Regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters

Abstract

The present invention discloses a regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters, including: performing probabilistic conflict detection based on ship AIS trajectory data to obtain the severity of conflicts between ships; constructing a composite distance measurement model based on the obtained conflict severity, and clustering using a constrained SSN clustering algorithm based on the measurement results; and obtaining an optimal clustering solution based on a hierarchical dual-objective optimization algorithm. The present invention handles the dynamics and disturbances of traffic movement through probabilistic conflict detection; ensures conflict connectivity and spatial compactness by designing an extended clustering algorithm so as to divide regional ship traffic into multiple clusters; and generates a robust clustering scheme that can adapt to dynamic changes in ship traffic by formulating a hierarchical dual-objective optimization algorithm, so as to be able to find traffic clusters with strong conflict connectivity and high density in high-density waters, have adaptive robustness to various traffic scenarios, and have good performance under the uncertainty of ship movement.

Description

Technical Field

[0001] The present invention relates to the field of maritime transportation safety technology, and in particular to a regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters. Background Art

[0002] Maritime safety management has always been regarded as one of the most important concerns. Because once a maritime traffic accident occurs, it will have intolerable consequences such as loss of life, economic losses and / or environmental pollution. Economic globalization associated with the rapid growth of transportation demand has made maritime traffic busier and riskier, especially in restricted waterways and ports with heavy traffic. This often brings major potential challenges to the traffic safety management of maritime regulators and increases the possibility of accidents to a certain extent. In order to meet these challenges, some advanced equipment and systems, such as vessel traffic service (VTS) systems, maritime traffic service networks and land radars are designed to assist regulators in maritime traffic monitoring. Although these facilities provide a variety of intelligent functions in monitoring maritime traffic behavior, they still have some defects in automatically and adaptively identifying real-time local conflict hotspots. Therefore, maritime traffic regulators must identify high-collision risk ship traffic clusters based on their experience and intuition, which greatly increases their workload and hinders the timely implementation of risk control strategies. Therefore, in the context of the development of intelligent traffic support systems, it is necessary to develop new advanced technologies and tools to further enhance the maritime traffic situational awareness capabilities.

[0003] Various quantitative risk analysis methods have been proposed to improve the operational safety of ship traffic in busy waters. These methods are prerequisites for maritime traffic safety monitoring and management, enabling regulators to issue collision warnings and provide important support for decision-making. In particular, the development of the Automatic Identification System (AIS) has helped to obtain a large amount of AIS-based trajectory data to enhance the ability to predict and assess ship collision risks. With the improvement of data quality and accessibility, the application of AIS data has expanded from navigation safety to a wider range. Among them, the detection of ship clusters based on real-time trajectory information has become an emerging research topic. It plays a vital role in identifying potential high-risk areas and understanding local ship traffic patterns. However, there are still some shortcomings in existing research, such as ignoring or simplifying ship dynamics, focusing only on traffic density, and difficulty in discovering clusters with different traffic densities. Therefore, the application of existing methods is ineffective and infeasible in complex scenarios involving different traffic densities and dynamic traffic behaviors. For real-time conflict hotspot identification in complex waters, the following factors need to be fully considered: 1) the spatiotemporal dynamics and uncertainty of ship motion; 2) the conflict connectivity and spatial compactness of encountering ships. Therefore, the development of more advanced models to capture high-risk traffic clusters needs to be urgently addressed. Summary of the invention

[0004] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters, so as to achieve the effect of automatically and adaptively discovering high-risk multi-ship encounter situations.

[0005] The present invention is achieved through the following technical solutions:

[0006] Regional traffic demarcation methods to identify hotspots of ship traffic conflicts in complex waters, including;

[0007] Probabilistic conflict detection is performed based on ship AIS trajectory data to obtain the severity of conflicts between ships;

[0008] A composite distance measurement model is constructed based on the obtained conflict severity, and based on the measurement results, the constrained SSN clustering algorithm is used for clustering;

[0009] Based on the hierarchical bi-objective optimization algorithm, the optimal clustering solution is obtained.

[0010] The above technical scheme first handles the dynamics and disturbances of traffic movement through probabilistic conflict detection; then ensures conflict connectivity and spatial compactness by designing an extended clustering algorithm, so as to facilitate the division of regional ship traffic into multiple clusters; then formulates a hierarchical bi-objective optimization algorithm to generate a robust clustering scheme that can adapt to the dynamic changes of ship traffic, so that traffic clusters with strong conflict connectivity and high density can be found in high-density waters, have adaptive robustness to various traffic scenarios, and perform well under the uncertainty of ship motion.

[0011] As a further technical solution, the method further includes: obtaining the AIS trajectory data of the ship; performing uncertainty pattern extraction; performing ship uncertain motion prediction; and performing probabilistic conflict detection. The AIS trajectory data covers static attributes and dynamic attributes. When performing uncertainty pattern extraction, kernel density estimation and Douglas-Peucker algorithm are used. Based on the extracted uncertainty pattern, combined with the ship's navigation plan, the ship's uncertain motion prediction is performed.

