Open Access
Subscription or Fee Access
Assessing the Likelihood of Damage at the Start of a Structural Health Monitoring Campaign
Abstract
In many structures that engineers are interested in monitoring the damage state of, the infeasibility of collecting data in each possible damage condition makes unsupervised learning algorithms an appealing approach. However, this learning method generally relies on access to an initial data set being representative of the undamaged behaviour of the structure, which often is not available, particularly when a monitoring system has been retrofitted to an operational structure. To overcome this issue, arbitrarily selecting data over an initial period of monitoring to represent the undamaged/normal condition is a common method to adopt, however, there are a variety of limitations with this approach, such as uncertainty in the damage condition of the structure prior to the installation of the monitoring system, in addition to the length of the required training period. It is therefore desirable to be able to infer a structure’s damage condition at the start of a monitoring period, explored here through the application of the Dirichlet Process Gaussian Mixture Model, a clustering method based on a Bayesian adaptation of the traditional Gaussian Mixture Model. In the context of condition assessment for wind turbine components, this paper combines the application of prior knowledge and population-based SHM in conjunction with this clustering method to provide a framework to assess the likelihood of damage at the start of a monitoring campaign.
DOI
10.12783/shm2019/32372
10.12783/shm2019/32372