Ship collision avoidance behavior and style identification method

Abstract

The invention discloses a ship collision avoidance behavior and style identification method. The method comprises the following steps: acquiring real AIS data of a water area to be identified; preprocessing the real AIS data of the water area to be identified; calculating inter-ship motion parameters of the two ships based on navigation parameters of the ship and the target ship in different groups, and screening meeting scenes of the ship and the target ship in different groups to form a preliminary meeting scene data set; based on the preliminary meeting scene data set, an improved sliding window algorithm is applied to identify collision avoidance feature points of the ship, the collision avoidance behavior of the ship is represented, and a collision avoidance track data set is constructed; selecting characteristic parameters in the collision avoidance track data set as collision avoidance characteristic indexes; and performing standardized processing on the collision avoidance characteristic indexes, and performing clustering analysis on the collision avoidance characteristic indexes by adopting a K-Means + + algorithm to obtain the collision avoidance style of the ship in the research water area. The method reveals the collision avoidance behavior styles of different ships.

Description

Technical Field

[0001] The invention belongs to the field of ship behavior feature analysis and relates to a method for identifying ship collision avoidance behavior and style. Background Art

[0002] With the rapid development of global maritime transportation, the issue of ship collision avoidance has become one of the hot issues that the industry pays close attention to. At present, the research on ship collision avoidance behavior mainly focuses on the identification of collision avoidance feature points, cluster analysis of ship track points (segments), and ship behavior prediction. When identifying the characteristics of ship track points, selecting appropriate threshold parameters based on different research purposes is a key issue.

[0003] With the scientific research needs of intelligent unmanned ship navigation risk assessment and collision avoidance, higher requirements have been put forward for the definition of ship encounter scenarios, the identification of collision avoidance behaviors, and the quantitative analysis of collision avoidance behavior styles. However, the driving styles of crew members are diverse, and their reactions to the external traffic environment are also different, resulting in different collision avoidance behaviors for crew members in the same encounter scenario. When ships with different driving styles encounter other ships, there may be significant differences in the timing and actions of collision avoidance. In current research, the impact of driving style differences on collision avoidance decisions is often ignored. In-depth research on the characteristic differences of different collision avoidance behaviors can better understand the relationship between driving style and collision avoidance decisions, thereby providing more effective guidance and measures for navigation safety. At present, experts in the field generally believe that ship driving styles can be roughly divided into three categories: aggressive, normal, and conservative. However, existing research has not yet conducted a quantitative category analysis of different styles of ship behavior. Therefore, further exploring and quantifying the impact of different driving styles on collision avoidance decisions is the key to improving navigation safety.

[0004] Based on the technical requirements of smart ships, smart shipping insurance, and smart shipping supervision, how to explore the implicit potential behavior patterns from the ship's track and analyze the characteristics of the ship's collision avoidance timing and collision avoidance actions has certain theoretical and practical significance. Summary of the invention

[0005] In order to solve the above problems, the technical solution adopted by the present invention is: a method for identifying ship collision avoidance behavior and style, characterized in that it includes the following steps:

[0006] Obtain the real AIS data of the waters to be identified;

[0007] Pre-process the real AIS data of the waters to be identified;

[0008] Based on different groups of navigation parameters of own ship and target ship, the inter-ship motion parameters of the two ships are calculated, and different groups of encounter scenarios of own ship and target ship are screened to form a preliminary encounter scenario data set;

[0009] Based on the preliminary encounter scenario dataset, an improved sliding window algorithm is used to identify the feature points of own ship's collision avoidance behavior and construct a collision avoidance trajectory dataset.

[0010] Select the characteristic parameters in the collision avoidance trajectory data set as the collision avoidance characteristic index;

[0011] Clustering algorithm is used to perform cluster analysis on collision avoidance characteristic indicators to obtain the collision avoidance style of ships in the study waters.

[0012] Furthermore: it also includes standardizing the collision avoidance feature indicators, reducing the dimension of the standardized collision avoidance feature indicators through factor analysis, and using a clustering algorithm to perform cluster analysis on the collision avoidance feature indicators based on the reduced dimension collision avoidance feature indicators.

[0013] Further: the process of applying the improved sliding window algorithm to identify the collision avoidance feature points of the own ship based on the preliminary encounter scene data set is as follows:

[0014] The trajectory segment of this ship is composed of N trajectory points. The sliding window is set to a window containing 3 trajectory points (P1, P2, P3). First, the first 3 points of the trajectory segment are taken as the initial sliding window, and a key feature point list key_points[] is created;

[0015] S1: Calculate the Euclidean perpendicular distance between the midpoint P2 of the window and the line connecting the two end points P1 and P3 of the window. If the Euclidean perpendicular distance is greater than the set threshold, go to S2; otherwise, discard point P2, update the sliding window to (P1, P3, P4), and repeat S1;

[0016] S2: Create a heading difference search window including the midpoint P2 of the window. The trajectory points in the window include all trajectory points whose interval with P2 does not exceed the time threshold, and go to S3;

[0017] S3: Calculate the difference in heading of each adjacent track point in the window, and search to find P i To P i+1 The heading difference is the largest, if P i If it is not in the key feature point key_points[], add P to the list i As the key feature point where the ship starts turning in the current search window, the updated sliding window is (P i ,P i+1 ,P i+2 ); otherwise, skip the current sliding window, update the sliding window to (P1, P3, P4), and go to S1;

[0018] Repeat the above steps until the midpoint of the sliding window is updated to the last point of the trajectory segment.

