Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper deeply investigates the critical challenge of humidity effects on Electromechanical Impedance (EMI) technology in Structural Health Monitoring (SHM) and proposes an innovative method utilizing machine learning for compensation. The research design is rigorous, the experimental data is rich, and the application of machine learning models is also highly significant. The paper's structure is clear, and its argumentation is logical, offering important reference value for the future development of the EMI SHM field. However, I believe revisions are necessary before publication. My recommendation is for Major Revisions. Specific review comments are provided below:

1. In the Introduction section on page 2, specifically lines 37-38, the emphasis on research novelty is insufficient. It is recommended that, after mentioning existing humidity studies, the paper more clearly articulate the specific novelty of utilizing machine learning (particularly 1D-CNN and the imaginary part of EMI) for humidity compensation, to highlight its contribution to the current literature.

2. In the Samples section on page 4, specifically lines 119-124, there is a lack of specific details regarding the samples. It is recommended to provide the concrete mix proportions (e.g., cement, aggregate, water-cement ratio), the specific model and dimensions of the PZT sensors, the material type, diameter, and surface treatment (if any) of the steel reinforcement, and the specific product model and curing conditions of the conductive epoxy.

3. In the Test setup - Data collection experiment section on page 5, specifically line 135, the description of humidity control is not precise enough. It is recommended to clarify how long the samples were kept under the new humidity conditions after each adjustment, and whether they reached a stable state before measurements were taken.

4. In the Test setup - Data collection experiment section on page 6, specifically lines 138-139, the details regarding temperature control are insufficient. It is recommended to confirm the temperature fluctuation range within the humidity chamber and discuss its potential impact on the results.

5. In the field of structural health monitoring, environmental factors such as temperature, humidity, and wind can affect monitoring results to varying degrees. While the author examines the influence of humidity in this paper, temperature variations can induce changes in structural responses—sometimes even exceeding those caused by structural damage (Damage Identification of Bridge Structures Considering Temperature Variations-Based SVM and MFO)—and may have a greater impact on structural responses than humidity. This viewpoint is widely acknowledged in areas such as damage identification (Damage Identification of a Steel Frame Based on Integration of Time Series and Neural Network under Varying Temperatures) and model updating. The author is requested to provide further clarification on this aspect.

6. In the Training the ML model section on page 6, specifically lines 152-154, the explanation of the experimental malfunction is inadequate. It is recommended to provide a more detailed explanation of the nature of the data unreliability between day #60 and day #90, and what corrective measures were taken.

7. In the Training the ML model section on page 6, specifically lines 155-159, the ML model architecture and training details are not sufficiently detailed. It is recommended to provide specific network architecture information for each layer (e.g., convolutional layers, pooling layers, fully connected layers, kernel size, stride, number of filters), training parameters such as optimizer, learning rate, batch size, and number of epochs, and to describe the method of splitting the dataset into training, validation, and test sets.

8. In the Real part (R) section on page 7, specifically line 171, the quantitative analysis of the RMSD results is insufficient. It is recommended to provide a quantitative correlation coefficient to support the observation of a strong correlation between RMSD and humidity.

9. In Table 2 on page 9, the explanation of prediction error is unclear. It is recommended to clarify whether "Avg. Error [RH %]" represents the average of absolute or relative error, and to provide the standard deviation or confidence interval for each model.

10. What are the advantages of the damage detection method based on the Electromechanical Impedance (EMI) technique compared to approaches utilizing swarm intelligence optimization algorithms? The author is asked to elaborate on this point, especially considering the widespread application of swarm intelligence optimization algorithms in recent years, such as MFO (Structural damage identification based on modal frequency strain energy assurance criterion and flexibility using enhanced Moth-Flame optimization) and SFO (Two-stage damage identification for bridge bearings based on sailfish optimization and element relative modal strain energy). The author is expected to provide a discussion based on the EMI method.

