Ship motion trajectory prediction method based on deep learning model T-LSTM

Abstract

The application discloses a ship motion trajectory prediction method based on a deep learning model T-LSTM. First, according to AIS data features, a trajectory prediction data set is constructed. For single ship motion, a time-sensitive T-LSTM neural network method is designed. By adding a new variable time interval and combining a time attention mechanism to determine the influence of the time interval on each layer of LSTM units, direct use of AIS data with different time intervals is realized. For multi-ship simultaneous motion prediction, on the basis of single-ship motion analysis, an adversarial neural network model is constructed. The generator part adopts a double-layer structure of T-LSTM and LTSM to predict ship data. Through adversarial learning network parameters, simultaneous prediction of multi-ship motion is realized. The method has the characteristics of high accuracy and easy implementation, can meet the actual sailing application requirements of ships, and has wide popularization and application value.

Description

Technical Field

[0001] This invention belongs to the field of ship motion prediction, specifically involving a method for predicting ship motion trajectories based on the deep learning model T-LSTM. Background Technology

[0002] The establishment of ship motion models is an important task in trajectory data analysis. It can provide a reference for abnormal ship behavior and collision risks, and can also serve as a technical indicator for route planning.

[0003] As maritime activities become increasingly complex, maritime traffic management and infrastructure protection face severe challenges. The increasingly complex maritime navigation environment is a potential obstacle and problem for the stable development of the maritime shipping industry. Reliable ship motion prediction can provide reliable data support. For example, Liu et al. proposed a ship trajectory prediction method based on a spatiotemporal multi-graph convolutional network using a mobile edge computing paradigm. This method fuses the time to the nearest point and the size data of surrounding ships, and uses a self-attention temporal convolutional layer to reduce parameters and improve feature extraction, ultimately achieving ship trajectory prediction based on the fused data. Murray B et al. proposed an automatic ship trajectory prediction method based on clustering single-point neighborhood search and multi-trajectory extraction. This data-driven method uses historical AIS data within a time range of 5-30 minutes to predict ship trajectories, estimating the ship's movement route through multi-trajectory extraction. In addition, Kanazawa M et al. proposed a multi-output hybrid predictor (MHP) ship trajectory prediction method. This method uses on-board sensor data and fuses black-box error compensation and LSTM neural networks to form a multi-output hybrid prediction method to predict the ship's position 30 seconds later.

[0004] However, in real-world ship navigation scenarios, ship motion prediction requires high accuracy and real-time performance. Traditional models can only predict ship trajectory points based on time interval data, and these intervals must be equal. However, real-world ship trajectory data is often irregular. The aforementioned methods require time-series calibration of the data before improving the algorithm to achieve motion analysis and prediction for a single ship. They cannot directly analyze the raw data, resulting in calibration data errors and data processing delays. Furthermore, they are difficult to analyze and predict the motion of multiple ships simultaneously. Summary of the Invention

[0005] This invention addresses the shortcomings of existing methods, such as the need for time-series calibration of raw data, resulting in calibration data errors and data processing delays, and the inability to perform predictive analysis on multiple ships. It proposes a ship motion trajectory prediction method based on the deep learning model T-LSTM, enabling analysis and research from the perspectives of single-ship and multi-ship motion.

[0006] This invention is achieved using the following technical solution: a method for predicting ship motion trajectories based on the deep learning model T-LSTM, comprising the following steps:

[0007] Step A: Collect ship AIS data and construct a single trajectory prediction dataset;

[0008] Step B: Construct a deep learning model T-LSTM to predict the trajectory of a single ship;

[0009] A time-aware module is introduced on the basis of the LSTM model. The single trajectory prediction dataset is input into the T-LSTM model for learning, and the deep learning model T-LSTM is trained. It is then applied to predict the motion trajectory of a single ship in real-world ship application scenarios with varying durations.

