Research on CNC Machine Tool Spindle Fault Diagnosis Method Based on Deep Residual Shrinkage Network with Dynamic Convolution and Selective Kernel Attention Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a fault diagnosis method for CNC machine tool spindles based on an enhanced Deep Residual Shrinkage Network with Dynamic Convolution and Selective Kernel Attention (DDRSN-SKA). The approach integrates the continuous wavelet transform (CWT) with SKAttention to improve feature extraction and diagnostic accuracy. The study is interesting and relevant. However, several issues need to be addressed:

1. The review of related works on condition monitoring should be expanded to include more recent advances. In particular, please discuss methods such as enhanced particle filter and cyclic spectral coherence-based prognostics, sensorless robust anomaly detection of roller chain systems using motor driver data, and attention-guided graph isomorphism learning for remaining useful life prediction.
2. The introduction would benefit from a clearer presentation of the existing problems and the key contributions of this work. It is recommended to summarize these points in bullet form to enhance readability and emphasize the main innovations.
3. The manuscript should be carefully proofread to eliminate typographical errors and ensure consistency. For example, please revise lines 2, 20, and 78 for grammar and clarity, check the indices in Equations (1) and (2), and correct inconsistencies in symbols in lines 128, 139, and 141. Consistent terminology should also be maintained throughout the paper.

Author Response

Comments 1: The review of related works on condition monitoring should be expanded to include more recent advances. In particular, please discuss methods such as enhanced particle filter and cyclic spectral coherence-based prognostics, sensorless robust anomaly detection of roller chain systems using motor driver data, and attention-guided graph isomorphism learning for remaining useful life prediction.

Response 1: Thank you for your insightful feedback on expanding the review of related works. We have revised the manuscript to incorporate your suggestions, adding a new discussion on these advanced topics to the Introduction section between lines 70 and 82. In this revised section, we now cover the shift towards prognostics (including enhanced particle filters and cyclic spectral coherence), the trend of sensorless anomaly detection using motor driver data, and the application of cutting-edge architectures like attention-guided graph isomorphism learning.

 

Comments 2: The introduction would benefit from a clearer presentation of the existing problems and the key contributions of this work. It is recommended to summarize these points in bullet form to enhance readability and emphasize the main innovations.

Response 2: Thank you for your valuable feedback on our manuscript. We have carefully revised the paper according to your suggestions, with key changes made between lines 108 and 134 of the original document. In this section, we've adopted your recommendation to use bullet points to clearly summarize the existing problems in the field and the main contributions of our work.

 

Comments 3: The manuscript should be carefully proofread to eliminate typographical errors and ensure consistency. For example, please revise lines 2, 20, and 78 for grammar and clarity, check the indices in Equations (1) and (2), and correct inconsistencies in symbols in lines 128, 139, and 141. Consistent terminology should also be maintained throughout the paper.

Response 3: Thank you for your thorough inspection. We change “faultdiagnosis” to “fault diagnosis” in line 2, and delete one dot in line 20. We have corrected the spacing issue on line 78 by adding a space between "example," and "Li."we have addressed your comments by correcting the indices in equations (1) and (2), resolving the symbol inconsistencies in lines 128, 139, and 141, and ensuring consistent terminology throughout the manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

The article entitled “Research on CNC machine tool spindle fault diagnosis method based on DDRSN-SKA model” for Journal Algorithms (ISSN 1999-4893) is a suitable article for publication in the journal. We will not hide that our article gained interest both from the scientific ethos of the authors, as well as from their scientific consistency and methodology. Everything seems neat, compatible, and consistent. The mathematical reference equations are properly structured and justified, and the whole project initially seems complete. Although your text focuses on an interesting topic, several parts require substantial revisions to improve the clarity, structure, and overall presentation of your research. From the formal part of the paper, we note some specific comments that we consider important for improving the overall image of the publication.

• Without ignoring the problem of length, we would suggest that all abbreviations be removed from the title and abstract. It is not customary to use abbreviations in titles, captions, and abstracts; these are for the main text. Even completely self-explanatory abbreviations should be replaced with reference words.

 

• Excessive paragraphing is not necessary. For example, in lines 71 to 76, there is a paragraph with two sentences! Limiting three paragraphs per page or page section is sufficient. At least three paragraphs per page or page section are adequate.

 

• At the end of the introduction, the authors should accurately add the scientific questions of the article, which will be answered in the conclusions.

 

• After the scientific questions, the authors should list the individual chapters of the article with a precise description and justification.

 

• Between the titles of the chapters and subchapters (for example, “2. Theoretical foundations and network structure and 2.1 Continuous Wavelet Transform”, line 111 – 112), a short introductory text should be written that concerns the separation of the parts of the section.

