Ship abnormal behavior identification method and device

Abstract

The invention provides a ship abnormal behavior identification method and device, and belongs to the technical field of ship transportation, and the method comprises the steps: extracting a ship track based on AIS data; compressing the ship trajectory, and discretizing ship attributes corresponding to the ship trajectory to obtain ship behavior feature data; based on the ship behavior characteristic data, generating ship behavior characteristic sequence embedding quantity; inputting the ship behavior feature sequence embedding quantity into the trained sequence-to-sequence network model to obtain ship behavior identification information; the network model comprises an encoder and a decoder; and determining an abnormal score of the ship behavior feature sequence embedding amount, and obtaining a ship abnormal behavior identification result based on the abnormal score and the ship behavior identification information. According to the method, the trained sequence-to-sequence network model is adopted, and the abnormal score of the embedding amount of the ship behavior feature sequence is combined, so that the technical problem that the ship abnormal behavior is difficult to accurately and reliably identify in the existing scheme is solved.

Description

Technical Field

[0001] The present invention relates to the field of ship transportation technology, and in particular to a method and device for identifying abnormal ship behavior. Background Art

[0002] Identification of abnormal ship behavior is an important research topic in the field of maritime safety and a key means to ensure the safety of ship navigation and control illegal marine activities. Efficient and accurate identification of abnormal ship behavior can assist maritime management personnel in achieving high-quality supervision of large-scale ships in target waters, and can also help researchers study the movement characteristics of abnormal ship behavior, which is of great significance to improving ship navigation safety and maritime traffic management.

[0003] There are methods based on statistics, clustering and deep learning for identifying abnormal ship behavior. Among them, the statistical method aims to establish a statistical model of random variables related to ship characteristics. This method often requires setting assumptions first, which is often inconsistent with the actual situation, resulting in inaccurate modeling. Clustering-based methods can be used to extract common movement patterns of ships and use this pattern to achieve detection tasks; they can also be directly coupled with detection tasks. This type of method has parameters that need to be defined in advance, and the values ​​of these parameters will have a great impact on clustering performance. In recent years, deep learning technology has been gradually applied to the field of water transportation due to its powerful feature extraction capabilities, end-to-end learning, nonlinear modeling, parallel computing and other advantages.

[0004] Due to the complex characteristics of ship motion, diverse navigation modes, and noise in the AIS data sent and received by ships, it is difficult for researchers to analyze and understand the navigation modes and laws of ships, which is not conducive to building a reliable framework for identifying abnormal ship behavior, limiting the improvement of maritime supervision and ship navigation safety. Therefore, existing solutions are difficult to accurately and reliably identify abnormal ship behavior. Summary of the invention

[0005] In view of this, it is necessary to provide a method and device for identifying abnormal behavior of a ship, so as to solve the technical problem that the existing solutions are difficult to accurately and reliably identify abnormal behavior of a ship.

[0006] In order to solve the above problems, on the one hand, the present invention provides a method for identifying abnormal behavior of a ship, comprising: Extract ship tracks based on AIS data; Compressing the ship trajectory and discretizing the ship attributes corresponding to the ship trajectory to obtain ship behavior characteristic data; Based on the ship behavior characteristic data, generating a ship behavior characteristic sequence embedding amount; The embedding amount of the ship behavior feature sequence is input into a trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; An abnormal score of the embedding amount of the ship behavior feature sequence is determined, and a ship abnormal behavior recognition result is obtained based on the abnormal score and the ship behavior recognition information.

[0007] In a possible implementation, extracting a ship track based on the AIS data includes: Filtering target attributes from the AIS data; Based on the target attribute, the AIS data is grouped according to the MMSI code to obtain grouped data; Sorting each set of data in the plurality of sets of data in the order of timestamps to obtain an original ship track; wherein each set of data corresponds to an original ship track; Segment the trajectories whose temporal and spatial intervals exceed a preset interval threshold among the multiple original ship trajectories to obtain preliminary extracted ship trajectories; The initially extracted ship trajectory is corrected to obtain the final ship trajectory.

[0008] In a possible implementation manner, screening target attributes from the AIS data includes: Selecting multiple basic attributes from the AIS data; the multiple basic attributes include: ship flag, ship type, ship length, ship width, MMSI code, timestamp, longitude, latitude, ground heading and ground speed; When it is determined that the basic attribute is a static attribute and there is an error in the static attribute, a preset ship information database is queried based on the MMSI code of the erroneous data in the static attribute to obtain correct data, and the erroneous data is replaced based on the correct data to obtain the target attribute; When it is determined that the basic attribute is a dynamic attribute and there is an error in the dynamic attribute, a corresponding data repair method is determined based on the error type of the dynamic attribute to repair the dynamic attribute to obtain a target attribute.

[0009] In a possible implementation, the initially extracted ship trajectory is corrected to obtain a final ship trajectory, including: Select two endpoints to be corrected from the initially extracted ship trajectory, and determine the neighborhoods corresponding to the two endpoints respectively; Filtering the ship trajectories passing through two neighborhoods from the initially extracted ship trajectories; The sub-trajectory of the screened ship trajectory between two neighborhoods is added to the candidate trajectory set; Clustering the candidate trajectory set to obtain a candidate trajectory class cluster set; Determine the support of each cluster in the candidate trajectory cluster set, and determine the cluster with the largest support; The representative trajectory of the cluster with the largest support is determined, and the initially extracted ship trajectory is corrected based on the representative trajectory to obtain the final ship trajectory.

