Ship trajectory prediction method, terminal device and storage medium

Abstract

The present invention discloses a ship trajectory prediction method, terminal device, and storage medium. The method comprises obtaining a first observation sequence of length g from a ground dataset, inputting the first observation sequence into a first machine learning model, and obtaining an intermediate prediction sequence of length l. A second observation sequence of length g is obtained from a satellite dataset, inputting the second observation sequence into a second machine learning model, and obtaining an intermediate prediction sequence of length l. A third observation sequence of length g representing the same trajectory segment is selected from a ground-satellite hybrid dataset, and the third observation sequence is fused with the intermediate prediction sequence to form new fused input data. The new fused input data is then used as input to a third machine learning model to obtain a prediction result. The present invention solves the problems of poor adaptability and low prediction accuracy associated with multi-source data in the prior art.

Description

Technical Field

[0001] The present invention relates to trajectory prediction technology, in particular to a ship trajectory prediction method, terminal equipment and storage medium. Background Art

[0002] The development of the maritime Internet of Things has promoted the intelligentization of maritime ship traffic services. Specifically, the integration of maritime data with the help of 6G networks promotes the intelligent perception of maritime situations[1], intelligent supervision of maritime traffic safety[2] and the development of intelligent unmanned surface vessels (USVs)[3]. The maritime Internet of Things system based on space-sea-ground integration obtains information such as the identification, location and behavior of maritime ships through the automatic identification system (AIS)[4]. However, the report "Safety and Shipping Review 2021"[5] released by Allianz Group stated that ship collisions are one of the important causes of maritime accidents. Therefore, in the complex maritime traffic environment, it is very important to improve the anti-collision capabilities of maritime ships in VTS and the safe deep-sea navigation of unmanned ships. Among them, the use of artificial intelligence technology to deeply mine ship AIS data to improve the ability to predict ship trajectories has always been the focus of attention in the industry and academia.

[0003] AIS data involves meaningful spatiotemporal maritime traffic information that supports various navigation operation decisions [6], and provides important data services for ship collision avoidance, automatic navigation, and ship trajectory prediction in maritime applications. However, there are differences in the spatiotemporal information records of ships between space satellite AIS data and ground AIS data. Satellites in space can break through distance limitations and collect global ship AIS data, but the periodic operation of satellites and the long distance from the water surface make the AIS data collected by satellites sparse and have high data loss rates [7]. Compared with satellite AIS data, ground AIS receivers are installed on the coastline and can quickly and periodically collect ship AIS data, making the ground data set data-intensive and with low loss rates [8]. However, the signal reception distance of ground AIS receivers is limited and cannot obtain deep-sea ship position information. Obviously, in the maritime Internet of Things system, the use of multi-source data sets to achieve reliable and robust marine ship trajectory prediction is not only beneficial to maritime route planning and collision warning, but also can improve the deep-sea ship trajectory prediction capability and promote the development of deep-sea navigation of intelligent unmanned ships.

[0004] Current trajectory prediction methods can be roughly divided into three categories: methods using kinematics, methods based on traditional models, and methods using deep learning. The first two methods have been widely used in practical tasks, such as the constant speed model [9], Gaussian mixture model

[10] , and Kalman filter model

[11] . However, with the rapid development of artificial intelligence technology, the use of deep learning methods for ship trajectory prediction has attracted the attention of researchers. The ship trajectory is essentially a series of time-stamped spatiotemporal points collected from AIS data. MLP

[12] and recurrent neural networks (RNN)

[13] have been successfully applied in time series data prediction. RNN variants have become the main methods for ship trajectory prediction, such as long short-term memory (LSTM)

[14] and gated recurrent unit (GRU)

[15] , and bidirectional recurrent neural networks

[16] can achieve robust prediction results. In the 6G-enabled maritime Internet of Things, more accurate ship trajectory prediction is crucial for collision avoidance and safe navigation of maritime traffic. In existing work, existing prediction methods are difficult to adapt to multi-source data sets, the model has weak adaptability, and there is still a lot of room for improvement in prediction accuracy.

[0005] Current methods for ship trajectory prediction primarily target a specific dataset. This not only relies heavily on the characteristics of the dataset, but also suffers from low accuracy when applied to datasets from diverse sources. Given the rapid development of the maritime Internet of Things (IoT), ship estimation points are diverse across sea, land, and air, making existing prediction methods difficult to apply to multiple datasets simultaneously.

