Unmanned surface ship cluster trajectory prediction method and system in uncertain environment

Abstract

The invention discloses an unmanned surface ship cluster trajectory prediction method and system in an uncertain environment, and the method comprises the steps: 1, employing a self-adaptive linear frequency modulation mode decomposition algorithm, decomposing original trajectory data into Q signal components, employing the first k signal components, including the kth signal component, in the Q signal components as useful signal components, and employing the first k signal components as the first k signal components; taking the residual signal components as noise of the ship trajectory data, and filtering the residual signal components; step 2, constructing a model for the time correlation of the ship trajectory based on Mama, and extracting a motion mode of the ship through the constructed model; step 3, constructing an interaction diagram according to the spatial positions of each ship and the surrounding ships in each time step, updating information of each node and edge of the interaction diagram, and obtaining interaction characteristics of the ships; and 4, fusing the motion mode in the step 2 and the interaction characteristics in the step 3, and predicting the future trajectory of the ship. According to the method, the influence of noise on ship trajectory prediction can be reduced, and the prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned surface ship cluster trajectory prediction, and in particular to a method and system for unmanned surface ship cluster trajectory prediction in an uncertain environment. Background Art

[0002] With the rapid development of artificial intelligence technology, unmanned ship swarms play an increasingly important role in maritime military operations. In order to ensure the safety and operational efficiency of unmanned ship surface swarms, it is crucial to understand the direction and location of the unmanned ship swarms at a specific time. Predicting the trajectories of one's own and friendly unmanned ship swarms through ship trajectory prediction methods can reduce the risk of collision and congestion and improve the safety and efficiency of maritime operations.

[0003] The ship trajectory prediction method based on deep learning is the mainstream method at present. Through deep neural networks, features are extracted from the historical trajectory of the ship, and its future movement path is accurately predicted. In order to predict the trajectory more accurately, it is necessary to extract the long-term time correlation in the trajectory. However, the existing methods have limited modeling capabilities for long-term predictions and forget the previously learned knowledge parameters. In addition, due to the high dynamics of the sea surface and the uncertainty of the environment, such as complex environment and long interval distance, there is a lot of noise in the trajectory data obtained from lidar, which greatly affects the accuracy of trajectory prediction.

[0004] For example, there is an existing ship trajectory prediction method based on AIS data and LSTM. This method obtains historical trajectory predictions based on the AIS system, makes full use of kinematic knowledge to finely process AIS data, and deeply captures the more subtle movement characteristics of ships. It also uses an LSTM-based network structure to improve the network structure's ability to predict data, and uses an improved Cell structure to provide the model with a more flexible feature extraction tool, effectively improving the accuracy of the prediction model.

[0005] This method uses an LSTM-based network structure to model the time dependency of the trajectory, but is unable to extract long-term motion features, resulting in large errors when predicting long-term ship trajectories.

[0006] Patent document CN117291295A discloses a JMIMHA - BiGRU ship trajectory prediction method based on an improved whale algorithm. Ship trajectories are obtained from the AIS system, data samples are constructed, and a ship trajectory time - series data set is constructed using the data samples. A JMIMHA - BiGRU ship trajectory prediction model based on a multi - head self - attention mechanism is constructed, and the model parameters of JMIMHA - BiGRU are trained using the ship travel information time - series data set and the whale optimization algorithm. The current ship trajectory is input into the trained JMIMHA - BiGRU ship trajectory prediction model to obtain the predicted ship trajectory.

[0007] This method uses the ship trajectories obtained from the AIS system to train and evaluate a JMIMHA - BiGRU ship trajectory prediction model based on a multi - head self - attention mechanism. There is a large amount of noise in the collected data, resulting in a large prediction error in the trained model. Summary of the Invention

[0008] The purpose of the present invention is to provide a method and system for predicting the trajectories of unmanned surface ship clusters in an uncertain environment, which can reduce the influence of noise on ship trajectory prediction and improve the prediction accuracy.

[0009] To achieve the above - mentioned purpose, the present invention provides a method for predicting the trajectories of unmanned surface ship clusters in an uncertain environment, which includes:

[0010] Step 1: Using the adaptive linear frequency - modulated mode decomposition algorithm, decompose the original trajectory data s(t) into Q signal components. The q - th signal component is denoted as s q (t), q ∈ (1, 2,..., Q). Take the first k signal components including the k - th signal component among the Q signal components as useful signal components S 1 ={s q (t)|q ∈ (1, 2,...k)}. Take the remaining signal components as the noise of the ship trajectory data S 2 ={s q (t)|q ∈ (k + 1, k + 2,..., Q)}, k < Q. Filter out the noise S 2 and output the denoised trajectory data

[0011] Step 2: Based on Mamba, construct a model for the time correlation of ship trajectories, and extract the motion pattern M of the ship through the constructed model;

[0012] Step 3: Construct an interaction graph G according to the spatial positions of each ship and the surrounding ships at each time step s , and update each node s in the interaction graph G and edge E s information, and obtain the interactive feature I of the ship;

[0013] Step 4: Fuse the motion pattern M from step 2 and the interactive feature I from step 3 to predict the future trajectory of the ship.

