ASSESSING THE INFORMATION CONTENT OF DATASETS FOR STRUCTURAL HEALTH MONITORING

Data-based SHM is highly dependent on the quality of the training data needed for machine learning algorithms. In many cases of engineering interest, data can be scarce, and this is a problem. However, in some cases, data are abundant and can create a computational burden. In data-rich situations, it is often desirable to select the subset(s) of the data which are of highest value (in some sense) for the problem of interest. In this paper, ‘value’ is interpreted in terms of information content, and entropy is used a measure of that content in order to condense training data without compromising useful information. Using the minimum covariance determinant, the dataset is first separated using inclusive outlier analysis. The entropies of the separated datasets are then assessed using parametric and nonparametric density estimators to identify the subset of data carrying most information. The Z24-Bridge dataset is used here to illustrate the idea, where the entropy values indicate that the subset containing data from environmental variations and damage is most rich in information. This subset was made up of half of the entire dataset, suggesting that it is possible to significantly reduce the amount of training data for an SHM algorithm whilst retaining the required information for analysis.