[0012] As a further technical solution, the composite distance metric model is defined as follows:

[0013]

[0014] Among them, dist ij represents the actual distance between ships i and j; C(γ) ijrepresents the conflict severity; R1 and R2 represent specific parameters to classify the conflict severity into three groups: intolerable, ALARP, and negligible. The first part of the above formula helps to merge ships with intolerable conflict severity together (the distance between them is 0), the second part uses weights to strengthen the connection between ships with ALARP conflict severity, and the third part shows that negligible conflict severity does not affect the distance metric.

[0015] As a further technical solution, after clustering using the constrained SSN clustering algorithm, a fine-tuning strategy is performed, including: outlier removal and outlier addition. During the clustering process, some ships are sometimes assigned to inappropriate clusters. This may be explained by the following facts: 1) The distribution of ship traffic in space is random due to various interferences, and in some cases has unclear cluster distribution characteristics. 2) The SNN metric considers the overlap between the k-nearest neighbor lists of objects, rather than a constant density value, to find clusters in non-uniform areas. For complex traffic scenarios, it may lead to the formation of clusters with significant density differences. Therefore, a fine-tuning strategy is proposed to further assign ships to appropriate clusters.

[0016] As a further technical solution, outlier point deletion further includes: determining the adjacency distance between each point and its second nearest neighbor in the same cluster, and sorting all the adjacency distances. If the point with the largest adjacency distance meets the following conditions,

[0017] (1) Its adjacent distance is greater than the preset times of the third largest adjacent distance,

[0018] (2) It has no intolerable or ALARP conflicts with peers in the same cluster,

[0019] The point is set as an outlier and deleted iteratively.

[0020] Usually, outlier removal is performed first, followed by outlier addition.

[0021] As a further technical solution, the outlier addition further includes: assuming that the outlier belongs to the cluster closest to the outlier, calculating the average distance between the outlier and other points in the cluster, and adding the outlier to the cluster when the average distance of the outlier is not the largest; for outliers with intolerable or ALARP conflict relationships, these outliers are added to the cluster only when these outliers meet the distance requirements at the same time.

[0022] As a further technical solution, the hierarchical dual-objective optimization algorithm further includes:

[0023] Generate candidate sets based on grid search strategy;

[0024] Infeasible solutions are eliminated in the preprocessing stage, mainly considering three constraints: 1) the noise / outlier ratio of the dataset should be less than noise_percent; 2) the number of generated clusters should be greater than min_nc; 3) the number of points in each cluster should be no less than min_np;

[0025] A non-dominated solution set is searched from the feasible solution pool to ensure the conflict connectivity and spatial compactness of the formed clusters.

[0026] As a further technical solution, conflict connectivity is measured based on the extent to which there are conflicting pairs of ships that are not assigned to the same cluster, as follows:

[0027]

[0028] Among them, N vc represents the number of conflicting pairs that are not assigned to the same cluster, C(γ) i Represents the severity of the conflict, V CC The smaller it is, the better the conflict connectivity;

[0029] The spatial compactness is measured based on the CH metric as follows:

[0030]

[0031] Where n is the number of samples, c is the number of clusters, and n k is the number of samples belonging to the kth cluster, z k is the center of the kth cluster, z is the center of the entire data set, the larger the CH index, the larger the inter-cluster distance, the smaller the intra-cluster distance, and the better the spatial compactness of the generated cluster.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] (1) The proposed regional traffic delineation method can effectively and reliably capture and identify local traffic clusters by integrating conflict connectivity and density compactness of encountering ships. Compared with existing related studies, this method helps regulators better understand and reveal local traffic conflict patterns.

[0034] (2) The present invention proposes a new density clustering method by effectively combining similarity measurement, constrained SNN clustering and fine-tuning strategy. It incorporates conflict relations into the distance measurement between ships to provide a powerful similarity measurement model, which can assign high-conflict ship pairs to the same cluster and can handle scenarios with different traffic densities, making it feasible and applicable in complex maritime traffic waters.

[0035] (3) In order to find the optimal clustering solution, the present invention designs a hierarchical dual-objective optimization algorithm based on a grid search strategy. The algorithm is practical and robust to various traffic scenarios by explicitly considering the hierarchical priorities between different objectives. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The figure is a flow chart of a method for regional traffic division for identifying hot spots of ship traffic conflicts in complex waters according to an embodiment of the present invention.

[0037] Figure 2 A schematic diagram of conflict definition between ships according to an embodiment of the present invention.

[0038] Figure 3 (a)-(c) are schematic diagrams of typical clusters that can be improved based on outlier deletion according to an embodiment of the present invention.

