Fault Diagnosis in Analog Circuits Using a Multi-Input Convolutional Neural Network with Feature Attention

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Multi-input constructive network model based on convolutional neural network and attention mechanism are proposed by the authors. The designed model demonstrates better comprehensive performance in fault diagnosis experiments of analog circuits, efficient classification and location of all faults. The experimental results of the Butterworth low-pass filter and two-stage four op-amp double-order low-pass filter are represented.

My following comments should be taken into consideration to revise the paper version.

1. The manuscript contains Figure 1 and Figure 2 with the same figure captions. There is no description of Figure 2 in text.
2. The blocks of Figure 3 don’t have names. There is no detail description of Figure 3 in text.
3. It would be better for clarity to describe the use of wavelet transform (WTF) and principal component analysis (PCA) in method WTF + PCA + ELM and etc. (Table 2, Table 4).
4. The quality of Figure 19 is bad.

Author Response

Response to reviewer 1

Dear Editor and Reviewers,

We sincerely thank you for all positive and constructive comments of the manuscript entitled “Fault diagnosis in analog circuits using a multi-input convolutional neural network with feature attention” (the manuscript number is computation-3545756). Your suggestions and comments will be of great significance to our next research and writing. We have studied the comments carefully and made all necessary modifications to the original version. In the revised manuscript, we have used different colors to mark the main changes according to the opinions or suggestions of different reviewers, so as to facilitate the review. We have resubmitted the revised manuscript to the Submission system, and we hope it could be considered for publication ultimately. Should you have any questions, please contact us without hesitation. The point-by-point replies to all reviewers are as follows:

Replies to Reviewer 1:

Reviewer 1:

Comments to the Author

Multi-input constructive network model based on convolutional neural network and attention mechanism are proposed by the authors. The designed model demonstrates better comprehensive performance in fault diagnosis experiments of analog circuits, efficient classification and location of all faults. The experimental results of the Butterworth low-pass filter and two-stage four op-amp double-order low-pass filter are represented.

My following comments should be taken into consideration to revise the paper version.

Comment 1:

1. The manuscript contains Figure 1 and Figure 2 with the same figure captions. There is no description of Figure 2 in text.

Response 1:

We sincerely appreciate your correction. We deeply regret using the same icons for Figure 1 and Figure 2. In the revised version, we have already made the necessary corrections. Additionally, we have provided a description of Figure 2.

(For specific revisions, please read the sections highlighted in red in the 2.2. MI-CNN of the revised manuscript.)

Comment 2:

2. The blocks of Figure 3 don’t have names. There is no detail description of Figure 3 in text.

Response 2:

Thank you for your valuable feedback. We sincerely apologize for the lack of detailed descriptions in the text and the absence of block names in Figure 3. In the revised manuscript, we have addressed this issue by adding a detailed textual explanation of Figure 3, explicitly describing each step involved in the ECA module. Additionally, we have updated Figure 3 by labeling each block to improve clarity and better illustrate the structure of the ECA module. We appreciate your insightful comments, which have helped us enhance the quality of our manuscript.

(For specific revisions, please read the sections highlighted in red in the 2.3. ECA module of the revised manuscript.)

Comment 3:

2. It would be better for clarity to describe the use of wavelet transform (WTF) and principal component analysis (PCA) in method WTF + PCA + ELM and etc. (Table 2, Table 4).

Response 3:

Thank you for your valuable feedback regarding the clarity of the "WTF + PCA + ELM" method. In response to your comments, we have revised the manuscript to provide a more detailed description. We first explained the necessity of applying WTF+PCA, mainly because traditional machine learning methods have limited feature extraction capabilities, and the reconstructed data often exhibits insufficient differences. Secondly, we elaborated on the distinct roles of WTF and PCA in the revised manuscript. Finally, we described how WTF and PCA are implemented. Specifically, WTF first applies the ‘wpdec’ function to perform a three-level wavelet packet decomposition of the original 9 fault signals, constructing the wavelet packet tree. PCA, on the other hand, retains the most significant components that explain the maximum variance in the data by comparing the contribution rate and cumulative contribution rate of each feature, resulting in a more compact and efficient feature representation.