 

• Congratulations on the scientific organization of the mathematical parts and diagrams. However, we would like to point out a publishing issue. The typologies, formulas, and symbols should be uniformly adapted to the text. We do not know whether this is the work of the publisher or the authors. We impose uniformity in our texts, so we believe that the authors will be able to do the same.

 

• Pay attention to some parts of the text (kernels.The feature – line 175), in which the punctuation and format specifications of the magazine are not adhered to.

 

• The typology should be marked intratextually (e.g., typology "4"), and not "can be represented as:" line 156, or "as follows", line 226. Also, the marking of the typology (e.g., (4) line 258 should be placed uniformly aligned to the right of the typology and not below! The same on line 228, and wherever else there are typologies. These are basic issues of standardization and are not in dispute.

 

• Between the titles of the chapters and subchapters (3.Analysis of experiments results, 3.1 Experimental environment and parameter settings, lines 267 and 268: experimental, or experiment not “experiments”), a short introductory text should be written that concerns the separation of the parts of the section.

 

• “Figure 6. Wavelet time-frequency plots for 10 health states at SNR of -8” and Figure 10. Wavelet time-frequency plots for four health states when SNR is -8: Explain and describe in detail how those images were obtained. What tool was used, how were the photos taken, and what exactly do they represent? Explain the content and background of each photo in detail. This specific reference is a piece of evidence, not an accompanying text; therefore, due attention must be given to it.

 

• Tables 3, 4, 6, and 7: Explain and describe in detail how those tables were obtained. What tool was used, how were the photos taken, and what exactly do they represent? Explain the calibration, subdivisions, physical parameters, measures, and sizes of each table in detail. This specific reference is a piece of evidence, not an accompanying text; therefore, due attention must be given to it.

 

• Figure 7 and Figure 8 (line 377 & line 379) and Figure 11 (line 465): Explain and describe in detail how these figures were obtained. What tool was used, how were the figures produced, and what exactly do they represent? Explain the content and background of each figure in detail. This specific reference is a piece of evidence, not an accompanying text; therefore, due attention must be given to it.

 

• In conclusion, in addition to the answers to the scientific questions, note the overall added value of the research, the difficulties in implementing the parts of the research, and finally, suggestions for future research work.

 

 

 

 

 

 

 

 

 

Author Response

Comments 1: Without ignoring the problem of length, we would suggest that all abbreviations be removed from the title and abstract. It is not customary to use abbreviations in titles, captions, and abstracts; these are for the main text. Even completely self-explanatory abbreviations should be replaced with reference words.

Response 1: Thank you for your valuable feedback on our manuscript. We have removed all abbreviations from the title and abstract as advised.

 

Comments 2: Excessive paragraphing is not necessary. For example, in lines 71 to 76, there is a paragraph with two sentences! Limiting three paragraphs per page or page section is sufficient. At least three paragraphs per page or page section are adequate.

Response 2: Thanks for your comment on paragraphing. We’ve adjusted by combining short paragraphs (like lines 83-88) to keep within 3 per page/section as suggested.

 

Comments 3: At the end of the introduction, the authors should accurately add the scientific questions of the article, which will be answered in the conclusions.

Response 3: Thank you for this valuable suggestion. We have revised the end of the introduction to clearly and accurately state the scientific questions of the article, which will be addressed in the conclusions.

 

Comments 4: After the scientific questions, the authors should list the individual chapters of the article with a precise description and justification.

Response 4: Thank you for your suggestion. Following your suggestion, after stating the scientific questions, we have added a section that lists each chapter of the article with precise descriptions and justifications for their structure and content.

 

Comments 5: Between the titles of the chapters and subchapters (for example, “2. Theoretical foundations and network structure and 2.1 Continuous Wavelet Transform”, line 111 – 112), a short introductory text should be written that concerns the separation of the parts of the section.

Response 5: Thanks for your guidance. We have revised the text to add brief introductory passages between chapter and subchapter titles.

 

Comments 6: Congratulations on the scientific organization of the mathematical parts and diagrams. However, we would like to point out a publishing issue. The typologies, formulas, and symbols should be uniformly adapted to the text. We do not know whether this is the work of the publisher or the authors. We impose uniformity in our texts, so we believe that the authors will be able to do the same.

Response 6: Thank you for pointing out the problems. We have carefully revised these elements to ensure consistency throughout the manuscript, adhering to the journal’s standards.

 

Comments 7: Pay attention to some parts of the text (kernels.The feature – line 175), in which the punctuation and format specifications of the magazine are not adhered to.

Response 7: Thank you for your feedback. We appreciate your constructive feedback and have carefully revised the manuscript to incorporate your suggestions.

 

Comments 8: The typology should be marked intratextually (e.g., typology "4"), and not "can be represented as:" line 156, or "as follows", line 226. Also, the marking of the typology (e.g., (4) line 258 should be placed uniformly aligned to the right of the typology and not below! The same on line 228, and wherever else there are typologies. These are basic issues of standardization and are not in dispute.

