Knot-TPP: A Unified Deep Learning Model for Process Incidence and Tool Wear Monitoring in Stacked Drilling

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Based on the scope, methodology, and limitations presented in the manuscript, the work demonstrates potential for contributing to the field of intelligent tool wear monitoring and process incidence prediction in composite-metal stack drilling. The proposed knot-TPP model appears to provide a unified framework for these tasks and shows promising adaptability across different sampling durations and frequencies. However, several critical concerns must be addressed before the manuscript can be considered for publication as listed below. Given these substantive issues affecting both the technical rigor and clarity of the manuscript, a major revision is required. 

• The Kistler 8152B AE sensor, depending on the specific variant, supports a minimum stable frequency measurement of either 50 kHz or 100 kHz. Using a sampling frequency of only 100 kHz is likely insufficient to capture the full spectrum of meaningful AE signals, especially since the data was further subsampled to a maximum of 2000 Hz. Please justify the choice of this sensor-sampling setup and discuss its impact on the completeness and quality of the acquired data.

• The manuscript should clarify the difference between sampling frequency and sample duration.

• In Figure 5, the orientation of the accelerometer, specifically the direction of the Y-axis, is unclear. Please explain the Y-axis direction and why it was chosen for monitoring.

• The influence of the dynamometer structure on the acquired vibration and AE data has not been addressed. Please elaborate on how the dynamometer design might affect the signals and what implications this has for applying the method in different setups.

• Please include images showing the cutting edges of both fresh and worn drill bits to provide visual confirmation of the tool wear states discussed in the study.

• It is not stated whether any preprocessing was applied to the acquired data before inputting it into the AI model. Please describe any preprocessing steps used and include this information in the manuscript.

• Provide graphs of the sampled data. Based on the way the data and results are currently presented, it appears that the insights in Section 4.1 might be obtainable directly from raw data, without the need for the developed AI model. Please clarify the unique added value the AI model brings to the analysis.

• The statement that tool wear assessment relies solely on signals captured during drilling of CFRP and aluminum layers requires clarification. How were these data segments identified? Was an automated technique developed to extract them? Please elaborate.

• Reference 6 (Sadek et al.) presents a cyber-physical adaptive control (CPAC) system that integrates real-time tool wear detection with advanced process modeling to enhance tool life and productivity. This CPAC system is detailed in:
  Sadek, A., Hassan, M., Attia, M.H.: A new cyber-physical adaptive control system for drilling of hybrid stacks. CIRP Ann. 69(1), 105–108 (2020)
  As both works address tool wear monitoring and process optimization in drilling hybrid stacks, please benchmark your findings against the approach presented by Sadek et al. and highlight the differences, advantages, or improvements offered by your method.

• The authors are encouraged to address the methodological concerns, enhance the completeness of the experimental and model validation details, and strengthen the discussion to clearly demonstrate the novelty and value of their approach in comparison to existing methods.

Author Response

We thank you for your time and effort in reviewing our manuscript. The feedback has been of great value to improve the content and presentation of the paper. We have revised our manuscript according to all the reviewers’ comments, and our point-by-point responses are given below. We have included a document that keeps track of the modifications made in response to the reviewers’ comments, which is attached at the end, with the added text highlighted in green and deleted sections highlighted in red. A clean file of the manuscript (without any highlighting) is also attached.

 

1. The Kistler 8152B AE sensor, depending on the specific variant, supports a minimum stable frequency measurement of either 50 kHz or 100 kHz. Using a sampling frequency of only 100 kHz is likely insufficient to capture the full spectrum of meaningful AE signals, especially since the data was further subsampled to a maximum of 2000 Hz. Please justify the choice of this sensor-sampling setup and discuss its impact on the completeness and quality of the acquired data.

Authors’ Response: We appreciate the reviewer's insightful comment and fully agree that energy waves below 50 kHz should not be referred to as acoustic emission (AE). The manuscript was revised accordingly to reflect this understanding. As referenced in the paper,  (https://doi.org/10.1016/j.ymssp.2025.112499)”, it is demonstrated that eight times the sampling frequency is sufficient to represent the events of process incidence for both thrust and torque. To maintain consistency with the dimensionality of the signals associated with thrust and torque, a sampling frequency of 2000 Hz was selected.

