Chapter 17 On the Application of Domain Adaptation in SHM X. Liu and K. Worden Abstract Machine learning has been widely and successfully used in many Structural Health Monitoring (SHM) applications. However, many machine learning models can only make accurate predictions when the training and test data are measured from the same system; this is because most traditional machine learning methods assume that all the data are drawn from the same distribution. Therefore, to train a robust predictor, it is often required to recollect and label new training data every time when considering a new structure, which can be significantly expensive, and sometimes impossible in the SHM context. In such cases, the idea of transfer learning may be employed, which aims to transfer knowledge between task domains to improve learners. In this paper, a subfield of transfer learning i.e. domain adaptation, is considered, and its utility in SHM applications is briefly investigated. Keywords Domain adaptation · Transfer learning · Structural health monitoring (SHM) 17.1 Introduction Machine learning technologies have already achieved significant success in many Structural Health Monitoring (SHM) applications, e.g. [1–5]. Most of these machine learning methods can perform well if the training and test data are drawn from the same distribution, i.e. all the data are measured from the same structure. However, in real-world applications, there are some cases where the training and the test distributions differ, often because damage state data are only available in one case. For example, in the case of a wind farm, suppose there is a sufficient amount of labelled feature data for a specific wind turbine about its normal and damage states (called the source domaindata) and a large number of unlabelled data of another wind turbine of the same model (called the target domain data). As there is a lack of labelled data for the target structure, and supervised learning is thus impossible, one may wish to train a model using the source domain data. However, due to the existence of the inevitable manufacturing and assembling errors for each individual, the data distributions of the two wind turbines may be different, and then the predictions of models trained using traditional machine learning methods can be significantly degraded. The direct solution is to collect and label new training data and rebuild models for each individual, which is expensive or sometimes impossible to implement. In such cases, it is desirable that the knowledge learnt from the related labelled source domain data can be transferred to the target domain. It will be clear that the context for the discussion here is population-based SHM [6], i.e. it is concerned with allowing inferences across populations of structures (even if the number in the population is only two). Furthermore, this paper is concerned with supervised learning; so the required diagnostics are at the higher levels in Rytter’s hierarchy [1] and thus directed at inferring damage location, type or severity. Transfer learning is a research subfield of machine learning which aims to improve a learner from one domain by transferring knowledge from a related domain [7]. Domain adaptionis one of the branches of transfer learning. The focus of domain adaptation is that the distribution across the source and target domain are different, which matches the issue of the aforementioned damage diagnosis situation, and indicates why the technique is potentially relevant for SHM. According to [8], the source and target domains of a transfer learning problem can be different in the input feature spaces (Xs =Xt ), the label spaces (Ys =Yt ), the marginal probability distributions of the input data (P(Xs) =P(Xt )) or conditional distributions (P(Ys|Xs) =P(Yt |Xt )). X. Liu · K. Worden ( ) Dynamics Research Group, Department of Mechanical Engineering, University of Sheffield, Sheffield, UK e-mail: k.worden@sheffield.ac.uk © Society for Experimental Mechanics, Inc. 2020 N. Dervilis (ed.), Special Topics in Structural Dynamics & Experimental Techniques, Volume 5, Conference Proceedings of the Society for Experimental Mechanics Series, https://doi.org/10.1007/978-3-030-12243-0_17 111
RkJQdWJsaXNoZXIy MTMzNzEzMQ==