Response 8: Thank you for your careful review. We have revised the relevant parts in accordance with the journal's standards as suggested.

 

Comments 9: Between the titles of the chapters and subchapters (3.Analysis of experiments results, 3.1 Experimental environment and parameter settings, lines 267 and 268: experimental, or experiment not “experiments”), a short introductory text should be written that concerns the separation of the parts of the section.

Response 9: Thank you very much for your suggestions. We added a brief intro between the chapter and subchapter to clarify their separation, as suggested.

 

Comments 10: “Figure 6. Wavelet time-frequency plots for 10 health states at SNR of -8” and Figure 10. Wavelet time-frequency plots for four health states when SNR is -8: Explain and describe in detail how those images were obtained. What tool was used, how were the photos taken, and what exactly do they represent? Explain the content and background of each photo in detail. This specific reference is a piece of evidence, not an accompanying text; therefore, due attention must be given to it.

Response 10: Thank you for asking these questions. 

Figure 6 presents the wavelet time-frequency plots of 10 states (as specified in Table 2) of bearings after adding Gaussian noise with a signal-to-noise ratio (SNR) of -8 dB.

These plots are generated through the following process:  

First, the original vibration signals from the CWRU dataset are injected with -8 dB Gaussian noise using the numpy.random.normal function from Python's NumPy library to simulate extreme industrial noise environments. 

Then, the noisy one-dimensional signals are converted into two-dimensional time-frequency representations via Continuous Wavelet Transform (CWT) using the pywt.cwt function from the PyWavelets library, with the "morl" wavelet basis selected for its excellent time-frequency localization. 

Finally, the time-frequency matrices are visualized using Matplotlib's imshow function, with legends and coordinate axes hidden to avoid interfering with model classification, and the images are normalized and compressed into a 128×128×3 pixel format.

The 10 subplots (a-j) in Figure 6 correspond to the 10 states respectively, (a) the normal state shows uniform energy distribution with only sparse noise dots, (b-d) the rolling element fault states (BA_1-BA_3) exhibit weak energy clusters near the rolling element passing frequency (~237 Hz), which are blurred by noise, with slightly stronger energy in larger fault diameters, (e-g) the inner ring fault states (IR_1-IR_3) have faint periodic energy bursts around the inner ring characteristic frequency (~305 Hz), scattered into irregular patches by noise,(h-j) the outer ring fault states (OR_1-OR_3) display diffuse energy bands near the outer ring frequency (~162 Hz), mostly dominated by noise.

These plots visually demonstrate that -8 dB noise obscures most fault features, highlighting the necessity of the DDRSN-SKA model's enhanced anti-noise capability. 

Figure 10 presents the wavelet time-frequency plots of four health states of bearings after adding Gaussian noise with a signal-to-noise ratio (SNR) of -8 dB.

The generation process is as follows:

First, the original vibration signals corresponding to the four health states of the 6007ZM deep groove ball bearings used in the experiment (parameters: inner diameter 35 mm, outer diameter 62 mm, width 14 mm, number of rolling elements 11) are injected with -8 dB Gaussian noise using the numpy.random.normal function from Python's NumPy library to simulate extreme industrial noise environments.

Then, the noisy one-dimensional signals are converted into two-dimensional time-frequency representations via Continuous Wavelet Transform (CWT) using the pywt.cwt function from the PyWavelets library, with the "morl" wavelet basis selected (scale parameters 1-128) for its excellent time-frequency localization, corresponding to a frequency range of approximately 78 Hz-10 kHz.

Finally, the time-frequency matrices are visualized using Matplotlib's imshow function, with legends and coordinate axes hidden to avoid interfering with model classification, and the images are normalized and compressed into a 128×128×3 pixel format.

The four subplots (a-d) in the figure correspond to the four states respectively: (a) the normal state shows uniform energy distribution with only sparse noise dots; (b) the inner ring fault state has faint periodic energy bursts around the inner ring characteristic frequency (~280 Hz), scattered into irregular patches by noise; (c) the rolling element fault state exhibits blurred energy clusters near the rolling element passing frequency (~220 Hz); (d) the outer ring fault state displays diffuse energy bands near the outer ring frequency (~150 Hz), mostly dominated by noise.

These images visually demonstrate that -8 dB noise obscures most fault features, highlighting the necessity of the DDRSN-SKA model's enhanced anti-noise capability.

 

Comments 11: Tables 3, 4, 6, and 7: Explain and describe in detail how those tables were obtained. What tool was used, how were the photos taken, and what exactly do they represent? Explain the calibration, subdivisions, physical parameters, measures, and sizes of each table in detail. This specific reference is a piece of evidence, not an accompanying text; therefore, due attention must be given to it.