11. In Table 2 and Figure 11 on page 9, regarding the issue of high error for the "Full |Z|" model in Table 2, while the Discussion mentions low error for "Peaks_Region |Z|". It is recommended to revise Table 2 or provide a clearer explanation of the distinction between these two in the Discussion to avoid confusion.

12. In the Discussion section on page 10, specifically line 200, the subsequent suggestions for addressing the limitation of small sample size are insufficient. It is recommended to provide further suggestions on how future research can address this issue.

13. In the Discussion section on page 10, specifically lines 205-209, the potential for distinguishing between damage and humidity effects could be explored further. It is recommended to discuss more specifically how these effects can be effectively separated through time-series analysis or multi-modal sensing, thereby enhancing the reliability of SHM.

14. In the Discussion section on page 10, specifically lines 209-210, there is a lack of practical application suggestions. It is recommended to discuss more specifically how these findings can be utilized to design more robust EMI SHM systems.

15. In Figure 11 on page 9, due to the excessive information in the legend, it is difficult to distinguish between different models. It is recommended to optimize the legend, for example, by grouping errors by range or highlighting the best-performing models, to improve readability.

Author Response

Comments 1: In the Introduction section on page 2, specifically lines 37-38, the emphasis on research novelty is insufficient. It is recommended that, after mentioning existing humidity studies, the paper more clearly articulate the specific novelty of utilizing machine learning (particularly 1D-CNN and the imaginary part of EMI) for humidity compensation, to highlight its contribution to the current literature.

Response 1: First of we wish to thank you for taking the time to review and comment our work. In line 38 and 40 we mention two studies and mention how the effect of temperature on the EMI signature has well studied, but the effect of humidity, while critically important, remains an understudied area. This work attempts to highlight the critical impact humidity has on the robustness of EMI based SHM. The work also argues that humidity more strongly affects the imaginary component as compared to damage or temperature and shows in Table 2 and the discussion that the imaginary component in the low frequency region of the EMI signature works well with a ML scheme, in this case a 1D-CNN which may not be the optimal choice as a comparison between models would be needed to determine the optimal model type and parameters.

Comments 2: In the Samples section on page 4, specifically lines 119-124, there is a lack of specific details regarding the samples. It is recommended to provide the concrete mix proportions (e.g., cement, aggregate, water-cement ratio), the specific model and dimensions of the PZT sensors, the material type, diameter, and surface treatment (if any) of the steel reinforcement, and the specific product model and curing conditions of the conductive epoxy.

Response 2: Section 2.1 has been updated to include the requested information. "In total, three concrete blocks were fabricated for the experiments. Each block contained a 316 stainless steel rod with an attached PZT patch, bonded using a conductive epoxy, and hollow 3D-printed end covers to limit vibration attenuation from the surrounding concrete, as shown in Fig. 3. The PZT transducers were 10 mm × 10 mm × 1.5 mm APC-840 patches (American Piezo), bonded with CW2400 conductive epoxy (Chemtronics), which was cured at room temperature for 72 hours. The rods were encased in fast-setting concrete from Skalflex, mixed at a water-to-cement ratio of approximately 1:11, and cured in the laboratory for 28 days."

Comments 3: In the Test setup - Data collection experiment section on page 5, specifically line 135, the description of humidity control is not precise enough. It is recommended to clarify how long the samples were kept under the new humidity conditions after each adjustment, and whether they reached a stable state before measurements were taken.

Response 3: The section has been updated, removing duplicate information and adding clarification about the time in the chamber: "The RH in the chamber was increased approximately once per week when the impedance measurements where stable, this was done to ensure the concrete had reached full saturation at the target RH level."

Comments 4: In the Test setup - Data collection experiment section on page 6, specifically lines 138-139, the details regarding temperature control are insufficient. It is recommended to confirm the temperature fluctuation range within the humidity chamber and discuss its potential impact on the results.

