Neighborhood component analysis to leverage the class label information during feature selection to enhance the damage classification performance

Abstract

In the health monitoring of large structures data from multiple sensors are fused to get reliable and robust results. Using data from multiple sensors leads to high dimensional feature vectors which result in computational inefficiency and overfitting. Therefore, it is necessary to reduce the dimension of these feature vectors, especially in the case of real-time health monitoring systems. Principal component analysis has become a go-to algorithm for dimensionality reduction in the structural health monitoring literature. However, it reduces the dimension by maximum variance criterion without considering the class labels of these initial feature vectors, this implies that the reduced dimensional feature vectors do not guarantee best class separability in damage classification problems. Using linear discriminant analysis to solve this issue poses singularity problems in the damage classification problems with small datasets having high dimensional feature vectors. This paper proposes the use of neighborhood component analysis for reducing the dimensionality of the feature vectors. As in the case of damage type identification problems the input signals have to be labeled, neighborhood component analysis makes use of the class label information and reduces the dimension by considering the classification performance. The algorithm’s performance has been validated by solving a damage classification problem by combining it with a K-nearest neighbor classifier. An experimental benchmark dataset from the American Society of Civil Engineers is utilized for this purpose. A comparison with the Principal Component Analysis on the same dataset depicts that the ability of Neighborhood Component Analysis to include the class label information indeed helps in enhancing the damage classification performance.