[0019] Further: The formula used in the standardization is as follows:

[0020]

[0021]

[0022] Where: x ij is the jth index of the i-th group of data, is the sample annual average of j evaluation indicators, s j is the sample standard deviation of the jth indicator.

[0023] Further: the process of reducing the dimension of the collision avoidance characteristic index after standardization through factor analysis is as follows:

[0024] First, we need to determine whether the sample data is suitable for factor analysis and perform Kaiser-Meyer-Olkin (KMO) and Bartlett sphericity tests on the sample data;

[0025] Combining the scree plot and the principal component table, the eigenvectors corresponding to the first p eigenvalues ​​when the cumulative variance contribution rate is greater than the contribution rate threshold are selected as the principal components to calculate the factor score of the sample.

[0026] Further: the process of clustering the collision avoidance feature indexes based on the factors after dimension reduction using a clustering algorithm to perform cluster analysis on the collision avoidance feature indexes to obtain the collision avoidance style of the ships in the research waters is as follows:

[0027] The clustering algorithm adopts K-Means++ algorithm;

[0028] The silhouette coefficient method and elbow diagram are combined to select the best k value and determine the optimal number of clusters;

[0029] The collision avoidance behavior styles of ships are analyzed quantitatively and qualitatively by combining the cluster centers and case summary reports of each category.

[0030] Further: The process of selecting the initial cluster center of the K-Means++ algorithm is as follows:

[0031] (1) Randomly select a point from the input data point set as the first cluster center;

[0032] (2) For each point x in the data set, calculate the Euclidean distance D(x) between each point and the nearest cluster center;

[0033] (3) Select a new data point as the new cluster center;

[0034] (4) Repeat (2) and (3) until k cluster centers are selected;

[0035] (5) Use these k initial cluster centers to run the standard K-Means algorithm.

[0036] Furthermore: the selection principle is: the point with larger D(x).

[0037] Furthermore: the collision avoidance characteristic indicators include but are not limited to: the amplitude of heading change; the DCPA, TCPA and relative distance between ships at the time corresponding to the first collision avoidance characteristic point in the collision avoidance phase; the DCPA and relative distance between ships when the relative distance between ships is the smallest.

[0038] A device for identifying ship collision avoidance behavior and style, comprising

[0039] Acquisition module: used to obtain the real AIS data of the waters to be identified;

[0040] Preprocessing module: used to preprocess the real AIS data of the waters to be identified;

[0041] Encounter scenario data set construction module: used to calculate the inter-ship motion parameters of the two ships based on the navigation parameters of different groups of own ship and target ship, screen different groups of encounter scenarios of own ship and target ship, and form a preliminary encounter scenario data set;

[0042] Collision avoidance trajectory dataset construction module: used to identify the feature points of own ship's collision avoidance behavior based on the preliminary encounter scenario dataset, and to filter and construct the collision avoidance trajectory dataset by applying the improved sliding window algorithm;

[0043] Selection module: used to select characteristic parameters in the collision avoidance trajectory data set as collision avoidance characteristic indicators;

[0044] Dimensionality reduction module: used to standardize the collision avoidance feature indicators and reduce the dimension of the standardized collision avoidance feature indicators through factor analysis;

[0045] Cluster analysis module: It is used to perform cluster analysis on the collision avoidance feature indicators based on the dimension reduction, and adopt clustering algorithm to obtain the collision avoidance style of ships in the study waters.

[0046] The present invention provides a method for identifying the collision avoidance behavior and style of ships, and provides a method for identifying the collision avoidance behavior in the process of ship encounter and performing cluster analysis on the behavior characteristics. Based on the real AIS data of the research waters, the inter-ship motion parameters of the two ships are calculated after preprocessing, the encounter scenes are screened, and the improved sliding window algorithm is used to identify the collision avoidance feature points of the ship; the characteristic parameters in the collision avoidance process are extracted as characteristic indicators, and the dimension is reduced by factor analysis after the data is standardized, and then the K-Means++ algorithm is used to perform cluster analysis on the characteristic indicators.

[0047] This method is suitable for extracting the characteristics of ship collision avoidance behavior and mining the ship behavior style. The specific application method is: based on real AIS data, a two-ship encounter scene dataset is constructed; an improved sliding window algorithm is used to identify the collision avoidance feature points of the ship, and six characteristic indicators are selected to construct a characteristic dataset of collision avoidance timing and action in combination with the four requirements of the International Convention on Regulations for Preventing Collisions at Sea for ship avoidance, namely "early, large, wide, and clear"; the characteristic indicator data is standardized, and the factor analysis method is used for dimensionality reduction. Finally, the K-Means++ algorithm is used to perform cluster analysis on the characteristic indicators, and it is demonstrated that the collision avoidance behaviors of different ships have different styles.

[0048] The present invention applies an improved sliding window algorithm to identify the collision avoidance feature points of the ship, performs cluster analysis on the feature indicators based on factor analysis and K-Means++ algorithm, extracts the characteristic parameters of the ship's collision avoidance behavior to characterize the collision avoidance behavior style, and explores whether different ships have style differences in collision avoidance decisions.

[0049] The present invention adopts a clustering algorithm to perform cluster analysis on the collision avoidance behavior characteristics during the ship encounter process, extracts the ship collision avoidance behavior characteristic parameters to characterize the collision avoidance behavior style, and demonstrates that different ships have style differences in collision avoidance decision-making.

