Statistical alignment in transfer learning to address the repair problem: An experimental case study

Abstract

Repair is a critical step in maintenance of civil structures to ensure safe operation. However, repair can pose a problem for data-driven approaches of long-term structural health monitoring, because repairs can change the underlying distributions of the data, which can invalidate models trained on pre-repair data. As a result, models previously trained on pre-repair information fail to generalise to post-repair data, reducing their performances and misrepresenting the actual behaviour of structures. This paper suggests a population-based structural health monitoring approach to address the problem of repair in long-term monitoring of a mast structure, by exploring domain adaptation techniques developed for transfer learning. A combined approach of normal condition alignment and Dirichlet process mixture models is adopted here for damage detection, that can operate unimpeded by post-repair shifts in distributions. The method is able correctly identify 99\% of the damage data with a false positive rate of around 1.6%. Moreover, it is able to detect environmental variations such as stiffening due to freezing conditions that can adversely affect the dynamic behaviour of structures.