Interpretable two-stage ship trajectory prediction method based on historical navigation pattern mining

Abstract

The present application relates to a two-stage ship trajectory prediction method based on the explainability of historical navigation mode mining, including two stages of feasible candidate trajectory set generation and candidate trajectory discrimination, the potential navigation environment and ship motion characteristic information in AIS data can be effectively extracted by mining the navigation mode in the historical AIS data of the ship; the navigation mode is applied to the current trajectory of the ship to realize the generation of the feasible candidate trajectory set of the ship based on the current navigation state and the historical behavior mode of the ship, so that the trajectory generation process and result have strong explainability; the correlation between the current trajectory of the ship and the generated trajectory is learned by using a neural network, and discrimination and selection are performed, which can effectively ensure the accuracy of the trajectory prediction model. The present application effectively fuses the ship navigation mode information and the characteristics of the current trajectory of the ship, the prediction result has good explainability, and provides technical support for ship behavior analysis, port safety supervision and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship trajectory intelligent prediction, and more particularly to an interpretable two-stage ship trajectory prediction method based on historical navigation pattern mining. BACKGROUND

[0002] Ship trajectory has the characteristics of obvious motion space freedom, complex and changeable environment, and many navigation features, and is difficult to predict.

[0003] Traditional ship trajectory prediction methods usually use ship kinematics models to predict the micro-motion behavior of ships. Trajectory prediction models are mainly divided into traditional mathematical statistical models and data-driven neural network models. However, the traditional kinematics model is easily affected by the error of the original data, and the model needs to be modified according to the specific scene and different ships. It is difficult to model in a complex environment, and the precision is low. Neural networks are widely used in ship trajectory prediction research due to their good nonlinear fitting ability and prediction accuracy. However, the black-box nature of neural networks makes the predicted trajectory obtained by only using neural networks lack interpretability.

[0004] Therefore, how to balance the prediction accuracy of the model and the interpretability of the prediction result needs further study. SUMMARY

[0005] In view of the above shortcomings of the prior art, the present application provides an interpretable two-stage ship trajectory prediction method based on historical navigation pattern mining. First, the AIS data is preprocessed, and then the historical navigation pattern mining is used to iteratively sample the feasible candidate trajectory set. Then, the nonlinear fitting ability of the neural network is used to select the feasible candidate trajectory, and the historical navigation pattern is used to generate an interpretable ship trajectory to compensate for the lack of interpretability of the neural network model.

[0006] To achieve the above purpose, the present application provides an interpretable two-stage ship trajectory prediction method based on historical navigation pattern mining, comprising:

[0007] (1) preprocessing the obtained AIS data to obtain the trajectory data set required by the model;

[0008] (2) in the feasible candidate trajectory set generation stage, the historical navigation pattern is mined to calculate the navigable range. The current motion state of the ship and the motion state of the next time of the ship sampled by the historical navigation pattern are used to iteratively generate an interpretable feasible candidate trajectory set by using kinematics formula;

[0009] (3) In the candidate trajectory discrimination stage, the LSTM network is used to encode the current trajectory of the ship and the set of feasible predicted trajectories respectively, the encoding results are spliced, and then the scoring network is used to calculate the score of each feasible trajectory in the set of feasible candidate trajectories, and the similarity between the label trajectory and the set of feasible candidate trajectories is used as the scoring label to realize training;

[0010] (4) In the prediction stage, the feasible trajectory with the highest score output by the scoring network is taken as the prediction result.

[0011] Further, the data preprocessing includes:

[0012] Extracting the data of the ship at sea and separating the data of different voyages of different ships, cleaning the repeated, missing key attributes and error data;

[0013] Resampling each trajectory so that the time intervals of the trajectory points on each trajectory are the same;

[0014] First-order difference is made on the latitude, longitude, speed and heading attributes of the trajectory points and the corresponding attributes of the previous time;

[0015] Selecting the first preset number of consecutive trajectory points as the model input trajectory, and the second preset number of consecutive trajectory points after that as the label trajectory, and using the sliding window method to sample the trajectory segment at a predetermined sampling interval to obtain the trajectory data set.

[0016] Further, the set of feasible candidate trajectories is generated, including:

[0017] For a given arbitrary current trajectory of a ship, all adjacent points within a predetermined spatial range of the candidate starting point are queried from the trajectory data set as a set of historical reference trajectory starting points, and the second preset sampling amount of trajectory points is obtained in time sequence from all trajectory points in the starting point set to form a set of historical reference trajectories.

