A ship track prediction method and system considering multi-ship coupling effects

Abstract

The present invention belongs to the technical field of ship track prediction and discloses a method and system for ship track prediction under multi-ship coupling influence. The method identifies neighboring ships that exert coupling influences on a target ship; extracts the potential motion feature vector of the target ship affected by the coupling at time #imgabs0#; uses a designed generator to output the predicted track of the target ship; and inputs the actual track and the predicted track into a discriminator, ultimately generating a probability that the input track is identified as the actual track. The competitive training of the generator and discriminator in the generative adversarial network of the present invention can further improve the accuracy of ship track prediction, making the constructed model more applicable to real-world ship navigation environments and reducing collision risks.

Description

Technical Field

[0001] The present invention belongs to the technical field of ship track prediction, and in particular relates to a ship track prediction method and system under the influence of multiple ships coupling. Background Art

[0002] With the growing demand for maritime trade and increasing traffic pressure in complex maritime areas, the risk of traffic accidents such as ship collisions is also increasing. To improve maritime traffic safety and prevent collisions, it is necessary to monitor and predict ship movement to achieve intelligent maritime surveillance. Existing technologies use a bidirectional long short-term memory (Bi-LSTM) network as the basic model for ship track prediction. This model more comprehensively considers the contextual information of time series data. Furthermore, to improve the accuracy and adaptability of complex tracks, an attention mechanism is incorporated into the Bi-LSTM model to increase the weight of key information. However, most current research on ship track prediction fails to consider the coupling effects caused by multi-ship encounters and retains redundant information during the prediction process, affecting prediction accuracy.

[0003] The above analysis reveals the following problems and shortcomings of existing technologies: While existing technologies use Bi-LSTM to more comprehensively consider the contextual information of time series data and incorporate an attention mechanism to increase the weight of key information, they retain redundant information during the prediction process. Furthermore, this research fails to consider the coupling effects caused by multi-ship encounters during navigation. In real-world navigation environments, ships may encounter head-on, crossing, and overtaking situations. In these situations, when two or more encounter scenarios occur simultaneously, neighboring ships can have a coupling effect on the navigation of the target vessel. Summary of the Invention

[0004] To overcome the problems existing in related technologies, the present invention discloses a method and system for ship track prediction under the influence of multiple ship coupling. This invention aims to address the problem that ship track prediction fails to consider the coupling effect caused by multiple ship encounters, or that, while considering the impact of encounters, only addresses one of the two ship encounters, namely, head-on encounters, crossing encounters, or overtaking, while simultaneously addressing two or more complex scenarios. The present invention constructs a ship track prediction model based on a generative adversarial network (GAN) that integrates multi-head probabilistic sparse self-attention and a multilayer perceptron (MLP). This prediction model uses a quantitative collision risk indicator to effectively identify neighboring ships that have coupling effects on a target ship and extracts the potential motion characteristics of the target ship under coupling effects. Based on this, the target ship's track is predicted, and through competitive training of the generator and discriminator in the model, a more accurate ship track is generated, providing technical support for ensuring maritime traffic safety and improving the efficiency of maritime traffic supervision.

[0005] The technical solution is as follows: A method for predicting ship tracks under the influence of multiple ship coupling, comprising:

[0006] S1, selects the closest approach distance DCPA and the closest approach time TCPA as evaluation indicators to identify the adjacent ships that have coupling effects on the target ship;

[0007] S2, MLP is used to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships, and the target ship is extracted. The potential motion feature vector affected by coupling at each moment ;

[0008] S3, the target ship The potential motion eigenvector affected by coupling at every moment The input is encoded into the Bi-LSTM-based G-encoder equipped with the generator, and the hidden feature representation of the target ship track is output. ; Represent the hidden features of the target ship track Input into the multi-head probabilistic sparse self-attention, remove redundant information, retain attention to key information, and obtain weighted hidden features ; The weighted hidden features The input is sent to the Bi-LSTM-based G-decoder on the generator for decoding, and the feature representation of the predicted trajectory is output. ; The characteristics of the predicted track are represented by The input is fed into the MLP on the generator, which outputs the predicted track of the target ship.

[0009] S4, input the actual track and predicted track into the D-encoder based on Bi-LSTM equipped with the discriminator to obtain the hidden state matrix output by Bi-LSTM , the hidden state matrix output by Bi-LSTM The input is fed into the MLP of the discriminator, which ultimately generates the probability that the input track is recognized as the real track. .

[0010] In step S1, the closest approach distance DCPA and the closest approach time TCPA are selected as evaluation indicators to identify adjacent ships that have coupling effects on the target ship, including:

[0011] The latitude, longitude, speed and heading of the target ship and adjacent ships are: and ; Calculate the relative distance between the target ship and the adjacent ship based on their latitude and longitude coordinates; as shown in formula (1)-formula (2):

[0012] (1)

[0013] (2)

[0014] Where, is the relative distance between the target ship and the adjacent ships, is the radius of the Earth;

[0015] According to the speed and heading of the target ship and the adjacent ship, the relative speed and heading between the two are calculated, as shown in formula (3)-formula (4):

[0016] (3)

[0017] (4)

[0018] Where, is the relative speed, is the relative heading, is the arccosine function;

[0019] According to the relative distance, relative speed and relative heading calculated above, the corresponding DCPA and TCPA are calculated as shown in formula (5)-formula (6):

[0020] (5)

[0021] (6)

[0022] Where, is the relative bearing of the target ship;

[0023] Vessels with DCPA less than 3 NM and TCPA less than 30 minutes are identified.

[0024] In step S2, MLP is used to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the neighboring ships, and the target ship is extracted. The potential motion feature vector affected by coupling at each moment , including: MLP establishes a complex nonlinear relationship model through hidden layers and activation functions, and automatically learns hierarchical feature representation from the original data by processing high-dimensional data, extracts the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the neighboring ships, and extracts the potential motion feature vector of the target ship based on the nonlinear relationship between them, as shown in formula (7)-formula (9):

[0025] (7)

[0026] Where, is the input feature matrix of MLP, For the target ship The feature vector at the moment, including latitude, longitude, speed, and heading information, ; The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, ; The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, ;

[0027] Formula (7) integrates the state information of the target ship and neighboring ships into an input matrix for subsequent feature extraction;

[0028] (8)

[0029] Where, For the The output of the hidden layer, For the The weight matrix of the layer, For the The output of the layer, For the The bias of the layer, is the total number of hidden layers in the MLP, is the activation function;

[0030] Formula (8) describes the first The calculation process of the hidden layer extracts features layer by layer through linear transformation and nonlinear activation function;

[0031] (9)

[0032] Where, For the target ship The potential motion eigenvector affected by the coupling at each moment, is the weight matrix of the output layer in MLP, is the output of the Lth hidden layer, is the bias of the output layer in MLP;

[0033] Formula (9) is the calculation process of the MLP output layer, which generates the target ship The potential motion eigenvector affected by the coupling at each moment.

