1. Introduction
Maritime transportation is favored and being rapidly developed due to its advantages of large freight volumes, low cost, low energy consumption, and low required investment. With the continuous innovation of shipbuilding technology, there are more than 4 million ships sailing within the coastal waters of China every year [
1]. Maritime transport undertakes 90% of the work of world trade and transportation [
2]. Therefore, it can be said that the development of seaborne trade is a barometer of the development trend of the global economy [
3]. At the same time, China’s national policy vigorously supports the development of the maritime transport industry. In 2013, President Xi Jinping proposed the cooperation initiatives to build the “New Silk Road Economic Belt” and the “21st Century Maritime Silk Road” [
4]. In 2016, China also issued the “Outline of the Development Plan for the Yangtze River Economic Belt” to promote the orderly and free flow of economic factors and build a new bidirectional corridor based on land and sea with the outside world. Therefore, the supervision of maritime traffic is crucial.
However, with the continuous development of the maritime shipping industry and continuous application of the integrated system of various marine services, the number of marine traffic accidents is also increasing. According to the “Transportation Safety Production Accident Report (2019)”, a total of 197 marine traffic accidents occurred during this period, resulting in the deaths of 115 people and the loss of up to 12 Gt of goods. Marine traffic accidents usually result from abnormal ship-related behaviors. In order to better supervise maritime safety, one needs to detect abnormal ship behaviors.
In order to ensure safety in navigation, the AIS base station came into being. International Convention for Safety of Life at Sea (SOLAS) regulation V/19 stipulates that all ships of 300 gross tonnage and above engaged in international voyages, cargo ships of 500 gross tonnage and above not engaged in international voyages, and all passenger ships (regardless of size) should be fitted with automatic identification systems (AISs). The requirement came into effect for all ships on 31 December 2004 [
5]. An AIS is a self-reporting information system mainly designed for collision avoidance that uses VHF radio information to track and monitor ship movements by exchanging position, heading, and speed information with other AIS-equipped ships and AIS base stations in the vicinity [
6]. There are 27 categories of AIS message types, which can be divided into three types: static information, dynamic information, and voyage-related information [
7]. However, this study focuses on the use of the dynamic information and static information of ships to further explore the law of ship behavior. The static attributes include the vessel’s IMO number, vessel name, Maritime Mobile Service Identity (MMSI), callsign number, and size and type of the vessel. The dynamic attributes include the ship position, time stamp, speed over ground (SOG), course over ground (COG), and other information [
8]. Meanwhile, vessels have different punctuation intervals due to differences in static information, dynamic information, and voyage-related information during the course of their journeys.
The AIS system provides data support for the analysis and modeling of maritime traffic and enhances the feasibility of research into abnormal behavior detection for ships. Ship behavior is not the specific behavior of a ship but refers to the mode and law of similar actions of ship groups, which not only depend on human consciousness, thinking, decision making, and manipulation but are also affected by the characteristics of the ship and the surrounding environment [
9]. Existing research has not effectively combined the characteristics of ship behavior to discriminate abnormal patterns. Since the occurrence of marine accidents can usually be attributed to the abnormal behavior of ships, the characteristics of marine accidents in this study were abstracted from three dimensions—spatial attributes, temporal attributes, and thematic attributes—to obtain the data expression of the abnormal behavior of ships. For example, on 14 May 2021, the accident investigation report of the Maritime Safety Administration of the People’s Republic of China showed that a ship drifted 2 km away from its original route due to a wind disaster and then finally capsized and sank. Based on the AIS data, it can be observed that the spatial position of the ship deviated from the trajectory point group after the wind accident. The schematic diagram is shown in
Figure 1. Therefore, it is feasible and necessary to detect abnormal behavior of ships based on AIS data. On the one hand, it can reduce the workload of maritime surveillance operators and improve the efficiency of maritime surveillance [
10]. On the other hand, it has guiding significance for the relevant departments to formulate management measures such as the planning of navigation routes and maritime traffic control. Additionally, it has great socioeconomic and environmental benefits.
The purpose of this study was to model and mine the typical motion patterns by analyzing the behavior of ships at sea and to detect and evaluate the abnormal behavior of ships based on the AIS data in the framework of the discovery patterns from the perspectives of spatial attributes and thematic attributes to comprehensively express the maritime traffic situation. First, we extracted the information of the maritime routes based on the AIS data extension main graph structure learning strategy to generate a data-driven representation of the real navigation routes. Then, we detected and evaluated abnormal ship behavior based on Rayda’s criterion and the isolated forest algorithm for the spatial attributes and thematic attributes, respectively.
