Damage Location Diagnosis of Frame Structure Based on a Novel Convolutional Neural Network

Round 1

Reviewer 1 Report

This manuscript proposed a novel deep learning-based approach for structural damage diagnosis， where a convolutional neural network was developed for the task of interest. First, CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise Analysis) was developed for processing original vibration signals for feature extraction. Then, the extracted features were inputted into hybridised CNN and LSTM for damage diagnosis. The proposed method was validated based on a 4-storey structure frame of Columbia University, with satisfactory results. Overall, the topic of this research is interesting, and the manuscript was well organised and written. The detailed comments are summarised as follows.

1.       Abstract: “LSTM” should be provided with its full name, when it appears for the first time.

2.       The contribution and innovation of the manuscript should be clarified clearly in abstract and introduction.

3.       Broaden and update literature review on convolutional neural network or deep learning and its engineering applications, such as structural damage diagnosis and behaviour evaluation. E.g. Torsional capacity evaluation of RC beams using an improved bird swarm algorithm optimised 2D convolutional neural network; Vision-based concrete crack detection using a hybrid framework considering noise effect.

4.       EEMD may have mode mixing problem. How did the authors avoid this problem in this application?

5.       The performance of deep learning model is heavily dependent on the setting of hyperparameters. How did the authors set them to achieve the optimal classification performance in this research?

6.       Fig. 9: the confusion matrix is not complete. Please revise it. In addition, the accuracy is 100%. It is very hard to achieve 100% prediction accuracy in the real application.

7.       How about the robustness of the proposed method against noise effect?

8.       More future research should be included in conclusion part.

Author Response

Q1.  Abstract: “LSTM” should be provided with its full name, when it appears for the first time.

Our response: Thank you for your comment. We have modified it in the original text. You can find it in the revised version.

Q2.  The contribution and innovation of the manuscript should be clarified clearly in abstract and introduction.

Our response: Thank you for your comment. We have emphasized the contribution and innovation of this article in the abstract and introduction. You can find it in the revised version.

Q3. Broaden and update literature review on convolutional neural network or deep learning and its engineering applications, such as structural damage diagnosis and behaviour evaluation. E.g. Torsional capacity evaluation of RC beams using an improved bird swarm algorithm optimised 2D convolutional neural network; Vision-based concrete crack detection using a hybrid framework considering noise effect.

Our response: Thank you for your comment. We have broadened and updated literature review on deep learning and its engineering applications. You can find it in the revised version.

Q4. EEMD may have mode mixing problem. How did the authors avoid this problem in this application? 

Our response: Thank you for your comment. When using EMD to decompose vibration signals, due to the existence of pulse effects or discontinuities, the distribution of extreme points in the signal segment is uneven, resulting in mode mixing problem. In order to eliminate the mode mixing problem of EMD, Huang et al（Ensemble Empirical Mode Decomposition: a noise-assisted data analysis method. Advances in Adaptive Data Analysis）added white Gaussian noise of the same amplitude to the source signal and proposed EEMD. Using the continuity of white noise and the uniform distribution of frequency, the extreme points of the source data are uniformly distributed, thus eliminating the mode mixing phenomenon. Although EEMD can effectively eliminate mode mixing, it is limited to Gaussian white noise, and mode mixing will still occur when other types of noise are used. It is difficult to obtain pure white Gaussian noise in the natural environment. In order to obtain pure white Gaussian noise, computer simulation is often used to add white Gaussian noise to the source signal. In order to avoid the problem of mode aliasing, we paper first used adaptive EEMD (CEEMDAN) to enhance its performance, in addition, the noise signal used in this paper is computer simulated Gaussian white noise.

Q5. The performance of deep learning model is heavily dependent on the setting of hyper parameters. How did the authors set them to achieve the optimal classification performance in this research? 