[0034] In step S3, the generator is equipped with: G-encoder, multi-head probabilistic sparse self-attention module, G-decoder and MLP;

[0035] Bi-LSTM is selected as the core component of G-encoder. Bi-LSTM consists of two layers of long short-term memory network LSTM stacked together. The first layer LSTM inputs data forward, and the second layer LSTM inputs data backward. The LSTM unit contains a forget gate, an input gate, a cell state, and an output gate. The above components control the information flow through the sigmoid activation function and point-by-point multiplication to achieve the modeling of long-term dependencies in sequence data. The long short-term memory network LSTM includes: the input vector enters the forget gate, the input gate, and the output gate, as shown in formula (10)-formula (12):

[0036] (10)

[0037] Where, is the forget gate vector, is the sigmoid function, Transform the weight of the historical state of the forget gate, For the target ship The potential motion eigenvector affected by the coupling at each moment, Transform the weights of the input features of the forget gate, for The hidden state of the moment, is the bias of the forget gate;

[0038] Formula (10) determines whether to remember the cell state information of the previous unit;

[0039] (11)

[0040] (12)

[0041] Where, is the input gate vector, is the historical state transformation weight of the input gate, Transform the weights of the input features of the input gate, is the bias of the input gate, is the output gate vector, is the historical state transformation weight of the output gate, Transform the weights of the input features of the output gate, is the bias of the output gate, is the activation function;

[0042] Formula (11) controls the new information that needs to be updated to the cell state; Formula (12) determines which information in the cell state is remembered as the output state;

[0043] Output the candidate cell state vector and update the cell state vector, as shown in formula (13)-formula (14):

[0044] (13)

[0045] (14)

[0046] Where, is the candidate cell state vector, is the historical state transformation weight of the candidate cell state, is the input feature transformation weight of the candidate cell state, is the cell state vector, is the bias of the candidate cell state, for The cell state vector at the moment;

[0047] Formula (13) generates new information that may be added to the cell state; Formula (14) combines the decisions of the forget gate and the input gate to update the cell state;

[0048] Update the hidden state vector according to the output gate vector and the cell state vector, as shown in formula (15):

[0049] (15)

[0050] Where, is the hidden state vector output by the unidirectional LSTM layer;

[0051] Formula (15) updates the hidden state vector based on the output gate vector and the cell state vector.

[0052] In step S3, in the Bi-LSTM, the forward propagation layer processes the sequence in time order to generate the hidden state , the backward propagation layer processes the sequence in reverse order to generate the hidden state ; The bidirectional hidden state at each moment is obtained by formula (16):

[0053] (16)

[0054] Where, is the bidirectional hidden state of the final output of Bi-LSTM;

[0055] Formula (16) will and Splicing to form a bidirectional hidden state containing context information for subsequent calculations.

[0056] Furthermore, the multi-head probabilistic sparse self-attention module combines the multi-head attention mechanism with the sparsification strategy to focus on the coupling effects of neighboring ships on the target ship, as shown in Formula (17)-Formula (24):

[0057] (17)

[0058] (18)

[0059] (19)

[0060] Where, are query matrix, key matrix and value matrix respectively, for the reason The hidden matrix consists of vectors, ; is the number of timestamps, is the weight matrix of query, key and value;

[0061] In order to reduce computational complexity and focus on processing the most relevant information, a sparsification strategy is adopted. The sparsity measure of each query is calculated as shown in formula (20):

[0062] (20)

[0063] Where, for No. row vectors, for No. row vectors, is the matrix transpose symbol, is a row vector The transpose of is the length of the key, is the feature dimension of the key vector in the model, is the sparsification metric score, is the bond matrix;

[0064] Formula (20) is used to determine the attention distribution between each query and all keys, thereby obtaining the sparse measurement value between each query and key. The obtained multiple sparse measurement values ​​are sorted from large to small, and the top The results are used in subsequent calculations; the calculation formula for sparse attention weight is shown in formula (21):

[0065] (twenty one)

[0066] Where, is the sparse attention weight, for indivual The corresponding query element set, is the bond matrix The transposed matrix of

[0067] Formula (21) is used to calculate the attention weight between the query and the key, and the probability distribution is obtained through softmax normalization to determine the degree of attention of the model at different positions;

[0068] The probabilistic sparse attention mechanism is extended to Attention heads are used to capture different features, as shown in formulas (22)-(24). The output captures the global dependencies and complex features in the data to analyze and predict future trajectories.

[0069] Calculate the The output of an attention head:

[0070] (twenty two)

[0071] Where, For the The output of an attention head, is the sparse attention weight, is the value matrix;

[0072] The outputs of all attention heads are concatenated and linearly transformed:

[0073] (twenty three)

[0074] Where, is the output of multi-head probabilistic sparse self-attention, For splicing operations, is the output of the first attention head, is the output of the second attention head, For the The output of an attention head, is the learnable weight matrix of the output layer;

[0075] Get the final output:

[0076] (twenty four)

[0077] Where, is the final output of the multi-head probabilistic sparse self-attention.

[0078] Furthermore, the G-decoder gradually generates the feature representation of the predicted trajectory based on the output of the G-encoder and the multi-head probabilistic sparse self-attention module. ; G-decoder also uses Bi-LSTM; then introduce MLP after the decoder to convert the hidden representation into Latitude, longitude, speed and heading data of the predicted track at all times , that is, the predicted track, as shown in formula (25)-formula (26):

[0079] (25)

[0080] (26)

[0081] Where, are the weight matrices of the Bi-LSTM and MLP networks respectively, are the biases of Bi-LSTM and MLP networks respectively.

[0082] In step S4, the discriminator is equipped with a D-encoder and an MLP;

[0083] The discriminator processes the input track data through the D-encoder and MLP, and outputs the probability that the input data is identified as a true track; as shown in formulas (27) and (28):

[0084] (27)

[0085] (28)

[0086] Where, is the hidden state matrix output by Bi-LSTM, are the weight matrices of Bi-LSTM and MLP respectively, are the biases of Bi-LSTM and MLP respectively, represents the real track sequence of the target ship, where For the target ship The actual latitude, longitude, speed and heading at each moment; Represents the predicted track sequence of the target ship output by the generator, where For the target ship The predicted latitude, longitude, speed and heading at each moment, is the probability that the input track is identified as a true track.

[0087] After step S4, the generator and discriminator compete with each other through an adversarial learning mechanism and gradually optimize their respective parameters; ultimately, predicted track data that is highly similar to the actual track data is generated, achieving high-precision track prediction.