(For specific revisions, please read the sections highlighted in red in the revised manuscript.)

Comment 4 :

2. 2. The quality of Figure 19 is bad.

Response 4:

Thank you for your valuable feedback regarding the quality of Figure 19. We apologize for the poor quality of the figure. In the revised manuscript, we have replaced the figure with a higher-resolution version to ensure better clarity and readability. We appreciate your attention to this detail.

(For specific revisions, please read the sections highlighted in red in the Figure 19 of the revised manuscript.)

Special thanks to you for your valuable comments.

 

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript presents an innovative approach to fault diagnosis in analog circuits using the ECA-MI-CNN model. The paper is well-structured and provides a comprehensive comparison of various models. However, there are several areas need to be improved.

 

The introduction provides a good overview of the challenges in analog circuit fault diagnosis. Could you elaborate on why traditional fault dictionary models are still relevant or not in certain scenarios?

For the introduction, deep learning methods are well-justified, it remains unclear why existing approaches (e.g., DBN, GAN) fall short in the context of analog circuit diagnosis.

For ECA module, could you provide more experimental evidence showing how the choice of k (interaction range) impacts model performance? For instance, does increasing k beyond a certain threshold lead to diminishing returns?

The MI-CNN's ability to fuse time-domain and frequency-domain features is a major strength. Could you include visual examples or feature maps showing how these fused features differ qualitatively from single-domain features extracted by traditional CNNs.

For the MI-CNN, could you clarify whether the fully connected layer in N3 introduces any bottlenecks or overfitting risks? If so, how are these mitigated?

For the butterworth low-pass filter, they may not fully showcase the robustness of the proposed model. Could you include additional experiments on circuits with varying levels of noise or environmental interference? Why an offset of 1064 was chosen?

Could you provide more detailed insights into why CBAM performs better than SENet but still falls short compared to ECA?

One key limitation of deep learning models in engineering applications is their "black-box" nature. Could you provide more mechanistic insights into how specific features (e.g., frequency-domain anomalies) correspond to physical faults in circuits?

Figure 19 appears to have a format issue, as part of the image is obscured.

Author Response

Response to reviewer 2

Dear Editor and Reviewers,

We sincerely thank you for all positive and constructive comments of the manuscript entitled “Fault diagnosis in analog circuits using a multi-input convolutional neural network with feature attention” (the manuscript number is computation-3545756). Your suggestions and comments will be of great significance to our next research and writing. We have studied the comments carefully and made all necessary modifications to the original version. In the revised manuscript, we have used different colors to mark the main changes according to the opinions or suggestions of different reviewers, so as to facilitate the review. We have resubmitted the revised manuscript to the Submission system, and we hope it could be considered for publication ultimately. Should you have any questions, please contact us without hesitation. The point-by-point replies to all reviewers are as follows:

Replies to Reviewer 2:

Reviewer 2:

Comments to the Author

The manuscript presents an innovative approach to fault diagnosis in analog circuits using the ECA-MI-CNN model. The paper is well-structured and provides a comprehensive comparison of various models. However, there are several areas need to be improved.

Comment 1:

1.The introduction provides a good overview of the challenges in analog circuit fault diagnosis. Could you elaborate on why traditional fault dictionary models are still relevant or not in certain scenarios？

Response 1:

We greatly appreciate your suggestion to further clarify the applicability of the traditional fault dictionary model. In the revised manuscript, we have elaborated on the scenarios where the Fault Dictionary model remains useful, such as in small-scale circuits with limited fault patterns, where it enables quick fault matching with reduced computational costs. Additionally, we have discussed its limitations in more complex and dynamic circuit environments, where constructing a complete and high-quality fault dictionary becomes impractical, and external factors like noise may significantly reduce diagnostic accuracy. These revisions provide a clearer comparison between traditional and modern fault diagnosis methods. We sincerely thank you for your valuable comments, which have helped improve the clarity and completeness of the introduction.