[0010] Step C: Construct a Generative Adversarial Network (GAN). The GAN includes a generator network and a discriminator network. The discriminator network uses an LSTM model, and the generator network uses a T-LSTM model. The GAN is combined with the T-LSTM model described in Step B to achieve simultaneous prediction of the motion trajectories of multiple ships.

[0011] Furthermore, in step B, T-LSTM decomposes the process of a time step into a long-term memory and a short-term memory. The short-term memory encodes the state in the past short period of time, capturing short-term changes in ship navigation; the long-term memory encodes the state in the past long period of time, capturing the overall trend of ship navigation. The time-aware module processes the time interval variable data, and the time difference weight generated by the time-aware module is integrated into the input gate structure, which directly affects the short-term memory of T-LSTM. The update of the long-term memory depends on the short-term memory of each time step, and the time difference weight indirectly affects the long-term memory.

[0012] The mathematical expression for a T-LSTM cell is as follows:

[0013] f t =σ(W f x t +U f h t-1 +b f (Forgotten Gate)

[0014] i t =σ(W i x t +U i h t-1 +b i (Input Gate)

[0015] o t =σ(W o xt +U o h t-1 +b o (Output Gate)

[0016]

[0017]

[0018] Where, x t h represents the input of the current hidden layer. t-1 It is the hidden state of the previous time step, {W f U f ,b f},

[0019] {W i U i ,b i}, {W o U o ,b o} and {W c U c b c} are the network parameters for the forget gate, input gate, output gate, and candidate memory, respectively.

[0020] Furthermore, in step B, if the time difference between time step t and t-1 is large, it indicates that the ship has not recorded new information for a long period of time. Therefore, in the prediction output at the current time step, the short-term memory needs to be adjusted proportionally to the time interval between time steps t and t-1, as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Among them, {W d b d} represents the network parameters, from which the current state information is extracted from the input signal at the current time step. And extract the information from the previous time step. The current state information is weighted and fused with the time difference weight to obtain the weighted state information. Separate and and short-term memory obtained after reintegration Δt′ represents the weight of the time interval.

[0026] Furthermore, in step B, the formulas for updating the hidden state output at the current time step and the long-term memory are as follows:

[0027]

[0028] h t =o t ·tanh(C t (Currently hidden)

[0029] Among them, C t-1 These are memory cells from the previous time step.

[0030] Furthermore, in step B, the time-aware module mainly extracts features from the input time-step sequence, mapping it to a state that accurately reflects the influence of the time step on the T-LSTM. The principle behind this is as follows:

[0031] First, for the input time interval sequence ΔT = Δt1, Δt2, ..., Δt n Linear normalization is adopted:

[0032]

[0033] Then, a decay function is constructed to balance the impact of excessively long or short time intervals on the T-LSTM output, and its formula is as follows:

[0034]

[0035]

[0036] Where g(x) is the decay function, e is the natural base, and Δt is the time interval between two ship trajectory points;

[0037] Finally, multiple convolutional layers balance the time interval sequence ΔT′=Δt′1,Δt′2,...,Δt′ n Deep feature extraction is performed to obtain time difference weights that can accurately reflect the strength of the input signal.

[0038] Furthermore, step C specifically includes the following steps:

[0039] Step C1: Use a generative network to predict the single trajectory prediction dataset to construct a new input dataset, and shuffle and clean it together with the single trajectory prediction dataset to form a new multi-trajectory prediction dataset.

[0040] Step C2: Calculate the loss of the judgment network using the multi-trajectory prediction dataset, and update the parameters of the discriminator network and the generator network accordingly. Repeat steps C1 and C2 to optimize the generator network and the discriminator network in sync and train the adversarial neural network.

[0041] Step C3: Use the GAN (Generative Adversarial Network) in conjunction with step B to train a T-LSTM model to achieve simultaneous prediction of multiple ships.

[0042] Furthermore, in step C2, when training the adversarial neural network GAN, a pre-training strategy is used to pre-train the generator and discriminator separately, and then perform joint training.