[0039] Figure 4 (a)-(b) are schematic diagrams of typical clusters that can be improved based on adding outliers according to an embodiment of the present invention.

[0040] Figure 5 Schematic diagram of the core port area of ​​Ningbo-Zhoushan Port according to an embodiment of the present invention.

[0041] Figure 6 (a)-(d) are schematic diagrams of regional ship traffic clustering results according to an embodiment of the present invention.

[0042] Figure 7 Schematic diagram of the evolution of regional and local maximum ship traffic density and conflict severity within one hour according to an embodiment of the present invention.

[0043] Figure 8 Schematic diagram of percentages of ship traffic scenarios with feasible solutions according to an embodiment of the present invention.

[0044] Fig. 9 Schematic diagram of multiple algorithm comparison according to an embodiment of the present invention.

[0045] Fig.10 (a)-(c) are schematic diagrams for comparing the extended clustering method with other methods in different traffic scenarios according to an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions of various embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0047] The present invention provides a regional traffic partitioning method for identifying conflict hotspots of ship traffic in complex waters to automatically and adaptively discover high-risk multi-ship encounter situations. Instead of focusing on the collision risk between two ships or the entire global traffic in the regulated waters, the present invention focuses on capturing and identifying local conflict hotspots. To achieve this goal, the present invention first proposes a probabilistic conflict detection method that can handle the dynamics and disturbances of traffic movement. Then, a clustering algorithm that can ensure conflict connectivity and spatial compactness extension is designed to partition regional ship traffic into multiple clusters. In addition, a hierarchical bi-objective optimization algorithm is formulated as an integral part of the method framework to generate a robust clustering scheme that can adapt to the dynamic changes of ship traffic. Finally, the performance of the proposed model is evaluated using AIS-based trajectory data collected at the port.

[0048] like Figure 1 As shown in FIG. 1 , the method proposed in the present invention consists of three modules: 1) performing probabilistic conflict detection, 2) performing an extended clustering algorithm, and 3) searching for hierarchical optimized clustering results. These modules are organized and integrated through a series of related steps to identify real-time multi-vessel clusters with high risk / high density. The detailed description of the key technologies in each module is specifically emphasized below.

[0049] (1) Perform probabilistic conflict detection

[0050] In order to adapt the collision detection method to complex waters with heavy traffic, this paper develops a probabilistic conflict method from the perspective of probabilistic risk. It considers the effects of the inherent dynamics and uncertainty of ship spatiotemporal motion, providing a basis for capturing high-risk hotspots and discovering hidden traffic conflict patterns. The proposed conflict detection method involves two important parts: 1) conflict definition and characterization; 2) conflict severity measurement.

[0051] Ship conflict usually refers to the possibility that a ship will violate the minimum safety distance in the future. The present invention adopts a classic ship domain model suitable for restricted waters to define conflicts between ships. Figure 2 An example of conflict identification is provided in . When formula (1) is satisfied within a certain period of time in the future, the two ships are considered to have a potential conflict:

[0052] Dist AB (t)≤SD A (t)+SD B (t) (1)

[0053] Among them, Dist AB Indicates the distance between ships, SD A and SD B Respectively represent the distance from each ship center to the boundary of its ship territory.

[0054] Since various uncertainties will affect the future trajectory of the ship, whether formula (1) is satisfied is a probability problem. Assume that the probability density function of the unsafe interval between ships (i.e., the distance between ships A and B is less than or equal to the safe interval requirement) is given by f L(t) Definition: The instantaneous probability of a conflict occurring at time t is as follows:

[0055]

[0056] Where L(t) represents Dist(t)-SD A (t)-SD B (t).

[0057] In order to effectively quantify the potential collision risk between ships that will encounter in the future, the present invention designs a practical conflict severity measurement method by combining the maximum conflict probability and the time of occurrence. These two indicators are similar to DCPA and TCPA, and they both play an important role in safe navigation. One reflects the maximum intensity of the conflict, while the other reflects the difficulty of conflict resolution, because the less reserved time, the less chance of resolving the conflict. Therefore, an exponential function is used to fuse them, as shown below:

[0058]

[0059] Among them, MPC is the maximum PC(t) within the prediction time range, t MPC represents the time of occurrence of the maximum conflict probability, and T represents the prediction time range.

[0060] In practical applications, probabilistic conflict estimation needs to focus on two factors: 1) modeling the uncertainty of future ship motion, that is, the probability density function of the future predicted trajectory; 2) real-time conflict probability calculation. In order to solve these problems, the present invention introduces a probabilistic conflict detection framework, in which a ship motion uncertainty model is used to process uncertain trajectory prediction, and a two-stage Monte Carlo simulation algorithm is designed to efficiently and accurately calculate the severity of the conflict. The specific steps of the two-stage Monte Carlo simulation algorithm are detailed in the content disclosed in the Chinese patent with publication number CN112966332A on June 15, 2021.