[0014] Furthermore, in step 1, the first k of the Q signal components and including the kth signal component are used as the useful signal component S 1 Specifically include:

[0015] Step 121, based on the initial signal, calculate the signal component s by formula (2) q (t), where the first signal component s is calculated 1 (t), the initial signal is the original trajectory data s(t):

[0016]

[0017] Where b q (t) is s q (t) time-varying amplitude at time step t, d q (t) is s q (t) The phase at time step t is calculated by equation (3), f q (t) They represent the signal decomposition process s q (t) Corresponding instantaneous frequency, reference frequency, and estimated frequency;

[0018]

[0019] Step 122, the initial signal and the current signal component s q The difference after the subtraction operation of (t) is used as the new initial signal, and the process returns to step S11 until the first k signal components up to and including the kth signal component are obtained.

[0020] Furthermore, step 2 specifically includes:

[0021] Step 21: Generate the embedding H of the ship trajectory at time step t through a nonlinear multilayer perceptron t =MLP(X t ,Θ N ), and then obtain the embedding H of the ship trajectory at all time steps, where They represent the horizontal and vertical coordinates of the nth ship cluster at time t, N represents the total number of ship clusters, Θ N is the learnable parameter of the multilayer perceptron;

[0022] Step 22, expanding the dimension of the embedding H in step 21 to obtain a first branch p and a second branch o for parallel processing;

[0023] Step 23, according to the first branch p of step 22, a high-dimensional representation p' of the ship motion is obtained by capturing the time dependency;

[0024] Step 24, according to the high-dimensional representation p' of the ship motion in step 23, the motion feature P of the ship is obtained by the selective state space model SelectiveSSM described by Mamba's formula (6), and according to the second branch o, the SiLU activation function, Sigmoid or tanh activation function described by formula (7) is used to generate the gating signal o';

[0025]

[0026] o'=SiLU(o) (7)

[0027] In the formula, Softplus represents the activation function, Parameter represents a constant, Broadcast represents the broadcast operation, A represents the structured state matrix, the input matrix B that projects p' into the state space, the output matrix C of the state space and the discretized time step Δ, B = Linear B (p'), C = Linear C (p'), Δ=Softplus(Parameter+Broadcast(Linear Δ (p'))), and then use Δ to discretize A and B respectively to get

[0028] Step 25, multiply the motion feature P of the ship in step 24 by the gated signal i', and then adjust the dimension through linear transformation Linear to obtain the motion mode M of the ship, as shown in the following formula (8):

[0029]

[0030] Furthermore, the interaction graph G is transformed into s The hidden features of each node and its neighbors are aggregated to obtain the dynamic spatial interaction of the ship, thereby obtaining the interaction features of the ship

[0031]

[0032] Where l represents the network layer number of the second graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the second graph convolutional network, The updated edge E of expression (10) smnThe normalized adjacency matrix is They represent the hidden features of the lth, l+1th, and Lth network layers of the second graph convolutional network, respectively. represents the learnable parameters of the second graph convolutional network, σ represents the nonlinear activation function of the second graph convolutional network, and R represents different relationship graphs A r A collection of;

[0033]

[0034] In the formula, Softmax represents the activation function, d k represents the scaling factor, T represents the matrix transpose, Q sm Denotes the feature vector h m The query vector, Q sm =f sQ (h m ), K sn Denotes the feature vector h n The key vector, K sn =f sK (h n ), V sm Indicates v m The eigenvector h m A vector of values, V sm =f sV (h m ), f sQ 、f sK and f sV Both represent linear transformations, h m 、h n They are v m and v n The eigenvector of

[0035] Update interaction graph G s Each node The information of is used to obtain the characteristic matrix in formula (9):

[0036]

[0037] Where l represents the network layer number of the first graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the first graph convolutional network, Represents edge E s The normalized adjacency matrix is They represent the hidden features of the lth and l+1th network layers of the first graph convolutional network, respectively. represents the parameters of the lth network layer of the first graph convolutional network, Represents the interaction graph G sThe initial feature vector of each node, σ represents the non-linear activation function of the first graph convolutional network, and R represents the set of different relational graphs A r of.

[0038] Furthermore, the fusion method in step 4 includes: predicting the future position of the nth ship at time step t through a temporal convolution function, and calculating the predicted ship speed;

[0039]

[0040] where concat represents the concatenation operation and TCN represents the temporal convolutional network;

[0041] The fusion method is trained by minimizing the negative log-likelihood loss:

[0042]

[0043] where W represents the trainable parameter, L n represents the loss function, T obs represents the length of the ship's observed trajectory, T pred represents the length of the predicted trajectory, and P represents the probability density function.

[0044] The present invention also provides an unmanned surface ship cluster trajectory prediction system under an uncertain environment, which includes:

[0045] A noise filtering module, which is used to decompose the original trajectory data s(t) into Q signal components by using the adaptive linear frequency modulation mode decomposition algorithm. The qth signal component is expressed as s q (t), q ∈ (1, 2,..., Q). The first k signal components including the kth signal component among the Q signal components are used as useful signal components S 1 = {s q (t)|q ∈ (1, 2,...k)}. The remaining signal components are used as the noise S of the ship trajectory data 2 = {s q (t)|q ∈ (k + 1, k + 2,..., Q)}, k < Q. The noise S 2 is filtered out, and the denoised trajectory data is output

[0046] A motion mode extraction module, which is used to construct a model based on Mamba for the temporal correlation of the ship trajectory, and extract the motion mode M of the ship through the constructed model;

[0047] A ship interaction extraction module, which is used to construct an interaction graph G according to the spatial positions of each ship and its surrounding ships at each time step s , and update the interaction graph Gs Each node and edge E s information, and obtain the interactive feature I of the ship;

[0048] The ship trajectory prediction module is used to fuse the motion pattern M of the motion pattern extraction module and the interaction feature I of the ship interaction extraction module to predict the future trajectory of the ship.