[0018] Mining historical navigation patterns from the set of historical reference trajectories;

[0019] Based on the historical navigation patterns, iteratively generate a set of feasible candidate trajectories starting from the predicted starting point.

[0020] Further, the historical navigation patterns are mined from the set of historical reference trajectories, including:

[0021] Based on all historical reference trajectories, the angle range of the trajectory point with respect to the starting point heading is calculated, and the maximum cumulative heading change and the maximum cumulative speed change of the trajectory point are calculated using all reference trajectories.

[0022] Calculate the maximum heading change of adjacent points using all trajectory points, and bin the speed of all trajectory points to obtain the speed increment distribution law and the speed decrement distribution law of each bin; Calculate the distribution law of the speed change state of adjacent points, including deceleration, constant speed and acceleration; Calculate the transition probability of the speed change state of adjacent points; Calculate the transition probability of the heading change state of adjacent points, including straight ahead, left turn and right turn.

[0023] Further, based on the historical navigation mode, iteratively generate a set of feasible candidate trajectories starting from the predicted starting point, including:

[0024] Limit the navigable range based on the angle range in the historical navigation mode, and iteratively generate trajectory points by the following steps:

[0025] Determine the heading change state based on the heading change of the current trajectory point, sample the next time point heading change state using the transition probability of the heading change state of adjacent points, generate N next time point headings based on the heading of the current trajectory point and the maximum heading change of adjacent points, and select the next time point feasible heading based on the maximum cumulative heading change of the trajectory point;

[0026] Determine the speed change state based on the speed change of the current trajectory point, sample the next time point speed change state using the transition probability of the speed change state of adjacent points; Select the appropriate speed change distribution law based on the speed of the current trajectory point and the next time point speed change state to sample the next time point speed change, limit the value of the speed change based on the maximum cumulative speed change, and calculate the speed of the next time point trajectory point;

[0027] Combine the heading and speed of the next time point obtained by the above steps with the current trajectory point speed vector to calculate the motion distance in the latitude and longitude direction of the ship at the adjacent time based on the uniform variable speed straight line motion, and convert the motion distance in the latitude and longitude direction into the latitude and longitude change to obtain the coordinates of the next time point trajectory point. If the next time point trajectory point is generated outside the navigable range, delete the entire trajectory corresponding to the point;

[0028] Iterate the above steps to generate several trajectories, and randomly sample the several trajectories to obtain a set of feasible candidate trajectories.

[0029] Further, the candidate trajectory discrimination includes:

[0030] Encode the features of the model input trajectory using the first LSTM network, and use the hidden layer state h L After dimension reduction by flatten and fully connected layer, the ship current trajectory encoding result h L* ;

[0031] The features of each feasible candidate trajectory in the feasible candidate trajectory set are encoded by using a second set of LSTM networks, and the hidden layer state h K Same as h L The feasible candidate trajectory encoding h Ki* Where i represents the i-th feasible candidate trajectory.

[0032] The two sets of encoding results are spliced to obtain a complete trajectory encoding.

[0033] The complete trajectory encoding result is input into a scoring network, and the scoring network outputs a scoring vector of the feasible candidate trajectory set through a softmax layer. In the training stage, a hybrid loss is calculated by using the scoring vector of the feasible candidate trajectory set and the similarity between the feasible candidate trajectory and the label trajectory to realize the training of the trajectory encoder and the scoring network.

[0034] Further, the hybrid loss function is composed of a cross-entropy loss and an L2 loss.

[0035] The advantages and beneficial effects of the present application are as follows:

[0036] The present application can effectively extract potential navigation environment and ship motion feature information in AIS data by mining the navigation mode in historical AIS data of the ship. The present application applies the navigation mode to the current trajectory of the ship to realize the generation of the feasible candidate trajectory set of the ship based on the current navigation state and the historical behavior mode of the ship, so that the trajectory generation process and result have strong interpretability, and the problem of insufficient interpretability of a single neural network model is overcome. Secondly, the present application learns the correlation between the current trajectory of the ship and the generated trajectory by using a neural network and discriminates and selects, which can effectively ensure the accuracy of the trajectory prediction model. Finally, the two-stage ship trajectory prediction method adopted by the present application effectively fuses the ship navigation mode information and the features of the current trajectory of the ship, and the prediction result has good interpretability, which provides technical support for ship behavior analysis, port safety supervision and the like. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0038] Figure 1 The method flowchart in the embodiment of the present application;

