Machine-Learning-Based Classification of Electronic Devices Using an IoT Smart Meter

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

1. Novelty: The research question is relatively original, as mentioned in Section 3 ("This work is an extension of the study titled..."). The article addresses the challenge of detecting and classifying electronic devices within smart electric networks. The results contribute to advancing current knowledge in this area.

2. Scope: The article aligns well with the scope of the journal.

3. Significance:

The "Abstract" section should be expanded to include a brief summary of the quantitative results for the various implementations and ML models across different scenarios.

The explanation of dataset manipulation and pre-processing in Section 4.3.2 is thorough. However, in the section titled "Analysis of Experiment Outcomes," it would be beneficial to include a separate subsection focused on evaluating the accuracy of predictive models. Specifically, this could cover performance metrics such as Root Mean Square Error (RMSE) and R² value, which would provide a clearer picture of the reliability of the different predictive models employed.

4. Quality: There are some minor writing and grammatical errors throughout the paper. Notably:

In the Introduction, the organization of paragraphs in Chapter 3 is unclear, and Chapter 3 itself is not explicitly referenced.

Abbreviations such as SM, IAIoSGT, and MQTT should be defined the first time they are mentioned in the text.

Section 4.2, "ML Models in Extreme Edge Devices," appears to be written in a language other than English, as does the description of Figure 15.

Section 4.3, "Figure 7. Testbed," not "Figure 7. Testbad."

In Section 5.2.2, "KNN on the ESP32," a table number is missing where the text references (as shown in Table ???).

Section 4.3.1, "Data Acquisition," could benefit from the inclusion of a table summarizing the key features used in the various datasets.

5. Scientific Soundness: The study’s design and technical execution require improvement:

In Sections 3.1 and 3.2, some concepts are repeated, and the writing should be revised to avoid redundancy.

Sections 4.2.1 and 4.2.2 would be stronger if the authors provided detailed descriptions of the architectures of the ML models used, such as the number of layers and neurons in each model. Additionally, it would be helpful to provide algorithmic steps for KNN, as well as for the other ML models discussed. These should be contextualized within the scope of the article and the proposed solution, rather than simply providing general definitions.

In Section 5.1, summarizing the discussion in a table could more effectively present the different ML models used in the study.

Comments on the Quality of English Language

English Level: The use of English is generally appropriate; however, the authors should avoid including non-English text in certain sections, as highlighted above.

Author Response

Comments 1: [The research question is relatively original, as mentioned in Section 3 ("This work is an extension of the study titled..."). The article addresses the challenge of detecting and classifying electronic devices within smart electric networks. The results contribute to advancing current knowledge in this area. The article aligns well with the scope of the journal.]

Response 1: [Thank you for your positive feedback. We appreciate your recognition of the originality of our research and its contribution to the field of smart electric networks. Our goal is to advance knowledge in detecting and classifying electronic devices within these networks, and we are pleased that our work has been recognized in this regard. Your insights and review have been invaluable in refining our manuscript.]

Comments 2: [The "Abstract" section should be expanded to include a brief summary of the quantitative results for the various implementations and ML models across different scenarios. The explanation of dataset manipulation and pre-processing in Section 4.3.2 is thorough. However, in the section titled "Analysis of Experiment Outcomes," it would be beneficial to include a separate subsection focused on evaluating the accuracy of predictive models. Specifically, this could cover performance metrics such as Root Mean Square Error (RMSE) and R² value, which would provide a clearer picture of the reliability of the different predictive models employed.]

Response 2: [We would like to thank you for your insightful suggestions, which have contributed to improving the manuscript. In response to your feedback, we have made the following modifications:

1. Abstract: The Abstract has been expanded to include a summary of the quantitative results, as suggested. The new version is as follows:

This study investigates the implementation of artificial intelligence algorithms on resourceconstrained edge devices, including ESP32 and Raspberry Pi, for smart grid applications. By employing Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) models, the research focuses on classifying IoT electronic devices within the IAIoSGT architecture. A literature review identifies critical gaps in existing methodologies, while experimental evaluations emphasize the pivotal role of data preprocessing, particularly normalization, in optimizing model performance. Results demonstrate distinct performance patterns between the MLP and KNN models across hardware platforms, with normalized data significantly influencing accuracy outcomes. Comparative analyses reveal inherent trade-offs between algorithmic complexity, preprocessing requirements, and hardware limitations. These findings underscore the potential for deploying lightweight AI models on extreme edge devices in smart grids and propose scalable strategies for enhancing IoT integration in energy systems.

1. Analysis of Experiment Outcomes: A new subsection titled "Quantitative Assessment of the Accuracy of Classification Models" has been added. In this subsection, we present a detailed quantitative analysis using metrics such as precision, recall, F1-score, and accuracy. Considering that the implemented models address classification problems (and not regression), these metrics were chosen to quantify the results more precisely.

Thank you again for your valuable feedback, which has certainly helped to enhance the completeness and robustness of our work.]

Comments 3: [There are some minor writing and grammatical errors throughout the paper. Notably: In the Introduction, the organization of paragraphs in Chapter 3 is unclear, and Chapter 3 itself is not explicitly referenced.]

Response 3: [We have addressed this issue by revising the last paragraph of the Introduction to include a description of Section 3 as follows:

Section 3 presents the proposed solution, detailing the implementation of artificial intelligence models for detecting and classifying electronic devices in smart grids. It describes the system architecture, data acquisition process, and the methodology applied to ensure efficient operation in extreme edge environments.]

Comments 4: [Abbreviations such as SM, IAloSGT, and MQTT should be defined the first time they are mentioned in the text.]

Response 4: [We have carefully reviewed the manuscript and ensured that all abbreviations - such as SM (Smart Meter), IAIoSGT (Artificial Intelligence in Intemet of Smart Grid Things), and MQTT (Message Queuing Telemetry Transport) - are properly defined upon their first occurrence in the text. We appreciate your suggestion and believe this adjustment improves the clarity and readability of the manuscript.]

Comments 5: [Section 4.2, "ML Models in Extreme Edge Devices," appears to be written in a language other than English, as does the description of Figure 15.]

Response 5: [We have made the necessary adjustment, and the description of Figure 15 has been revised to "Confusion Matrix of the MLP with Unnormalized Data." This ensures consistency with the rest of the manuscript.]

Comments 6: [Section 4.3, "Figure 7. Testbed," not "Figure 7. Testbad."]

Response 6: [We have fixed this typo in Section 4.3, changing "Figure 7. Testbad." to "Figure 7. Testbed" accordingly.]

Comments 7: [In Section 5.2.2, "KNN on the ESP32," a table number is missing where the text references (as shown in Table ???).]

Response 7: [Thank you for your valuable feedback. The missing table number in Section 5.2.2 has been corrected. The reference now properly cites Table 2. This adjustment ensures clarity and consistency in the manuscript.]

Comments 8: [Section 4.3.1, "Data Acquisition," could benefit from the inclusion of a table summarizing the key features used in the various datasets.]

Response 8: [Thank you for your suggestion. To enhance the clarity and comprehensiveness of Section 4.3.1, a new table summarizing the database has been included, which overviews the devices utilized during data acquisition. The modification can be found in Table 1, along with the text added to introduce the respective table:

To provide a clearer overview of the dataset used in this study, Table ?? summarizes the key features extracted from the acquired data. These features were used to classify the electrical and electronic devices based on their consumption signatures.]

Comments 9: [Scientific Soundness: The study's design and technical execution require improvement: In Sections 3.1 and 3.2, some concepts are repeated, and the writing should be revised to avoid redundancy.]

Response 9: [Thank you for your insightful feedback. We have revised Sections 3.1 and 3.2, adjusting several paragraphs to eliminate redundancy and improve clarity. These modifications ensure a more concise and structured presentation of the concepts while maintaining technical accuracy. We appreciate your suggestion and believe these refinements enhance the overall readability of the manuscript.

In section 3.1: For the Raspberry Pi 3B+, the classification process leverages the pre-trained MLP and KNN models, reducing the local processing load and ensuring consistency in the use of training data. On the ESP32, the KNN is implemented using the ArduinoKNN library, with weights configured locally at each initialization. However, periodic synchronization with the cloud ensures the retrieval of preprocessed data necessary for classification.

In section 3.2: After acquisition, the raw signals are transmitted to the cloud via MQTT, where they undergo preprocessing steps such as filtering and normalization. These transformations prepare the data for training the Machine Learning models (MLP and KNN). The complementary characteristics of these models justify their use: MLP generalizes well, handling complex patterns robustly, while KNN enables fast classifications with moderate resource consumption.

To optimize classification performance, the Raspberry Pi utilizes the pre-trained MLP model, updated in the cloud and deployed to the device. On the other hand, the ESP32 runs the KNN model locally, leveraging the ArduinoKNN library, where weights are reconfigured at each initialization to optimize the device's limited resources.]

Comments 10: [Sections 4.2.1 and 4.2.2 would be stronger if the authors provided detailed descriptions of the architectures of the ML models used, such as the number of layers and neurons in each model. Additionally, it would be helpful to provide algorithmic steps for KNN, as well as for the other ML models discussed. These should be contextualized within the scope of the article and the proposed solution, rather than simply providing general definitions.]