[0006] The ship trajectory prediction method based on kinematics is simple and efficient. Compared with other methods, the kinematic method can generate prediction results at a faster speed. Constant velocity model [9], Gaussian mixture model (GMM)

[10] , Ornstein Uhlenbeck (OU) model

[17] and Kalman filter variant

[18] . However, this type of method depends on the application scenario with limited conditions, and the prediction accuracy is still insufficient. Now some methods based on building motion models can improve the accuracy of ship trajectory prediction. They can build appropriate motion models according to specific scenarios, thereby improving the prediction accuracy. Such as support vector machine

[19] , clustering model

[20] , similarity search model

[21] and knowledge-based model

[22] . However, this type of method is highly dependent on manually designed objective functions and data rules, and cannot deeply mine the pattern features in the historical trajectory data of ships. Therefore, it is not suitable for application scenarios with complex conditions. Existing methods based on deep neural networks have powerful feature expression capabilities, are not constrained by specific scenarios, and have strong advantages in processing and learning massive data. For example, based on recurrent neural networks

[13] and their naive LSTM

[23] and GRU

[24] , graph neural networks

[25] , deep learning methods are used to learn large amounts of data and use neural networks to learn more movement rules. However, in existing work, existing prediction methods are difficult to adapt to sea-land-air multi-source datasets, the model's adaptability is weak, and there is still much room for improvement in prediction accuracy. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a ship trajectory prediction method, terminal device and storage medium to address the shortcomings of the existing technology, so as to solve the problems of poor adaptability and low prediction accuracy caused by multi-source data in the existing technology.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is: a ship trajectory prediction method, comprising the following steps:

[0009] S1. Obtain the first observation sequence of length g from the ground dataset, input the first observation sequence into the first machine learning model, and obtain the intermediate prediction sequence of length l

[0010] A second observation sequence of length g is observed from the satellite dataset, and the second observation sequence is input into the second machine learning model to obtain an intermediate prediction sequence of length l

[0011] S2, select the third observation sequence of length g of the same trajectory segment from the ground-satellite mixed data set, and compare the third observation sequence with the intermediate prediction sequence Intermediate prediction sequence Mix to form a new fusion input data of length g+2l The ground-satellite hybrid dataset is a mixture of a ground dataset and a satellite dataset.

[0012] S3, the new fusion input data As input to the third machine learning model, the prediction results are obtained.

[0013] This method extracts trajectory data features from ground and satellite datasets using ground and satellite sub-blocks, respectively. The fusion block then simultaneously and effectively learns these sub-blocks, along with the mixed dataset. This allows the overall solution to adapt to prediction tasks for datasets from diverse sources. Furthermore, the method leverages adjacency distance information to constrain the distance characteristics between time-series points in trajectory segments, enhancing the model's ability to learn and perceive features from time-series data from diverse sources, further improving prediction accuracy.

[0014] The first machine learning model is a multi-layer perceptron; the second machine learning model and the third machine learning model are both BiGRU neural networks.

[0015] In the first machine learning model, due to the linear and dense nature of the ground dataset, directly using a multilayer perceptron network yielded better performance for linear datasets. In the second and third machine learning models, although the satellite datasets are time-series data, the datasets have fewer samples and the distances between the adjacent trajectory points vary significantly. Using a BiGRU network effectively extracts stronger data feature relationships from the dataset samples as a whole, thereby enhancing the effectiveness of the overall solution. Extensive experimental results have confirmed that the use of a multilayer perceptron in the first machine learning model and a BiGRU neural network in the second and third machine learning models are the optimal choices.

[0016] In step S1, the intermediate prediction sequence of length l The specific acquisition process includes:

[0017] The observation sequence in the ground dataset The elements of are sequentially input into the first machine learning model, where t represents the position in the current trajectory sequence;

[0018] Update the output h of the jth hidden layer of the first machine learning model according to the following formula j :h j =σ j (h j-1 w j +b j ;θ), where the hidden state h1 of the first hidden layer is σ is the activation function of the first hidden layer, n is the total number of inputs, bt is the bias of the first hidden layer, w t is the weight of the connection layer, σ j is the nonlinear activation function of the jth hidden layer with learnable parameters θ, b j is the bias of the jth hidden layer, w tj is the weight of the j-th connection layer;

[0019] The vector composed of the output of all hidden layers is input into the fully connected layer to obtain an intermediate prediction sequence of length l

[0020] In step S1, the intermediate prediction sequence of length l The specific acquisition process includes: The elements of are sequentially input into the first machine learning model, where t represents the position in the current trajectory sequence;

[0021] Obtain the forward hidden sequence of the second machine learning model according to the following formula and the backward hidden sequence