[0039] Figure 2 The ship trajectory generation stage flowchart in the embodiment of the present application;

[0040] Figure 3 This is a schematic diagram of historical trajectory data extraction in an embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram of the discriminator model for the ship trajectory discrimination stage in an embodiment of the present invention. Detailed Implementation

[0042] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0043] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0044] like Figure 1 The diagram illustrates the overall flowchart of a two-stage ship trajectory prediction method based on historical navigation pattern mining, as described in an embodiment of the present invention. The method comprises two stages: a feasible candidate trajectory set generation stage and a trajectory discrimination stage. The feasible candidate trajectory set generation stage involves AIS data preprocessing and generating feasible candidate ship trajectories. The trajectory discrimination stage involves trajectory discrimination model training and ship trajectory prediction. Specifically, it includes the following steps:

[0045] S1. Acquire data on ships in transit and preprocess the acquired data to obtain a trajectory dataset;

[0046] Among them, the data of ships under navigation can be obtained from the dynamic data of AIS (Automatic Identification System, which was developed from the IFF (Identification Friend or Foe) of ships and aircraft. It works with the Global Positioning System (GPS) to broadcast ship dynamics such as position, speed, rate of change of course and course, combined with ship static data such as name, call sign, draft and dangerous goods, to ships in nearby waters and shore stations via VHF channels. This allows nearby ships and shore stations to keep abreast of the dynamic and static information of all ships in the nearby sea, enabling them to communicate and coordinate immediately and take necessary avoidance actions, which greatly helps ship safety).

[0047] Data preprocessing specifically includes the following steps:

[0048] S11. Data Cleaning. First, navigation data of vessels is extracted from AIS dynamic data using speed thresholds. Then, navigation data of different vessels is divided using MMSI (Maritime Mobile Service Identifier, a nine-digit code transmitted by a ship's radio communication system on its radio channel to uniquely identify various stations and groups of calling stations). Different voyages are divided based on trajectory point time interval thresholds. Overlapping trajectory points, trajectory points missing key attributes (speed, heading, latitude and longitude coordinates), and trajectory points with erroneous attribute data obtained by threshold discrimination are deleted from the trajectory data of each voyage. Each trajectory data is resampled using interpolation to ensure that the time interval is Δt. For adjacent trajectory points with heading changes less than a set threshold, linear interpolation is used; otherwise, Hermitian interpolation is used.

[0049] S12. Calculate the changes in latitude and longitude, and changes in speed and heading of each trajectory point in each segment of trajectory data compared to the previous moment. Then, the characteristics of the ship's trajectory point at each moment can be expressed as: p = {LAT} t ,LON t SOG t COG t ,ΔLAT t ,ΔLON t ,ΔSOG t ,ΔCOG t}, LAT t Latitude, LON t For longitude, SOG t For speed, COG t For heading, ΔLAT t ΔLON represents the change in latitude. t ΔSOG is the change in longitude. t For the change in speed, ΔCOG t This represents the change in heading.

[0050] S13, construct a trajectory dataset Tra_S, each trajectory sample tra_s in the trajectory dataset is composed of two parts: L continuous time trajectory points as input trajectory (original trajectory), and K continuous time trajectory points as label trajectory, that is, the original trajectory contains a ship trajectory of LxAt length, and the label trajectory contains a ship trajectory of KxAt length. The features used for input trajectory data are P in ={SOG t , COG t , ALAT t , ALON t}, and the features used for label trajectory are P pre ={ALAT t , ALON t}, and the trajectory sample is obtained by sampling from each voyage trajectory data at a certain sampling interval using a sliding window method to form the trajectory dataset Tra_S.

[0051] S2, generate a feasible candidate trajectory set based on the ship motion state through historical navigation pattern mining and kinematic model.

[0052] The trajectory generation process is shown in Figure 2 , and specifically includes the following steps:

[0053] S21, for a given arbitrary ship current trajectory, query all adjacent points within a certain spatial range from Tra_S as historical reference trajectory starting point set his_p_S, and further sample K continuous time trajectories from each trajectory point in the starting point set his_p_S in time sequence to form a historical reference trajectory set his_Tra_S, as shown in Figure 3 The query space range is a spatial rectangle with the short side parallel to the predicted starting point heading direction, and the extracted trajectory data is a trajectory segment containing K+1 trajectory points with the starting point in the query space range.