[0088] Another object of the present invention is to provide a ship track prediction system for multi-ship coupling effects, which implements the ship track prediction method for multi-ship coupling effects, and the system includes:

[0089] The ship identification module based on multi-ship coupling influence is used to select the closest approach distance DCPA and the closest approach time TCPA as evaluation indicators to identify adjacent ships that have a coupling influence on the target ship;

[0090] The feature extraction module uses MLP to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships, and extracts the target ship's The potential motion eigenvector affected by coupling at every moment ;

[0091] Generator for placing the target ship in The potential motion eigenvector affected by coupling at every moment The input is encoded into the Bi-LSTM-based G-encoder equipped with the generator, and the hidden feature representation of the target ship track is output. ; Represent the hidden features of the target ship track Input into the multi-head probabilistic sparse self-attention, remove redundant information, retain attention to key information, and obtain weighted hidden features ; The weighted hidden features The input is sent to the Bi-LSTM-based G-decoder on the generator for decoding, and the feature representation of the predicted trajectory is output. ; The characteristics of the predicted track are represented by The input is fed into the MLP on the generator, which outputs the predicted track of the target ship.

[0092] The discriminator inputs the real track and the predicted track into the D-encoder based on Bi-LSTM carried by the discriminator to obtain the hidden state matrix output by Bi-LSTM , the hidden state matrix output by Bi-LSTM The input is fed into the MLP of the discriminator, which ultimately generates the probability that the input track is recognized as the real track. .

[0093] In combination with all the above technical solutions, the beneficial effects of the present invention are as follows:

[0094] First, this paper constructs a novel ship track prediction model based on a GAN that simultaneously integrates multi-head probabilistic sparse self-attention and multi-layered prediction (MLP) by extracting the coupling effects of neighboring vessels on a target vessel in various encounter scenarios. This model overcomes the limitations of previous studies that either fail to consider the coupling effects caused by multi-ship encounters or, while considering the effects of ship encounters, only consider encounters, crossings, or overtakings between two ships, while omitting the possibility of two or more complex scenarios. This model enables ship track prediction under the coupling effects of neighboring ships in multi-ship encounters. This paper employs an MLP to quantify the multi-ship coupling effects and extract the latent motion characteristics of the target vessel. Bi-LSTM is also selected as the core component of the encoder and decoder to capture the characteristics of bidirectional time series. Multi-head probabilistic sparse self-attention is embedded in the model. Through parallel attention computation and a probabilistic sparse selection mechanism, it can more efficiently capture the intrinsic connections between hidden representations and eliminate redundant information. The competitive training of the generator and discriminator in the generative adversarial network further improves the accuracy of ship track prediction, making the constructed model more applicable to real-world ship navigation environments and reducing collision risk.

[0095] Second, the solution presented in this paper can be widely applied in intelligent shipping management systems, port scheduling, and autonomous ship driving. By accurately predicting the paths of ships during multi-ship encounters, it can effectively reduce the incidence of ship collisions, minimize economic losses from insurance claims and ship repairs, and lower business operating costs. It can also improve waterway utilization and shipping efficiency, thereby boosting logistics and transportation efficiency.

[0096] Third, existing ship track prediction technologies, both domestically and internationally, primarily focus on single-ship motion or simple two-ship encounter scenarios, lacking systematic research on ship track prediction in complex encounters involving two or more ships. This invention proposes, for the first time, a ship track prediction model that integrates multi-head probabilistic sparse self-attention, MLP, and GAN, achieving key breakthroughs at both the theoretical and algorithmic levels and providing a new technical path for multi-ship collaborative navigation research. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure;

[0098] Figure 1 This is a schematic diagram of a ship track prediction system with multi-ship coupling effects provided by an embodiment of the present invention;

[0099] Figure 2 This is an example diagram of DCPA and TCPA calculation provided by an embodiment of the present invention;

[0100] Figure 3 This is a diagram of the Bi-LSTM structure provided by an embodiment of the present invention;

[0101] Figure 4 This is a comparison diagram of the prediction errors when the time interval is 30s;

[0102] Figure 5 This is a comparison diagram of the prediction errors when the time interval is 1 minute;

[0103] Figure 6 This is a comparison diagram of the prediction errors when the time interval is 2 minutes;

[0104] Figure 7 This is a comparison diagram of the prediction errors when the time interval is 5 minutes;

[0105] Figure 8 This is a comparative trajectory diagram of the present invention with a time interval of 30s;

[0106] Figure 9 Schematic diagram of longitude prediction using different methods when the time interval is set to 30s;

[0107] Figure 10 Schematic diagram of latitude prediction using different methods when the time interval is set to 30s;

[0108] Figure 11 This is a comparative trajectory diagram of the present invention with a time interval of 1 minute;

[0109] Figure 12 This is the longitude prediction graph of different methods when the time interval is set to 1 minute;

[0110] Figure 13 This is the latitude prediction map of different methods when the time interval is set to 1 minute;

[0111] Figure 14 This is a comparative trajectory diagram of the present invention with a time interval of 2 minutes;

[0112] Figure 15 This is the longitude prediction graph of different methods when the time interval is set to 2 minutes;

[0113] Figure 16 This is the latitude prediction map of different methods when the time interval is set to 2 minutes;

[0114] Figure 17 This is a comparative trajectory diagram of the present invention with a time interval of 5 minutes;

[0115] Figure 18 This is the longitude prediction graph of different methods when the time interval is set to 5 minutes;

[0116] Figure 19 This is the latitude prediction map of different methods when the time interval is set to 5 minutes. DETAILED DESCRIPTION

[0117] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth numerous specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0118] The innovation of the present invention lies in: in view of the excellent performance of the generative adversarial network in generating predicted tracks that are infinitely close to the real track through game adversarial training between the generator and the discriminator, a new ship track prediction model based on the generative adversarial network is constructed by integrating multi-head probabilistic sparse self-attention and MLP, which is used to explore the track prediction of target ships under the influence of multi-ship coupling; the model introduces MLP for feature extraction to quantify the influence of multi-ship coupling and explore the potential navigation behavior characteristics of target ships under the multi-ship coupling effect; at the same time, the multi-head probabilistic sparse self-attention mechanism is embedded, which can more efficiently capture the intrinsic connection between hidden representations through parallel attention calculation and probabilistic sparse selection mechanism, eliminate redundant information, and determine the weights of continuous variables in the hidden state.

[0119] Example 1, as Figure 1 As shown, the ship track prediction system for multi-ship coupling influence provided by an embodiment of the present invention includes:

[0120] The ship identification module based on multi-ship coupling effects selects the closest approach distance (DCPA) and the closest approach time (TCPA) as evaluation indicators to identify adjacent ships that have a coupling effect on the target ship. The DCPA is less than 3 NM and the TCPA is less than 30 minutes.

[0121] The feature extraction module uses MLP to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships, and extracts the target ship's The potential motion feature vector affected by coupling at each moment ;

[0122] Generator for placing the target ship in The potential motion feature vector affected by coupling at each moment The input is encoded into the Bi-LSTM-based G-encoder equipped with the generator, and the hidden feature representation of the target ship track is output. ; Represent the hidden features of the target ship track Input into the multi-head probabilistic sparse self-attention, remove redundant information, retain attention to key information, and obtain weighted hidden features ; The weighted hidden features The input is sent to the Bi-LSTM-based G-decoder on the generator for decoding, and the feature representation of the predicted trajectory is output. ; The characteristics of the predicted track are represented by The input is fed into the MLP on the generator, which outputs the predicted track of the target ship.