[0049] Furthermore, the noise filtering module specifically includes:

[0050] A signal component calculation unit, which is used to calculate the signal component s according to the initial signal through formula (2) q (t), where the first signal component s is calculated 1 (t), the initial signal is the original trajectory data s(t):

[0051]

[0052] Where b q (t) is s q (t) time-varying amplitude at time step t, d q (t) is s q (t) The phase at time step t is calculated by equation (3), f q (t) They represent the signal decomposition process s q (t) Corresponding instantaneous frequency, reference frequency, and estimated frequency;

[0053]

[0054] The difference calculation unit is used to calculate the difference between the initial signal and the current signal component s q The difference after the subtraction operation of (t) is used as a new initial signal and then input into the signal component calculation unit until the first k and including the kth signal components are obtained.

[0055] Furthermore, the motion pattern extraction module specifically includes:

[0056] A nonlinear multilayer perceptron that receives the denoised trajectory data X t , generating the embedding H of the ship trajectory at time step t t =MLP(X t ,Θ N ), and then obtain the embedding H of the ship trajectory at all time steps, where They represent the horizontal and vertical coordinates of the nth ship cluster at time t, N represents the total number of ship clusters, Θ N is the learnable parameter of the multilayer perceptron;

[0057] A dimension expansion unit, which is used to expand the dimension of the embedding H of the nonlinear multilayer perceptron to obtain a first branch p and a second branch o for parallel processing;

[0058] A time dependency capturing unit, which is used to obtain a high-dimensional representation p' of the ship motion by capturing the time dependency according to the first branch p of the dimensional expansion unit;

[0059] A motion feature acquisition unit, which is used to obtain the high-dimensional representation p' of the ship's motion according to the time-dependent capture unit, obtain the ship's motion feature P through the selective state space model SelectiveSSM described by Mamba's formula (6), and generate a gating signal o' according to the second branch o using the SiLU activation function, Sigmoid or tanh activation function described by formula (7);

[0060]

[0061] o'=SiLU(o) (7)

[0062] In the formula, Softplus represents the activation function, Parameter represents a constant, Broadcast represents the broadcast operation, A represents the structured state matrix, the input matrix B that projects p' into the state space, the output matrix C of the state space and the discretized time step Δ, B = Linear B (p'), C = Linear C (p'), Δ=Softplus(Parameter+Broadcast(Linear Δ (p'))), and then use Δ to discretize A and B respectively to get

[0063] The motion mode acquisition unit is used to multiply the motion feature P of the ship obtained by the motion feature acquisition unit with the gate signal o', and then perform dimensionality adjustment through linear transformation Linear to obtain the motion mode M of the ship, as shown in the following formula (8):

[0064]

[0065] Furthermore, the interaction graph G is transformed into s The hidden features of each node and its neighbors are aggregated to obtain the dynamic spatial interaction of the ship, thereby obtaining the interaction features of the ship

[0066]

[0067] Where l represents the network layer number of the second graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the second graph convolutional network, The updated edge E of expression (10) smn The normalized adjacency matrix is They represent the hidden features of the lth, l+1th, and Lth network layers of the second graph convolutional network, respectively. represents the learnable parameters of the second graph convolutional network, σ represents the nonlinear activation function of the second graph convolutional network, and R represents different relationship graphs A r A collection of;

[0068]

[0069] In the formula, Softmax represents the activation function, d k represents the scaling factor, T represents the matrix transpose, Q sm Denotes the feature vector h m The query vector, Q sm =f sQ (h m ), K sn Denotes the feature vector h n The key vector, K sn =f sK (hn), V sm Indicates v m The eigenvector h m A vector of values, V sm =f sV (h m ), f sQ 、f sK and f sV Both represent linear transformations, h m 、h n They are v m and v n The eigenvector of

[0070] Update interaction graph G s Each node The information of is used to obtain the characteristic matrix in formula (9):

[0071]

[0072] Where l represents the network layer number of the first graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the first graph convolutional network, Represents edge E s The normalized adjacency matrix is They represent the hidden features of the lth and l+1th network layers of the first graph convolutional network, respectively. represents the parameters of the lth network layer of the first graph convolutional network, Represents the interaction graph G s The initial feature vector of each node in , σ represents the nonlinear activation function of the first graph convolutional network, and R represents different relationship graphs A r A collection of .

[0073] Furthermore, the fusion method of the ship trajectory prediction module includes: using a time convolution function to calculate the future position of the nth ship at time step t Make predictions and calculate predicted ship speeds;

[0074]

[0075] In the formula, concat represents the concatenation operation, and TCN represents the temporal convolutional network;

[0076] The fusion method is trained by minimizing the negative log-likelihood loss:

[0077]

[0078] Where W represents the trainable parameter, L n represents the loss function, T obs represents the length of the ship's observation track, T pred represents the length of the predicted trajectory, and P represents the probability density function.

[0079] The present invention has the following advantages due to the adoption of the above technical solution:

[0080] 1) Adaptive linear frequency modulation mode decomposition can effectively filter out the noise in the trajectory data, thereby improving the quality and reliability of the data. This efficient noise reduction process can significantly improve the performance of the trajectory prediction model, ensuring that the method is more reliable and accurate in complex marine environments.

[0081] 2) Using Mamba to model the long-term temporal correlation of ship trajectories can not only effectively capture the navigation mode of ships over a long period of time, but also has the ability to autonomously select valid historical data, and can intelligently filter out high-quality historical data to further reduce the impact of noise, thereby improving the accuracy and reliability of the model.

