mapping signal features to damage categories. Simulations were conducted on composite material specimens with delamination damage, with validation performed using laboratory-derived CFRP damage experimental data. Continuous wavelet transform was employed to process Lamb wave signals recorded from a specified sensor network on the composite material panel, extracting time–frequency scale representations. A cross-workpiece deep transfer learning (CWTL) model was proposed to address the interdependence of the machine learning (ML) model on a large set of labeled damage data for different composite material structures. The CWTL, by seeking an appropriate initial range with minimal data, alters the direction of gradient descent, thereby identifying initial parameters more sensitive to the task. This process allows the ML model to fit to a limited damage dataset quickly. To assess the robustness of this method, considering environmental variability as well as damage localization and quantification, further extensions of the study were explored. The results demonstrate the efficacy of CWTL in accurately classifying both undamaged and delaminated damage categories, with high accuracy, suggesting the ML model’s potential for practical applications in such structural frameworks.