[0061] Regarding trajectory prediction, interactive sensing prediction is considered the most accurate method. This can be attributed to the fact that it allows communication between ships and each ship has a better understanding of its own intentions. The intention information usually refers to the planned trajectory of the ship, including a series of waypoints such as WP i＝1,2,…,n+1 Therefore, we assume that the encountering ships will exchange their sailing plans through interaction.

[0062] In order to consider the effects of ship motion dynamics and uncertainty, we combine the deterministic motion associated with the voyage plan and the uncertain motion associated with various disturbances to construct a ship motion model. It consists of three main parts: 1) continuous dynamics showing the physical motion of the ship; 2) discrete dynamics consistent with the voyage plan; 3) random components caused by various disturbances (such as environmental, mechanical and human factors). On this basis, we can use formula (4) to predict the instantaneous position of the ship at time T.

[0063]

[0064] Among them, t c It is the present moment, represents the initial position of ship A, defines the ship's nominal speed, represents the rotation matrix related to the ship's calibrated heading, Represents the uncertainty component of the predicted position of the ship. Based on the predicted uncertain trajectory, the probability of instantaneous conflict can be further calculated.

[0065] Regarding the conflict probability calculation, the present invention provides a two-stage Monte Carlo simulation algorithm to allow for a fast and accurate estimation of the conflict severity. It is designed based on the fact that we focus on the maximum conflict probability within the prediction horizon. Therefore, before performing a large number of Monte Carlo iterations, we roughly extract time points with high conflict probabilities. In addition, the present invention integrates the Hoeffding inequality into the model to determine the quantitative bounds of the approximation effect. The results show that the algorithm achieves an efficient and robust calculation of the conflict severity.

[0066] (2) Execute the extended clustering algorithm

[0067] As described above, the present invention aims to develop a method for dividing regional ship traffic into multiple clusters to achieve two basic requirements: (1) conflict connectivity and (2) spatial compactness. Conflict-connected ship traffic clusters mean that ships with high conflict severity are assigned to the same cluster, while spatially compact clusters mean that ships are organized together based on their spatial proximity. These two goals are crucial for degrading regional ship traffic risks and further local situational awareness.

[0068] In essence, these two criteria sometimes conflict with each other. This is because the severity of conflict between ships is not entirely related to density. It is also affected by the convergence trend, spatial approach rate, size, speed, etc. between the encountering ships. Therefore, the present invention introduces an extended clustering method to solve this problem. First, a composite distance metric is designed to combine conflict connectivity and density compactness. Based on the distance metric results, the constrained SSN clustering algorithm is further used for clustering. Finally, a fine-tuning strategy is applied to ensure that all ships are assigned to appropriate clusters.

[0069] Distance metric is an important component of implementing clustering algorithms. Therefore, it is necessary to develop a new distance metric to help clustering methods meet the two division criteria mentioned above. In practice, conflict connectivity is more concerned by regulators than density compactness. This is because high conflict severity clearly means that there is a potential risk of collision, while high traffic density only represents the complexity of traffic conditions. In addition, risk states can be divided into three categories based on probability and / or consequences: intolerable, as low as possible (ALARP), and negligible. The intolerable state means that measures must be taken to eliminate or control the risk; the ALARP state means that the risk can be accepted when some solutions are taken to reduce the risk to the lowest level reasonably practicable; the negligible state indicates that no further strategy is required if the existing risk is maintained during operation.

[0070] Based on the risk attributes and the goal of the present invention to reveal the actual conflict pattern, the present invention proposes a composite distance measurement model to support robust distance measurement by embedding the conflict severity into the actual distance, which is defined as follows:

[0071]

[0072] Among them, dist ij represents the actual distance between ships i and j, R1 and R2 represent specific parameters to classify the conflict severity into three groups (i.e., intolerable, ALARP, and negligible). The first part of equation (5) helps to merge ships with intolerable conflict severity together (the distance between them is 0), the second part uses weights to strengthen the connection relationship between ships with ALARP conflict severity, and the third part shows that negligible conflict severity does not affect the distance metric.

[0073] Clustering is a typical unsupervised learning technique that divides a dataset into multiple clusters based on similarity measures. It has been applied in many fields, including environmental science, geographical science, information technology, astronomy, and business intelligence. Among the existing methods, DBSCAN is one of the most popular data mining methods in the marine field. This is because it has ideal characteristics, such as the ability to cluster datasets of arbitrary shapes, the ability to identify and exclude noise samples, and the ability to automatically determine the number of clusters. Therefore, it was chosen to perform the task of real-time hotspot area identification in complex waters.

[0074] Nevertheless, maritime traffic is complex and dynamic, and directly applying DBSCAN may not produce ideal traffic clusters. On the one hand, DBSCAN cannot handle datasets with uneven density because it uses a fixed density criterion to globally search all clusters. Therefore, DBSCAN is difficult to apply to all ship traffic scenarios, especially complex scenarios involving different traffic densities and changing traffic behaviors. On the other hand, DBSCAN may not always guarantee that high-conflict ship pairs are assigned to the same cluster. However, it is necessary to group ship pairs with intolerable conflict severity into the same cluster to capture the true local conflict patterns.

