Knock Detection with Ion Current and Vibration Sensor: A Comparative Study of Logistic Regression and Neural Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Title: Knock detection with ion current and vibration sensor: A Comparative study of logistic regression and neural networks

Authors: O. Björnsson, P. Tunestål

The authors propose using vibration sensors and ion currents to enhance the accuracy of knock detection in spark ignition engines. They also compare the logistic regression and neural network methodologies. The results indicate that the CNN model achieves the best performance. However, I cannot recommend the publication in Agriculture in its current form. Several criticisms arise, which are as follows.

(1) In the abstract, the authors state, “The dual-input logistic regression and CNN models reduced cylinder variation.” From the reviewer’s point of view, this sentence is somehow misleading and should be revised. Readers may misunderstand how the models can “reduce the cylinder-to-cylinder variations.”

(2) The engine specification presented in Table 1 is insufficient. More details should be included, such as rated power and torque, and speed range.

(3) The experimental setup and typical raw sample data should be provided in the paper.

(4) More information about the machine learning models utilized in the study should be included. For clarity and brevity, at least some flowcharts should be added.

(5) The definition of the abbreviations ROC and AUC in the paper is informal and should be revised. Furthermore, the calculation procedure for these two key parameters should be included.

(6) Figs. 3-5 present similar curves for different cylinders and should be grouped together.

Comments on the Quality of English Language

Please see the comments above.

Author Response

Comments 1: In the abstract, the authors state, “The dual-input logistic regression and CNN models reduced cylinder variation.” From the reviewer’s point of view, this sentence is somehow misleading and should be revised. Readers may misunderstand how the models can “reduce the cylinder-to-cylinder variations.”

Response 1: We appreciate the reviewer’s feedback regarding this sentence's clarity. We have revised the wording to specifically refer to the variation in model performance across cylinders, as follows (page 1, lines 10–12):

"The dual-input logistic regression and CNN models demonstrated reduced cylinder-to-cylinder variation in classification performance, providing more consistent knock detection accuracy across all cylinders."

 

Comments 2: The engine specification presented in Table 1 is insufficient. More details should be included, such as rated power and torque, and speed range.

Response 2: We appreciate the reviewer’s suggestion to provide additional engine specifications. As this is a research engine, it does not have an official rated power; instead, operational guidelines specify safe ranges for its operation. Based on these guidelines, we have now included the maximum torque and speed range in Table 1, found on page 3.

 

Comments 3: The experimental setup and typical raw sample data should be provided in the paper.

Response 3: We appreciate the reviewer’s suggestion. In response, we have added a figure of the experimental setup in Section 2, which has been renamed “Engine Specifications & Experimental Setup.” Section 2 begins on page 3, while the corresponding figure is placed on page 4. Additionally, we have introduced a new section, “Data Characteristics” (Section 7, starting at the bottom of page 9), that includes typical raw sample data along with explanatory details.

 

Comments 4: More information about the machine learning models utilized in the study should be included. For clarity and brevity, at least some flowcharts should be added.

Response 4: We appreciate the reviewer’s feedback. In response, we have added a schematic representation of the fully connected neural network in Section 6.2 (Figure 2, page 7), and a flowchart illustrating the dual-input CNN process in Section 6.3 (Figure 3, page 8). Additionally, we rewrote and expanded Section 6.2 (end of page 6 start of page 7, lines 223-248) to include further details about the network architecture.

 

Comments 5: The definition of the abbreviations ROC and AUC in the paper is informal and should be revised. Furthermore, the calculation procedure for these two key parameters should be included.

Response 5: We agree with the reviewer’s suggestion. In response, we revised the sentence introducing the abbreviations in the first paragraph of Section 6.1, “Logistic Regression”  (page 6, lines 197-199) for clarity. The revised sentence is as follows:

“In this study, we employ logistic regression to formalize this process, offering a more interpretable analysis of knock detection by evaluating model performance using the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve.”

Additionally, we included formal definitions of ROC and AUC in Section 6.5, “Receiver Operating Characteristic” (page 9), where we now also provide the detailed calculation procedure for these metrics.

 

Comments 6: Figs. 3-5 present similar curves for different cylinders and should be grouped together.

