Big Data-Based Methods and Systems for Analyzing Vessel Navigation Behavior in Waterways

Abstract

This invention discloses a method and system for analyzing the navigation behavior of ships in waterways based on big data, relating to the field of ship technology. This invention collects dynamic ship data and static waterway data, constructs an environmental semantic grid, and maps the dynamic data to generate semantic trajectory sequences. Based on the sequences, it calculates the basic spatiotemporal correlation value, extracts trajectory direction entropy using a local minimum spanning tree, and constructs a ship behavior feature vector. Based on physical limits, it constructs a dynamic pseudo-label dataset and uses an evolutionary algorithm based on a weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function. Using the optimal parameter set, it constructs a decision function to identify abnormal ship behavior, calculates risk potential energy, and generates chain reaction warnings based on the risk transmission coefficient. This invention effectively integrates environmental semantics and entropy features, solving the problems of scarce abnormal samples and complex nonlinear feature identification without manual annotation, and achieving proactive and precise prevention and control of waterway collision risks.

Description

Technical Field

[0001] This invention relates to the field of ship technology, and in particular to a method and system for analyzing ship navigation behavior in waterways based on big data. Background Technology

[0002] With the rapid development of inland waterway and coastal shipping, the density of ship traffic is increasing and the waterway environment is becoming more complex. In order to ensure maritime traffic safety, sensing devices such as Automatic Identification Systems (AIS) and video surveillance systems are widely used, accumulating massive amounts of ship navigation dynamic data. How to utilize these multi-source heterogeneous big data resources to achieve intelligent analysis and risk warning of ship navigation behavior has become a key link in the construction of smart shipping. By mining ship trajectory data, abnormal ship behavior can be detected in a timely manner, which is of great practical significance for preventing collision accidents and improving the efficiency of maritime supervision.

[0003] Chinese invention patent application CN116740137A discloses a method, device, equipment, and medium for identifying and warning of dangerous behaviors in ship navigation. This method mainly extracts the first ship motion feature parameters based on similar and different motion parameters of ships, uses principal component analysis algorithm to reduce the dimensionality of the feature parameters, and inputs them into a multi-algorithm fusion model constructed by BP neural network, support vector machine, and random forest for learning and classification, thereby realizing the identification of normal and dangerous behaviors of ships. Existing methods mostly focus on the statistical analysis of the ship's own kinematic parameters, often ignoring the constraints of the static environment of the waterway on ship behavior, making it difficult to combine specific geographical environment semantics to differentiate the compliance of ship behavior. Existing technologies mainly rely on linear or basic statistical features such as mean and variance for feature extraction. For complex abnormal behaviors with high nonlinearity and disorder, such as ship drifting out of control and spinning in place, there is a lack of deep feature descriptions that can effectively characterize the degree of trajectory disorder. Existing warning mechanisms mostly focus on the abnormal state of a single ship and lack consideration for the risk transmission mechanism in multi-ship interaction scenarios, making it difficult to effectively quantify the potential collision threat posed by abnormal ships to surrounding normal ships. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies are difficult to differentiate the compliance of ship behavior by combining specific geographical environment semantics. For complex abnormal behaviors with high nonlinearity and disorder, such as ship drifting out of control and spinning in place, there is a lack of deep feature descriptions that can effectively characterize the degree of trajectory disorder, and it is difficult to effectively quantify the potential collision threat posed by abnormal ships to surrounding normal ships.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a method for analyzing the navigation behavior of vessels in waterways based on big data, comprising the following steps:

[0006] Step S1: Collect dynamic data of ships and static data of inland waterways, construct an environmental semantic grid based on the static data of the waterways, and map the dynamic data of ships to the environmental semantic grid to generate a semantic trajectory sequence;

[0007] Step S2: Based on the semantic trajectory sequence, calculate the basic spatiotemporal correlation value and trajectory direction entropy of the trajectory segment, and construct the ship behavior feature vector;

[0008] Step S3: Based on the distribution characteristics of the ship behavior feature vector in the feature plane, construct a dynamic pseudo-label dataset, construct a nonlinear classification decision function, and use an evolutionary algorithm based on the weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function in combination with the dynamic pseudo-label dataset to obtain the optimal decision parameter set.

[0009] Step S4: Construct the final nonlinear classification decision function using the optimal decision parameter set to determine the navigation behavior of ships in the water area at the current moment and generate an early warning.

[0010] Preferably, step S1 includes the following sub-steps:

[0011] The ship dynamic data includes the ship's MMSI, latitude and longitude coordinates, speed over land, and heading over land;

[0012] The static data of the waterway includes functional areas on the electronic chart and the maximum speed limit stipulated in the waterway.

[0013] The process of constructing the environmental semantic grid includes: dividing the waterway into discrete grid units, defining the functional attributes of each grid unit according to the electronic nautical chart, and assigning a corresponding environmental weight value to each grid unit based on the functional attributes;

[0014] The ship dynamic data within the preset sliding window is mapped to the environmental semantic grid, and the semantic trajectory sequence of the ship is constructed by combining the environmental weight values.

[0015] Preferably, in step S2, the process of constructing the ship behavior feature vector specifically includes:

[0016] The ship behavior feature vector in the semantic trajectory sequence is used as the trajectory point. The trajectory points in the semantic trajectory sequence are traversed. Two trajectory points are arbitrarily selected to form a trajectory point pair. The latitude and longitude coordinates of the trajectory point pair are extracted, converted into spatial projection coordinates, and the Euclidean distance and time difference between the two points are calculated. Based on the Euclidean distance and time difference, the basic spatiotemporal correlation value of each pair of trajectory points is calculated.

[0017] Construct a local minimum spanning tree of trajectory points within a sliding window, statistically analyze the distribution probability of the direction angles of each connecting line in the local minimum spanning tree, and calculate the trajectory direction entropy using the Shannon entropy formula.

[0018] Preferably, in step S3, the process of constructing the dynamic pseudo-label dataset specifically includes:

[0019] Map the mean of the basic spatiotemporal correlation and the trajectory direction entropy of each ship in the current sliding window to sample points in a two-dimensional feature plane;

[0020] Preset absolute safe poles and absolute outlier poles in a two-dimensional feature plane;

[0021] In a two-dimensional feature plane, calculate the first distance from each sample point to the absolute safe pole and the second distance to the absolute abnormal pole;

[0022] Based on the ratio of the difference between the first distance and the second distance, generate the corresponding pseudo-label of the ship and the corresponding pseudo-label confidence level;

[0023] The process of generating the pseudo-tags includes:

[0024] If the first distance is less than the second distance, the navigation status of the corresponding vessel will be marked as a suspected normal label;

[0025] If the first distance is greater than or equal to the second distance, the navigation status of the corresponding vessel will be marked as a suspected anomaly.

[0026] The process of generating the pseudo-label confidence score includes:

[0027] The ratio of the absolute value of the difference between the first distance and the second distance to the sum of the first distance and the second distance is used as the pseudo-label confidence score.

[0028] If the confidence level of the pseudo-tag of the corresponding vessel is greater than the preset execution threshold, then the pseudo-tag of the vessel's navigation status will be recorded as a high-confidence pseudo-tag.

[0029] If a ship's navigation status is marked as a suspected normal label and the pseudo label is a high-confidence pseudo label, then the sample points corresponding to the ship constitute a dynamic pseudo label dataset.

[0030] Preferably, in step S3, the mathematical expression of the nonlinear classification decision function is:

[0031] ;

[0032] in, It is a vector composed of the basic spatiotemporal correlation value and the trajectory direction entropy. The number of radial basis functions. For the first The kernel center location of each radial basis function. For the first The width coefficients of the radial basis functions, For the first The weighting coefficients of the radial basis functions, This is the bias threshold.