Response 4: Details regarding the mean and standard deviation of the temperature over the duration of test has been added on line 140. "The chamber was indirectly cooled and heated by air from a walk-in storage room maintained at a mean temperature of 25.24 C◦ with a standard deviation of 0.37 C◦ over the duration of the experiment. Because the observed temperature variation was minimal, the effect of temperature variance was assumed to have a negligible effect and excluded from subsequent analysis. "

Comments 5: In the field of structural health monitoring, environmental factors such as temperature, humidity, and wind can affect monitoring results to varying degrees. While the author examines the influence of humidity in this paper, temperature variations can induce changes in structural responses—sometimes even exceeding those caused by structural damage (Damage Identification of Bridge Structures Considering Temperature Variations-Based SVM and MFO)—and may have a greater impact on structural responses than humidity. This viewpoint is widely acknowledged in areas such as damage identification (Damage Identification of a Steel Frame Based on Integration of Time Series and Neural Network under Varying Temperatures) and model updating. The author is requested to provide further clarification on this aspect.

Response 5: It is widely acknowledged that enviromental factors have a large effect on the robustness of SHM techniques. Especially the effect of temperature and damage has been widely studied in the context of the EMI technique, indeeded one of our previous works looks at this very topic. The effect of humidity in regard to the EMI technique in reinforced concrete is a under studied area. Our work confirms the importance of considering humidity to get robust SHM using the EMI technique. Future work should focus on larger sample sizes and analyzing combined temperature, damage and humidity effects.

Comment 6: In the Training the ML model section on page 6, specifically lines 152-154, the explanation of the experimental malfunction is inadequate. It is recommended to provide a more detailed explanation of the nature of the data unreliability between day #60 and day #90, and what corrective measures were taken.

Response 6: A more detailed explanation has been added: "On day #60 a physical connection on Sample1 became loose, on day #79 the samples were taken out of the chamber and the problem corrected. Moving the samples caused the dip in RMSD score as seen on Fig. 7. For this reason, data collected between day #60 and day #90 were excluded from the analysis."

Comments 7: In the Training the ML model section on page 6, specifically lines 155-159, the ML model architecture and training details are not sufficiently detailed. It is recommended to provide specific network architecture information for each layer (e.g., convolutional layers, pooling layers, fully connected layers, kernel size, stride, number of filters), training parameters such as optimizer, learning rate, batch size, and number of epochs, and to describe the method of splitting the dataset into training, validation, and test sets.

Response 7: Section 2.3 has been rewritten and a new table has been included showing the hyperparameter search space "The dataset was randomly split into 80% training and 20% validation subsets using the random_split() function from PyTorch. All models share the same network topology, consisting of a three-layer one-dimensional convolutional neural network (1D-CNN) followed by a fully connected regression head. Each convolutional layer is followed by a ReLU activation function, and global feature aggregation is performed using adaptive average pooling. The fully connected part consists of a hidden layer with ReLU activation and dropout regularization, followed by a single linear output neuron.

Hyperparameter optimization was performed using Optuna. The optimized hyperparameters include the number of convolutional filters, kernel sizes, stride lengths, hidden layer dimension, dropout probability, and learning rate. The full hyperparameter search space is summarized in Table 1, wide hyperparameter ranges were used to avoid biasing the model design allowing Optuna to find suitable configurations for each frequency range and feature. For each experiment, the model configuration with the lowest validation loss was selected.

Models were trained using the Adam optimizer with a batch size of 64 for 300 epochs. This configuration was chosen to ensure stable optimization and convergence across all hyperparameter configurations. In total, 16 models were trained, corresponding to a 4*4 combination of frequency ranges and impedance data representations, as summarized in Table 3."

Comments 8:  In the Real part (R) section on page 7, specifically line 171, the quantitative analysis of the RMSD results is insufficient. It is recommended to provide a quantitative correlation coefficient to support the observation of a strong correlation between RMSD and humidity.

Response 8: The Correlation coefficient has been added to the figure and the figure text has been updated "Output of RMSD algorithm applied to R in the peaks region (135kHz-150kHz) showing a strong correlation coefficient between RMSD and humidity."