[0050] The present invention identifies the characteristic points of collision avoidance behavior based on the ship trajectory, reveals the collision avoidance behavior styles of different ships, and helps to understand the decision-making process of ships in complex marine environments, thereby improving the intelligence level of the ship's autonomous navigation system, which is specifically reflected in the following aspects:

[0051] 1. When constructing a test scenario for a collision avoidance algorithm, considering the ship's behavior style can make the test scenario more realistic;

[0052] 2. Behavioral styles reveal the rules behind the collision avoidance behavior of ships and can be integrated into autonomous navigation algorithms, enabling the system to more accurately predict the behavior of other ships and adjust its own navigation path based on this, thereby helping to reduce collision accidents, optimize routes, reduce fuel consumption, and improve shipping efficiency.

[0053] 3. The clustering results help improve the collision avoidance algorithm of autonomous ships, enabling them to better understand and predict the intentions and behaviors of surrounding ships and make more reasonable navigation decisions;

[0054] 4. Understanding the navigation characteristics of ship groups in the waters can help maritime management departments formulate more effective traffic management strategies and improve maritime traffic efficiency;

[0055] 5. The analysis of ship collision avoidance behavior style can be used as input to the decision support system to assist on-board duty personnel in making quick and accurate judgments in complex marine environments.

[0056] The present invention proposes a research method for ship collision avoidance behavior recognition and style recognition, which has the following beneficial effects.

[0057] Firstly, the data of the present invention adopts real AIS data, which provides a real basis for the quantitative analysis of the collision avoidance behavior style of ships, conforms to the real behavior characteristics of ships in the research waters, and has reference and guiding value for the formulation of traffic management strategies of maritime administration departments.

[0058] Secondly, when mining the characteristic points of ship collision avoidance behavior from the ship track, the present invention adopts an improved sliding window method to identify the scene heading change characteristic points as the ship collision avoidance characteristic points. Considering the influence of factors such as wind, wave and current, and shipborne equipment errors on the heading, the present invention does not use the heading difference as the only indicator to identify the characteristic points, but uses the ship track presented by the latitude and longitude coordinates in the AIS data to express the result after the heading change. However, since the present invention emphasizes the characteristics of the timing and action of ship collision avoidance, and the ship track as the result of the collision avoidance action has a time delay, in order to avoid the error between the characteristic point moment identified by the ship track and the actual occurrence time of the collision avoidance action, the present invention improves the sliding window algorithm. On the basis of identifying the characteristic points, a new heading difference search window that does not exceed the set time threshold is created in the algorithm, and the maximum heading difference between two adjacent track points in the search window is replaced by the corresponding track point as a new characteristic point. In addition, the algorithm requires a reasonable threshold determination. Since the goal of the ship's heading change in the encounter situation is to increase the distance between the ship and the other ship when passing each other, and the passing distance is often related to the ship's ship width, the present invention sets the threshold in the algorithm to λ times the ship width. This method not only retains the influence of ship width on target passing distance, but also can more accurately identify the timing of the ship taking collision avoidance action, reducing the error caused by the characteristic data of collision avoidance behavior to the subsequent cluster analysis.

[0059] Finally, the present invention adopts K-Means++ algorithm for clustering, and combines the elbow rule and silhouette coefficient to determine the optimal number of clusters, which makes up for the defects of K-Means in the following aspects: ① The number of clusters K needs to be specified in advance: K-Means algorithm needs to determine the number of clusters in advance, which is difficult for unknown number of clusters; ② Sensitive to the selection of initial cluster centers: The selection of initial cluster centers will affect the final clustering results, and different initial selections may lead to different local optimal solutions; ③ Sensitive to noise and outliers: K-Means algorithm is sensitive to noise and outliers, and may misclassify them into a certain cluster. K-Means++ improves the initial centroid selection strategy of K-Means, ensures the dispersion between centroids, reduces the risk of local optimality, and improves clustering quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0061] Figure 1 It is a quantitative schematic diagram of the situation that the ship will encounter;

[0062] Figure 2 It is a flowchart of improving the sliding window method to identify feature points;

[0063] Figure 3 It is a flowchart for constructing collision avoidance trajectory dataset;

[0064] Figure 4 It is the trajectory feature point recognition diagram based on different scenes, where (a) the encounter scene is 2044, (b) the encounter scene is 58283;

[0065] Figure 5 It’s a gravel map;

[0066] Figure 6 It is the elbow diagram;

[0067] Figure 7 is the silhouette coefficient plot;

[0068] Figure 8 It is a visualization diagram of clustering results, where (a) is a three-dimensional scatter plot of original data points, and (b) is a three-dimensional scatter plot of clustering results;

[0069] Fig. 9 are box plots of classification results, including (a) box plot of classification data in the factor “large”, (b) box plot of classification data in the factor “wide and clear”, (c) box plot of classification data in the factor “early”, (d) box plot of classification data in the CRI when the first collision avoidance action is taken, and (e) box plot of classification data in the CRI when the relative distance between ships is the smallest. DETAILED DESCRIPTION

[0070] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0071] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] A method for identifying ship collision avoidance behavior and style comprises the following steps:

[0073] S1: Obtain the real AIS data of the waters to be identified;

[0074] S2: Preprocessing the real AIS data of the water area to be identified;

[0075] S3: Based on the navigation parameters of the own ship and the target ship in different groups, the inter-ship motion parameters of the two ships are calculated, and the encounter scenes of the own ship and the target ship in different groups are screened to form a preliminary encounter scene data set;

[0076] S4: Based on the preliminary encounter scenario dataset, the improved sliding window algorithm is used to identify the collision avoidance feature points of own ship, characterize the collision avoidance behavior of own ship, and construct the collision avoidance trajectory dataset;

[0077] S5: Select characteristic parameters in the collision avoidance trajectory data set as collision avoidance characteristic indicators;

[0078] S6: Standardizing the characteristic indexes, and reducing the dimension of the collision avoidance characteristic indexes after standardization through factor analysis;

[0079] S7: Based on the collision avoidance feature indicators after dimensionality reduction, the K-Means++ algorithm is used to perform cluster analysis on the collision avoidance feature indicators to obtain the collision avoidance style of ships in the study waters.

