A Novel Feature Selection Framework to Deal with Outliers in SHM Problems

Predictive modeling becomes increasingly reliant on feature selection algorithms as dataset dimensions grow. Feature selection methods can estimate biased feature importance values when datasets present outliers. This can compromise the effectiveness of classification algorithms by reducing their accuracy. This paper primarily targets feature selection for solving damage classification problems in civil and mechanical engineering. We propose a novel framework for feature selection based on the minimum covariance determinant and PCA. The proposed feature selection algorithm obtains a metric for scoring features’ importance based on the loading matrix obtained from the robust PCA algorithm applied to the feature matrix. The covariance matrix of the robust PCA algorithm is obtained from the minimum covariance determinant algorithm. This way, features considered outliers in the feature matrix are discarded from further analysis. The proposed feature extraction framework is tested on several damage classification problems of wood materials. Its superiority is demonstrated by comparing its results with a PCA-based feature selection algorithm. The results obtained from the proposed unsupervised feature selection method demonstrate its robustness to outliers, rendering it a viable application technique in complex problems involving datasets containing outliers.

• File upload: SHM_ES_2023__15_April_-2657da95