Prediction of Structural Damage Trends Based on the Integration of LSTM and SVR
Round 1
Reviewer 1 Report
Overall the article is interesting and well-written. Following comments must be addressed before it can be accepted
1. Abstract should contain more quantitative information.
2. Fig. 4 is not appropriate. Please improve the quality and presentation of the figure
3. Why data has been standardized not normalized say [-1,1]. Give proper reasoning.
4. In the obtained transient frequency, 1000 data were taken for the fusion algorithm 334 prediction. The first 700 were used as training data, and the last 300 were used as test data. Have you tried any other splits? Like first 600 for training and last 400 for testing. Else state within the manuscript that most optimized results were obtained when first 700 were used as training and last 300 for validation
5. Only three statistical indices were chosen to show the accuracy which is not enough. More indices such as Willmott index, VAF, etc. must be estimated to conclude the superiority of LSTM-SVR method. Help can be taken from the following article as citation “Predicting the settlement of geosynthetic-reinforced soil foundations using evolutionary artificial intelligence technique. Geotextiles and Geomembranes, 49(5), pp.1280-1293.”
6. Remove the inconsistency R2 and R2.
7. Check for all the symbols and their description
8. Some latest references must be included to strengthen the AI literature review such as SVR . e,g., https://www.sciencedirect.com/science/article/pii/S1674775522001093
Must be improved and checked by native speaker.
Author Response
Response to Reviewer 1 Comments
Dear Reviewer:
Thank you for your comments on our manuscript entitled “Prediction of Structural Damage Trends Based on the Integration of LSTM and SVR” (ID：applsci-2381283). Those comments are very helpful for revising and improving our paper, as well as the important guiding significance to other research. We have studied the comments carefully and made corrections which we hope meet with approval. The main corrections are in the manuscript and the responses to the reviewers’ comments are as follows (the replies are highlighted in red ).
Responds to the reviewers' comments：
Point 1: Abstract should contain more quantitative information.
Response 1: We appreciate it very much for this good suggestion.The quantitative information has been added:Compared with individual LSTM and SVR models, the integration model has higher prediction accuracy for small samples in a chaotic time series, that is 6.56%,2.56%,3.7% respectively. The standard deviation of absolute percentege error(SDAPE) values of the three operating conditions under the integrated method have decreased 0.0994,0.0869and0.0921, can improve the stability of prediction.
Point 2: Fig. 4 is not appropriate. Please improve the quality and presentation of the figure.
Response 2: I am very sorry for my incorrect drawing and I have corrected the Fig.4 as follow.
Fig. 4. R2 corresponding to weight coefficient.
Point 3: Why data has been standardized not normalized say [-1,1]. Give proper reasoning.
Response 3: Thank you for pointing this out. The data is standardized data, and if I use ordinate [-1,1] in the figure, it does not look very clear, so only part of the figure is used.
Point 4: In the obtained transient frequency, 1000 data were taken for the fusion algorithm 334 prediction. The first 700 were used as training data, and the last 300 were used as test data. Have you tried any other splits? Like first 600 for training and last 400 for testing. Else state within the manuscript that most optimized results were obtained when first 700 were used as training and last 300 for validation.
Response 4: We appreciate it very much for your suggestion. I have tried many times before, and the results were not satisfactory. If the training data is less than 700, the model can not complete sufficient training, and the prediction result is unstable, with large error and low accuracy.
Point 5: Only three statistical indices were chosen to show the accuracy which is not enough. More indices such as Willmott index, VAF, etc. must be estimated to conclude the superiority of LSTM-SVR method. Help can be taken from the following article as citation “Predicting the settlement of geosynthetic-reinforced soil foundations using evolutionary artificial intelligence technique. Geotextiles and Geomembranes, 49(5), pp.1280-1293.”
Response 5: Thank you for pointing this out. Standard deviation of absolute percentege error(SDAPE)[32] and refined Willmott index (RWI) have been added in line 240 and line 244,the results have been added in the paper.
Point 6: Remove the inconsistency R2 and R2.
Response 6: I am very sorry for my incorrect writing, I have changed all the R2 into R2 in line 390,409,429.
Point 7: Check for all the symbols and their description.
Response 7: Thank you for pointing this out.I have checked for all the symbols and their description in the whole paper.
Point 8: Some latest references must be included to strengthen the AI literature review such as SVR . e,g., https://www.sciencedirect.com/science/article/pii/S1674775522001093
Response 8: Thank you for pointing this out. Some latest references have been changed in the paper:the No. of referance is:25,26,27,30,32,33,34,37,40,41,42.
Once again, thank you very much for your constructive comments and suggestions which would help us both in English and in depth to improve the quality of the paper.
