An Integrated Deep-Learning-Based Approach for Energy Consumption Prediction of Machining Systems
Round 1
Reviewer 1 Report
Dear Authors,
Comments:
- the text needs to be corrected for linguistic correctness
- the purpose of the study is not clear and needs improvement
- methodology and case study are unclear and chaotic
- equations - abbreviations and symbols in each equation are not developed
- figure 2 - I would recommend changing the direction of information - from top to bottom, but this is just a suggestion
- figure 3 - additional arrows that are not related to anything
- figure 4 - needs improvement, misspellings need to be correct
- Figure 6 - the axes of your graphs are not signed
Author Response
Point 1: The text needs to be corrected for linguistic correctness.
Response 1: Thanks to the reviewers for their patient review and meticulous evaluation. We have checked the paper carefully and corrected the grammatical errors in the paper. At the same time, some improper expressions have been corrected.
Point 2: The purpose of the study is not clear and needs improvement.
Response 2: Thanks to the reviewers for their patient review. In the preface, we have added the research purpose and content of the paper, which more clearly indicates the urgency of our work. Additions are highlighted in yellow.
Point 3: Methodology and case study are unclear and chaotic.
Response 3: Thank the reviewers for their careful review. In this paper, we establish an energy consumption prediction framework and an energy consumption prediction model for machining systems based on deep learning. The energy consumption forecasting framework consists of four modules, namely, data acquisition layer, data processing layer, data analysis layer and application layer. Secondly, based on this framework model, we propose a deep learning-driven energy consumption prediction method, which includes six steps. On the basis of the theory, we set up an experimental platform for energy consumption prediction, and carried out experimental verification. If the reviewer has questions about our modifications or is not satisfied with them, we are also very welcome to raise them, and we will actively modify them.
Point 4: Equations - abbreviations and symbols in each equation are not developed.
Response 4: Thanks to the reviewers for their patient review. We have checked the abbreviations and expression problems in the previous version and corrected them (the symbols in Equation 3 to Equation 8 have been explained one by one). If the reviewer thinks there is anything that needs to be modified, he is also welcome to point it out, and we will actively modify it.
Point 5: Figure 2 - I would recommend changing the direction of information - from top to bottom, but this is just a suggestion.
Response 5: Thanks to the reviewers for their patient review and meticulous evaluation. We have changed the direction of the information flow in Figure 2, from top to bottom. The sequence from top to bottom is :1) Data acquisition and storage layer; 2) Data preprocessing layer; 3) Data analysis layer; 4) Application layer.
Point 6: Figure 3 - additional arrows that are not related to anything.
Response 6: Thanks to the reviewers for their patient review. We have gone over it and removed the redundant arrows in Figure 3.
Point 7: Figure 4 - needs improvement, misspellings need to be correct.
Response 7: Thank the reviewers for their carefull review. We have reviewed and corrected the error in Figure 4.
Point 8: Figure 6 - the axes of your graphs are not signed.
Response 8: Thank the reviewers for their patient review. I'm terribly sorry for our oversight. We have gone over and corrected the missing axis symbols in Figure 6.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper proposes a deep learning-based approach for predicting the energy consumption of machining system. In the proposed approach, an improved LSTM network is developed on real collected test data. The advantage of the proposed approach is demonstrated by comparing it with other LSTM-style neural networks. This paper is well organized; but requires significant technical editing. The followings are suggested comments to improve the manuscript quality:
1. In section 3.1 data preprocessing, it is mentioned that the pre-processed EC data sets from two ways were randomly mixed at a ratio of 8:2. In section 4.2 and Table 1.1, however, it is stated that the data collected by both methods are mixed 7:3 to build a new EC dataset. Could the author clarify how the data from the two methods are combined?
