Damage diagnosis under environmental and operational variations using unsupervised support vector machine

The goal of structural health monitoring is to provide reliable information regarding damage states that include damage presence, location, and severity. Damage diagnosis is performed by utilizing measurements that are obtained from a structure being monitored. However, time-varying environmental and operational conditions such as temperature and external loading may produce an adverse effect on damage detection within the structure exposed to these changes. Therefore, in order to achieve successful structural health monitoring goals, it is necessary to develop data normalization techniques which distinguish the effects of damage from those caused by environmental and operational variations. In this study, nonlinear principal component analysis based on unsupervised support vector machine is introduced and incorporated with a discrete-time prediction model and a hypothesis test for data normalization. The proposed nonlinear principal component analysis characterizes the nonlinear relationship between extracted damage-sensitive features and unmeasured environmental and operational parameters by employing kernel functions and by solving a simple eigenvalue problem. The performance of the proposed method is compared with that of another nonlinear principal component analysis realized by auto-associative neural network. It is demonstrated that the proposed method is a promising data normalization tool that is capable of detecting damage in the presence of environmental and operational variations.