[0010] In a possible implementation, the ship trajectory is compressed, and the ship attributes corresponding to the ship trajectory are discretized to obtain ship behavior characteristic data, including: Determine the distance from each trajectory point in the target trajectory point set to the line segment connecting the starting point and the end point of the target trajectory point set as the first type distance, and determine the maximum distance from the first type distances corresponding to each trajectory point as the first target distance; the target trajectory point set is a set of trajectory points on the ship trajectory; When the first target distance is greater than a preset distance threshold, determining the trajectory point corresponding to the first target distance as a feature point, and dividing the target trajectory point set into two target trajectory point subsets based on the trajectory points corresponding to the first target distance; Determine the distance between each trajectory point in the target trajectory point subset and the line segment connecting the starting point and the end point of the target trajectory point subset as the second type distance, and determine the maximum distance from the second type distances corresponding to each trajectory point as the second target distance; When the second target distance is less than or equal to the preset distance threshold, determine the anchor point at the current moment, and calculate the speed difference from the track point behind the anchor point at the current moment to the anchor point, and determine the track point whose speed difference is greater than the preset distance threshold as the anchor point and feature point at the next moment, until all feature points in the target track point set are determined; The ship attributes of all feature points in the target trajectory point set are discretized to obtain ship behavior feature data.

[0011] In a possible implementation, the anchor point at the initial moment is the starting point of the target trajectory point subset.

[0012] In a possible implementation, generating a ship behavior feature sequence embedding amount based on the ship behavior feature data includes: One-hot encoding is performed on the ship behavior characteristic data to obtain a corresponding one-hot vector; The one-hot vector is multiplied by the embedding matrix to obtain the embedding amount of the ship behavior feature sequence.

[0013] In a possible implementation, determining the abnormal score of the embedding amount of the ship behavior feature sequence includes: The embedding amount of the ship behavior feature sequence is used as a label of the network model, and based on the BLEU algorithm and the label, the abnormal score of the embedding amount of the ship behavior feature sequence is calculated.

[0014] In a possible implementation, the embedding amount of the ship behavior feature sequence is input into a trained sequence-to-sequence network model to obtain ship behavior recognition information, including: Inputting the ship behavior feature sequence embedding amount into the encoder, so as to encode the ship behavior feature sequence embedding amount according to the element order based on the encoder to generate a context vector with a fixed dimension; The context vector is input into the decoder to obtain ship behavior recognition information; the decoder is added with an attention mechanism.

[0015] On the other hand, the present invention also provides a device for identifying abnormal behavior of a ship, comprising: Track extraction module, used to extract ship tracks based on AIS data; A feature extraction module, used to compress the ship trajectory and discretize the ship attributes corresponding to the ship trajectory to obtain ship behavior feature data; A sequence generation module, used for generating a ship behavior feature sequence embedding amount based on the ship behavior feature data; An information recognition module is used to input the embedding amount of the ship behavior feature sequence into a trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; The anomaly determination module is used to determine the anomaly score of the embedding amount of the ship behavior feature sequence, and obtain the ship abnormal behavior identification result based on the anomaly score and the ship behavior identification information.

[0016] The beneficial effects of adopting the above implementation method are as follows: the ship abnormal behavior identification method and device provided by the present invention compresses the ship trajectory and discretizes the ship attributes corresponding to the ship trajectory to obtain ship behavior feature data, effectively filters out redundant trajectory points, and effectively retains the speed information of the track while significantly reducing the data scale. Then, the ship attributes are discretized to improve the data granularity, improve the model training effect and generalization performance; based on the ship behavior feature data, the ship behavior feature sequence embedding is generated; the ship feature sequence embedding is input into the trained sequence-to-sequence network model to obtain ship behavior identification information, and the embedding technology is used to deeply mine the potential information of the ship behavior characteristics, and the encoding and decoding of the ship trajectory characteristics is realized based on the sequence-to-sequence network model, so as to understand the ship behavior pattern in complex scenarios and accurately capture the ship behavior law. Improve the accuracy of ship behavior recognition, and judge abnormal behavior in combination with the abnormal score of the ship behavior feature sequence embedding, thereby solving the technical problem that the existing solution is difficult to accurately and reliably identify abnormal ship behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A flow chart of an embodiment of a method for identifying abnormal ship behavior provided by the present invention; Figure 2 A flow chart of another embodiment of the method for identifying abnormal behavior of a ship provided by the present invention; Figure 3 A schematic diagram of the visualization results before and after the ship trajectory correction algorithm provided by the present invention is applied to a certain trajectory; Figure 4 A schematic diagram of a ship trajectory compression algorithm based on speed perception provided by the present invention; Figure 5 The compression rate, DTW distance, and average speed error box plot of the speed-aware ship trajectory compression algorithm provided by the present invention; Figure 6 A schematic diagram of embedding ship behavior characteristics provided by the present invention; Figure 7 A schematic diagram of the Seq2Seq model architecture provided by the present invention; Figure 8 A principle block diagram of an embodiment of a device for identifying abnormal behavior of a ship provided by the present invention; Fig. 9A schematic structural diagram of an embodiment of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0020] In the description of the embodiments of the present application, unless otherwise specified, “plurality” means two or more than two.

[0021] The terms "including" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product or equipment comprising a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules not clearly listed or inherent to these processes, methods, products or equipment.

[0022] The naming or numbering of the steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time / logical sequence indicated by the naming or numbering. The execution order of the named or numbered process steps can be changed according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved.

[0023] Reference to an "embodiment" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0024] The present invention provides a method and device for identifying abnormal behavior of a ship, which are described below respectively.

[0025] like Figure 1 As shown, the present invention provides a method for identifying abnormal behavior of a ship, comprising: S101, extracting ship tracks based on Automatic Identification System (AIS) data; S102, compressing the ship trajectory, and discretizing the ship attributes corresponding to the ship trajectory to obtain ship behavior feature data; S103, generating a ship behavior feature sequence embedding amount based on the ship behavior feature data; The embedding amount of the ship behavior feature sequence is input into a trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; S104: Determine an abnormal score of the embedded amount of the ship behavior feature sequence, and obtain a ship abnormal behavior identification result based on the abnormal score and the ship behavior identification information.