The remainder of this paper is organized as follows.
Section 2 is a review of the relevant research.
Section 3 introduces the study area and the data and describes the detection method for abnormal behavior based on AIS data.
Section 4 provides the experimental results and analysis.
Section 5 summarizes the content of this article and puts forward future outlooks.
2. Related Work
Most of the existing studies have reviewed the abnormal behavior of ships based on research methods from their respective fields. Classification according to the research methods can help to elucidate the relevance and continuity of the method in the field, but because the research methods are prone to repeated use, and the existing methods tend to be complex and comprehensive, it is impossible to review the progress of the research with a single method. Therefore, we chose a classification method based on the results of the study to classify the abnormal behavior of ships. According to the literature, the research on abnormal behavior detection for ships can be divided into the following four categories according to the results: only spatial position abnormalities, spatial position and thematic attribute abnormalities, specific situation abnormalities, and no specific experimental results, with only a description of the detection framework instead.
Soleimani et al. [
11] compared the input trajectory with the approximate optimal path generated by the A* algorithm to obtain the ship’s abnormal score for the spatial position. This method belongs to unsupervised learning, which can effectively deal with large datasets. Rong et al. [
12], based on the Douglas–Peucker algorithm and density-based clustering algorithm, abstracted the ship route into leg sections and turning sections and then probabilistically characterized marine traffic based on the lateral distance of the trajectory points and other attributes. Pallotta et al. [
2] used unsupervised incremental learning to cluster and visualize routes and then used the knowledge of the routes to trajectory points for route classification and anomaly detection. Xiong et al. [
13] used the multi-kernel non-parametric estimation method to estimate the probability density of the motion patterns of ferry vessels and set the criterion of abnormality to detect a ferry with abnormal behavior. Laxhammar [
14] trained the Gaussian mixture model by taking the ship point space information and the attribute information as eigenvalues and marked the outliers by calculating the probability of the new trajectory points generated by the trained model. Ristic et al. [
8] used the kernel density estimation method to extract the ship motion pattern from the real AIS data and identify the abnormal points. Bomberger et al. [
15] used the incremental learning method based on a fuzzy neural network to forecast and evaluate the position of a ship. Smith et al. [
16] combined the Gaussian method and extreme value theory to identify abnormal behavior in the trajectory flow, such as deviation, mooring, and wandering. Venskus et al. [
17] performed unsupervised learning on the trajectory flow data obtained by the sensor based on the SOM network to detect abnormal ship movements in maritime traffic.
Lei [
6] proposed a comprehensive anomaly detection framework, MT-MAD, to mine the spatial characteristics, sequence characteristics, and behavior characteristics of the ship’s trajectory. Finally, the three features were combined with the cooperative value to determine an abnormal situation for the ship. Shahir et al. [
18] proposed an interactive detection system based on the DBSCAN algorithm and Markov algorithm for the multi-ship encounter problem. Based on the two-dimensional Gaussian distribution, Holst et al. [
10] carried out probability modeling for the normal ship-running trajectory and detected the ship characteristics by establishing a model. Pallotta et al. [
19] extracted abnormal ship traffic by correlating the ship trajectory with the existing route and then scored the degree of abnormality of the point with respect to the course, bow direction, and lateral distance of the trajectory point. Osekowska et al. [
20] represented the ship’s behavior using the charge value and built a model to detect abnormal behavior of a ship by defining the relationship between the ship’s movement mode and the change, accumulation, and distribution of the charge value. The method proposed by Venskus et al. [
21] is similar to the charge theory proposed by Osekowska et al. [
20], which uses the principle of a bionic animal pheromone to detect the standard path of the ship and then uses it to detect the nonstandard movement of marine traffic.
