Image Classification Method Based on Improved Deep Convolutional Neural Networks for the Magnetic Flux Leakage (MFL) Signal of Girth Welds in Long-Distance Pipelines
Round 1
Reviewer 1 Report
Authors need to justify why CNN is used.
Network parameters need to be discussed in details
performance analysis of the network model would add value
process of selection of data is not provided
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The manuscript entitled “Image Classification Method Based on Improved Deep Convolutional Neural Networks for the Magnetic Flux Leakage（MFL）Signal of Girth Welds in Long-distance Pipelines” has been investigated in detail. The topic addressed in the manuscript is potentially interesting and the manuscript contains some practical meanings, however, there are some issues which should be addressed by the authors:
1) The Introduction section needs a major revision in terms of providing more accurate and informative literature review and the pros and cons of the available approaches and how the proposed method is different comparatively. Also, the motivation and contribution should be stated more clearly.
2) What makes the proposed method suitable for this unique task? What new development to the proposed method have the authors added (compared to the existing approaches)? These points should be clarified.
3) “Discussion” section should be added in a more highlighting, argumentative way. The author should analysis the reason why the tested results is achieved.
4) The authors should clearly emphasize the contribution of the study. Please note that the up-to-date of references will contribute to the up-to-date of your manuscript. The studies related with machine learning approaches named-The effects on classifier performance of 2d discrete wavelet transform analysis and whale optimization algorithm for recognition of power quality disturbances, Crude oil time series prediction model based on lstm network with chaotic henry gas solubility optimization, Classification of power quality disturbances by 2d-riesz transform, multi-objective grey wolf optimizer and machine learning methods, Investigation of power quality disturbances by using 2d discrete orthonormal s-transform, machine learning and multi-objective evolutionary algorithms-can be used to explain the method in the study or to indicate the contribution in the “Introduction” section.
5) It will be helpful to the readers if some discussions about insight of the main results are added as Remarks.
This study may be proposed for publication if it is addressed in the specified problems.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors presented a research related to the automatic classification of girth welds in long-distance pipelines. For modelling, they used some of the state-of-the art methods and performed fined tuning. Experimental setup is fine in general. However, the paper could be improved in the following aspects:
1) The authors stated in the Introduction that there is a minimal research currently available regarding the classification of girth weld defect images using a deep convolutional neural networks. However, the authors could compare the obtained results of their model with the performance of other available models if possible, or just to mention performance of other models on other datasets if it is impossible to make a direct comparison.
2) The authors have stated that the training test was verified ten times. Does it mean that the authors have performed 10-cross-fold validation or they just repeated training on the same training, validation and test sets? It should be clarified in the manuscript.
3) The authors stated that the image data set of girth welds MFL signal was established with the radiographic testing results as labels. The authors should clarify if the data set is collected by the authors. Also, is it publicly available? Is it possible to access the data set or to test the model on some other data set? Also, why do after data set augmentation there are no images with incomplete penetration, crack, pit, and undercut labels?
4) Conclusions and Discussion section is very brief and have to be rewritten.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Authors have tried to present some clarity on the suggestions provided by the reviewers.
However the revision is not satisfactory. Like still it is not clear why use CNN over other algorithms.
Algorithms related data is not provided with full clairty.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
They have completed needed rearrangements of the paper. The projection for future work is very useful for the researchers to work in this topic. Please cite Figures 1 and 2 if you haven't drawn them yourself. In my opinion, this study can be proposed for the publication.
Author Response
Thank you for your encouragement. We will continue to try our best to study in the next research.
Reviewer 3 Report
Although I have encouraged the authors in the first round of review as the topic is very interesting, I have to reject the manuscript as the response is not adequate. However, I suggest to the authors to resubmit the manuscript and consult additional machine learning expert to perform experiments again.
Here are comments to the authors' response:
1) The authors have included an overview regarding CNN networks in the field of face recognition, speech etc, but that was not required following my comments. It was required to make an overview of other models related to the girth weld detection and possibly to compare the results.
2) The authors have stated the following:
"Then, the data set after enhancement is divided into training set and validation set in the ratio 9:1, conducted 10-cross-fold validation with each training result tested on the same testing set."
This is not accurate, it is not valuable if we have the same testing set each time and that is not k-folds cross validation. Experimental setup should be as follows:
The k-fold cross-validation procedure divides a limited dataset into k non-overlapping folds. Each of the k folds is given an opportunity to be used as a held-back test set, whilst all other folds collectively are used as a training dataset. A total of k models are fit and evaluated on the k hold-out test sets and the mean performance is reported.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 1 Report
Authors have satisfactorily addressed the observation indicated by the reviewers
Reviewer 3 Report
Description of k-cross fold validation can still be improved by discussing why do we need a separate test set when we take each fold as a test and calculate average performance, but can be accepted in the present form, too. Other issues are fixed.