[0123] The discriminator inputs the real track and the predicted track into the D-encoder based on Bi-LSTM carried by the discriminator to obtain the hidden state matrix output by Bi-LSTM , the hidden state matrix output by Bi-LSTM The input is fed into the MLP of the discriminator, which ultimately generates the probability that the input track is recognized as the real track. .

[0124] The generator and discriminator are trained competitively to further improve the accuracy of ship track prediction.

[0125] Example 2: The key to ship track prediction based on the influence of multi-ship coupling is how to explore the potential navigation characteristics of the target ship under the multi-ship coupling effect. In addition, the generator in GAN is responsible for generating the future track of the ship, while the discriminator continuously distinguishes the generated track from the real track. Through continuous adversarial training, the two can jointly improve the model performance, thereby achieving high-precision ship track prediction. Therefore, the present invention proposes a new ship track prediction model based on GAN that simultaneously integrates multi-head probabilistic sparse self-attention and MLP. The specific framework principle is as follows Figure 1As shown in the figure, the distance to closest point of approach (DCPA) and time to close point of approach (TCPA) are used as evaluation indicators to identify the adjacent ships that have coupling effects on the target ship and MLP is used for feature extraction to obtain the potential navigation feature vector of the target ship. . By G-encoder Encode and output the hidden feature representation of the target ship track Embedding multi-head probabilistic sparse self-attention, through parallel attention calculation and probabilistic sparse selection mechanism, more effectively explores the relationship between hidden features, and obtains weighted hidden features . The G-decoder will The decoded data is converted into context information to generate a predicted track. The real track and the predicted track are input into the discriminator to identify the authenticity of the track.

[0126] Specifically, the ship track prediction method for multi-ship coupling effects provided by the embodiment of the present invention includes:

[0127] S1, selects the closest approach distance DCPA and the closest approach time TCPA as evaluation indicators to identify the adjacent ships that have coupling effects on the target ship;

[0128] The closest approach distance DCPA is less than 3 NM and the closest approach time TCPA is less than 30 minutes;

[0129] S2, MLP is used to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships, and the target ship is extracted. The potential motion eigenvector affected by coupling at every moment ;

[0130] S3, the target ship The potential motion eigenvector affected by coupling at every moment The input is encoded into the Bi-LSTM-based G-encoder equipped with the generator, and the hidden feature representation of the target ship track is output. ; Represent the hidden features of the target ship track Input into the multi-head probabilistic sparse self-attention, remove redundant information, retain attention to key information, and obtain weighted hidden features ; The weighted hidden features The input is sent to the Bi-LSTM-based G-decoder on the generator for decoding, and the feature representation of the predicted trajectory is output. ; The characteristics of the predicted track are represented by The input is fed into the MLP on the generator, which outputs the predicted track of the target ship.

[0131] S4, input the actual track and predicted track into the D-encoder based on Bi-LSTM equipped with the discriminator to obtain the hidden state matrix output by Bi-LSTM , the hidden state matrix output by Bi-LSTM The input is fed into the MLP of the discriminator, which ultimately generates the probability that the input track is recognized as a true track. ;

[0132] Among them, the generator and discriminator are trained competitively to further improve the accuracy of ship track prediction.

[0133] Exemplarily, in step S1, ship identification based on multi-ship coupling impact: DCPA and TCPA are selected as evaluation indicators, the DCPA and TCPA of the target ship and other ships are calculated, and ships with DCPA less than 3NM and TCPA less than 30 minutes are regarded as adjacent ships that have an impact on the target ship.

[0134] Specifically, ship identification based on multi-ship coupling effects includes:

[0135] During navigation, the navigation of the target ship will be affected by the coupling caused by multiple ships. How to identify the adjacent ships that have a coupling effect on the target ship is the basis for the subsequent work of this invention. Considering that DCPA and TCPA are usually used in collision risk assessment and can effectively determine which are dangerous ships and which are non-dangerous ships, this invention uses DCPA and TCPA as a method for identifying adjacent ships. Figure 2 As shown in the DCPA and TCPA calculation example, the latitude, longitude, speed and heading of the target ship and the adjacent ships are and The relative distance between the target ship and the adjacent ship is calculated based on their latitude and longitude coordinates as shown in Formula (1)-Formula (2). The relative speed and relative heading between the target ship and the adjacent ship are calculated based on their speed and heading as shown in Formula (3)-Formula (4). Finally, the corresponding DCPA and TCPA are calculated based on the relative distance, relative speed, and relative heading calculated above as shown in Formula (5)-Formula (6).

[0136] (1)

[0137] (2)

[0138] Where, is the relative distance between the target ship and the adjacent ships, is the radius of the Earth;

[0139] According to the speed and heading of the target ship and the adjacent ship, the relative speed and heading between the two are calculated, as shown in formula (3)-formula (4):

[0140] (3)

[0141] (4)

[0142] Where, is the relative speed, is the relative heading, is the arccosine function;

[0143] According to the relative distance, relative speed and relative heading calculated above, the corresponding DCPA and TCPA are calculated as shown in formula (5)-formula (6):

[0144] (5)

[0145] (6)

[0146] Where, is the relative bearing of the target ship;

[0147] Vessels with DCPA less than 3 NM and TCPA less than 30 minutes are identified.

[0148] Generally speaking, ship operators typically define a safe distance of 2 NM based on experience to avoid other vessels. However, in real-world navigation scenarios, various physiological, psychological, and environmental factors can affect a driver's judgment. Therefore, the present invention defines a safe distance of 3 NM. A DCPA of less than 3 NM and a TCPA of less than 30 min are used as evaluation indicators in the present invention.

[0149] It can be understood that the above formula innovatively proposed in the present invention plays a unique and critical role in the application background and specific application scenarios of the present invention, has significant technical effects, and plays a positive role in improving the accuracy of track prediction, enhancing adaptability to complex environments and multi-ship interactions, and solving the technical difficulties of existing technologies in track prediction in complex port waters.

[0150] For example, in step S2, feature extraction is used to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships using MLP, and extract the target ship's t The potential motion eigenvector affected by coupling at every moment , as the input to the Bi-LSTM-based G-encoder on the generator.

