Physics-Informed Feature Selection for CFRP Damage Assessment Using ECS and GPR

Abstract

Aims: To develop a physics-based feature selection algorithm to be used in detecting structural damage of CFRP with ECS capacitance data, and to perform a statistical analysis of a Gaussian Process Regression (GPR) model with uncertainty quantification to be tested on a small sample scale. Dataset: The ECS-CFRP Altabey (2022) dataset (DOI: 10.17632/c9v4zy3555b.1) of 11 progressive fatigue damage conditions (D0-D10) consisting of 78 raw capacitance data (12 self-capacitance and 66 mutual-capacitance) of a CFRP composite pipeline system was used. Methodology: The research utilized a structured approach to feature selection and validation, including Principal Component Analysis for dimensionality reduction, Random Forest feature selection with cross-validation, and physics-based validation of angular span effects. Gaussian Process Regression was utilized with a Matern 3/2 kernel, and Leave-One-Out Cross-Validation was employed. Results: The results demonstrate that a reduced, geometry-independent representation, based upon delta capacitance, was found to be more accurate than the entire feature set, with higher values of R² (0.986 vs. 0.974) and lower values of RMSE (0.380 vs. 0.508). The response was also shown to increase with angular span, as expected physically. Conclusion: A physically interpretable, small five-feature ECS subset, provides better damage-level prediction accuracy with the uncertainty calibrated. The delta-capacitance formulation allows a cross-configuration transferability, which has been a major gap in ECS-based SHM, which has been based on ad-hoc rules of empirical thresholds without interpretable ML explanation.