[0043] Furthermore, in step C3, when predicting multiple ship trajectories simultaneously, the last time step in the time series of all ships is aligned along the time dimension, as follows:

[0044] First, assume a sequence of ship trajectory labels. Where X represents the longitude coordinate sequence and Y represents the latitude coordinate sequence, among the latest message information received by the base station from all ships, the ship whose message transmission time T is located at the median of all ships is selected as the benchmark. Therefore, the time difference between the latest message transmission time received by the base station from other ships and the time of the benchmark ship is Δt = |Tt| n |;

[0045] Let R i =[x i -x i-1 y i -y i-1 ], where R i It is a displacement vector, module |R i | represents the algebraic value of the displacement vector, R. i The sine and cosine values ​​of the angle between the x-axis and the x-axis (cos) <R i ,i>and sin <R i ,i> represents displacement R i The direction, v i Let the velocity of the ship be the latitude and longitude coordinates of the regularized trajectory points at the i-th time step. The formula for updating the latitude and longitude coordinates of the regularized trajectory points is as follows:

[0046]

[0047] Furthermore, step C3 specifically employs the following method to predict the motion trajectories of multiple ships:

[0048] First, for time-series data of multiple ships, T-LSTM is used to encode the motion features of the ships. The encoded motion features obtained from each time step of T-LSTM are then combined into a temporally ordered feature sequence H = {h1, h2, ..., h...}. n};

[0049] Then, the time series sequence H is input into a multidimensional LSTM, which decodes H to generate the ship's trajectory sequence.

[0050] Furthermore, step A specifically includes the following steps:

[0051] Step A1: Preprocess the AIS data with unequal step sizes by removing outliers and performing data interpolation, and then normalize the data to form a preprocessed dataset.

[0052] Step A2: Extract the temporal feature information of the preprocessed dataset;

[0053] Step A3: Combine the preprocessed dataset with temporal feature information to construct a single trajectory prediction dataset.

[0054] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0055] This solution proposes a novel time-aware T-LSTM neural network architecture for single-ship operational analysis. It adds a new variable, time interval, and uses a time attention mechanism to determine the impact of the time interval on each LSTM unit, enabling direct utilization of AIS data with different time intervals. By introducing time awareness and time attention mechanisms to focus on the relationships between data from different time intervals, it achieves direct analysis and prediction of raw data, solving the problem that traditional analysis methods require prior data calibration and cannot directly analyze raw data.

[0056] Furthermore, based on the analysis of a single ship's motion, this method combines adversarial neural networks to achieve simultaneous prediction of multiple ship motions. Multiple T-LSTMs are used as inputs, and the generator part adopts a two-layer structure of T-LSTM and LSTM to predict ship data. Simultaneous prediction of multiple ship motions is achieved through adversarial learning network parameters. Furthermore, based on the single-ship operation analysis method, the method realizes the prediction of simultaneous motion of multiple ships. This approach has the characteristics of high accuracy and ease of implementation, can meet the actual navigation application needs of ships, and has broad application value. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the ship motion prediction method according to an embodiment of the present invention;

[0058] Figure 2This is a schematic diagram of the T-LSTM principle framework described in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the time attention module structure according to an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the structure of the GAN module described in an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of the prediction results of the prediction method described in the embodiments of the present invention;

[0062] Figure 6 This is a schematic diagram illustrating the principle of updating the latitude and longitude coordinates of the trajectory points after regularization in an embodiment of the present invention. Detailed Implementation

[0063] To better understand the above-described objects, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention may be practiced in other ways than those described herein, and therefore, the present invention is not limited to the specific embodiments disclosed below.

[0064] This embodiment proposes a deep learning-based method for predicting ship motion, such as... Figure 1 As shown, it includes the following steps:

[0065] Step A: Collect ship AIS data and construct a single trajectory prediction dataset;

[0066] Step A1: Preprocess the AIS data with unequal step sizes;

[0067] The purpose of preprocessing is to filter out information in the Automatic Identification System (AIS) that has a significant impact on ship trajectory prediction and to remove abnormal data with obvious errors in the AIS. Each data entry contains 40 points. In this embodiment, the last 10 points of each ship are interpolated according to the motion pattern at fixed time intervals of 5 minutes. Finally, the data is normalized to form the preprocessed dataset.

