Tool Wear State Identification Based on the IWOA-VMD Feature Selection Method

Round 1

Reviewer 1 Report

- Authors should strive to write a shorter, clearer title.

- What is the type of stainless steel used during machining tests. Your designation and nominal chemical composition must be added.

- Regarding the cutting tool, type, material, number of teeth and diameter must be included.

Author Response

1. Authors should endeavor to write shorter, clearer titles.
RESPONSE: I thank the reviewers for their questions and apologize for failing to write a shorter, clearer title. The revised title is added below:
Tool wear state identification based on IWOA-VMD feature selection method.
2. what type of stainless steel was used in the machining test. Your name and nominal chemical composition must be added.
RESPONSE: Thanks to the reviewer for the question and I apologize for not describing in detail the type of tool used. The types of tools used are as follows:
The dry milling cutter was a three-flute ball-ended carbide tool cutting a square stainless steel workpiece with a hardness of HRC52.
3.Regarding the cutting tool, it is important to include the type, material, number of teeth and diameter.
RESPONSE: Thank you for your reviewer question and I apologize for not describing in detail the type, material, number of teeth and diameter of the tools used. The specific description of the tools used is as follows:
Milling cutter: 6mm three-flute carbide ball milling cutter; Material of the milling cutter: stainless steel (HRC52); Also, detailed information about the milling cutter can be found on the official website of the 2010 PHM Society Conference Data Challenge - PHM Society as the data used in this paper is the 2010 PHM open dataset; therefore, the milling cutter information is not described in detail in this paper.

These are my answers to the questions posed by the experts and I would like to ask the editors and reviewers to evaluate them.
With this sincere
Greetings!

Reviewer 2 Report

The authors have proposed a milling tool wear estimation approach using CNN.

The work is interesting and novel. The paper is structured well, has a clear logical flow and has good readability. I however have the following questions/comments/suggestions to improve the article further.

First, why IWOA? there are plenty of optimization algorithms and a lot many standard ones (GA, PSO, ABC etc) have been thoroughly investigated and have proven their capabilities. Isn't it better to go with such algorithms than IWOA? Otherwise, a performance comparison of IWOA vs. the state of the art is recommended. 

The same question for 1D CNN, I can see the direct selection of one of the deep learning models without any rationale for its selection over other algorithms. The readers would benefit if you gave a comprehensive explanation of the logic of your selection- (or a performance comparison with respect to other standard ML models)

I hope the ground truth class is manually labelled from microscope images at the training stage. How do you account for data and model drifts over the course of time?

Have you analysed the possible reasons for the low classification accuracy of the initial condition compared to others? How do you plan to better the model?

Why just three class labels and why do we need a classifier for this problem? The terms 'initial', 'normal' and 'severe' are very qualitative. And there can be many transition stages in between. So why can't the wear value be predicted using a regression model?

The work would have been more interesting if the real-time monitoring tool wear using cutting force signals had been demonstrated and validated, especially if the transition from one class to another had been demonstrated.

Nevertheless, I think the work is overall good and I hope it has the potential to evolve into a good tool wear prediction and adaptive control system. I hope the authors can consider and incorporate my suggestions in the next revision, all the best.

English is fine

Author Response

First of all, why choose IWOA?There are many optimization algorithms out there, and many of the standard algorithms (GA, PSO, ABC, etc.) have been thoroughly researched and proven to have their capabilities. Wouldn't using such algorithms be better than IWOA? Otherwise, it is suggested to compare IWOA with existing techniques.

Response: I thank the reviewer for the question, and I am very sorry for not clearly presenting the advantages of the IWOA algorithm over other optimization algorithms for milling cutter wear state identification. The description is added below:

In this paper, the superiority of IWOA in optimizing VMD parameters is demonstrated by comparing several standard algorithms (GA, PSO, GWO, etc.) and verifying them in Section 4.2.

For the same problem for 1D CNNs, I can see a direct choice of one of the deep learning models and no reason to choose it over other algorithms. Readers would benefit from a full explanation of the logic of your choice (or a performance comparison with other standard ML models).

Response: Thank you to the reviewer for the question, and I am very sorry for not clearly presenting the advantages of the 1D CNN model over other deep learning models. The description is added below:

Since the main research focus of this paper is in the direction of milling cutter wear feature extraction, it does not focus on the selection of the model for milling cutter wear state recognition. However, this paper verifies the superiority of the 1DCNN model in recognizing milling cutter wear states by comparing the 1DCNN and BP neural networks in subsection 4.5. Meanwhile, it will take the problem you raised as the next research focus.