[0026] It is understandable that the method provided by the present invention can be summarized as the following steps: Filter out redundant AIS data attributes, fix AIS data errors, and extract ship tracks; A ship trajectory compression algorithm based on speed perception is used to compress the ship trajectory and discretize the ship attributes according to the ship's heading and speed attributes; Construct ship behavior feature sequence; Using word embedding technology to generate a ship behavior feature sequence embedding amount based on the ship behavior feature sequence; The ship behavior feature sequence is embedded as the input of a sequence-to-sequence (Seq2Seq) encoder, each element is encoded to generate a context vector, and the context vector is decoded by a Seq2Seq decoder. An attention mechanism is introduced during decoding to incorporate the state information of the Seq2Seq encoder at each moment into the decoding process; The ship behavior feature sequence is used as a label of Seq2Seq, and the BLEU (Bilingual Evaluation Understudy) algorithm is used to calculate the abnormal score of the ship behavior feature sequence, so as to determine the ship abnormal behavior recognition result. In some embodiments, extracting a ship track based on the AIS data includes: Filtering target attributes from the AIS data; Based on the target attribute, the AIS data is grouped according to the MMSI code to obtain grouped data; Sorting each set of data in the plurality of sets of data in the order of timestamps to obtain an original ship track; wherein each set of data corresponds to an original ship track; Segment the trajectories whose temporal and spatial intervals exceed a preset interval threshold among the multiple original ship trajectories to obtain preliminary extracted ship trajectories; The initially extracted ship trajectory is corrected to obtain the final ship trajectory.

[0027] In some embodiments, filtering target attributes from the AIS data includes: Selecting multiple basic attributes from the AIS data; the multiple basic attributes include: ship flag, ship type, ship length, ship width, MMSI code, timestamp, longitude, latitude, ground heading and ground speed; When it is determined that the basic attribute is a static attribute and there is an error in the static attribute, a preset ship information database is queried based on the MMSI code of the erroneous data in the static attribute to obtain correct data, and the erroneous data is replaced based on the correct data to obtain the target attribute; When it is determined that the basic attribute is a dynamic attribute and there is an error in the dynamic attribute, a corresponding data repair method is determined based on the error type of the dynamic attribute to repair the dynamic attribute to obtain a target attribute.

[0028] It can be understood that, based on the AIS data, extracting the ship track specifically includes: From the AIS data, 10 attributes including ship registration, ship type, ship length, ship width, MMSI (Maritime Mobile Service Identity), timestamp, longitude, latitude, ground course and ground speed were selected as research attributes; If there is an error in the static attribute in the AIS data, the ship information database is queried according to the MMSI of the erroneous data, and the erroneous attribute is replaced with the correct attribute to obtain the target attribute; If there are errors in the dynamic attributes in the AIS data, the corresponding repair method is adopted according to the specific error type to obtain the target attribute; The AIS data are grouped by MMSI and sorted in timestamp order to obtain original ship trajectories, and the original ship trajectories whose time-space interval exceeds a threshold are segmented to generate preliminary extracted ship trajectories; The initially extracted ship trajectory is corrected using a data-driven ship trajectory correction method.

[0029] In some embodiments, the initially extracted ship trajectory is corrected to obtain a final ship trajectory, including: Select two endpoints to be corrected from the initially extracted ship trajectory, and determine the neighborhoods corresponding to the two endpoints respectively; Filtering the ship trajectories passing through two neighborhoods from the initially extracted ship trajectories; The sub-trajectory of the screened ship trajectory between two neighborhoods is added to the candidate trajectory set; Clustering the candidate trajectory set to obtain a candidate trajectory class cluster set; Determine the support of each cluster in the candidate trajectory cluster set, and determine the cluster with the largest support; The representative trajectory of the cluster with the largest support is determined, and the initially extracted ship trajectory is corrected based on the representative trajectory to obtain the final ship trajectory.

[0030] It is understandable that the data-driven ship trajectory correction method is used to correct the initially extracted ship trajectory. The specific steps are as follows: a. The endpoint of the part of the trajectory to be corrected is the center of the circle, r is the radius, and the neighborhood is determined respectively ; b. From the initially extracted ship trajectories, filter out the The sub-trajectories of these trajectories between the two neighborhoods are added as new trajectories to the candidate trajectory set ; c. Yes The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm is executed, in which the DTW (Dynamic Time Wrapping) algorithm is used to calculate the distance between ship trajectories to obtain candidate trajectory clusters. ; d. Calculation C Each cluster Ci The support degree sup( Ci ), and obtain the cluster with the largest support , to determine the most likely path of the ship's trajectory C Which cluster to drive in? e. Calculation Representative trajectory of , to determine the specific route of the target ship; f. Generate the completion part of the target trajectory according to the representative trajectory , and the corrected trajectory is obtained.

[0031] In some embodiments, the ship trajectory is compressed, and the ship attributes corresponding to the ship trajectory are discretized to obtain ship behavior feature data, including: Determine the distance from each trajectory point in the target trajectory point set to the line segment connecting the starting point and the end point of the target trajectory point set as the first type distance, and determine the maximum distance from the first type distances corresponding to each trajectory point as the first target distance; the target trajectory point set is a set of trajectory points on the ship trajectory; When the first target distance is greater than a preset distance threshold, determining the trajectory point corresponding to the first target distance as a feature point, and dividing the target trajectory point set into two target trajectory point subsets based on the trajectory points corresponding to the first target distance; Determine the distance between each trajectory point in the target trajectory point subset and the line segment connecting the starting point and the end point of the target trajectory point subset as the second type distance, and determine the maximum distance from the second type distances corresponding to each trajectory point as the second target distance; When the second target distance is less than or equal to the preset distance threshold, the anchor point at the current moment is determined, and the speed difference from the trajectory point behind the anchor point at the current moment to the anchor point is calculated, and the trajectory point with a speed difference greater than the preset distance threshold is determined as the anchor point and feature point at the next moment, until all feature points in the target trajectory point set are determined; the anchor point at the initial moment is the starting point of the target trajectory point subset; The ship attributes of all feature points in the target trajectory point set are discretized to obtain ship behavior feature data.