[0033] Preferably, step S3, the process of obtaining the optimal decision parameter set, specifically includes:

[0034] The kernel center position, width coefficient, weight coefficient, and bias threshold of the nonlinear classification decision function are used as the individuals in the population to be optimized.

[0035] Multiple population individuals are randomly generated within a preset value range, and multiple classification decision functions are obtained using the randomly generated population individuals;

[0036] The classification decision function is used to predict the dynamic pseudo-label dataset to obtain the pseudo-label prediction results. Based on the comparison between the pseudo-label prediction results of the decision function and the pseudo-labels of the sample points, a weighted statistic is calculated in combination with the pseudo-label confidence of the sample points. The weighted statistic includes weighted true positives and weighted false positives.

[0037] The weighted true instance is as follows: for all sample points whose pseudo-labels are suspected to be abnormal and whose prediction results are abnormal, the confidence weights of the pseudo-labels corresponding to the sample points are accumulated.

[0038] The weighted false positives are as follows: for all sample points whose pseudo-labels are suspected to be normal but whose prediction results are abnormal, the confidence weights of the pseudo-labels corresponding to the sample points are accumulated.

[0039] The weighted true positive rate is the ratio of the weighted true positive rate to the sum of the confidence weights of all pseudo-labeled sample points that are suspected of being abnormal, and the weighted false positive rate is the ratio of the weighted false positive rate to the sum of the confidence weights of all pseudo-labeled sample points that are suspected of being normal.

[0040] Preferably, in step S3, the process of obtaining the optimal decision parameter set further includes:

[0041] The weighted true positive rate and weighted false positive rate of each individual in the population are used as the coordinate points of the individual in the ROC space;

[0042] The convex hull algorithm is used to identify the set of dominant individuals located at the vertices of the ROC convex hull in the ROC space. Based on the coordinates of the individuals in the set of dominant individuals, the area of ​​the convex hull is calculated using the trapezoidal quadrature formula.

[0043] Based on the coordinates of the elements in the dominant individual set in the ROC space, a virtual convex hull curve is constructed using the tangent projection method, which encloses the current population and expands towards the ideal point.

[0044] The individuals remaining in the current population, excluding the dominant individuals set, are non-convex hull individuals. For any non-convex hull individual, calculate the Euclidean distance from the non-convex hull individual to the virtual convex hull curve, sort them in ascending order, and select the top few non-convex hull individuals from the sorting results.

[0045] Construct a preferred parent population from several non-convex hull individuals and individuals from the set of dominant individuals. Perform crossover and mutation operations on the preferred parent population, and repeat the iterative optimization until the preset stopping iteration condition is reached, and output the final set of dominant individuals.

[0046] Iterate through each individual in the final set of dominant individuals, calculate the Euclidean distance from the individual to the ideal point, select the individual with the smallest Euclidean distance, and extract the kernel center position, width coefficient, weight coefficient, and bias threshold as the optimal decision parameter set.

[0047] Preferably, in step S4, the process of generating the early warning specifically includes:

[0048] Using the optimal decision parameter set, a nonlinear classification decision function is constructed to calculate the state feature vector composed of the basic spatiotemporal correlation value and trajectory direction entropy of all ships in the water area at the current moment;

[0049] If the value of the nonlinear classification decision function is greater than 0, the ship's navigation behavior is determined to be abnormal.

[0050] If the value of the nonlinear classification decision function is less than or equal to 0, the ship's navigation behavior is determined to be normal.

[0051] Map each ship in the current waters to a graph node, and assign attribute values ​​to the graph nodes according to the classification decision function.

[0052] If a ship's navigation behavior is abnormal, the ship is marked as a risk source node, and the ship's trajectory direction entropy is used as the node's initial risk potential energy.

[0053] If the ship's navigation behavior is normal, mark the ship as the victim target node and set the initial risk potential energy of the node corresponding to the ship to zero.

[0054] Preferably, in step S4, the process of generating the early warning further includes:

[0055] For each victim node, with the victim node as the center and the preset safety warning distance as the radius, search for all risk source nodes within the radius.

[0056] Calculate the physical distance and relative approach speed between the victim target node and each risk source node found in the search;

[0057] The relative approach speed is calculated as follows: obtain the ground velocity vector of the risk source node and the ground velocity vector of the victim target node, calculate the residual between the ground velocity vector of the risk source node and the ground velocity vector of the victim target node to obtain the relative velocity vector, calculate the unit direction vector of the victim target node pointing to the risk source node, calculate the projection magnitude of the relative velocity vector on the unit direction vector, and use the projection magnitude as the relative approach speed.

[0058] Calculate the risk transmission coefficient from each risk source node to the affected target node based on physical distance and relative approach speed;

[0059] The risk transmission coefficient is calculated as follows: the square of the physical distance between the risk source node and the victim target node is calculated, and the square is added to a preset distance correction factor to obtain a distance attenuation factor; the ratio of the relative approach speed to the distance attenuation factor is calculated, and the ratio is determined as the risk transmission coefficient.

[0060] The initial risk potential of all risk source nodes within the search range is weighted by the risk transmission coefficient and then summed to obtain the total external risk potential of the victim target node.

[0061] If the sum of normalized external risk potential energy exceeds the preset risk threshold, an emergency collision warning will be generated.

[0062] Preferably, it includes a data acquisition module, a feature construction module, an iteration module, and an analysis module;

[0063] The acquisition module is used to collect dynamic data of ships and static data of inland waterways, construct an environmental semantic grid based on the static data of the waterways, and map the dynamic data of ships to the environmental semantic grid to generate a semantic trajectory sequence.

[0064] The feature construction module is used to calculate the basic spatiotemporal correlation value and trajectory direction entropy of the trajectory segment based on the semantic trajectory sequence, and construct the ship behavior feature vector;

[0065] The iterative module is used to construct a dynamic pseudo-label dataset based on the distribution characteristics of the ship behavior feature vector in the feature plane, construct a nonlinear classification decision function, and use an evolutionary algorithm based on the weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function in combination with the dynamic pseudo-label dataset to obtain the optimal decision parameter set.

[0066] The analysis module is used to construct the final nonlinear classification decision function using the optimal decision parameter set, to determine the navigation behavior of ships in the water area at the current moment, and to generate an early warning.

[0067] The beneficial effects of this invention are as follows: By constructing an environmental semantic grid and generating semantic trajectory sequences, this invention can deeply integrate discrete ship dynamic data with the static environment of the waterway. This invention utilizes the characteristic that information entropy can effectively measure the disorder of the system to accurately capture abnormal behaviors with highly nonlinear characteristics, such as ship drifting out of control and spinning in place. This invention proposes an evolutionary optimization strategy based on dynamic pseudo-label datasets and weighted ROC convex hull guidance. By using weighted ROC convex hull guidance to iterate the parameters of the decision function, this invention can automatically focus on high-confidence samples while taking into account the division of fuzzy boundaries, thereby obtaining the optimal nonlinear classification boundary under the condition of sample imbalance, which greatly improves the robustness of behavior determination. Attached Figure Description

[0068] Figure 1 A flowchart illustrating the steps of a big data-based method for analyzing the navigation behavior of ships in waterways, as provided in one embodiment of the present invention. Detailed Implementation

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0070] Example 1, referring to Figure 1 This paper provides a method for analyzing the navigation behavior of ships in waterways based on big data, including the following steps:

[0071] Step S1: Collect dynamic data of ships and static data of inland waterways, construct an environmental semantic grid based on the static data of the waterways, and map the dynamic data of ships to the environmental semantic grid to generate a semantic trajectory sequence;

[0072] Step S2: Based on the semantic trajectory sequence, calculate the basic spatiotemporal correlation value and trajectory direction entropy of the trajectory segment, and construct the ship behavior feature vector;

[0073] Step S3: Based on the distribution characteristics of ship behavior feature vectors in the feature plane, construct a dynamic pseudo-label dataset, construct a nonlinear classification decision function, and use an evolutionary algorithm based on the weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function in combination with the dynamic pseudo-label dataset to obtain the optimal decision parameter set.