[0022] in is the output set of the fully connected layer that represents the hidden state connected to the satellite sub-block model. The hidden layer state dimension size is q, which is the same as the dimension size of the model input data. The trajectory points of the sequence input at the current moment, i.e., the input sample points of the second machine learning model;

[0023] Represents the parameter set in the forward layer mapping process of the second machine learning model:

[0024] The i-th time point in the trajectory sequence of length g in the layer, Represents the parameter set in the backward layer mapping process of the second machine learning model: represents the i-th time point in the trajectory sequence of length g input to the backward layer, L represents the quantization error, N represents the total number of training samples, represents the i-th time point in the mapping sample of length h, and θ represents the parameter value in each input-to-output mapping process;

[0025] M g,h Indicates that given an input sequence X of length g g Next, the output sequence Y of predicted length h h , thus maximizing the conditional probability: M g,l =argmaxY p(Y h |X g ), p(Y h |X g ) indicates that given g observation sequences X g Mapped to the future h predicted trajectory sequence Y h probability;

[0026] Splice the forward hidden sequence and the backward hidden sequence to obtain a spliced ​​state sequence, input the spliced ​​state sequence into the fully connected layer to obtain an intermediate prediction sequence of length l

[0027] In step S3, the prediction result of the jth trajectory sequence point Expressed as: Among them, u j =BiGRU(μ t ,u j-1 ;θ), θ represents the parameter value in the mapping process of the third machine learning model, W y and b y Is to output the third machine learning model Map to the next predicted position The trainable parameters of the third machine learning model, is the output set representing the hidden state connected to the third machine learning model. The hidden layer state dimension size is q, 2q represents the total data dimension size in the third machine learning model, d j is the Cartesian plane distance between the jth trajectory sequence point and the next trajectory sequence point j+1, that is, the adjacent distance, d j The calculation formula is as follows: lat j+1 Indicates the latitude of the j+1th trajectory sequence point, lon j+1 Indicates the longitude of the j+1th trajectory sequence point.

[0028] As an inventive concept, the present invention further provides a terminal device, comprising:

[0029] one or more processors;

[0030] A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the above method of the present invention.

[0031] A computer-readable storage medium stores a computer program, which implements the steps of the above method of the present invention when executed by a processor.

[0032] Compared with the existing technology, the beneficial effects of the present invention are: the adaptive fusion data prediction method based on neural network designed in the embodiment of the present invention can perform feature learning and clever fusion learning on multi-source data sets, greatly increasing the adaptability of the model to the data set and improving the accuracy of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the prediction method according to an embodiment of the present invention;

[0034] Figure 2 This is a diagram of the MLP network structure according to an embodiment of the present invention;

[0035] Figure 3 This is a BiGRU network structure diagram of an embodiment of the present invention;

[0036] FIG4 (a) and FIG4 (b) are d of the satellite sub-block at each predicted time sequence trajectory point according to the embodiment of the present invention. H Difference, Figure 4(a) is for the straight trajectory prediction task, and Figure 4(b) is for the curved trajectory prediction task;

[0037] Figure 5 This is an experimental data diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0039] In the embodiments of the present invention, the terms "first", "second" and other similar words are not intended to imply any order, quantity and importance, but are only used to distinguish different elements. In the embodiments of the present invention, the terms "one", "an" and other similar words are not intended to indicate that there is only one of the things described, but rather to indicate that the relevant description is only for one of the things described, and the things described may have one or more. In the embodiments of the present invention, the terms "comprise", "include" and other similar words are intended to indicate logical relationships, and cannot be regarded as indicating relationships in spatial structure. For example, "A includes B" is intended to indicate that B logically belongs to A, but does not mean that B is spatially located inside A. In addition, the meanings of the terms "comprise", "include" and other similar words should be regarded as open, not closed. For example, "A includes B" is intended to indicate that B belongs to A, but B does not necessarily constitute the whole of A, and A may also include other elements such as C, D, and E.

[0040] The machine learning model training process employed in this embodiment of the present invention is as follows: This embodiment utilizes existing historical trajectory data from the AIS dataset and divides it into a training set and a test set. Those skilled in the art will appreciate that, when applying this method in practice, the test set should be obtained from the time series dataset for ship trajectory prediction. Since this embodiment is merely intended to illustrate and validate this method, time series data from an existing dataset can be used as the test set.

[0041] We downloaded 68 ship data sets from the official website. 1,339,183 ship trajectory data sets collected by ground-based AIS receivers served as the ground AIS sample set, while 108,684 ship trajectory data sets collected by airborne satellite receivers served as the satellite AIS sample set. The combined dataset consisted of 1,447,867 data sets. Each dataset type was further divided into a training set and a test set in a 7:3 ratio.