Response 10: [We appreciate your comment! Our goal in sections 4.2.1 and 4.2.2 was to provide an overview of the models, especially for readers who are not yet familiar with them. However, considering your suggestion, we have added two new figures to enhance the explanation:

Two figures in section 5.1.1, illustrating the layers and the number of neurons used in the models.

And in section 5.3.1, the explanation of KNN was expanded.

The experiment systematically evaluated different values of k (k=1, k=3, k=5, and k=7) and two types of weighting schemes (uniform and distance-based) in the KNN classifier. The goal was to assess the impact of these parameters on the model's classification performance and computational efficiency.

For each configuration, the model's accuracy was measured on the test set to determine how well it classified new data points. Additionally, the execution time for both training and prediction was recorded, providing insights into the computational cost associated with different values of k and weighting strategies.

To further analyze the model's effectiveness, confusion matrices were generated for each tested configuration. These matrices provided a visual representation of how well the model distinguished between different classes, highlighting correctly classified instances as well as misclassifications. This allowed for a deeper evaluation of the classifier's strengths and potential weaknesses.

Figure 20 presents the confusion matrix for the KNN algorithm with k=1, applied to non-normalized data. The results clearly demonstrate the model's performance, with only a single instance being misclassified. This suggests that for this specific dataset and preprocessing condition, KNN with k=1 achieved high accuracy. However, further analysis is required to determine the stability of this performance across different parameter settings and data distributions.

We hope these additions make the content clearer.]

Comments 11: [In Section 5.1, summarizing the discussion in a table could more effectively present the different ML models used in the study.]

Response 11: [We appreciate your suggestion to summarize the discussion on different ML models in a table in Section 5.1. However, we would like to point out that the synthesis of the models used in the study is already provided in Table 3, titled "KNN and MLP Performance on ESP32 and Raspberry Pi 3B+ for Normalized and Non-Normalized Data", which can be found in Section 6 (Conclusion and Future Work). This table consolidates the key performance metrics of the models, offering a comprehensive comparison.]

 

Reviewer 2 Report

Comments and Suggestions for Authors

This paper mainly presents the ideas to apply Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) machine learning models for classification of the electrical and electronic devices installed in Smart Grid, based on the current, voltage signatures of different devices installed in Smart grid.  The ideas are implemented in both ESP32 and Raspberry Pi platforms. The paper clearly presents the classification accuracy and the response time with normalized data and non-normalized data of two ML models and two implementation platforms. This paper is a nice application of ML and edge computing for smart grid. It is meaningful to precisely for the development of Smart Grid.

 

The research in this paper are further developments based on the existing research.  The conclusions in this paper have practical values for the possible engineering applications of MLP and KNN machine learning models, especially the detailed the comparisons. All references are appropriate. 

Besides, it is suggested to slightly revise the abstract to highlight the detailed MLP and KNN machine learning models applied. 

Comments on the Quality of English Language

 

There are some editorial and English mistakes, such as Figure 7, Figure 15, section 4.2.  Suggest the authors review this paper in detail to correct some mistakes.

Author Response

Comments 1: [This paper mainly presents the ideas to apply Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) machine learning models for classification of the electrical and electronic devices installed in Smart Grid, based on the current, voltage signatures of different devices installed in Smart grid. The ideas are implemented in both ESP32 and Raspberry Pi platforms. The paper clearly presents the classification accuracy and the response time with normalized data and non-normalized data of two ML models and two implementation platforms. This paper is a nice application of ML and edge computing for smart grid. It is meaningful to precisely for the development of Smart Grid.

The research in this paper are further developments based on the existing research. The conclusions in this paper have practical values for the possible engineering applications of MLP and KNN machine learning models, especially the detailed the comparisons. All references are appropriate.

Besides, it is suggested to slightly revise the abstract to highlight the detailed MLP and KNN machine learning models applied.]

Response 1: [We appreciate the valuable feedback regarding the abstract. Following the recommendations, we have revised the abstract to explicitly highlight the MLP and KNN machine learning models applied in our study. The updated abstract is as follows:

This study investigates the implementation of artificial intelligence algorithms on resource-constrained edge devices, including ESP32 and Raspberry Pi, for smart grid applications. By employing Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) models, the research focuses on classifying IoT electronic devices within the IAIoSGT architecture. A literature review identifies critical gaps in existing methodologies, while experimental evaluations emphasize the pivotal role of data preprocessing, particularly normalization, in optimizing model performance. Results demonstrate distinct performance patterns between the MLP and KNN models across hardware platforms, with normalized data significantly influencing accuracy outcomes. Comparative analyses reveal inherent trade-offs between algorithmic complexity, preprocessing requirements, and hardware limitations. These findings underscore the potential for deploying lightweight AI models on extreme edge devices in smart grids and propose scalable strategies for enhancing IoT integration in energy systems.]

Comments 2: [There are some editorial and English mistakes, such as Figure 7, Figure 15, section 4.2. Suggest the authors review this paper in detail to correct some mistakes.]

Response 2: [Thank you for your feedback. We have made the necessary corrections as suggested. In Figure 7, we have revised the label to "Figure 7. Testbed" instead of "Figure 7. Testbad." Additionally, in Figure 15 and Section 4.2, we have updated the description to "Confusion Matrix of the MLP with Unnormalized Data." These adjustments ensure consistency and clarity throughout the manuscript. We appreciate your careful review and the opportunity to enhance our work.]

Reviewer 3 Report

Comments and Suggestions for Authors

In this article, authors have proposed the implementation of artificial intelligence algorithm on devices located at the extreme edge, such as ESP32 and Raspberry Pi. It is claimed that the proposal aims to develop practical and scalable solutions for the extreme edge using the IAIoSGT architecture to test and validate these algorithms’ applications in a smart grid environment. The efforts made by the authors are appreciable but followings are the observations regarding the paper.

1. In the last paragraph of the introduction section, the authors have not described about the section-3 while giving an overview of the paper. Further, the word ‘section’ is to be used in place of ‘chapter’.
2. In the same paragraph the authors claim that “Chapter 2 presents a review of related works, highlighting key contributions from literature in the field of smart grids. This chapter provides an overview of the state of the art, identifying gaps this work aims to fill.” But what is observed that this section is very short. Authors have not been able to identify the gaps that the paper aims to fill. Inclusion of some more recent related papers are highly desirable.
3. In section-3.3 (Creation of Load Signature) the approach is not described clearly.
4. Similarly, in the section-4 (Protyping) Figure 6. (IAIoSGT Architecture) needs to be described with more clarity.
5. In section 4.3 “Two distinct testbeds were conducted to evaluate the performance of the proposed system.” But only one is described. Authors to clarify regarding this.
6. In overall, the paper has interesting results but the analysis of these results have not been up to the mark.
7. The scalability of the proposed approach needs to be discussed. The authors need to mention a scheme for its implementation at a bigger scale so that it can used for SG applications as envisaged.

Comments on the Quality of English Language

Authors are advised to write in passive voice always.

Author Response

Comments 1: [In this article, authors have proposed the implementation of artificial intelligence algorithm on devices located at the extreme edge, such as ESP32 and Raspberry Pi. It is claimed that the proposal aims to develop practical and scalable solutions for the extreme edge using the IAIoSGT architecture to test and validate these algorithms’ applications in a smart grid environment. The efforts made by the authors are appreciable but followings are the observations regarding the paper.]

Response 1: [We appreciate the reviewer's efforts to carefully review the manuscript and give positive feedback. We are grateful for the time and energy you took to review our work. In the following, you will find our responses and corrections to each of your points and suggestions.]

Comments 2: [In the last paragraph of the introduction section, the authors have not described about the section-3 while giving an overview of the paper. Further, the word ‘section’ is to be used in place of ‘chapter’.]

Response 2: [We have addressed this issue by revising the last paragraph of the Introduction to include a description of Section 3 as follows:

Section 3 presents the proposed solution, detailing the implementation of artificial intelligence models for detecting and classifying electronic devices in smart grids. It describes the system architecture, data acquisition process, and the methodology applied to ensure efficient operation in extreme edge environments.”

Additionally, we have replaced the word "chapter" with "section" in the mentioned paragraph to ensure consistency throughout the manuscript. We appreciate your careful review and the opportunity to refine our work.]

Comments 3: [In the same paragraph the authors claim that “Chapter 2 presents a review of related works, highlighting key contributions from literature in the field of smart grids. This chapter provides an overview of the state of the art, identifying gaps this work aims to fill.” But what is observed that this section is very short. Authors have not been able to identify the gaps that the paper aims to fill. Inclusion of some more recent related papers are highly desirable.]

Response 3: [In response to your comment regarding the brevity of Section 2 and the identification of research gaps, we have revised and expanded this section to provide a more comprehensive discussion of related works. Specifically, we have strengthened the literature review by incorporating two recent studies that further contextualize our research and clarify the existing gaps that this work aims to address.

These additions are now clearly marked in Section 2 for your review. We believe these improvements enhance the academic rigor of our paper and align with your recommendation for a more thorough discussion of recent advancements in the field.

A framework combining edge computing and machine learning to enhance the detection and prevention of cyber intrusions in Intelligent Electronic Devices (IEDs) within smart grids was recently proposed by [19]. Modified LGBM and One-Class SVM models were utilized in the study, implementing an intelligent prioritization mechanism that assesses the severity of threats based on their behavior, thereby enabling more effective responses and optimized resource allocation. This approach underscores the effectiveness of edge computing and advanced machine learning techniques in protecting critical infrastructures against emerging cyber threats.