[0074] Step S4: Construct the final nonlinear classification decision function using the optimal decision parameter set to determine the navigation behavior of ships in the water at the current moment and generate an early warning.

[0075] This invention deeply integrates the semantics of the static environment of the waterway with the dynamic trajectory data of ships and introduces entropy features that can characterize the degree of disorder in the system. This effectively overcomes the problem that traditional methods are not capable of identifying complex nonlinear abnormal behaviors (such as uncontrolled drift). By using iterative optimization of the nonlinear decision function based on evolutionary algorithm, it can still maintain efficient and accurate judgment of abnormal behavior of ships in waterways in cold start scenarios where there are no historical accident labels.

[0076] Step S1 includes the following sub-steps:

[0077] Ship dynamic data includes the ship's MMSI, latitude and longitude coordinates, speed over land, and heading over land;

[0078] Static channel data includes functional areas on electronic charts and the maximum speed limits specified in the waterways.

[0079] The process of constructing an environmental semantic grid includes: dividing the waterway into discrete grid cells, defining the functional attributes of each grid cell according to the electronic nautical chart, and assigning a corresponding environmental weight value to each grid cell based on the functional attributes;

[0080] Ship dynamic data within a preset sliding window is mapped to an environmental semantic grid, and a semantic trajectory sequence of the ship is constructed by combining environmental weight values.

[0081] In one specific embodiment of the present invention, the static channel data is derived from electronic nautical charts. Electronic nautical chart data is vector data, which includes the polygonal boundary coordinates and functional attributes of geographic entities (coastlines, main channels, anchorages, warning zones, and shoals).

[0082] Ship dynamic data is received in real time from AIS base stations deployed along the coast, including MMSI (Unique Ship Identifier), UTC timestamp, longitude, latitude, speed to ground, and heading to ground.

[0083] Based on the geographic coordinate system of the waterway, a raster index is established, dividing the entire monitored water area into uniformly sized discrete rectangular grid cells (in this embodiment...). For a grid cell that covers multiple electronic chart functional areas (such as a grid cell that is located at the edge of the main channel and also involves the bridge area), the grid cell is marked as the functional area with the highest priority by setting the priority order of the functional areas.

[0084] The priority order for determining functional areas is: restricted area > bridge area waters > main channel > warning area > anchorage (including berthing area and service area). If a grid cell contains multiple overlapping functional areas, the grid cell will be forcibly marked as the functional area with the highest priority.

[0085] A preset environmental weight mapping table is used to convert functional attributes into numerical environmental weights. The environmental weight value represents the sensitivity of the area to abnormal ship behavior.

[0086] The environmental weight for the main channel, bridge area waters, and restricted navigation area is 1.0; the environmental weight for the warning area and confluence area is 0.6; and the environmental weight for the anchorage, berthing area, and service area is 0.2.

[0087] After gridding is completed, the ship dynamic data is mapped to the corresponding grid cells. Using the spatial indexing algorithm (R-Tree), the grid cell in which the ship dynamic data falls is quickly retrieved based on latitude and longitude. The corresponding environmental weight is read from the attributes of that grid cell and concatenated as a new feature dimension into the original ship dynamic data to form a ship behavior feature vector. ;

[0088] For the same vessel (same MMSI), its continuously generated vessel behavior feature vectors Arranged according to the chronological order of UTC timestamps to form a semantic trajectory sequence;

[0089] The expression for the semantic trajectory sequence is:

[0090] ;

[0091] in, These represent the 1st, 2nd, and 3rd semantic trajectory sequences, respectively. The feature vector of ship behavior at each moment, for The feature vector at time step, The expression is:

[0092] ;

[0093] in, for The dimension of time, for Longitude of time for The speed at ground level at any given moment. for The constant heading towards the ground. for The environmental weight value corresponding to the grid cell where the ship is located at any given time;

[0094] n is the number of sampling points within the sliding window. In this embodiment, the number of sampling points is 30, covering a ship's voyage of 5 to 10 minutes.

[0095] This invention constructs an environmental semantic grid and maps ship dynamic data to this grid. By assigning differentiated environmental weights to different functional areas (such as the main channel, bridge area, and restricted navigation area), it can dynamically adjust the sensitivity to abnormal behavior according to the specific water environment in which the ship is located. This data processing method, which combines prior knowledge of electronic charts, effectively filters out misjudgments caused by environmental factors (for example, low speed in the anchorage is normal, but low speed in the main channel may be abnormal), and greatly improves the amount of information contained in the feature vector and the scene adaptability.

[0096] Step S2, the process of constructing the ship behavior feature vector specifically includes:

[0097] The ship behavior feature vector in the semantic trajectory sequence is used as the trajectory point. The trajectory points in the semantic trajectory sequence are traversed. Two trajectory points are arbitrarily selected to form a trajectory point pair. The latitude and longitude coordinates of the trajectory point pair are extracted, converted into spatial projection coordinates, and the Euclidean distance and time difference between the two points are calculated. Based on the Euclidean distance and time difference, the basic spatiotemporal correlation value of each pair of trajectory points is calculated.

[0098] Construct a local minimum spanning tree of trajectory points within a sliding window, statistically analyze the distribution probability of the direction angles of each connecting line in the local minimum spanning tree, and calculate the trajectory direction entropy using the Shannon entropy formula.

[0099] In one specific embodiment of the present invention, any two points are selected within the sliding window. and (in These points form trajectory point pairs, which are used to capture the positional constraints between any two points within the window and prevent the omission of abnormal trajectories.

[0100] When calculating Euclidean distance, only extract and The latitude and longitude are calculated, but since latitude and longitude are spherical coordinates, direct calculation will have errors. (latitude), (Longitude) is converted to coordinates in a Cartesian coordinate system using Mercator projection. The Euclidean distance between trajectory point pairs is calculated using the final coordinates after Mercator projection transformation.

[0101] The mathematical expression for the basic spatiotemporal correlation value is:

[0102] ;

[0103] in, For trajectory points and The basic spatiotemporal correlation value, The distance sensitivity coefficient is 0.01 in this embodiment. For trajectory points and Euclidean distance, For trajectory points and Time difference, To prevent the denominator from being 0, this embodiment uses a preset decimal number. It is 0.0001;

[0104] Represents the average displacement velocity, if the Euclidean distance between two points is... Very small (position unchanged), but time difference It is very large (after a long time), then Approaching 0, the final basic spatiotemporal correlation value is very small, which means that there is a strong spatiotemporal correlation between the two points, that is, the ship is stuck in the area during this time period.

[0105] Within the sliding window, utilize the spatial projection coordinates of all trajectory points. Constructing a local minimum spanning tree can capture the topological structure of a point set in space. The local minimum spanning tree algorithm calculates the Euclidean distance between all trajectory points and connects all trajectory points with the minimum total path length.

[0106] When a ship is sailing normally, it sails in sequence according to the track points, which are arranged in a line and all the connecting lines point in the same direction. When a ship is sailing abnormally, drifting out of control, or spinning in place, it sways randomly in place and all the connecting lines point in a random direction.

[0107] Extract the coordinates of two trajectory points connected by a line segment in the local minimum spanning tree. Obtain the direction vector of the connecting line based on the coordinates, and calculate the azimuth angle of the connecting line relative to true north. The range of the azimuth angle is... ;

[0108] Histogram statistics were used to divide the 360 ​​degrees into... Each interval (in this embodiment) (i.e., every 20 degrees is an interval), count the number of lines falling into each interval, and calculate the angular distribution probability. The trajectory direction entropy is calculated using the Shannon entropy formula;

[0109] If the ship's sailing direction is straight, the direction of the connecting lines is concentrated, the angle distribution probability is concentrated in a few intervals, and the trajectory direction entropy value is low. If the ship's sailing direction is disordered drifting and the behavior is abnormal, the direction of the connecting lines is disordered, the angle distribution probability is disordered, and the trajectory direction entropy is high.