[0068] Step A2: Extract the temporal feature information of the dataset from Step A1 by slicing;

[0069] Step A3: Combine the preprocessed dataset from Step A1 with the temporal feature information extracted in Step A2 to form a single trajectory prediction dataset, and divide the single trajectory prediction dataset into a training set and a test set in an 8:2 ratio.

[0070] Step B: Construct a deep learning model T-LSTM to predict the trajectory of a single ship.

[0071] Step B1: Based on the LSTM model, a time-aware module is introduced. The single trajectory prediction dataset is input into the T-LSTM model for learning to obtain the implicit ship motion patterns in the trajectory data and complete the training of the deep learning model T-LSTM.

[0072] Step B2: Use the T-LSTM deep learning model trained in Step B1 to predict the trajectory of a single ship in real-world ship application scenarios with varying durations.

[0073] Step C: Construct an adversarial neural network and combine it with a GAN-T-LSTM model to simultaneously predict the motion trajectories of multiple ships.

[0074] Step C1: Construct an adversarial neural network, which includes a generator network and a discriminator network. In this embodiment, LSTM will be used as the discriminator network, and the T-LSTM model from step B will be used as the generator network.

[0075] Generative networks are used to predict single-trajectory prediction datasets to construct new input datasets (generated samples), and these datasets are then shuffled and cleaned together with the single-trajectory prediction datasets (real samples) to form new multi-trajectory prediction datasets.

[0076] Step C2: Train the adversarial neural network. Calculate the loss of the judgment network using the multi-trajectory prediction dataset and update the parameters of the discriminator and generator networks accordingly. Repeat steps C1 and C2 to simultaneously optimize the generator and discriminator networks and train the adversarial neural network.

[0077] Step C3: Use the GAN (Generative Adversarial Network) in conjunction with step B to train a T-LSTM model to achieve simultaneous prediction of multiple ships.

[0078] To better understand the present invention, the prediction of the motion trajectories of single and multiple ships is explained in detail:

[0079] In step B, when making predictions for the single-ship motion model:

[0080] In the AIS system, each ship's equipment sends AIS messages at regular intervals. These messages contain basic ship information and navigation status information, forming a time series of ship trajectory data. The time records in the ship trajectory data are unevenly distributed, with time intervals ranging from several minutes to tens of minutes. The irregularity of the time intervals between adjacent data points directly affects the prediction of the ship's speed and position.

[0081] A standard LSTM unit consists of a forget gate, an input gate, an output gate, and a storage unit. This structure requires the assumption that the time intervals between sequence elements are uniformly distributed, making it unsuitable for handling irregular time series. For ship motion tasks, the time interval between two consecutive AIS messages is a crucial factor affecting prediction results. This embodiment proposes a novel LSTM structure called Time-aware LSTM (T-LSTM) neural network, introducing time interval data as a new input variable to enhance the ability to predict ship trajectories with unequal time steps on top of the LSTM architecture.

[0082] (1) As Figure 2 The diagram illustrates the framework of T-LSTM. T-LSTM decomposes the process of a time step into a long-term memory and a short-term memory. The short-term memory encodes the state over a short period of time in the past, capturing short-term changes in ship navigation; the long-term memory encodes the state over a long period of time in the past, capturing the overall trend of ship navigation. The key to T-LSTM lies in introducing a time-aware module to process time interval variable data. The time difference weights generated by the module are integrated into the input gate structure, directly affecting the short-term memory of T-LSTM. The update of the long-term memory depends on the short-term memory at each time step, and the time difference weights can indirectly affect the long-term memory.