[0054] S22, mine the historical navigation pattern from the obtained his_Tra_S. Calculate the angle range of the trajectory point with respect to the starting point heading vector based on all historical reference trajectories , where the variables represent the maximum angle of the trajectory point connection with the starting point heading in the counterclockwise direction and the clockwise direction, respectively; calculate the maximum cumulative heading change and the maximum cumulative speed change , which represents the maximum value of the cumulative change of ship speed or heading in the time dimension in all trajectory segments, used to limit the ship motion state change in subsequent trajectory generation; calculate the maximum heading change of the adjacent point using the trajectory point speed, and count the speed increment distribution law and the speed decrement distribution law PΔSOG+ , P ΔSOG- ; utilize ASOG t attribute statistics K time corresponding to the speed change state distribution law P k,ΔSOG = {p dec , p 0 , p acc}, the corresponding k time ship deceleration, uniform speed, acceleration probability; and statistics adjacent point speed change state transition probability represent the ship in adjacent time ship speed change in deceleration, uniform speed, acceleration state transition probability; utilize ACOG t attribute statistics adjacent point heading change direction state transition probability

[0055] S23, from the predicted starting point to generate a feasible candidate trajectory set. First based on limit the navigable range R, through the following steps to generate trajectory points iteratively:

[0056] based on the current trajectory point pcurACOG cur by threshold method to determine the heading change state S ΔCOG,cur , using calculate the conditional probability P (S ΔCOG,next |S ΔCOG,cur ) and to sample the next time heading change state, that is, to determine whether the ship next time to keep straight or turn left or right, if keep straight based on heading change state threshold to generate N next time heading, if left or right turn based on COG cur and generate N next time point heading to filter the next time feasible heading;

[0057] based on the current trajectory point ASOG cur by threshold method to determine the speed change state S ΔSOG,cur , using P ΔSOG and P k,ΔSOG calculate the conditional probability P (S ΔSOG,next |S ΔSOG,cur ) and sample to obtain the next time point speed change state S ΔSOG,next ; based on the current trajectory point speed SOG cur and S ΔSOG,next select the appropriate speed change distribution law and each COG next sample to select the speed change, to limit the value of the speed change, if the selected speed change makes the cumulative speed change exceeds then resample, finally calculate the next time trajectory point corresponding to the speed SOG next;

[0058] with COG cur and the obtained COG next SOG cur and SOG next decompose into latitude and longitude directions, calculate AL using the decomposed speed and the set time interval At in latitude and longitude directions lon and AL lat convert the distance into latitude and longitude direction increments to obtain the latitude and longitude coordinates pos of the next moment next fuse the attributes to finally obtain the trajectory point p of the next moment next , if p next is generated outside R, it is explained that the generated trajectory point does not conform to the historical navigation mode, and the entire trajectory corresponding to the point is deleted;

[0059] iteratively generate M init trajectories, randomly sample M init trajectories to obtain a set of M feasible candidate trajectories, if M init is less than M, the oversampling method is used, otherwise the undersampling method is used; if used for the training process, the similarity label is also calculated using the generated trajectory and the label trajectory for the discriminant network training, and the normalized similarity is calculated by accumulating the error of the corresponding trajectory point coordinates of the feasible candidate trajectory and the label trajectory, as shown in the following formula:

[0060]

[0061] wherein score j represents the score of the jth generated feasible candidate trajectory, ||·|| represents the Euclidean distance of the corresponding trajectory point, and the similarity score label true_score is finally obtained.

[0062] S3, using the current trajectory of the ship and the set of feasible candidate trajectories to realize the training of the trajectory discriminant model based on the neural network.

[0063] As shown in Figure 4 , the trajectory discriminant model based on the neural network includes three parts: a ship current trajectory and generated trajectory encoding module, a trajectory encoding splicing module, and a generated trajectory scoring module.

[0064] The ship current trajectory and generated trajectory encoding module is composed of two groups of LSTM networks; the splicing module realizes the broadcasting of the ship current trajectory encoding and the splicing of the two groups of trajectory encodings, and does not involve parameters that need to be trained; the scoring module realizes the scoring of the encoding results by three layers of MLP.

