Improving Damage Identification Reliability by Combining Classification on Distinct Feature Spaces

In the last decades, vibration-based dynamic identification and modal properties tracking have become a common strategy for damage detection purposes. Nonetheless, assessing the unbiased sensitivity to damage of the modal properties and their derivatives requires further study in order to reduce the number of false negatives and ensure a reliable recognition of the anomaly onset, location and extent. To this end, in the present work, a recently developed version of a Deterministically Generated Negative Selection Algorithm (DGNSA) is used to perform damage detection considering several pairwise combinations of dynamic and environmental features. The capability of correctly identifying an anomaly due to an unknown emerging abnormal behaviour is improved by exploiting the distinct performance, against a set of possible damages, of the classifiers trained in different two-dimensional feature spaces. To ensure a full control over all the parameters that might affect the structural response and the performance of the investigated algorithm, an artificial dataset is generated by simulating the response of a three-storey concrete frame building in different scenarios. The analyses confirm that the identification of the monitored features is a fundamental step for a correct implementation of a damage identification strategy. They also show that the distinct sensitivity of the classifiers can be exploited to improve the reliability and responsiveness of the DGNSA against many possible unknown damage scenarios.