[0032] It is understandable that the step of obtaining the ship behavior characteristic data may specifically include the following steps: Step (1) calculates the distance from the target ship's trajectory from the next trajectory point after the starting point to the previous trajectory point of the end point to the line segment connecting the starting point and the end point of the target ship's trajectory, and obtains the maximum distance d max and corresponding trajectory points p max. The distance from the trajectory point to the line segment is the Euclidean distance; Step (2) If d If max is greater than the threshold, execute step (3), otherwise execute step (4); Step (3): Determine p max is the feature point, and takes the starting point as p max is the new starting point 1 and the new end point 1. p max, the end point is the new starting point No. 2, the new end point No. 2, and the algorithm is executed recursively; Step (4), take the starting point as the anchor point and the point next to the starting point as the current point; Step (5) Calculate the speed difference between the current point and the anchor point . Track Point The speed difference is expressed as .in, Represents trajectory points p Speed ​​over ground; Step (6) If If the value is less than the threshold, execute step (7); otherwise, execute step (8); Step (7): If the current point is the previous point of the end point, the algorithm ends or jumps out of the current recursion, otherwise, execute step (9); Step (8), taking the current point as the feature point and anchor point; Step (9), taking the next point of the current point as the new current point, execute step (5).

[0033] In some embodiments, generating a ship behavior feature sequence embedding amount based on the ship behavior feature data includes: One-hot encoding is performed on the ship behavior characteristic data to obtain a corresponding one-hot vector; The one-hot vector is multiplied by the embedding matrix to obtain the embedding amount of the ship behavior feature sequence.

[0034] In some embodiments, determining an anomaly score of the embedding amount of the ship behavior feature sequence includes: The embedding amount of the ship behavior feature sequence is used as a label of the network model, and based on the BLEU algorithm and the label, the abnormal score of the embedding amount of the ship behavior feature sequence is calculated.

[0035] It can be understood that the BLEU algorithm uses the sequence composed of elements corresponding to the maximum a posteriori probability component in the output of the Seq2Seq decoder as the generated sequence, and the ship behavior feature sequence as the reference sequence.

[0036] In some embodiments, the embedding amount of the ship behavior feature sequence is input into a trained sequence-to-sequence network model to obtain ship behavior recognition information, including: Inputting the ship behavior feature sequence embedding amount into the encoder, so as to encode the ship behavior feature sequence embedding amount according to the element order based on the encoder to generate a context vector with a fixed dimension; The context vector is input into the decoder to obtain ship behavior recognition information; the decoder is added with an attention mechanism.

[0037] It can be understood that the input of the Seq2Seq encoder is the corresponding element embedding in the ship behavior feature sequence and the hidden state of the Seq2Seq encoder at the previous moment, and the context vector is the hidden state of the Seq2Seq encoder at the last moment.

[0038] The input of the Seq2Seq decoder includes the embedding with the maximum posterior probability component in its previous output, the hidden state of the Seq2Seq decoder at the previous moment, and the attention vector summarized from the hidden states of the Seq2Seq encoder at different moments.

[0039] The hidden state of the Seq2Seq decoder at the previous moment is used as a query vector, and the attention distribution of the hidden state of the Seq2Seq encoder at different moments is calculated in combination with the scoring function and the Softmax function, and finally the attention vector is generated through the attention distribution.

[0040] In one embodiment, the Chengshanjiao waters are used as the research waters, and the geographical range is 122°0′E-124°40′E, 36°12′N-38°22′N. The time range of the collected AIS data is from 0:00 on May 1, 2023 to 24:00 on May 31, 2023. By collecting AIS data in this time and space range and executing the Seq2Seq (sequence to sequence)-based ship abnormal behavior identification method and system, abnormal ship behavior is identified.

[0041] See also Figure 2 : The method and system for identifying abnormal behavior of ships based on Seq2Seq in this embodiment are described in detail, and the method includes S201-S206.

[0042] S201, filter out redundant AIS data attributes, repair AIS data errors, and extract ship tracks. The specific steps are as follows: In step (1), 10 attributes including ship registration, ship type, ship length, ship width, MMSI code, timestamp, longitude, latitude, course over ground and speed over ground were selected from AIS data as research attributes; Step (2), if there is an error in the static attribute in the AIS data, the ship information database is queried according to the MMSI of the erroneous data and the erroneous attribute is replaced with the correct attribute; Step (3), if there are errors in the dynamic attributes in the AIS data, the corresponding repair method is adopted according to the specific error type. Specifically, the errors in dynamic attributes include attribute missing, attribute out of bounds and data duplication, and the corresponding repair methods are linear interpolation, linear interpolation and deletion; Step (4) groups the AIS data by MMSI and sorts them in timestamp order to obtain original ship trajectories, segments the original ship trajectories whose time-space interval exceeds a threshold, and generates preliminary extracted ship trajectories; Step (5) uses a data-driven ship trajectory correction method to correct the initially extracted ship trajectory. The specific steps are as follows: a. The endpoint of the part of the trajectory to be corrected is the center of the circle, r is the radius, and the neighborhood is determined respectively ; b. From the initially extracted ship trajectories, filter out the The sub-trajectories of these trajectories between the two neighborhoods are added as new trajectories to the candidate trajectory set ; c. Yes The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm is executed, in which the DTW (Dynamic Time Wrapping) algorithm is used to calculate the distance between ship trajectories to obtain candidate trajectory clusters. ; d. Calculation C Each cluster Ci The support degree sup( Ci ), and obtain the cluster with the largest support , to determine the most likely path of the ship's trajectory C Which cluster to drive in? Class Cluster The support degree sup( Ci ) is calculated as follows:

[0043] in, f Indicates features, including 8 types: position, course, speed, time, ship type, length, width and ship registration; wf Represents predefined features f The weight of , the larger the weight, f The more significant the difference is in terms of support; Represents the target ship trajectory and cluster Ci Middle j Ship tracks tj In Features f The difference on Representation characteristics f The corresponding threshold. Indicator function The definition is as follows:

[0044] In addition, the calculation methods of the 8 characteristic differences are as follows: set up ps and tj Distance is , pe and tj Distance is The position difference is the sum of the above two distances, that is:

[0045] ps and tj The difference in heading Calculate as follows:

[0046] in, Represents trajectory points The heading difference is calculated as follows:

[0047] ps and tj Speed ​​difference Calculate as follows:

[0048] in, Represents trajectory points The speed difference is calculated as follows:

[0049] ps and tj Time difference Calculate as follows:

[0050] in, Represents trajectory points The timestamp difference.