A comprehensive review of AI techniques applied to smart grids was provided by [20], emphasizing how AI has been utilized to enhance demand forecasting, optimize energy distribution, and improve fault detection. The authors highlight that the integration of AI into energy systems results in more reliable and efficient operations, aligning with the objectives of the present study, in which machine learning models are applied to optimize the management and security of smart grids.

In the work [21], a machine learning-based approach for fraud detection in smart grids is presented, employing time-series classifiers applied to data collected from IEDs. The effectiveness of these techniques in identifying non-technical losses, such as illegal connections and measurement errors, is demonstrated. In general, the field of smart grids with embedded IoT and AI encompasses a broad research domain, with multiple approaches and solutions addressing specific challenges. Unlike the focus on fraud detection, the proposed work applies machine learning at the extreme edge to classify and identify electronic devices connected to the power grid, considering the computational constraints of embedded devices such as ESP32 and Raspberry Pi. This approach contributes to advancing energy reliability by enabling real-time analysis and monitoring directly on edge devices, eliminating the need for centralized cloud processing.]

Comments 4: [In section-3.3 (Creation of Load Signature) the approach is not described clearly.]

Response 4: [We go into more detail about creating of load signature.

This approach creates a unique load signature for each device using only its fundamental current signal. Instead of relying on multiple electrical parameters such as power, power factor, phase angle, or harmonics, the identification process focuses solely on the current signal and its amplitude variations in each cycle. This simplification reduces the need for complex processing, making identification more efficient and accessible without compromising accuracy. Additionally, by requiring fewer computational resources, this approach enables deployment on extreme edge devices, allowing execution on hardware with limited resources, such as microcontrollers.]

Comments 5: [Similarly, in the section-4 (Protyping) Figure 6. (IAIoSGT Architecture) needs to be described with more clarity.]

Response 5: [We sincerely thank the reviewer for the valuable feedback regarding the clarity of Figure 6 (IAIoSGT Architecture) in Section 4 (Prototyping). We have taken the following steps to address this concern:

1. Figure Update: The figure has been revised to better represent the IAIoSGT architecture, with a focus on the layers most relevant to this work (Extreme Edge and Edge-Cloud DC). The updated figure now highlights the key components and data flows, making it easier to understand the interaction between modules.
2. Enhanced Description in the Text: We have added a more detailed description of the figure in the text. Specifically, we included a breakdown of each layer and its components, along with a list of the modules in the Edge-Cloud DC layer, accompanied by brief explanations of their roles. This should help readers better follow the figure and understand the architecture.
3. Focus on Relevant Layers: Since the Central Cloud DC layer is not covered in this work, we have emphasized the description of the Extreme Edge and Edge-Cloud DC layers, which are the primary focus of our solution.

We believe these changes address the reviewer's request for greater clarity in the description of the architecture. We appreciate the feedback, which has helped us improve the quality of the manuscript.]

Comments 6: [In section 4.3 “Two distinct testbeds were conducted to evaluate the performance of the proposed system.” But only one is described. Authors to clarify regarding this.]

Response 6: [Thank you for your observation. In Section 4.3, we state that two distinct testbeds were conducted to evaluate the performance of the proposed system. These testbeds are indeed described in the section as follows:

First testbed: The ESP32, preloaded with a CSV file containing previously captured energy reading samples, performed data classification using MLP and KNN algorithms, simulating an application at the extreme edge.

Second testbed: The ESP32 captured real-time energy data, stored it in a CSV file, and later made it available to a Raspberry Pi, which classified the parameters using the same algorithms. This second experiment enabled a comparative analysis of the algorithms' performance on devices that can be deployed at the extreme edge.

For clarity, below is the exact revised text included in Section 4.3:

In this section, we present the methodology used in the development of this work. To evaluate the performance of the proposed system, we conducted two distinct experiments. In the first, the ESP32 microcontroller, preloaded with a CSV file containing previously captured energy reading samples, performed data classification using MLP and KNN algorithms, simulating an application at the extreme edge. In the second experiment, the ESP32 captured real-time energy data, stored it in a CSV file, and later made it available to a Raspberry Pi, which classified the parameters using the same algorithms. This second experiment enabled a comparative analysis of the algorithms' performance on devices that can be deployed at the extreme edge.]

Comments 7: [In overall, the paper has interesting results but the analysis of these results have not been up to the mark.]

Response 7: [We sincerely appreciate your valuable feedback, which helped us enhance the clarity and depth of our research presentation. In line with your suggestions, a new subsection titled ``Quantitative Assessment of the Accuracy of Classification Models'' has been incorporated to offer a detailed quantitative analysis of our results, incorporating metrics such as precision, recall, F1-score, and accuracy. To further illustrate model performance, we have included the following figures:

ROC Curves for MLP with normalized and non-normalized data,

ROC Curves for KNN with k = 1 and k = 3, using normalized and non-normalized data.

Additionally, to provide a concise overview of the experimental outcomes, we have referenced the existing summary table in the conclusion – “KNN and MLP Performance on ESP32 and Raspberry Pi 3B+ for Normalized and Non-Normalized Data.” Further information and data supporting these findings can be found in Section-5: Analysis on the Experiment Outcomes. Once again, thank you for your insightful feedback and for guiding us toward a more comprehensive analysis.]

Comments 8: [The scalability of the proposed approach needs to be discussed. The authors need to mention a scheme for its implementation at a bigger scale so that it can used for SG applications as envisaged.]

Response 8: [We appreciate the reviewer's comment. To address this issue, we have expanded the discussion on scalability in the conclusion. We explain how the proposed approach integrates into the IAIoSGT architecture, where extreme edge devices, such as the ESP32, perform local inference and send only essential data to edge servers or the cloud, reducing communication overhead. Additionally, we detail a hybrid scheme where lightweight models operate on resource-constrained devices, while more complex models run at the network edge. This strategy enhances scalability by optimizing resource usage and enabling the application in large-scale Smart Grids. These changes have been included in the manuscript to clarify the feasibility of the proposed solution.]

Comments 9: [Authors are advised to write in passive voice always.]

Response 9: [We thank the reviewer for the valuable feedback regarding the use of passive voice. In response, we have revised the entire manuscript to ensure consistent use of the passive voice, replacing active constructions and personal pronouns where necessary. All sections, including the introduction, methodology, results, and conclusion, have been adjusted to adhere to this guideline. We believe these changes improve the manuscript's adherence to academic writing standards, and we hope the revised version now meets the reviewer's expectations.]

Reviewer 4 Report

Comments and Suggestions for Authors

“Extreme Edge-Premised Machine Learning-Based Classification and Detection of IoT Electronic Devices” is a good initiative made by the authors. The paper contains some contents but they are not properly organized. A major revision is recommended therefore.

The authors shall check the title relevancy, Extreme Edge-Premised Machine Learning-Based Classification and Detection of IoT Electronic Devices.

Throughout the manuscript, the authors have not mentioned the Detection of IoT Electronic Devices. But utilized the devices for deploying.

           In the abstract section, authors must elaborate on the AI algorithms used, clearly state the objective of the proposed work, and clearly state the outcome obtained. The present abstract is general and does not include the author's contribution. Rewriting the abstract is recommended.

Keywords Section: Machine Learning, Extreme Edge Computing, IoT are used, but the same is not seen in the abstract or not expanded.

Throughout the manuscript there exists an ambiguity in using the word AI and ML. What are the differences between AI and ML? There should be a clear understanding of AI and ML.

(Line 47-48) What are the differences between edge and extreme edge? How is IoT used?

In line 52, it is mentioned as improved data security. How is it achieved? What type of security is required in the edge devices when ML algorithms are deployed? No references are made and the  way of implementing security is not mentioned. …

It is mentioned it transmits large volumes of data to central servers; what will be the data rate and the mode.

 With 9 reference papers, including the author's paper, the authors have concluded the need for the proposal. Can also refer to more related works in edge cloud and IoT.

Line 61-63 ; Authors have not mentioned about the ML models need to be deployed in the edge devices. How ESP32 Performs the computation w.r.t the ML model. Is time complexity and performance analysed for a large volume of data.?

Line68-69 ; Chapter 3 is not mentioned. Why have authors mentioned as chapters in the place of sections.

Line 104-109; The statements are not clear to the readers.

          In my opinion, literature survey is written as a generic, and the authors have to discuss each work with the related methods, along with their merits and demerits. Needs to be Fine-tuned to make readers understand why this proposed work will be useful.

Figure 1-3: There is no clear idea how the edge is connected to the cloud. Are there any cloud service providers involved? How Network connectivity is established. In Fig-3, the edge node seems to be overloaded.

              If data is sent to the cloud, what is the latency between the edge and cloud? What will be the performance of the proposed systems if there is high latency?

(MLP and KNN) Algorithm is not elaborated properly in the context of the proposed work.

MQTT has to be explained.

Figure 6. IAIoSGT Architecture is utilized from the earlier work https://journals-sol.sbc.org.br/index.php/jisa/article/view/3076; (Check with the editor for using the same image) but this work includes the ESP32 and Raspberry Pi. There is no representation in the architecture diagram for these devices and the network connectivity links to edge and the central cloud.

How the Bluetooth module helps in this architecture as mentioned in line 254.

Section 4.2 is not in the ENGLISH Language.

