UAV Anomaly Detection Method Based on Convolutional Autoencoder and Support Vector Data Description with 0/1 Soft-Margin Loss

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper, the authors address the increasing use of UAVs across various industries and the growing concerns about their safety and dependability. It emphasizes the difficulties in constructing accurate anomaly detectors due to the rarity of abnormal flight data and the spatiotemporal characteristics of the data.  They propose an anomaly detection framework that combines a CAE model for feature extraction with an SVDD model. The CAE model is a neural network that uses one-dimensional convolutional neural network 1DCNNs as the encoder–decoder layers, combined with linear layers. The accuracy and robustness of anomaly detection are enhanced by the introduction of a nonlinear model of support vector data description (SVDD) with a 0/1 soft-margin loss, called L0/1-SVDD. The framework involves data preprocessing, training a CAE model to extract spatiotemporal features, and using the residuals to train the L0/1-SVDD model for anomaly detection. The Bregman ADMM algorithm is proposed to solve the L0/1-SVDD model, enhancing the adaptability of anomaly detection thresholds. The proposed method is validated using real flight data such as the Air Laboratory Failure and Anomaly (ALFA) dataset and shows superior performance compared to existing methods like AE, LSTM, and LSTM-AE across five evaluation metrics namely accuracy, precision, recall, F1-score, and G-mean.

The topic is timely and compelling. The paper exhibits good writing quality and demonstrates a meticulous organizational structure. Figures and tables are well designed. The reference list is appropriate. The cited references are mostly recent.  The proposed methodology appears to be sound. The experimental results are convincing. However, as stated by the authors, the introduction of the 0/1 loss function in L0/1-SVDD leads to a non-convex optimization problem, increasing computational complexity. In addition, the training time may significantly increase when using large datasets, necessitating the use of a distributed computing framework. Moreover, the model’s performance is sensitive to the selection of parameters, which may require intelligent optimization algorithms for better results.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes an anomaly detection framework that fully considers the temporal correlation and distribution characteristics of flight data. This study proposed a support vector data description (SVDD) method based on a 0/1 loss function and the L0/1-SVDD threshold de-termination module was trained using residuals, allowing for a more accurate threshold determination and further improving the anomaly detection performance. Experimental results show that, compared with a series of methods, the proposed method exhibited superior performance across five evaluation metrics. In general, although this investigation is of interest, the following questions and comments should be addressed before recommending for publication.

 

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Please check the detailed information in the attached reviews

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

This study proposes an anomaly detection framework that combines a 1DCNN and an autoencoder to extract spatiotemporal features from flight data. It also introduces an SVDD model with 0/1 soft margin loss, solved by the Bregman ADMM algorithm, to improve detection accuracy. The results show superior performance compared to other methods. This study is of great interest, but its current writing quality needs to be improved.

1- Authors must separate the introduction and the related work. A separate introduction helps to better contextualize the research before diving into the details of related work. This allows readers to understand the study's motivation and overall framework without being overwhelmed by references. When combining the introduction and the literature review, there is a risk of overloading the first part of the manuscript with too much information. Separation allows the information to be structured in a more digestible way for the reader.

2- The authors should also highlight the following sections: Materials and methods, Experimentation, Discussion, and Conclusion.

3- To improve the understanding of data preprocessing, providing a concrete example illustrating the reconstruction of flight data would be helpful. Such an example will help clarify how the sliding window reconstruction method is applied and how it contributes to temporal data analysis. This will help visualize the process, understand parameter choices such as window length and step, and appreciate the impact of this technique on the quality and accuracy of subsequent analyses. A real-world example will highlight the benefits of this method by showing how it can optimize temporal feature extraction and improve anomaly detection.

4- A presentation of the data sample is necessary to follow the study better. To monitor the study more effectively, it is essential to present the sample data used. This will make it possible to contextualize the results, understand the characteristics of the data, and verify the relevance and representativeness of the sample to the study's objectives.

 

5- I recommend using a table to describe the parameters of the CAE model, which will improve clarity and presentation.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

Authors should add contributions and the manuscript's organization at the end of the introduction rather than in the related work section. This would help to better structure the document, clarifying from the outset the objectives of the study and its organization.

The authors satisfactorily addressed my concerns. I recommend publishing this article after the requested minor corrections have been implemented.

Author Response

Please see the attachment.

Author Response File: Author Response.docx