Improving damage detection by combining multiple classifiers in different feature spaces

Abstract

Vibration-based damage identification methods have been explored in several engineering fields. However, the effective damage-sensitivity of vibration-based features and the balance between false alarms and readiness of the anomaly detection are still open issues. Here, a cost-effective damage detection methodology based on a Deterministically Generated Negative Selection Algorithm is implemented and tested over a wide dataset generated through the numerical simulation of a three-storey concrete building in different damage scenarios. The approach allows training the classifiers only based on undamaged samples and ensures a prompt response to damage outbreak by fusing the classification conducted in parallel on different feature spaces, exploiting their distinct sensitivity to damage. A strategy to improve the reliability of the alarm is implemented by introducing counters and thresholds on consecutive classification outputs. A satisfactory trade-off between false and true alarms is achieved by comparing their probabilities, obtained through Markov chains, for different values of such thresholds.