[0042] In the training set, the trajectory segments corresponding to each ship in the ground dataset and the satellite dataset are preprocessed and input into the deep features of the ship time series data in the ground sub-block and the satellite sub-block respectively. The tanh function is used for prediction in the output layer, and the error between the predicted trajectory label and the true trajectory label is calculated using the root mean square error function. The weights and biases of each layer of the neural network are calculated through the back propagation algorithm. The neural networks in the ground sub-block and the satellite sub-block are continuously iterated and trained until the loss function converges. The optimal training ground sub-model and the optimal training satellite sub-model are obtained and saved, and the intermediate output of the training data of the training set is obtained at the same time.

[0043] The intermediate outputs of the optimal training ground sub-model and satellite sub-model are combined with the trajectories of the same ship in the mixed data set to form a new fusion training set. The new fusion training set is then input into the fusion prediction block. The fusion prediction block network is used to learn and extract deep features from the fusion training set. The tanh function is used for prediction at the output layer. The root mean square error function is used to calculate the error between the predicted trajectory label and the true trajectory label. The weights and biases of each layer of the neural network are calculated through the back propagation algorithm. The neural network in the fusion prediction block is continuously iterated and trained until the loss function converges. The model of the optimal training fusion prediction block is obtained and saved.

[0044] Example 1

[0045] The embodiment of the present invention utilizes the multi-layer perceptron network and BiGRU network obtained after training to achieve the adaptability of the prediction model to multi-source data and better prediction performance.

[0046] Figure 1The structure of the proposed adaptive data fusion ship trajectory prediction method is shown in Figure 1, where the satellite sub-block, ground sub-block, fused data block, and fused prediction block form an integrated prediction framework for trajectory prediction. The key features of the invented fusion prediction method are as follows:

[0047] First, the fusion prediction method of this embodiment of the present invention takes g observed existing trajectory points as input and, after learning and fusing through four key blocks, outputs a future prediction of h trajectory points. The values ​​of g and h can be adjusted based on actual business needs.

[0048] Secondly, the fusion prediction method of the embodiment of the present invention constructs prediction sub-blocks based on different data sets, and uses the prediction sub-blocks to perform feature learning and prediction on data sets from different sources. For example, the satellite sub-block uses the BiGRU network to learn and make intermediate predictions on the satellite data set, and the ground sub-block uses the MLP network to learn and make intermediate predictions on the ground data set. At this time, different blocks use the network structure that best suits the data set, so that the overall fusion prediction model achieves the best effect. The adaptive fusion trajectory prediction process based on MLP and BiGRU networks includes the following steps, such as Figure 1 As shown:

[0049] Step (1): First, the ground sub-block receives a sequence of length g observed from the ground dataset and inputs the observation sequence into the MLP network for learning, where the data types included in each sequence point in the ground dataset are: time, latitude, longitude and adjacent distance.

[0050] Step (2): Secondly, the satellite sub-block receives a sequence of length g observed from the satellite dataset and inputs the observed sequence into the BiGRU network for learning, where the data types included in each sequence point in the satellite dataset are: time, latitude, longitude and adjacent distance.

[0051] Step (3): Then, the ground sub-block is trained and learned through the MLP network structure to output an intermediate prediction sequence of length l

[0052] Step (4): Similarly, the satellite sub-block is trained and learned through the BiGRU network structure to output an intermediate prediction sequence of length l

[0053] Step (5): Then, the observation sequence of length g of the same trajectory segment is selected from the ground-satellite mixed data set, and then fused with the sub-prediction sequence of length l of the ground sub-block and satellite sub-block to form a new fused input data of length g+2l This fused data serves as input to the fusion block. A ground-satellite hybrid dataset is a blend of a ground dataset and a satellite dataset. In this embodiment of the present invention, the blending process uses the same Maritime Mobile Service Identity (MMSI) to correlate the trajectories of the same vessel in different datasets, blending them together in chronological order.

[0054] For example, the process of mixing the trajectory data of a ship with MMSI = 218826000 is as follows: first, find the corresponding trajectory data segments in the ground and satellite datasets according to MMSI = 218826000, and then mix the two trajectory data segments in an orderly manner according to the order of the time fields to form a mixed data segment with MMSI = 218826000.

[0055] Step (6): The new fused data Input into the fusion prediction block, the fusion prediction block uses the BiGRU neural network structure for learning, and finally outputs the overall prediction model result Y h .