(For specific revisions, please read the sections highlighted in red in the Abstract of the revised manuscript.)

Comment 2:

2.For the introduction, deep learning methods are well-justified, it remains unclear why existing approaches (e.g., DBN, GAN) fall short in the context of analog circuit diagnosis.

Response 2:

Thank you for your thoughtful comment. We appreciate your observation regarding the lack of clarity on why existing deep learning approaches, such as DBN and GAN, fall short in the context of analog circuit diagnosis. In the revised manuscript, we have provided further explanation of the specific challenges that DBN and GAN face in this domain.

For DBN, while it shows promise in extracting features from time-series signals, its limitations in handling complex fault patterns and optimal classification within the same category can affect diagnostic accuracy, especially in dynamic and diverse fault scenarios.

Similarly, although GANs are effective in reducing overfitting and capturing nonlinear relationships between fault sources and features, they still encounter issues related to stability and optimization in real-world environments, particularly when dealing with noisy or incomplete data. These challenges hinder their effectiveness in fully addressing the complex and precise diagnostic requirements of analog circuits.

We hope these additions clarify the reasons why DBN and GAN may not fully meet the needs of analog circuit fault diagnosis. Your feedback has been invaluable in improving the depth and clarity of the manuscript..

(For specific revisions, please read the sections highlighted in red in the Abstract of the revised manuscript.)

Comment 3:

3.For ECA module, could you provide more experimental evidence showing how the choice of k (interaction range) impacts model performance? For instance, does increasing k beyond a certain threshold lead to diminishing returns？

Response 3:

Thank you for your valuable comment regarding the impact of the interaction range parameter k on the performance of the ECA module. In response to your suggestion, we have conducted a series of experiments to further investigate how the choice of k affects model performance, as described in 4.1. ECA parameter selection experiment of the manuscript.

The experiments were designed to evaluate different values of k, including k=1, k=3, k=5, k=7, in order to analyze the impact of k on both accuracy and F1-Measure. The results, as shown in Table 5, reveal that increasing k from 1 to 3 leads to a significant improvement in model performance, with accuracy rising from 98.24% to 99.21% and the F1-Measure improving from 0.979 to 0.990. However, further increasing k to 5 and 7 results in diminishing returns, with performance slightly decreasing (accuracy of 99.02% for k=5 and 98.42% for k=7, F1-Measure of 0.987 and 0.982, respectively).

These findings indicate that while increasing k initially enhances model performance by capturing broader feature dependencies, beyond a certain threshold, the returns diminish and the model’s performance starts to plateau or even decline. Additionally, a larger k leads to increased computational complexity without substantial gains in performance. Therefore, we conclude that k=3 is the optimal value, as it achieves a good balance between capturing feature interactions and avoiding unnecessary complexity.

We hope that these additional experimental results provide the detailed analysis you requested and clarify the relationship between k and model performance. Thank you again for your constructive feedback, which has greatly helped improve the quality of our manuscript.

(For specific revisions, please read the sections highlighted in red in the 4.1. ECA parameter selection experiments of the revised manuscript.)

Comment 4:

4.The MI-CNN's ability to fuse time-domain and frequency-domain features is a major strength. Could you include visual examples or feature maps showing how these fused features differ qualitatively from single-domain features extracted by traditional CNNs.

Response 4:

Thank you for your valuable comment regarding the fusion of time-domain and frequency-domain features in the MI-CNN model. We understand the value of visualizing how these fused features compare qualitatively with the single-domain features extracted by traditional CNNs. However, due to the architecture of our MI-CNN model, which consists of three sub-networks, we face certain limitations in providing feature maps directly. Specifically, Networks 1 and 2 in the MI-CNN are responsible for extracting time-domain and frequency-domain features individually, and feature fusion is performed through a fully connected layer. This architecture does not directly generate intermediate feature maps that can be easily compared.