[0080] The collision avoidance styles include: slightly aggressive, normal, conservative predictive, and large-scale adaptive.

[0081] The slightly aggressive type is characterized by: a higher risk level during collision avoidance action, a smaller collision avoidance range, a late collision avoidance timing, and passing by at a smaller distance to let other ships clear;

[0082] The normal type is characterized by: moderate risk during collision avoidance action, small collision avoidance range, early collision avoidance timing, and passing by at a small distance to let other ships clear;

[0083] The conservative prediction type is characterized by: low risk, small collision avoidance range, moderate collision avoidance timing, and passing by at a large safe distance to let other ships clear;

[0084] The large-scale response type is manifested as: the risk level is relatively high during collision avoidance actions, the collision avoidance range is large, the collision avoidance timing is random, and the ship passes by at a smaller distance to clear other ships.

[0085] Steps S1 / S2 / S3 / S4 / S5 / S6 / S7 are executed in sequence;

[0086] The preprocessing of the real AIS data of the water area to be identified includes AIS data cleaning and interpolation processing.

[0087] Furthermore, based on different groups of navigation parameters of the own ship and the target ship, the inter-ship motion parameters of the two ships are calculated, and different groups of encounter scenes of the own ship and the target ship are screened to form a preliminary encounter scene data set as follows:

[0088] Based on different groups of navigation parameters of the own ship and the target ship, the inter-ship motion parameters of the two ships are calculated, including the relative distance, relative orientation, relative heading, relative speed, DCPA (closest approach distance), TCPA (minimum approach time), and collision risk index (CRI) of the target ship relative to the own ship.

[0089] The definition of the encounter situation between ships in the "International Regulations for Preventing Collisions at Sea" is: "A situation when two power-driven vessels meet on nearly opposite or opposite courses so as to constitute a risk of collision." The quantitative expression of this application is that the target ship's bearing relative to the ship is between 354° and 360° or 0° and 6°, and the difference in the headings of the two ships is between 174° and 186°. Figure 1 Middle yellow area. The screening conditions for the two-ship encounter sample are:

[0090]

[0091] After screening according to the above conditions, the ship track pairs with potential collision risks in the study waters are identified. Since the AIS data of ships in some scenes are missing, it is judged whether the TCPA at the end point of the scene is less than 0. If it is less than 0, the trajectory segment is considered complete; otherwise, the trajectory segment is considered incomplete due to the lack of AIS data, and the trajectory segment is discarded.

[0092] Based on different groups of encounter scenarios of own ship and target ship, an encounter scenario dataset is formed. The quantitative schematic diagram of the ship encounter situation is shown in the figure below: Figure 1 As shown;

[0093] Improve the process of identifying feature points using the sliding window method Figure 2 As shown; Figure 3 It is a flowchart for constructing collision avoidance trajectory dataset;

[0094] Further: the process of applying the improved sliding window algorithm to identify the collision avoidance feature points of the own ship based on the preliminary encounter scene data set is as follows:

[0095] The trajectory segment of this ship is composed of N trajectory points. The sliding window is set to a window containing 3 trajectory points (P1, P2, P3). First, the first 3 points of the trajectory segment are taken as the initial sliding window, and a key feature point list key_points[] is created.

[0096] S1: Calculate the Euclidean perpendicular distance between the midpoint P2 of the window and the line connecting the two end points P1 and P3 of the window. If the Euclidean perpendicular distance is greater than the set threshold, go to S2; otherwise, discard point P2, update the sliding window to (P1, P3, P4), and repeat S1.

[0097] S2: Create a new heading difference search window including the window midpoint P2, and the track points in the window include all track points whose interval with P2 does not exceed the time threshold. Go to S3.

[0098] S3: Calculate the difference in heading of each adjacent track point in the window, and search to find P i To P i+1 The heading difference is the largest. i If it is not in the key feature point key_points[], add P to the list i As the key feature point where the ship starts turning in the current search window, the updated sliding window is (P i ,P i+1 ,P i+2 ); otherwise, skip the current sliding window and update the sliding window to (P1, P3, P4). Go to S1.

[0099] Repeat the above steps until the midpoint of the sliding window is updated to the last point of the trajectory segment.

[0100] The present invention adopts an improved sliding window method to identify the scene heading change feature points as the ship collision avoidance feature points. Considering the influence of factors such as wind, wave and current, and shipboard equipment errors on the heading, the present invention does not use the heading difference as the only indicator to identify the feature points, but uses the ship trajectory presented by the latitude and longitude coordinates in the AIS data to express the result after the heading change. Since the present invention emphasizes the characteristics of the timing and action of ship collision avoidance, it focuses on the timing of taking collision avoidance actions, and the ship trajectory has a time delay as the result of the collision avoidance action. In order to avoid the error between the feature point moment identified by the ship trajectory and the actual time of the collision avoidance action, the present invention improves the sliding window algorithm, and creates a heading difference search window based on the identification of feature points. In addition, the algorithm requires a reasonable threshold determination. Since in the encounter situation, the goal of the ship's heading change is to increase the distance between the other ship when passing each other, and the passing distance is often related to the ship's ship width, the threshold of the present invention in the algorithm is set to λ times the ship width.