[0075] For the first problem, one solution is to use the SNN similarity metric. It is relatively insensitive to changes in density and is very effective for discovering clusters in non-uniform areas. Unlike the distance metric in DBSCAN, it calculates the similarity between two individuals based on the number of common elements between their k-nearest neighbor lists, which is defined by the following equation:

[0076] Simlalirity ij = intersect(NN i (k)∩NN j (k)) (6)

[0077] where NN i (k) is the subset of the k nearest neighbors of individual i. Since the distance (i.e., dissimilarity) is used as the input of DBSCAN, we use k minus Simlaritity ij To represent the dissimilarity between individuals i and j.

[0078] As for the second question, hard constraints are added to DBSCAN to ensure that pairs of ships with intolerable conflict severity are grouped into the same cluster. We use graph theory to identify pairs of ships that must be connected. We set the ships as nodes and use edges to connect the ships with intolerable conflict severity. Then we can get the i ,…,icc sThe must-connected closure (MLC) composed of icc i Represents a subgraph. icc i Every pair of ships in can be connected by at least one sequence edge.

[0079] The detailed process of constrained SNN is described in Algorithm 1. Its program is an extension of DBSCAN. Three hyper parameters (i.e., EPS, MinPts, and k) and a must-be-connected closure (MLC) are taken as input. The hyper parameters EPS and MinPts come from DBSCAN, representing the maximum neighborhood radius of a sample point and the minimum number of points contained in this neighborhood. Based on these two parameters, the relationship between two points can be defined as: directly density reachable, density reachable, and density connected. In addition, the core point can be determined according to the number of directly density reachable points (Algorithm 1, line 6). The hyper parameter k means that the k most similar neighbors of each point are retained for the SNN similarity measure. Compared with DBSCAN, the improvement of constrained SNN can involve two aspects:

[0080] 1) The SNN metric is embedded in the program to support a more robust similarity metric that is applicable to traffic scenarios with different densities (Algorithm 1, line 3).

[0081] 2) During the loop iteration, if a point belongs to MLC, all points in MLC that belong to the same subgraph are arranged into the current cluster to satisfy the must-be-connected constraint (lines 10 and 15 of the algorithm).

[0082]

[0083] I: sample; N c : The serial number of the cluster; N ε (i): The neighborhood of point i based on the radius of Eps; v i : 1 means point i has been used, 0 means the opposite; MLC(i): point i belongs to the graph.

[0084] During the clustering process, some ships are sometimes assigned to inappropriate clusters. This may be explained by the following facts: 1) The distribution of ship traffic in space is random due to various interferences and has unclear cluster distribution characteristics in some cases. 2) The SNN metric considers the overlap between the k-nearest neighbor lists of objects instead of a constant density value to find clusters in uniform areas. For complex traffic scenarios, it may lead to the formation of clusters with significant density differences. Therefore, this paper proposes a fine-tuning strategy to further assign ships to appropriate clusters. It consists of two stages: 1) deleting points far away from the cluster, and 2) adding some outliers to the appropriate clusters.

[0085] Figure 3(a)-(c) show 3 typical clusters that can be improved using outlier removal. The vertical axis "d" represents the distance between a point and its second nearest neighbor, and the horizontal axis "p" represents the sequence number of the point. Obviously, clustering performance can be improved by removing triangles. To achieve this goal, we determine the distance between each point and its second nearest neighbor (2-nn) in the same cluster and sort them. Figure 3 The bottom shows the 2-nn distance of each point. Figure 3 In these graphs, the triangles have much larger 2-nn distances than other points. Therefore, we remove these triangles in an iterative manner. In each iteration, if the point with the largest 2-nn distance satisfies the following conditions, it is set as an outlier.

[0086] (1) Its 2-nn distance is greater than th1 times the third largest 2-nn distance, where th1 is a custom coefficient.

[0087] (2) It has no intolerable or ALARP conflicts with peers in the same cluster.

[0088] The procedure will continue to execute on the current cluster until no other points are deleted. The reason for using 2-nn instead of 1-nn (first nearest neighbor) is that Figure 3 In (c), when two points far from the cluster are close to each other, the latter has no effect. The detailed algorithmic steps for outlier removal are provided in Algorithm 2.

[0089]

[0090] C i : A subset belonging to the i-th cluster; MHCC: A dataset that has intolerable or ALARP conflict relationships with other points.