However, in Section 4.2 Feature distribution experiment, we compare the clustering performance of the traditional CNN and the MI-CNN. The experimental results indicate that the MI-CNN, which incorporates multi-input feature fusion, significantly outperforms the traditional CNN that uses only a single domain. These results highlight the effectiveness of our model in combining time-domain and frequency-domain features to enhance performance. Additionally, our team has published papers in Energy Conversion and Management titled "Fault diagnosis of photovoltaic array with multi-module fusion under hyperparameter optimization" and in the Journal of Membrane Science titled "Membrane fouling diagnosis of membrane components based on multi-feature information fusion" both of which demonstrate that multi-domain fusion outperforms single-domain approaches.

We hope that the experimental results in Section 4.2 Feature distribution experiment along with our published papers, can help illustrate the advantages of the MI-CNN in feature fusion. Thank you again for your valuable feedback.

Comment 5:

5.The MI-CNN's ability to fuse time-domain and frequency-domain features is a major strength. Could you include visual examples or feature maps showing how these fused features differ qualitatively from single-domain features extracted by traditional CNNs.

Response 5:

Thank you for your valuable question regarding the fully connected layer in the N3 network of the MI-CNN model. The fully connected layer in N3 consists of two parts: one layer with 2048 units and another with 50 units. While fully connected layers are known to have the potential to introduce bottlenecks or overfitting risks due to their high number of parameters, we have taken several measures to mitigate these risks.

Firstly, the 2048-unit layer enables the model to learn a comprehensive representation of the features from the previous layers, ensuring that important information from both time-domain and frequency-domain features is captured. The subsequent 50-unit layer is designed to reduce the dimensionality of the feature space, helping the model focus on the most relevant features for classification.

To mitigate the risk of overfitting, we primarily use BN layers, pooling layers, and regularization techniques during training. Additionally, the model employs an early stopping mechanism to monitor the validation loss during training and prevent overfitting.

We hope this explanation clarifies the functionality of the fully connected layers in N3 and how we address potential bottlenecks or overfitting issues. Thank you again for your valuable suggestions.

Comment 6:

6.For the butterworth low-pass filter, they may not fully showcase the robustness of the proposed model. Could you include additional experiments on circuits with varying levels of noise or environmental interference? Why an offset of 1064 was chosen?

Response 6:

Thank you for your valuable comments regarding the robustness of the proposed model and the use of the Butterworth low-pass filter. We greatly appreciate your suggestion to include additional experiments on circuits under different noise levels or environmental interferences.  

In response, we would like to clarify that we have already conducted experiments on the Butterworth low-pass filter in Section 4.4 Model Robustness Verification. Since the Butterworth low-pass filter was set to nine different modes, displaying all of them would result in excessively large figures. Due to space limitations, we have redrawn Figure 20, removing the original power spectrum and envelope spectrum, and instead, we have included more signal curves under different fault modes and noise levels. Additionally, Table 6 still presents the experimental results for all nine modes at 5dB, 15dB, and 20dB noise levels to evaluate the model's performance under different noise conditions. Moreover, considering real-world applications, we also built an experimental platform for the Butterworth low-pass filter and validated the model using data obtained from both short-term and long-term operations. The long-term experiments were specifically designed to study the accumulation of noise over extended periods of use.  

Furthermore, to expand the robustness evaluation and in line with your valuable suggestion, we also conducted noise experiments on a second-order low-pass filter with two stages and four operational amplifiers under the same signal-to-noise ratio conditions in Section 4.4 Model Robustness Verification. These additional experiments provide further insights into the model’s ability to handle noise in different circuit configurations.  

Regarding the offset value of 1064 in overlapping sampling, this choice was made to optimize the sampling process based on the specific characteristics of the circuit and data acquisition setup. The offset ensures that the sampling window aligns with the signal characteristics, thereby avoiding aliasing issues and ensuring accurate signal representation during processing. This decision was based on empirical analysis of the system’s performance under various sampling conditions.  

We hope these explanations clarify the robustness experiments we conducted and the rationale behind the offset selection. Once again, thank you for your constructive feedback, which has greatly helped improve the clarity and completeness of our manuscript.