Figure 7. Testbad (Check the spelling); Give real-time photos of the testbed.

The main objective is to classify electrical and  electronic devices based on their consumption signatures (why the word signature is used)

Dataset is too small to test the system.

Figure 15, check language

Generally, a Confusion matrix is not needed as the screenshots. It can be in the table if needed to represent.

ROC graph is missing

The conclusion is not properly written. Table 2 should be in Section 5.

615 – 620 has to be justified when authors comment over the languages and the processes.

Conclusion needs re-write

References are not properly written, check 23.

References should be more related to the IoT, Edge, and Cloud.

                    In my opinion, the paper is not properly structured, needs re-write, has more typos, punctuation, and grammatical errors. Moreover, some places, other than the English Language, statements are noted. The References section and their citation need to be improved related to this proposed context. Architecture diagram should cover all aspects. The Test Bed can be elaborated well. Methodology should be elaborated with ML methods. Results and discussion should be properly justified with the large volume of data. Finally, the Conclusion section should be re-written with the author's problem statement and outcome, and comment on the future works.

The authors shall check the title relevancy, Extreme Edge-Premised Machine Learning-Based Classification and Detection of IoT Electronic Devices.  (The title needs to be justified)

Comments on the Quality of English Language

Needs Improvement

Author Response

Comments 1: [Extreme Edge-Premised Machine Learning-Based Classification and Detection of IoT Electronic Devices” is a good initiative made by the authors. The paper contains some contents but they are not properly organized. A major revision is recommended therefore. The authors shall check the title relevancy, Extreme Edge-Premised Machine Learning-Based Classification and Detection of IoT Electronic Devices. Throughout the manuscript, the authors have not mentioned the Detection of IoT Electronic Devices. But utilized the devices for deploying.]

Response 1: [We appreciate the reviewer’s feedback regarding the title’s clarity. To better align the title with the content of the manuscript, we have restructured it to emphasize that the goal of the study is not the detection of IoT devices themselves, but rather the classification of electronic devices using an IoT-based smart meter. The revised title now more accurately reflects the study’s focus and methodology.

We've updated the title to: Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge]

Comments 2: [In the abstract section, authors must elaborate on the AI algorithms used, clearly state the objective of the proposed work, and clearly state the outcome obtained. The present abstract is general and does not include the author's contribution. Rewriting the abstract is recommended.]

Response 2: [We appreciate the valuable feedback on our manuscript. As recommended, we have revised the abstract to elaborate on the AI algorithms used, explicitly state the objective of the proposed work, and clearly present the outcomes obtained. The updated abstract is provided below.

The implementation of artificial intelligence algorithms on resource-constrained edge devices is investigated in this study, including ESP32 and Raspberry Pi, for smart grid applications. Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) models are employed to classify IoT electronic devices within the IAIoSGT architecture. A literature review identifies critical gaps in existing methodologies, while experimental evaluations emphasize the pivotal role of data preprocessing, particularly normalization, in optimizing model performance. Results demonstrate distinct performance patterns between the MLP and KNN models across hardware platforms, with normalized data significantly influencing accuracy outcomes. Comparative analyses reveal inherent trade-offs between algorithmic complexity, preprocessing requirements, and hardware limitations. These findings underscore the potential for deploying lightweight AI models on extreme edge devices in smart grids and propose scalable strategies for enhancing IoT integration in energy systems.]

Comments 3: [Keywords Section: Machine Learning, Extreme Edge Computing, IoT are used, but the same is not seen in the abstract or not expanded.]

Response 3: [After a thorough analysis, we have revised both the abstract and the keywords to ensure better alignment between them. The keywords have been updated to reflect the core aspects of our study. The revised keywords are:

Keywords: Artificial Intelligence; Extreme Edge; IoT; Smart Grid.

Thank you for your valuable suggestions, which have helped improve the clarity and coherence of our manuscript.]

Comments 4: [Throughout the manuscript there exists an ambiguity in using the word AI and ML. What are the differences between AI and ML? There should be a clear understanding of AI and ML.]

Response 4: [Thank you for your observation. We have added a new paragraph to the Introduction to clarify the distinction between Artificial Intelligence (AI) and Machine Learning (ML). AI is defined as the broader field encompassing systems that perform tasks requiring human intelligence, while ML is a subset of AI focused on learning from data to improve performance without explicit programming.]

Comments 5: [(Line 47-48) What are the differences between edge and extreme edge? How is IoT used?]

Response 5: [Thank you for your observation. A paragraph has been added to the Introduction to clarify the differences between edge and extreme edge computing, as well as the role of IoT in the proposed architecture. Specifically, it explains that edge computing involves processing data closer to the source than centralized cloud servers, while extreme edge computing pushes processing directly onto resource-constrained devices, such as microcontrollers (e.g., ESP32) or single-board computers (e.g., Raspberry Pi). Additionally, the role of IoT devices, such as smart meters, in collecting and transmitting data for real-time analysis is highlighted.]

Comments 6: [In line 52, it is mentioned as improved data security. How is it achieved? What type of security is required in the edge devices when ML algorithms are deployed? No references are made and the way of implementing security is not mentioned.]

Response 6: [We appreciate the reviewer's observation regarding the mention of improved data security. Since the current work does not focus on security aspects, we have removed the reference to “improved data security” from the manuscript to avoid any misinterpretation. Future studies may explore security mechanisms for edge and extreme edge devices, but this topic is beyond the scope of the present research.]

Comments 7: [It is mentioned it transmits large volumes of data to central servers; what will be the data rate and the mode.]

Response 7: [We appreciate the reviewer's comment. While the Central Cloud DC is part of the proposed IAIoSGT architecture, this study focuses exclusively on data processing at the extreme edge (using devices like ESP32 and Raspberry Pi) and the Edge-Cloud DC. In this implementation, no data is transmitted to the Central Cloud, as all processing is performed locally at the edge or within the Edge-Cloud DC. The MQTT protocol is used for efficient communication between edge devices and the Edge-Cloud, ensuring minimal bandwidth usage.

To address this point more clearly, we have revised the manuscript to emphasize that the Central Cloud is not utilized in this implementation and that the focus is entirely on edge and Edge-Cloud processing.]

Comments 8: [With 9 reference papers, including the author's paper, the authors have concluded the need for the proposal. Can also refer to more related works in edge cloud and IoT.]

Response 8: [Three more works were included in the article reference, totaling 26 related works.

A framework combining edge computing and machine learning to enhance the detection and prevention of cyber intrusions in Intelligent Electronic Devices (IEDs) within smart grids was recently proposed by [18]. Modified LGBM and One-Class SVM models were utilized in the study, implementing an intelligent prioritization mechanism that assesses the severity of threats based on their behavior, thereby enabling more effective responses and optimized resource allocation. This approach underscores the effectiveness of edge computing and advanced machine learning techniques in protecting critical infrastructures against emerging cyber threats.

A comprehensive review of AI techniques applied to smart grids was provided by [19], emphasizing how AI has been utilized to enhance demand forecasting, optimize energy distribution, and improve fault detection. The authors highlight that the integration of AI into energy systems results in more reliable and efficient operations, aligning with the objectives of the present study, in which machine learning models are applied to optimize the management and security of smart grids.

In the work [20], a machine learning-based approach for fraud detection in smart grids is presented, employing time-series classifiers applied to data collected from IEDs. The effectiveness of these techniques in identifying non-technical losses, such as illegal connections and measurement errors, is demonstrated. In general, the field of smart grids with embedded IoT and AI encompasses a broad research domain, with multiple approaches and solutions addressing specific challenges. Unlike the focus on fraud detection, the proposed work applies machine learning at the extreme edge to classify and identify electronic devices connected to the power grid, considering the computational constraints of embedded devices such as ESP32 and Raspberry Pi. This approach contributes to advancing energy reliability by enabling real-time analysis and monitoring directly on edge devices, eliminating the need for centralized cloud processing.]

Comments 9: [Line 61-63 ; Authors have not mentioned about the ML models need to be deployed in the edge devices. How ESP32 Performs the computation w.r.t the ML model. Is time complexity and performance analysed for a large volume of data.?]

Response 9: [Thank you for your valuable review comments. I appreciate your attention to the implementation details of ML models on edge devices.

Regarding your concerns about deploying ML models on edge devices like ESP32 and analyzing their time complexity and performance with large volumes of data, I would like to respectfully note that these aspects are comprehensively addressed in the revised paper:

ESP32 ML model implementation details: In the section "MLP on ESP32", we explicitly detail the implementation approach: “The implementation of the MLP neural network on the ESP32 was carried out using TensorFlow Lite Micro. Since the network's architecture was relatively simple, quantization of the weights was not necessary when transferring the model to the microcontroller.” We also provide specific memory metrics: “The network used 1936 bytes of memory for the model with normalized data and 1968 bytes for the model with non-normalized data.”

Performance analysis: In the “KNN and MLP Performance on ESP32 and Raspberry Pi 3B+” section, we present a comprehensive table comparing both algorithms across platforms, showing that “When running the network with normalized data on the ESP32, an accuracy of 100% was achieved” and “The average execution time per prediction was approximately 0.31 milliseconds.”

Data volume considerations: In the “Dataset manipulation and Pre-processing” section, we address memory constraints: “The reduced dataset was used to run the KNN algorithm on the ESP32, where memory limitation was a constraint... Thus, reducing the number of examples in the 'ProcessedRAND' file was a crucial step in using KNN on the ESP32.”