[0110] Based on the trajectory direction entropy and basic spatiotemporal correlation value of the local minimum spanning tree, a vector that can deeply characterize the micro-motion characteristics of ships is constructed. By using the statistical distribution probability of the connection direction of the local minimum spanning tree and calculating the Shannon entropy, it can keenly capture the nonlinear motion characteristics of ships when they are out of control, spinning, or drifting disorderly. Compared with traditional linear statistical indicators such as speed and heading, trajectory direction entropy can quantify the degree of trajectory disorder, thereby accurately identifying those hidden dangers with normal speed but extremely unstable heading, significantly enhancing the monitoring capability of complex abnormal behavior.

[0111] Step S3, the process of constructing the dynamic pseudo-label dataset specifically includes:

[0112] Map the mean of the basic spatiotemporal correlation and the trajectory direction entropy of each ship in the current sliding window to sample points in a two-dimensional feature plane;

[0113] Preset absolute safe poles and absolute outlier poles in a two-dimensional feature plane;

[0114] In a two-dimensional feature plane, calculate the first distance from each sample point to the absolutely safe pole and the second distance to the absolutely outlier pole;

[0115] Based on the ratio of the difference between the first distance and the second distance, generate the corresponding pseudo-label of the ship and the corresponding pseudo-label confidence level;

[0116] The process of generating pseudo-tags includes:

[0117] If the first distance is less than the second distance, the navigation status of the corresponding vessel will be marked as a suspected normal label;

[0118] If the first distance is greater than or equal to the second distance, the navigation status of the corresponding vessel will be marked as a suspected anomaly.

[0119] The process of generating pseudo-label confidence scores includes:

[0120] The ratio of the absolute value of the difference between the first distance and the second distance to the sum of the first distance and the second distance is used as the pseudo-label confidence score.

[0121] If the confidence level of the pseudo-tag of the corresponding vessel is greater than the preset execution threshold, then the pseudo-tag of the vessel's navigation status will be recorded as a high-confidence pseudo-tag.

[0122] If a ship's navigation status is marked as a suspected normal label and the pseudo label is a high-confidence pseudo label, then the sample points corresponding to the ship constitute a dynamic pseudo label dataset.

[0123] In a specific embodiment of the present invention, a two-dimensional rectangular coordinate system is constructed as a two-dimensional feature plane with the basic spatiotemporal correlation value as the horizontal axis and the trajectory direction entropy as the vertical axis.

[0124] Within the current detection area, the mean of basic spatiotemporal correlation and trajectory direction entropy of each vessel within the current sliding window are extracted. The mean of basic spatiotemporal correlation and trajectory direction entropy of each vessel are mapped to coordinate points on a two-dimensional feature plane, and the coordinate points of each vessel are one sample point.

[0125] Pre-defined safety poles within the two-dimensional feature plane. (Representing absolute stillness and order) and outlier poles (Represents the maximum displacement and the most chaotic situation);

[0126] The process of obtaining abnormal poles is as follows:

[0127] Basic spatiotemporal correlation value The core of the formula is proportional to the ratio of displacement to time, essentially reflecting the generalized velocity of the ship. This is crucial for determining abnormal poles. ;

[0128] Extract the maximum physical speed that ships are allowed to navigate in the current waters. The set sliding window time length is obtained (5-10 minutes in this embodiment). If the ship... When moving in a straight line, the distance between any two points is... With time difference The ratio approaches ,Will Substituting the value into the formula for calculating the basic spatiotemporal correlation value in step S2, the obtained value is... ;

[0129] ;

[0130] in, The physical limit value of the basic spatiotemporal correlation value, The distance sensitivity coefficient, The length of the sliding window. The maximum speed limit stipulated for waterways;

[0131] In step S2, the circumferential direction (360 degrees) is divided into... discrete intervals ( The directional entropy reaches its maximum when the ship's motion direction is completely uniformly distributed across all intervals (i.e., exhibiting isotropic Brownian motion, completely disordered). ;

[0132] For any sample point within the current window Calculate its to Euclidean distance and to Euclidean distance ;

[0133] like The corresponding vessel's navigation status is marked as "suspected normal," and the label is applied. ;

[0134] like The navigation status of the corresponding vessel is marked as a suspected anomaly. ;

[0135] Calculate the confidence level of pseudo-labels The mathematical expression for the confidence level of pseudo-labels is:

[0136] ;

[0137] in, For false label confidence, The first distance, This is the second distance;

[0138] Based on the above calculations, pseudo-labels were generated for each ship within the current sliding window. and the corresponding pseudo-label confidence weights ;

[0139] If the confidence level of the ship's status is greater than the preset confidence threshold (0.6~0.8 in this embodiment), then the pseudo-label of the ship's navigation status is recorded as a high confidence pseudo-label.

[0140] when When the value approaches 1, it indicates that the ship's motion state is in an extreme region of physical characteristics (extremely ordered stillness or extremely disordered motion), and its corresponding pseudo-label... These samples have extremely high confidence and are therefore marked as high-confidence samples.

[0141] when When the value approaches 0, it indicates that the ship is in a gray area between normal and abnormal, and its false label... The reference value is low, and we await the optimal decision function in subsequent steps for final classification.

[0142] By presetting absolute safe poles and absolute outlier poles, the distance ratio of sample points in the feature plane is automatically calculated, thereby generating pseudo-labels with confidence for unlabeled data. By using the physical boundary of ship motion as a weak supervision signal, it not only solves the problem of extreme scarcity of negative samples in practical applications, but also ensures the high purity and reliability of subsequent model training samples through the confidence screening mechanism.

[0143] In step S3, the mathematical expression of the nonlinear classification decision function is:

[0144] ;

[0145] in, It is a vector composed of the basic spatiotemporal correlation value and the trajectory direction entropy. The number of radial basis functions. For the first The kernel center location of each radial basis function. For the first The width coefficients of the radial basis functions, For the first The weighting coefficients of the radial basis functions, This is the bias threshold.

[0146] In a specific embodiment of the present invention, considering that the distribution of abnormal ship behavior in the spatiotemporal correlation and directional entropy feature space has nonlinear characteristics, a radial basis function is used as a decision function to determine the ship's navigation status.

[0147] Decision function The mathematical expression for it is defined as:

[0148] ;

[0149] in, The input feature vector, in this embodiment, is a vector composed of the basic spatiotemporal correlation value and the trajectory direction entropy. , The number of radial basis functions is preset in this embodiment. It is a fixed integer (in this embodiment) ), representing the existence in this feature space A potential normal state; For the first The kernel center location of each radial basis function. It is a two-dimensional vector , representing the first The coordinate center of the typical behavior pattern in the feature space;

[0150] For the first The width coefficient (or broadening parameter) of each radial basis function controls the radius of influence of that radial basis function. For the first The weighting coefficients of the radial basis functions, This is the bias threshold used to adjust the overall offset of the decision boundary;

[0151] The nonlinear classification decision function is used to determine the navigation status of a ship. The decision logic is as follows:

[0152] like The ship's navigation status is determined to be abnormal (corresponding to pseudo-label 1). The ship's navigation status is determined to be normal (corresponding to pseudo-label 0).

