Assessment of Earthquake Destructive Power to Structures Based on Machine Learning Methods

Round 1

Reviewer 1 Report

The subject of this article is very interesting and it deserves to be published

Author Response

Response to Reviewer 1 Comments

Point 1: The subject of this article is very interesting and it deserves to be published.

Response 1: We appreciate your interest and recognition of our work and thank you for reviewing our manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear authors,

Your paper on “Assessment of Earthquake Destructive Power to Structures Based on Machine Learning Methods” is comparing machine learning based approaches to assess the destructive power of earthquake to structures. The paper and the comparison of methods are in principle very interesting and relevant. However, at present, the paper is rather difficult to follow. Therefore, I would have the following main recommendations to improve the presentation of the work:

• The abstract is not very well written and often stays rather vague about the main findings. In particular, a) about the selection of an architecture “ after discussion” b) add more specific findings and c) the final sentence “ The outcomes of this study are expected to provide a useful reference for timely emergency response and disaster relief after earthquake“ is too unspecific.
• I would suggest to add a bit more information about the benefits of the THA-based seismic damage simulation method as part of the introduction section.
• In the introduction neither the research gap nor the innovation of your work is clearly stated. Please make this more explicit.
• The structure of the paper needs to be improved, the methodology is scattered across several sections and does not provide an easy overview of the methods. Furthermore, the results are mixed with the discussion, which also does not provide an easy overview of the results. Moreover, the discussion is too short and does not allow the fully understand the relevance of your findings. For example, I would find it very interesting to understand how you see that your findings can support to“assist in emergency response and rapid disaster relief”. Finally, the conclusion chapter is partially a summary of results (e.g., “The architecture of BPNN is discussed by building multiple models with different number of hidden layers and nodes. As a result, a BPNN model with two hidden layers and four nodes in each layer is adopted”) and could be more specific about the major conclusions.
• As part of the restructuring, also several sections could be made more concise.

Some more specific comments:

• The section title “Methodology framework” could be improved – please specific a framework of what?
• Table 1. “Structural parameters” of what – please add some context to the figure caption.
• Line 115 etc. “PEER ground motion database [27]” – you need to give some more details about this data as this are the base for your analysis.
• Section 3 the split in training, validation and testing seems to have very few testing data. But I also do not fully understand this split.
•  

Author Response

Response to Reviewer 2 Comments

Point 1: The paper and the comparison of methods are in principle very interesting and relevant. However, at present, the paper is rather difficult to follow. Therefore, I would have the following main recommendations to improve the presentation of the work.

Response 1: The authors appreciate the reviewer’s constructive comments on our manuscript. The manuscript has been improved according to the comments, and all comments have been responded correspondingly.

Point 2: The abstract is not very well written and often stays rather vague about the main findings. In particular, a) about the selection of an architecture “after discussion” b) add more specific findings and c) the final sentence “The outcomes of this study are expected to provide a useful reference for timely emergency response and disaster relief after earthquake” is too unspecific.

Response 2: The authors appreciate the reviewer’s comment. The abstract has been revised accordingly.

1. a) The sentence about the selection of an architecture “after discussion” has been modified to “The optimized BPNN architecture is obtained by discussing the influence of different number of hidden layers and nodes”.
2. b) The finding about computational efficiency has been added to the abstract. Specifically, the CNN model exhibits remarkable computational efficiency, the prediction of 1000 structures based on the CNN model takes 0.762 second, while 507.81 seconds are required for the conventional THA-based simulation.
3. c) The last sentence has been modified to “The proposed method can assist in fast prediction of engineering demand parameters of a large number structures, which facilitates the damage or loss assessments of regional structures for timely emergency response and disaster relief after earthquake”.