[0091] Figure 4 (a)-(b) show two typical clusters that can be improved by adding outliers, where the vertical axis "d" represents the average distance between a point and other points in the same cluster, and the horizontal axis "p" represents the serial number of the point. The triangle represents the outlier point obtained based on the constrained SNN algorithm. It can be seen that the distance between the outlier point and the points in the cluster is relatively close. We assume that the outlier point belongs to the cluster and calculate the average distance between each point and other points in the cluster. From Figure 4 As can be seen in (b), the average distance of the outliers is not the largest. Therefore, we add outliers to make the cluster more compact. For outliers with intolerable or ALARP conflict relationships ( Figure 4 (b)), only when they satisfy the distance requirement at the same time, they are added to the cluster. Algorithm 3 provides the procedure for adding outliers. It should be noted that outliers are preferentially added to closer clusters (Algorithm 3, lines 2-3).

[0092]

[0093] o: outlier data set; The point that has an intolerable or ALARP conflict relationship with the i-th abnormal point; v oi : 1 means the i-th outlier has been processed, otherwise 0.

[0094] (3) Searching for hierarchical optimization clustering results

[0095] One difficulty in implementing extended clustering methods is how to adaptively determine the values ​​of hyperparameters to ensure conflict connectivity of clusters while maintaining spatial compactness of clusters. More importantly, these two criteria conflict with each other. Therefore, their non-dominated relationship must be considered to allow simultaneous optimization. In addition, using different hyperparameter combinations leads to a variety of clustering results, many of which are infeasible solutions. This study adopts a hierarchical bi-objective optimization algorithm to solve it.

[0096]

[0097] CR i The percentage of noise points; CR i The number of clusters; CR i The number of points in the smallest cluster.

[0098] Algorithm 4 gives the overall procedure of hierarchical bi-objective optimization. The algorithm first generates candidate sets based on a grid search strategy. Then, infeasible solutions are eliminated in the preprocessing stage, mainly considering three constraints: 1) The noise / outlier ratio of the dataset should be less than noise_percent, because a larger number of outliers means that ship traffic clusters cannot be accurately detected; 2) The number of clusters generated should be greater than min_nc, because a smaller number of clusters cannot effectively degrade the regional traffic risk in high-density waters; 3) The number of points in each cluster should be no less than min_np, because clusters with a certain number of points can be regarded as a cluster. In order to ensure that the formed clusters are conflict-connected and spatially compact, we further search for non-dominated solution sets from the feasible solution pool. Regarding the judgment of these two criteria, two indicators are used to evaluate. Conflict connectivity is measured based on the extent to which conflicting ship pairs are not assigned to the same cluster, and is defined as follows:

[0099]

[0100] Among them, N vc represents the number of conflicting pairs that are not assigned to the same cluster, C(γ) i Represents the corresponding conflict severity. CCThe smaller the value, the more effective the conflict connectivity is. The spatial compactness is measured based on the Calinski–Harabasz (CH) index, as shown below:

[0101]

[0102] Where n is the number of samples, c is the number of clusters, and n k is the number of samples belonging to the kth cluster, z k is the center of the kth cluster, z is the center of the entire data set, trace B represents the inter-cluster distance, and trace W represents the intra-cluster distance. The larger the CH index, the larger the inter-cluster distance and the smaller the intra-cluster distance. This means that the spatial compactness of the generated clusters is well satisfied.

[0103] Example

[0104] This example uses Ningbo-Zhoushan Port as the research area to test the applicability and effectiveness of the proposed regional traffic division method. It is the largest port in the world in terms of cargo throughput. At the same time, the waters are busy with traffic, the navigable area is limited, and the traffic behavior is dynamic and uncertain. Therefore, it is a typical representative of complex waters. Figure 5 The port area is shown, which is between latitudes 29°43N and 30°02N, and longitudes 121°52E and 122°22E. One month of AIS data records were collected from this area between November 1, 2018 and November 30, 2018.

[0105] Although the amount of information in the received raw AIS data is large, it has inherent weaknesses, such as data noise or errors in position and speed information caused by data transmission or technical failures. Therefore, it is necessary to clean and filter the data before conducting experimental analysis. In this study, we adopted a systematic cleaning procedure, including noise removal of each ship attribute, track extraction and separation, track consistency confirmation, and track interpolation.

[0106] To test the performance of the proposed method, a case study is used to illustrate its online application in high-risk hotspot identification. Figure 6 An example of vessel traffic cluster identification within a one-hour period in the study area is shown. Figure 6 In (a)-(b), visualization of ship traffic clusters at two moments is shown, where points represent the spatial distribution of ship traffic and lines represent intolerable or ALARP conflicts between ships. Most ships with intolerable or ALARP conflicts are assigned to one cluster, and ships in the same cluster are spatially compact. This means that by dividing ship traffic into multiple clusters, local hotspots can be effectively detected. Therefore, this method provides higher resolution for local risk identification by degrading regional ship traffic risks by considering multi-ship relationships.