(For specific revisions, please read the sections highlighted in red in the 4.4 Model robustness verification of the revised manuscript.)

Comment 7:

7.Could you provide more detailed insights into why CBAM performs better than SENet but still falls short compared to ECA?

Response 7:

Thank you for your valuable question regarding the performance comparison between CBAM, SENet, and ECA. The main reasons why CBAM performs worse than SENet but still outperforms ECA are as follows:

(1) Comparison between CBAM and SENet

CBAM (Convolutional Block Attention Module) applies attention mechanisms in both the spatial and channel dimensions. In some tasks, CBAM has advantages over SENet because it can more precisely focus on the importance of specific regions or features, optimizing information extraction during the process.

SENet (Squeeze-and-Excitation Networks) models channel attention through "squeeze" and "excitation" operations, but it mainly focuses on inter-channel dependencies, somewhat neglecting spatial information. For tasks that require the integration of spatial and channel information, CBAM may provide a more comprehensive feature representation.

(2) Comparison between CBAM and ECA

ECA model processes local cross-channel information, which results in higher efficiency and accuracy compared to CBAM. The ECA module reduces the computational cost of global information interactions, maintaining low computational complexity while effectively enhancing the model's expressive power, leading to better performance in certain tasks.

ECA excels at channel attention mechanisms without the need for complex convolution operations. Although CBAM has both spatial and channel attention mechanisms, it is less efficient and computationally complex than ECA.

(3) Explaining the performance differences

The lightweight structure and efficient feature extraction capability of the ECA model give it an advantage in handling complex datasets, whereas CBAM still delivers good performance in some cases. However, due to its relatively complex structure, CBAM carries a heavier computational burden, making it less efficient than ECA in certain tasks.

Additionally, we have already provided a detailed analysis in Section 5 Overview of the study of the revised manuscript. In that section, we specifically discussed the advantages of CBAM over SENet, mainly due to CBAM's joint attention mechanism that considers both spatial and channel features, enhancing the model's ability to extract relevant information. However, despite CBAM's superiority over SENet, it still falls short compared to ECA due to CBAM's relatively complex structure, which leads to higher computational costs and lower efficiency. In contrast, ECA optimizes the channel attention mechanism more efficiently, maintaining performance while improving computational efficiency, which explains its superior classification accuracy and faster computational speed.

(For specific revisions, please read the sections highlighted in red in the 5 Overview of the revised manuscript.)

Comment 8:

8.For the butterworth low-pass filter, they may not fully showcase the robustness of the proposed model. Could you include additional experiments on circuits with varying levels of noise or environmental interference? Why an offset of 1064 was chosen?

Response 8:

Thank you for raising the important question regarding the "black-box" nature of deep learning models. In our study, to address this issue and enhance the interpretability of the model, we provide more mechanistic insights by analyzing the diagnostic features of the Butterworth low-pass filter. Specifically, we used the Canny operator to process the frequency-domain feature maps, which helps highlight key anomalies in the signal. By plotting and analyzing the diagnostic feature maps for nine different fault modes, we observed significant differences among the faults, which correspond to different physical fault types in the circuit.

On this basis, we can see how the deep learning model extracts features from the data that are helpful for fault classification. Although the model still retains some "black-box" characteristics, this approach provides a clearer connection between specific anomalies and physical faults in the circuit. Therefore, while deep learning models may not be entirely transparent, by combining signal processing and feature extraction techniques, we have effectively improved the interpretability of the model for physical faults. As for research directly addressing interpretability in relation to physical mechanisms, we have not yet delved into this, but we highly appreciate your suggestion, and it will be a focus of our future research.

Comment 9:

9.Figure 19 appears to have a format issue, as part of the image is obscured.

Response 9:

Thank you for pointing out the issue with Figure 19. We apologize for the formatting problem that caused part of the image to be obscured. We have revised the figure to ensure that the entire content is visible and properly formatted. The updated version of Figure 19 has been included in the revised manuscript for your review. We appreciate your attention to detail, and we hope the updated figure resolves the issue.