The revised abstract is presented as follow:

This study presents a machine learning-based method for the destructive power assessment of earthquake to structures. First, the analysis procedure of the method is presented and the back propagation neural network (BPNN) and convolutional neural network (CNN) are used as the machine learning algorithms. Second, the optimized BPNN architecture is obtained by discussing the influence of different number of hidden layers and nodes. Third, the CNN architecture is proposed based on several classical deep learning networks. To build the machine learning models, 50570 time history analysis results of a structural system subjected to different ground motions are used as training, validation and test samples. The results of BPNN indicate that the features extraction method based on short-time Fourier transform (STFT) can well reflect the frequency-/time-domain characteristics of ground motions. The results of CNN indicate that the CNN exhibits better accuracy (R² = 0.8737) compared with that of BPNN (R² = 0.6784). Furthermore, the CNN model exhibits remarkable computational efficiency, the prediction of 1000 structures based on the CNN model takes 0.762 second, while 507.81 seconds are required for the conventional THA-based simulation. Feature visualization of different layers of CNN reveals that the shallow to deep layers of CNN can extract the high to low frequency features of ground motions. The proposed method can assist in fast prediction of engineering demand parameters of a large number structures, which facilitates the damage or loss assessments of regional structures for timely emergency response and disaster relief after earthquake.

The manuscript has been revised in Lines 12 to 29.

Point 3: I would suggest to add a bit more information about the benefits of the THA-based seismic damage simulation method as part of the introduction section. 


Response 3: The authors appreciate the reviewer’s comment. Supplementary information about the benefits of the THA-based seismic damage simulation method has been added to the introduction section. The revised content is as follow:

The THA-based seismic damage simulation method can adequately consider the frequency-/time-domain characteristics of ground motions and the nonlinear seismic performance of structures, so as to reasonably reflect the destructive power of earthquake to structures. Furthermore, the THA-based seismic damage simulation method can be applied to different types of structural systems and reveal the whole process of failure mechanism of a structure [11–12]. Therefore, the THA-based seismic damage simulation is adopted as the benchmark method in this study.

The above description can be found in Lines 47 to 53.

Point 4: In the introduction neither the research gap nor the innovation of your work is clearly stated. Please make this more explicit.

Response 4: The authors appreciate the reviewer’s comment. According to the suggestion of the reviewer, different machine learning methods on the seismic response simulation of structures are compared to demonstrate the research gap and the innovation of this work.

To date, extensive research works have been reported on the machine learning-based seismic response simulation (Table 1). Mangalathu et al. [23–24] adopts various machine learning methods to predict the damage state of a structure. It is a classification problem and the computational workload of the simulation is light. Nevertheless, in some cases, obtaining engineering demand parameters (EDPs) is more favorable for damage or loss assessment of a structure. In the study of Zhang et al. [25], long-short-term memory (LSTM) [26] is used to predict the structural time-history response. The LSTM network requires a relatively large computational workload [27] and is not suitable for the response prediction of a large number of structures. This study focuses on the fast prediction of EDP under seismic excitation. Considering that BPNN [28] and CNN [13] models are capable of simulating complex nonlinear problems and exhibit good computational efficiency, these two models are used to predict the EDP of structures subjected to earthquake, so as to reflect the destructive power of earthquake to structures.

The above description can be found in Lines 68 to 81.

Table 1. Comparison of machine learning-based seismic response simulation methods.

Point 5: The structure of the paper needs to be improved, the methodology is scattered across several sections and does not provide an easy overview of the methods. Furthermore, the results are mixed with the discussion, which also does not provide an easy overview of the results. Moreover, the discussion is too short and does not allow the fully understand the relevance of your findings. For example, I would find it very interesting to understand how you see that your findings can support to “assist in emergency response and rapid disaster relief”. Finally, the conclusion chapter is partially a summary of results (e.g., “The architecture of BPNN is discussed by building multiple models with different number of hidden layers and nodes. As a result, a BPNN model with two hidden layers and four nodes in each layer is adopted”) and could be more specific about the major conclusions.

Response 5: The authors appreciate the reviewer’s comment. The structure of the paper is elaborated in Section 2, where the simulation mainly consists of two modules: sample preparation and machine learning simulation. The sample preparation module presents details of the structural system, ground motion records and how each sample for model training are obtained. In the machine learning module, BPNN and CNN are adopted as the specific machine learning methods, of which the model architecture and simulation results are given in Section 4 and Section 5 respectively.