[0107] Based on the detected ship traffic clusters, key traffic clusters can be further identified. Two indicators are used to describe the characteristics of each cluster. One is the sum of conflict severity, and the other is the density index, which is expressed as:

[0108]

[0109] Wherein, density(x) represents the density of x, N(x,k) represents the subset of its k nearest neighbors, and distance(x,y) represents the distance between point x and point y. In this embodiment, the density of each cluster is calculated using the average density of ships in the same cluster. Figure 6 (c)-(d) show the sum of the density and conflict severity of each cluster in the above figure. It can be clearly seen which clusters are more worthy of attention. For example, it is observed that Figure 6 Clusters 1 and 2 in (c) have high conflict severity, while Figure 6 Cluster 5 in (d) has high traffic density. Therefore, the regional traffic segmentation method of this embodiment provides valuable insights into where to enhance situational awareness and provide hazard warnings.

[0110] Figure 7 The evolution of the density and conflict severity of regional and local ship traffic is further shown, showing the maximum density and conflict severity of local clusters for each time slice. Two phenomena can be found in the figure: (1) The global traffic density varies slightly over time, while the local maximum traffic density fluctuates considerably. These results indicate that regional traffic density does not play a significant role in enhancing maritime traffic monitoring capabilities. In contrast, detecting local high-density hotspots can help regulators better understand traffic conditions and improve their ability to handle complex traffic scenarios. (2) It was observed that the global and local maximum conflict severity showed a consistent trend. This has important practical significance for local conflict risk management because local conflict clusters are effectively captured. In summary, the regional traffic division method proposed in this embodiment is expected to be applied to intelligent transportation service systems to alleviate the safety monitoring pressure of future regulators.

[0111] Model comparison and verification are the key to ensure the superiority and effectiveness of model results. Therefore, this embodiment first analyzes the performance of the proposed regional traffic division method by comparing it with the DBSCAN algorithm.

[0112] The two hyper parameters (i.e., EPS and MinPts) in DBSCAN are taken as [1, 1.5, …, 6 km] and [2, 3, …, 10] to search for the best clustering results. Figure 8The percentage of ship scenarios with feasible cluster identification solutions when using these two algorithms is shown (Con_DBSCAN_CSM and Con_SNN_CSM represent the algorithm corresponding to DBSCAN and the algorithm proposed in this study, respectively). It is observed that almost all traffic scenarios can obtain feasible solutions by the proposed method, while only 33.39% of traffic scenarios can obtain feasible solutions by using DBSCAN. Even if we adjust the tolerance of the outlier percentage to 20%, there is still a large part of traffic scenarios that cannot obtain feasible clustering solutions based on DBSCAN. These results show that DBSCAN has limitations in handling complex traffic scenarios with different densities.

[0113] In order to demonstrate the functionality and practicality of the key modules in the extended clustering method, this embodiment compares it with the following algorithms: standard SNN, standard SNN using a composite distance metric, and constrained SNN without a composite distance metric. Fig. 9 Comparative results on their dominance relations are provided, where “outperforms” means that the proposed model outperforms other algorithms and “inferior” is the opposite. The results show that the proposed method significantly outperforms the standard SNN and the standard SNN using the composite distance metric. This means that constructing hard connection constraints for intolerable conflicting ship pairs can significantly improve the clustering performance. At the same time, it can be observed that the extended clustering model outperforms the constrained SNN to a certain extent without using the composite distance metric. This is because the composite distance metric is designed in a more precise way to allow for simultaneous consideration of conflicting connectivity and spatial compactness of clustered ship traffic. However, traditional similarity / distance metrics cannot handle multiple clustering objectives and therefore perform poorly when multiple clustering attributes are considered. Therefore, hard constraints and composite distance metrics are crucial to improving clustering performance.

[0114] Fig. 9 In Fig. 3, A1, A2, A3 and A4 represent the standard SNN (SNN), the standard SNN using composite distance metric (SNN_CSM), the constrained SNN without composite distance metric (Con_SNN) and the proposed method (Con_SNN_CSM), respectively.

[0115] In addition, this embodiment further compares the performance of different models in different traffic scenarios to evaluate the generalization ability and stability of the extended clustering method. Fig.10 As shown in (a)-(c), Fig.10 (a) Con_SNN_CSM vs SNN; Fig.10 (b) Con_SNN_CSM vs SNN-CSM; Fig.10(c) Con_SNN_CSM vs Con_SNN. The proposed method outperforms all other models when the number of ships varies. These results confirm the applicability and robustness of the extended clustering method, which can adapt to various traffic scenarios.

[0116] Identifying potential high-risk ship traffic groups in complex waters is crucial for maritime traffic safety management. This paper proposes a systematic ship traffic segmentation method to automatically and adaptively capture high-risk multi-ship encounter situations. It effectively integrates composite distance metrics, constrained SNN clustering, fine-tuning strategy and hierarchical bi-objective optimization, so that regional ship traffic can be divided into multiple conflict-connected and spatially compact clusters. The method has the following characteristics: (1) It can adaptively discover traffic clusters with strong conflict connectivity and high density in high-density waters; (2) It allows robust processing of traffic situations with different densities; (3) It shows good performance in the presence of ship motion uncertainty.