[0083] like Figure 2 As shown, the mathematical expression for a T-LSTM unit is as follows:

[0084] f t =σ(W f x t +U f h t-1 +b f (Forgotten Gate)

[0085] i t =σ(W i x t +U i h t-1 +b i (Input Gate)

[0086] o t =σ(W o x t +U o h t-1 +b o (Output Gate)

[0087]

[0088]

[0089] Where, x t h represents the input of the current hidden layer. t-1 It is the hidden state of the previous time step, {W f U f ,b f},{W i U i ,b i}, {W c U c ,b c} and {W c U c b c} are the network parameters for the forget gate, input gate, output gate, and candidate memory, respectively.

[0090] If the time difference between time step t and t-1 is large, it indicates that the ship has not recorded new information for a long period. Therefore, in the prediction of the output at the current time step, short-term memory should not rely excessively on the current state, but rather needs to be adjusted proportionally to the time interval between time steps t and t-1 to avoid over-reliance on outdated information. Similarly, when the time interval between time steps t and t-1 is short, the input signal of the current time step should be amplified to avoid causing lag in the time series. The hidden information h from the previous time step... t-1 Obtained directly from the internet:

[0091]

[0092]

[0093]

[0094]

[0095] Among them, {W d b d} represents network parameters. Since the update of the input signal at the current time step depends on the integration of the short-term memory of the previous time step and the input at the current time step, it is undesirable to affect past memory information while adjusting the strength of the input signal at the current time step. Therefore, this embodiment extracts the current state information from the input signal at the current time step. And extract the information from the previous time step. The current state information is weighted and fused with the time difference weight to obtain the weighted state information. Separate and and short-term memory obtained after reintegration Δt′ represents the weight of the time interval.

[0096] The formulas for the hidden state output and long-term memory update at the current time step are as follows:

[0097]

[0098] h t =o t ·tanh(C t (Currently hidden)

[0099] Among them, C t-1 These are the memory cells from the previous time step. The time difference weighting helps maintain the effectiveness of short-term memory, thus T-LSTM can more accurately reflect the importance and trends of historical information.

[0100] (2) In ship trajectory sequences, the time intervals between adjacent time steps can be highly irregular. For example, the time interval between several AIS data receptions by a base station may be several seconds, minutes, or tens of minutes. When the time interval between adjacent time steps is too large, the output of the current time step should not rely excessively on the current state; therefore, the signal strength of the input signal at the current time step should be reduced. The time-aware module can effectively capture the hidden feature information in irregular time-series data, and its structure is as follows: Figure 3 As shown:

[0101] The main component of time awareness is to extract features from the input time step sequence and map it to a state that can accurately reflect the degree of influence of the time step on the T-LSTM unit.

[0102] First, for the input time interval sequence ΔT = Δt1, Δt2, ..., Δt n Linear normalization is adopted:

[0103]

[0104] Then, a decay function is constructed to balance the impact of excessively long or short time intervals on the T-LSTM unit output. Specifically, when the time interval is long, the long short-term memory information from the previous time step is reduced; when the time interval is short, the long short-term memory information from the previous time step is retained. The formula is as follows:

[0105]

[0106]

[0107] Here, g(x) is an attenuation function that reduces the influence of short-term memory over time, e is the natural base, and Δt is the time interval between two ship trajectory points.

[0108] Finally, multiple convolutional layers balance the time interval sequence ΔT′=Δt′1, Δt′2, ..., Δt′ n Deep feature extraction is performed to obtain time difference weights that can accurately reflect the strength of the input signal.

[0109] In addition, during model training, the predicted results are compared with the true values. Mean squared error (MSE) and Dynamic Time Series Warp Matching (DTW) are used to evaluate the accuracy of the predictions. MSE is the average of the sum of squares of the differences between the predicted and true values. The expression for MSE is:

[0110]

[0111] Where N represents the total prediction duration, t is the time series number, pret represents the predicted value of the trajectory, and realt represents the actual value of the trajectory. Dynamic time series distortion matching refers to the time series similarity between predicted and actual values:

[0112] A[1,1]=D[1,1]