[0051] If the target ship is tj If the corresponding ship types are the same, the ship type difference 0 if yes, 1 otherwise. For target ships and tj The difference in length of the corresponding ship. For target ships and tj The difference in width between the target ship and tj If the corresponding ships have the same nationality, the difference in ship registration is 0 if yes, otherwise it is 1.

[0052] e. Calculation Representative trajectory of , to determine the specific route of the target ship. The representative trajectory refers to a trajectory that can represent the overall motion pattern of the group trajectory. The present invention adopts a simple and effective method to determine the representative trajectory. The specific method is: traverse For each trajectory of The maximum DTW distance of other trajectories, and finally select the trajectory with the smallest maximum DTW distance as the representative trajectory, that is:

[0053] in, for t and distance.

[0054] f. Generate the completion part of the target trajectory according to the representative trajectory , and the corrected trajectory is obtained.

[0055] contrast Figure 3 It can be seen from (a) and (b) that the corrected trajectory is more consistent with the law of ship motion and its geometric features are closer to the real trajectory, which verifies that the data-driven ship trajectory correction algorithm has certain effectiveness.

[0056] S202, execute the ship trajectory compression algorithm based on speed perception and the discretization of ship attributes. It includes two sub-steps: the ship trajectory compression algorithm based on speed perception and the discretization of ship attributes. The specific steps are as follows: The speed-aware ship trajectory compression algorithm is based on the classic Douglas-Peuker (DP) compression algorithm. In the DP compression algorithm, only the direction of the trajectory is considered, but not the speed information of the trajectory. Therefore, the track processed by the DP compression algorithm will lose a lot of speed information, which will affect the accuracy of subsequent abnormal speed detection of ships.

[0057] The speed-aware ship trajectory compression algorithm effectively retains the speed information of the track while reducing the size of the trajectory data. The basic idea is to find those trajectory points that are considered to have undergone a sudden change in speed and set them as trajectory feature points. The specific steps are as follows: Step (1) calculates the distance from the target ship's trajectory from the next trajectory point after the starting point to the previous trajectory point of the end point to the line segment connecting the starting point and the end point of the target ship's trajectory, and obtains the maximum distance d max and corresponding trajectory points p max. The distance from the trajectory point to the line segment is the Euclidean distance; Step (2) If d If max is greater than the threshold, execute step (3), otherwise execute step (4); Step (3): Determine p max is the feature point, and takes the starting point as p max is the new starting point 1 and the new end point 1. p max, the end point is the new starting point No. 2, the new end point No. 2, and the algorithm is executed recursively; Step (4), take the starting point as the anchor point and the point next to the starting point as the current point; Step (5) Calculate the speed difference between the current point and the anchor point . Track Point The speed difference is expressed as .in, Represents trajectory points p Speed ​​over ground; Step (6) If If the value is less than the threshold, execute step (7); otherwise, execute step (8); Step (7): If the current point is the previous point of the end point, the algorithm ends or jumps out of the current recursion, otherwise, execute step (9); Step (8), taking the current point as the feature point and anchor point; Step (9), taking the next point of the current point as the new current point, execute step (5).

[0058] This embodiment sets the distance and heading threshold Speed ​​threshold .

[0059] Figure 4 (a) shows an original ship trajectory, which contains 9 track points. First, according to the DP compression algorithm process, calculate p 2 to p 8 to line segment distance and get the point p Maximum distance of 5 d max, found , then determine p 5 is the feature point ( Figure 4 Next, we transform the trajectory from p 5 is divided into two sub-trajectories, front and back. First, process the front half of the sub-trajectory. According to the DP compression algorithm process, p 2 to Maximum distance d max, and have ( Figure 4 (c) in the above example). At this time, calculate the p 2, p 3 to the anchor point (currently p 1) The speed difference, and and , then determine p 2 is a non-feature point, p 3 is the feature point and new anchor point ( Figure 4 Next, calculate p 4 To the new anchor point p 3 speed difference, and there is , so p 4 is set as a non-feature point ( Figure 4Then process the second half of the trajectory. According to the DP compression algorithm process, we get p 7 to Maximum distance d max, and have ( Figure 4 Calculate p 6 to the anchor point p 5 speed difference, and , so p 6 is set as the feature point, and the current anchor point is updated to p 6 ( Figure 4 Finally, calculate With new anchor p 6, and determine the speed based on the calculation results p 7 is a non-feature point, p 8 is the feature point ( Figure 4 (h) in the figure).

[0060] This embodiment uses compression rate, DTW distance and average speed error as evaluation indicators, and takes the DP compression algorithm as the benchmark method to verify the performance of the speed-aware ship trajectory compression algorithm.

[0061] The compression rate is used to evaluate the degree to which the algorithm compresses the trajectory data. To and the original trajectory Tc , compression ratio Rc Defined as follows:

[0062] in, Represents the number of trajectory points of the compressed trajectory and the original trajectory respectively. The larger the compression rate, the fewer the number of trajectory points after compression and the higher the compression degree.

[0063] The average speed error is used to evaluate the ability of the ship trajectory compression algorithm to retain trajectory speed information and is calculated as follows:

[0064] in, n represents the number of track points of the ship track before compression; Represents the number of feature points of the compressed ship trajectory; The compressed ship trajectory i The sequence number of the feature points in the original trajectory; pi Indicates the original trajectory i Track points.

[0065] observe Figure 5From (a) and (b) in Figure 1, we can see that the upper and lower quartiles of the compression rate of the speed-aware ship trajectory compression algorithm are slightly lower than those of the DP compression algorithm, and the DTW distance of the trajectory before and after compression of the speed-aware ship trajectory compression algorithm is also slightly smaller than that of the DP compression algorithm. This result is basically consistent with expectations. Because the speed-aware ship trajectory compression algorithm considers the speed of the trajectory point during the compression process, some additional feature points are added on the basis of the results obtained by the DP compression algorithm, resulting in a lower compression rate and a smaller DTW distance.