Time complexity: The “Analysis on the Experiment Outcomes” section provides detailed execution time analysis across configurations. For example, in “KNN on the ESP32” we note that “The average prediction time for the non-normalized data was slightly higher, ranging from 4.21 ms to 4.30 ms—times that can be considered acceptable given the limited hardware of the ESP32.”

The paper's “Conclusion and Future Work”' section further synthesizes these findings, noting the trade-offs between processing requirements and memory constraints when implementing machine learning algorithms on extreme edge devices.

I hope this clarifies your concerns, and I thank you again for your thoughtful review.]

Comments 10: [Line68-69; Chapter 3 is not mentioned. Why have authors mentioned as chapters in the place of sections.]

Response 10: [We appreciate the reviewer’s feedback regarding the terminology inconsistency. The manuscript has been revised to replace 'chapters' with 'sections' for clarity and consistency. Additionally, we have included the description of Section 3 as follows:

Section 3 presents the proposed solution, detailing the implementation of artificial intelligence models for detecting and classifying electronic devices in smart grids. It describes the system architecture, data acquisition process, and the methodology applied to ensure efficient operation in extreme edge environments.]

Comments 11: [Line 104-109; The statements are not clear to the readers.]

Response 11: [The techniques used in the cited studies were detailed in more depth, and we hope the ideas are now clearer.

Several recent studies demonstrate the effectiveness of artificial intelligence in monitoring and identifying electronic devices using advanced machine learning techniques. For example, residual convolutional neural networks (ResNet), which utilize residual connections to enable efficient learning in deep networks, have been used to enhance device recognition in non-intrusive load identification scenarios, achieving high performance in distinguishing different electrical devices based on their energy consumption patterns [13]. The residual connections allow the network to learn the difference (or “residue”) between the input and output of a layer, helping the network avoid the vanishing gradient problem and enabling the training of deeper networks to learn complex energy consumption patterns, even with noisy data.

Furthermore, efficient hybrid models have been developed for the classification of electrical devices, exploiting features extracted from consumption time series. These models combine different learning techniques, such as convolutional neural networks (which capture temporal patterns), with other supervised learning techniques, such as support vector machines or decision trees, to improve accuracy and reduce computational complexity [14]. The advantage of hybrid models is that they can combine the strengths of different approaches, allowing the system to benefit from various characteristics of the consumption data, thus increasing classification efficiency.

Another relevant approach is the use of siamese neural networks for detecting unidentified devices in non-intrusive load monitoring systems. Siamese networks use two identical networks that share weights and parameters and are trained to compare inputs and measure the similarity between them. This architecture allows the system to identify new devices by comparing energy consumption patterns with those of known devices [15]. Even without pre-existing examples of unidentified devices, the network can determine if a new device has a consumption behavior similar to that of an already registered appliance. This facilitates the detection of new devices without the need for a large labeled dataset.

These studies demonstrate the potential of AI algorithms in analyzing energy consumption patterns, making them essential tools for the accurate and efficient identification of electrical devices in both domestic and industrial environments.]

Comments 12: [In my opinion, literature survey is written as a generic, and the authors have to discuss each work with the related methods, along with their merits and demerits. Needs to be Fine-tuned to make readers understand why this proposed work will be useful.]

Response 12: [Thank you for your valuable feedback regarding the literature survey section of our manuscript. We appreciate your recommendation to enhance this section by discussing relevant works with their associated methods, merits, and limitations, as well as clarifying the unique contribution of our proposed approach.

In response to your suggestions, we have thoroughly revised the literature review section to include more targeted discussions of related works. Specifically, we have added the following paragraphs:

A framework combining edge computing and machine learning to enhance the detection and prevention of cyber intrusions in Intelligent Electronic Devices (IEDs) within smart grids was recently proposed by [18]. Modified LGBM and One-Class SVM models were utilized in the study, implementing an intelligent prioritization mechanism that assesses the severity of threats based on their behavior, thereby enabling more effective responses and optimized resource allocation. This approach underscores the effectiveness of edge computing and advanced machine learning techniques in protecting critical infrastructures against emerging cyber threats.

A comprehensive review of AI techniques applied to smart grids was provided by [19], emphasizing how AI has been utilized to enhance demand forecasting, optimize energy distribution, and improve fault detection. The authors highlight that the integration of AI into energy systems results in more reliable and efficient operations, aligning with the objectives of the present study, in which machine learning models are applied to optimize the management and security of smart grids.

In the work [20], a machine learning-based approach for fraud detection in smart grids is presented, employing time-series classifiers applied to data collected from IEDs. The effectiveness of these techniques in identifying non-technical losses, such as illegal connections and measurement errors, is demonstrated. In general, the field of smart grids with embedded IoT and AI encompasses a broad research domain, with multiple approaches and solutions addressing specific challenges. Unlike the focus on fraud detection, the proposed work applies machine learning at the extreme edge to classify and identify electronic devices connected to the power grid, considering the computational constraints of embedded devices such as ESP32 and Raspberry Pi. This approach contributes to advancing energy reliability by enabling real-time analysis and monitoring directly on edge devices, eliminating the need for centralized cloud processing.

These additions provide better context for our research, highlight the distinctive aspects of our contribution, and demonstrate how our work differs from and builds upon existing approaches. We believe these revisions address your concerns by creating a clearer narrative that positions our work within the broader research landscape while emphasizing its unique value and practical significance.]

Comments 13: [Figure 1-3: There is no clear idea how the edge is connected to the cloud. Are there any cloud service providers involved? How Network connectivity is established. In Fig-3, the edge node seems to be overloaded.]

Response 13: [Thank you for your review comments regarding Figures 1-3. I would like to address the concerns about cloud-edge connectivity and implementation:

Edge-Cloud Connectivity: As shown in Figure 1, the connectivity between edge devices and the cloud is established through the MQTT protocol. To clarify this important aspect, we have updated the paper with additional text: “The communication between the Smart Meter and the data acquisition module uses the MQTT protocol, enabling efficient data transmission to the cloud, as illustrated in Figure 1.”

Cloud Implementation: The paper describes our implementation using the FIWARE ecosystem rather than commercial cloud service providers. As detailed in the “Edge-Cloud DC” section: “These modules are organized within the FIWARE ecosystem and encapsulated in Docker containers, which promotes scalability.”

Network Architecture: We have enhanced the explanation of how network connectivity is established by adding: “This network connectivity is established through a layered architecture where edge devices connect to the cloud infrastructure as detailed in our IAIoSGT architecture (Figure 6). Figure 3 specifically demonstrates how trained models are deployed to extreme edge devices.”

Edge Node Processing: Regarding your concern about the ESP32 appearing overloaded in Figure 3, we have clarified that despite the multiple functions depicted, our implementation ensures “computational efficiency at the edge.” This is supported by our results showing the ESP32 handles classification tasks efficiently with “average execution time per prediction was approximately 0.31 milliseconds” for MLP and reasonable times for KNN as well.

I appreciate your careful review which prompted these clarifications.]

Comments 14: [If data is sent to the cloud, what is the latency between the edge and cloud? What will be the performance of the proposed systems if there is high latency?]

Response 14: [Thank you for your valuable comment. To address your concern, we measured the latency between the ESP32 (Extreme Edge device) and the Edge-Cloud infrastructure in our experimental setup. The measured latency for sending data via the MQTT protocol was approximately 110.164 ms (0.110164 seconds) under typical network conditions. This value has been incorporated into the revised manuscript in Section 3.1 (System Architecture) to provide clarity on the data transmission process.

Regarding the impact of latency on system performance, we would like to emphasize that the proposed IAIoSGT architecture is designed to mitigate the effects of network latency. The machine learning models (MLP and KNN) are executed locally on the Extreme Edge devices (ESP32 and Raspberry Pi 3B+), enabling real-time classification of electronic devices independently of the network connection to the Edge-Cloud. The data transmission to the cloud is primarily used for asynchronous model training and periodic synchronization of preprocessed data, which does not affect the real-time inference performance at the edge. Furthermore, training and synchronization can be scheduled during periods of low system usage (e.g., overnight), ensuring minimal interference with operational efficiency even in scenarios of high latency. This design choice ensures that the system remains robust and scalable, maintaining high performance regardless of network conditions. We have added a brief explanation of this in the revised manuscript (Section 3.1) to address your question comprehensively.]

Comments 15: [(MLP and KNN) Algorithm is not elaborated properly in the context of the proposed work.]

Response 15: [We appreciate the reviewer’s comment regarding the need for further elaboration on the selection of MLP and KNN algorithms in the context of our proposed work. In the introduction, we have already stated that this study builds upon a previous work, which conducted a comparative analysis of multiple machine learning models for this specific application. The findings from that study guided the selection of MLP and KNN as the most suitable algorithms considering both classification performance and computational constraints of extreme edge devices.

To enhance clarity and reinforce this point, we have now added an explicit reference to the selection of MLP and KNN in the introduction, emphasizing their effectiveness for the application while also considering the computational cost of their implementation on resource-constrained devices.

The previous study evaluated different machine learning models and identified the most effective ones for this specific application. Building upon these findings, the current study selected KNN and MLP, considering both their classification performance and the computational cost required for deployment on extreme edge devices.

We believe this addition strengthens the manuscript by providing a clearer justification for the choice of these models. We thank the reviewer for this valuable suggestion.]