2. Data leakage is a big problem in developing predictive deep learning models. It could happen that the training dataset is so similar to the testing and validation dataset that the deep learning just memorize the results in the training dataset. In this study, the dataset is divided into training, testing, and validation. However, data leakage is highly possible because the data is all collected on one machine. It is suggested to elaborate more on how to avoid this data leakage problem and guarantee accurate predictions on new and unseen data.
3. To make it easier to understand, the author may consider redrawing Figure 2. This figure should follow the sequence up to down: 1) data acquisition and storage layer; 2) data preprocessing layer; 3) data analysis layer; 4) application layer.
4. In the proposed LSTM network, the inputs of the network are spindle speed, feed speed, cut depth, number of cutting edges current, voltage. The outputs of the network are the predicted sequence energy values. Some information on the inputs and outputs are missing, such as the magnitude, unit and sampling frequency. It is recommended to summarize the information in a table and provide some examples in figures.
5. In this study, the proposed DIWPSO-LSTM method is compared with four other deep learning methods, namely the conventional PSO-LSTM, LSTM, IPSO-LSTM and ACMPSO-LSTM. Although the results show the proposed DIWPSO-LSTM method performs better than others, it is not clear if it is a fair comparison. It could happen that the performance of the four other methods will be improved by choosing better hyperparameters and fine tuning the models. Could the author provide some evidence that the comparison is reasonable?
6. Please provide units of optimization time and training time in Table 4. Similarly, what is the unit of energy data in Figure 8, Figure 9 and Figure 10?
7. Figure 11, please display the RMSE and Loss in log scale. It is not clear when the training starts to converge from the current figure.
Author Response
Response to Reviewer 2 Comments
Point 1: In section 3.1 data preprocessing, it is mentioned that the pre-processed EC data sets from two ways were randomly mixed at a ratio of 8:2. In section 4.2 and Table 1.1, however, it is stated that the data collected by both methods are mixed 7:3 to build a new EC dataset. Could the author clarify how the data from the two methods are combined?
Response 1: Thanks to the reviewers for their comments. The construction of the data set is to mix the collected energy consumption data of different categories in accordance with 7:3 to establish a comprehensive energy consumption data set. In terms of the input of the energy consumption prediction model, training data and verification data are obtained from the energy consumption data set. A total of 80% of the energy consumption data set is used for training, and 20% of data is used for prediction. 200 sets of data are taken from the training data set for verification.
Point 2: Data leakage is a big problem in developing predictive deep learning models. It could happen that the training dataset is so similar to the testing and validation dataset that the deep learning just memorize the results in the training dataset. In this study, the dataset is divided into training, testing, and validation. However, data leakage is highly possible because the data is all collected on one machine. It is suggested to elaborate more on how to avoid this data leakage problem and guarantee accurate predictions on new and unseen data.
Response 2: Thanks to the reviewers for their comments. In this study, we construct our data sets in two ways: by using power testers and by simulated data. The data collected by the power tester is also the data collected by different personnel in different processing periods of different workpieces. A data breach is highly unlikely. At present, few data sets about processing energy consumption have been disclosed. Our next research content is to collect workpiece processing data with different characteristics on multiple CNC machine tools. This will enable us to build a more accurate energy consumption prediction data set for the processing system. We will publish our data set and report the latest research results in a timely manner.
Point 3: To make it easier to understand, the author may consider redrawing Figure 2. This figure should follow the sequence up to down: 1) data acquisition and storage layer; 2) data preprocessing layer; 3) data analysis layer; 4) application layer.
Response 3: Thanks to the reviewers for their patient review and meticulous evaluation. We have changed the direction of the information flow in Figure 2, from top to bottom, The sequence from top to bottom is :1) Data acquisition and storage layer; 2) Data preprocessing layer; 3) Data analysis layer; 4) Application layer.
Point 4: In the proposed LSTM network, the inputs of the network are spindle speed, feed speed, cut depth, number of cutting edges current, voltage. The outputs of the network are the predicted sequence energy values. Some information on the inputs and outputs are missing, such as the magnitude, unit and sampling frequency. It is recommended to summarize the information in a table and provide some examples in figures.
