Machine Learning of Surface Layer Property Prediction for Milling Operations

Round 1

Reviewer 1 Report

Overall, this is a well-written, well-prepared study. This reviewer has the following minor requests.

Figure 5 presents a good layout since it has its legend and Figure 4c is unclear. Legend of the data should be clearly presented for all figures.

The paper has no Conclusions. Why?

Figure 1 is not a good representation of the performed study. Better representation is needed. What are ‘trigger experiments’?

The study talks about the absolute errors in two places. However, there is no depth on the comparative error/accuracy of the results.

In Figure 5, explain the GMDH Neural Network 50 data point.

Author Response

The authors thank the reviewer for the thelpful and constructive comments.

1) Legend for Figure 4c is given in the textbox and in the axis descriptions to the left and the right of the figure.

2) The conclusions are drawn from line 306-314. The text has been adjusted to highlight the conclusions and conclusions have been added.

3) The Figure 1 and text in line 155 has been adjusted to make clear that the Python script triggers the measurement.

4) In line 218-228, performance of the tool wear estimator is quantified by calculating the R² value and the mean average error MAE for different Δr ranges. Calculating the relative error therefore appears to be redundant.

5) In lines 282-283, it is stated that the R² value drops significantly when the training data set is too small. When only 50 % of the data set are used, R² value drops to around R²= 0.6. The figure 5 gives the reader an insight on the necessary magnitude of a dataset for training a robust model. The detailed mathematical reasons for the drop at this specific value are not within the scope of this paper.

Reviewer 2 Report

Dear Authors,

In this work, authors developed, employed and evaluated  a fully automated system for machine learning of the relation between tool wear, cutting parameters and SLP.  Monitoring of SLP was performed based on micromagnetic and coercive force measurements, while feed per tooth fz, tool radius reduction Δr, cutting speed vc, average power  consumption PM, Mean and the standard deviation of power measurement PM, Std  were used as the  input parameters of the SLP model. Based on the monitoring results, the feed rate was adjusted according to changes in the cutter radius due to wear. The developed automated approach, which is based on the analysis of 826 experiments conducted in this study, has obvious advantages over the existing manual ones, and, could possibly be applied in the real-life industry in the nearest future.  The paper is physically sound, technically solid and can be accepted for publication after the following minor issues are addressed:

1. On line 26, after the keywords, please use dot.
2. In the list of references, DOIs  were not  included. Please, add.
3. In lines 128-129, it is recommended to indicate the type of cutter coating, the angle of the helical groove and, if possible, the rake and clearance angles.
4. The shape of the workpiece surface being machined was not specified. Please, specify. 
5. In lines 136-137, it is stated that the CNC command G41 provides an equidistant path to the profile. However, in such conditions, the cut width can vary in the case, when the trajectory is not a straight line. Please comment.
6. In lines 208-209, it is necessary to specify which sum of currents is  meant by “three phase currents’.
7. In lines 224-227, it is recommended to specify wear on the cutter flank caused by a decrease in the radius Δr.
8. On lines 262-269, it is necessary  to explain in detail why  and how the specific limits for coercive force and micromagnetic measurements were chosen because their use causes a significant productivity degradation. In particular, a cutting speed of 45 m / min and a feed per tooth of 0.01 mm are very low when machining steel with a hardness of 241 HB. For machinability group P (alloyed heat-treatable steels), the recommended cutting speed is 180-207 m / min and the recommended feed per tooth is  0.06 - 0.08 mm (Ap = 5 , Ae = 10 for shoulder milling).
9. Lines 313-333, the description of the  further research is too long and general. Please, make it shorter, more specific and concise.  
10. It is recommended to include in the Conclusion section a comprehensive description of gains in the performance achieved due to the adjustments in the feed rate made based on the application of the automated system developed in this work.

Author Response

The authors thank the reviewer for the helpful and constructive comments.

1. Done.

2. Done.

3. As stated in line 129, an uncoated cutting tool was used. Helix angle, rake angle and clearance angle were added in line 130.

4. Shape of the workpiece surface being machined was added in line 140.

5. The information, that the trajectory was a straight line was added in line 139.

6. The information was added in line 210.

7. The information was added in line 229.

8. As stated in line 235, inital cutting parameters for the algorithm have been chosen by recommendation of the tool supplier. The tool supplier recommended v_c = 40-80 m/min and f_z = 0.05 mm. The micromagnetic values were chosen to keep dominating material properties within a tolerance band by avoiding re-hardening zones, which would be indicated by an increase in coercive force, as stated in lines 258-261. Since the developed algorithm only considers the MM, as stated in lines 164-180, and not the productivity, productivity degradations are possible. A multi-criteria optimisation of the MM as well as the productivity are not within the scope of this paper.

9. The list of further research has been adjusted according to the reviewers comments.

10. A conclusion section has been added and the gains have been detailed in line 312.