Comments 9: In Table 2 on page 9, the explanation of prediction error is unclear. It is recommended to clarify whether "Avg. Error [RH %]" represents the average of absolute or relative error, and to provide the standard deviation or confidence interval for each model.

Response 9: The term average error has been removed and replaced be the correct term, Mean Absolute Error (MAE) to improve clarity. The STD between the target and the predictions have been added to table 2.

Comments 10: What are the advantages of the damage detection method based on the Electromechanical Impedance (EMI) technique compared to approaches utilizing swarm intelligence optimization algorithms? The author is asked to elaborate on this point, especially considering the widespread application of swarm intelligence optimization algorithms in recent years, such as MFO (Structural damage identification based on modal frequency strain energy assurance criterion and flexibility using enhanced Moth-Flame optimization) and SFO (Two-stage damage identification for bridge bearings based on sailfish optimization and element relative modal strain energy). The author is expected to provide a discussion based on the EMI method.

Response 10: The EMI technique is a well documented non-destructive SHM technique, especially in industrial and areospace engineering. In the last few years there has been an increased intrest in using the EMI technique for concrete structures. This work does not aim to make a comparison between various SHM techniques but rather attempt to improve robustness in systems utilizing the EMI technique, specifically for reinforced concrete structures.

Comments 11: In Table 2 and Figure 11 on page 9, regarding the issue of high error for the "Full |Z|" model in Table 2, while the Discussion mentions low error for "Peaks_Region |Z|". It is recommended to revise Table 2 or provide a clearer explanation of the distinction between these two in the Discussion to avoid confusion.

Response 11: Table 1 fully explains what is meant by each "Frequency range" and each "Feature" a line has been added to section 3.1.5 referring back to table 1 "Refer to table 1 for explanation of each Frequency Range and Feature."

Comments 12: In the Discussion section on page 10, specifically line 200, the subsequent suggestions for addressing the limitation of small sample size are insufficient. It is recommended to provide further suggestions on how future research can address this issue.

Response 12: We have already adressed the limitation of small sample size. It is clear that future research can benefit from larger sample sizes, but we are unable to recommend an optimal sample size for future works.

Comments 13: In the Discussion section on page 10, specifically lines 205-209, the potential for distinguishing between damage and humidity effects could be explored further. It is recommended to discuss more specifically how these effects can be effectively separated through time-series analysis or multi-modal sensing, thereby enhancing the reliability of SHM.

Response 13: A line has been added highlighting how temporal analysis likely can be used to enhance the reliability of SHM. "Additionally, a temporal analysis may increase the reliability of damage detection as, similarly to PZT debonding, damage will only increase."

Comments 14: In the Discussion section on page 10, specifically lines 209-210, there is a lack of practical application suggestions. It is recommended to discuss more specifically how these findings can be utilized to design more robust EMI SHM systems.

Response 14: More details about practical advice and the direction of future work has been added to the dicussion "Our results indicate that the robustness of the EMI technique can be improved by explicitly considering the imaginary component X. In cases where a large RMSD is observed together with large variations in X, particularly in the lower frequency range, the response is likely influenced by ambient humidity changes and not only structural damage. Future work should focus on quantifying and generalizing the relationship between humidity and the EMI signature through experiments on a large and diverse set of samples, with the ultimate goal of developing a robust humidity compensation strategy, similar to the existing temperature compensation approaches."

Comments 15: In Figure 11 on page 9, due to the excessive information in the legend, it is difficult to distinguish between different models. It is recommended to optimize the legend, for example, by grouping errors by range or highlighting the best-performing models, to improve readability.

Response 15: The figure has been updated to only include the top 5 best performing models to improve readability.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The presented results shows use of electro-mechanial impedance observed by the pzt sensor attached to the steel bar imbedded in the concrete block to monitor humidity. This approach has been a research topic to monitor concrete curing. In this aspect it would be important to compare the sensitivity with the previously suggested methodology to verify the advantage of the proposed method. For publication, following aspects should be discussed.