In the context of signal processing for the deep learning model, the raw signal is directly utilised. Choosing 50 kHz would result in an excessively high-dimensional input, making it impractical for the model. As highlighted in many studies, the energy waves 2000 Hz still encapsulate valuable information for tool condition monitoring. Although the complete spectrum of the energy wave cannot be reconstructed with perfect fidelity, the lower-frequency components still carry significant information that supports process incidence identification. This is further substantiated in “(https://doi.org/10.1007/s00170-024-14867-z)” where even lower sampling frequencies for acceleration in the y-axis were shown to be effective.

 

2. The manuscript should clarify the difference between sampling frequency and sample duration.

Authors’ Response: Within the context of the manuscript, sampling frequency is how often samples are taken (Hz) and sample duration is the total time recorded, which is consistent with standard terminology in signal processing. We have made changes to the text where we felt the original wording could have caused some confusion.

3. In Figure 5, the orientation of the accelerometer, specifically the direction of the Y-axis, is unclear. Please explain the Y-axis direction and why it was chosen for monitoring.

Authors’ Response: A coordinate system has been added to Figure 5 to clearly illustrate the orientation of the accelerometer in relation to the experimental setup. The selection of the Y-axis acceleration is based on findings from previous experiments, which showed frequent signal saturation in both the X and Z axes, whereas the Y axis remained unsaturated. Consequently, the acceleration signal from the Y-axis-oriented accelerometer was chosen.

4. The influence of the dynamometer structure on the acquired vibration and AE data has not been addressed. Please elaborate on how the dynamometer design might affect the signals and what implications this has for applying the method in different setups.

Authors’ Response: According to the datasheet for the dynamometer provided by Kistler, the natural frequency is stated with 3.1 kHz. Attaching the sample holder with workpiece to the dynamometer will definitely reduce the natural frequency of the setup, but it is reasonable to assume that the reduced natural frequency of the system is still significantly higher than the tool engagement frequency of 132 Hz. It is therefore safe to say that the structure of the dynamometer including sample holder and workpiece did not affect the signals.

5. Please include images showing the cutting edges of both fresh and worn drill bits to provide visual confirmation of the tool wear states discussed in the study.

Authors’ Response: Figure 6 on Page 9 presents images of a single cutting edge for both new and worn drills. To enhance visibility of the flank wear along the cutting edges for the reader, the worn drills were tilted to align the cutting edges parallel to the viewing plane.

6. It is not stated whether any preprocessing was applied to the acquired data before inputting it into the AI model. Please describe any preprocessing steps used and include this information in the manuscript.

Authors’ Response: Apart from subsampling to reduce the burden on deep learning computation as a result of excessive dimension, no other preprocessing steps were applied. This was clarified in subsection 2.3 on Page 6.

7. Provide graphs of the sampled data. Based on the way the data and results are currently presented, it appears that the insights in Section 4.1 might be obtainable directly from raw data, without the need for the developed AI model. Please clarify the unique added value the AI model brings to the analysis.

Authors’ Response: Graphs containing samples of the recorded data have been included. As discussed in the introduction, a single feature alone cannot fully capture the occurrence of a process incidence, making direct prediction from raw data ineffective. This limitation necessitates the implementation of more complex rules and the extraction of patterns from the signals. Traditional feature extraction methods, such as machine learning models dependent on feature engineering, often require substantial prior knowledge and expertise. Moreover, manual feature extraction is prone to human bias, which can lead to incomplete coverage or redundancy.

In this study, both tool wear and process incidence are predicted, presenting significant challenges for feature engineering to predict two separate objectives using one set of . The integration of AI models enables direct processing of raw signals without the need for manual preprocessing or feature extraction, allowing for adaptive and accurate predictions for multiple objectives. This approach enhances both prediction accuracy and reliability, as compared with other conventional machine learning model in Table 2.

