An Effective Damage Identification Method Combining Double-Window Principal Component Analysis with AutoGluon

Abstract

In recent years, Double Window Principal Component Analysis (DWPCA) has been proposed. The spatial windows exclude damage-insensitive data from the analysis, while the temporal window improves the discrimination between healthy and damaged states. As a result, the DWPCA method exhibits higher sensitivity and resolution in damage identification compared to traditional PCA methods, as well as other traditional signal processing methods such as wavelet analysis. However, existing research on DWPCA has mainly focused on using the first-order eigenvector for damage identification, while the potential of higher order DWPCA eigenvectors remains unexplored. Therefore, the objective of this paper is to investigate the damage identification capabilities of higher-order DWPCA eigenvectors. Furthermore, we propose three types of damage-sensitive features based on DWPCA eigenvectors and use them as inputs to artificial intelligence (AI) algorithms for damage localization and quantification. The AI algorithms considered include AutoGluon and Transformer, which are powerful machine learning (ML) and deep learning (DL) algorithms proposed in recent years, respectively. In addition, classical ML algorithms such as Decision Tree (DT), Random Forest (RF) and Extreme Gradient Boost (XGBoost) are considered for comparison. Extensive benchmark experiments are performed and the numerical results obtained show that the combination of AutoGluon with DWPCA features achieves remarkable performance in terms of damage localization and quantification. This performance exceeds that of DT, RF, XGBoost and Transformer algorithms. Specifically, the prediction accuracies for damage localization and quantification exceed 90%. These results highlight the great potential of integrating AutoGluon with DWPCA features, particularly by combining AutoGluon with the first and second DWPCA eigenvectors, for real-world applications in structural health monitoring.

Keywords

• structural health monitoring
• damage detection
• machine learning algorithm
• principal component analysis