[0082] 3) Using spatial graphs to model the interactions between ships can better capture and represent the complex interactions between ships. The data structure of the graph can establish more accurate spatial topological position relationships, which not only improves the understanding of the interaction relationship, but also significantly reduces the impact of noise in the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1It is a flow chart of the trajectory prediction method of unmanned surface ship clusters in an uncertain environment.

[0084] Figure 2 It is a trajectory prediction system for unmanned surface ship clusters in uncertain environments. DETAILED DESCRIPTION

[0085] In the drawings, the same or similar reference numerals are used to represent the same or similar elements or elements with the same or similar functions. The embodiments of the present invention are described in detail below in conjunction with the drawings.

[0086] In the description of the present invention, the terms "center", "longitudinal", "lateral", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the scope of protection of the present invention.

[0087] like Figure 1 As shown, the unmanned surface ship cluster trajectory prediction method under an uncertain environment provided by an embodiment of the present invention includes:

[0088] Step 1: Adopt the adaptive linear frequency modulation mode decomposition algorithm to transform the noise S in the original trajectory data s(t) 2 Filter out and output the denoised trajectory data X t The method for obtaining the original trajectory data s(t) includes: obtaining the trajectory information of the ship cluster through the AIS system, using the average position of each ship in the ship cluster as the trajectory data, and constructing a data sample to provide input for subsequent algorithms.

[0089] Step 2: Build a model based on the time correlation of Mamba's ship trajectory, extract the ship's motion mode M through the built model, and autonomously select valid historical data to reduce the impact of noise.

[0090] Step 3: Construct an interaction graph G based on the spatial position of each ship and its surrounding ships at each time step. s , update the interaction graph G s Each node and edge E s The information of the ship is used to obtain the interaction characteristics I of the ship.

[0091] Step 4: Fuse the motion pattern M from step 2 and the interactive feature I from step 3 to predict the future trajectory of the ship.

[0092] In one embodiment, step 1 specifically includes:

[0093] Step 11, the original trajectory data s(t) can be decomposed into a set of intrinsic mode functions (IMFs) of different frequency components of the signal by the adaptive chirp mode decomposition (ACMD) algorithm described in the following formula (1), that is, Q signal components in this paper, and the qth signal component is represented by s q (t), q∈(1,2,...,Q).

[0094]

[0095] In the formula, A q (t) is s q (t) is the amplitude at time step t, f q (t) is s q (t) is the frequency at time step t, φ q is the initial phase.

[0096] Step 12: The first k signal components including the kth signal component among the Q signal components are used as the useful signal component S 1 ={s q (t)|q∈(1,2,...k)}, and the remaining signal component is used as the noise S of the ship trajectory data. 2 ={s q (t)|q∈(k+1,k+2,...,Q)},k <Q。

[0097] Step 13: convert the noise S in step 12 2 Filter out the useful signal component S according to step 12 1 , get the denoised trajectory data X t : And output.

[0098] In one embodiment, the step 12 of “taking the first k of the Q signal components and including the kth signal component as the useful signal component S” 1 " specifically includes step 121 and step 122:

[0099] Step 121, based on the initial signal, calculate the signal component s by formula (2) q (t), where the first signal component s is calculated 1 (t), the initial signal is the original trajectory data s(t):

[0100]

[0101] Where bq (t) is s q (t) time-varying amplitude at time step t, d q (t) is s q (t) The phase at time step t is calculated by equation (3), f q (t) They represent the signal decomposition process s q (t) Corresponding instantaneous frequency, reference frequency, and estimated frequency.

[0102]

[0103] Step 122, the initial signal and the current signal component s q The difference after the subtraction operation of (t) is used as the new initial signal, and the process returns to step S11 until the first k signal components up to and including the kth signal component are obtained.

[0104] Adaptive linear frequency modulation mode decomposition can effectively filter out the noise in the trajectory data, thereby improving the quality and reliability of the data. This efficient noise reduction process can significantly improve the performance of the trajectory prediction model, ensuring that the method is more reliable and accurate in complex marine environments.

[0105] In one embodiment, step 2 specifically includes:

[0106] Step 21: Generate the embedding H of the ship trajectory at time step t through a nonlinear multilayer perceptron t =MLP(X t ,Θ N ), and then the embedding H of the ship trajectory at all time steps is obtained. Among them, They represent the horizontal and vertical coordinates of the nth ship cluster at time t, N represents the total number of ship clusters, Θ N is the learnable parameter of the multilayer perceptron. The specific value is defined by the deep learning framework (pytorch) after MLP is passed through the back propagation process. N The value of is learned.

[0107] Step 22, expanding the dimension of the embedding H in step 21 to obtain a first branch p and a second branch o for parallel processing.

[0108] For example, the dimension of the embedding H can be expanded by the linear layer Linear described in formula (4), and the dimensions before and after the expansion are 64 and 256 respectively. Of course, the dimension can also be expanded by using a multi-layer perceptron MLP or other existing methods.

[0109]

[0110] Step 23, according to the first branch p of step 22, a high-dimensional representation p' of the ship motion is obtained by capturing the time dependency.

[0111] For example, the first branch p can be processed by the one-dimensional convolution Conv1d and SiLU activation function described in formula (5) to obtain a high-dimensional representation p' of the ship motion. The SiLU activation function can also be replaced by the Sigmoid or tanh activation function.