[0066] Depend on Figure 5 As shown in (c), the most obvious difference between the two algorithms is the average speed error. The average speed error of the speed-aware ship trajectory compression algorithm is significantly lower than that of the DP compression algorithm in terms of both the median and the upper and lower quartiles. Since the speed-aware ship trajectory compression algorithm retains the trajectory points where the speed changes suddenly by calculating the speed difference between the speed points during the compression process, it retains most of the speed information of the original trajectory and achieves a lower average speed error.

[0067] Furthermore, the discretization of ship attributes is used to improve the generalization performance of subsequent detection models. Different ship attributes all have the problem of too small attribute granularity, which leads to a decrease in the generalization ability of the detection model and ultimately affects the detection accuracy. The present invention uses a discretization method to extract six high-level attributes of the ship, namely, time, position, heading, speed, length, and width, to increase the attribute granularity and improve the model training effect and generalization performance.

[0068] S203, constructing a ship behavior feature sequence. The specific steps are as follows: refer to Figure 6 ,The ship behavior feature sequence includes the length, width, registry, ship type, and the ship's current longitude, latitude, ground course, ground speed, and timestamp. The length of a ship behavior feature sequence is greater than or equal to 7. The first three elements of the sequence represent the ship's registry, length, and width corresponding to the ship's trajectory, and the remaining elements represent the position, time, course, and speed corresponding to the ship's trajectory feature points.

[0069] S204, generate ship behavior feature sequence embedding. The specific steps are as follows: The method of directly using the original ship behavior feature sequence as the model input and label to train and test the model has certain defects. The original ship data can hardly ensure that the model can deeply explore the complex semantic information and internal connections hidden behind the ship behavior. The potential dependency between these features often affects the specific performance of the ship behavior.

[0070] Drawing on the word embedding technology widely used in the field of natural language processing, the original ship feature data is encoded using the one-hot encoding method and embedding matrix to obtain an embedding vector that is easy for the model to understand.

[0071] Furthermore, the unique hot coding technique is used to uniformly encode different ship characteristics. Specifically, suppose The unique hot vector of the ship feature is ,in Indicates k The one-hot vector with dimension 1, then the unified one-hot vector can be expressed as ,in, .

[0072] After uniformly one-hot encoding all the features of the ship, the uniform one-hot vector With the embedding matrix Multiply to get time step t Embed :

[0073] in, is the kth column vector of M; M participates in the training as a model parameter.

[0074] S205, encode the ship behavior feature sequence, generate the context vector and decode it, and introduce the attention mechanism. The specific steps are as follows: refer to Figure 7 The Seq2Seq encoder is designed to use a recurrent neural network to embed the ship behavior feature sequence elements input in time steps and finally convert them into a vector with a fixed dimension. Formally speaking, the Seq2Seq encoder embeds the input sequence into a vector using a recurrent neural network composed of LSTM. Encode in element order and finally generate a context vector ,Right now:

[0075] in, Represents an LSTM network; Indicates s The hidden state of the encoder at time steps summarizes the previous state of the sequence. s Information about each element; represents the network parameters of the encoder. C is the last moment S The hidden state of the encoder is the vector that can be regarded as the encoding of the entire ship behavior feature sequence, which can reflect the intrinsic structural information of the ship behavior to a certain extent.

[0076] set up s The hidden state of the LSTM unit of the decoder at time -1 is , Ship behavior characteristics sequence No. s The embedding of -1 element is , then s Hidden state of the moment decoder for:

[0077] in, represents the decoder network parameters, Represents the attention vector. The hidden state at time 0 , the embedding of the 0th element The termination symbol .

[0078] Furthermore, the decoder’s hidden state It is not only used as the network history information for calculating the hidden state of the next time step, but also used to calculate the output of the current time step. s Output of time steps It is expressed as:

[0079] in, Indicates a feedforward neural network with the last layer being the Softmax function; are the parameters of the feedforward neural network; The posterior probability of each value of different ship features. The decoder input contains the ship feature embedding output by the model at the previous moment , and this vector is the output of the decoder at the previous moment It is determined by the component with the largest posterior probability.

[0080] By introducing the attention mechanism to consider the degree of attention to the hidden states of the encoder at different time steps, the model can overcome the problem of being unable to establish long-interval dependencies of ship features due to excessive sequence information of ship behavior features.

[0081] To calculate the hidden state of the encoder at each time step For the contribution during decoding, query vector and scoring are introduced Function. In this model, the query vector is the hidden state of the decoder at the previous moment The scoring function is used to calculate the attention distribution of the hidden state of the encoder at each time step , the specific calculation formula is as follows:

[0082] Among them, the attention scoring function adopts the dot product model, that is:

[0083] The dot product model is used as the attention scoring function because it works in the form of vector product, which can give full play to the computing power of the GPU. The attention distribution is normalized based on the Softmax function, which measures the degree of attention paid by the current decoder hidden state to the encoder hidden state at each moment.

[0084] After obtaining the attention distribution of each hidden state, the distribution is regarded as a weight and the weighted average of the hidden states at all times is performed, that is:

[0085] in, It is called the attention vector, which is the summary of the attention information obtained from the input sequence.

[0086] The integration of the attention mechanism for the input ship behavior feature sequence can fully capture the key historical behavior information of the ship, which is of great significance for improving the model's ability to understand the input data and the detection performance.

[0087] S206, using the BLEU algorithm to calculate the anomaly score. The ship behavior feature sequence is used as a label of Seq2Seq, and the BLEU algorithm is used to calculate the anomaly score of the ship behavior feature sequence. The specific steps are as follows: In the Seq2Seq model, the cross entropy loss function is used for model training. The loss function is defined as follows:

[0088] in, represents the one-hot vector sequence corresponding to the input sequence of ship features, Output sequence for the model.

[0089] The Seq2Seq model uses the BLEU algorithm to calculate the anomaly score of the ship behavior feature sequence. is the candidate sequence consisting of elements with the maximum posterior probability in the output distribution of this model, and the reference sequence is the input sequence of the model , then the N-tuple combination accuracy can be expressed as:

[0090] in, Represents the candidate sequence All N-tuple combinations generated; and Represents N-tuple combinations exist and The number of occurrences in .