I would like to manually mark the ground truth courses based on microscope images during the training phase. How do you explain the drift of the data and model over time?

Response: Thank you to the reviewer for the question, and I am very sorry for not being able to clearly explain the changes in the data and model over time. The explanation is added below:

Since this paper uses a publicly available dataset, it was not possible to manually label the ground truth conditions, and doing experiments and collecting your own data will be the focus of the next steps.

Have you analyzed the possible reasons for the low classification accuracy of the initial condition compared to the other conditions? How do you plan to improve the model?

Response: Thank you to the reviewer for the question, and I am very sorry for not being able to clearly indicate the reason for the low accuracy of the initial condition components. The analyzed reasons are added below:

The reason for the low classification accuracy in the initial condition may be due to the fact that the training of the 1D CNN model requires a large number of data samples, while in the initial condition there are fewer data samples, so the low classification accuracy occurs when the method is used to recognize the initial condition. And we are ready to collect multiple sets of experimental data and train the model using a large amount of data to improve the recognition accuracy in the next plan.

Why are there only three class labels, and why do we need a classifier to solve this problem? The terms "initial,"  "normal," and "severe" are very qualitative. There may be many transitions in between. So, why can't we use a regression model to predict wear values?

RESPONSE: Thank you for the reviewer's question, and I apologize for not being able to clearly indicate the reason for the three categories and the choice of classifiers rather than regression models. The reasons are added below:

The reason is that the milling cutter has three wear stages in its whole life cycle, corresponding to the "initial,"  "normal," and "severe" wear states, and the purpose of this paper is to improve the surface quality of the machined parts and the wear of the milling cutter. The aim of this paper is to improve the surface quality of the machined parts and determine whether the milling cutter is about to be scrapped or not. Therefore,  a classifier is used to identify the wear state of the milling cutter machining.

It is possible to use a regression prediction model to predict the wear value when identifying the wear state of the milling cutter, but accurate prediction of the wear value consumes a lot of arithmetic power and time, which is not cost-effective for the problem to be solved in this paper. By monitoring the milling cutter wear status in real time, it is already possible to ensure the surface quality accuracy of the machined parts and whether the milling cutter is about to be scrapped or not. Therefore, there is no need for further regression prediction of the milling cutter wear value.

Reviewer 3 Report

The work investigated the Wear state identification of milling cutters based on parameter optimised variational modal decomposition, multi-scale arrangement entropy and 1D Convolutional Neural Network.  The results provide a good guidance for machining thin-walled parts. The paper can be accepted after minor revision. 

(1) The grammar and expression can be improved. 

(2) How to use the prediction results to guide the actual production?

(3) What are the advantages of the proposed methods?

The expression and typo need to be modified. 

Author Response

(1) Grammar and presentation skills need to be improved.
RESPONSE: We thank the reviewers for their questions and apologize for not being able to describe this paper in short and scholarly terms. Improvements have now been made to the grammar and expression of the whole paper.
(2) How to utilize the prediction results to guide the actual production?
RESPONSE: We thank the reviewers for their questions and apologize for not presenting how the results of this paper are used to guide actual production. The description is added as follows:
In this paper, the real-time and accurate monitoring of the tool wear state during the machining of thin-walled parts is helpful to improve the machining accuracy and machining efficiency of thin-walled parts, so as to complete the guidance of the actual production. At the same time, the milling tool wear state identification model constructed in this paper provides a reference for professionals to manage the milling tool in accordance with the machining requirements, which contributes to the further construction of the intelligent workshop.
(3) What are the advantages of the proposed method?
RESPONSE: Thank you for the reviewer's question, and I am very sorry for not being able to synthesize and evaluate the advantages and disadvantages of the existing methods. The analyzed overview is now added as follows:
First, A VMD parameter optimization method based on IWOA optimization is proposed. Aiming at the problem that the WOA algorithm is slow to converge and prone to fall into local optimum, an improved IWOA algorithm based on elite inverse learning and golden sine algorithm is proposed. Taking the minimum of the joint correlation coefficient of modal components as the optimization objective with optimization, IWOA is used to search the VMD parameter combinations.
Second, combining its advantages, multi-scale arrangement entropy is introduced into the feature extraction of milling cutter to better characterize the feature information.
Third, 1D CNN is used to screen the wear features and identify the wear state of the milling cutter. This method can reduce the training parameters of 1D CNN while ensuring the accuracy of milling cutter wear monitoring.
These are my answers to the questions raised by the experts, and I would like to ask the editors and reviewers to evaluate them.
Sincerely!

Reviewer 4 Report

Comments in the attachment.