[0113] A[i,1]=A[i-1,1]+D[i.1] for i=2,3,...M

[0114] A[1, i] = A[1, j-1] + D[1, j] for j = 2, 3, ... M

[0115] A[i,j]=D[i,j]+min(A[i-1,j],A[i,j-1],A[i-1,j-1])

[0116] For i = 2, 3, ..., M and j = 2, 3, ..., N

[0117] DTW_distance = A[M, N]

[0118] Where D[i,j] represents the distance between the i-th element of the predicted value sequence and the j-th element of the true value sequence. A[i,j] represents the minimum cumulative distance from the first element to the i-th element of the predicted value sequence, and from the first element to the j-th element of the true value sequence. The DTW distance is the value of the bottom right element of the cumulative matrix A, representing the minimum cumulative distance between the predicted value sequence and the true value sequence.

[0119] Step C: GAN-based analysis of multiple ship motions:

[0120] Multi-ship motion analysis refers to the analysis and prediction of information such as the motion trajectory, speed, and heading of multiple ships. This embodiment proposes a novel method that integrates Generative Adversarial Network (GAN) and a time-aware LSTM model. The overall network structure is as follows: Figure 4 As shown, a Generative Adversarial Network (GAN) consists of a generator network and a discriminator network: the generator network is responsible for generating simulated data; the discriminator network is responsible for determining whether the input data is real or generated. The generator network is continuously optimized to confuse the discriminator network, while the discriminator network is simultaneously optimized to accurately distinguish between generated and real data.

[0121] In step C2, traditional GANs perform all loss calculations at the discriminator's output during network training. The discriminator's output is typically a judgment of whether the input data is real or fake; therefore, a binary cross-entropy function is generally used to calculate the overall loss, defined as follows:

[0122]

[0123] Training GANs presents numerous challenges, including training instability, pattern collapse, vanishing gradients, and exploding gradients. This study employs a pre-training strategy, separately pre-training the generator and discriminator before joint training. This approach helps improve model stability and convergence speed, addressing common GAN training issues.

[0124] When training the model, the dataset needs to be processed because the existing dataset cannot be directly used to train the model. In this embodiment, the original dataset is interpolated with equally spaced data points through sliding window interpolation as the prediction points of the model. Each sliding window processing will generate a data sample for training the model.

[0125] In step C3, when predicting multiple ship trajectories simultaneously, the time series presented by the AIS message data of multiple ships have inconsistent beats, causing the predicted data points to lose their practical meaning. A common method proposed by researchers is to interpolate or shift trajectory points at different timestamps to align all trajectory points in time. In fact, it is only necessary to align the last time step in the time series of all ships in the time dimension. The advantage of doing so is to minimize errors and ensure the accuracy of the data. Specific implementation details are as follows:

[0126] First, assume a sequence of ship trajectory labels. Where X represents the longitude coordinate sequence and Y represents the latitude coordinate sequence, among the latest message information received by the base station from all ships, the ship whose message transmission time T is located at the median of all ships is selected as the benchmark. Therefore, the time difference between the latest message transmission time received by the base station from other ships and the time of the benchmark ship is Δt = |Tt| n |

[0127] Let R i =[x i -x i-1 y i -y i-1 ], where R i It is a displacement vector, module |R i | represents the algebraic value of the displacement vector. To reflect the magnitude and direction of the displacement vector, the displacement vector R... i The sine and cosine values ​​of the angle between the x-axis and the x-axis (cos) <R i ,i>and sin <R i ,i> represents displacement R i The direction, v i Let the velocity of the ship be the latitude and longitude coordinates of the regularized trajectory points at the i-th time step. The formula for updating the latitude and longitude coordinates of the regularized trajectory points is as follows:

[0128]

[0129]

[0130]

[0131] The principle of the update method is as follows Figure 6 As shown in this embodiment, in multi-ship missions, the interaction of ship motions leads to complex motion characteristics, and no simulation data is generated. Therefore, a two-layer multidimensional T-LSTM and LSTM network is used to generate ship trajectory information. The purpose is to simulate more complex time series characteristics.