[0101] The improved sliding window algorithm is used to traverse the data point set, and the comprehensive threshold of the vertical Euclidean distance and the heading difference is combined to determine whether the current point is a key feature point, and finally output a list of key feature points. Combined with the feature point list, it is determined whether the ship has a collision avoidance feature point. If so, the trajectory segment is retained, otherwise it is discarded. Take the TCPA at the moment of the first collision avoidance feature point. If TCPA is greater than 0, it means that the two ships still have a collision risk when the first collision avoidance action is taken, and the trajectory segment is retained; otherwise, it is considered that the ship has not taken collision avoidance action due to the target ship, and the trajectory segment is discarded. After screening according to the conditions, a collision avoidance trajectory data set is formed.

[0102] Furthermore, the process of selecting the characteristic parameters in the collision avoidance trajectory data set as the collision avoidance characteristic index is as follows:

[0103] Combined with the four requirements of "early, large, wide and clear" for ship avoidance proposed by the "International Convention on the Regulations for Preventing Collisions at Sea", the present invention selects the following six eigenvalue indicators for subsequent cluster analysis:

[0104] (1) Heading change range: The heading of the initial trajectory point is recorded as C0, and the heading C of each trajectory point is calculated by traversing. i The maximum absolute value of the difference between C0 and C0 is the course change amplitude of the ship during the collision avoidance phase. The formula is:

[0105] Δcog=max|C i -C0|

[0106] (2) The first collision avoidance action, that is, the DCPA, TCPA and relative distance between ships at the moment when the first collision avoidance feature point in the collision avoidance phase corresponds to.

[0107] (3) DCPA and relative distance between ships when the relative distance between ships is the smallest.

[0108] Furthermore, the characteristic indexes are standardized, and the process of reducing the dimension of the collision avoidance characteristic indexes after standardization through factor analysis is as follows:

[0109] The six selected characteristic indicators were standardized and unified in dimension. SPSS software was then used to perform factor analysis, dimensionality reduction, and extract principal components to prepare data for subsequent cluster analysis.

[0110] The formula used for the standardization is as follows:

[0111]

[0112] Where: x ij is the jth index of the i-th group of data, is the sample annual average of j evaluation indicators, s j is the sample standard deviation of the jth indicator.

[0113] First, we need to determine whether the sample data is suitable for factor analysis. We need to perform Kaiser-Meyer-Olkin (KMO) and Bartlett sphericity tests on the sample data. The KMO test is used to test whether the partial correlation coefficient between variables is too small. Generally, when the KMO is greater than 0.9, the effect is best, and when it is less than 0.5, it is not suitable for factor analysis.

[0114] Bartlett's sphericity test is used to test whether the correlation coefficient matrix is ​​a unit matrix. If the conclusion is that the hypothesis is not rejected, it means that each variable is independent of each other. Generally, if the result is less than 0.05, the sphericity hypothesis is rejected, and there is a correlation between the original variables, which is suitable for factor analysis.

[0115] Combining the scree plot and the principal component table, the eigenvectors corresponding to the first p eigenvalues ​​when the cumulative variance contribution rate is greater than the contribution rate threshold are selected as the principal components. Using SPSS software, the factor score calculation method selects the "regression method" to calculate the factor score of the sample. The contribution rate threshold can be 0.85.

[0116] The traditional K-Means algorithm has obvious shortcomings, so the present invention uses the K-Means++ algorithm to analyze the ship collision avoidance behavior style. The silhouette coefficient method and elbow diagram are combined to select the optimal k value and determine the optimal number of clusters.

[0117] Different from the K-Means algorithm, the K-Means++ algorithm selects the initial cluster center as follows:

[0118] (1) Randomly select a point from the input data point set as the first cluster center;

[0119] (2) For each point x in the data set, calculate the Euclidean distance D(x) between each point and the nearest cluster center;

[0120] (3) Select a new data point as the new cluster center; the selection principle is: the point with larger D(x). The larger D(x), the greater the probability of selecting it as the cluster center.

[0121] (4) Repeat (2) and (3) until k cluster centers are selected;

[0122] (5) Use these k initial cluster centers to run the standard K-Means algorithm;

[0123] Combined with the cluster center and the case summary report of each classification (generated by SPSS software), a quantitative and qualitative analysis of the ship collision avoidance behavior style was conducted, that is, the characteristic indicators were quantitatively calculated, and a qualitative comparative analysis was conducted based on the quantitatively calculated data.

[0124] Embodiment 1:

[0125] A method for identifying ship collision avoidance behavior and style, comprising the following steps:

[0126] S1: Get real AIS data;

[0127] Obtain the real AIS data of the research waters (Laotieshan Waterway) in May 2019;

[0128] S2: Preprocessing the real AIS data of the water area to be identified;

[0129] Including preprocessing steps such as data cleaning and interpolation.