Comments 16: [MQTT has to be explained.]

Response 16: [We added a brief paragraph explaining what MQTT is and why it was chosen.

MQTT is a communication protocol for IoT that follows the asynchronous publish/subscribe model, enabling efficient message exchange between connected devices. It has a low overhead, meaning it adds only a minimal amount of extra data to messages for communication control, optimizing bandwidth usage. Designed to be extremely lightweight, MQTT works well for resource-constrained devices and low-bandwidth networks, ensuring efficient and reliable transmission. This characteristic makes MQTT an excellent choice for various applications, including monitoring systems, industrial automation, smart homes, and energy consumption meters.]

Comments 17: [Figure 6. IAIoSGT Architecture is utilized from the earlier work https://journals-sol.sbc.org.br/index.php/jisa/article/view/3076; (Check with the editor for using the same image) but this work includes the ESP32 and Raspberry Pi. There is no representation in the architecture diagram for these devices and the network connectivity links to edge and the central cloud.]

Response 17: [Thank you for your feedback. We have updated Figure 6 to include the ESP32 and Raspberry Pi, as well as the network connectivity links, ensuring the architecture diagram aligns with the described system.]

Comments 18: [How the Bluetooth module helps in this architecture as mentioned in line 254.]

Response 18: [We appreciate the reviewer observation. The mention of the Bluetooth module was solely to highlight one of the features of the ESP32 microcontroller. However, this functionality was not explored in the current study, as it falls outside the scope of our work. Nonetheless, Bluetooth connectivity can be valuable in similar solutions that require communication with mobile devices or even between multiple ESP32 units using this protocol.]

Comments 19: [Section 4.2 is not in the ENGLISH Language.]

Response 19: [We appreciate the reviewer’s feedback. The sections that were not in English have been revised and corrected to ensure consistency throughout the manuscript.]

Comments 20: [Figure 7. Testbad (Check the spelling); Give real-time photos of the testbed.]

Response 20: [Thank you for your feedback. The figure has been updated as requested.]

Comments 21: [The main objective is to classify electrical and electronic devices based on their consumption signatures (why the word signature is used)]

Response 21: [We appreciate the reviewer question. The term 'signature' is widely used in the literature on similar studies, as different works may use specific electrical characteristics for analysis and identification. These characteristics can include power consumption, as in our case, or other parameters such as harmonic components, power factor, or additional electrical features. Each study defines its own load signature based on the available data, computational resources, and the intended application.]

Comments 22: [Dataset is too small to test the system.]

Response 22: [Thank you for your important observation regarding the dataset size. I appreciate the opportunity to address this methodological concern:

Dataset constraints in extreme edge computing: While we acknowledge that larger datasets generally lead to more robust models, our research specifically focuses on the feasibility of deploying ML models on highly resource-constrained devices like the ESP32. As detailed in the paper: “The reduced dataset was used to run the KNN algorithm on the ESP32, where memory limitation was a constraint... Thus, reducing the number of examples in the 'ProcessedRAND' file was a crucial step in using KNN on the ESP32.” This limitation represents a real-world constraint that embedded systems researchers must address.

Consistent findings despite dataset size: Despite the dataset's size limitations, we were able to demonstrate consistent and meaningful performance differences between algorithms and preprocessing techniques. For example, our results show clear patterns where “normalization was crucial to achieving high performance when using the MLP, reaching 100% accuracy” while KNN showed “better performance with non-normalized data, achieving up to 96.23% accuracy for k = 1.” These distinct patterns suggest the dataset was sufficient to draw meaningful conclusions about algorithm behavior on edge devices.

Focus on computational efficiency: The primary contribution of our work is not a state-of-the-art classification system, but rather demonstrating the “trade-off between processing and memory when implementing these algorithms” on extreme edge devices. Our detailed analysis of execution times (e.g., “0.31 milliseconds”} for MLP) and memory usage (e.g., “1936 bytes”) offers important computational insights for embedded systems researchers that remain relevant despite the dataset limitations.

Careful dataset preparation: As described in the “Data Acquisition” section, we ensured data quality through meticulous collection procedures and preprocessing steps, including cleaning and organization: “A transpose matrix was applied to the data to enable the analysis of patterns in the sinusoidal forms of voltage and current signals, overcoming the limitation of point data.” This approach helped maximize the information content of the available data.

We appreciate this valuable feedback and will consider expanding our dataset in future work to strengthen the generalizability of our findings. However, we believe the current dataset adequately serves the paper's focus on computational efficiency and feasibility of ML implementation on extreme edge devices.]

Comments 23: [Figure 15, check language.]

Response 23: [We have made the necessary adjustment, and the description of Figure 15 has been revised to “Confusion Matrix of the MLP with Unnormalized Data.” This ensures consistency with the rest of the manuscript.]

Comments 24: [Generally, a Confusion matrix is not needed as the screenshots. It can be in the table if needed to represent.]

Response 24: [We appreciate the reviewer’s suggestion regarding the representation of the confusion matrices. As requested, all confusion matrix figures have been replaced with tables to present the classification results more concisely and clearly.]

Comments 25: [ROC graph is missing.]

Response 25: [Thank you for noting the absence of ROC graphs in the initial submission. We have addressed this in the revised version of the paper by implementing ROC curve analysis for both MLP and KNN models. The “ROC Curves for MLP with normalized and non-normalized data” and “ROC Curves for KNN with k = 1 and k = 3, using normalized and non-normalized data” now present these visualizations, showing the classifier performance with both preprocessing approaches across different configurations.]

Comments 26: [The conclusion is not properly written. Table 2 should be in Section 5.]

Response 26: [We appreciate the reviewer’s feedback regarding the conclusion and the placement of Table 2. The necessary adjustment has been made, and the table is now correctly referenced in Section 5.3.2 KNN on the ESP32. Additionally, we have reviewed and refined the conclusion to ensure it effectively summarizes the key findings and contributions of this study. Thank you for your valuable suggestions.]

Comments 27: [615 – 620 has to be justified when authors comment over the languages and the processes.]

Response 27: [We've added more details about the overhead caused by a non-bare-metal operating system and using an interpreted language like Python.

Although the Raspberry Pi has more powerful hardware in terms of processing and memory, its inference times are significantly higher than those of the ESP32 due to the operating system overhead. The Raspberry Pi runs a Linux-based system, such as Raspbian, which manages multiple processes simultaneously, consuming resources and introducing latency. In contrast, the ESP32 runs bare-metal firmware, allowing model inference to take top priority without significant resource competition.

Additionally, the choice of programming language directly impacts performance. On the Raspberry Pi, inference runs in Python, an interpreted language that adds execution overhead since it dynamically translates each instruction into machine code. In contrast, the ESP32 operates in C or C++, compiled languages that generate optimized code for direct hardware execution, eliminating interpretation overhead and making inference far more efficient.]

Comments 28: [Conclusion needs re-write]

Response 28: [We have added a few more paragraphs to the conclusion, reinforcing the focus of the proposed architecture, highlighting the positive aspects of its use, and explaining the results in more detail. I hope this change meets your expectations. If not, please provide more details on the improvements needed.

Although the Raspberry Pi has more powerful hardware in terms of processing and memory, its inference times are significantly higher than those of the ESP32 due to the operating system overhead. The Raspberry Pi runs a Linux-based system, such as Raspbian, which manages multiple processes simultaneously, consuming resources and introducing latency. In contrast, the ESP32 runs bare-metal firmware, allowing model inference to take top priority without significant resource competition.

Additionally, the choice of programming language directly impacts performance. On the Raspberry Pi, inference runs in Python, an interpreted language that adds execution overhead since it dynamically translates each instruction into machine code. In contrast, the ESP32 operates in C or C++, compiled languages that generate optimized code for direct hardware execution, eliminating interpretation overhead and making inference far more efficient.

Finally, resource management plays a crucial role. The Raspberry Pi runs multiple processes simultaneously, sharing CPU and memory with other system tasks, which can cause inference delays. On the ESP32, execution is exclusively dedicated to inference, ensuring more predictable and efficient performance. Therefore, despite the Raspberry Pi’s greater processing power, selecting the right platform and execution environment is key to optimizing performance for embedded and IoT applications.

To enable large-scale implementation within a SG, the proposed approach can be integrated into the IAIoSGT architecture, in which AI and SM are combined to optimize the performance of the power grid. In this structure, inference is performed locally by extreme-edge devices, such as the ESP32, and only essential information is sent to network edge or cloud servers, reducing latency and communication load. Additionally, scalability and computational efficiency are improved by adopting a hybrid model, in which lightweight algorithms are used on embedded devices, while more robust models are employed at the network edge. Through this strategy, real-time detection of consumption patterns, identification of connected devices, and optimization of energy distribution are enabled, ensuring a more efficient and resilient operation of the Smart Grid, even in scenarios with connectivity and computational capacity constraints.]

Comments 29: [References are not properly written, check 23.]

Response 29: [The references have been reviewed and corrected. The reference has been reformatted to ensure consistency with the academic citation style used in the article.]

Comments 30: [References should be more related to the IoT, Edge, and Cloud.]

Response 30: [As explained in question 7, we have added more references to the work and improved the explanation of some existing references. We hope this meets your expectations.]

