Confounder-adjusted Covariances of System Outputs and Applications to Structural Health Monitoring

Many existing methods for correcting confounding effects are based on different types of mean regression.

“Many existing methods for correcting confounding effects are based on different types of mean regression.”

• cites Worden and Cross, 2018 sourceSupported
• cites Maes et al., 2022 sourceNot checked — no full text of the source

The scores of the second to fourth components closely follow a normal distribution.

“The bottom subplot clarifies that the scores of the second to fourth components closely follow a normal distribution.”

The optimal bandwidth varies depending on whether acceleration, strain, or inclination data is examined.

“The optimal bandwidth varies depending on whether acceleration, strain, or inclination data is examined; see Table 2.”

The changes in the correlation functions are more pronounced for acceleration and strain measurements than for inclination measurements.

“The changes in the correlation functions are more pronounced for acceleration and strain measurements than for inclination measurements”

Almost all correlation functions are affected by temperature changes.

“The most important finding is that almost all correlation functions are affected by temperature changes, which justifies the methodological contributions in this paper.”

The scores deviate substantially from the normal distribution for the first four components.

“It can be observed that the scores deviate substantially from the normal distribution for the first four components, meaning the first four scores are affected by temperature changes.”

The conditional PCA shows that both the direction and the variance of the principal components change with changing temperature.

“The conditional PCA shows that both the direction and the variance of the principal components change with changing temperature.”

The first eigenvalue is much larger for standard PCA than for conditional PCA.

“The finding that, in particular, the first eigenvalue is much larger for standard PCA than for conditional PCA can be explained by the fact that conditional PCA removes the temperature-induced variation from the data by using the conditional mean instead of the marginal one.”

The standard scores of the first principal component exhibit a strong and almost linear association with temperature.

“The top subplot clarifies that the standard scores of the first principal component exhibit a strong and almost linear association with temperature.”

Normality cannot be assumed for the standard scores.

“Moreover, normality cannot be assumed.”

The first component exhibits a bimodal distribution.

“Only the first component exhibits a bimodal distribution, but this may be due to a second, unobserved confounder.”

The results for the variance function are similar with distinct minima at h = 1.5 and h = 1 depending on the sample size.

“Ultimately, the results for the variance function are similar with distinct minima at h = 1.5 and h = 1 depending on the sample size.”

A chosen bandwidth that is too large results in substantial bias if the true correlation indeed depends on the temperature ('over-smoothing').

“On the other hand, a chosen bandwidth that is too large results in substantial bias if the true correlation indeed depends on the temperature ('over-smoothing').”

The larger the bandwidth, the more the estimated correlation function is pushed toward a constant value, in particular for the extreme case with h → ∞.

“In other words, the larger the bandwidth, the more the estimated correlation function is pushed toward a constant value, in particular for the extreme case with h → ∞.”

The estimates of the target function become smoother for higher bandwidths and larger sample sizes.

“On one hand, the estimates of the target function become smoother for higher bandwidths and larger sample sizes.”

For conventional PCA, temperature changes affect the distribution of the first three to four scores.

“So, temperature changes affect not only the distribution of the first scores but also the first three to four scores.”

• Also stated in: § Standard Deviation Correlation · plain

  “In summary, for conventional PCA, the distribution of the first three to four scores is affected by temperature changes.”

The mean of the conditional scores is zero across the entire temperature range for all four components.

“In contrast to the standard scores, the mean of the conditional scores is zero across the entire temperature range for all four components (top plot), indicating no association between temperature and the scores.”

For low temperatures, the main directions of the variation in the strain measurements are substantially different from those obtained from standard PCA, ignoring potential temperature effects on covariances.

“In other words, for low temperatures, the main directions of the variation in the strain measurements are substantially different from those we would (wrongly) obtain from standard PCA, ignoring potential temperature effects on covariances.”

The output covariances depend on temperature readings.

“The following results clearly show that the output covariances depend on temperature readings.”

For the logistic correlation function, the curves in Figure 4 have well-defined minima as well.

“For the logistic correlation function, the curves in Figure 4 have well-defined minima as well.”

For a constant correlation function, the optimum is achieved for a very large bandwidth.

“On close inspection, it can be seen that, for a constant correlation function, the optimum is achieved for a very large bandwidth.”

The proposed non-parametric approach can identify confounder effects on conditional covariances.

“The two main takeaways are that the proposed non-parametric approach can identify confounder effects on conditional covariances”

Particularly for the subzero temperature value of -1°C, the conditional principal components are distinctly different from the standard ones.

“Particularly for the subzero temperature value of -1°C, the conditional principal components are distinctly different from the standard ones (as seen from the top left subplots of Figure 7).”

The second to fourth components also show a clear and non-linear relationship with temperature.

“However, the second to fourth components also show a clear and non-linear relationship with temperature.”

The developed method for confounder removal leads to system outputs that approximate normal distributions more closely, provided all confounding variables are measured.

“Furthermore, the developed method for confounder removal leads to system outputs that approximate normal distributions more closely, provided all confounding variables are measured.”