[0130] S3: Based on the navigation parameters of the own ship and the target ship in different groups, the inter-ship motion parameters of the two ships are calculated, and the encounter scenarios of the own ship and the target ship in different groups are screened to form a preliminary encounter scenario data set;

[0131] Based on the navigation parameters of the own ship and the target ship, the inter-ship motion parameters of the two ships are calculated, including the relative distance, relative position, relative heading, relative speed, DCPA, TCPA, and CRI of the target ship relative to the own ship. The definition of the encounter situation between ships in the "Convention on the International Regulations for Preventing Collisions at Sea" is: "A situation when two motor vessels meet on nearly opposite or opposite headings so that there is a risk of collision." The quantitative expression of the present invention is that the heading of the target ship relative to the own ship is between 354° and 360° or 0° and 6°, and the heading difference between the two ships is between 174° and 186°. Specifically, Figure 1 Middle yellow area. The screening conditions for the two-ship encounter sample are:

[0132]

[0133] After screening according to the above conditions, the ship trajectory pairs with potential collision risks in the study waters are identified. Since the AIS data of ships in some scenes are missing, it is judged whether the TCPA at the end point of the scene is less than 0. If it is less than 0, the trajectory segment is considered complete; otherwise, the trajectory segment is considered incomplete due to the lack of AIS data, and the trajectory segment is discarded. Finally, a preliminary encounter scene dataset is formed, totaling 402 encounter scenes.

[0134] S4: Based on the preliminary encounter scenario dataset, the improved sliding window algorithm is used to identify the feature points of the own ship's collision avoidance behavior, characterize the own ship, and construct a collision avoidance trajectory dataset;

[0135] The thresholds are set to 0.2, 0.25, 0.3, 0.35, and 0.4 times the ship width, respectively, and the identified feature points are analyzed. It is found that the number of feature points based on 0.3 times the ship width as the threshold follows the theoretical analysis. Therefore, the threshold in the algorithm of the present invention is set to 0.3 times the ship width.

[0136] In this step, the improved sliding window algorithm is used to traverse the data point set, and the comprehensive threshold of the vertical Euclidean distance and the heading difference is combined to determine whether the current point is a key feature point, and finally the key feature point list is output. Combined with the feature point list, it is determined whether the ship has a collision avoidance feature point. If so, the trajectory segment is retained, otherwise it is discarded. Take the TCPA at the moment of the first collision avoidance feature point. If TCPA is greater than 0, it means that the two ships are still at risk of collision when the collision avoidance action is taken for the first time, and the trajectory segment is retained; otherwise, it is considered that the ship has not taken collision avoidance action due to the target ship, and the trajectory segment is discarded. After screening according to the conditions, a collision avoidance trajectory data set is formed, with a total of 232 valid scenes. Some data are shown in Table 1.

[0137] Table 1. Collision avoidance feature point dataset

[0138] Scene number Number of feature points Feature point number 2044 2 [3,12] 2665 1

[19] 5412 4 [9,10,13,22] 6944 1 [8] 9075 2 [8,33] 58283 3 [17,23,28]

[0139] Trajectory feature point recognition based on different scenarios, such as Figure 4 As shown, (a) encounter scene 2044, (b) encounter scene 58283; the blue is the track of the own ship, and the red is the track of other ships.

[0140] S5: Select characteristic parameters in the collision avoidance trajectory data set as collision avoidance characteristic indicators;

[0141] Combined with the four requirements of "early, large, wide and clear" for ship avoidance proposed by the "International Convention on the Regulations for Preventing Collisions at Sea", the present invention selects the following six eigenvalue indicators for subsequent cluster analysis:

[0142] (1) Heading change range: The heading of the initial trajectory point is recorded as C0, and the heading C of each trajectory point is calculated by traversing. iThe maximum absolute value of the difference between C0 and C0 is the course change amplitude of the ship during the collision avoidance phase. The formula is:

[0143] Δcog=max|C i -C0|

[0144] (2) The first collision avoidance action, that is, the DCPA, TCPA and relative distance between ships at the moment when the first collision avoidance feature point in the collision avoidance phase corresponds to.

[0145] (3) DCPA and relative distance between ships when the relative distance between ships is the smallest.

[0146] S6: Standardizing the characteristic indexes, and reducing the dimension of the collision avoidance characteristic indexes after standardization through factor analysis;

[0147] The six selected characteristic indicators were standardized and unified in dimension. SPSS software was then used to perform factor analysis, dimensionality reduction, and extract principal components to prepare data for subsequent cluster analysis.

[0148] Standardization formula:

[0149] Where: x ij is the jth index of the i-th group of data, is the sample annual average of j evaluation indicators, s j is the sample standard deviation of the jth indicator.

[0150] First, KMO and Bartlett's sphericity test were performed, and the results are shown in Table 2.

[0151] Table 2 KMO and Bartlett test table

[0152]

[0153] As can be seen from the table, the statistic value of Bartlett's sphericity test is 987.015, and the corresponding probability P value is 0. At the significance level α = 0.05, the null hypothesis should be rejected, and it is believed that there is a significant difference between the correlation coefficient matrix and the unit matrix. At the same time, the KMO value is 0.653, so factor analysis can be performed.

[0154] Table 3 Common factor variance table

[0155] initial extract Collision avoidance range 1.000 1.000 first_DCPA 1.000 .980 first_TCPA 1.000 .966 first_Rdistance 1.000 .964 min_DCPA 1.000 .995 min_Rdistance 1.000 .995

[0156] In Table 3, the extraction value indicates how much each variable is expressed by the common factor. The larger the extraction value, the better the variable can be expressed by the common factor. Generally, a value greater than 0.5 means that it can be expressed, and a value greater than 0.7 is sufficient to indicate that the variable can be reasonably expressed by the common factor. As can be seen from the table, the extraction values ​​of all variables are greater than 0.96, and the variables can be reasonably expressed by the common factor.