[0056] The adaptive fusion data prediction method based on neural networks designed in the embodiment of the present invention can perform feature learning and clever fusion learning for multi-source data sets, greatly increasing the adaptability of the model to the data set and improving the accuracy of the prediction model.

[0057] The ground sub-block uses a multi-layer perceptron network MLP. Multi-layer perceptron (MLP) is a neural network with a supervised learning technique using the back-propagation method. Figure 2 As shown in the figure, the MLP network consists of three layers: input layer, hidden layer and output layer. The different layers of the MLP neural network are fully connected.

[0058] The embodiment of the present invention uses the observation sequence defined in the ground dataset Represents a general sequence of length g, and t represents the position in the current trajectory sequence. The MPL network reads each element x of the input sequence in turn. t And update the internal hidden state according to the following formula

[0059]

[0060] Where σ is the activation function of the first hidden layer, x represents the input layer, x t is the input variable, n is the total number of inputs, b t is the bias of the layer, w t is the weight of the connection layer. Then, the following hidden layer updates the internal hidden state by the following formula

[0061] h j =σ j (hj-1 w j +b j ;θ), (2)

[0062] where σ j is the nonlinear activation function of the jth hidden layer with learnable parameters θ, b j is the bias of the jth hidden layer, w tj is the weight of the j-th layer connection. Finally, an output layer is added to accept the hidden state h in formula (2) j As input, to make predictions in sequence.

[0063] In the ground sub-block, the MLP network structure training process is to input the sequence Mapped into an output sequence, ie, a hidden sequence By doing the following:

[0064]

[0065] Among them, MLP represents the operation process of formula (1) and (2), using Output vector to the hidden layer Perform another full connection layer operation and output a sequence of length l, where each element The encoding extracts spatiotemporal information from the input ground-truth observation sequence in the t-th component of the sequence.

[0066] The satellite sub-block uses the BiGRU network structure to learn and train in the positive and negative directions. The BiGRU network is composed of a GRU network, which contains two structures: a reset gate and an update gate. The reset gate is used to reduce the information that is considered irrelevant in the previous unit, and the update gate is used to determine how much information from the previous unit needs to be passed to the next unit. The embodiment of the present invention uses the observation sequence that defines the satellite data set. Represents a general sequence of length, and t represents the position in the current trajectory sequence. Input sequence Calculate the hidden vector sequence sequentially through the GRU network The specific GRU model is controlled by the following formulas (4)-(7)

[0067] r t =σ(U r x t +W r h t-1 +b r ), (4)

[0068] z t =σ(U z x t +W z h t-1 +bz ), (5)

[0069]

[0070]

[0071] Where σ represents the sigmoid activation function, tanh is the hyperbolic tangent function, r t ,z t represents the output of the reset gate and update gate, h t represents candidate outputs and actual outputs, Indicates element-wise multiplication, Us and Ws are weight matrices, and bs is the bias term. The BiGRU model has an additional set of reverse propagation GRU models compared to the unidirectional GRU model, which enables BiGRU to explore past and future information in the observation sequence, thereby providing more effective prediction results.

[0072] BiGRU takes the input sequence Mapped into two output sequences, namely the forward hidden sequence Backward Hidden Sequence And operate through the following formulas (8)-(10):

[0073]

[0074]

[0075]

[0076] Each GRU function is a recurrent network of formulas (4)-(7), which adjusts the parameters To learn long-term patterns in two time directions. t and β t Represents the weights of the hidden layer states of the forward and backward neural networks of BiGRU at time t, b t is the bias term of the time series. The forward latent state and reverse latent state calculated by the two unidirectional GRU networks through formulas (8)-(10) are concatenated into a compact bidirectional representation, and finally the output layer of the satellite sub-prediction block is obtained.

[0077] Finally, use For the output vector Do another full connection layer calculation and output a sequence of length l, where each element encodes the spatiotemporal information extracted from the ship trajectory of the input satellite block in the t-th component of the sequence.

[0078] The embodiment of the present invention fuses the output sequence of the ground-satellite sub-block with the observation time sequence in the ground-satellite hybrid data set. It is the observation time series data in the ground-satellite hybrid data set. The embodiment of the present invention fuses the hybrid data set with the output sequences of the two sub-blocks according to the ship's water mobile communication service identification code and time field, that is, Finally, the unique value characteristics of the water mobile communication service identification code are used to identify the Output to the ground sub-prediction component and the output of the satellite sub-prediction component Combine to form the input observation sequence of the final fusion prediction component The length of the observation time series is g+2l.