[0151] Specifically, feature extraction includes:

[0152] As described in step S1, determining the neighboring ships that produce multi-ship coupling effects on the target ship is the basis of the research. Then how to quantify the coupling effects produced by neighboring ships is the key issue. MLP is able to establish complex nonlinear relationship models through hidden layers and activation functions, and automatically learn hierarchical feature representations from raw data by processing high-dimensional data. Therefore, MLP is embedded in the feature extraction framework. It is used to extract the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the neighboring ships. Based on the nonlinear relationship between them, the potential motion feature vector of the target ship is extracted as shown in Formula (7)-Formula (9):

[0153] (7)

[0154] Where, is the input feature matrix of MLP, For the target ship The feature vector at the moment, including latitude, longitude, speed, and heading information, ; The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, ; The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, ;

[0155] Formula (7) integrates the state information of the target ship and neighboring ships into an input matrix for subsequent feature extraction;

[0156] (8)

[0157] Where, For the The output of the hidden layer, For the The weight matrix of the layer, For the The output of the layer, For the The bias of the layer, is the total number of hidden layers in the MLP, is the activation function;

[0158] Formula (8) describes the first The calculation process of the hidden layer extracts features layer by layer through linear transformation and nonlinear activation function;

[0159] (9)

[0160] Where, For the target ship The potential motion eigenvector affected by the coupling at each moment, is the weight matrix of the output layer in MLP, is the output of the Lth hidden layer, is the bias of the output layer in MLP;

[0161] Formula (9) is the calculation process of the MLP output layer, which generates the target ship The potential motion eigenvector affected by the coupling at each moment.

[0162] It is understandable that the usage scenarios and functions of MLP in the prior art are different from those of the present invention. Many existing studies focus only on the track prediction of a single ship. In these scenarios, MLP is mainly used to extract and process the features of the track data of a single ship itself. For example, MLP is used to analyze historical data such as the speed, heading, and position of a ship to explore its own motion laws and patterns. It does not involve the interaction between multiple ships, nor does it extract the nonlinear relationship between the motion characteristics of the target ship and neighboring ships. The problem of ship track prediction under the multi-ship coupling effect is uniquely complex. Multiple ships influence each other during navigation, and their motion relationships are nonlinear and complex and changeable. The methods of single ship feature processing or general data feature conversion in the prior art cannot capture such complex interactions and coupling effects between multiple ships.

[0163] Therefore, the formulas (7) to (9) constructed in the present invention are used to extract potential motion feature vectors, which is an innovative method for ship track prediction in multi-ship coupling scenarios.

[0164] Exemplarily, in step S3, the generator is equipped with a G-encoder, a multi-head probabilistic sparse self-attention module, a G-decoder, and an MLP.

[0165] The G-encoder converts input data into a high-dimensional latent representation to capture the time-dependent contextual information within the data. In fact, for ship track prediction, the ship's track at the next moment will be influenced by the combined influence of the data from the previous and next moments. Because Bi-LSTM can capture both forward and reverse sequential information, improving the model's ability to analyze sequential data, this paper selects Bi-LSTM as the core component of the G-encoder.

[0166] like Figure 3 As shown in the Bi-LSTM structure diagram, Bi-LSTM consists of two layers of long short-term memory (LSTM) networks stacked together. The first layer of LSTM inputs data forward, and the second layer of LSTM inputs data backward. The LSTM unit contains a forget gate, an input gate, a cell state, and an output gate. These components control the flow of information through a sigmoid activation function and point-by-point multiplication, thereby modeling long-term dependencies in sequence data. The details of the LSTM network are as follows. The input vector enters the forget gate, input gate, and output gate, as shown in Formula (10)-Formula (12). Then, the candidate cell state vector is output and the cell state vector is updated, as shown in Formula (13)-Formula (14). Finally, the hidden state vector is updated based on the output gate vector and the cell state vector, as shown in Formula (15).

[0167] (10)

[0168] Where, is the forget gate vector, is the sigmoid function, Transform the weight of the historical state of the forget gate, For the target ship The potential motion eigenvector affected by the coupling at each moment, Transform the weights of the input features of the forget gate, for The hidden state of the moment, is the bias of the forget gate;

[0169] Formula (10) determines whether to remember the cell state information of the previous unit;

[0170] (11)

[0171] (12)

[0172] Where, is the input gate vector, is the historical state transformation weight of the input gate, Transform the weights of the input features of the input gate, is the bias of the input gate, is the output gate vector, is the historical state transformation weight of the output gate, Transform the weights of the input features of the output gate, is the bias of the output gate, is the activation function;

[0173] Formula (11) controls the new information that needs to be updated to the cell state; Formula (12) determines which information in the cell state is remembered as the output state;

[0174] Output the candidate cell state vector and update the cell state vector, as shown in formula (13)-formula (14):

[0175] (13)

[0176] (14)

[0177] Where, is the candidate cell state vector, is the historical state transformation weight of the candidate cell state, is the input feature transformation weight of the candidate cell state, is the cell state vector, is the bias of the candidate cell state, for The cell state vector at the moment;

[0178] Formula (13) generates new information that may be added to the cell state; Formula (14) combines the decisions of the forget gate and the input gate to update the cell state;

[0179] Update the hidden state vector according to the output gate vector and the cell state vector, as shown in formula (15):

[0180] (15)

[0181] Where, is the hidden state vector output by the unidirectional LSTM layer;

[0182] Formula (15) updates the hidden state vector based on the output gate vector and the cell state vector.

[0183] In Bi-LSTM, the forward propagation layer processes the sequence in time order to generate the hidden state , the backward propagation layer processes the sequence in reverse order to generate the hidden state Finally, the bidirectional hidden state at each moment is obtained by formula (16).

[0184] (16)

[0185] Where, is the bidirectional hidden state of the final output of Bi-LSTM;

[0186] Formula (16) will and Splicing to form a bidirectional hidden state containing context information for subsequent calculations.

[0187] It can be understood that in the field of ship track prediction, encoders in the existing technology mostly use a single LSTM, unidirectional GRU or ordinary feedforward neural network structure: a single LSTM can only capture sequence dependencies in the forward direction and cannot utilize future information; the unidirectional GRU is similar and has limited long-distance modeling capabilities; and ordinary feedforward neural networks cannot process time series features. The G-encoder in the present invention uses Bi-LSTM as a core component in a breakthrough way. By superimposing two layers of forward and backward LSTM, it realizes the bidirectional time series modeling of the ship track sequence for the first time - it not only captures the impact of historical tracks on the current state, but also explores the potential constraints of future trends on current behavior. Compared with the existing technology, this encoder solves the technical bottleneck that traditional unidirectional models cannot capture the "future-present" reverse time series correlation and long-distance features are easily lost. It provides a more comprehensive spatiotemporal feature base for track prediction in complex multi-ship interaction scenarios, so that the model can accurately portray the dynamic behavior logic of ships in the multi-ship coupling effect.

[0188] An exemplary multi-head probabilistic sparse self-attention module. To further improve the model's ability to capture the complex dependencies between the target ship and its neighboring ships, a multi-head probabilistic sparse self-attention mechanism is employed. This mechanism combines the multi-head attention mechanism with a sparsification strategy. By retaining important motion features, it dynamically reduces attention to irrelevant motion features, thereby facilitating attention to the coupling effects of neighboring ships on the target ship, as shown in Equations (17) to (24).

