Robust training databases for supervised learning algorithms in structural health monitoring applications

Abstract

Machine learning algorithms are progressively used in structural health monitoring (SHM) applications. However, damage identification in a supervised learning context is challenging due to insufficient training data for various damage states of the structure. It may be feasible to acquire unhealthy sensor data for low-cost structures, but not for high-value complex structures, such as aircraft. In this work, numerical modelling is employed to simulate hard to attain damage scenarios that are in turn used to create training databases for damage identification through machine learning. In order to take into account modelling uncertainty, a comprehensive set of models representing different damaged states with varying parameters is created. Clustering techniques are then used to explore robust features between the model and the physical structure which are finally used to train artificial neural networks. The framework is validated by an experimental case study of damage detection, in the form of loosened bolts on a laboratory tested structure.