A numerical and easily reproducible case study demonstrated that the underlying covariance function (the ground truth) can be retrieved correctly.

“Moreover, a numerical and easily reproducible case study was presented (a) to demonstrate that the underlying covariance function (the ground truth) can be retrieved correctly and (b) to validate the optimization scheme for the bandwidth.”

For the entire Phase I set combined, the false alarm rates are 6.3%, 4.9%, and 4%, respectively.

“For the entire Phase I set combined, the false alarm rates are 6.3%, 4.9%, and 4%, respectively.”

• AssumptionRests on an assumption: The first 200 days before retrofitting are in-control (undamaged).

The temperature-induced artifacts around observation number 1,500 are removed using the new conditional approach.

“At the same time, it is worth pointing out again that the temperature-induced artifacts (i.e., implausibly large MD values) around observation number 1,500 are removed using the new conditional approach.”

• AssumptionRests on an assumption: The first 200 days before retrofitting are in-control (undamaged).

The (co-)variances can change substantially for varying temperatures, in particular for subzero temperatures.

“Based on two well-established benchmark data sets from real bridges, it was demonstrated that the (co-)variances can change substantially for varying temperatures, in particular for subzero temperatures.”

The classical covariances are substantially larger compared to the conditional approach for warmer temperatures.

“This is because the classical covariances are substantially larger compared to the conditional approach for warmer temperatures, which makes the test less sensitive.”

Existing approaches for the removal of confounding effects (e.g., response surface modeling) may lead to a considerable number of false alarms, particularly for low temperatures in the case of the KW51 bridge.

“By applying existing approaches for the removal of confounding effects (e.g., response surface modeling) to the bridge case studies, it could be demonstrated that those methods may lead to a considerable number of false alarms-particularly for low temperatures in the case of the KW51 bridge.”

• AssumptionRests on an assumption: The first 200 days before retrofitting are in-control (undamaged).

For temperatures above 2°C, the false alarm rates are 1.6%, 1.5%, and 3.1%, respectively.

“For temperatures above 2 • C, the false alarm rates are 1.6%, 1.5%, and 3.1%, respectively, meaning the conditional MD leads to a larger number of false alarms.”

• AssumptionRests on an assumption: The first 200 days before retrofitting are in-control (undamaged).

Mean-regression-based correction neglects potential changes in higher-order statistical moments.

“This neglects potential changes in higher-order statistical moments”

Civil engineering structures such as bridges are exposed to changing environmental and operational conditions, causing most system response quantities to fluctuate even in the absence of damage.

“Civil engineering structures, such as bridges, are exposed to changing environmental and operational conditions, causing most system response quantities to fluctuate even in the absence of damage, e.g., due to changes in temperature, humidity, traffic, wind, and solar radiation.”

The assumption that covariances do not change throughout training, validation, and testing is often unjustified in civil engineering.

“However, this article will demonstrate that this assumption is often unjustified in civil engineering.”

Dependencies of system response quantities on environmental and operational conditions are inevitable for static response measurements such as strain and inclination, and for dynamic ones such as acceleration and natural frequencies.

“These dependencies are inevitable for static response measurements such as strain (Kromanis and Kripakaran, 2016;Liang et al., 2009) and inclination (Han et al., 2021;Ju et al., 2023;Lee et al., 2019), and dynamic ones (e.g., acceleration, natural frequencies).”

• cites Kromanis and Kripakaran, 2016 sourceNot checked — no full text of the source
• cites Han et al., 2021 sourceNot checked — no full text of the source

Output covariances are essential for generating reliable diagnostics for damage detection.

“the output covariances are essential for generating reliable diagnostics for damage detection.”

• cites Döhler et al., 2014 sourceNot checked — no full text of the source
• cites Montgomery, 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Rejecting the first conditional score and reconstructing the system outputs based on the remaining conditional scores would have been sufficient to remove the effect of temperature changes on the mean and the covariance of the outputs.

“Consequently, rejecting the first conditional score and reconstructing the system outputs based on the remaining conditional scores would have been sufficient to remove the effect of temperature changes on the mean and the covariance of the outputs.”

The results appeared robust against the exact choice of the smoothing parameters.

“However, the results appeared robust against the exact choice of the smoothing parameters, similar to the trends in Figure 2 and 3.”

If a damage identification tool uses only mean regression and residualization to account for environmental effects but later uses covariances of system outputs, an alarm might be issued due to changes in the covariance caused by the confounder variable.

“Consequently, if a damage identifica-tion tool only uses mean regression and residualization to account for environmental effects but covariances of system outputs at some later point along the diagnostic chain, an alarm might be issued due to changes in the covariance caused by the confounder variable.”

In the case of the Munich Test Bridge, conditional covariances are particularly important for acceleration and strain sensors.

“meaning that, in the case of the Munich Test Bridge, conditional covariances are particularly important for acceleration and strain sensors.”

The main contribution of the paper is to show how conditional covariances of system outputs can be estimated using a kernel-based smoothing operator.