[0079] The fusion prediction block decomposes the joint probability into ordered conditions to compute the conditional probability of the prediction sequence given the fused observation sequence. These assume conditional independence between past and future sequences, given the prediction sequence of the joint ground-satellite subcomponent and the observation sequence of the hybrid dataset, i.e.

[0080]

[0081] Each conditional probability in formula (14) can be modeled by the F function in the BiGRU network and expressed in the following form

[0082]

[0083] Among them, the predicted ship trajectory Calculated by F, given the neighbor distance d, the observation sequence at time step j, and the hidden state u of the BiGRU in the fusion prediction component j and its fusion data The advantage of a fusion network using a bidirectional recursive BiGRU architecture is that it can be designed to model sequences of arbitrary length and use observations from the past and future to predict the target time series. In the experiments of this embodiment of the present invention, an output sequence of length h is considered. The fusion subcomponent iterates the j = 1, ..., h bidirectional network to predict the future target sequence. The calculation process is as follows:

[0084]

[0085]

[0086]

[0087] in, W y and b y Is to output BiGRU Map to the next predicted position The first operation is to integrate the time series state Adjacency distance d j Connected to the input vector μ t In, μ t The dimension of depends on the output dimension of the ground-satellite sub-prediction component, Represents the output value of the jth trajectory sequence point in the output sequence of length h.

[0088] The following introduces the prediction scenario and application model of the ADF (Adaptive Data Fusion Model) prediction method of an embodiment of the present invention, and then analyzes the effectiveness of the ADF model in this scenario.

[0089] Prediction Model

[0090] (1) It is assumed that the user knows the trajectory observation points of the ship at sea for a certain period of time, and the trajectory observation sequence can come from different AIS (Automatic Identification System) datasets, such as ground datasets and satellite datasets.

[0091] (2) The user can input the known observation sequence into the ADF method, and the ADF method will pass the observation sequence through the ground, satellite sub-blocks, fusion data blocks and fusion prediction blocks to obtain the final future predicted trajectory sequence.

[0092] (3) In the ADF prediction method of the embodiment of the present invention, the number of trajectory points in the input observation sequence is not allowed to be less than 1.

[0093] Effectiveness Analysis

[0094] This section takes the ship trajectory sequence in the west coast of the United States as an example to analyze the effectiveness of the invented ADF.

[0095] (1) Comparison with existing work. The embodiment of the present invention compares the ADF method with the naive LSTM, GRU network and the five most advanced existing works. At the same time, it is verified on the ground dataset, satellite dataset and ground-satellite mixed dataset. Under the test of the mixed dataset, the ADF method proposed in the embodiment of the present invention outperforms most existing methods. Among them, it outperforms the existing research by an average of 46.51% in the RMSE evaluation index. The work of the embodiment of the present invention outperforms the existing work by up to 99.7% in MAE performance evaluation, and outperforms the existing work by an average of 52.7%. In order to better verify the effect of the ADF method, the embodiment of the present invention verifies the ADF method on ground and satellite datasets respectively. On the ground dataset, the ADF method outperforms the existing research by an average of 67.08% in RMSE evaluation index and by an average of 30.12% in MAE evaluation index. Similarly, on the satellite dataset, the ADF method outperforms the existing research by an average of 23.1% in RMSE evaluation index and by an average of 53.48% in MAE evaluation index. Therefore, the ADF method proposed in the embodiment of the present invention not only has good prediction accuracy, but also performs well on different datasets.

[0096] (2) The comparison of the prediction results of the ADF as a whole and the key sub-blocks further proves the effectiveness of the ADF method. The embodiment of the present invention uses ground and satellite datasets to compare the performance differences between the key sub-blocks and the overall ADF method. Specifically, RMSE and MAE are used to evaluate the differences between the ground sub-blocks and the ADF method for the ground dataset. As shown in Table 1, the performance results of the ground sub-block are better than the fusion model, which is consistent with the prominent linear characteristics of the ground dataset. The reason is that the ADF method combines the characteristics of both the ground dataset and the satellite dataset. Since the ground dataset is denser than the mixed dataset and the adjacent distances between two adjacent trajectory points are relatively uniform, the ADF method fits the time series trajectory faster and better. At the same time, the embodiment of the present invention also compares the performance differences between the ADF method and the satellite sub-blocks in the satellite dataset. The ADF method is better than the satellite sub-block in both RMSE and MAE. The reason is that the ADF method not only combines the dense law of the ground dataset to make the adjacent distance characteristics smoother, but also expands the random step size for the mixed dataset and supports the prediction of time series trajectories of any length, thereby enhancing the adaptability of the ADF method. Table 1 compares the overall performance differences of the ground sub-block, satellite sub-block and ADF methods under ground and satellite datasets.