[0091] The BLEU score is obtained by weighted summing the length and precision of different N-gram combinations, namely:

[0092] in, Indicates the length of the longest N-tuple combination; is the weight of the length of the N-tuple combination. The range of BLEU scores is The higher the score, the more similar the ship behavior feature sequence generated by the model is to the input sequence, and the lower the abnormality of the ship behavior, and vice versa. The embedding vector dimension and the parameter settings of this model are shown in Table 1, including the settings of the embedding vector dimension, LSTM unit dimension, N-ary combination weight and other parameters: Table 1: Embedding vector dimensions and parameter settings of this model

[0093] This embodiment uses precision, recall, and F1-Score as evaluation indicators of the method.

[0094] This example selects Seq2Seq-RNN, GeoTrackNet, and MADVCN as the benchmark methods. Table 2 shows the experimental results of Seq2Seq, GeoTrackNet, MADVCN, and Seq2Seq-RNN: Table 2: Evaluation metrics of different models

[0095] It can be seen that Seq2Seq surpasses the other three methods in terms of precision, recall and F1 score, which verifies the excellent performance of the present invention.

[0096] In summary, the present invention aims at the problems of poor accuracy and low efficiency in identifying abnormal behaviors of ships caused by the limited ability to extract ship behavior features in previous methods, and proposes a method for identifying abnormal behaviors of ships based on a sequence-to-sequence model. First, a speed-aware ship trajectory compression algorithm is designed to effectively filter out redundant trajectory points, effectively retain the speed information of the track while significantly reducing the data scale, and then discretize the ship attributes to improve the data granularity, improve the model training effect and generalization performance; then, by combining the static and dynamic characteristics of the ship, the embedding technology is used to deeply mine the potential information of the ship's behavior characteristics, and the encoding and decoding of the ship's trajectory features are realized based on the Seq2Seq model and LSTM network to understand the ship's behavior pattern; finally, the attention mechanism is introduced in Seq2Seq, so that the model can capture the hidden information of the ship's behavior characteristics and effectively improve the detection accuracy. The present invention successfully realizes the automatic identification of abnormal behaviors of ships in water traffic supervision.

[0097] like Figure 8 As shown, the present invention also provides a ship abnormal behavior identification device 800, comprising: The track extraction module 801 is used to extract the ship track based on the AIS data; A feature extraction module 802 is used to compress the ship trajectory and discretize the ship attributes corresponding to the ship trajectory to obtain ship behavior feature data; A sequence generation module 803 is used to generate a ship behavior feature sequence embedding amount based on the ship behavior feature data; The information recognition module 804 is used to input the embedding amount of the ship behavior feature sequence into the trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; The anomaly determination module 805 is used to determine the anomaly score of the embedding amount of the ship behavior feature sequence, and obtain the ship abnormal behavior identification result based on the anomaly score and the ship behavior identification information.

[0098] The ship abnormal behavior identification device provided in the above embodiment can implement the technical solution described in the above ship abnormal behavior identification method embodiment. The specific implementation principles of the above modules or units can refer to the corresponding contents in the above ship abnormal behavior identification method embodiment, which will not be repeated here.

[0099] like Fig. 9 As shown, the present invention also provides an electronic device 900. The electronic device 900 includes a processor 901, a memory 902 and a display 903. Fig. 9 Only some components of the electronic device 900 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0100] In some embodiments, the memory 902 may be an internal storage unit of the electronic device 900, such as a hard disk or memory of the electronic device 900. In other embodiments, the memory 902 may also be an external storage device of the electronic device 900, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc., equipped on the electronic device 900.

[0101] Furthermore, the memory 902 may include both an internal storage unit of the electronic device 900 and an external storage device. The memory 902 is used to store application software installed in the electronic device 900 and various data.

[0102] In some embodiments, the processor 901 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run program codes or process data stored in the memory 902, such as the ship abnormal behavior identification method of the present invention.

[0103] In some embodiments, the display 903 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 903 is used to display information on the electronic device 900 and to display a visual user interface. The components 901-903 of the electronic device 900 communicate with each other via a system bus.

[0104] In some embodiments of the present invention, when the processor 901 executes the ship abnormal behavior identification program in the memory 902, the following steps may be implemented: Extract ship tracks based on AIS data; Compressing the ship trajectory and discretizing the ship attributes corresponding to the ship trajectory to obtain ship behavior characteristic data; Based on the ship behavior characteristic data, generating a ship behavior characteristic sequence embedding amount; The embedding amount of the ship behavior feature sequence is input into a trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; An abnormal score of the embedding amount of the ship behavior feature sequence is determined, and a ship abnormal behavior recognition result is obtained based on the abnormal score and the ship behavior recognition information.

[0105] It should be understood that: when the processor 901 executes the ship abnormal behavior identification program in the memory 902, in addition to the above functions, other functions can also be implemented. For details, please refer to the description of the corresponding method embodiment above.

[0106] Furthermore, the embodiment of the present invention does not specifically limit the type of the electronic device 900 mentioned, and the electronic device 900 may be a portable electronic device such as a mobile phone, a tablet computer, a personal digital assistant (PDA), a wearable device, a laptop computer, etc. Exemplary embodiments of portable electronic devices include but are not limited to portable electronic devices equipped with IOS, Android, Microsoft or other operating systems. The above-mentioned portable electronic device may also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, the electronic device 900 may not be a portable electronic device, but a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0107] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for identifying abnormal behavior of a ship provided by the above methods is implemented, and the method comprises: Extract ship tracks based on AIS data; Compressing the ship trajectory and discretizing the ship attributes corresponding to the ship trajectory to obtain ship behavior characteristic data; Based on the ship behavior characteristic data, generating a ship behavior characteristic sequence embedding amount; The embedding amount of the ship behavior feature sequence is input into a trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; An abnormal score of the embedding amount of the ship behavior feature sequence is determined, and a ship abnormal behavior recognition result is obtained based on the abnormal score and the ship behavior recognition information.