[0112] p' = SiLU(Conv1d(p)) (5)

[0113] Step 24: According to the high-dimensional representation p' of the ship motion in step 23, the ship motion feature P is obtained by using the selective state space model SelectgiveSSM described by Mamba's formula (6). According to the second branch o, the SiLU activation function, Sigmoid or tanh activation function described by formula (7) is used to generate a gating signal o' to further filter the trajectory data X output by the noise filtering module based on step 1. t The irrelevant information in the output is then used to output the motion mode M.

[0114]

[0115] o'=SiLU(o) (7)

[0116] Where Softplus represents the activation function, Parameter represents a constant, Broadcast represents the broadcast operation, A represents the structured state matrix, and p' is passed through the linear layer of deep learning (the following Linear B , Linear C , Linear Δ Both represent linear layers) and other linear projections to the input matrix B of the state space, the output matrix C of the state space, and the discretized time step Δ, B = Linear B (p'), C = Linear C (p'), Δ=Softplus(Parameter+Broadcast(Linear Δ (p'))), and then use Δ to discretize A and B respectively to get

[0117] Step 25, multiply the motion feature P of the ship in step 24 by the gate signal o', and then adjust the dimension through linear transformation Linear to obtain the motion mode M of the ship, as shown in the following formula (8), and the dimensions before and after the adjustment are 64 and 256 respectively:

[0118]

[0119] By using Mamba to model the long-term temporal correlation of ship trajectories, it can not only effectively capture the navigation patterns of ships over a long period of time, but also has the ability to autonomously select valid historical data, intelligently filter out high-quality historical data, and further reduce the impact of noise, thereby improving the accuracy and reliability of the model.

[0120] In one embodiment, in step 3, G s =(V s ,E s ) represents the position of the ship at time step t, and s has no actual physical meaning.

[0121] in, represents the coordinates of the nth ship at time step t That is, the interaction graph G s The nth node of Vs represents the interaction graph G s The set of midpoints, T obs Indicates the length of the ship's observation track, i.e. X t N represents the number of time steps in a specific time step, for example, when the time step t = 1s, N = 10 ships.

[0122] Represents the interaction graph G s The edge of E s Represents the interaction graph G s The set of middle edges, 1 represents the mth node and In connected state, 0 means and In disconnected state, the node The set of m ,node The set of n .

[0123] Update interaction graph G s Each node The information of is used to obtain the characteristic matrix in formula (9):

[0124]

[0125] Where l represents the network layer number of the first graph convolutional network such as graph convolutional network, graph neural network (GNN) or graph attention network (GAT), l∈{0,1,…,L}, L represents the total number of network layers of the first graph convolutional network, Represents edge E s The normalized adjacency matrix is They represent the hidden features of the lth and l+1th network layers of the first graph convolutional network, respectively. Represents the parameters of the lth network layer of the first graph convolutional network. Its specific value can be obtained through deep learning back-propagation update. The initial value is generally randomly assigned. Represents the interaction graph G s The initial feature vector of each node in , σ represents the nonlinear activation function of the first graph convolutional network, and R represents different relationship graphs A r A collection of .

[0126] According to the interaction graph G s For each node in the graph, the self-attention function described by formula (10) is used to update the interaction graph G s The edge E s Information:

[0127]

[0128] In the formula, E smn represents the updated edge, Softmax represents the activation function, and d k represents the scaling factor, which is obtained by dividing the input dimension by the number of attention heads to prevent the result from being too large or too small. T is the superscript, indicating the matrix transpose, Q sm Denotes the feature vector h m The query vector, Q sm =f sQ (h m ), K sn Denotes the feature vector h n The key vector, K sn =f sK (h n ), v sm Indicates v m The eigenvector h m A vector of values, V sm =f sV (h m ), f s Q, f sK and f sV Both represent linear transformations, h m 、h n They are v m and v n The feature vector of .

[0129] The interaction graph G is transformed into s The hidden features of each node and its neighbors are aggregated to obtain the dynamic spatial interaction of the ship, thereby obtaining the interaction features of the ship

[0130]

[0131] Where l represents the network layer number of the second graph convolutional network such as graph convolutional network, graph neural network (GNN) or graph attention network (GAT), l∈{0,1,…,L}, L represents the total number of network layers of the second graph convolutional network, E of expression (10) smn The normalized adjacency matrix is They represent the hidden features of the lth, l+1th, and Lth network layers of the second graph convolutional network, respectively. represents the learnable parameters of the second graph convolutional network, σ represents the nonlinear activation function of the second graph convolutional network, and R represents different relationship graphs A r A collection of .

[0132] In one embodiment, in step 4, the fusion method may be, for example: using a time convolution function to calculate the future position of the nth ship at time step t. Make a prediction and calculate the predicted ship speed.

[0133]

[0134] In the formula, concat represents the concatenation operation and TCN represents the temporal convolutional network.

[0135] In one embodiment, this embodiment trains the method by minimizing the negative log-likelihood loss:

[0136]

[0137] Where W represents the trainable parameter, L n represents the loss function, T obs represents the length of the ship's observation track, T pred represents the length of the predicted trajectory, and P represents the probability density function.

[0138] In the above embodiment, when predicting the future trajectory of the ship, a method based on the time convolution function is adopted. Other recurrent neural network methods may also be adopted, such as RNN (Recurrent Neural Network, RNN), LSTM (Long Short-Term Memory, LSTM), GRU (gated recurrent unit, GRU), etc.

[0139] By using a spatial graph to model the interactions between ships, the present invention can better capture and represent the complex interaction relationships between ship clusters. The data structure of the graph can establish a more accurate spatial topological position relationship, which not only improves the understanding of the interaction relationship but also significantly reduces the influence of noise in the data.