[0097]

[0098] (3) Prediction is performed on straight and curved trajectories to verify the effectiveness of the ADF method. In order to quantify the difference in prediction effect of the ADF method on different trajectory lines, the embodiment of the present invention calculates the haversine distance d between the predicted value and the true value for each predicted trajectory point. H In the ship trajectory prediction task of the embodiment of the present invention, the specific trajectory points are represented as geographic coordinate pairs. The embodiment of the present invention uses the following haversine formula to calculate the distance d between any two longitude and latitude position points p1 and p2: H ,

[0099]

[0100] in i=1,2, r is the radius of the earth, is the latitude of the position points p1 and p2, λ1, λ2 is the longitude of the position points p1 and p2, and the unit is meter (m). In the straight trajectory prediction task, as shown in Figure 4 (a) and Figure 4 (b), the ADF method is closer to the true trajectory than the prediction results of the satellite sub-block, regardless of whether it is a straight trajectory or a curved trajectory. In the curved trajectory prediction task, the two key blocks d H The maximum difference in the resulting values ​​is as high as 90.04%. It can also be seen that the training values ​​differ significantly from the true values. This is because the satellites' periodic movement in space causes them to intermittently receive AIS messages. This trajectory trend of large differences in adjacent distances d makes it prone to large prediction errors when using only satellite sub-blocks. However, the model of the embodiment of the present invention significantly improves prediction results by fusing multi-source datasets and utilizing smoothed adjacent distance d information constraints.

[0101] For a more detailed description, the embodiment of the present invention takes the trajectory segment (36.9955, -123.7687), (36.9947, -123.7695), (36.9947, -123.7695) as an example, and uses the first two trajectory points to predict the third trajectory point to illustrate the effectiveness of the invented ADF method. Figure 5 shown.

[0102] (1) Initialization. The ship's trajectory points are composed of the earth's longitude and latitude, i.e. (lat, lon). If the input x = (36.9955, -123.7687), (36.9947, -123.7695) is loaded into the ADF method as the initial state, the input length g = 2. And assume that the input x of the ground sub-block is terra = x, and the input x of the satellite sub-block sat =x.

[0103] (2) Key sub-block prediction. The ground sub-block receives input xterra , the trained MLP network generates the sub-prediction output y terra =(37.0024,-123.7715), the satellite sub-block receives input x sat , the trained BiGRU network generates the sub-prediction output y sat =(37.0822,-123.4826), the length of each is l=1.

[0104] (3) Fusion data set. In the fusion data block, the ground sub-block intermediate prediction output y terra , satellite sub-block intermediate prediction output y sat And the data x in the mixed data set is fused. That is, x fused =x+y terra +y sat , that is, x fused =(36.9955,-123.7687),(36.9947,-123.7695),(37.0024,-123.7715),(37.0822,-123.4826), at this time the input length of the fusion block is g+2l.

[0105] (4) Fusion prediction block. At this time, the input of the fusion prediction block is x fused The trained BiGRU network produces the final prediction y = (36.9963, -123.7559), with length h = 1. The prediction result is obtained by fusing the network structures of different sources. This prediction result not only has an advantage in prediction accuracy, but also enables the ADF method to accept trajectory prediction tasks based on any AIS data source.

[0106] Example 2

[0107] Embodiment 2 of the present invention provides a terminal device corresponding to the above-mentioned embodiment 1. The terminal device can be a processing device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the method of the above-mentioned embodiment.

[0108] The terminal device of this embodiment includes a memory, a processor, and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method in the above-mentioned embodiment 1.

[0109] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage.

[0110] In other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors, which are not limited herein.

[0111] Example 3

[0112] Embodiment 3 of the present invention provides a computer-readable storage medium corresponding to the above-mentioned embodiment 1, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the method of the above-mentioned embodiment 1 are implemented.

[0113] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0114] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0115] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0116] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1A step that specifies a function in one or more boxes.

[0117] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0118] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

[0119] References

[0120] [1]Y.Zhou, W.Daamen, T.Vellinga, and S.Hoogendoorn, "Review of maritimetraffic models from vessel behavior modeling perspective," Transp Res Part CEmerg Technol, vol.105, no.December 2018, pp.323–345, Aug.2019, doi:10.1016 / j.trc.2019.06.004.

[0121] [2] Navig, vol.74, no.6, pp.1252–1266, Nov.2021, doi:10.1017 / S0373463321000540.