[0189] (17)

[0190] (18)

[0191] (19)

[0192] Where, are query matrix, key matrix and value matrix respectively, for the reason The hidden matrix consists of vectors, ; is the number of timestamps, is the weight matrix of query, key and value;

[0193] In order to reduce computational complexity and focus on processing the most relevant information, a sparsification strategy is adopted. The sparsity measure of each query is calculated as shown in formula (20):

[0194] (20)

[0195] Where, for No. row vectors, for No. row vectors, is the matrix transpose symbol, is a row vector The transpose of is the length of the key, is the feature dimension of the key vector in the model, is the sparsification metric score, is the bond matrix;

[0196] Formula (20) is used to determine the attention distribution between each query and all keys, thereby obtaining the sparse measurement value between each query and key. The obtained multiple sparse measurement values ​​are sorted from large to small, and the top The results are used in subsequent calculations; the calculation formula for sparse attention weight is shown in formula (21):

[0197] (twenty one)

[0198] Where, is the sparse attention weight, for indivual The corresponding query element set, is the bond matrix The transposed matrix of

[0199] Formula (21) is used to calculate the attention weight between the query and the key, and the probability distribution is obtained through softmax normalization to determine the degree of attention of the model at different positions;

[0200] The probabilistic sparse attention mechanism is extended to Attention heads are used to capture different features, as shown in Equations (22) to (24). This final output captures the global dependencies and complex features in the data, enabling the model to effectively analyze and predict future trajectories.

[0201] Calculate the The output of an attention head is shown in formula (22):

[0202] (twenty two)

[0203] Where, For the The output of an attention head, is the sparse attention weight, is the value matrix;

[0204] The outputs of all attention heads are concatenated and linearly transformed as shown in formula (23):

[0205] (twenty three)

[0206] Where, is the output of multi-head probabilistic sparse self-attention, For splicing operations, is the output of the first attention head, is the output of the second attention head, For the The output of an attention head, is the learnable weight matrix of the output layer;

[0207] The final output is shown in formula (24):

[0208] (twenty four)

[0209] Where, is the final output of the multi-head probabilistic sparse self-attention.

[0210] It is understandable that the multi-head self-attention and sparsification strategies in the existing technology mostly focus on computational efficiency optimization or general sequence modeling, and are not combined with the feature modeling requirements of multi-ship coupling scenarios. The multi-head probabilistic sparse self-attention mechanism of the present invention has made a breakthrough in deeply integrating the quantification of ship coupling effects with attention calculation: after mapping the ship motion features into query (Q), key (K), and value (V) matrices through equations (17)-(19), the sparse metric of equation (20) is used to dynamically filter features that are strongly associated with the target ship and eliminate redundant information; through the sparse attention weight calculation of equation (21), different attention heads are adaptively focused on the multi-dimensional features of the coupling relationship, realizing dynamic weight allocation and hierarchical ranking of feature importance in the hidden state of continuous variables; finally, through multi-head concatenation and linear transformation of equations (22)-(24), the local sensitivity and global dependency of multi-ship interactions are uniformly captured. Compared with existing technologies, the formula operation of this mechanism has been upgraded from "general feature association" to "coupling effect-guided semantic perception", which solves the problems of redundant calculation and dilution of key features in traditional attention in multi-ship scenarios, and significantly improves the model's efficiency and accuracy in analyzing complex multi-ship interaction patterns.

[0211] Specifically, the G-decoder gradually generates the feature representation of the predicted trajectory based on the output of the G-encoder and the multi-head probabilistic sparse self-attention module. Like the G-encoder, the G-decoder also uses Bi-LSTM. Then, MLP is introduced after the decoder to transform the hidden representation into Latitude, longitude, speed and heading data of the predicted track at all times , that is, the predicted track. The specific formulas are shown in Equations (25)-(26).

[0212] (25)

[0213] (26)

[0214] Where, are the weight matrices of the Bi-LSTM and MLP networks respectively, are the biases of Bi-LSTM and MLP networks respectively.

[0215] It is understandable that in the prior art, decoders mostly use unidirectional LSTM or traditional fully connected layers, which can only process unidirectional time series or local features and are difficult to accurately map to latitude and longitude coordinates and speed and heading with complex spatiotemporal dependencies. The G-decoder of this application innovatively continues the bidirectional modeling capability of Bi-LSTM, and synchronously captures the causal relationship of historical tracks and the potential constraints of future tracks through the forward and backward LSTM layers, solving the problem that the unidirectional model is insufficient in modeling "reverse time series interaction" in multi-ship coupling scenarios. At the same time, the second MLP introduced after the decoder directly maps the high-dimensional hidden features output by Bi-LSTM to latitude and longitude coordinates and speed and heading (such as Equations (25)-(26)) through multi-layer nonlinear transformations, breaking through the information loss bottleneck of the prior art that requires multi-layer complex networks to convert coordinates - it not only improves the coherence of track generation through bidirectional time series modeling, but also enhances the prediction regression accuracy by utilizing the nonlinear fitting capability of MLP.

[0216] Exemplarily, in step S4, the generator and the discriminator are trained competitively with each other, which can further improve the accuracy of ship track prediction.

[0217] Specifically, the discriminator is equipped with a D-encoder and MLP;

[0218] The discriminator is a traditional binary classification model. In this invention, the discriminator distinguishes between real ship track data and predicted ship track data generated by the generator. The discriminator processes the input track data through a D-encoder and an MLP, outputting the probability that the input data is identified as a real track. The formulas are shown in Equations (27) and (28).

[0219] (27)

[0220] (28)

[0221] Where, is the hidden state matrix output by Bi-LSTM, are the weight matrices of Bi-LSTM and MLP respectively, are the biases of Bi-LSTM and MLP respectively, represents the real track sequence of the target ship, where For the target ship The actual latitude, longitude, speed and heading at each moment; Represents the predicted track sequence of the target ship output by the generator, where For the target ship The predicted latitude, longitude, speed and heading at each moment, is the probability that the input track is identified as a true track.

[0222] During training, the generator and discriminator compete with each other through an adversarial learning mechanism, gradually optimizing their respective parameters. Ultimately, the generator is able to generate predicted track data that is highly similar to the actual track data, thus achieving high-precision track prediction.

[0223] To further illustrate the effects of the embodiments of the present invention, the following experiments were conducted.

[0224] The model proposed in this invention is applicable to various typical navigation scenarios such as waters near ports, narrow waterways, and converging waters with complex ship traffic flow characteristics, and is verified using the waters near a certain port as an example.

[0225] 1. Evaluation indicators

[0226] The proposed model is applicable to typical navigation scenarios with complex ship traffic flow characteristics, such as near-port waters, narrow waterways, and confluences. To evaluate the effectiveness and accuracy of the GAN-MM model in track prediction, AIS data from the waters near Qingdao Port was used to validate the proposed model. Model performance was also compared using prediction methods such as GRU, MLP, SVR, ARIMA, LSTM, Bi-LSTM, LSTM-Attention, and GAN.