[0140] As Figure 2 shown, an unmanned surface ship cluster trajectory prediction system under an uncertain environment provided by an embodiment of the present invention includes a noise filtering module, a motion pattern extraction module, a ship interaction extraction module, and a ship trajectory prediction module, where:

[0141] The noise filtering module is used to decompose the original trajectory data s(t) into Q signal components by using an adaptive linear frequency modulation mode decomposition algorithm. The qth signal component is represented as s q (t), q ∈ (1, 2,..., Q). The first k signal components including the kth signal component among the Q signal components are used as useful signal components S 1 ={s q (t)|q ∈ (1, 2,..., k)}, and the remaining signal components are used as the noise S 2 ={s q (t)|q ∈ (k + 1, k + 2,..., Q)}, k < Q. The noise S 2 is filtered out, and the denoised trajectory data is output The noise filtering module decomposes the complex signal into several simpler components, uses these components to analyze and process the useful information in the original trajectory data, and filters out the remaining information as noise, that is, the unnecessary noise components.

[0142] The motion pattern extraction module is used to construct a model based on Mamba for the time correlation of the ship trajectory, and extract the motion pattern M of the ship through the constructed model. The motion pattern extraction module adopts an algorithm based on Mamba to model the long-term time correlation of the ship trajectory, capture the navigation pattern of the ship in a long time range, and has the ability to autonomously select effective historical data, and intelligently screen out high-quality historical data.

[0143] The ship interaction extraction module is used to construct an interaction graph G according to the spatial positions of each ship and its surrounding ships at each time step s and update each node s in the interaction graph G and the edge E sThe ship interaction extraction module constructs an interaction graph based on the trajectory data to model the interaction behavior between ships. The nodes represent the positions of the ships, and the edges represent the interaction relationships between the ships. The graph convolution layer is used to aggregate the information of the neighboring nodes and update the state of each node. Then, the attention mechanism is used to update the edge information and highlight the important interaction relationships.

[0144] The ship trajectory prediction module is used to fuse the motion pattern M of the motion pattern extraction module and the interaction feature I of the ship interaction extraction module to predict the future trajectory of the ship. The main work of the trajectory prediction module is to fuse the motion pattern and ship interaction information obtained above, and use the temporal convolutional network to predict the future trajectory of the ship.

[0145] In one embodiment, the noise filtering module specifically includes a signal component calculation unit and a difference calculation unit, wherein:

[0146] The signal component calculation unit is used to calculate the signal component s according to the initial signal through formula (2) q (t), where the first signal component s is calculated 1 (t), the initial signal is the original trajectory data s(t).

[0147] The difference calculation unit is used to calculate the difference between the initial signal and the current signal component s q The difference after the subtraction operation of (t) is used as a new initial signal and then input into the signal component calculation unit until the first k and including the kth signal components are obtained.

[0148] In one embodiment, the motion pattern extraction module specifically includes a nonlinear multilayer perceptron, a dimension expansion unit, a time dependency capture unit, a motion feature acquisition unit, and a motion pattern acquisition unit, wherein:

[0149] The nonlinear multilayer perceptron is used to receive the denoised trajectory data X t , generating the embedding H of the ship trajectory at time step t t =MLP(X t ,Θ N ), and then obtain the embedding H of the ship trajectory at all time steps, where They represent the horizontal and vertical coordinates of the nth ship cluster at time t, N represents the total number of ship clusters, Θ N are the learnable parameters of the multilayer perceptron.

[0150] The dimension expansion unit is used to expand the dimension of the embedding H of the nonlinear multilayer perceptron to obtain a first branch p and a second branch o for parallel processing.

[0151] The time dependency capturing unit is used to obtain a high-dimensional representation p' of the ship motion by capturing the time dependency according to the first branch p of the dimensional expansion unit.

[0152] The motion feature acquisition unit is used to capture the high-dimensional representation p' of the ship's motion according to the time dependency, and obtain the ship's motion feature P through the selective state space model SelectiveSSM described by Mamba's formula (6). According to the second branch o, the SiLU activation function, Sigmoid or tanh activation function described by formula (7) is used to generate the gating signal o'.

[0153] The motion mode acquisition unit is used to multiply the motion feature P of the ship obtained by the motion feature acquisition unit with the gate signal o', and then perform dimensionality adjustment through linear transformation Linear to obtain the motion mode M of the ship, as shown in formula (8).

[0154] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Those skilled in the art should understand that the technical solutions described in the above embodiments may be modified, or some of the technical features thereof may be replaced by equivalents; these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the trajectory of an unmanned surface ship cluster in an uncertain environment, characterized in that: include: Step 1, using the adaptive linear frequency modulation mode decomposition algorithm, decompose the original trajectory data s(t) into Q signal components. The q-th signal component is denoted as s q (t), where q ∈ (1, 2, ..., Q). Consider the first k signal components (including the k-th component) among the Q signal components as the useful signal components S 1 = {s q (t) | q ∈ (1, 2, ..., k)}, and regard the remaining signal components as the noise of the ship trajectory data S 2 = {s q (t) | q ∈ (k + 1, k + 2, ..., Q)}, where k < Q. Filter out the noise S 2 and output the denoised trajectory data Step 2, construct a model based on the time correlation of Mamba to the ship trajectory, and extract the motion mode M of the ship through the constructed model; Step 3: Construct an interaction graph G based on the spatial position of each ship and its surrounding ships at each time step. s , update the interaction graph G s Each node and edge E s information, and obtain the interactive feature I of the ship; Step 4: Fuse the motion pattern M from step 2 and the interactive feature I from step 3 to predict the future trajectory of the ship.

