Enhancing Interpretability in Drill Bit Wear Analysis through Explainable Artificial Intelligence: A Grad-CAM Approach

Round 1

Reviewer 1 Report

Nicely written paper, good flow, background information and state of the art.

113-114: Highlighting the input regions that are estimated to contribute most to an output is only one XAI technique/approach. I  encourage authors to clarify and change the general statement.

116-118: "In this study we propose a novel approach that integrates....": Grad-Cam together with CNN is not a new approach, what is new is their application to drill bit wear detection. This should be clarified in the text.

121-123: I propose to use "can enable" instead of "enables" since there are many ifs that need to be evaluated. For instance, training data which does not capture certain situations, etc. In general throughout the paper, I propose the statements should be fitted to the case and not be given as strong general statements, since their validity outside the use case presented was not assessed. For example one could state: "in the use case addressed in the paper it was observed an increased interpretability, ...".

148: typo lth filter

253-254: not necessarily, one could use a different approach. But I assume what was meant is that that there is a need to employ an approach that can help in assessing the cause. It will improve the readibility of the paper to re-phrase.

313-330: A description of how the unseen data used for the validation in the should be provided together with a discussion: how far away from the training data, etc.

347-349: Same comment as before, I propose to not use general statements and related to the actual case.

360-367: In this case, but it is always the case? In general, couldn't these be triggered due to specific context of the operation and not necessarily related to bit failures which would lead to false positives?

Author Response

Please see the attachment. Thank you

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper combines signal analysis with artificial intelligence to analyze drill bit wear, which has certain application value and feasibility.

1. Different drill bits, each with distinct conditions, were used for data collection. After analyzing the signals collected under different drill bits and conditions, it is possible that some parameters obtained due to different conditions may compensate for the difference between normal bit and abnormal bit, affecting judgment.

2. “Real-world drilling datasets” are significantly influenced by formation lithology, drill string combination and length, and circulating media, which can lead to different data signals. How are these considered to distinguish them in models or algorithms. Different drilling environments can lead to different information in the signal. Can these factors potentially mask the difference (damage) between the two types of drill bits.

3. The pictures in the paper are not clear.

Normal

Author Response

Please see the attachment. Thank you

Author Response File: Author Response.pdf

Reviewer 3 Report

1. The research deals with the analysis of vibrations in the drilling process. Further, these vibrations, with the help of a hybrid model based on CNN and CAI, model the relevant vibration frequencies, which generate the state of the tool (drill bit) in terms of wear.

2. The article itself is original because it is about experimental data that the authors measured. The originality is in the model. CNN modeling is well-known, however, frequency classification using the XAL algorithm is original.

3. The experiment is original and has not been published so far. Therefore, the model was created for the first time on the original data.

4. As I wrote in the first analysis of this article. I need an explanation for the range of frequencies from 4000 to 6000Hz, which are not defined in figure 10. And I ask the authors for an explanation.

5. I agree with the conclusions. Possibly another example can be presented as in the discussion. For the frequency range where drill bit wear is detected.
6. References are appropriate.

7. Figure 10 should be clarified, for the frequency range of 4-6000Hz.

Author Response

Please see the attachment. Thank you

Author Response File: Author Response.pdf