Our response: Thank you for your comment. Hyper parameters in deep learning can generally be divided into three categories: network parameters, optimization parameters and regularization parameters. Network parameters generally include the number and size of convolution kernels, the number of network layers and activation function. Optimization parameters generally refer to learning rate, batch size, and parameters of different optimizers. Regularization parameters generally include weight attenuation coefficient and dropout. The most important of these parameters is the learning rate, which directly controls the magnitude of network gradient update in training and directly affects the performance of the model. In order to better set the learning rate to obtain the optimal classification performance, Adam was adopted as the optimization algorithm in this paper. It is an adaptive optimization algorithm for learning rate. It dynamically adjusts the learning rate of each parameter by using the first order moment estimation and the second order moment estimation of the gradient. Adam is usually robust to the selection of hyper parameters, so it is very helpful to the parameter adjustment of neural networks. Therefore, the setting of the learning rate is mainly based on the initial value suggested by Adam. Batch size is also a very important hyper parameter in depth learning, which determines the direction of quantity gradient decline. Too large or too small batch size is not conducive to the performance of the network model. The general reference range of batch size is [1:1024]. This paper set the batch size according to the size of the training data and the empirical value of WDCNN, and compared the performance of the model under different batch sizes, so as to obtain the appropriate batch size. For the setting of other hyper parameters, the original parameters of WDCNN were mostly used, because the advantages of WDCNN have been verified. For other hyper parameters that cannot be given by WDCNN, the main setting principle is: first take the empirical value, and then take different values for performance comparison, so as to select the parameters that make the network model perform optimally.

 Q6. Fig. 9: the confusion matrix is not complete. Please revise it. In addition, the accuracy is 100%. It is very hard to achieve 100% prediction accuracy in the real application. 

Our response: Thank you for your comment. We have updated Fig.9. You can find it in the revised version. In addition, for the case that the accuracy rate is 100%, we explained as follows:

First of all, WDCNN has a high accuracy in feature extraction and classification of vibration signals. Secondly, on the premise of using CEEMDAN to preprocess the data, this paper adopted a WDCNN network with strong performance, and used LSTM to capture the additional features between the features before and after the vibration signal, which not only improved the purity of training sample data, but also increased the form of features extracted by the convolution layer, thus making the network model very powerful. Finally, there are only five diagnosis classifications of the frame structure conducted in this paper, so a higher accuracy rate can be obtained in the case of less training times, as shown in Figure 8, a higher accuracy rate had been obtained after 150 times of training. In addition, in the actual convolutional neural network training, the accuracy rate is only infinitely close to 100%, and will not really reach 100%, but the computer will default to display the accuracy rate of infinitely close to 100% as 100%. Therefore, many scholars regard the accuracy in this case as 100%, such as reference [29].

Q7. How about the robustness of the proposed method against noise effect?.

Our response: Thank you for your comment. The method proposed in this paper has strong robustness in terms of against noise. This is because in this study, we have diagnosed the damage location of the frame structure under five different signal-to-noise ratios (the signal-to-noise ratio is - 4dB, - 2dB, 0dB, 2dB and 4dB respectively). From the diagnosis results in Table 3, we can know that the method proposed in this paper have high accuracy under different noise environments. Only when the signal-to-noise ratio is - 4 dB, the accuracy decreases to 99.97%, and in other cases, it is 100%. This shows that when the noises fluctuate in a large range, it has little impact on the accuracy of the method proposed in this paper. Therefore, the method proposed in this paper has strong robustness in terms of against noise.

Q8. More future research should be included in conclusion part. Our response: Thank you for your comment. The conclusion of this paper is obtained when Gaussian white noise is added to the original data, but in the actual environment, the noise is complex. Therefore, in the next research, we will study the damage location diagnosis of the frame structure in different noise environments. In addition, only five damage locations of frame structures are diagnosed in this paper, and there are at least nine damage situations in the open source data provided by Columbia University. Therefore, the types of damage locations will be expanded in the next research.

Author Response File: Author Response.doc

Reviewer 2 Report

The comments mentioned in the pdf file must be included. It is necessary to add more reference and give more datails about the experiment

Comments for author File: Comments.pdf

Author Response

Q1. This paper needs more details about experiment.

Our response: Thank you for your comment. We have added relevant details in the experiment section; we add the structural parameters of the network model, the running environment of the experiment and the relevant hyper parameters. You can find it in the revised version.

Q2. It needs to include more reference about the state of the art.

Our response: Thank you for your comment. We have added relevant reference about the state of the art in the introduction section. You can find it in the revised version.

Q3. A flow diagram in the experimental section will be useful, Would it be useful in marine structures, buildings?, explain other possible applications?

Our response: Thank you for your comment. Is there any misunderstanding here? There is no flow diagram in the experimental section of this article. However, the method proposed in this paper is a general fault diagnosis method based on vibration data. This method can not only be used to diagnose the damage location of frame structures, but also be applicable to other structural health monitoring and fault diagnosis objects based on vibration data, such as the health monitoring of bridges, grandstands and other buildings, and the fault diagnosis of gear boxes, bearings and other machinery.

Author Response File: Author Response.doc

Round 2

Reviewer 1 Report

All the technical issues have been well addressed by the authors. I do not have further comments.