[0065] The ship current trajectory and generated trajectory encoding module part, the trajectory point feature in the ship current trajectory is P in ={SOG t , COGt ,ΔLAT t ,ΔLON t The feasible candidate trajectory set contains features P of M trajectories and K trajectory points. pre ={ΔLAT t ,ΔLON t The model uses two LSTM networks for encoding. Before inputting the trajectory data into the encoder, each dimension of the attribute is normalized using the Max-Min method. Furthermore, the hidden state of the LSTM encoder at the last time step is used as the trajectory encoding result; therefore, the current trajectory encoding result of the ship is h. L The coding results of each trajectory in the feasible candidate trajectory set are [h K1 ,.,h Ki ,..,h KM ] T Where i represents the i-th feasible candidate trajectory, the two sets of encoding results are flattened and then the high-dimensional features of the encoding are extracted by a fully connected layer as the final encoding result L_encode = h L * ,K_encode=[h K1 * ,...,h KM * ] T .

[0066] The purpose of the trajectory encoding concatenation part is to merge the two sets of codes to obtain the complete trajectory code for use by the scoring network. After L_encode generates M sets via a broadcast mechanism, it is concatenated with each segment of the code in K_encode to obtain the scoring network input vector F_encode = [[h L *,h K1 *],...[h L *,h KM *]] T .

[0067] The trajectory scoring module is implemented using an MLP network. LeakyReLU is chosen as the activation function for both the input and intermediate layers, with dropout added to the intermediate layers. The output is obtained through a softmax function. Each encoding group in F_encode is processed using a multi-step input and weight-sharing approach to obtain the scoring output result pre_score. During training, the loss function is calculated using true_score to train the encoding module and the scoring network according to the following formula:

[0068]

[0069] The loss function consists of two parts: the cross-entropy loss between the true score and the discriminator score, and the L2 regularization term.

[0070] S4, when predicting the current trajectory of a given ship, repeating the trajectory data preprocessing procedure in S1 above, through the step in the S2 feasible candidate trajectory set generation stage, inputting the given trajectory and the generated feasible candidate trajectory set into the neural network model obtained by training in the S3 trajectory discrimination stage to calculate the scores of each trajectory in the feasible candidate trajectory set, and selecting the generated trajectory with the highest score as the final prediction result.

[0071] The two-stage ship trajectory prediction method adopted in the above embodiments effectively fuses the ship navigation mode information and the features of the current ship trajectory, and the prediction result has good interpretability, which provides technical support for important tasks such as ship behavior analysis and port safety supervision.