(For specific revisions, please read the sections highlighted in red in the Figure 19 of the revised manuscript.)

 

Special thanks to you for your valuable comments.

 

Reviewer 3 Report

Comments and Suggestions for Authors

The paper discusses the application of multi-input constructive ANN model based on attention mechanism to the analog circuits fault detection. The fault detection problem is an important issue in a modern circuit design. The paper is well-written and generally makes sense despite the relatively limited novelty. However, I have several questions and suggestions to the Authors. Please, find them as follows.

1. I recommend rewriting the Abstract with clear problem statement and proposed solution descriptions. The current version is far from being perfect.
2. In the introduction section, some alternative fault detection methods like chaos-based methods and other ANN-based detectors, are to be mentioned. I recommend referencing to feature extraction by return map analysis with application to fault detection tasks, approximation of peak-to-peak amplitudes distribution, and papers comparing reservoir artificial and spiking neural networks in machine fault detection tasks.
3. The choice of analog filters as test circuits is to be explained further. I believe, really challenging fault detection cases appear in highly nonlinear circuits, e.g. circuits with vacuum tubes, memristors, Chua diodes etc. I recommend considering a bit more complex test circuit in the revised paper.
4. How many real analog circuits constructed for your study? Did you use any simulated faults to investigate the proposed approach in simulation? These issues should be clarified in the manuscript text.
5. "can combine time domain information and frequency domain information" - this feature can be very important for analyzing chaotic systems. I recommend considering such opportunity for your future studies.

Nevertheless, I highly appreciate a high quality and scientific rigor of the presented study, and believe it can be accepted for publication after minor revisions.

Author Response

Response to reviewer 3

Dear Editor and Reviewers,

We sincerely thank you for all positive and constructive comments of the manuscript entitled “Fault diagnosis in analog circuits using a multi-input convolutional neural network with feature attention” (the manuscript number is computation-3545756). Your suggestions and comments will be of great significance to our next research and writing. We have studied the comments carefully and made all necessary modifications to the original version. In the revised manuscript, we have used different colors to mark the main changes according to the opinions or suggestions of different reviewers, so as to facilitate the review. We have resubmitted the revised manuscript to the Submission system, and we hope it could be considered for publication ultimately. Should you have any questions, please contact us without hesitation. The point-by-point replies to all reviewers are as follows:

Replies to Reviewer 3:

Reviewer 3:

Comments to the Author

The paper discusses the application of multi-input constructive ANN model based on attention mechanism to the analog circuits fault detection. The fault detection problem is an important issue in a modern circuit design. The paper is well-written and generally makes sense despite the relatively limited novelty. However, I have several questions and suggestions to the Authors. Please, find them as follows.

Nevertheless, I highly appreciate a high quality and scientific rigor of the presented study, and believe it can be accepted for publication after minor revisions.

Comment 1:

1.I recommend rewriting the Abstract with clear problem statement and proposed solution descriptions. The current version is far from being perfect.

Response 1:

Thank you for your suggestion regarding the Abstract. We have carefully revised it to make the problem statement clearer and to provide a more precise description of the proposed solution. The updated Abstract explicitly highlights the importance of analog circuit fault detection and the challenges associated with this problem. Additionally, we have refined the wording to improve clarity and conciseness. We sincerely appreciate your valuable feedback and hope that the revised version better meets expectations.

(For specific revisions, please read the sections highlighted in red in the Abstract of the revised manuscript.)

Comment 2:

2.In the introduction section, some alternative fault detection methods like chaos-based methods and other ANN-based detectors, are to be mentioned. I recommend referencing to feature extraction by return map analysis with application to fault detection tasks, approximation of peak-to-peak amplitudes distribution, and papers comparing reservoir artificial and spiking neural networks in machine fault detection tasks.

Response 2:

Thank you for your valuable feedback on our paper. Regarding your suggestion, we have added some alternative fault detection methods in the revised manuscript, including feature extraction methods using return map analysis for fault detection tasks, approximation of peak-to-peak amplitude distribution, and papers comparing reservoir artificial neural networks and spiking neural networks in mechanical fault detection tasks (see References [8-11]). These additions will help enhance the comprehensiveness of the paper and provide broader contextual information. Once again, we greatly appreciate your valuable suggestions.

