Approximating Dynamic Elastic Modulus of Concrete for an Old Aqueduct Using Dynamic Tests and BP Neural Network

Round 1

Reviewer 1 Report

This paper presents a study for a monitoring an aqueduct structure and identifying its stiffness properties using a back-propagation neural network. 

Overarching thoughts and questions I had when reading through the paper:

Q1 - How do we know that the FEM model is valid with respect to the real aqueduct?

Q2 - Why does the trained NN predict a value for bearing stiffness that is on the extreme value of the trained range (indicates this may not be a good fit after all)?

Q3 - What is the value of the NN compared to classical optimization methods in this scenario?

Q4 - What is the sensitivity of the frequencies to each of the stiffness values?

Q5 - Why these three stiffness values, and why are they constant throughout the entire structure?

Specific Comments:

1. Line 93: How can these three lateral modes be separated in the field data? They are very closely spaced, and we do not have any mode shape or phase  information to be able to distinguish them. It seems they may be subtly different based on FEM results, but how do we know we can trust that?

2. Figures 4-7: What do the different colored lines mean? Different sensors? Should contain a legend to clarify which line means what.

3. Figure 7: Figure caption seems like stock text. Please provide correct caption.

4. Lines 134-137: Please expand on this "m" method for those who are not familiar with the Chinese design codes. What is "m", why does it have units of force/length^4, and are there any source references (i.e., not secondary references like the Chinese code) that provide values of "m" for different subgrade materials?

5. Line 142: Clarify what is meant by "vertical ... rotation." You mean "rotation about the vertical axes" right?

6. Line 150-153: What conversion or equation is used (reference please) from concrete design strength to dynamic modulus? Where did the shear stiffness of the asphaltic bearings come from? Please provide a reference or justification for the selected stiffness.

7. Lines 154-158 and Table 2: The sensitivity for parameter m is shown in Table 2. However, the sensitivity for all other parameters is NOT shown. Please include a sensitivity analysis for all other parameters: pile modulus (authors say this is insensitive, but do not show any results), K_flume, E_d-arch, and E_d-frame.

8. Related to previous comment: Does the flume itself also have a stiffness value (like E_d-flume), or is it the same as E_d-arch? I only see stiffness for the asphaltic bearings, but not the actual flumes.

9. Also related to previous comment: How likely is it that the stiffness values are the same at all locations? A single parameter is used to capture all parts of the arch trusses or all parts of the bent frames and moment frames. It seems that these parameters may vary throughout the structure, though I'm not convinced the BP model could handle that many outputs.

10. Lines 185-189: Reorder this paragraph. Because it discusses the longitudinal modes and refers to figures 11 and 12, it should go BEFORE the discussion of the vertical modes in Figure 13.

11. Figures 8-10: It is very difficult to see the differences between the lateral vibration modes. Is there some way to present a "baseline" or "undeformed line" so that the mode shapes can be more easily compared?

12. Line 203: Minor comment, but you should define "BP" the first time you use it: back propagation.

13. Line 213: Please justify the range for K_flume. When taking the natural log, the range of this parameter is actually rather narrow... ln(K-flume) varies from 13.08 to 18.63. This is a ratio of 18.63/13.08 = 1.42, which is less than the range considered by the two E_d values (60/35 = 1.71 and 55/30 = 1.83). The other issue is that, using the field-measured frequencies, the NN returns the extreme value of K_flume. This is suspect; we cannot be sure if this is an accurate estimation of stiffness or an artifact of fitting. The authors should investigate a higher upper bound for K_flume to see if the NN might converge to values within (not at the edges) of the modeled range.

14. Table 3: Please define purelin, tansig, and logsig transfer functions.

15. Line 261 and 276: Is there any way to validate these predictions for the stiffness parameters with real-world information? Why should these numbers be trusted? Particularly the K_flume parameter is suspect for the reasons explained above.

16. Related to above comment: Could the authors present some other optimization scheme for estimating these stiffness parameters? It seems like it would be straight-forward to do some sort of gradient descent or classic optimization, changing the stiffness values to minimize the error in frequency. What exactly is the benefit of training a NN as opposed to classic optimization?

17. Line 314: This is not an example of "transfer learning", which normally refers to transferring the "knowledge" from a trained NN to some other domain. Rather, this is just a simple application of an assumption that model and reality are consistent with each other (which may or may not be true...). How do we know that the model matches reality? What sorts of model validation were performed?

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Reviewer 2 Report

In the paper, the dynamic elastic modulus Ed-frame, Ed-arch and Kflume are adopted as the parameters to define the dynamic behavior of the aqueduct. a full-scale 3D FE model of the entire aqueduct is created and the analytical dataset for the network are obtained, then the primary modal frequencies of the aqueduct obtained from in-situ dynamic tests are put into the network to get the final approximations for Kflume, Ed-arch and Ed-frame. The paper exhibits clear writing logic and rigorous numerical and laboratory experimental designs. However, to enhance the academic quality and persuasive power, some revisions are suggested. The specific suggestions are as follows:

1. In the first section, the introduction does not provide sufficient background and include all relevant references. In the field of health monitoring and damage detection, BP neural network is a widely used method in the field of structural health monitoring ( Damage Identification of a Steel Frame Based on Integration of Time Series and Neural Network under Varying Temperatures), please highlight the improvement and advantage of the BP neural network in the paper. if possible, please introduce the advantage of the methods compared to swarm intelligent algorithms for the optimization of updating, such as GA(Vibration-based structural damage identification under varying temperature effects), MFO(Damage Identification of Bridge Structures Considering Temperature Variations-Based SVM and MFO) and WOA(Structural damage identification based on substructure method and improved whale optimization algorithm).

2. In the opinion of the reviewer, the modulus of the concrete will alter under the variation of temperature, please clarify how the temperature is considered in the paper. 

3. If possible, please provide the primary mode shapes of the aqueduct.

4. Please introduce the software used in the paper to establish the finite element model and the details of finite element model.

5. line 322-323, the authors conclude that the make-shift asphaltic bearings of U-shape flumes 322 basically can be treated as a three-directional hinge in the FE model, please clarify it.

Author Response

Please see the attachment.

Author Response File: Author Response.doc

Round 2

Reviewer 2 Report

Please provide the detailed response to the comments.

Author Response

Sorry, the reviewer's comments were not successfully submitted last time. Please refer to the attachment.

Author Response File: Author Response.doc

Round 3

Reviewer 2 Report

Please accept as is.