Comments 31: [In my opinion, the paper is not properly structured, needs re-write, has more typos, punctuation, and grammatical errors. Moreover, some places, other than the English Language, statements are noted. The References section and their citation need to be improved related to this proposed context. Architecture diagram should cover all aspects. The Test Bed can be elaborated well. Methodology should be elaborated with ML methods. Results and discussion should be properly justified with the large volume of data. Finally, the Conclusion section should be re-written with the author's problem statement and outcome, and comment on the future works. The authors shall check the title relevancy, Extreme Edge-Premised Machine Learning-Based Classification and Detection of IoT Electronic Devices. (The title needs to be justified)]

Response 31: [We appreciate the reviewer feedback. All comments have been carefully analyzed and addressed, with revisions improving structure, clarity, references, methodology, results, and conclusions. The title has also been reviewed for alignment with the study’s scope. Detailed explanations of the modifications have been provided in previous responses. We believe these changes enhance the manuscript's quality and clarity.]

 

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have addressed all the points raised by me. However, the following point must be addressed.

1. The previous version of the paper included 23 references and  5 out of
   these  23 cited references belong to the authors (Refs. 1,9,19,20), which
   is a self-citation rate of about 21.73%. The present revised paper has 26 references and  5 out of these  26 cited references belong to the authors (Refs. 1,20,21,22,23 ), which
   is a self-citation rate of about 19.23%. Authors are advised to reduce this number of self citations as far as possible.

Author Response

Comments 1: [The previous version of the paper included 23 references and 5 out of these 23 cited references belong to the authors (Refs. 1,9,19,20), which is a self-citation rate of about 21.73%. The present revised paper has 26 references and 5 out of these 26 cited references belong to the authors (Refs. 1,20,21,22, 23), which is a self-citation rate of about 19.23%. Authors are advised to reduce this number of self citations as far as possible.]

Response 1: [We appreciate your feedback regarding the self-citation rate in our manuscript. To address your request to reduce the self-citation rate from ~19.23% (5/26 references), we identified that references [1] and [21] cited the same article, which was an oversight. This has been corrected, reducing the number of unique self-citations to 4. Additionally, we included two new references in Section 2 (Related Work) to broaden the literature review. The first [18] examines the ESP32’s Xtensa LX6 processor for TinyML neural network applications, highlighting its efficiency in resource-constrained environments. The second [19] proposes an IoT healthcare monitoring solution using machine learning on the ESP32, emphasizing real-time classification capabilities.

With these changes, the total number of references increased to 27, resulting in a self-citation rate of 4/27 (~14.81%), which aligns with the recommended threshold of 15% or lower. These adjustments enhance the manuscript’s comprehensiveness while maintaining the relevance of our prior contributions, such as the IAIoSGT architecture and smart meter development, which remain essential for contextualizing this work. We believe these revisions fully address your concerns and are open to further suggestions.]

Reviewer 4 Report

Comments and Suggestions for Authors

The revised script entitled Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge submitted is reviewed and the comments are given below.

1. In the abstract, the authors have mentioned that the implementation of artificial intelligence (AI) algorithms on resource-constrained edge devices is investigated in this study, including ESP32 and Raspberry Pi, for smart grid applications. But in the title, it is mentioned as Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge. There is no clear implementation for the use of classifying and detecting IoT devices for the smart meter devices.
2. In the abstract, Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) models are employed to classify IoT electronic devices within the IAIoSGT architecture. What is the use in classifying and detecting IoT devices for the smart meter devices ? Are the authors pointing the performances of both the algorithms and choosing the best among ESP and Raspberry PI.
3. Clarification is needed for the title and the same has to be incorporated in the manuscript.
4. In the response to qn no.6, the authors have mentioned NO Security studies are involved. If so, why this paper is citeAlgarni, A.; Ahmad, Z.; Alaa Ala’Anzy, M. An Edge Computing-Based and Threat Behavior-Aware Smart Prioritization Framework for Cybersecurity Intrusion Detection and Prevention of IEDs in Smart Grids With Integration of Modified LGBM and One Class-SVM Models. IEEE Access 2024, 12, 104948–104963. https://doi.org/10.1109/ACCESS.2024.3435564.
5.  In the response to qn 28,29; Still the reference section is not clear.

                   Reference No: 1 AND 21 are the same and from the same author.

6. Even though the following is the continued work. The self-citation may increase and therefore adhere to the guidelines of the journal. (Ref 1,20,21,22)

1.Marques, L.; Eugênio, P.; Bastos, L.; Santos, H.; Rosário, D.; Nogueira, E.; Cerqueira, E.; Kreutz, M.; Neto, A. Analysis of electrical Signals by Machine Learning for Classification of Individualized Electronics on the Internet of Smart Grid Things (IoSGT) architecture. Journal of Internet Services and Applications 2023, 14, 124–135. https://doi.org/10.5753/jisa.2023.3076.

20. Bastos, L.; Martins, B.; Santos, H.; Medeiros, I.; Eugênio, P.; Marques, L.; Rosário, D.; Nogueira, E.; Cerqueira, E.; Kreutz, M.; et al. Predictive Fraud Detection: An Intelligent Method for Internet of Smart Grid Things Systems. Journal of Internet Services and Applications 2023, 14, 1–15. https://doi.org/10.5753/jisa.2023.3077.
21. Marques, L.; Eugênio, P.; Bastos, L.; Santos, H.; Rosário, D.; Nogueira, E.; Cerqueira, E.; Kreutz, M.; Neto, A. Analysis of Electrical Signals by Machine Learning for Classification of Individualized Electronics on the Internet of Smart Grid Things (IoSGT) architecture. Journal of Internet Services and Applications 2023, 14, 124–135
22. Santos, H.; Eugênio, P.; Marques, L.; Oliveira, H.; Rosário, D.; Nogueira, E.; Neto, A.; Cerqueira, E. Internet of Smart Grid Things (IoSGT): Prototyping a Real Cloud-Edge Testbed. In Proceedings of the Proceedings of the 14th Brazilian Symposium on Ubiquitous and Pervasive Computing (SBCUP), Porto Alegre, RS, Brazil, 2022; pp. 111–120. https://doi.org/10.5753/sbcup.2022 860.223205.

 

7. The following references are incomplete; Please check the references and write it in the required format and cite it accordingly. For example Ref.No.26 (Cited in 4.2.2) is cited more than a time, and their relation is not clear in the reference section. Also check ref.no 4,5,9.      Haykin, S. Redes Neurais: Princípios e Prática; Bookman Editora: Porto Alegre, Brasil, 2001. 3.IEEE Innovation at Work. Smart Grid and Renewable Energy 2023. Disponível em: https://innovationatwork.ieee.org. 813IEEE Smart Grid. Smart Grid and IoT for Sustainable Smart Cities: Potential, Applications and Future Research Directions 2021.Disponível em: https://smartgrid.ieee.org. GovInsider. How Thailand and New York are Modernizing the Smart Grid for Renewables and EVs 2023. Disponível em: https://govinsider.asia.de Freitas Velozo, L.; Mota, L.T.M. Bases de dados para Monitoramento Não-intrusivo da Carga: uma revisão.

 

As of now, 9/26 reference papers have to be verified as mentioned above.

 

In general typos, punctuation, and grammatical errors have to be verified.  The References section and their citation need to be improved, related to this proposed context.

Comments on the Quality of English Language

Needs improvement

Author Response

Comments 1: [In the abstract, the authors have mentioned that the implementation of artificial intelligence (AI) algorithms on resource-constrained edge devices is investigated in this study, including ESP32 and Raspberry Pi, for smart grid applications. But in the title, it is mentioned as Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge. There is no clear implementation for the use of classifying and detecting IoT devices for the smart meter devices.]

Response 1: [Thank you for your valuable observation. We understand the concern regarding the alignment between the title, abstract, and the core contributions of the work.

In the revised manuscript, we have clarified that the classification and detection of IoT electronic devices are indeed carried out using data collected by the smart meter and processed directly on extreme edge devices, namely the ESP32 and Raspberry Pi 3B+. The smart meter developed and used in this work is responsible for acquiring current and voltage signals from connected electronic devices, which are then used to train and execute lightweight machine learning models (MLP and KNN). These models were deployed and validated on the edge devices, demonstrating the feasibility of local classification.

To make this contribution clearer, we have revised the abstract to explicitly state the role of the smart meter as the data acquisition source and how the classification task is executed using the edge devices based on the signals it provides.

We hope these clarifications resolve the misunderstanding and reinforce the practical implementation of classification and detection using the proposed architecture.]

 

Comments 2: [In the abstract, Multi-Layer Perceptron (MLP) and K-Nearest Neighbors (KNN) models are employed to classify IoT electronic devices within the IAIoSGT architecture. What is the use in classifying and detecting IoT devices for the smart meter devices ? Are the authors pointing the performances of both the algorithms and choosing the best among ESP and Raspberry PI.]

Response 2: [A paragraph was added to the introduction discussing the importance of device identification and classification via smart meters, highlighting its potential for consumption profiling, equipment detection, and broader applications in energy management.

Identifying and classifying devices through smart meters allows researchers and utilities to map consumption profiles, detect the installation of new equipment or the replacement of existing ones, and thereby anticipate increases in energy usage. This approach also enables a range of other applications, such as supply optimization, anomaly detection, and support for intelligent energy management

In the conclusion, the results of both models and the hardware platforms used in the study are presented and discussed. Overall, the KNN model delivered excellent performance on both hardware platforms, while the MLP model performed well only on the Raspberry Pi. Therefore, the choice between one solution or the other should be based primarily on the available hardware and the specific requirements of the project.]