Response 4: Thanks to the reviewers for their patient review. We added Table 2 to clarify input and output data information, such as the names and units of input and output data. Input data include: machining time, machining duration, current, voltage, spindle speed, feed speed, power, cutting depth, tool edge number, and machining energy consumption. Sampling frequency is described in Section 4.1. The sampling frequency is set to 100ms.
Point 5: In this study, the proposed DIWPSO-LSTM method is compared with four other deep learning methods, namely the conventional PSO-LSTM, LSTM, IPSO-LSTM and ACMPSO-LSTM. Although the results show the proposed DIWPSO-LSTM method performs better than others, it is not clear if it is a fair comparison. It could happen that the performance of the four other methods will be improved by choosing better hyperparameters and fine tuning the models. Could the author provide some evidence that the comparison is reasonable?
Response 5: Thanks to the reviewers’ carefull evaluation. We have carried out hundreds of experiments on five methods (LSTM, PSO-LSTM, IPSO-LSTM, ACMPSO-LSTM, DIWPSO-LSTM). When hidden layer neurons, learning rate, learning rate decline factor and learning rate decline period are selected as the hyperparameters to be optimized, the proposed method (DIWPSO-LSTM) has better predictive performance than the other four methods. Choosing more hyperparameters for experimental analysis would be our next research direction. If the performance of other methods is improved by increasing the number of hyperparameters to be optimized, we will report our findings in time.
Point 6: Please provide units of optimization time and training time in Table 4. Similarly, what is the unit of energy data in Figure 8, Figure 9 and Figure 10?
Response 6: Thanks to the reviewers for their patient review and meticulous evaluation. We have added units of optimization time and training time to Table 4, and unit of energy consumption value to Figure. 8, Figure. 9 and Figure. 10.
Point 7: Figure 11, please display the RMSE and Loss in log scale. It is not clear when the training starts to converge from the current figure.
Response 7: Thanks to the reviewers’ carefull evaluation. Figure 11 shows content generated automatically by the training process. We also tried to express it with a log scale, but unfortunately it failed. The modification time of the article is too late if we rewrite the program. We then considered the necessity of Figure 11. We found that what it was intended to show was already shown in Figures 8 through Figures 10, as well as in Table 5. Therefore, we consider removing Figure 11 from the paper. Of course, if the reviewer feels that this is not appropriate or that Figure 11 should be retained, we are willing to discuss this further.
Author Response File: Author Response.docx
Reviewer 3 Report
Dear authors,
I have carefully read and analyzed the paper. It is about the study of using LSTM networks to predict energy consumption in mechanical systems. The results show the framework for optimizing the hypermeter values of the double layer LSTM.
Abstract – last line: Performance improvement compared to what? Please clarify.
Figure 2: Perhaps it would be appropriate to add the numbers from Section 3.1. or reverse the order in the figure, because intuitively the reader would expect the first thing they see in the figure to be the first step, but here the last step is at the top.
Section 3.1, Data preprocessing (2): 4:1 instead of 8:2?
Figure 3: Two arrows are coming from the right side of the figure and point to the input layer, without their purpose being clear.
Section 3.2: What are the settings for the particle swarm (swarm size, etc.)? Have you tried varying the settings and how do you justify the settings you chose? Do you have some references that support the settings you chose?
Section 3.2: Is there variation in the accuracy and computational cost of the hyperparameter optimization procedure as the particle swarm settings change.
Figure 7: is this true for the test data or? Please indicate.
Figure 7: The performance metrics are unclear and difficult to read, e.g., the R2 appears to be the same for all models. Also, the y-axis should read MAE, ME, RMSE, R2 instead of LSTM. This figure is also redundant as the same data could be observed in Table 3. Consider removing one.
Table 3: Why did the performance metrics deteriorate from LSTM to PSO-LSTM or IPSO-LSTM? Please comment.