[0132] First, for time-series data of multiple ships, T-LSTM is used to encode the motion features of the ships. The encoded motion features obtained from each time step of T-LSTM are then combined into a temporally ordered feature sequence H = {h1, h2, ..., h...}. n}

[0133] Then, the temporal feature H is input into a multidimensional LSTM, which decodes H to generate the ship's trajectory sequence.

[0134] T-LSTM and LSTM interact to abstract and extract features from the input data. Simultaneously, a weighted averaging operation reduces the error of a single-layer LSTM, improving the model's robustness and generalization ability. Ideally, the data generated by the generative network should be accurate and regular, while the discriminative network uses one-dimensional convolution to extract and discriminate the trajectory information features generated by the generative network.

[0135] To better understand the solution of this invention, the effectiveness of this solution is verified below through experimental simulation:

[0136] This embodiment uses the AIS dataset of ships in the Bohai Sea, totaling 5000 time series data points. Each time series contains 40 ship tracking points with irregular time intervals. In single-ship motion tasks, the first 30 tracking points of each time series are left unprocessed, and the last 10 data points are interpolated using a sliding window interpolation method to generate 10 data points with a fixed time interval of three minutes as the real data for the sample. In multi-ship motion tasks, GAN-T-LSTM performs motion analysis on ten time series simultaneously each time. The first 29 tracking points are left unprocessed, and the 30th tracking point is spatiotemporally calibrated according to the method of this embodiment. Then, the last 10 data points are interpolated using a sliding window interpolation method to generate data points with a time interval of three minutes as the real data for the current sample. During experimental testing, 70% of each dataset is used as the training set, 20% as the test set, and 10% as the validation set.

[0137] In this experiment, the GAN-T-LSTM model was used to simultaneously predict the trajectories of multiple ocean vessels, such as... Figure 5 As shown, the GAN-T-LSTM model's performance in predicting the simultaneous trajectories of multiple ships is demonstrated. The pentagrams represent the latest data points for all ships after spatiotemporal calibration, and the incrementing numbers on the predicted points represent the time offset of subsequent time steps based on the green star, in minutes. Experimental results show that our model performs well in multi-ship prediction tasks, indicating that the GAN-T-LSTM model can be successfully applied to multi-ship prediction tasks and possesses high versatility and reliability.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A ship motion trajectory prediction method based on a deep learning model T-LSTM, characterized in that, The method comprises the following steps: Step A, collecting ship AIS data and constructing a single trajectory prediction data set; Step B, constructing a deep learning model T-LSTM to predict a single ship motion trajectory; On the basis of the LSTM model, a time perception module is introduced, the single trajectory prediction data set is input into the T-LSTM model for learning, the deep learning model T-LSTM is trained, and the T-LSTM model is applied to the actual ship application scenario with different time lengths to predict a single ship motion trajectory; Step C, constructing an adversarial neural network GAN, the adversarial neural network GAN comprises a generation network and a judgment network, the judgment network adopts the LSTM, the generation network adopts the T-LSTM model, and the simultaneous prediction of multiple ship motion trajectories is realized by combining the T-LSTM model of step B and the adversarial neural network GAN. 2.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 1, wherein: In step B, the T-LSTM decomposes a time step process into a long-term memory and a short-term memory, the short-term memory encodes the state in the past short time and captures the short-term changes of ship navigation, the long-term memory encodes the state in the past long time and captures the overall change trend of ship navigation, the time perception module processes the time interval variable data, the time difference weight generated by the time perception module is integrated into the input gate structure, directly affecting the update of the short-term memory and the long-term memory of the T-LSTM, and the time difference weight indirectly affects the long-term memory. The mathematical expression of the T-LSTM unit is as follows: , is a forget gate; , for the input gate; , is an output gate; , for time memory; , is a hidden feature; wherein, denotes the input to the current hidden layer, is the hidden state of the previous time step, , and are network parameters for the forget gate, input gate, output gate, and candidate memory, respectively.​ 3.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 2, characterized in that: In step B, if the time difference between the time steps t and t-1 is large, it indicates that the ship has not recorded new information for a long time, therefore, the short-term memory needs to be adjusted in proportion to the time interval between the time steps t and t-1 in the prediction output of the current time step, and the specific adjustment is as follows: , is the current state information; , is a weighted current state; , is the information from the previous time step; , is short-term memory; wherein, is a network parameter, the current state information is extracted from the input signal of the current time step , and the information of the previous time step is separated , the current state information is weighted and fused with the time difference weight to obtain the weighted state information , the separated and are re-integrated to obtain the short-term memory , represents a time interval weight.

