Automatic selection of optimal structures for population-based structural health monitoring

For structural health monitoring based on machine learning techniques, population-based structural health monitoring (PBSHM) has been developed to address problems caused by environmental variations or issues caused by data scarcity. PBSHM utilises information from one set of structures in a population to predict the responses of another set or to transfer knowledge between them. However, how to choose a reference set of structures is still a problem to be solved. This paper introduces an unsupervised strategy to rank the candidate structure sets based on canonical correlation coefficients and principal component analysis. Furthermore, depending on the characteristics of the data from these structures, two different fast selection approaches based on a greedy search can be employed. The supervisory control and data acquisition (SCADA) data collected in the Lillgrund offshore wind farm will be used to demonstrate the proposed strategy. The information of the selected subset of wind turbines (WTs) is used to train a Gaussian process-based model to predict the power outputs across the entire wind farm. The prediction results corresponding to the proposed method are compared with those corresponding to two other methods—a supervised selection method and a random selection method. The results show that the proposed technique can help find effective structure sets for PBSHM efficiently and automatically, which is beneficial to establishing an online monitoring system.