[0157] Table 4 Total variance explained

[0158]

[0159] Combined with the scree plot and the principal component table, the cumulative variance contribution rate α was selected p The eigenvectors corresponding to the first p eigenvalues ​​greater than 0.85 are taken as principal components. In this embodiment, the first three principal components are selected, as shown in Table 5. The factor scores of the samples are calculated to prepare data for subsequent cluster analysis.

[0160] Table 5 Rotated component matrix

[0161]

[0162] As can be seen from the table, the following explanations are given to each factor in combination with the International Convention for Preventing Collisions at Sea: "first_DCPA", "min_DCPA" and "min_Rdistance" are located on the first factor with a higher load, so the first factor is named "wide and clear". The final result of the action taken to avoid collision should be to pass by and clear other ships at a sufficient safe distance. "first_TCPA" and "first_Rdistance" are located on the second factor with a higher load, so the second factor is named "early". The ship should judge the situation as early as possible, and the actions taken to avoid should leave sufficient room in both time and distance. "Δcog" is located on the third factor with a higher load, so the third factor is named "large". To avoid collision, large-scale actions should be taken as much as possible so that other ships can quickly understand the intention and action of avoidance.

[0163] Finally, the scores of each factor are calculated and the variables are saved as "wide and clear", "early" and "large".

[0164] S7: Based on the collision avoidance feature indicators after dimensionality reduction, the K-Means++ algorithm is used to perform cluster analysis on the collision avoidance feature indicators to obtain the collision avoidance style of ships in the study waters.

[0165] The traditional K-Means algorithm has obvious shortcomings, so the present invention uses the K-Means++ algorithm to analyze the ship collision avoidance behavior style. This example combines the silhouette coefficient method and the elbow diagram to select the best k value and determines that the optimal number of clusters is 4. Figure 6 is the elbow diagram, Figure 7 is the silhouette coefficient plot;

[0166] Combined with the final cluster center and the case summary report analysis of each classification, the ship collision avoidance behavior style is quantitatively and qualitatively analyzed. Combined with the final cluster center and the case summary report analysis of each classification, the ship collision avoidance behavior style is quantitatively and qualitatively analyzed, and the classification result of collision avoidance behavior style is 4 categories. The results are as follows: Figure 8 As shown, (a) is a three-dimensional scatter plot of original data points, and (b) is a three-dimensional scatter plot of clustering results;

[0167] like Fig. 9 As shown, (a) the box plot of the classified data in the factor "large", (b) the box plot of the classified data in the factor "wide and clear", (c) the box plot of the classified data in the factor "early", (d) the box plot of the CRI of the classified data when the first collision avoidance action is taken, and (e) the box plot of the CRI of the classified data when the relative distance between ships is the smallest. The clustering results of the three factors after dimensionality reduction are visualized and analyzed. Combined with the three-dimensional scatter plot of the clustering results and the CRI when the first collision avoidance action is taken in different categories and the CRI at the moment of the minimum relative distance, the analysis is as follows:

[0168] (1) Class 0 “slightly aggressive” has a total of 74 cases, which are characterized by: a high risk level when taking the first collision avoidance action, with an average CRI of 0.75; a small collision avoidance margin, with an average collision avoidance margin of 5.5° and a maximum collision avoidance margin of 15.58°; a late collision avoidance timing, with an average TCPA of 4.6 minutes when taking the first collision avoidance action; passing by other ships at a small distance, with an average minimum relative distance between ships of 0.69 nautical miles;

[0169] (2) Class 1 “normal”, a total of 119 cases, were characterized by: moderate risk when taking the first collision avoidance action, with an average CRI of 0.57; small collision avoidance margin, with an average collision avoidance margin of 7.6° and a maximum collision avoidance margin of 19.08°; early collision avoidance timing, with an average TCPA of 12.5 minutes when taking the first collision avoidance action; passing by other ships at a small distance, with an average minimum relative distance between ships of 0.69 nautical miles;

[0170] (3) Class 2 “conservative prediction type” has a total of 22 cases, which are characterized by: low risk when taking the first collision avoidance action, with an average CRI of 0.33; small collision avoidance amplitude, with an average collision avoidance amplitude of 7.2° and a maximum collision avoidance amplitude of 18.60°; moderate collision avoidance timing, with an average TCPA of 9.2 minutes when taking the first collision avoidance action; passing by other ships at a large safety distance, with an average minimum relative distance between ships of 3.14 nautical miles;

[0171] (4) There were 17 Class 3 “large response types”, which were characterized by: high risk when taking the first collision avoidance action, with an average CRI of 0.67; large collision avoidance amplitude, with an average collision avoidance amplitude of 26.4° and a maximum collision avoidance amplitude of 39.84°; random collision avoidance timing, with TCPA at the first collision avoidance action randomly distributed within 2.1-25.4 minutes; and passing by other ships at a smaller distance, with the average minimum relative distance between ships being 0.92 nautical miles.

[0172] The present invention is based on AIS data, and through the steps of data preprocessing, construction of encounter scenario data set, collision avoidance feature point identification, index selection, cluster analysis and the like, the collision avoidance behavior of ships in a specific water area (Laotieshan water area) is identified and extracted, and a cluster analysis is performed on the collision avoidance behavior style of ships in combination with the collision avoidance timing and actions of the ships.