[0108] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0109] The above is a detailed introduction to the ship abnormal behavior identification method and device provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for identifying abnormal behavior of a ship, characterized in that: include: Extract ship tracks based on AIS data; Compressing the ship trajectory and discretizing the ship attributes corresponding to the ship trajectory to obtain ship behavior characteristic data; Based on the ship behavior characteristic data, generating a ship behavior characteristic sequence embedding amount; The embedding amount of the ship behavior feature sequence is input into the trained sequence-to-sequence network model to obtain the ship behavior recognition information; The network model includes an encoder and a decoder; An abnormal score of the embedding amount of the ship behavior feature sequence is determined, and a ship abnormal behavior identification result is obtained based on the abnormal score and the ship behavior identification information.

2. The method for identifying abnormal ship behavior according to claim 1, characterized in that: Based on the AIS data, the ship track is extracted, including: Filtering target attributes from the AIS data; Based on the target attribute, the AIS data is grouped according to the MMSI code to obtain grouped data; Sorting each set of data in the plurality of sets of data in the order of timestamps to obtain an original ship track; wherein each set of data corresponds to an original ship track; Segment the trajectories whose temporal and spatial intervals exceed a preset interval threshold among the multiple original ship trajectories to obtain preliminary extracted ship trajectories; The initially extracted ship trajectory is corrected to obtain the final ship trajectory.

3. The method for identifying abnormal ship behavior according to claim 2, characterized in that: Target attributes are screened from the AIS data, including: Selecting multiple basic attributes from the AIS data; the multiple basic attributes include: ship flag, ship type, ship length, ship width, MMSI code, timestamp, longitude, latitude, ground heading and ground speed; When it is determined that the basic attribute is a static attribute and there is an error in the static attribute, a preset ship information database is queried based on the MMSI code of the erroneous data in the static attribute to obtain correct data, and the erroneous data is replaced based on the correct data to obtain the target attribute; When it is determined that the basic attribute is a dynamic attribute and there is an error in the dynamic attribute, a corresponding data repair method is determined based on the error type of the dynamic attribute to repair the dynamic attribute to obtain a target attribute.

4. The method for identifying abnormal ship behavior according to claim 2, characterized in that: The initially extracted ship trajectory is corrected to obtain the final ship trajectory, including: Select two endpoints to be corrected from the initially extracted ship trajectory, and determine the neighborhoods corresponding to the two endpoints respectively; Filtering the ship trajectories passing through two neighborhoods from the initially extracted ship trajectories; Add the sub-trajectory of the screened ship trajectory between two neighborhoods to the candidate trajectory set; Clustering the candidate trajectory set to obtain a candidate trajectory class cluster set; Determine the support of each cluster in the candidate trajectory cluster set, and determine the cluster with the largest support; The representative trajectory of the cluster with the largest support is determined, and the initially extracted ship trajectory is corrected based on the representative trajectory to obtain the final ship trajectory.

5. The method for identifying abnormal ship behavior according to claim 1, characterized in that: The ship trajectory is compressed, and the ship attributes corresponding to the ship trajectory are discretized to obtain ship behavior characteristic data, including: Determine the distance from each trajectory point in the target trajectory point set to the line segment connecting the starting point and the end point of the target trajectory point set as the first type distance, and determine the maximum distance from the first type distances corresponding to each trajectory point as the first target distance; the target trajectory point set is a set of trajectory points on the ship trajectory; When the first target distance is greater than a preset distance threshold, determining the trajectory point corresponding to the first target distance as a feature point, and dividing the target trajectory point set into two target trajectory point subsets based on the trajectory points corresponding to the first target distance; Determine the distance between each trajectory point in the target trajectory point subset and the line segment connecting the starting point and the end point of the target trajectory point subset as the second type distance, and determine the maximum distance from the second type distances corresponding to each trajectory point as the second target distance; When the second target distance is less than or equal to the preset distance threshold, determine the anchor point at the current moment, and calculate the speed difference from the track point behind the anchor point at the current moment to the anchor point, and determine the track point whose speed difference is greater than the preset distance threshold as the anchor point and feature point at the next moment, until all feature points in the target track point set are determined; The ship attributes of all feature points in the target trajectory point set are discretized to obtain ship behavior feature data.

6. The method for identifying abnormal ship behavior according to claim 5, characterized in that: The anchor point at the initial moment is the starting point of the target trajectory point subset.

7. The method for identifying abnormal ship behavior according to claim 1, characterized in that: Based on the ship behavior feature data, generating a ship behavior feature sequence embedding amount includes: One-hot encoding is performed on the ship behavior characteristic data to obtain a corresponding one-hot vector; The one-hot vector is multiplied by the embedding matrix to obtain the embedding amount of the ship behavior feature sequence.

8. The method for identifying abnormal ship behavior according to claim 1, characterized in that: Determining an anomaly score of the embedding amount of the ship behavior feature sequence includes: The embedding amount of the ship behavior feature sequence is used as a label of the network model, and based on the BLEU algorithm and the label, the abnormal score of the embedding amount of the ship behavior feature sequence is calculated.

9. The method for identifying abnormal ship behavior according to any one of claims 1 to 8, characterized in that: The embedding amount of the ship behavior feature sequence is input into the trained sequence-to-sequence network model to obtain the ship behavior recognition information, including: Inputting the ship behavior feature sequence embedding amount into the encoder, so as to encode the ship behavior feature sequence embedding amount according to the element order based on the encoder to generate a context vector with a fixed dimension; The context vector is input into the decoder to obtain ship behavior recognition information; the decoder is added with an attention mechanism.

10. A device for identifying abnormal behavior of a ship, characterized in that: include: Track extraction module, used to extract ship tracks based on AIS data; A feature extraction module, used to compress the ship trajectory and discretize the ship attributes corresponding to the ship trajectory to obtain ship behavior feature data; A sequence generation module, used for generating a ship behavior feature sequence embedding amount based on the ship behavior feature data; An information recognition module is used to input the embedding amount of the ship behavior feature sequence into a trained sequence-to-sequence network model to obtain ship behavior recognition information; the network model includes an encoder and a decoder; The anomaly determination module is used to determine the anomaly score of the embedding amount of the ship behavior feature sequence, and obtain the ship abnormal behavior identification result based on the anomaly score and the ship behavior identification information.

Images