Advancing Fault Detection in HVAC Systems: Unifying Gramian Angular Field and 2D Deep Convolutional Neural Networks for Enhanced Performance

Round 1

Reviewer 1 Report

This paper utilizes a GAF-CNN method to diagnose the faults of HVAC systems. However, this method is not novel, since the method has been published in the journal Applied Energy [1]. Moreover, the authors claim that there are five contributions in this study. In my opinion, these statements are not accurate. More details can be found as follows:

A.     An existing study has reported the GAF-CNN method for HVAC fault diagnosis [1]. This study utilizes the same method without any improvement. Hence, I think this study is not novel.

B.     The authors state five contributions in the introduction. However, I disagree with these statements:

l  The authors state that the first contribution is that they use the simulation data to implement the data-driven FDD methods. However, it is very common to utilize simulation data to train and verify data-driven HVAC FDD methods. Beside, if a data-driven FDD model is trained using simulation data, it is quite hard to apply this model in practical HVAC systems.

l  As the authors state, the second contribution of this study is that their method can achieve automated feature extraction. However, some feature extraction methods have been applied in previous studies, such as PCA and autoencoder. The authors don't mention these methods in this manuscript, which is cheating.

l  The authors claim that the fourth contribution of their study is the utilization of GAF. However, a similar method has already been proposed by Gao et al. [1], suggesting that this contribution is not entirely accurate.

l  For the fifth contribution, the authors state "Existing FDD systems fail to detect various HVAC faults, but the proposed method considers ten significant HVAC faults…" As far as I know, existing data-driven FDD method can diagnose various HVAC faults well. The nature of the GAF-CNN is a multi-classification model. Lots of multi-classification FDD models has been studied in existing studies.

l  The authors tell the readers that the last contribution of this study is that the proposed method can reduce the number of required sensors. However, I cannot find enough evidence for this statement in the section of "Result and Discussion".

[1] Gao Y, Miyata S, Akashi Y. How to improve the application potential of deep learning model in HVAC fault diagnosis: Based on pruning and interpretable deep learning method. Applied Energy 2023;348:121591. https://doi.org/10.1016/j.apenergy.2023.121591.

The written English needs to be improved.

Author Response

Thanks for the comments. Please see the attached response letter.

Author Response File: Author Response.pdf

Reviewer 2 Report

Summary:

This study addresses the need for efficient fault detection and diagnosis in HVAC systems to enhance reliability and reduce maintenance costs. The proposed approach combines convolutional neural networks (CNNs) and gramian angular fields (GAF) to transform sensor signals into 2D images, enabling automatic feature extraction and deep fault classification. More specifically, the study highlights GAF-CNN's superiority over other machine learning models. The model achieves 97% accuracy in classifying HVAC faults without requiring extensive HVAC expertise, offering potential for operational stability and cost reduction.

Comments:

1. The reviewer raises a question about the potential impact of varying sample amounts (simulated data) on result accuracy. The study utilizes training and testing data from a 24-hour simulation period. An interesting aspect to consider is how altering the simulation duration, such as reducing it to 12 hours or extending it to 48 hours, will it influence the outcomes. Similarly, what is the fundamental reasoning behind the simulation configuration in HVACSIM+ as employed in this study?  

2. It is not clear that how to generate the fault types in the proposed fault-modeling approach in HVACSIM+.

3. What is the intended definition of "real-time" in this context? Are there any specific time constraints imposed on the Fault Detection and Diagnosis (FDD) system? Furthermore, how can the ability of the proposed method to achieve "real-time diagnosis" be substantiated?

4. What is the additional computational burden associated with converting HVAC sensors' time series signals into images using the GAF approach? The accuracy comparison in Table 6 reveals that the proposed method outperforms the 1D-CNN approach by an average of 3%. While the 1D-CNN primarily employs raw data for training and inference, it is important for the paper to analyze the computational overhead introduced by GAF encoding to better illustrate the benefits of the GAF-based method.

5. The novelty of this work is not clear. As outlined in Section 1.1, there exist prior works that have also introduced machine learning-based approaches for fault detection in HVAC systems. It is not clear that what is the main difference between these works and the proposed method.

6. The reviewer found that there is a similar work published in this year as follows. If it is necessary, please discuss the difference between your work and this work.

Yuan Gao, Shohei Miyata, Yasunori Akashi, “how to improve the application potential of deep learning model in HVAC fault diagnosis: Based on pruning and interpretable deep learning method” Applied Energy, Volume 348, 2023, 121591, ISSN 0306-2619,

https://doi.org/10.1016/j.apenergy.2023.121591.

7. The evaluation solely focuses on a single existing work. Is it possible that the related works referenced in Section 1.1, particularly [19] and [20], could be utilized for comparison? Please provide reasons for the decision not to include a comparison with these mentioned references.

1. The conclusion is too lengthy. It is better to shorten the paragraph and try to only emphasize the findings and results of this work.

Author Response

Thanks for the comments. Please see the attached response letter.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Thanks for your responses to my comments. The authors have clarified the contributions of this study in the revised manuscript. All my comments have been addressed well.

Reviewer 2 Report

I appreciate the authors' efforts. All my concerns have been addressed. Thank you very much.