[0153] This invention employs radial basis functions to construct a nonlinear classification decision function, utilizing its excellent nonlinear mapping capability to fit the complex boundaries of normal and abnormal ship behaviors in the feature space. Since ship behavior characteristics typically exhibit a nonlinear distribution, simple linear classification struggles to distinguish samples in ambiguous zones. However, radial basis functions, through the combination of kernel center, width, and weights, can flexibly approximate the decision surface. This allows the decision function to more precisely classify safe and dangerous behaviors, effectively reducing the false alarm and false negative rates of samples in critical states.

[0154] Step S3, the process of obtaining the optimal decision parameter set specifically includes:

[0155] The kernel center position, width coefficient, weight coefficient, and bias threshold of the nonlinear classification decision function are used as the individuals in the population to be optimized.

[0156] Multiple population individuals are randomly generated within a preset value range, and multiple classification decision functions are obtained using the randomly generated population individuals;

[0157] The classification decision function is used to predict the dynamic pseudo-label dataset and obtain the pseudo-label prediction results. Based on the comparison between the pseudo-label prediction results of the decision function and the pseudo-labels of the sample points, the weighted statistics are calculated in combination with the pseudo-label confidence of the sample points. The weighted statistics include weighted true positives and weighted false positives.

[0158] The weighted true example is: for all sample points whose pseudo-labels are suspected to be abnormal and whose prediction results are abnormal, the confidence weights of the pseudo-labels corresponding to the sample points are accumulated.

[0159] The weighted false positives are as follows: for all sample points whose pseudo-labels are suspected to be normal but whose prediction results are abnormal, the confidence weights of the pseudo-labels corresponding to the sample points are accumulated.

[0160] The weighted true positive rate is the ratio of the weighted true positive rate to the sum of the confidence weights of all pseudo-labeled sample points that are suspected of being abnormal, and the weighted false positive rate is the ratio of the weighted false positive rate to the sum of the confidence weights of all pseudo-labeled sample points that are suspected of being normal.

[0161] In a specific embodiment of the present invention, since the decision function contains an unknown parameter kernel center position Width coefficient Weighting coefficients Bias threshold An evolutionary algorithm is used to iteratively optimize the unknown parameters in the decision function;

[0162] The unknown parameters are encoded as one-dimensional real number vectors, which represent the individuals in the population. The initial population is randomly generated using real number encoding, with an initial population size of 150. Two-dimensional coordinates The initial range of the horizontal axis is set to , The physical limit value of the basic spatiotemporal correlation value is used, and the initial range of the ordinate is set to... , The maximum directional entropy value;

[0163] Width coefficient The initialization range is set to Weighting coefficient The initialization range is set to Using a symmetrical interval, the decision function is allowed to either positively incentivize (classify as abnormal) or negatively inhibit (classify as normal) the final result, with a bias threshold. The initialization range is set to ;

[0164] The number of radial basis functions in this embodiment Then the coding structure of each individual in the population is:

[0165] ;

[0166] because Two-dimensional coordinates Therefore, an individual actually contains the following gene loci (parameters) to be optimized:

[0167] ;

[0168] In each iteration of the evolutionary algorithm, for each individual in the population (i.e., a set of determined decision function parameters), a dynamic pseudo-label dataset is used for testing. In order to eliminate the interference of low-confidence samples located in the intermediate fuzzy region on the optimization of the decision function, this embodiment abandons the traditional sample counting method and instead adopts a weighted statistical method based on confidence weight for performance evaluation.

[0169] Weighted true positives reflect the decision function's ability to capture high-certainty risks, while weighted false positives reflect the decision function's degree of false alarms for obviously normal vessels.

[0170] The weighted true positive rate represents the proportion of absolute abnormal targets detected by the decision function after eliminating the interference of fuzzy samples, while the weighted false positive rate represents the proportion of false positives detected by the decision function on absolutely safe samples.

[0171] This invention introduces a weighted statistical evaluation mechanism (weighted true positive rate and weighted false positive rate) based on pseudo-label confidence during parameter optimization. This evaluation method fully utilizes the uncertainty information in pseudo-labels, enabling the optimization algorithm to automatically focus on high-quality samples with obvious physical characteristics and high confidence during iteration, while reducing the interference of low-confidence samples in the ambiguous zone on the model's convergence direction. This gives the algorithm extremely strong noise resistance, ensuring that the final decision model is trained based on the most reliable data features, thus improving the model's robustness.

[0172] Step S3, the process of obtaining the optimal decision parameter set also includes:

[0173] The weighted true positive rate and weighted false positive rate of each individual in the population are used as the coordinate points of the individual in the ROC space;

[0174] The convex hull algorithm is used to identify the set of dominant individuals located at the vertices of the ROC convex hull in the ROC space. Based on the coordinates of the individuals in the set of dominant individuals, the area of ​​the convex hull is calculated using the trapezoidal quadrature formula.

[0175] Based on the coordinates of the elements in the dominant individual set in the ROC space, a virtual convex hull curve is constructed using the tangent projection method, which encloses the current population and expands towards the ideal point.

[0176] The individuals remaining in the current population, excluding the dominant individuals set, are non-convex hull individuals. For any non-convex hull individual, calculate the Euclidean distance from the non-convex hull individual to the virtual convex hull curve, sort them in ascending order, and select the top few non-convex hull individuals from the sorting results.

[0177] Construct a preferred parent population from several non-convex hull individuals and individuals from the set of dominant individuals. Perform crossover and mutation operations on the preferred parent population, and repeat the iterative optimization until the preset stopping iteration condition is reached, and output the final set of dominant individuals.

[0178] Iterate through each individual in the final set of dominant individuals, calculate the Euclidean distance from the individual to the ideal point, select the individual with the smallest Euclidean distance, and extract the kernel center position, width coefficient, weight coefficient, and bias threshold as the optimal decision parameter set.

[0179] In a specific embodiment of the present invention, based on the aforementioned weighted statistics, the coordinates of individuals in the ROC space are calculated to achieve a noise-resistant evaluation of the decision function performance. The coordinates in the ROC space are composed of the weighted true positive rate and the weighted false positive rate. The coordinates of all individuals in the current population are mapped to the ROC plane, and the set of dominant individuals currently at the vertices of the ROC convex hull is identified using the convex hull algorithm. The specific process includes:

[0180] Sort the coordinates of all individuals in the population on the ROC plane in ascending order of weighted false positive rate.

[0181] Establish a candidate set, iterate through the sorted coordinate points, and use the vector cross product method to determine the direction of the line connecting the current point and the last two points in the candidate set.

[0182] If the newly added coordinate point causes the line to show an inward concave trend (i.e., the new point is located to the lower right of the line connecting the first two points), it indicates that the coordinate point is a poor individual and it is removed from the candidate set. Repeat the above removal process until the newly added coordinate point causes the line to maintain an outward convex trend. Add the new coordinate point to the candidate set. After the traversal is completed, all the points retained in the candidate set are the set of dominant individuals that constitute the vertices of the ROC convex hull.

[0183] The area of ​​a convex hull is calculated as follows:

[0184] To quantify the overall classification performance of the current population and monitor the convergence state of the algorithm, after identifying the set of dominant individuals that constitute the true ROC convex hull, the effective coverage area under the convex hull is calculated.

[0185] Assume the set of dominant individuals includes For each individual, the corresponding weighted false positive rate and weighted true positive rate are calculated, and the weighted false positive rate is calculated according to the weighted false positive rate (horizontal axis). Sort by size from smallest to largest And forcibly added to the origin. and the end point The area of ​​the convex hull is calculated using the trapezoidal quadrature formula. The mathematical expression for the area of ​​the convex hull is:

[0186] ;

[0187] in, Representing the The weighted false positive rate for each individual Representing the The weighted true rate corresponding to each individual;

[0188] Based on the coordinates of elements in the dominant individual set in the ROC plane, a tangent projection method is used to construct a plane that encloses the current population and points towards the ideal point. A virtual convex hull curve that expands in a specific direction;

[0189] Tangent projection method specifically includes:

[0190] Obtain the true ROC convex hull polyline segment consisting of the dominant individuals in the current population. For each vertex on the true convex hull... Calculate its relationship with the ideal point Difference vectors between ;

[0191] Set expansion coefficient (This embodiment is) Virtual projection points are generated along the direction of the difference vector. ;

[0192] All generated virtual projection points and ideal point Connecting them sequentially, the resulting outer envelope polygonal line is the virtual convex hull curve, which represents the theoretically best classification performance boundary that can be achieved under the current physical environment.