Since it is difficult to consider all the aspects in anomaly detection, it is necessary to consider an anomaly in the context of the situation after it is detected [
22]. For example, the deviation of the ship may be a simple response to a particular scenario such as an iceberg or hurricane [
22]. Additionally, in some cases, seemingly normal and benign behavior, after combination with the information of the scene, may turn out to be deceptive or concealed abnormal behavior. Kraiman et al. [
23] built a model based on a Gaussian mixture model and self-organizing neural network and used the ship attribute information and related environmental information as input data for abnormal behavior detection. Riveiro et al. [
24] used an interactive data mining module including SOM, GMM, and Bayesian theory to build the model and judge the abnormal behavior of a ship. This method can obtain situational knowledge and artificial feedback online as well as use visual means to continuously improve the model update. Mascaro et al. [
25] learned the data by generating static and dynamic Bayesian networks and added meteorological factors and ship interaction information to enhance the learning ability of the model. Finally, the average scores of 12 abnormalities and the difference between the score and normal behavior were obtained. Radon et al. [
26] combined “context information”, including marine environmental factors such as the wind direction and speed, to discriminate abnormal results based on the density clustering algorithm and finally interpreted an abnormal situation through the prior knowledge of experts. The study of Jakob et al. [
27] focused on the detection of pirate threat behavior, combining environmental factors and ship attribute characteristics to create a ship movement model based on game theory, the Nash equilibrium principle, and reinforcement learning for the ship route and attack simulation evaluation.
Riveiro et al. [
22] expounded on and discussed methods of marine anomaly detection based on four aspects—marine traffic data, methods, systems, and users—and put forward strong insights for future development. Laxhammar et al. [
28] evaluated two commonly used methods of maritime anomaly detection: the Gaussian mixture model and adaptive kernel density estimator. They stated that both methods have certain limitations. Therefore, a clustering method combining scenario information and ship motion attribute features was proposed at the end of the article, but this was not verified by experiments. Arguedas et al. [
29] created a multi-layer interactive platform called Blue Hub and proposed several spatiotemporal mining techniques to enhance situational awareness. The authors opined that there is still much work to be conducted in the consolidation of past studies. Jasinevicius et al. [
30] used a rule-based fuzzy expert knowledge graph to monitor the port safety system. The system combined the characteristics of ship attributes, personnel, and cargo risk factors to detect and identify the behavior of a ship, but this was not verified by experiments.
From the review of the research on abnormal behavior detection for ships, we can observe the following problems: (1) most of the existing studies have defined the “abnormal behavior of ships” in their respective fields without elaborating on the reasons for these definitions; (2) while previous studies have usually been based on using historical data to extract relevant knowledge about the behavior of ships, existing studies have focused more narrowly on the behavior of ships on specific routes; and (3) the existing research on detection is not comprehensive, with the research methods only considering the single attribute characteristics of the ship and not a fusion of characteristics. Some studies still only identify normal or abnormal conditions in a binary fashion, not considering the degree of abnormality for the impacting factors.
In this study, the existing research was combined with the actual geographical environmental factors for a single ship’s motion behavior to create a relatively comprehensive framework for detecting abnormal behavior. Based on the graph structure learning method and the tree linking strategy, we detected the structure of maritime routes to represent the ship’s travel route. The extracted network structure characteristics maintained the real geographic mapping between nodes and could reduce the storage requirements for the maritime routes. Then, using each segment as a unit, the probability characteristics of the trajectory points were detected from the perspectives of spatial attributes and thematic attributes, and the score value for the degree of abnormality of the trajectory points was obtained to evaluate the ship’s motion mode.
5. Conclusions
The marine environment is complex and variable, and the unconstrained motion of a ship’s trajectory increases the uncertainty of the data. Based on the characteristics of the target, this study explored a ship’s motion behavior and further detected its abnormal behavior based on the two aspects of spatial attributes and thematic attributes.
The factors affecting a ship during the course of travel, including the range price, weather, traffic control, and geographical environment, usually follow the following factually standardized maritime lanes, so we extended the inversed graph-embedding technology based on historical AIS data to detect and characterize the structure of the maritime routes. Then, based on the extracted structure of the maritime routes, we detected the departure of the ship from the route from the spatial attributes. Next, we refined the data, grouped the ship trajectory groups according to the network structure based on the maritime routes, and detected abnormal ship thematic attributes and scored their degrees based on the isolation forest algorithm. Based on the AIS data for ship navigation between the East China Sea and the Bohai Sea, we detected and scored the abnormal ship motion with spatial attributes and thematic attributes. After detection, we were able to identify a number of abnormal points with high scores. We combined these with the actual physical environment of the trajectory points to find and explain them and then identified more meaningful points and phenomena.
AIS data are an important tool for global sea state monitoring, but their regulation and privacy remain to be discussed. Relevant maritime authorities believe that the privacy and sovereignty of maritime vessels require the encryption of AIS data. At the same time, there is a lack of research on the feature extraction and correlation of high-dimensional datasets, such as those for the weather and the marine environment. The non-spatiotemporal information of ships and the situational information in driving usually contain other behavior patterns of ships. Therefore, considering the static information and environmental information of ships could provide new ideas for the detection of abnormal ship behaviors.