8. The statement that tool wear assessment relies solely on signals captured during drilling of CFRP and aluminum layers requires clarification. How were these data segments identified? Was an automated technique developed to extract them? Please elaborate.

Authors’ Response: In the identification process, the process incidences are continuously predicted from the corresponding signal segments. As the model continuously processes the signal, the input—representing a segment window—slides over time, allowing it to determine the exact material the drill bit is currently engaged in. Consequently, when the process incidence prediction indicates the cutting of CFRP or Al, this information is passed as an embedded vector input to the tool wear prediction module. This integration, achieved in the form of knot structure of the proposed model, provides valuable context for enhancing prediction accuracy. This clarification has been further emphasized in the last paragraph of Section 2.4 on Page 6.

9. Reference 6 (Sadek et al.) presents a cyber-physical adaptive control (CPAC) system that integrates real-time tool wear detection with advanced process modeling to enhance tool life and productivity. This CPAC system is detailed in:
   Sadek, A., Hassan, M., Attia, M.H.: A new cyber-physical adaptive control system for drilling of hybrid stacks. CIRP Ann. 69(1), 105–108 (2020)
   As both works address tool wear monitoring and process optimization in drilling hybrid stacks, please benchmark your findings against the approach presented by Sadek et al. and highlight the differences, advantages, or improvements offered by your method.

Authors’ Response: The differences in the structure of the hybrid stack (sandwich CFRP/Al/CFRP vs CFRP/Al) and workpiece material might explain the differences in wear progression: their tool wear curve exhibits quasi-linearity, while ours follows a typical S-curve. Consequently, a direct comparison between the two approaches is not straightforward. Nonetheless, both approaches achieve similar outcomes in terms of root mean square errors, with the one in our case being slightly smaller (5.09 microns vs 6 microns and 12 microns respectively). Consequently, we feel that our paper is a valuable addition to and an extension of the knowledge provided by the paper authored by Sadek, Hassan and Attia. This comparison is further detailed in Table 4 on Page 13.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

attached

Comments for author File: Comments.pdf

Author Response

We thank you for your time and effort in reviewing our manuscript. The feedback has been of great value to improve the content and presentation of the paper. We have revised our manuscript according to all the reviewers’ comments, and our point-by-point responses are given below. We have included a document that keeps track of the modifications made in response to the reviewers’ comments, which is attached at the end, with the added text highlighted in green and deleted sections highlighted in red. A clean file of the manuscript (without any highlighting) is also attached.

 

1-The abstract needs to give more information. The current version only explains some results, and it would be better if quantitative results supported the explanations.

Authors’ Response: As suggested by the reviewer, we have incorporated quantified statistics in the abstract to further support the explanations already provided.

 

2- In the introduction section, discuss the formation of knots, their physical characteristics, and challenges when it comes to drilling.

Authors’ Response: We appreciate the reviewer's interest in the knot structure. The term knot structure refers to the deep learning architecture proposed in this study, designed to enable the unified prediction of two distinct objectives—tool wear and process incidences—within the drilling process. It is a conceptual framework unique to this model and does not represent a physical structure or possess any tangible attributes.

 

3-In the introduction section, mention the application of TPP techniques.

Authors’ Response: The original TPP architecture was initially developed for video-based action recognition rather than signal processing, underscoring the necessity for its adaptation and application in the context of tool condition monitoring. In response to the reviewer's suggestion, the discussion of the TPP architecture has been relocated from the Methodology section to the Introduction (Section 1) for better context and clarity.

 

4-The authors note that prediction variability increases with significant tool wear, but no clear mitigation strategies (e.g., ensemble methods, noise-robust techniques) are tested.

Authors’ Response: Since the model directly processes raw machining signals, excessive noise can affect prediction quality. However, its overall impact on final predictions remains limited. Despite this variability, the knot-TPP model consistently demonstrates high accuracy, suggesting that additional noise mitigation strategies are not currently necessary. Analysis across different combinations of signal duration and sampling frequency indicates that extending the sampling duration while maintaining the minimum viable frequency can further enhance the accuracy of tool wear prediction. This has been addressed in Section 4.3 on Page 14.

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I am satisfied with the authors' response.