4. The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 3, characterized in that: In step B, the formula for updating the current time step hidden state output and the long-term memory is as follows: , is long-term memory; , is the current hidden state; wherein, is the memory cell of the previous time step. 5.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 1, wherein: In step B, the time perception module mainly extracts features from the input time step sequence, so as to map to a state that can correctly reflect the influence degree of the time step on the T-LSTM, and the principle is introduced as follows: First, for the input sequence of time intervals Take the linear normalization: , is a linear normalization; Then, a decay function is constructed to balance the influence of too long or too short time interval on the output of the T-LSTM, and the formula is as follows: , is a decay function; , is a balance time interval; wherein, is a decay function, e is the natural base, is the time interval between two ship trajectory points; Finally, multiple convolutional layers are applied to the balanced time interval sequence to extract deep features to obtain time difference weights that can accurately reflect the intensity of the input signal. 6.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 1, wherein: Step C specifically comprises the following steps: Step C1, predicting the single trajectory prediction data set by using the generation network to construct a new input data set, and shuffling and cleaning the single trajectory prediction data set and the new input data set to form a new multi-trajectory prediction data set; Step C2, calculating the loss of the judgment network by using the multi-trajectory prediction data set, and feeding back the parameters of the judgment network and the generation network to update the parameters, repeating steps C1 and C2 to optimize the generation network and the judgment network synchronously, and training the adversarial neural network; Step C3: realizing the simultaneous prediction of multiple ships by combining the adversarial neural network GAN with the T-LSTM model trained in step B.

7. The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 6, characterized in that: In step C2, when training the adversarial neural network GAN, a pre-training strategy is adopted to separately pre-train the generator and the discriminator, and then jointly train them. 8.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 6, wherein: In the step C3, the last time step in the time sequence of all ships is aligned in the time dimension during the simultaneous prediction of multi-ship trajectories, specifically as follows: First, assume a ship trajectory label sequence , where represents a longitude coordinate sequence, represents a latitude coordinate sequence, among all the latest message information of the ships received by the base station, the ship whose message sending time T is located in the median of all the ships is selected as the reference, and thus the time difference between the latest message sending time of the other ships and the time of the reference ship received by the base station is ; Let where is the displacement vector, the modulus is the algebraic value of the displacement vector, the displacement vector and the sine and cosine values of the angle between the displacement vector and the X-axis and denotes the direction of the displacement , is the speed of the ship at the i-th time step, the longitude and latitude coordinates of the regularized trajectory points are updated as follows: 。 9.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 6, wherein: In the step C3, the prediction of the multi-ship motion trajectory is specifically performed in the following manner: First, for the time series data of multiple ships, the motion characteristics of the ships are encoded using T-LSTM, and the encoded motion characteristics obtained from each time step of T-LSTM are combined into a time series feature sequence ; The time series are then input into a multi-dimensional LSTM, which decodes to generate a sequence of trajectories for the ship. 10.The ship motion trajectory prediction method based on the deep learning model T-LSTM according to claim 1, wherein: The step A specifically comprises the following steps: Step A1, pre-processing the unequal step AIS data, after the operations of abnormal data elimination and data interpolation, performing normalization processing to form a pre-processed data set; Step A2, extracting time feature information of the pre-processed data set; Step A3, combining the pre-processed data set and the time feature information to construct a single trajectory prediction data set.

Images