Table 4: Please explain how the optimization time for DIWPSO-LSTM is lower when the training time is almost twice as long compared to PSO-LSTM? Are fewer iterations used? Why do you not use the same number of iterations?
Conclusion: please remove figures from the conclusions and add them to the results or appendix.
General: Please put all performance metrics (RMSE, MAE, etc.) in italics
Author Response
Response to Reviewer 3 Comments
Point 1: Abstract – last line: Performance improvement compared to what? Please clarify.
Response 1: Thank the reviewers for their careful review. The proposed method improves the mean absolute error (MAE) and mean error (ME) by more than 30%, and is better than other methods in terms of RMSE and R2 indicators.
Point 2: Figure 2: Perhaps it would be appropriate to add the numbers from Section 3.1. or reverse the order in the figure, because intuitively the reader would expect the first thing they see in the figure to be the first step, but here the last step is at the top.
Response 2: Thanks to the reviewers for their patient review and meticulous evaluation. We have changed the direction of the information flow in Figure 2, from top to bottom. The sequence from top to bottom is :1) Data acquisition and storage layer; 2) Data preprocessing layer; 3) Data analysis layer; 4) Application layer.
Point 3: Section 3.1, Data preprocessing (2): 4:1 instead of 8:2?
Response 3: I would like to thank the reviewers for their detailed feedback. Section 3.1 describes how to obtain data. To create a comprehensive energy consumption data set, it is necessary to mix the collected data from different categories in a ratio of 7:3. Section 4.2 describes the proportional segmentation of training data, validation data and prediction data in the prediction model. 80% of the energy consumption data set is for training and 20% for forecasting. 200 groups of data were selected from the training data set for verification. The ratio of training to prediction data is 80 per cent and 20 per cent, and a change to 4:1 is not recommended.
Point 4: Figure 3: Two arrows are coming from the right side of the figure and point to the input layer, without their purpose being clear.
Response 4: Thank you for your careful review. These two arrows are redundant and we have carefully changed Figure 3. Due to our oversight, these two arrows are redundant. We have reviewed and modified Figure 3.
Point 5: Section 3.2: What are the settings for the particle swarm (swarm size, etc.)? Have you tried varying the settings and how do you justify the settings you chose? Do you have some references that support the settings you chose?
Response 5: Thank you for your paitent review. The number of initialized particle swarms is 20, the population dimension is 4, and the maximum number of iterations is 100, inertia weights ,, acceleration constant . For example, the setting of particle population size, inertia weight and acceleration constant has been studied extensively (reference 33) and proved to be the most appropriate choice. Therefore, the empirical method is adopted in this paper. During the experiment, we also tried to modify the settings of particle population size, inertia weight and acceleration constant. However, after many tests, we found that the initial settings were reasonable and achieved acceptable results.
Point 6: Section 3.2: Is there variation in the accuracy and computational cost of the hyperparameter optimization procedure as the particle swarm settings change.
Response 6: Thanks to the reviewers for their patient review and meticulous evaluation. The accuracy and calculation cost of hyperparameter optimization process vary with the change of particle swarm Settings. In this paper, four hyperparameters (hidden layer neurons, learning rate, learning rate decline factor and learning rate decline period) are selected to be optimized, namely, particle dimension is 4. If the particle dimension is too large, the precision and calculation cost of the optimization process will be further increased. Therefore, only four hyperparameters optimization are considered in this paper. In the next step, we will optimize more hyperparameters and then conduct comparative experiments.
Point 7: Figure 7: is this true for the test data or? Please indicate.
Response 7: Thank you for your careful review. The predicted results are the average of 10 runs on the server, and the data is real.
Point 8: Figure 7: The performance metrics are unclear and difficult to read, e.g., the R2 appears to be the same for all models. Also, the y-axis should read MAE, ME, RMSE, R2 instead of LSTM. This figure is also redundant as the same data could be observed in Table 3. Consider removing one.