1. Figure 3 shows the bar imbedded in the concrete block. Actual picture of the experiemental apparatus is required to be included.
2. The results in Figures 1 and 2 are not required to be included. Rather, it is recommended to include the results as a comparison with the measured results. 
3. Some numerical model that has influence on the measrued impedance is required to overcome the limitation of the small specimen size used in the experiments.
4. The predicted humidity shown in Fig. 11 is rather show limited accuracy to be used as a sensor. Since the machine learning is used, the accuracy should be improved for the prediticted results published. 

Author Response

Comments 1: Figure 3 shows the bar imbedded in the concrete block. Actual picture of the experiemental apparatus is required to be included.

Response 1: We agree and have added a picture of the steel sensor to figure 3.

Comments 2: The results in Figures 1 and 2 are not required to be included. Rather, it is recommended to include the results as a comparison with the measured results. 

Response 2: The results in Figures 1 and 2 use the well established model of the EMI technique using 'Liang's model'. With the figures we attempt to show how damage and humidity have similar effect on the real part of the impedance, but only the humidity affects the imaginary part. This is used to argue that the imaginary part of the impedance can be used to seperate the effects of humidity from the effects of damage leading to more robust SHM.

Comments 3: Some numerical model that has influence on the measrued impedance is required to overcome the limitation of the small specimen size used in the experiments.

Response 3: Computing a numerical model is an important task. The effect of humidity is still an under represented area in the literature and based on the observations in our work, there are still too many unknown variables to accurately create a model. Future work will undoubtedly focus on this problem

Comments 4: The predicted humidity shown in Fig. 11 is rather show limited accuracy to be used as a sensor. Since the machine learning is used, the accuracy should be improved for the prediticted results published. 

Response 4: The results summeraised in table 2 show a minimum Mean Absolute Error of 2%RH, which is very accurate. It is clear however that these results may not generalize, but we mention this in the discussion and future work will focus on large sample size tests.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript entitled

 

"Electromechanical Impedance Sensing Under Humid Conditions: Experimental Insights and Compensation using Machine Learning."

 

by Mads Kofod Dahl et al.

 

investigates the influence of ambient humidity on electromechanical impedance (EMI) measurements in steel-reinforced concrete and explores the use of machine learning to infer relative humidity from EMI signatures. The study focuses on embedded piezoelectric sensors bonded to reinforcement bars and combines theoretical considerations based on Liang’s model with long-term experimental data collected under controlled humidity conditions..

 

The topic is relevant to the field of structural health monitoring (SHM), particularly for reinforced concrete infrastructure exposed to variable environmental conditions. The emphasis on humidity effects—an aspect often overshadowed by temperature compensation studies—is timely, and the use of embedded transducers rather than surface-mounted sensors is a notable strength. The manuscript also provides extensive experimental data and a clear qualitative demonstration that humidity can significantly affect common damage indices such as RMSD.

 

However, despite these strengths, the manuscript presents several substantive issues related to experimental assumptions, methodological rigor, machine-learning validation, and the scope of the conclusions. As a result, the work requires major revision before it can be considered for publication. The main concerns are detailed below.

 

Major remarks:

