Suggested Citation:
Lopez, Israel and Nesrin Sarigul-Klijn (2009) A Novel Dimensional Reduction Approach for Structural Damage Diagnosis using Feature Similarity . Proceedings of the SPIE 7295

Dimensionality reduction is an essential data preprocessing technique for feature extraction, clustering and data classification in the area of Structural Health Monitoring (SHM). This paper presents a novel data-driven model for feature extraction and its application to damage identification by means of experimental case studies. The method obtains similarity matrix indices for individual dimensional reduction techniques whereby maximum compression of information is obtained and redundancy therein is removed by creating an ensemble of these indices. A systematic comparison of this novel technique to existing linear and nonlinear dimensional reduction methods is given. First case study investigates the aeroacoustic properties of a scaled wing model with penetrating impact damage. In the experimental vibration case study, we use the response of surface mounted accelerometers to detect and quantify damage of an aluminum plate. The dimensional reduction methods will be used to improve the efficiency and effectiveness of damage classifier. In this study, damage identification performances are evaluated using a one-class k-Nearest Neighbor classifier. Classification performance is measured in terms of rate of detection and false alarm via receiver operating characteristic (ROC) curves. The robustness of the damage detection approach to uncertainty in the input data is investigated using probabilistic-based confidence bounds of prediction accuracy. Experimental results show that proposed approach yields significant reduction of false-diagnosis and increasing confidence levels in damage detection.