[0193] The individuals remaining in the current population, excluding the dominant individuals, are non-convex hull individuals. For any non-convex hull individual, calculate the distance from the non-convex hull individual to each line segment of the virtual convex hull curve. The minimum distance from the convex hull individual to each line segment of the virtual convex hull curve is the Euclidean distance from the non-convex hull individual to the boundary of the virtual convex hull. Sort the Euclidean distances from all non-convex hull individuals to the boundary of the virtual convex hull in ascending order.

[0194] Set a fixed population size as (This embodiment is) ), the number of individuals in the dominant individual set ;

[0195] Calculate the number of replacements needed for the next generation population. );

[0196] Based on the sorting results, select the one with the smallest distance. A set of non-convex hull individuals is merged with the set of dominant individuals to form a set containing... The preferred parent population of each individual;

[0197] Perform crossover (exchanging core centers and weights) and mutation (fine-tuning width coefficients) operations on the retained preferred parent population to generate a new generation population;

[0198] After each generation of evolution, the change in the convex hull area of ​​the current generation is calculated. The process of obtaining the weighted false positive rate and weighted true positive rate, constructing the convex hull, and performing population crossover and mutation is repeated until any of the following termination conditions are met. The optimization is then stopped. The termination conditions include:

[0199] The preset maximum number of iterations is reached (50 times in this embodiment);

[0200] continuous If the area change over a generation (5 generations in this embodiment) is less than the preset convergence threshold (10⁻⁴ in this embodiment), it indicates that the population performance has reached the theoretical limit under the current physical environment, and the algorithm automatically converges and terminates.

[0201] Iterate through each individual in the final output set of dominant individuals and calculate the distance from each point to the ideal point. The Euclidean distance is used to select the individual with the smallest Euclidean distance, and the individual is decoded to obtain the optimal decision parameter set containing the kernel center, width and weight.

[0202] This invention introduces an evolutionary optimization strategy guided by weighted ROC convex hull. A virtual convex hull is constructed using the tangent projection method and used to select dominant individuals. The optimal solution is found directly in the multi-objective optimization space. The geometric properties of the ROC convex hull guide the population to evolve towards the theoretically optimal performance point (ideal point), avoiding the local optimum trap that traditional single-objective optimization may fall into. By maximizing the area under the ROC convex hull, the algorithm can automatically balance the sensitivity and specificity of the model, ensuring that the final set of optimal decision parameters can achieve the best balance of comprehensive performance under various working conditions.

[0203] Step S4, the process of generating an early warning specifically includes:

[0204] A nonlinear classification decision function is constructed using the optimal decision parameter set to calculate the state feature vector composed of the basic spatiotemporal correlation values ​​and trajectory direction entropy of all ships in the water area at the current moment;

[0205] If the value of the nonlinear classification decision function is greater than 0, the ship's navigation behavior is determined to be abnormal.

[0206] If the value of the nonlinear classification decision function is less than or equal to 0, the ship's navigation behavior is determined to be normal.

[0207] Map each ship in the current waters to a graph node, and assign attribute values ​​to the graph nodes according to the classification decision function.

[0208] If a ship's navigation behavior is abnormal, the ship is marked as a risk source node, and the ship's trajectory direction entropy is used as the node's initial risk potential energy.

[0209] If the ship's navigation behavior is normal, mark the ship as the victim target node and set the initial risk potential energy of the node corresponding to the ship to zero.

[0210] In one specific embodiment of the present invention, the final nonlinear classification decision function is constructed using the optimal decision parameter set output in step S3. For all ships in the waters at the current moment (not just high-confidence samples), their feature vectors are... Substitute into function Perform calculations;

[0211] like The system determines that the vessel's navigation behavior is abnormal, marks it as a risk source node, and generates an abnormal vessel warning.

[0212] like The vessel's navigation behavior was determined to be normal, and it was marked as a victim target node.

[0213] For ships determined to be abnormal, the trajectory direction entropy calculated in step S2 is directly used as the initial risk potential energy of the ship. The larger the trajectory direction entropy, the more chaotic the ship's motion (such as uncontrolled drift), and the greater the collision uncertainty and destructive energy it contains. For ships determined to be normal, their initial risk potential energy is zero.

[0214] This invention achieves risk quantification from behavior classification by using graph nodes and converting the trajectory direction entropy of ships into the initial risk potential energy of the nodes. Ships identified as abnormal are marked as risk sources, and the degree of disorder of their motion (entropy value) is directly used as a measure of potential destructive energy.

[0215] In step S4, the process of generating an early warning also includes:

[0216] For each victim node, with the victim node as the center and the preset safety warning distance as the radius, search for all risk source nodes within the radius.

[0217] Calculate the physical distance and relative approach speed between the victim target node and each risk source node found in the search;

[0218] The relative approach speed is calculated as follows: obtain the ground velocity vector of the risk source node and the ground velocity vector of the victim target node, calculate the residual between the ground velocity vector of the risk source node and the ground velocity vector of the victim target node to obtain the relative velocity vector, calculate the unit direction vector of the victim target node pointing to the risk source node, calculate the projection magnitude of the relative velocity vector on the unit direction vector, and use the projection magnitude as the relative approach speed.

[0219] Calculate the risk transmission coefficient from each risk source node to the affected target node based on physical distance and relative approach speed;

[0220] The risk transmission coefficient is calculated as follows: the square of the physical distance between the risk source node and the victim target node is calculated, and the square is added to a preset distance correction factor to obtain the distance attenuation factor; the ratio of the relative approach speed to the distance attenuation factor is calculated, and the ratio is determined as the risk transmission coefficient.

[0221] The initial risk potential of all risk source nodes within the search range is weighted by the risk transmission coefficient and then summed to obtain the total external risk potential of the victim target node.

[0222] If the sum of normalized external risk potential energy exceeds the preset risk threshold, an emergency collision warning will be generated.

[0223] In one specific embodiment of the present invention, for each victim target node, with the victim target node as the center and a preset safety warning distance as the radius (100 meters in this embodiment), all risk source nodes within the search radius are searched;

[0224] Calculate the physical distance and relative approach speed between the victim target node and each risk source node found in the search;

[0225] The method for calculating the relative proximity speed is as follows: obtain the speed and heading of the risk source node and the victim target node, perform vector decomposition on the speed, and obtain the ground velocity vector of the risk source node as ( , The ground velocity vector of the damaged target node is ( , );

[0226] Vector decomposition includes: The risk source node is the product of its ground speed and the sine of its ground heading. The product of the ground speed and ground heading of the risk source node is the cosine of the product. and Obtained using the same method;

[0227] Calculate the unit direction vector from the victim target node to the risk source node, calculate the projection magnitude of the relative velocity vector onto the unit direction vector, and use this projection magnitude as the relative approach velocity;

[0228] The method for calculating the unit direction vector from the victim target node to the risk source node is as follows:

[0229] The longitude difference is obtained by subtracting the longitude of the victim target node from the longitude of the Mercator projection of the risk source node, and the latitude difference is obtained by subtracting the latitude of the victim target node from the latitude of the Mercator projection of the risk source node. The physical distance between the risk source node and the victim target node is then obtained.

[0230] Divide the latitude difference by the physical distance to obtain the latitude component of the unit direction vector. The vector composed of this latitude and longitude components is the unit direction vector pointing from the victim target node to the risk source node.