 

Comments 3: [Clarification is needed for the title and the same has to be incorporated in the manuscript.]

Response 3: [Thank you for your observation regarding the need for clarification related to the manuscript title. We agree that the connection between the title and the content should be explicit and well-aligned.

To address this, we have revised the abstract to clearly highlight that the classification and detection of IoT electronic devices are performed using data acquired by a smart meter and processed directly on resource-constrained edge devices, such as ESP32 and Raspberry Pi. The updated abstract now makes it explicit that the smart meter plays a central role in acquiring the electrical signals used for training and executing lightweight machine learning models for real-time device classification at the extreme edge.

Regarding the title, “Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge,” it accurately reflects the core contribution of the paper:

Machine Learning-Based Classification and Detection: Refers to the application of MLP and KNN models to identify connected IoT devices;

Using an IoT Smart Meter: Specifies that the data for this classification originates from a smart meter designed by the authors;

At the Extreme Edge: Emphasizes the novelty of deploying and executing the models directly on low-power devices (ESP32 and Raspberry Pi), instead of relying on cloud-based inference.

We hope this clarification and the revised abstract now effectively communicate the manuscript’s scope and ensure consistency with the title.]

 

Comments 4: [In the response to qn no.6, the authors have mentioned NO Security studies are involved. If so, why this paper is citeAlgarni, A.; Ahmad, Z.; Alaa Ala’Anzy, M. An Edge Computing-Based and Threat Behavior-Aware Smart Prioritization Framework for Cybersecurity Intrusion Detection and Prevention of IEDs in Smart Grids With Integration of Modified LGBM and One Class-SVM Models. IEEE Access 2024, 12, 104948–104963. https://doi.org/10.1109/ACCESS.2024.3435564.]

Response 4: [Thank you for your feedback regarding the citation of Algarni et al. (2024). Our study focuses on machine learning for device classification in smart grids using extreme edge devices, and does not address cybersecurity. The original paragraph in Section 2 (Related Work) discussing their cybersecurity framework was included to note related edge computing applications, but its specific focus on security could imply an unintended scope.

We have removed that paragraph and replaced it with a new one citing Algarni et al. (2024) as an example of edge-based solutions for smart grid enhancements, while clarifying that our work concentrates solely on device classification and energy management. This revision retains the citation to acknowledge complementary applications of edge computing and aligns with our study’s objectives. We appreciate your comment, which has improved the manuscript’s.]

 

Comments 5: [In the response to qn 28,29; Still the reference section is not clear.

Reference No: 1 AND 21 are the same and from the same author.]

Response 5: [Incorrect references have been corrected and duplicate entries have been removed.]

 

Comments 6: [Even though the following is the continued work. The self-citation may increase and therefore adhere to the guidelines of the journal. (Ref 1,20,21,22)

1.Marques, L.; Eugênio, P.; Bastos, L.; Santos, H.; Rosário, D.; Nogueira, E.; Cerqueira, E.; Kreutz, M.; Neto, A. Analysis of electrical Signals by Machine Learning for Classification of Individualized Electronics on the Internet of Smart Grid Things (IoSGT) architecture. Journal of Internet Services and Applications 2023, 14, 124–135. https://doi.org/10.5753/jisa.2023.3076.

20. Bastos, L.; Martins, B.; Santos, H.; Medeiros, I.; Eugênio, P.; Marques, L.; Rosário, D.; Nogueira, E.; Cerqueira, E.; Kreutz, M.; et al. Predictive Fraud Detection: An Intelligent Method for Internet of Smart Grid Things Systems. Journal of Internet Services and Applications 2023, 14, 1–15. https://doi.org/10.5753/jisa.2023.3077.
21. Marques, L.; Eugênio, P.; Bastos, L.; Santos, H.; Rosário, D.; Nogueira, E.; Cerqueira, E.; Kreutz, M.; Neto, A. Analysis of Electrical Signals by Machine Learning for Classification of Individualized Electronics on the Internet of Smart Grid Things (IoSGT) architecture. Journal of Internet Services and Applications 2023, 14, 124–135
22. Santos, H.; Eugênio, P.; Marques, L.; Oliveira, H.; Rosário, D.; Nogueira, E.; Neto, A.; Cerqueira, E. Internet of Smart Grid Things (IoSGT): Prototyping a Real Cloud-Edge Testbed. In Proceedings of the Proceedings of the 14th Brazilian Symposium on Ubiquitous and Pervasive Computing (SBCUP), Porto Alegre, RS, Brazil, 2022; pp. 111–120. https://doi.org/10.5753/sbcup.2022 860.223205.]

Response 6: [We appreciate your comment and the opportunity to clarify this point.

We have carefully reviewed the self-citations and made adjustments to reduce redundancy. Specifically, References [1] and [21] referred to the same prior work, and we have consolidated them to avoid duplication. As a result, the revised manuscript now includes a total of four self-citations.

These references were retained to ensure proper contextualization and to reflect the chronological development of the proposed solution:

One reference presents the initial architecture;

Another details the development of the smart meter used in this study;

The remaining two demonstrate previous applications that integrate the smart meter with the proposed architecture, validating both the system's evolution and practical applicability.

We believe these references are essential to demonstrate the technical continuity and validate the foundation upon which this current work is built.]

 

Comments 7: [The following references are incomplete; Please check the references and write it in the required format and cite it accordingly. For example Ref.No.26 (Cited in 4.2.2) is cited more than a time, and their relation is not clear in the reference section. Also check ref.no 4, 5, 9.

Haykin, S. Redes Neurais: Princípios e Prática; Bookman Editora: Porto Alegre, Brasil, 2001.

3. IEEE Innovation at Work. Smart Grid and Renewable Energy 2023. Disponível em: \\ https://innovationatwork.ieee.org. 813

IEEE Smart Grid. Smart Grid and IoT for Sustainable Smart Cities: Potential, Applications and Future Research Directions 2021.Disponível em: https://smartgrid.ieee.org.

GovInsider. How Thailand and New York are Modernizing the Smart Grid for Renewables and EVs 2023. Disponível em: https://govinsider.asia.de

Freitas Velozo, L.; Mota, L.T.M. Bases de dados para Monitoramento Não-intrusivo da Carga: uma revisão.

As of now, 9/26 reference papers have to be verified as mentioned above.

In general typos, punctuation, and grammatical errors have to be verified.  The References section and their citation need to be improved, related to this proposed contexto.]

Response 7: [All the referenced citations have been corrected and the others have been verified.

3. IEEE Innovation at Work. The Smart Grid and Renewable Energy, 2023. Acessado em: 13 abr. 2025.

4. Aggarwal, G.; Al-Greer, M.; Packiaraj, M.J.C.T. Smart Grid and IoT for Sustainable Smart Cities: Potential, Applications and 825 Future Research Directions, 2023. Acessado em: 13 abr. 2025.

5. Basu, M. How Thailand will integrate renewables and EVs into the grid, 2019. Acessado em: 13 abr. 2025.

9. de Freitas Velozo, L.; Mota, L.T.M. Bases de dados para Monitoramento Não-intrusivo da Carga: uma revisão. In Proceedings of the Anais do Brazilian Technology Symposium (BTSym’22), Campinas, Brasil, 2022. Acessado em: 13 abr. 2025.

26. Meyer-Baese, A.; Schmid, V. Chapter 7 - Foundations of Neural Networks. In Pattern Recognition and Signal Analysis in Medical Imaging, 2nd ed.; Meyer-Baese, A.; Schmid, V., Eds.; Academic Press: Oxford, 2014; pp. 197–243. https://doi.org/10.1016/B978-0-12-409545-8.00007-8.

27. Haykin, S. Redes Neurais: Princípios e Prática; Bookman Editora: Porto Alegre, Brasil, 2001. Acessado em: 13 abr. 2025.]

Round 3

Reviewer 4 Report

Comments and Suggestions for Authors

Appreciate the authors effort

Still, the title looks broader, suggesting that simplifying it will be good.

Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge

The revised manuscript now includes a total of four self-citations, and the authors have mentioned the need, but suggested confirming with the editor for the journal guidelines

Finally, check for typo errors. Recommended for strong proof read.

Comments on the Quality of English Language

Needs Improvement

Author Response

Comments 1: [Still, the title looks broader, suggesting that simplifying it will be good. Machine Learning-Based Classification and Detection of Electronic Devices Using an IoT Smart Meter at the Extreme Edge]
Response 1: [We appreciate the suggestion to simplify the title. The title has been revised to "Machine Learning-Based Classification of Electronic Devices Using an IoT Smart Meter", which we believe better reflects the scope and focus of the study.]

Comments 2: [The revised manuscript now includes a total of four self-citations, and the authors have mentioned the need, but suggested confirming with the editor for the journal guidelines]
Response 2: [Thank you for your observation. We have included four self-citations, which are essential to contextualize the current work within our previous research. As recommended, we will contact the editor to confirm compliance with the journal’s guidelines regarding self-citations.]

Comments 3: [Finally, check for typo errors. Recommended for strong proof read.]
Response 3: [We have conducted a detailed review of the manuscript to identify and correct potential typographical errors. We believe this revision improves the overall clarity and quality of the text.]