Engineers monitor structures with many sensors. Changes in sensor readings do not always mean damage. They can come from changing environmental or operational variables, like temperature, wind, or traffic. This paper shows that those variables can change not only average sensor values but also the variances and the correlations between sensors. The authors present methods to estimate and remove those multivariate effects so that monitoring focuses more on true structural changes.

The central idea is to estimate the conditional covariance matrix. That is the matrix of variances and covariances of the sensor outputs given the measured environmental and operational variables. The paper discusses four ways to get that matrix. One is a nonparametric, kernel-based estimator that averages nearby observations in the space of environmental measurements. The kernel method uses a bandwidth to set how “nearby” is defined and selects it by validation. The authors note this approach weakens as the number of covariates grows, a problem known as the curse of dimensionality.

A second approach adapts random forests. A random forest builds many decision trees to split the data into subgroups. In each subgroup the sample covariance can be estimated. By combining many trees, the method aims to find regions of the covariate space that have distinct covariance patterns. The paper also proposes a semiparametric additive model and a deep learning approach as alternative ways to model how covariates affect covariances.

The estimated conditional covariance matrices are then used in standard structural health monitoring steps. The authors describe how to create covariate-adjusted versions of principal component analysis and of the Mahalanobis distance, which are common tools for detecting anomalies. They test and compare the four approaches on simulated data and on two real datasets: load-test data from the Vahrendorfer Stadtweg bridge in Hamburg, Germany, and modal (eigenfrequency) data from the KW51 railway bridge near Leuven, Belgium.