[0231] The specific calculation method for the projection modulus is as follows:

[0232] Multiply the longitude component of the relative velocity vector by the longitude component of the unit direction vector to obtain the longitude product; multiply the latitude component of the relative velocity vector by the latitude component of the unit direction vector to obtain the latitude product; add the longitude product and the latitude product together, and the resulting value is the projection magnitude of the relative velocity vector onto the unit direction vector.

[0233] For each victim node, with the victim node as the center and the preset safety warning distance as the radius, search for all risk source nodes within the radius.

[0234] The mathematical expression for the risk transmission coefficient is:

[0235] ;

[0236] in, For risk transmission coefficient, For relative approximation speed, Physical distance The preset distance correction factor (a non-zero small constant, 0.0001 in this embodiment) is used to prevent computational singularities where the denominator is zero due to zero physical distance.

[0237] The risk transmission weight indicates that the faster the relative speed and the closer the distance, the higher the risk transmission efficiency, that is, the greater the collision risk.

[0238] For each victim node, traverse all risk source nodes within the search radius of the victim node, and calculate the total external risk potential energy received by the victim node. The mathematical expression for the total external risk potential energy is:

[0239] ;

[0240] in, The sum of external risk potential energy. This refers to the set of neighboring risk source nodes within a preset search radius of the victim target node. For the first The risk transmission coefficient from each risk source node to the current victim target node. For the first The initial risk potential energy of each risk source node (i.e., the trajectory direction entropy of that node).

[0241] The sum of external risk potential energy quantifies the comprehensive threat posed to the target ship when all out-of-control vessels (high trajectory direction entropy) approach the target ship at the current speed and distance.

[0242] The sum of external risk potential energy is normalized and compared with a preset risk threshold. If it exceeds the risk threshold (set to 0.85 in this embodiment), a chain reaction warning is generated, indicating to the VTS center that the normal vessel is about to face a passive collision risk.

[0243] By calculating the projection of the relative approach velocity onto the unit direction vector and the attenuation effect of physical distance, the propagation process of risk potential energy in the water space is accurately simulated. It not only considers spatial proximity but also emphasizes the impact of relative motion trends on risk. When the total external risk potential energy exceeds the threshold, a chain reaction warning is triggered, which can inform normally sailing ships of the impending passive collision risk in advance, realizing the functional extension from monitoring anomalies to active collision avoidance.

[0244] A big data-based waterway vessel navigation behavior analysis system includes a data acquisition module, a feature construction module, an iteration module, and an analysis module.

[0245] The data acquisition module is used to collect dynamic data of ships and static data of inland waterways. It constructs an environmental semantic grid based on the static data of the waterways and maps the dynamic data of ships to the environmental semantic grid to generate semantic trajectory sequences.

[0246] The feature construction module is used to calculate the basic spatiotemporal correlation value and trajectory direction entropy of trajectory segments based on semantic trajectory sequences, and construct ship behavior feature vectors.

[0247] The iterative module is used to construct a dynamic pseudo-label dataset based on the distribution characteristics of ship behavior feature vectors in the feature plane, construct a nonlinear classification decision function, and use an evolutionary algorithm based on the weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function in combination with the dynamic pseudo-label dataset to obtain the optimal decision parameter set.

[0248] The analysis module is used to construct the final nonlinear classification decision function using the optimal decision parameter set, to determine the navigation behavior of ships in the water at the current moment, and to generate an early warning.

[0249] This invention constructs an environmental semantic grid and generates semantic trajectory sequences, enabling deep integration of discrete ship dynamic data with the static environment of the waterway. This overcomes the shortcomings of traditional methods that rely solely on kinematic parameters while neglecting geographical constraints. The invention introduces trajectory direction entropy based on local minimum spanning trees as a key feature. Leveraging the ability of information entropy to effectively measure system disorder, it accurately captures highly nonlinear abnormal behaviors such as ship drift and spinning in place, significantly improving the sensitivity to complex hazards and solving the problem that conventional linear statistical features are insufficient to characterize chaotic ship motion. Furthermore, this invention proposes a method based on dynamic pseudo-label datasets and weighted ROC convex hull entropy... The evolutionary optimization strategy guided by this invention automatically generates pseudo-labels and confidence levels by setting absolute safety and absolute anomaly extremes, effectively solving the problem of scarce abnormal samples in real-world scenarios. By using weighted ROC convex hull to guide the parameter iteration of the decision function, this invention can automatically focus on high-confidence samples while also considering the division of fuzzy boundaries. This allows it to obtain the optimal nonlinear classification boundary even in cases of sample imbalance, significantly improving the robustness of behavior determination. It can quantify abnormal ships as risk sources with specific energy and calculate their risk transmission coefficient to surrounding normal ships by combining relative approach speed and physical distance. This not only enables the identification of single-ship anomalies but also generates chain reaction warnings for the affected targets.

[0250] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for analyzing the navigation behavior of vessels in waterways based on big data, characterized in that: Includes the following steps: Step S1: Collect dynamic data of ships and static data of inland waterways, construct an environmental semantic grid based on the static data of the waterways, and map the dynamic data of ships to the environmental semantic grid to generate a semantic trajectory sequence; Step S2: Based on the semantic trajectory sequence, calculate the basic spatiotemporal correlation value and trajectory direction entropy of the trajectory segment, and construct the ship behavior feature vector; Step S3: Based on the distribution characteristics of the ship behavior feature vector in the feature plane, construct a dynamic pseudo-label dataset, construct a nonlinear classification decision function, and use an evolutionary algorithm based on the weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function in combination with the dynamic pseudo-label dataset to obtain the optimal decision parameter set. Step S4: Construct the final nonlinear classification decision function using the optimal decision parameter set to determine the navigation behavior of ships in the water at the current moment and generate an early warning. In step S3, the process of constructing the dynamic pseudo-label dataset specifically includes: Map the mean of the basic spatiotemporal correlation and the trajectory direction entropy of each ship in the current sliding window to sample points in a two-dimensional feature plane; Preset absolute safe poles and absolute outlier poles in a two-dimensional feature plane; In a two-dimensional feature plane, calculate the first distance from each sample point to the absolute safe pole and the second distance to the absolute abnormal pole; Based on the ratio of the difference between the first distance and the second distance, generate the corresponding pseudo-label of the ship and the corresponding pseudo-label confidence level; The process of generating the pseudo-tags includes: If the first distance is less than the second distance, the navigation status of the corresponding vessel will be marked as a suspected normal label; If the first distance is greater than or equal to the second distance, the navigation status of the corresponding vessel will be marked as a suspected anomaly. The process of generating the pseudo-label confidence score includes: The ratio of the absolute value of the difference between the first distance and the second distance to the sum of the first distance and the second distance is used as the pseudo-label confidence score; If the confidence level of the pseudo-tag of the corresponding vessel is greater than the preset execution threshold, then the pseudo-tag of the vessel's navigation status will be recorded as a high-confidence pseudo-tag. If a ship's navigation status is marked as a suspected normal label and the pseudo label is a high-confidence pseudo label, then the sample points corresponding to the ship constitute a dynamic pseudo label dataset. In step S3, the mathematical expression of the nonlinear classification decision function is: ； in, It is a vector composed of the basic spatiotemporal correlation value and the trajectory direction entropy. The number of radial basis functions. For the first The kernel center location of each radial basis function. For the first The width coefficients of the radial basis functions, For the first The weighting coefficients of the radial basis functions, This is the bias threshold.