[0117] The performance of the proposed regional vessel traffic segmentation method was evaluated and tested based on actual AIS data from Ningbo Zhoushan Port. The results show that the proposed method performs well in identifying local hotspot areas in complex traffic scenarios and provides valuable insights into where to enhance situational awareness and provide hazard warnings. At the same time, the functionality and practicality of the key modules in the proposed traffic segmentation method were tested and verified by comparison with other clustering methods. In summary, the developed method shows great potential in improving the maritime traffic monitoring capabilities of regulators and facilitating the implementation of risk mitigation measures.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters, characterized by: include; Probabilistic conflict detection is performed based on ship AIS trajectory data to obtain the severity of conflicts between ships; A composite distance measurement model is constructed based on the obtained conflict severity, and based on the measurement results, constrained SSN clustering algorithm is used for clustering; The composite distance metric model is defined as follows: Among them, dist ij represents the actual distance between ships i and j; C(γ) ij Indicates the severity of the conflict; R1 and R2 represent specific parameters to classify the severity of the conflict into three groups: intolerable, ALARP, and negligible; The constrained SNN clustering algorithm is improved on the basis of the DBSCAN data mining algorithm as follows: 1) The SNN metric is embedded in the program to support a more robust similarity metric that is applicable to traffic scenarios with different densities; 2) During the loop iteration, if a point belongs to the must-connect closure MLC, all points in the must-connect closure MLC that belong to the same subgraph are arranged into the current cluster to satisfy the must-connect constraint; Based on the hierarchical bi-objective optimization algorithm, the optimal clustering solution is obtained, wherein the hierarchical bi-objective optimization algorithm includes: Generate candidate sets based on grid search strategy; Infeasible solutions are eliminated in the preprocessing stage, mainly considering three constraints: 1) the noise / outlier ratio of the dataset should be less than noise_percent; 2) the number of clusters generated should be greater than min_nc; 3) the number of points in each cluster should be no less than min_np; A non-dominated solution set is searched from the feasible solution pool to ensure the conflict connectivity and spatial compactness of the formed clusters.

2. The regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters according to claim 1 is characterized in that: The method further comprises: acquiring AIS track data of the ship; performing uncertainty pattern extraction; performing ship uncertain motion prediction; and performing probabilistic conflict detection.

3. The regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters according to claim 1 is characterized in that: After clustering using the constrained SSN clustering algorithm, fine-tuning strategies are performed, including: outlier deletion and outlier addition.

4. The regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters according to claim 3 is characterized in that: The outlier point removal further includes: determining the adjacency distance between each point and its second nearest neighbor in the same cluster, and sorting all the adjacency distances. If the point with the largest adjacency distance meets the following conditions, (1) Its adjacent distance is greater than the preset times of the third largest adjacent distance, (2) It has no intolerable or ALARP conflicts with peers in the same cluster, The point is set as an outlier and deleted iteratively.

5. The regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters according to claim 3 is characterized in that: The outlier point addition further includes: assuming that the outlier point belongs to the cluster closest to the outlier point, calculating the average distance between the outlier point and other points in the cluster, and adding the outlier point to the cluster when the average distance of the outlier point is not the largest; for outliers with intolerable or ALARP conflict relationships, these outliers are added to the cluster only when these outliers meet the distance requirements at the same time.

6. The regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters according to claim 1 is characterized in that: The hierarchical bi-objective optimization algorithm further includes: Generate candidate sets based on grid search strategy; Infeasible solutions are eliminated in the preprocessing stage, mainly considering three constraints: 1) the noise / outlier ratio of the dataset should be less than noise_percent; 2) the number of clusters generated should be greater than min_nc; 3) the number of points in each cluster should be no less than min_np; A non-dominated solution set is searched from the feasible solution pool to ensure the conflict connectivity and spatial compactness of the formed clusters.

7. The regional traffic division method for identifying hot spots of ship traffic conflicts in complex waters according to claim 6 is characterized in that: Conflict connectivity is measured based on the extent to which conflicting pairs of ships are not assigned to the same cluster, as follows: Among them, N vc represents the number of conflicting pairs that are not assigned to the same cluster, c(γ) i Represents the severity of the conflict, V CC The smaller it is, the better the conflict connectivity; The spatial compactness is measured based on the CH metric as follows: Where n is the number of samples, c is the number of clusters, and n k is the number of samples belonging to the kth cluster, z k is the center of the kth cluster, z is the center of the entire data set, the larger the CH index, the larger the inter-cluster distance, the smaller the intra-cluster distance, and the better the spatial compactness of the generated cluster.

Images