2. The method for predicting the trajectory of unmanned surface ship clusters in an uncertain environment as claimed in claim 1, characterized in that: Step 1: "Use the first k of the Q signal components and including the kth signal component as the useful signal component S 1 Specifically include: Step 121, based on the initial signal, calculate the signal component s by formula (2) q (t), where when calculating the first signal component s1(t), the initial signal is the original trajectory data s(t): Where b q (t) is s q (t) time-varying amplitude at time step t, d q (t) is s q (t) The phase at time step t is calculated by equation (3), f q (t) They represent the signal decomposition process s q (t) Corresponding instantaneous frequency, reference frequency, and estimated frequency; Step 122, the initial signal and the current signal component s q The difference after the subtraction operation of (t) is used as the new initial signal, and the process returns to step S11 until the first k signal components up to and including the kth signal component are obtained.

3. The method for predicting the trajectory of unmanned surface ship clusters in an uncertain environment as claimed in claim 1, characterized in that: Step 2 specifically includes: Step 21: Generate the embedding H of the ship trajectory at time step t through a nonlinear multilayer perceptron t =MLP(X t ,Θ N ), and then obtain the embedding H of the ship trajectory at all time steps, where They represent the horizontal and vertical coordinates of the nth ship cluster at time t, N represents the total number of ship clusters, Θ N is the learnable parameter of the multilayer perceptron; Step 22, expanding the dimension of the embedding H in step 21 to obtain a first branch p and a second branch o for parallel processing; Step 23, according to the first branch p of step 22, a high-dimensional representation p' of the ship motion is obtained by capturing the time dependency; Step 24, according to the high-dimensional representation p' of the ship motion in step 23, the motion feature P of the ship is obtained by the selective state space model SelectiveSSM described by Mamba's formula (6), and according to the second branch o, the SiLU activation function, Sigmoid or tan h activation function described by formula (7) is used to generate the gating signal o'; o'=SiLU(o) (7) Where Spftplus represents the activation function, Parameter represents a constant, Broadcast represents the broadcast operation, A represents the structured state matrix, the input matrix B that projects p' into the state space, the output matrix C of the state space and the discretized time step Δ, B = Linear B (p'), C = Linear C (p'), Δ=Softplus(Parameter+Broadcast(Linear Δ (p'))), and then use Δ to discretize A and B respectively to get Step 25, multiply the motion feature P of the ship in step 24 by the gated signal o', and then adjust the dimension through linear transformation Linear to obtain the motion mode M of the ship, as shown in the following formula (8):

4. The method for predicting the trajectory of an unmanned surface ship cluster in an uncertain environment according to any one of claims 1 to 3, characterized in that: The interaction graph G is transformed into s The hidden features of each node and its neighbors are aggregated to obtain the dynamic spatial interaction of the ship, thereby obtaining the interaction features of the ship Where l represents the network layer number of the second graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the second graph convolutional network, The updated edge E of expression (10) smn The normalized adjacency matrix is They represent the hidden features of the lth, l+1th, and Lth network layers of the second graph convolutional network, respectively. represents the learnable parameters of the second graph convolutional network, σ represents the nonlinear activation function of the second graph convolutional network, and R represents different relationship graphs A r A collection of; In the formula, Softmax represents the activation function, d k represents the scaling factor, T represents the matrix transpose, Q sm Denotes the feature vector h m The query vector, Q sm =f sQ (h m ), K sn Denotes the feature vector h n The key vector, K sn =f sK (h n ), V sm Indicates v m The eigenvector h m A vector of values, V sm =f sV (h m ), f sQ 、f sK and f sV Both represent linear transformations, h m 、h n They are v m and v n The eigenvector of Update interaction graph G s Each node The information of , obtains the characteristic matrix in formula (9): Where l represents the network layer number of the first graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the first graph convolutional network, Represents edge E s The normalized adjacency matrix is They represent the hidden features of the lth and l+1th network layers of the first graph convolutional network, respectively. represents the parameters of the lth network layer of the first graph convolutional network, Represents the interaction graph G s The initial feature vector of each node in , σ represents the nonlinear activation function of the first graph convolutional network, and R represents different relationship graphs A r A collection of .

5. The method for predicting the trajectory of unmanned surface ship clusters in an uncertain environment as claimed in claim 4, characterized in that: The fusion method of step 4 includes: using the time convolution function to calculate the future position of the nth ship at time step t Make predictions and calculate predicted ship speeds; In the formula, concat represents the concatenation operation, and TCN represents the temporal convolutional network; The fusion method is trained by minimizing the negative log-likelihood loss: Where W represents the trainable parameter, L n represents the loss function, T obs represents the length of the ship's observation track, T pred represents the length of the predicted trajectory, and P represents the probability density function.

6. A trajectory prediction system for unmanned surface ship clusters in an uncertain environment, characterized in that: include: A noise filtering module, which is used to decompose the original trajectory data s(t) into Q signal components by using an adaptive linear chirp mode decomposition algorithm. The q-th signal component is expressed as s q (t), q ∈ (1, 2,..., Q). The first k signal components among the Q signal components, including the k-th signal component, are used as useful signal components S 1 ={s q (t)|q ∈ (1, 2,...k)}. The remaining signal components are used as the noise of the ship trajectory data S 2 ={s q (t)|q ∈ (k + 1, k + 2,..., Q)}, k < Q. The noise S 2 is filtered out to output the denoised trajectory data A motion mode extraction module is used to construct a model based on the time correlation of Mamba to the ship's trajectory, and extract the ship's motion mode M through the constructed model; The ship interaction extraction module is used to construct an interaction graph G according to the spatial position of each ship and its surrounding ships at each time step. s , update the interaction graph G s Each node and edge E s information, and obtain the interactive feature I of the ship; The ship trajectory prediction module is used to fuse the motion pattern M of the motion pattern extraction module and the interaction feature I of the ship interaction extraction module to predict the future trajectory of the ship.