To answer the question of how our findings can support emergency response and rapid disaster relief, the authors have supplemented more contents as follow:

The CNN model can make significantly faster prediction of engineering demand parameters compared with THA-based method. For example, using a laptop platform (i5-4210H, RAM 8G, GTX 960M), the prediction of 1000 structures using the CNN model takes 0.762 second, which is over 650 times faster than that using the THA-based method (507.81 seconds). The predicted engineering demand parameters can be used to assess the damage, economic loss and downtime of regional structures based on damage assessment methods [9, 57], so as to assist in emergency response and rapid disaster relief.

The authors have also modified the conclusion part according to the reviewer’s suggestions. The modified contents are as follow:

In this study, a machine learning-based destructive power assessment method of earthquake to structures is proposed. The method is proposed based on the back propagation neural network (BPNN) and convolutional neural network (CNN). The results of 50,570 numerical simulations are used as samples to train the machine learning models and some conclusions can be drawn as follow:

1. For the feature extraction of ground motion time history data, STFT can well reflect the frequency-/time-domain features of ground motions. The prediction using STFT as feature extractor yield higher accuracy (R² = 0.6784) than that using FFT (R² = 0.4277).
2. The selection of the number of layers and nodes of BPNN is related to the features. In this study, in the case of 40 features as input, the optimal model is the BPNN with 2 hidden layers and 3 nodes in each layer. And for the case of 10 features as input, the optimal model is the BPNN with 2 hidden layers and 4 nodes in each layer.
3. The model based on CNN exhibits better prediction accuracy (R² = 0.8737) than the BPNN model (R² = 0.6784). This is because the convolution layer of CNN can identify the frequency-/time-domain information of a signal, and avoid the bias caused by artificial feature extraction at the meantime. And the advantage of CNN is magnified for its deep network architecture and multiple filters in each layer.
4. Feature visualization of convolutional layers reveals that the shallow convolutional layers of the CNN model mainly extract the high-frequency features of a signal, while the deep convolutional layers primarily extract the low-frequency features of a signal. In addition, the features extracted by CNN model can reflect the time-domain features of a signal.
5. The CNN model exhibits remarkable computational efficiency compared with THA-based method. For example, using a laptop platform (i5-4210H, RAM 8G, GTX 960M), the prediction of 1000 structures using the CNN model takes 0.762 second, which is over 650 times faster than that using the THA-based method (507.81 seconds). The high computational efficiency of the CNN-based seismic response prediction makes it promising for use in timely assessment of regional structures.

The research proves that CNN model can make significantly faster prediction of engineering demand parameters compared with THA-based method. The predicted engineering demand parameters can be used to assess the damage, economic loss and downtime of regional structures based on damage assessment methods [9, 57], so as to assist in emergency response and rapid disaster relief.

The above description can be found in Lines 94 to 98; Lines 314 to 316 and Lines 372 to 402.

Point 6: The section title “Methodology framework” could be improved – please specific a framework of what?

Response 6: The authors appreciate the reviewer’s comment. The section title “Methodology framework” has been modified to “Analysis procedure of the proposed method”. The manuscript has been revised in Line 89.

Point 7: Table 1. “Structural parameters” of what – please add some context to the figure caption.

Response 7: The authors appreciate the reviewer’s comment. The caption has been modified to “Structural parameters of the elasto-plastic SDOF model”. The manuscript has been revised in Line 131.

Point 8: Line 115 etc. “PEER ground motion database [27]” – you need to give some more details about this data as this are the base for your analysis.

Response 8: The authors appreciate the reviewer’s comment. More details about the data have been added. The modified contents are as follow:

Considering that the characteristics of different ground motions vary significantly, to cover different types of ground motions as much as possible, 5057 ground motion records were collected from the pacific earthquake engineering research (PEER) center ground motion database [33]. The obtained ground motions were recorded from 1935 to 2003, of which earthquake magnitude ranges from 4.2 to 7.9. Most of the ground motions were recorded in US and Japan. Some typical earthquakes are Chi-Chi, El Centro, Northridge, Landers and Kobe earthquakes.

