Ship Abnormal Behavior Detection Method Based on Optimized GRU Network
Round 1
Reviewer 1 Report
The present work consists on the use of the neural network to detect the abnormal behavior of ships. The aim of this work is the use of the GRU network model with attention mechanism. The proposed subject is very interesting,
Authors, should present well the originality of their work. In fact, the ship anomalies detection using the GRU network model already exist in the literature. Authors, should explain also what is the scientific contribution of the present work and how it can be used by operators?
I have some questions for authors :
- Why do we use the neural networks to identify the abnormal ship behavior, while it is possible and easy to do that by a simple Matlab program using the AIS data and anomalies conditions (ship speed, course, trajectory…)?
- For accuracy identification it is indispensable to add the environmental operating conditions (current, and wind). Because, these parameters have an significant impact on the ship drift. Is it possible to add these two parameters in neural network s and how can be managed?
- The paper structure and manuscript quality can be considerably improved for an easy and fluid reading. I invite authors to reorganize different sections.
- Lot of functions are not defined and explained (SIGMOID, Optimize ADAM, Softmax…).
- The article deals with an applied subject which is the abnormal behavior of ships which concerns directly the coast guards, however, the manuscript remains very theoretical (mathematical), it is really a pity. I would have liked that the authors make more effort to write a manuscript that can be read ed by engineers and not necessarily by specialized researchers!
- Mathematical notation should be written in equation format.
- Figure 2 not very clear. From my point of view authors can improve the section 2.
- Figure 3, Authors should define the probability parameter « Pi »
- Section 2.2 : It is indispensable to make the link between the GA algorithm use and the application which is the detection of anomalies
- Section 3 : How the initial ship anomalies are obtained ? I thought that the anomalies are detected from the AIS data using different algorithms?
From my point of view the article has potential, however, I invite authors to review the structure of their manuscript well and also by answering the questions given above.
Author Response
Please see the attachment
Author Response File: 
Author Response.docx
Reviewer 2 Report
- In introduction the claim that VTS are only in coastal waters is not entirely accurate
- The data accuracy of the AIS system used is not sufficient for navigation in restricted port accesses and ports, therefore the version of the use of primary data accepted by the authors is not sufficiently accurate
- The system does not assess the maneuverability of the ship itself, but differs significantly for different types of ships, so it is recommended that the proposed solution assess the potential impact of the ship's maneuverability on the final result (based on probabilities).
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 3 Report
The article addresses the challenge of ship anomaly detection, which is a highly relevant topic in maritime situation awareness. The authors present a GRU network with attention to identify anomalous ship behavior, based on historical ship traffic. Furthermore, the network is optimized using a genetic algorithm to select the optimal hyper-parameters, aiding the predictive performance. The outlined approach achieved high accuracy in detecting anomalous behavior and is further improved via hyper-parameter optimization.
However, I find the article to be lacking in various areas. Firstly, the article is littered with grammatical errors. In general, the English language needs to be improved to increase the readability of the article. Furthermore, the formatting is in some cases not consistent e.g. starting with "0." for the introduction. In general, the readability of the article needs to be improved.
The size of the data set used to train the network is also of concern. It is stated that one month of historical trajectories is utilized. The authors should state how many anomalies, and normal trajectories are in the training and test sets. The data should also be split into training, validation and test sets, where the validation set ensures that overfitting does not occur (as it appears the authors have currently done with their test set), and the test set is not evaluated until the model is trained. I am slightly concerned that the test set in the article has very few "anomalies" present, and that the high accuracy is due to the overwhelming presence of "normal" behavior. The authors should present the data set clearly, where the number of "normal" and "anomalous" trajectories in each data set (training, validation, test) are clearly presented. Furthermore, in Section 4.4 it is stated 270 anomalies are selected from the data set. Are these anomalies the test set, or do they include trajectories that were part of the training set as well? If the latter is the case, the results are invalid. The authors should clarify this and present what is being tested more clearly in the article.
In general, the readability of the article should be improved where the methodology and results are more clearly presented to the reader.
Below are specific comments for the text:
In the text, anomaly detection is referred to in various ways e.g. ship abnormal behavior detection, anomaly detection, ship track anomaly detection. I would recommend choosing one term, and providing consistency throughout the article to avoid confusion.
In line 32, I believe the authors have misunderstood the meaning of meticulous, and may want to review this.
I was unable to find many of the references e.g. [12], [18], [27]. Please improve all references to include doi/links where relevant as well as more detailed information within each reference, as many were found lacking.
In line 50, is "density spaced" supposed to be DBSCAN?
In line 96, the authors state Vessel Traffic System (VTS). In the maritime domain, VTS is often in reference to Vessel Traffic Services (VTS) and should not be confused. See https://www.imo.org/en/OurWork/Safety/Pages/VesselTrafficServices.aspx.
Figure 2 is also not consistent with the equations in (1), and should be checked.
In line 153, the authors state that the number of layers of each neural network should not exceed 4. Why is this? Please provide a relevant explanation or reference.
It may be useful to present how the anomaly labels have been determined. The only information that could be found was in Figure 6 where it states human tags. A section should be included where this is addressed.
In line 226, the features are stated to include both course and velocity. Velocity is a vector inherently including the course. This should be changed to speed (I assume this is speed over ground from AIS data).
In lines 230-232, it is mentioned that 20 "moments" of trajectory features are used. What does this correspond to in terms of time (how long are the trajectories)? Is this constant, i.e. are the trajectories pre-processed to have equal spacing between time steps, or does this vary? It may also be better to refer "moments" as states or time steps.
In line 232, it x_20 should be fixed, and in 233 instead of "pieces of trajectory data" it is sufficient to say that the batch size is 40.
In line 246, "one-dimensional vector" is stated. A one-dimensional vector will only contain one element. This may be altered to the general term vector, alternatively to a N-dimensional vector where N is the number of classes in the output.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 4 Report
Insufficient explanation of the (2.2. GA Optimizes GRU Network)
Please explain in detail the Selection method(Proportional, Ranking, etc.) and Cross-over methods (single point, multi-point, etc.) are applied and parameters (population size, crossover probability, mutation probability, etc.)
Present the fitness function formula of the GA.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Thanks to the authors for the improvements made. Some answers are very clear. However, the article remains difficult to understand, I invite the authors for the second time to take their time to bring more clarification and improvement to the current manuscript and not for future works.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
I attached the review comments.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 3
Reviewer 4 Report
1. Vector notation should be represented in bold font.
So, please represent the bias vector in bold font.
The weight matrix just represented with upper character, not in bold font.
2. in line 294, 'International Maritime Collision Avoidance Rules'
means 'COLREGS - International Regulations for Preventing Collisions at Sea'?
If correct, it should be revised with 'COLREGS - International Regulations for Preventing Collisions at Sea.'
3. Write down the threshold value for the ship speed and course.
4. in Figure 8, draw the boundary channel (= ship fairway or route) of Figure 1.
Author Response
Please see the attachment
Author Response File: Author Response.docx