[0173] 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 cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying ship collision avoidance behavior and style, characterized in that: The following steps are involved: Obtain the real AIS data of the waters to be identified; Pre-process the real AIS data of the waters to be identified; Based on different groups of navigation parameters of own ship and target ship, the inter-ship motion parameters of the two ships are calculated, and different groups of encounter scenarios of own ship and target ship are screened to form a preliminary encounter scenario data set; Based on the preliminary encounter scenario dataset, an improved sliding window algorithm is used to identify the feature points of own ship's collision avoidance behavior and construct a collision avoidance trajectory dataset. Select the characteristic parameters in the collision avoidance trajectory data set as the collision avoidance characteristic index; Clustering algorithm is used to perform cluster analysis on collision avoidance characteristic indicators to obtain the collision avoidance style of ships in the study waters.

2. A method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: It also includes standardizing collision avoidance feature indicators, reducing the dimension of the standardized collision avoidance feature indicators through factor analysis, and using a clustering algorithm to perform cluster analysis on the collision avoidance feature indicators based on the reduced dimension collision avoidance feature indicators.

3. A method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: The process of using the improved sliding window algorithm to identify the collision avoidance feature points of the own ship based on the preliminary encounter scene data set is as follows: The trajectory segment of this ship is composed of N trajectory points. The sliding window is set to a window containing 3 trajectory points (P1, P2, P3). First, the first 3 points of the trajectory segment are taken as the initial sliding window, and a key feature point list key_points[] is created; S1: Calculate the Euclidean perpendicular distance between the midpoint P2 of the window and the line connecting the two end points P1 and P3 of the window. If the Euclidean perpendicular distance is greater than the set threshold, go to S2; otherwise, discard point P2, update the sliding window to (P1, P3, P4), and repeat S1; S2: Create a heading difference search window including the midpoint P2 of the window. The trajectory points in the window include all trajectory points whose interval with P2 does not exceed the time threshold, and then go to S3; S3: Calculate the difference in heading of each adjacent track point in the window, and search to find P i To P i+1 The heading difference is the largest, if P i If it is not in the key feature point key_points[], add P to the list i As the key feature point where the ship starts turning in the current search window, the updated sliding window is (P i ,P i+1 ,P i+2 ), go to S1; otherwise, skip the current sliding window, update the sliding window to (P1, P3, P4), and go to S1; Repeat the above steps until the midpoint of the sliding window is updated to the last point of the trajectory segment.

4. A method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: The formula used for the standardization is as follows: Where: x ij is the jth index of the i-th group of data, is the sample annual average of j evaluation indicators, s j is the sample standard deviation of the jth indicator.

5. The method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: The process of reducing the dimension of the collision avoidance characteristic index after standardization through factor analysis is as follows: First, we need to determine whether the sample data is suitable for factor analysis and perform Kaiser-Meyer-Olkin (KMO) and Bartlett sphericity tests on the sample data; Combining the scree plot and the principal component table, the eigenvectors corresponding to the first p eigenvalues ​​when the cumulative variance contribution rate is greater than the contribution rate threshold are selected as the principal components to calculate the factor score of the sample.

6. A method for identifying ship collision avoidance behavior and style according to claim 2, characterized in that: The process of clustering the collision avoidance characteristic indexes based on the factors after dimension reduction using a clustering algorithm to obtain the collision avoidance style of the ships in the research waters is as follows: The clustering algorithm adopts K-Means++ algorithm; The silhouette coefficient method and elbow diagram are combined to select the best k value and determine the optimal number of clusters; The collision avoidance behavior styles of ships are analyzed quantitatively and qualitatively by combining the cluster centers and case summary reports of each category.

7. A method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: The process of selecting the initial cluster center of the K-Means++ algorithm is as follows: (1) Randomly select a point from the input data point set as the first cluster center; (2) For each point x in the data set, calculate the Euclidean distance D(x) between each point and the nearest cluster center; (3) Select a new data point as the new cluster center; (4) Repeat (2) and (3) until k cluster centers are selected; (5) Use these k initial cluster centers to run the standard K-Means algorithm.

8. A method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: The selection principle is: the point with larger D(x).

9. A method for identifying ship collision avoidance behavior and style according to claim 1, characterized in that: The collision avoidance characteristic indicators include but are not limited to: the amplitude of heading change; the DCPA, TCPA and relative distance between ships at the time corresponding to the first collision avoidance characteristic point in the collision avoidance phase; the DCPA and relative distance between ships when the relative distance between ships is the minimum.

10. A device for identifying ship collision avoidance behavior and style, characterized in that: include Acquisition module: used to obtain the real AIS data of the waters to be identified; Preprocessing module: used to preprocess the real AIS data of the waters to be identified; Encounter scenario data set construction module: used to calculate the inter-ship motion parameters of the two ships based on the navigation parameters of different groups of own ship and target ship, screen different groups of encounter scenarios of own ship and target ship, and form a preliminary encounter scenario data set; Collision avoidance trajectory dataset construction module: used to identify the feature points of own ship's collision avoidance behavior based on the preliminary encounter scenario dataset, and to filter and construct the collision avoidance trajectory dataset by applying the improved sliding window algorithm; Selection module: used to select characteristic parameters in the collision avoidance trajectory data set as collision avoidance characteristic indicators; Dimensionality reduction module: used to standardize the collision avoidance feature indicators and reduce the dimension of the standardized collision avoidance feature indicators through factor analysis; Cluster analysis module: It is used to perform cluster analysis on the collision avoidance feature indicators based on the dimension reduction, and adopt clustering algorithm to obtain the collision avoidance style of ships in the study waters.

Images