Abstract Summary
Vibration-based Structural Health Monitoring (SHM), exploits a variety of approaches for novelty detection. In particular, some data-based methods try to recognise patterns by exploiting analogies with the human body’s natural defences at a cellular level. These algorithms often require the use of a large variety of data correlated to different environmental and operation conditions to ensure performance robustness and to aid interpretation of results. Regardless, the scarcity of such data often limits the extent of their applicability. In the framework presented here, a possible solution is provided by a novel approach based on modelling and sharing already acquired knowledge between sufficiently similar structures, i.e., population-based structural health monitoring (PBSHM). This study investigates the process of damage detection in a group of three different structures, obtained by applying structural modifications to a small-scale glider model, which follows the characteristic geometry of the GARTEUR benchmark project. Damage identification is performed by exploiting the Negative Selection Algorithm, (NSA), already applied by some of the Authors on numerically simulated case studies, and chosen for its capability of self/non-self discrimination under varying operational or environmental conditions. The research is expanded by using sparse autoencoders for feature dimensionality reduction. The method is applied to three experimental datasets acquired by laser vibrometer measurements, to identify consistent damage-sensitive features from the frequency response functions, and to obtain a reliable fault-detection performance.