Response 11: Thank you for asking these questions. 

Tables 3 and 4 present the results of experiments conducted on the publicly available Case Western Reserve University (CWRU) dataset, providing a comprehensive quantitative comparison of the performance of five different deep learning models (ResNet, DRSN-CS, DRSN-Transformer, 1Dproposed, and the proposed DDRSN-SKA model) under strong noise conditions. The diagnostic accuracy data in these two tables were obtained by training and testing the five models separately in Python and calculating their accuracy under various conditions. Each row in these tables represents a specific model, and each column represents a noise level. The values in the tables precisely reflect each model's noise resistance capability on the standard dataset. 

Tables 7 and 8 present experiments conducted on a private laboratory dataset to quantitatively compare the performance of five different deep learning models (ResNet, DRSN-CS, DRSN-Transformer, 1Dproposed, and the proposed DDRSN-SKA model) under strong noise conditions. The diagnostic accuracy data in these two tables were obtained by training and testing the five models separately in Python and calculating their accuracy under various conditions. Each row in these two tables represents a specific model, and each column represents a noise level. The values in the table precisely reflect each model's noise resistance capability on the standard dataset.

 

Comments 12: Figure 7 and Figure 8 (line 377 & line 379) and Figure 11 (line 465): Explain and describe in detail how these figures were obtained. What tool was used, how were the figures produced, and what exactly do they represent? Explain the content and background of each figure in detail. This specific reference is a piece of evidence, not an accompanying text; therefore, due attention must be given to it.

Response 12: Thank you for asking these questions.

Figures 7 and 11 show the results of training the model using the Case Western Reserve University (CWRU) datase (corresponding to Figure 7) and the laboratory bearing dataset (corresponding to Figure 11), respectively. After model training was completed, the confusion matrix function in the Python library was used to compare the predicted results with the actual labels, and the results were visualized using the matplotlib and seaborn libraries in Python. A confusion matrix diagram was then plotted and automatically saved by the computer. The confusion matrix directly displays the relationship between the true labels and the model's predicted labels for all fault types in matrix form. Confusion Matrix Figure 7 is a 10×10 matrix (labels 0–9), corresponding to ten bearing states (as shown in Table 2 of the paper). The vertical axis of the matrix represents the true labels, the horizontal axis represents the predicted labels, the values on the main diagonal indicate the number of fault types correctly classified, and the remaining values indicate the number of fault types misclassified as other types. Confusion Matrix Figure 11 is a 4×4 matrix (labels 0–3), corresponding to four bearing fault states (as shown in Table 6 of the paper), illustrating the performance of each model. As can be clearly seen from the figure, although all models face the challenge of misclassifying minor inner ring faults as normal states, the model proposed in this paper (Figure 11d) has the fewest misclassifications across all categories. This visualization demonstrates the model's superior robustness and diagnostic accuracy, fully aligning with the highest quantitative accuracy rate of 97.50% reported in Table 7. 

Figure 8 shows the bearing dataset input into the model proposed in this paper. The t-SNE algorithm is used to reduce these high-dimensional vectors to two dimensions, and Matplotlib in Python is used for coloring and plotting based on the true labels of the samples. The figure clearly shows the clustering of fault types corresponding to each color, demonstrating the model's clustering effect and classification performance. In the Gaussian noise environment of Figure 8(a), features of different categories form distinct, independent clusters. The t-SNE visualization technique validates the model's feature learning effectiveness and feature discrimination capability in an -8 dB strong noise environment. This analysis reveals the separability of the internal feature space, complementing the quantitative accuracy results provided by the confusion matrix. In the Laplace noise environment shown in Figure 8(b), there is some overlap between feature clusters, which aligns with the increased classification difficulty and reduced accuracy rate to 91.77% observed in the confusion matrix. However, most categories remain effectively separable, strongly demonstrating that the model retains robust feature extraction and discriminative capabilities even under adverse noise conditions.

 

Comments 13：In conclusion, in addition to the answers to the scientific questions, note the overall added value of the research, the difficulties in implementing the parts of the research, and finally, suggestions for future research work.

Response 13: Thank you for your suggestion. We have revised the conclusion to include, alongside answers to the scientific questions, the research’s overall added value, implementation difficulties, and suggestions for future work.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

From the beginning of our undertaking of this particular journal, we appreciated its value and provided the best comments to highlight this value. Fortunately for all of us, the authors, as experienced and reliable researchers, fully understood the positivity of our perspective and restored 100% of our comments. The feedback was extremely successful. They deserve congratulations. Well done for the comments, and good luck with their future research work.

Do not be afraid to add to the text any material that emerged from the corrections, especially the answers to comments 10, 11, and 12. Typically are good answers constitute important elements for the proven material of the manuscript, and they are of interest to readers.