ON NORMALISATION FOR DOMAIN ADAPTATION IN POPULATION-BASED STRUCTURAL HEALTH MONITORING

Abstract

The practical application of structural health monitoring (SHM) is often limited by the unavailability of labelled data. Transfer learning - specifically in the form of domain adaptation – gives rise to the possibility of leveraging information from a population of physical or numerical structures, by inferring a mapping that aligns the feature spaces. There are a number of approaches to domain adaptation that minimise some distribution discrepancy metric. However, it is found that under high initial distribution discrepancy, these methods may be prone to performance degradation. In this paper, guided normalisation is proposed as a solution to the initial distribution discrepancy problem. Several case studies demonstrate how normalisation can itself perform powerful adaptation and facilitate further adaptation, with more sophisticated domain adaptation methods.