7. The unmanned surface ship cluster trajectory prediction system in an uncertain environment as claimed in claim 6, characterized in that: The noise filtering module specifically includes: A signal component calculation unit, which is used to calculate the signal component s according to the initial signal through formula (2) q (t), where when calculating the first signal component s1(t), the initial signal is the original trajectory data s(t): Where b q (t) is s q (t) time-varying amplitude at time step t, d q (t) is s q (t) The phase at time step t is calculated by equation (3), f q (t) They represent the signal decomposition process s q (t) Corresponding instantaneous frequency, reference frequency, and estimated frequency; The difference calculation unit is used to calculate the difference between the initial signal and the current signal component s q The difference after the subtraction operation of (t) is used as a new initial signal and then input into the signal component calculation unit until the first k and including the kth signal components are obtained.

8. The unmanned surface ship cluster trajectory prediction system in an uncertain environment as claimed in claim 6, characterized in that: The motion pattern extraction module specifically includes: A nonlinear multilayer perceptron that receives the denoised trajectory data X t , generating the embedding H of the ship trajectory at time step t t =MLP(X t ,Θ N ), and then obtain the embedding H of the ship trajectory at all time steps, where They represent the horizontal and vertical coordinates of the nth ship cluster at time t, N represents the total number of ship clusters, Θ N is the learnable parameter of the multilayer perceptron; A dimension expansion unit, which is used to expand the dimension of the embedding H of the nonlinear multilayer perceptron to obtain a first branch p and a second branch o for parallel processing; A time dependency capturing unit, which is used to obtain a high-dimensional representation p' of the ship motion by capturing the time dependency according to the first branch p of the dimensional expansion unit; A motion feature acquisition unit, which is used to obtain the high-dimensional representation p' of the ship's motion according to the time-dependent capture unit, obtain the ship's motion feature P through the selective state space model SelectiveSSM described by Mamba's formula (6), and generate a gating signal o' according to the second branch o using the SiLU activation function, Sigmoid or tanh activation function described by formula (7); o'=SiLU(o) (7) Where Spftplus represents the activation function, Parameter represents a constant, Broadcast represents the broadcast operation, A represents the structured state matrix, the input matrix B that projects p' into the state space, the output matrix C of the state space and the discretized time step Δ, B = Linear B (p'), C = Linear C (p'), Δ=Softplus(Parameter+Broadcast(Linear Δ (p'))), and then use Δ to discretize A and B respectively to get The motion mode acquisition unit is used to multiply the motion feature P of the ship obtained by the motion feature acquisition unit with the gate signal o', and then perform dimensionality adjustment through linear transformation Linear to obtain the motion mode M of the ship, as shown in the following formula (8):

9. The unmanned surface ship cluster trajectory prediction system in an uncertain environment according to any one of claims 6 to 8, characterized in that: The interaction graph G is transformed into s The hidden features of each node and its neighbors are aggregated to obtain the dynamic spatial interaction of the ship, thereby obtaining the interaction features of the ship Where l represents the network layer number of the second graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the second graph convolutional network, The updated edge E of expression (10) smn The normalized adjacency matrix is They represent the hidden features of the lth, l+1th, and Lth network layers of the second graph convolutional network, respectively. represents the learnable parameters of the second graph convolutional network, σ represents the nonlinear activation function of the second graph convolutional network, and R represents different relationship graphs A r A collection of; In the formula, Softmax represents the activation function, d k represents the scaling factor, T represents the matrix transpose, Q sm Denotes the feature vector h m The query vector, Q sm =f sQ (h m ), K sn Denotes the feature vector h n The key vector, K sn =f sK (h n ), V sm Indicates v m The eigenvector h m A vector of values, V sm =f sV (h m ), f sQ 、f sK and f sV Both represent linear transformations, h m 、h n They are v m and v n The eigenvector of Update interaction graph G s Each node The information of is used to obtain the characteristic matrix in formula (9): Where l represents the network layer number of the first graph convolutional network, l∈{0,1,…,L}, L represents the total number of network layers of the first graph convolutional network, Represents edge E s The normalized adjacency matrix is They represent the hidden features of the lth and l+1th network layers of the first graph convolutional network, respectively. represents the parameters of the lth network layer of the first graph convolutional network, Represents the interaction graph G s The initial feature vector of each node in , σ represents the nonlinear activation function of the first graph convolutional network, and R represents different relationship graphs A r A collection of .

10. The method for predicting the trajectory of unmanned surface ship clusters in an uncertain environment as claimed in claim 9, characterized in that: The fusion method of the ship trajectory prediction module includes: using the time convolution function to predict the future position of the nth ship at time step t Make predictions and calculate predicted ship speeds; In the formula, concat represents the concatenation operation, and TCN represents the temporal convolutional network; The fusion method is trained by minimizing the negative log-likelihood loss: Where W represents the trainable parameter, L n represents the loss function, T obs represents the length of the ship's observation track, T pred represents the length of the predicted trajectory, and P represents the probability density function.

Images