Response 8: Thanks to the reviewers for their patient review and meticulous evaluation. We modified Figure 7, and the Y axis is the specific value of the evaluation index. In Figure 7, MAE, ME, RMSE and R2 of the model in this paper are better than those of the control model. However, the difference in R2 in Figure 7 is small, so we represent it in Table 3. If the reviewer thinks it needs to be modified, we are also willing to alter it actively.
Point 9: Table 3: Why did the performance metrics deteriorate from LSTM to PSO-LSTM or IPSO-LSTM? Please comment.
Response 9: Thanks to the reviewers for their meticulous review. The evaluation of performance indicators is divided into two categories. One type is energy consumption prediction using single-layer LSTM, and the other type is energy consumption prediction using PSO algorithm to optimize single-layer LSTM hyperparameters. By comparing five methods, DIWPSO-LSTM, LSTM, PSO-LSTM, IPSO-LSTM, and ACMPSO-LSTM, we found that the prediction effect of single-layer LSTM model is better than that of PSO-optimized LSTM, but the prediction model of PSO-optimized hyperparameters of double-layer LSTM has better prediction effect than that of single-layer LSTM.
Point 10: Table 4: Please explain how the optimization time for DIWPSO-LSTM is lower when the training time is almost twice as long compared to PSO-LSTM? Are fewer iterations used? Why do you not use the same number of iterations?
Response 10: Thanks to the reviewers for their patient review. In this paper, the iteration times of prediction model DIWPSO-LSTM and other comparison models (PSO-LSTM/IPSO-LSTM/ACMPSO-LSTM) are set to 300. The running results show that the DIWPSO-LSTM has a shorter optimization time, and the training time is almost twice as long as the other aspects, because the dynamic inertia weight optimization method is used to reduce the hyperparameter selection time, but the DIWPSI-LSTM uses a double-layer LSTM structure, thus increasing the training time.
Point 11: Conclusion: please remove figures from the conclusions and add them to the results or appendix.
Response 11: Thanks to the reviewers for their patient review. Upon careful examination, we found no figures in the conclusion. If the reviewer thinks that the conclusion needs to be modified, we welcome the reviewer to point out the specific points that need to be modified. We are willing to make active modifications.
Point 12: General: Please put all performance metrics (RMSE, MAE, etc.) in italics.
Response 12: Thanks to the reviewers for their patient review. We have checked carefully and changed the evaluation metrics (MAE, MA, RMSE, R2) to italics. And highlighted in yellow in the text.
Author Response File: Author Response.docx
Reviewer 4 Report
The paper is well written.
The data analysis flow is complete and well conducted. The results achieved are more than satisfactory and are very interesting.
Author Response
Thank you for your careful review. We checked the paper carefully and revised part of the content. The specific changes were highlighted in yellow in the paper.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Dear authors,
I have not noticed any changes in the latest version of your article. Please provide changes which I suggested in my review and what you said you have done in your resposes
Author Response
Please refer to the attachment for replies to reviewer comments, and see the yellow highlighted marks in the pdf version.
Author Response File: Author Response.docx
Reviewer 2 Report
I am glad to see all my comments have been well addressed. However, I did not see any change in the manuscript. I am wondering if an old version of the manuscript has been mistakenly uploaded?
Author Response
Please refer to the attachment for replies to reviewer comments, and see the yellow highlighted marks in the pdf version.
Author Response File: Author Response.docx
Reviewer 3 Report
Authors addressed my concerns in the 'Author's notes', but the uploaded paper seems to be the same as the old version. Please upload the new version.
Author Response
Please refer to the attachment for replies to reviewer comments, and see the yellow highlighted marks in the pdf version.
Author Response File: Author Response.docx
Round 3
Reviewer 1 Report
Dear authors,
thank you for provided changes
Reviewer 2 Report
I am glad to see all my comments have been well addressed. Good job!