1. A central assumption of the machine-learning framework is that the internal relative humidity of the concrete specimens increases linearly from 20% to 80% over the duration of the experiment. This assumption is not experimentally validated and contradicts well-established nonlinear moisture diffusion and sorption behavior in concrete. Since this assumed internal humidity is used as the regression target, the reported prediction errors cannot be interpreted as true accuracy with respect to the physical state of the material. This limitation significantly weakens the quantitative conclusions and must be addressed more rigorously, either through direct measurement, physical modeling, or a clear reframing of the results as correlations with exposure time rather than true internal humidity.
2. The experimental study is based on only three concrete specimens, all produced with similar geometry, materials, and sensor configurations. While the authors acknowledge this limitation, the small sample size severely restricts the generalizability of the findings, particularly for machine-learning applications. Sensor-to-sensor variability, bonding layer effects, and microstructural differences in concrete are known to strongly influence EMI signatures, yet these factors cannot be meaningfully assessed with the current dataset.
3. The choice of a three-layer one-dimensional convolutional neural network is not adequately justified. No comparison is provided against simpler baseline models (e.g., linear regression, principal-component-based regression, or tree-based methods), making it difficult to assess whether the CNN architecture is necessary or beneficial. Moreover, the manuscript does not clearly describe the data splitting strategy for training, validation, and testing. Given the strong temporal correlation in the dataset, improper splitting could lead to data leakage and overly optimistic performance estimates.
4. While the manuscript convincingly demonstrates that humidity strongly affects EMI signatures and RMSD values, it does not implement or validate a true humidity compensation strategy for damage detection. Claims related to “robust SHM systems” and damage–humidity separation are therefore premature. In particular, the assertion that the imaginary part of the impedance is minimally affected by damage is not experimentally validated within this study and relies primarily on theoretical arguments and prior literature.
5. The proposed approach shows several complementarities with the framework of self-sensing concrete. Thus, a brief state-of-the-art review in this sense can be introduced, referring, for instance, at Recent advances in embedded technologies and self-sensing concrete for structural health monitoring.
6. There are several minor but recurring language and terminology issues, such as the use of “dampening” instead of “damping,” duplicated section titles, and occasional awkward phrasing. While these do not obscure the technical content, careful language editing would improve clarity and professionalism.
7. The introduction can also be extended in terms of current research works in humidity-induced effects in building materials, e.g., Estimating Internal Moisture Content Effects for the Vibration-Based Damage Assessment of Cross-Laminated Timber Buildings – Preliminary Results.
8. Machine-learning results are reported primarily in terms of average prediction error. Additional metrics such as standard deviation, confidence intervals, or sensor-specific performance would provide a more complete picture of model robustness.

Concerning the English, the manuscript is generally understandable and the technical meaning can be followed throughout; however, the quality of the English requires moderate but systematic revision to meet journal standards.

Author Response

Comments 1: A central assumption of the machine-learning framework is that the internal relative humidity of the concrete specimens increases linearly from 20% to 80% over the duration of the experiment. This assumption is not experimentally validated and contradicts well-established nonlinear moisture diffusion and sorption behavior in concrete. Since this assumed internal humidity is used as the regression target, the reported prediction errors cannot be interpreted as true accuracy with respect to the physical state of the material. This limitation significantly weakens the quantitative conclusions and must be addressed more rigorously, either through direct measurement, physical modeling, or a clear reframing of the results as correlations with exposure time rather than true internal humidity.

Response 1: First off all, we would like to thank you for your time and effort in reading, understanding and commenting on our work. This is a good point as it targets a key assumption in this work and is likely at least somewhat inaccurate. A line has been added to the conclusion discussing this and how it relates to our results and future work "From the results presented on Fig. 11 it is clear that the majority of the errors appear at the most dry and most humid ends of the dataset. This is likely a results of assuming a linear absorption curve in the RC where an S-formed absorption curve may be more accurate. Future works should further investigate the nonlinear absorption curves of concrete to improve the accuracy of the predictions.  "

Comments 2: The experimental study is based on only three concrete specimens, all produced with similar geometry, materials, and sensor configurations. While the authors acknowledge this limitation, the small sample size severely restricts the generalizability of the findings, particularly for machine-learning applications. Sensor-to-sensor variability, bonding layer effects, and microstructural differences in concrete are known to strongly influence EMI signatures, yet these factors cannot be meaningfully assessed with the current dataset.

Response 2: The small sample size is a limitation of the study and future work should focus on larger sample size data collection experiments as well as focusing on extending the generalizability of the findings. As for our work, we test samples using the same geometry, concrete, steel and PZT type. We mention how even between 3 very similar samples, the EMI signatures appear different, but even with these differences it was possible to extract meaningful patterns between the EMI signature dataset and ambient relative humidity. As mentioned, this is an understudied area and more research is needed.