Response 6: We appreciate the reviewer’s suggestion. We believe the comment may refer to Figures 2–4 (Figures 5-8 in the revised version), as these figures present ROC curves from different models rather than different cylinders. Figures 5 and 8 do indeed show the same cylinder, as it was the best and worst performing cylinder for each respective model. While we understand the rationale for grouping the figures, we believe that presenting each figure within its respective section better maintains the flow and allows readers to view relevant results in context. For these reasons, we have opted to keep the figures separated.

Reviewer 2 Report

Comments and Suggestions for Authors

Paper considers knock detection methodology for SI engines. Paper well structured and organized. All illustrations correspond to paper contents. There are few comments about paper:

1. It is important to provide character pressure signals for knocking and not knocking cases and provide examples how new methodology process it.

Author Response

Comments 1: It is important to provide character pressure signals for knocking and not knocking cases and provide examples how new methodology process it.

Response 1: Thank you for this feedback. We agree and have added Section 7, “Data Characteristics” starting at the bottom of page 9, which details the different signals for knocking and non-knocking cases, along with a brief explanation of how the CNN might process these signals.

Reviewer 3 Report

Comments and Suggestions for Authors

It is a good paper in my opinion.  My suggestion is Authors should extend conclusion and references sections in the future.

Author Response

Comments 1: My suggestion is Authors should extend conclusion and references sections in the future.

Response 1: Thank you for your suggestion. We made slight extensions to the second paragraph of the conclusion (page 20, lines 610-618). Regarding the references section, we would appreciate any specific suggestions or areas you felt were underrepresented so we can ensure a more thorough citation of relevant literature. Additionally, if there are particular aspects of the conclusion that you believe require further expansion or elaboration, we would be grateful for your input.

Reviewer 4 Report

Comments and Suggestions for Authors

Summary:

The authors conducted an analysis of the selection of the optimal solution for knock detection in SI engines. Models based on logistic regression, FCN and CNN were tested. As input to the models, the authors created their own data set based on vibration sensor readings and ion current measurements. The models were compared through an ROC curve including AUC. In addition, the linear regression model was compared to the CNN by using confusion matrix and swarm plots. The best metrics were obtained by dual-input CNN. The authors also provided an insightful discussion of the aspects involved in their solution.

Comments:

In line 103, I would suggest giving the abbreviation in parentheses, like this: “compressed natural gas (CNG)”. From a reader's perspective, the occurrence of an abbreviation in a table 1 that has not been previously explained may reduce the readability of the article.

I would recommend a more detailed description of the architecture of  final CNN dual-input model. It is not clear from the content of the article exactly what dual-input means and how it is operationalized in the architecture. While it is quite clear in the case of logistic regression, a number of different solutions can be used in the CNN architecture, such as processing the input tensors through 2 separate branches and then concatenating them before the FC layer, or using 2 different data and combining them as one input tensor to a single branch.

Author Response

Comments 1: In line 103, I would suggest giving the abbreviation in parentheses, like this: “compressed natural gas (CNG)”. From a reader's perspective, the occurrence of an abbreviation in a table 1 that has not been previously explained may reduce the readability of the article.

Response 1: Thank you for pointing that out. We have now included the abbreviation in the text and added it to the list of abbreviations to improve readability.

 

Comments 2: I would recommend a more detailed description of the architecture of  final CNN dual-input model. It is not clear from the content of the article exactly what dual-input means and how it is operationalized in the architecture. While it is quite clear in the case of logistic regression, a number of different solutions can be used in the CNN architecture, such as processing the input tensors through 2 separate branches and then concatenating them before the FC layer, or using 2 different data and combining them as one input tensor to a single branch.

Response 2: The final paragraph of Section 6.3, Convolutional Neural Network, conveys this information in non-technical terms. We intentionally avoided technical terminology, such as ‘tensors,’ to ensure clarity for readers without a machine learning background. However, we agree that clarification is needed and have revised the section to clarify how the tensors are concatenated for the dual-input model. We also included a flowchart to clarify the process further. The revised paragraph can be found on page 8, lines 273-283; the flowchart is just below this paragraph.