[0072] In several embodiments provided in the present application, it should be understood that the disclosed technical content can be implemented by other ways. Among them, the above-described device embodiments are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division way, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0073] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed to multiple units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0074] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The above integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0075] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0076] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An interpretable two-stage ship trajectory prediction method based on historical navigation pattern mining, characterized in that, The method comprises the following steps: Obtaining in-service ship data and preprocessing the obtained data to obtain a trajectory data set; For a given arbitrary ship current trajectory, generating a feasible candidate trajectory set based on the trajectory data set through historical navigation mode mining and kinematic model; Building and training a trajectory discrimination model based on a neural network; The input of the trajectory discrimination module is the ship current trajectory and the feasible candidate trajectory set, and the output is the scoring result of the feasible trajectory set; The trajectory discrimination model comprises a trajectory encoding module, a trajectory encoding splicing module and a generated trajectory scoring module; the trajectory encoding module comprises two groups of LSTM networks, respectively used for encoding the current trajectory and the trajectories in the feasible candidate trajectory set; the trajectory encoding splicing module is used for merging the two groups of trajectory encodings obtained by the trajectory encoding module; the generated trajectory scoring module is realized by a three-layer MLP network and is used for scoring the encoding result obtained by the trajectory encoding splicing module; wherein the feasible candidate trajectory set is generated, comprising: For a given arbitrary ship current trajectory, querying all adjacent points within a predetermined spatial range from the trajectory data set as historical reference trajectory starting point set, and obtaining second preset sampling amount of trajectory points from all trajectory points in the starting point set in time sequence to form a historical reference trajectory set; Mining the historical navigation mode from the historical reference trajectory set, comprising: calculating the angle range of the trajectory point relative to the starting point heading based on all historical reference trajectories, and calculating the maximum cumulative heading change and the maximum cumulative speed change; calculating the maximum heading change of the adjacent point using all trajectory points, and dividing the speed of all trajectory points into bins to count the speed increment distribution law and the speed decrement distribution law of each bin; counting the distribution law of the adjacent point speed change state, including deceleration, constant speed and acceleration; counting the adjacent point speed change state transition probability; counting the adjacent point heading change state transition probability, including straight ahead, left turn and right turn; Based on the historical navigation mode, iteratively generate a feasible candidate trajectory set from the predicted starting point; input the ship current trajectory and the feasible candidate trajectory set into the trained trajectory discrimination model for candidate trajectory discrimination, and take the feasible trajectory with the highest score as the prediction result. 2.The two-stage explainable ship trajectory prediction method based on history sailing pattern mining of claim 1, wherein, Data preprocessing comprises: Extracting in-service ship data and separating different ship voyage data, cleaning repeated, missing key attributes and incorrect data; Resampling each trajectory to make the time interval of the trajectory points on each trajectory the same; Taking the first-order difference of the latitude, longitude, speed and heading attributes of the trajectory points and the corresponding attributes of the previous time; Selecting continuous first preset sampling number of trajectory points as model input trajectory, and selecting continuous second preset sampling number of trajectory points after the input trajectory as label trajectory; using a sliding window method to sample trajectory segment samples from each trajectory data at a predetermined sampling interval to form a trajectory data set. 3.The two-stage interpretable ship trajectory prediction method based on history sailing pattern mining of claim 1, wherein, Based on the historical navigation mode, iteratively generate a feasible candidate trajectory set from the predicted starting point, comprising: A navigable range is defined based on the included angle range in the historical navigation mode, and a trajectory point is iteratively generated by the following steps: A heading change state is determined based on a heading change amount of the current trajectory point, a next-time point heading change state is obtained by sampling a neighbor point heading change state transition probability, a next-time point heading is generated based on a heading of the current trajectory point and a maximum heading change amount of the neighbor point, and a next-time point heading is selected based on a maximum cumulative heading change amount of the trajectory point; A speed change state is determined based on a speed change amount of the current trajectory point, a next-time point speed change state is obtained by sampling a neighbor point speed change state transition probability, a suitable speed change amount distribution law is sampled based on a speed of the current trajectory point and the next-time point speed change state to obtain a next-time point speed change amount, a maximum cumulative speed change amount is used to limit the value of the speed change amount, and a speed of a next-time point corresponding to the trajectory point is calculated and obtained; The heading and speed of the next-time point obtained in the above steps are combined with a speed vector of the current trajectory point to calculate a motion distance in the latitude and longitude directions of the ship at the adjacent time by uniform variable speed straight line motion, the motion distance in the latitude and longitude directions is converted into latitude and longitude change amounts to obtain coordinates of the next-time point, and if the next-time point is generated outside the navigable range, the entire trajectory corresponding to the point is deleted. The above steps are iterated to generate a plurality of trajectories, and a feasible candidate trajectory set is obtained by randomly sampling the plurality of trajectories. 4.The two-stage interpretable ship trajectory prediction method based on history sailing pattern mining of claim 1, wherein, Candidate trajectory discrimination includes: The feature of the model input trajectory is encoded by using a first set of LSTM networks, and the hidden layer state h L After being flattened and dimensionally reduced by a fully connected layer, the current trajectory encoding result h of the ship is obtained L ; The features of each feasible trajectory in the set of feasible candidate trajectories are encoded using a second set of LSTM networks, again outputting the hidden state h K The same as h L The feasible candidate trajectory encoding h Ki where i denotes the i-th feasible candidate trajectory; The two sets of encoding results are spliced to obtain complete trajectory encoding; The complete trajectory encoding result is input into a scoring network, the scoring network outputs a scoring vector of the feasible candidate trajectory set through a softmax layer, and in the training stage, a hybrid loss is calculated using the scoring vector of the feasible candidate trajectory set and the similarity between the feasible candidate trajectory and the label trajectory to realize training of the trajectory encoder and the scoring network. 5.The two-stage interpretable ship trajectory prediction method based on history sailing pattern mining of claim 4, wherein, The hybrid loss function is composed of a cross-entropy loss and an L2 loss.

Images