Comments for author File: Comments.pdf

Author Response

We thank the reviewers for their suggestions. However, due to the non-stationary and time-varying nature of the cutting force signals generated during milling process of a milling cutter, their features are usually embedded in multi-scale sequences, and their extraction by alignment entropy alone will be limited. Multi-scale permutation entropy (MPE) can extract multi-scale features from the signals according to the time series and comprehensively characterize the internal information of the cutting force signals. Therefore, the multiscale permutation entropy is used in this paper.

1. What material was milled? Designation. HRC is just the hardness of the material.
RESPONSE: Thank you to the reviewer for the question, and I apologize for not being able to describe in detail the milling cutter and the material of the part being machined. The relevant information is added below:
Milling cutter: 6mm three-flute tungsten carbide ball end milling cutter; the cutting material is a square stainless steel workpiece with a hardness of HRC52. The data used in this paper is the 2010 PHM open data set, so detailed information about the milling cutter can be found in the 2010 PHM Society Conference Data Challenge - PHM Society official website, so the milling cutter information is not described in detail in this paper.
2. What index was used to assess cutting edge wear?
Response: Thank you for the reviewer's question, and I apologize for not being able to provide the metrics for assessing milling cutter wear. Here is a list of metrics for assessing milling cutter wear:
Experiments were conducted using a three-flute ball-end milling cutter for 315 repetitions of face milling, with a fixed length of 108 mm for each pass. sensor data were recorded during each pass, and the amount of backface wear (VB) of the three cutting edges was measured using a microscope, LEICA MZ12, at the end of the pass as an indicator of tool wear.
In 1~33 times milling, in the tool surface roughness, coating uneven and other reasons, in the beginning of cutting a short period of time wear faster; in 34~166 times milling, the tool is in a stable wear stage, wear trend is relatively gentle; in the number of milling times greater than 166 times, the tool appeared to be violent wear phenomenon, basically reached the service life. Comprehensive three-edged tool wear, the wear process can be divided into three wear stages .
Since the data used in this paper is a public dataset.
3. There is no axis description in Figure 5. Please correct.
RESPONSE: Thank you for the reviewer's question, and I apologize for not being able to add the descriptions of the axes. The labels for the axes have now been added to the text.
4. What forces were analyzed? In the X, Y, Z direction? Maybe a tangential force? I suggest showing a drawing with the distribution of forces in the milling process. 
RESPONSE: Thank you for the reviewer's question, and I apologize for not being able to present a graph of the forces in the milling process in the text. The distribution of forces during milling is now added to the text as shown in Figure 6.
5. Figure 14, no axis description. 
RESPONSE: Thank you for the reviewer's question, and I apologize for not being able to add a description of the axes. The labels for the axes have now been added to the text.
6. Please write equation 3 correctly. There are errors.
RESPONSE: Thank you to the reviewer for the question, and I apologize for not writing Equation 3 correctly; the correct equation has now been added to the text.
7. Paragraph 180: correct the equation.
RESPONSE: Thank you to the reviewer for raising the issue, and I apologize for not writing Equation 3 correctly; the correct equation has been added to the text.
8. Conclusion 6: how did the authors determine the effectiveness of the proposed method? Tested on other studies?
Response: We thank the reviewer for his question and apologize for not clearly describing the validity of the proposed method, which is now added below:
This paper confirms the validity of the method by using the 2010PHM public dataset for validation; however, since this paper uses a public dataset, it has not been tested in other studies, so doing experiments and collecting our own data will be the focus of the next step;
9. How to monitor CTF with the proposed method?
RESPONSE: Thanks to the reviewer for the question, the short steps on how to monitor CTF using the method of this paper are added below:
At the same time, the CTF is monitored by collecting and analyzing the cutting force signal of the milling cutter. When the cutting force signal shows different and unusual signal fluctuation for a certain length of time, it can be assumed that the milling cutter is ruptured or has other quality problems at this time, and the professional can be reminded to overhaul the machine tool and change the cutter in time in order to avoid greater economic losses.
When the milling cutter wear state is in serious wear, the milling cutter wear state recognition model proposed herein can be used to identify its state, and timely warnings prompted by professionals to replace the inserts, thus avoiding the machine tool downtime and due to the milling cutter wear due to serious surface roughness of the machined parts can not reach the preset zero degree and other major economic losses.
The above methods are accomplished by the cutting force signal acquisition and visualization, AVMD-MPE cutting force signal feature extraction and 1D CNN milling cutter wear state recognition model in this paper, respectively.

To
Salutation!

Round 2

Reviewer 2 Report

This version has addressed all reviewer comments and can be considered for publication—no further comments/suggestions from my side. well done