Comments 3: The choice of a three-layer one-dimensional convolutional neural network is not adequately justified. No comparison is provided against simpler baseline models (e.g., linear regression, principal-component-based regression, or tree-based methods), making it difficult to assess whether the CNN architecture is necessary or beneficial. Moreover, the manuscript does not clearly describe the data splitting strategy for training, validation, and testing. Given the strong temporal correlation in the dataset, improper splitting could lead to data leakage and overly optimistic performance estimates.

Response 3: : In the introduction we mention how the effect of temperature on the EMI signature is well studied, but the effect of humidity, while critically important, remains an understudied area. This work attempts to highlight the critical impact humidity has on the robustness of EMI based SHM. The work also argues that humidity more strongly affects the imaginary component as compared to damage or temperature and shows in Table 2 and the discussion that the imaginary component in the low frequency region of the EMI signature works well with a ML scheme, in this case a 1D-CNN which may not be the optimal choice as a comparison between models would be needed to determine the optimal model type and parameters. A line about the data splitting strategy has been added to section 2.3 "The data collected was split in an 80%/20% training/validation set using the random_split() function from the PyTorch library."

Comments 4: While the manuscript convincingly demonstrates that humidity strongly affects EMI signatures and RMSD values, it does not implement or validate a true humidity compensation strategy for damage detection. Claims related to “robust SHM systems” and damage–humidity separation are therefore premature. In particular, the assertion that the imaginary part of the impedance is minimally affected by damage is not experimentally validated within this study and relies primarily on theoretical arguments and prior literature.

Response 4: In the work we highlight the importance of considering humidity by demonstrating the effect on RMSD values. As indicated by table 2 and mentioned in the discussion, our results show that the imaginary part of the impedance has a strong link with humidity and robust humidity compensation likely should focus on using the imaginary part. In the literature, it is a well observed phenomenon that the imaginary part, especially in the lower frequency range, is minimally affected by damage, we use this fact to strengthen our claim that the imaginary part is a good candidate for use in a future humidity compensation strategy. We also do not experimentally validate that the real part is sensitive to damage, but rely on theoretical arguments and prior works.

Comments 5: The proposed approach shows several complementarities with the framework of self-sensing concrete. Thus, a brief state-of-the-art review in this sense can be introduced, referring, for instance, at Recent advances in embedded technologies and self-sensing concrete for structural health monitoring.

Response 5: An excellent review article that is now being mentioned in the introduction "Within the broader context of smart structures with self-sensing capabilities, much work has been done using piezoelectric sensors [4], yet embedded EMI-based approaches remain comparatively understudied. This is despite it's advantages, EMI-based SHM still faces challenges that limit its broader adoption."

Comments 6: There are several minor but recurring language and terminology issues, such as the use of “dampening” instead of “damping,” duplicated section titles, and occasional awkward phrasing. While these do not obscure the technical content, careful language editing would improve clarity and professionalism.

Response 6: Minor changes have been made to the manuscript that hopefully improve the language and terminology use.

Comments 7: The introduction can also be extended in terms of current research works in humidity-induced effects in building materials, e.g., Estimating Internal Moisture Content Effects for the Vibration-Based Damage Assessment of Cross-Laminated Timber Buildings – Preliminary Results.

Response 7: The work mentioned, while interesting in it's own, investigates cross-laminated timber using vibration-based Structural Health Monitoring techniques. The work is not related to concrete, steel reinforcements or EMI based SHM. We feel that the work should not be mentioned due to a lack of overlap or relevance.

Comments 8: Machine-learning results are reported primarily in terms of average prediction error. Additional metrics such as standard deviation, confidence intervals, or sensor-specific performance would provide a more complete picture of model robustness.

Response 8: Table 2 has been updated to include the standard deviation between the predictions and targets for each frequency range and feature.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

(1) What do the quotation marks on the first page mean? Has the author not carefully revised and checked the paper's formatting?

(2) As mentioned in previous comment #5 from Round 1 of the review, the author only addressed the influence of environmental factors such as temperature and humidity in the response but did not supplement the paper with relevant current hotspot literature. Please make the necessary revisions.

