Value of Information of Structural Health Monitoring in Asset Management of Flood Defences
Round 1
Reviewer 1 Report
This paper presents Value of Information for SHM in Asset Managemen of flood defences. The results were presented in suitable forms of graph and table. The overall flow of the manuscript is acceptable.
Author Response
Thank you for your comments.
Reviewer 2 Report
The monitoring and management of flood defence systems is of high scientific and societal interest, especially in the Nordic countries in Europe. With the current paper it is attempted to provide the basis for quantifying and optimising the monitoring and management of flood defence systems with the Bayesian decision theory. Whereas the dike reliability, monitoring and management models are of high standard, the use of the Bayesian decision theory requires a major revision in the following aspects:
1. The foundations of the use of Bayesian decision analyses for quantifying the value of monitoring should be more directly addressed and the paper should be better contextualised in recent works on value of monitoring information e.g. with the COST Action TU1402 Guidelines and:
a. Thöns, S. (2018). On the Value of Monitoring Information for the Structural Integrity and Risk Management. Computer-Aided Civil and Infrastructure Engineering 33(1): 79-94. DOI: 10.1111/mice.12332
b. Fauriat, W. and E. Zio (2018). An Importance Measure to Assess the Value of a Component Inspection Policy. 2018 3rd International Conference on System Reliability and Safety (ICSRS).
c. Thöns, S. and M. Kapoor (2019). Value of information and value of decisions 13th International Conference on Applications of Statistics and Probability in Civil Engineering (ICASP), Seoul, Korea, May 26-30, 2019.
2. The Bayesian decision analysis and value of information approach needs to be more explicitly formulated.
a. How is the information dependent optimal action found by maximisation of the utility as in the original works on Bayesian decision theory (as claimed by the authors with e.g. “In this study we use a Bayesian decision model …” - line 113, “Bayesian pre-posterior decision model” line 390)?
b. Which assumptions are made in regard to the original formulation?
c. It should be distinguished between a model parameter (e.g. a threshold) conditional value of information analysis and a conditional value of information analysis according to the Bayesian decision theory.
i. A conditional value of information analysis constitutes the difference between the maximised expected value of the utilities calculated with a posterior decision analysis (with posterior probabilities calculated with Bayesian updating) and a prior decision analysis.
d. It should be explicitly written how the probabilities are calculated for a pre-posterior decision analysis.
i. A “Bayesian pre-posterior decision model” (line 390) requires the modelling and forecasting of monitoring outcomes.
ii. The calculation of the posterior probability is appropriate for a posterior decision analysis.
3. In Equ. 8, the dike reinforcement and the monitoring seem to occur at the same point in time. Further explanation is needed.
4. Formulations like “… by the state of the flood defence in that individual draw …” (line 219) need revision for clarity.
Author Response
We thank you for your very useful comments. We agree that Bayesian decision theory needed better embedding in existing literature. We believe that we have made the necessary changes to ensure this.
For point-by-point comments please see the attached Word document.
Author Response File: 
Author Response.pdf
Reviewer 3 Report
This paper has explored the benefits of Structural Health Monitoring for flood defences for effective infrastructure management based on a Bayesian pre-posterior decision model and the concept of Value of Information have been used to quantify the benefits of different monitoring strategies. Overall the paper is well written and should be published in the Infrastructures Journal.
Author Response
Thank you for your comments.
Round 2
Reviewer 2 Report
The authors have improved the manuscript according to the review comments.
The description of the modelling and the assumptions of monitoring outcomes (comment 2.d) in the context of a pre-posterior and posterior probabilities and decision analysis should be further improved in the paper.
Despite the claim of the authors of having addressed comment no. 3, there is no relevant adjustment. Comment 3 needs to be resolved.
Author Response
We are glad that the improvements made are recognized and thank you for making them once again.
With regards two the two remaining points:
The description of the modelling and the assumptions of monitoring outcomes (comment 2.d) in the context of a pre-posterior and posterior probabilities and decision analysis should be further improved in the paper.
It is a bit unclear to us what improvement you are looking for precisely. The modelling and forecasting of observations is described in section 3.3.2. We model the effect of observations by updating our belief distribution for hc which subsequently leads to a change in belief failure probability. The observations are samples from the state θhc,j in a sample of a possible posterior state j. To update the belief distribution we use the conjugate distribution in eq. 11. We sample many states j to get a proper preposterior estimate.
We've attempted to clarify the notation and make it more in line with the preceding sections in terms of terminology and symbols. The main changes are that we have more clearly described the link between the approach outlined and the preposterior analysis by introducing this in the beginning of the section. Additionally we have more explicitly described the time dimension in the equations as this was not entirely consistent.
Hopefully this resolves the mentioned comments.
Despite the claim of the authors of having addressed comment no. 3, there is no relevant adjustment. Comment 3 needs to be resolved.
We have added a remark to address this in the lines after the formula:
where cEAD(t) is the cost component of the EAD at time t and cm(t) and cr(t) are the costs for monitoring and reinforcement at time t. Note that these are equal to 0 if no reinforcement or monitoring is done at a specific time step t. r is the discount rate, for which a value of 3% is prescribed in the Netherlands.
We hope this solves the issue.