[0227] At the same time, in order to verify the accuracy of ship track prediction, MAE, MAPE, RMSE, MSE and SMAPE are used as evaluation indicators, as shown in Equations (29) to (33):

[0228] (29)

[0229] (30)

[0230] (31)

[0231] (32)

[0232] (33)

[0233] 2. Analysis of experimental results

[0234] To verify the performance of the GAN-MM model, Tables 1 to 4 list the total prediction errors when the time interval is set to 30s, 1min, 2min, and 5min. Figures 4 to 7 The improvement in prediction accuracy of the proposed model compared with other algorithms is detailed. Figure 4 Comparison chart of prediction errors when the time interval is 30s, Figure 5 Comparison chart of prediction errors when the time interval is 1 minute, Figure 6 Comparison chart of prediction errors when the time interval is 2 minutes. Figure 7 Comparison chart of prediction errors when the time interval is 5 minutes;

[0235] Table 1 Prediction error when the time interval is set to 30 seconds

[0236]

[0237] Table 2 Prediction error when the time interval is set to 1 minute

[0238]

[0239] Table 3 Prediction error when the time interval is set to 2 minutes

[0240]

[0241] Table 4 Prediction error when the time interval is set to 5 minutes

[0242]

[0243] Analyze Tables 1-4 and Figure 4-Figure 7 Across all time intervals, GAN-MM achieved the lowest total prediction error, followed by GAN, and SVR achieved the highest total prediction error. Compared to the other compared algorithms, GAN-MM achieved a maximum improvement in prediction error of nearly 100%, while the minimum improvement was 8.43%. Furthermore, while GAN-MM achieved the best prediction performance, its performance varied across time intervals. GAN-MM achieved the best prediction performance at a 2-minute time interval and the worst performance at a 5-minute time interval.

[0244] Figures 8-19 The figure further shows the comparison between the predicted track and the actual track of each algorithm, as well as the comparison between each algorithm in terms of longitude and latitude. Figures 8-19It can be seen that in the numerical experiments of the four time intervals, the deviation between GAN-MM and the true track is the smallest, followed by GAN, while SVR has the largest deviation from the true track. Therefore, compared with other algorithms, GAN-MM has the best performance.

[0245] In summary, considering the influence of multi-ship coupling, the model proposed in this invention has the best track prediction performance and can significantly improve the accuracy and stability of track prediction.

[0246] The above description is only a preferred specific implementation method of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A ship track prediction method with multi-ship coupling effect, characterized in that: The method includes: S1, selects the closest approach distance DCPA and the closest approach time TCPA as evaluation indicators to identify the adjacent ships that have coupling effects on the target ship; S2, MLP is used to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships, and the target ship is extracted. The potential motion eigenvector affected by coupling at every moment ; S3, the target ship The potential motion eigenvector affected by coupling at every moment The input is encoded into the Bi-LSTM-based G-encoder equipped with the generator, and the hidden feature representation of the target ship track is output. ; Represent the hidden features of the target ship track Input into the multi-head probabilistic sparse self-attention, remove redundant information, retain attention to key information, and obtain weighted hidden features ; The weighted hidden features The input is sent to the Bi-LSTM-based G-decoder on the generator for decoding, and the feature representation of the predicted trajectory is output. ; The characteristics of the predicted track are represented by The input is fed into the MLP on the generator, which outputs the predicted track of the target ship. S4, input the actual track and predicted track into the D-encoder based on Bi-LSTM equipped with the discriminator to obtain the hidden state matrix output by Bi-LSTM , the hidden state matrix output by Bi-LSTM The input is fed into the MLP of the discriminator, which ultimately generates the probability that the input track is recognized as a true track. .

2. The ship track prediction method for multi-ship coupling effect according to claim 1 is characterized in that: In step S1, the closest approach distance DCPA and the closest approach time TCPA are selected as evaluation indicators to identify adjacent ships that have coupling effects on the target ship, including: The latitude, longitude, speed and heading of the target ship and adjacent ships are: and ; Calculate the relative distance between the target ship and the adjacent ship based on their latitude and longitude coordinates; as shown in formula (1)-formula (2): （1） （2） Where, is the relative distance between the target ship and the adjacent ships, is the radius of the Earth; According to the speed and heading of the target ship and the adjacent ship, the relative speed and heading between the two are calculated, as shown in formula (3)-formula (4): （3） （4） Where, is the relative speed, is the relative heading, is the arccosine function; According to the relative distance, relative speed and relative heading calculated above, the corresponding DCPA and TCPA are calculated as shown in formula (5)-formula (6): （5） （6） Where, is the relative bearing of the target ship; Vessels with DCPA less than 3 NM and TCPA less than 30 minutes are identified.

3. The ship track prediction method for multi-ship coupling effect according to claim 1, characterized in that: In step S2, MLP is used to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the neighboring ships, and the target ship is extracted. The potential motion eigenvector affected by coupling at every moment , including: MLP establishes a complex nonlinear relationship model through hidden layers and activation functions, and automatically learns hierarchical feature representation from the original data by processing high-dimensional data, extracts the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the neighboring ships, and extracts the potential motion feature vector of the target ship based on the nonlinear relationship between them, as shown in formula (7)-formula (9): （7） Where, is the input feature matrix of MLP, For the target ship The feature vector at the moment, including latitude, longitude, speed, and heading information, ; The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, ; The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, The first one that produces coupling effects on the target ship Neighboring ships The feature vector at the moment, including latitude, longitude, speed, and heading information, ; Formula (7) integrates the state information of the target ship and neighboring ships into an input matrix for subsequent feature extraction; （8） Where, For the The output of the hidden layer, For the The weight matrix of the layer, For the The output of the layer, For the The bias of the layer, is the total number of hidden layers in the MLP, is the activation function; Formula (8) describes the first The calculation process of the hidden layer extracts features layer by layer through linear transformation and nonlinear activation function; （9） Where, For the target ship The potential motion eigenvector affected by the coupling at each moment, is the weight matrix of the output layer in MLP, is the output of the Lth hidden layer, is the bias of the output layer in MLP; Formula (9) is the calculation process of the MLP output layer, which generates the target ship The potential motion eigenvector affected by the coupling at each moment.