The above description can be found in Lines 132 to 137.

Point 9: Section 3 the split in training, validation and testing seems to have very few testing data. But I also do not fully understand this split.

Response 9: The authors appreciate the reviewer’s comment. The explanation for the split of dataset is as follow:

According to the open source book of Ng [34], test set should be large enough to give high confidence in the overall performance of model. Therefore, in the early work of this research, the authors tried the following sample splits:

Different sample splits and their average R²

It can be seen from the comparison that the R² values are very close when size of the test set ranges from 1000 to 5000. This means the test set size of 1000 can reasonably reflect the performance of the prediction model, and the authors chose to keep more samples in the training set to improve the generalization of the prediction model. Therefore, the original sample set is divided into 48,570, 1000 and 1000 samples for the training, validation and test sets, respectively. It’s worth noting that the original 50,570 sample set is generated using 5057 ground motion records in ten different PGAs. To makes sure that the ground motions in the test and validation sets are completely different from those in the training set, 4857, 100 and 100 ground motion records are randomly selected to generate the training, validation and test sets, respectively.

The above discussion can be found in Lines 147 to 155 of the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors propose a method to assess the destructive power of earthquakes to structures based on machine learning techniques. The topic is interesting and the paper is well organized. The methods are described with proper details, and the obtained results seem to be promising and support their conclusions.

I have two comments which can improve the manuscript even further.

- It would be of interest if the authors consider employing the Long Short Term Memory (LSTM) networks as they are better at handling time series or sequence data.

- It would be useful if you compare your obtained results with the previous work in a table when applicable.

Author Response

Response to Reviewer 3 Comments

Point 1: The topic is interesting and the paper is well organized. The methods are described with proper details, and the obtained results seem to be promising and support their conclusions. I have two comments which can improve the manuscript even further.

Response 1: The authors appreciate the reviewer’s constructive comments on our manuscript. The manuscript has been improved according to the comments, and all comments have been responded correspondingly.

Point 2: It would be of interest if the authors consider employing the Long Short Term Memory (LSTM) networks as they are better at handling time series or sequence data.

Response 2: The authors appreciate the reviewer’s comment. LSTM network is a very powerful machine learning method for processing time series, and its application is also a hot research topic at present. In fact, in the work of Zhang et al. [25], LSTM network was used to predict the structural seismic response, which has great potential. But LSTM network is more suitable for the continuous prediction of time series, rather than the prediction of an engineering demand parameter such as the maximum displacement of a structure. Meanwhile, the computational workload of LSTM network is relatively heavy [27], which is not suitable for the response prediction of a large number of structures. This study focuses on the fast prediction of EDP under seismic excitation, the application of LSTM network in the scenario of this work may need more considerate investigation. The authors are very grateful to the reviewer for providing us with this idea, which has great potential and is important for our future work.

Point 3: It would be useful if you compare your obtained results with the previous work in a table when applicable.

Response 3: The authors appreciate the reviewer’s comment. The table that compares this study with previous works have been added to the Introduction.

To date, extensive research works have been reported on the machine learning-based seismic response simulation. Mangalathu et al. [23–24] adopts various machine learning methods to predict the damage state of a structure. It is a classification problem and the computational workload of the simulation is light. Nevertheless, in some cases, obtaining engineering demand parameters (EDPs) is more favorable for damage or loss assessment of a structure. In the study of Zhang et al. [25], long-short-term memory (LSTM) [26] network is used to predict the structural time-history response. The LSTM network requires a relatively large computational workload [27] and is not suitable for the response prediction of a large number of structures. This study focuses on the fast prediction of EDP under seismic excitation. Considering that back propagation neural network (BPNN) [28] and convolutional neural network (CNN) [13] models are capable of simulating complex nonlinear problems and exhibit good computational efficiency, these two models are used to predict the EDP of structures subjected to earthquake, so as to reflect the destructive power of earthquake to structures. The manuscript has been revised in Lines 68 to 81.

Table 1. Comparison of machine learning-based seismic response simulation methods.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Dear authors,

thank you for you clear and detailed replies. The paper has been substantially improved. Congrats!