(3) As mentioned in previous comment #10 from Round 1 of the review, the author has shown insufficient understanding of the advantages of the damage detection method based on the Electromechanical Impedance (EMI) technique compared to approaches using swarm intelligence optimization algorithms. Relevant references have not been consulted or reviewed. Please supplement this section accordingly.

(4) Lines 63–117 should be moved to the section of Methods.

(5) At the end of the Introduction chapter, please add a summary of the paper's main contributions and an overview of the research content in each section.

Author Response

Comments 1: What do the quotation marks on the first page mean? Has the author not carefully revised and checked the paper's formatting?

Response 1: This is a formatting issue due to the MDPI sensors template, we have sent an e-mail to them asking for help in correcting the issue after submission.

Comments 2: As mentioned in previous comment #5 from Round 1 of the review, the author only addressed the influence of environmental factors such as temperature and humidity in the response but did not supplement the paper with relevant current hotspot literature. Please make the necessary revisions.

Response 2: In the manuscript we briefly mention the effect of temperature and our previous work on it to contrast the relatively low amount of attention that humidity has recieved despite how large of an impact humidity has on the EMI signature. Some works have looked at the drying process of wet cement to monitor the strength gain, this is tangentiel to our work, but is not mentioned as we work with set concrete. The manuscript deals with the effect of humidity with the temperature being stable doing the experiment it's effect is assumed to be neglible, many works detailing the effect of temperature excists and the reviewer is encouraged to see our previous work for more information about this and the hotspot literature in temperature effects on EMI.

Comment 3: As mentioned in previous comment #10 from Round 1 of the review, the author has shown insufficient understanding of the advantages of the damage detection method based on the Electromechanical Impedance (EMI) technique compared to approaches using swarm intelligence optimization algorithms. Relevant references have not been consulted or reviewed. Please supplement this section accordingly.

Response 3: The EMI technique is a well established and well researched technique, this manuscript is not a review of currently used metods for strucutral health monitoring and does not seek out to argue why EMI is better or worse than other techniques, for example swarm intelligence optimization algorithms. This work sets out to improve on the EMI technique, and does not involve any direct comparisons with other SHM techniques, mentioning swarm intelligence optimization algorithms is not relevant.

Comments 4: Lines 63–117 should be moved to the section of Methods.

Response 4: We respectfully disagree, the lines 63-117 is related to the background theory of the EMI technique and not the methods used for samples or analysis, therefore does not belong in the Methods section.

Comments 5: At the end of the Introduction chapter, please add a summary of the paper's main contributions and an overview of the research content in each section.

Response 5: A paragraph has been added to the introduction summerazing the papers main contributions "This work provides an experimental investigation of the influence of ambient humidity on the EMI technique, demonstrating that robust SHM based on EMI must account for humidity fluctuations. Using the collected experimental data, we show that different impedance representations and frequency ranges exhibit different sensitivities to humidity. In particular, we demonstrate that the real part of the impedance is a poor indicator of humidity, whereas the imaginary part shows significantly stronger predictive capability. This conclusion is supported by the results of a ML prediction. Overall, this work highlights humidity as a critical challenge for EMI-based SHM and suggest a path forward through the development of humidity compensation schemes using machine learning models trained on the imaginary component of the impedance."  The research content in each section is covered in line 58-62 "The structure of the paper is as follows: I. Introduction and theoretical exploration using Liang’s model [ 13], illustrating the effect of humidity on the EMI technique. II. Methods describing the samples, humidity chamber setup, and ML model training. III. Results analyzing the humidity effect and evaluating ML model performance. IV. Discussion of findings and limitations. V. Conclusions."

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Reviewer comments were addressed with relevant revisions. 

Author Response

Reviewer comments were addressed with relevant revisions. 

Reviewer 3 Report

Comments and Suggestions for Authors

This Reviewer is satisfied with the replies by the Authors and the current state of the manuscript.

Author Response

This Reviewer is satisfied with the replies by the Authors and the current state of the manuscript.