4. The ship track prediction method for multi-ship coupling effect according to claim 1, characterized in that: In step S3, the generator is equipped with: G-encoder, multi-head probabilistic sparse self-attention module, G-decoder and MLP; Bi-LSTM is selected as the core component of G-encoder. Bi-LSTM consists of two layers of long short-term memory network LSTM stacked together. The first layer LSTM inputs data forward, and the second layer LSTM inputs data backward. The LSTM unit contains a forget gate, an input gate, a cell state, and an output gate. The above components control the information flow through the sigmoid activation function and point-by-point multiplication to achieve the modeling of long-term dependencies in sequence data. The long short-term memory network LSTM includes: the input vector enters the forget gate, the input gate, and the output gate, as shown in formula (10)-formula (12): （10） Where, is the forget gate vector, is the sigmoid function, Transform the weight of the historical state of the forget gate, For the target ship The potential motion eigenvector affected by the coupling at each moment, Transform the weights of the input features of the forget gate, for The hidden state of the moment, is the bias of the forget gate; Formula (10) determines whether to remember the cell state information of the previous unit; （11） （12） Where, is the input gate vector, is the historical state transformation weight of the input gate, Transform the weights of the input features of the input gate, is the bias of the input gate, is the output gate vector, is the historical state transformation weight of the output gate, Transform the weights of the input features of the output gate, is the bias of the output gate, is the activation function; Formula (11) controls the new information that needs to be updated to the cell state; Formula (12) determines which information in the cell state is remembered as the output state; Output the candidate cell state vector and update the cell state vector, as shown in formula (13)-formula (14): （13） （14） Where, is the candidate cell state vector, is the historical state transformation weight of the candidate cell state, is the input feature transformation weight of the candidate cell state, is the cell state vector, is the bias of the candidate cell state, for The cell state vector at the moment; Formula (13) generates new information that may be added to the cell state; Formula (14) combines the decisions of the forget gate and the input gate to update the cell state; Update the hidden state vector according to the output gate vector and the cell state vector, as shown in formula (15): （15） Where, is the hidden state vector output by the unidirectional LSTM layer; Formula (15) updates the hidden state vector based on the output gate vector and the cell state vector.

5. The ship track prediction method for multi-ship coupling effect according to claim 4 is characterized in that: In step S3, in the Bi-LSTM, the forward propagation layer processes the sequence in time order to generate the hidden state , the backward propagation layer processes the sequence in reverse order to generate the hidden state ; The bidirectional hidden state at each moment is obtained by formula (16): （16） Where, is the bidirectional hidden state of the final output of Bi-LSTM; Formula (16) will and Splicing to form a bidirectional hidden state containing context information for subsequent calculations.

6. The ship track prediction method for multi-ship coupling effect according to claim 4 is characterized in that: The multi-head probabilistic sparse self-attention module combines the multi-head attention mechanism with the sparsification strategy, focusing on the coupling effects of neighboring ships on the target ship, as shown in Formula (17)-Formula (24): （17） （18） （19） Where, are query matrix, key matrix and value matrix respectively, for the reason The hidden matrix consists of vectors, ; is the number of timestamps, is the weight matrix of query, key and value; In order to reduce computational complexity and focus on processing the most relevant information, a sparsification strategy is adopted. The sparsity measure of each query is calculated as shown in formula (20): （20） Where, for No. row vectors, for No. row vectors, is the matrix transpose symbol, is a row vector The transpose of is the length of the key, is the feature dimension of the key vector in the model, is the sparsification metric score, is the bond matrix; Formula (20) is used to determine the attention distribution between each query and all keys, thereby obtaining the sparse measurement value between each query and key. The obtained multiple sparse measurement values ​​are sorted from large to small, and the top The results are used in subsequent calculations; the calculation formula for sparse attention weight is shown in formula (21): （21） Where, is the sparse attention weight, for indivual The corresponding query element set, is the bond matrix The transposed matrix of Formula (21) is used to calculate the attention weight between the query and the key, and the probability distribution is obtained through softmax normalization to determine the degree of attention of the model at different positions; The probabilistic sparse attention mechanism is extended to Attention heads are used to capture different features, as shown in formulas (22)-(24). The output captures the global dependencies and complex features in the data to analyze and predict future trajectories. Calculate the The output of an attention head: （22） Where, For the The output of an attention head, is the sparse attention weight, is the value matrix; The outputs of all attention heads are concatenated and linearly transformed: （23） Where, is the output of multi-head probabilistic sparse self-attention, For splicing operation, is the output of the first attention head, is the output of the second attention head, For the The output of an attention head, is the learnable weight matrix of the output layer; Get the final output: （24） Where, is the final output of the multi-head probabilistic sparse self-attention.

7. The ship track prediction method for multi-ship coupling effect according to claim 4 is characterized in that: The G-decoder gradually generates the feature representation of the predicted trajectory based on the output of the G-encoder and the multi-head probabilistic sparse self-attention module. ; G-decoder also uses Bi-LSTM; then introduce MLP after the decoder to convert the hidden representation into Latitude, longitude, speed and heading data of the predicted track at all times , that is, the predicted track, as shown in formula (25)-formula (26): （25） （26） Where, are the weight matrices of the Bi-LSTM and MLP networks respectively, are the biases of Bi-LSTM and MLP networks respectively.

8. The ship track prediction method for multi-ship coupling effect according to claim 1, characterized in that: In step S4, the discriminator is equipped with a D-encoder and an MLP; The discriminator processes the input track data through the D-encoder and MLP, and outputs the probability that the input data is identified as a true track; as shown in formulas (27) and (28): （27） （28） Where, is the hidden state matrix output by Bi-LSTM, are the weight matrices of Bi-LSTM and MLP respectively, are the biases of Bi-LSTM and MLP respectively, represents the real track sequence of the target ship, where For the target ship The actual latitude, longitude, speed and heading at each moment; Represents the predicted track sequence of the target ship output by the generator, where For the target ship The predicted latitude, longitude, speed and heading at each moment, is the probability that the input track is identified as a true track.

9. The ship track prediction method for multi-ship coupling effect according to claim 1, characterized in that: After step S4, the generator and discriminator compete with each other through an adversarial learning mechanism and gradually optimize their respective parameters; ultimately, predicted track data that is highly similar to the actual track data is generated, achieving high-precision track prediction.

10. A ship track prediction system with multi-ship coupling effect, characterized in that: The method for predicting ship tracks under multi-ship coupling effects as described in any one of claims 1 to 9 is implemented, and the system comprises: The ship identification module based on multi-ship coupling influence is used to select the closest approach distance DCPA and the closest approach time TCPA as evaluation indicators to identify adjacent ships that have a coupling influence on the target ship; The feature extraction module uses MLP to quantify the nonlinear relationship between the motion characteristics of the target ship and the motion characteristics of the adjacent ships, and extracts the target ship's The potential motion eigenvector affected by coupling at every moment ; Generator for placing the target ship in The potential motion eigenvector affected by coupling at every moment The input is encoded into the Bi-LSTM-based G-encoder equipped with the generator, and the hidden feature representation of the target ship track is output. ; Represent the hidden features of the target ship track Input into the multi-head probabilistic sparse self-attention, remove redundant information, retain attention to key information, and obtain weighted hidden features ; The weighted hidden features The input is sent to the Bi-LSTM-based G-decoder on the generator for decoding, and the feature representation of the predicted trajectory is output. ; The characteristics of the predicted track are represented by The input is fed into the MLP on the generator, which outputs the predicted track of the target ship. The discriminator inputs the real track and the predicted track into the D-encoder based on Bi-LSTM carried by the discriminator to obtain the hidden state matrix output by Bi-LSTM , the hidden state matrix output by Bi-LSTM The input is fed into the MLP of the discriminator, which ultimately generates the probability that the input track is recognized as a true track. .

Images