2. The method for analyzing the navigation behavior of vessels in waterways based on big data as described in claim 1, characterized in that, Step S1 includes the following sub-steps: The ship dynamic data includes the ship's MMSI, latitude and longitude coordinates, speed over land, and heading over land; The static data of the waterway includes functional areas on the electronic chart and the maximum speed limit stipulated in the waterway. The process of constructing the environmental semantic grid includes: dividing the waterway into discrete grid units, defining the functional attributes of each grid unit according to the electronic nautical chart, and assigning a corresponding environmental weight value to each grid unit based on the functional attributes; The ship dynamic data within the preset sliding window is mapped to the environmental semantic grid, and the semantic trajectory sequence of the ship is constructed by combining the environmental weight values.

3. The method for analyzing the navigation behavior of ships in waterways based on big data as described in claim 2, characterized in that, In step S2, the process of constructing the ship behavior feature vector specifically includes: The ship behavior feature vector in the semantic trajectory sequence is used as the trajectory point. The trajectory points in the semantic trajectory sequence are traversed. Two trajectory points are arbitrarily selected to form a trajectory point pair. The latitude and longitude coordinates of the trajectory point pair are extracted, converted into spatial projection coordinates, and the Euclidean distance and time difference between the two points are calculated. Based on the Euclidean distance and time difference, the basic spatiotemporal correlation value of each pair of trajectory points is calculated. Construct a local minimum spanning tree of trajectory points within a sliding window, statistically analyze the distribution probability of the direction angles of each connecting line in the local minimum spanning tree, and calculate the trajectory direction entropy using the Shannon entropy formula.

4. The method for analyzing the navigation behavior of ships in waterways based on big data as described in claim 3, characterized in that, In step S3, the process of obtaining the optimal decision parameter set specifically includes: The kernel center position, width coefficient, weight coefficient, and bias threshold of the nonlinear classification decision function are used as the individuals in the population to be optimized. Multiple population individuals are randomly generated within a preset value range, and multiple classification decision functions are obtained using the randomly generated population individuals; The classification decision function is used to predict the dynamic pseudo-label dataset to obtain the pseudo-label prediction results. Based on the comparison between the pseudo-label prediction results of the decision function and the pseudo-labels of the sample points, a weighted statistic is calculated in combination with the pseudo-label confidence of the sample points. The weighted statistic includes weighted true positives and weighted false positives. The weighted true instance is as follows: for all sample points whose pseudo-labels are suspected to be abnormal and whose prediction results are abnormal, the confidence weights of the pseudo-labels corresponding to the sample points are accumulated. The weighted false positives are as follows: for all sample points whose pseudo-labels are suspected to be normal but whose prediction results are abnormal, the confidence weights of the pseudo-labels corresponding to the sample points are accumulated. The weighted true positive rate is the ratio of the weighted true positive rate to the sum of the confidence weights of all pseudo-labeled sample points that are suspected of being abnormal, and the weighted false positive rate is the ratio of the weighted false positive rate to the sum of the confidence weights of all pseudo-labeled sample points that are suspected of being normal.

5. The method for analyzing the navigation behavior of vessels in waterways based on big data as described in claim 4, characterized in that, In step S3, the process of obtaining the optimal decision parameter set further includes: The weighted true positive rate and weighted false positive rate of each individual in the population are used as the coordinate points of the individual in the ROC space; The convex hull algorithm is used to identify the set of dominant individuals located at the vertices of the ROC convex hull in the ROC space. Based on the coordinates of the individuals in the set of dominant individuals, the area of ​​the convex hull is calculated using the trapezoidal quadrature formula. Based on the coordinates of the elements in the dominant individual set in the ROC space, a virtual convex hull curve is constructed using the tangent projection method, which encloses the current population and expands towards the ideal point. The individuals remaining in the current population, excluding the dominant individuals set, are non-convex hull individuals. For any non-convex hull individual, calculate the Euclidean distance from the non-convex hull individual to the virtual convex hull curve, sort them in ascending order, and select the top few non-convex hull individuals from the sorting results. Construct a preferred parent population from several non-convex hull individuals and individuals from the set of dominant individuals. Perform crossover and mutation operations on the preferred parent population, and repeat the iterative optimization until the preset stopping iteration condition is reached, and output the final set of dominant individuals. Iterate through each individual in the final set of dominant individuals, calculate the Euclidean distance from the individual to the ideal point, select the individual with the smallest Euclidean distance, and extract the kernel center position, width coefficient, weight coefficient, and bias threshold as the optimal decision parameter set.

6. The method for analyzing the navigation behavior of vessels in waterways based on big data as described in claim 5, characterized in that, In step S4, the process of generating the early warning specifically includes: Using the optimal decision parameter set, a nonlinear classification decision function is constructed to calculate the state feature vector composed of the basic spatiotemporal correlation value and trajectory direction entropy of all ships in the water area at the current moment; If the value of the nonlinear classification decision function is greater than 0, the ship's navigation behavior is determined to be abnormal. If the value of the nonlinear classification decision function is less than or equal to 0, the ship's navigation behavior is determined to be normal. Map each ship in the current waters to a graph node, and assign attribute values ​​to the graph nodes according to the classification decision function. If a ship's navigation behavior is abnormal, the ship is marked as a risk source node, and the ship's trajectory direction entropy is used as the node's initial risk potential energy. If the ship's navigation behavior is normal, mark the ship as the victim target node and set the initial risk potential energy of the node corresponding to the ship to zero.

7. The method for analyzing the navigation behavior of ships in waterways based on big data as described in claim 6, characterized in that, In step S4, the process of generating the early warning also includes: For each victim node, with the victim node as the center and the preset safety warning distance as the radius, search for all risk source nodes within the radius. Calculate the physical distance and relative approach speed between the victim target node and each risk source node found in the search; The relative approach speed is calculated as follows: obtain the ground velocity vector of the risk source node and the ground velocity vector of the victim target node, calculate the residual between the ground velocity vector of the risk source node and the ground velocity vector of the victim target node to obtain the relative velocity vector, calculate the unit direction vector of the victim target node pointing to the risk source node, calculate the projection magnitude of the relative velocity vector on the unit direction vector, and use the projection magnitude as the relative approach speed. Calculate the risk transmission coefficient from each risk source node to the affected target node based on physical distance and relative approach speed; The risk transmission coefficient is calculated as follows: the square of the physical distance between the risk source node and the victim target node is calculated, and the square is added to a preset distance correction factor to obtain a distance attenuation factor; the ratio of the relative approach speed to the distance attenuation factor is calculated, and the ratio is determined as the risk transmission coefficient. The initial risk potential of all risk source nodes within the search range is weighted by the risk transmission coefficient and then summed to obtain the total external risk potential of the victim target node. If the sum of normalized external risk potential energy exceeds the preset risk threshold, an emergency collision warning will be generated.

8. A big data-based waterway vessel navigation behavior analysis system, applied in the big data-based waterway vessel navigation behavior analysis method as described in any one of claims 1-7, characterized in that, It includes a data acquisition module, a feature construction module, an iteration module, and an analysis module; The acquisition module is used to collect dynamic data of ships and static data of inland waterways, construct an environmental semantic grid based on the static data of the waterways, and map the dynamic data of ships to the environmental semantic grid to generate a semantic trajectory sequence. The feature construction module is used to calculate the basic spatiotemporal correlation value and trajectory direction entropy of the trajectory segment based on the semantic trajectory sequence, and construct the ship behavior feature vector; The iterative module is used to construct a dynamic pseudo-label dataset based on the distribution characteristics of the ship behavior feature vector in the feature plane, construct a nonlinear classification decision function, and use an evolutionary algorithm based on the weighted ROC convex hull guidance strategy to iteratively optimize the parameters of the nonlinear classification decision function in combination with the dynamic pseudo-label dataset to obtain the optimal decision parameter set. The analysis module is used to construct the final nonlinear classification decision function using the optimal decision parameter set, to determine the navigation behavior of ships in the water area at the current moment, and to generate an early warning.

Images