“The main contribution of this paper is to show how conditional covariances of system outputs can be estimated in this case using a kernel-based smoothing operator.”

The conditional covariance approach inevitably affects the performance of any covariance-based damage detection algorithm.

“Obviously, the conditional covariance approach is more accurate in capturing the uncertainties in the system outputs at varying temperature values, which inevitably affects the performance of any covariance-based damage detection algorithm, but more about this is given in the next section.”

Conditional PCA removes temperature-induced variation from the data by using the conditional mean instead of the marginal one.

“The finding that, in particular, the first eigenvalue is much larger for standard PCA than for conditional PCA can be explained by the fact that conditional PCA removes the temperature-induced variation from the data by using the conditional mean instead of the marginal one.”

Mean regression and the subtraction method may miss effects on higher-order moments, such as covariances.

“The problem with mean regression and the subtraction method is that it may miss effects on higher-order moments, such as covariances.”

The estimate is biased if the true correlation is not constant.

“Consequently, the estimate is biased if the true correlation is not constant.”

The conditional covariance matrix can be used to adjust the Mahalanobis distance and principal component analysis for confounding effects.

“The conditional covariance matrix presented above can be used to adjust the Mahalanobis distance (MD) and principal component analysis (PCA) for confounding effects.”

If the standard version were used, the fact that principal components change with temperature would be disregarded.

“If the standard version were used, this fact would be disregarded.”

Monitoring systems often use higher-order statistical moments such as covariances for uncertainty quantification and damage detection.

“and thus, monitoring systems often use higher-order statistical moments such as covariances for uncertainty quantification and damage detection (Döhler et al., 2014;Montgomery, 2007).”

• cites Döhler et al., 2014 sourceNot checked — no full text of the source
• cites Montgomery, 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

The first four scores are affected by temperature changes.

“It can be observed that the scores deviate substantially from the normal distribution for the first four components, meaning the first four scores are affected by temperature changes.”

The conditional covariance approach is more accurate in capturing the uncertainties in the system outputs at varying temperature values.

“Obviously, the conditional covariance approach is more accurate in capturing the uncertainties in the system outputs at varying temperature values, which inevitably affects the performance of any covariance-based damage detection algorithm, but more about this is given in the next section.”

There is no association between temperature and the conditional scores.

“In contrast to the standard scores, the mean of the conditional scores is zero across the entire temperature range for all four components (top plot), indicating no association between temperature and the scores.”

The principal components change substantially due to changes in the covariances, particularly at lower temperatures.

“In summary, the principal components change substantially due to changes in the covariances, particularly at lower temperatures.”

Uncertainty quantification, meaning the estimation and analysis of measurement errors and variability, is equally important as evaluating mean values.

“Experience with large engineering systems, however, has shown that uncertainty quantification (the estimation and analysis of the measurement errors and variability) is equally important as evaluating the mean values”

• cites Döhler et al., 2014 sourceNot checked — no full text of the source
• cites Montgomery, 2007Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Only with the newly developed conditional PCA is it possible to fully remove the effect of temperature changes on the mean and the (co-)variance of the scores.

“Only with the help of the newly developed conditional PCA, it is possible to fully remove the effect of temperature changes on the mean and the (co-)variance of the scores.”

The (co-)variances and correlations cannot be accounted for by standard response surface modeling, such as the often-used (bi-)linear approach or any other method for mean regression.

“The latter is impossible to account for by standard response surface modeling, such as the often-used (bi-)linear approach or any other method for mean regression.”

The developed method can be universally applied to static response measurements (strains, inclinations), dynamic response measurements (acceleration), or damage-sensitive features that are extracted from measurements (natural frequencies).

“As shown in the paper, the developed method can be universally applied to static response measurements (strains, inclinations), dynamic response measurements (acceleration), or damage-sensitive features that are extracted from measurements (natural frequencies).”

The larger classical covariances for warmer temperatures make the test less sensitive.

“This is because the classical covariances are substantially larger compared to the conditional approach for warmer temperatures, which makes the test less sensitive.”

This was demonstrated using data from the Munich Test Bridge and the KW51 railway bridge in Leuven.

“This was demonstrated using data from the Munich Test Bridge and the KW51 railway bridge in Leuven.”

The conditional MD leads to a larger number of false alarms for temperatures above 2°C.

“For temperatures above 2 • C, the false alarm rates are 1.6%, 1.5%, and 3.1%, respectively, meaning the conditional MD leads to a larger number of false alarms.”

• AssumptionRests on an assumption: The first 200 days before retrofitting are in-control (undamaged).

The optimization scheme for the bandwidth was validated.

“Moreover, a numerical and easily reproducible case study was presented (a) to demonstrate that the underlying covariance function (the ground truth) can be retrieved correctly and (b) to validate the optimization scheme for the bandwidth.”

When combined with confounder-adjusted covariances, PCA turns into a supervised method that is able to exploit the additional information provided through measurements of the confounding variables.

“However, when combined with confounder-adjusted covariances, it turns into a supervised method that is able to exploit the additional information provided through measurements of the confounding variables.”