(For specific revisions, please read the sections highlighted in red in the Abstract and Reference of the revised manuscript.)

Comment 3:

3.The choice of analog filters as test circuits is to be explained further. I believe, really challenging fault detection cases appear in highly nonlinear circuits, e.g. circuits with vacuum tubes, memristors, Chua diodes etc. I recommend considering a bit more complex test circuit in the revised paper.

Response 3:

Thank you for your valuable feedback on our manuscript. Regarding the choice of analog filters as test circuits, we have provided a detailed explanation in the revised manuscript (3.1. Experimental subjects). We selected the Butterworth low-pass filter and two-stage quad op-amp dual second-order low-pass filter as test circuits because these filters are widely used in practical engineering and have a clear theoretical foundation and engineering application value. The Butterworth low-pass filter has a flat passband response, making it suitable for scenarios requiring precise filtering characteristics, while the second-order four-op-amp dual second-order low-pass filter can simulate more complex filtering behaviors and offers higher adjustability. These filters effectively reflect common fault patterns in circuits during fault diagnosis, and their design and analysis methods have important demonstration significance in fault detection research.

However, we also acknowledge that more complex circuits (such as vacuum tubes, memristors, Chua diodes, etc.) may present greater challenges. Therefore, we will consider discussing and further exploring such highly nonlinear circuits in the revised manuscript. Once again, thank you for your suggestions.

(For specific revisions, please read the sections highlighted in red in the 3.1. Experimental subjects of the revised manuscript.)

Comment 4:

4.How many real analog circuits constructed for your study? Did you use any simulated faults to investigate the proposed approach in simulation? These issues should be clarified in the manuscript text.

Response 4:

Thank you for your valuable comments. In response to your query, this study conducted experimental verification on a real analog circuit, specifically the Butterworth low-pass filter. We used the fault modes listed in Table 1 to test the performance of the ECA-MI-CNN diagnostic model. Regarding the selection of fault modes, based on preliminary sensitivity analysis, we found that the component faults listed in Table 1 (such as deviations in resistors and capacitors) have a significant impact on circuit performance. Therefore, we focused on simulating faults in these components and did not consider other components.

As for why we chose the Butterworth low-pass filter for the experiments, the main reason is its relative simplicity and widespread use in practical engineering, making it suitable for verifying the effectiveness of the proposed method. Although the two-stage four-op-amp dual second-order low-pass filter is also a highly adjustable filter, its circuit structure is more complex, involving more components and fault modes. Therefore, in the initial phase of this study, we opted for the simplified Butterworth low-pass filter to ensure the feasibility and robustness of the diagnostic model. Future research will include experimental verification on more complex circuits, including the two-stage four-op-amp dual second-order low-pass filter..

(For specific revisions, please read the sections highlighted in red in the 5. Overview of the study of the revised manuscript.)

Comment 5:

5."can combine time domain information and frequency domain information" - this feature can be very important for analyzing chaotic systems. I recommend considering such opportunity for your future studies.

Response 5:

Thank you very much for your valuable suggestion. We fully recognize the importance of combining time-domain and frequency-domain information in the analysis of chaotic systems, especially in complex systems, where this integrated approach can provide new insights into the dynamic behavior and nonlinear characteristics of the system. Although the focus of this study is primarily on fault diagnosis in linear and nonlinear circuits and does not specifically address chaotic systems, we acknowledge that your suggestion presents a potential research direction. In the revised manuscript, we have included an analysis of chaotic systems in the outlook section, exploring the potential of combining time-domain and frequency-domain information in chaotic system studies. We plan to further investigate the application prospects of this approach in future research. Once again, thank you for your profound insights into our study.

(For specific revisions, please read the sections highlighted in red in the 6. Conclusion of the revised manuscript.)

 

Special thanks to you for your valuable comments.