Kind regards,
Yiyan Liu
E-mail: yyliu1@chd.edu.cn
Reviewer 2 Report
The article is concerned with the actual problem of predicting time-series using modern methods of support vector regression and long term memory networks. The literature review includes 42 sources from various countries. All references relevant to the research. To demonstrate the novel predictive method, 3 vibration signals corresponding to 3 stages of structural damage were used, based on experiments from the early 2000s. The prediction was made for the instantaneous (also called by the authors transient) frequency of the vibration signal. The integration model proposed by the authors has demonstrated a better quality of prediction of the 3 time-series data than the SVR and LTSM models.
The title of the article suggests application of the prediction to the structural damage trend, but in the work for each damaged state of the structure separately predicts the final part of the vibration response based on its first part. The purpose of prediction of the second part of this time series is not obvious. Why do you need an extra 5 seconds of signal forecasting? It is of practical interest to predict the time series, during which the damaged state of the structure changes, but this is not the case.
Also it is necessary to eliminate some shortcomings which are listed in what follows:
- on page 3, line 122. the i index at the top of the sum is unnecessary
-on page 4, line 154. typo in "SVector Regression"
- on page 9-11, figures 9,10,11,12. the vertical axis caption has the wrong reflection.
The article can be published after minor revision.
Author Response
Response to Reviewer 2 Comments
Dear Reviewer:
Thank you for your comments on our manuscript entitled “Prediction of Structural Damage Trends Based on the Integration of LSTM and SVR” (ID：applsci-2381283). Those comments are very helpful for revising and improving our paper, as well as the important guiding significance to other research. We have studied the comments carefully and made corrections which we hope meet with approval. The main corrections are in the manuscript and the responses to the reviewers’ comments are as follows (the replies are highlighted in red ).
Responds to the reviewers' comments：
Point 1: The title of the article suggests application of the prediction to the structural damage trend, but in the work for each damaged state of the structure separately predicts the final part of the vibration response based on its first part. The purpose of prediction of the second part of this time series is not obvious.
Response 1: We appreciate it very much for this good suggestion. The purpose of prediction of the second part of this time series is added as follows: Most of the main structural damages are caused by the stiffness decrease from the perspective of general structural damage characterization. When the structure has different degrees of stiffness damage,the system output acceleration is used to identify the gradual damage of the stiffness changes.The LSTM -SVR is used as the structure health prediction model in this paper. Since the instantaneous frequency can effectively reflect the health status changes of the structure[38], the model is divided into signal processing, data preprocessing and LSTM -SVR network prediction.
Point 2: Why do you need an extra 5 seconds of signal forecasting? It is of practical interest to predict the time series, during which the damaged state of the structure changes, but this is not the case.
Response 2:
We appreciate it very much for this good suggestion. An extra 5 seconds of signal prediction is done after the training process is complete, by applying the model to a future time series data. This can evaluate the model's generalization performance and provide a more accurate understanding of the model's performance in real-world applications, thereby offering valuable feedback for optimizing and tuning the model. But in practical, the damage of structures is a gradual process that requires a lot of time and data to predict. I will strive to improve this in my subsequent research work. Thank you very much for your suggestions
Point 3:On page 3, line 122. the i index at the top of the sum is unnecessary
Response 3: Thank you for pointing this out. I have revised : Where, is the function set of each modal component;is the central frequency set of each component.
Point 4: on page 4, line 154. typo in "SVector Regression"
Response 4: Thank you for pointing this out. “Support Vector Regression (SVR)”instead of “Support Sector Regression (SVR)”
Point 5: on page 9-11, figures 9,10,11,12. the vertical axis caption has the wrong reflection.
Response 5: ALL the vertical axis caption has been revised as follows:
(a) LSTM prediction (b) SVR prediction
Fig. 5. Comparison of short-term prediction effects of a chaotic time series.
Fig. 6. Prediction effects of a chaotic time series of the integration method.
Fig. 9. Acceleration vibration signals under three working conditions.
Fig. 10. The transient frequency corresponding to acceleration under three working conditions.
(a) LSTM model (b) Integration method
Fig. 11. Short-term prediction results of non-destructive conditions.
(a) LSTM model (b) Integration method
Fig. 12. Short-term prediction results after removing all the diagonal supports on the southeast side.
(a) Prediction results of LSTM (b) Integration method
Fig. 13. Short-term prediction results of bolts loosening at both ends of the first and second floor beams on the northeast side without any diagonal support.
Once again, thank you very much for your constructive comments and suggestions which would help us both in English and in depth to improve the quality of the paper.
Kind regards,
Yiyan Liu
E-mail: yyliu1@chd.edu.cn
Author Response File: 
Author Response.docx