[0122] [3]T.Wang,G.Li,L.I.Hatledal,R.Skulstad,V.AEsoy,and H.Zhang,“Incorporating Approximate Dynamics Into Data-Driven Calibrator:ARepresentative Model for Ship Maneuvering Prediction,”IEEE Trans IndInformatics,vol.18,no.3,pp.1781–1789,Mar.2022,doi:10.1109 / TII.2021.3088404.

[0123] [4]T.Qiu,Z.Zhao,T.Zhang,C.Chen,and C.L.P.Chen,“Underwater Internet ofThings in Smart Ocean:System Architecture and Open Issues,”IEEE Trans IndInformatics,vol.16,no.7,pp.4297–4307,Jul.2020,doi:10.1109 / TII.2019.2946618.

[0124] [5]X.Chen,R.Yan,S.Wu,Z.Liu,H.Mo,and S.Wang,“A fleet deployment modelto minimise the covering time of maritime rescue missions,”Marit PolicyManag,p.3088839,2022,doi:10.1080 / 03088839.2021.2017042.

[0125] [6]X.Chen et al.,“Ship Trajectory Reconstruction from AIS SensoryData via Data Quality Control and Prediction,”Math Probl Eng,vol.2020,pp.1–9,Aug.2020,doi:10.1155 / 2020 / 7191296.

[0126] [7]T.Wei,W.Feng,Y.Chen,C.-X.Wang,N.Ge,and J.Lu,“Hybrid Satellite-Terrestrial Communication Networks for the Maritime Internet of Things:KeyTechnologies,Opportunities,and Challenges,”IEEE Internet Things J,vol.8,no.11,pp.8910–8934,Jun.2021,doi:10.1109 / JIOT.2021.3056091.

[0127] [8]Y.Huang,Y.Li,Z.Zhang,and R.W.Liu,“GPU-Accelerated Compression andVisualization of Large-Scale Vessel Trajectories in Maritime IoT Industries,”IEEE Internet Things J,vol.7,no.11,pp.10794–10812,Nov.2020,doi:10.1109 / JIOT.2020.2989398.

[0128] [9]C.Scholler,V.Aravantinos,F.Lay,and A.Knoll,“What the ConstantVelocity Model Can Teach Us About Pedestrian Motion Prediction,”IEEE RobotAutomLett,vol.5,no.2,pp.1696–1703,Apr.2020,doi:10.1109 / LRA.2020.2969925.

[10] B.R.Dalsnes,S.Hexeberg,A.L.Flaten,B.-O.H.Eriksen,and E.F.Brekke,

[0129] “The Neighbor Course Distribution Method with Gaussian Mixture ModelsforAIS-Based Vessel Trajectory Prediction,”in 2018 21st InternationalConferenceon Information Fusion(FUSION),Jul.2018,pp.580–587,doi:

[0130] 10.23919 / ICIF.2018.8455607.

[0131]

[11] K.Jaskólski,“Automatic Identification System(AIS)Dynamic DataEstimationBased on Discrete Kalman Filter(KF)Algorithm,”Sci J Polish NavAcad,vol.

[0132] 211,no.4,pp.71–87,Dec.2017,doi:10.5604 / 01.3001.0010.6747.

[0133]

[12] L.P.Perera,P.Oliveira,and C.Guedes Soares,“Maritime TrafficMonitoringBased on Vessel Detection,Tracking,State Estimation,andTrajectoryPrediction,”IEEE Trans Intell Transp Syst,vol.13,no.3,pp.1188–1200,Sep.

[0134] 2012,doi:10.1109 / TITS.2012.2187282.

[0135]

[13] S.Capobianco,L.M.Millefiori,N.Forti,P.Braca,and P.Willett,“DeepLearning Methods for Vessel Trajectory Prediction Based on RecurrentNeuralNetworks,”IEEE Trans Aerosp Electron Syst,vol.57,no.6,pp.4329–4346,Dec.

[0136] 2021,doi:10.1109 / TAES.2021.3096873.

[0137]

[14] D.Gao,Y.Zhu,J.Zhang,Y.He,K.Yan,and B.Yan,“A novel MP-LSTMmethodfor ship trajectory prediction based on AIS data,”Ocean Eng,vol.228,

[0138] no.March,p.108956,May 2021,doi:10.1016 / j.oceaneng.2021.108956.

[0139]

[15] P.Han,W.Wang,Q.Shi,and J.Yue,“A combined online-learning modelwithK-means clustering and GRU neural networks for trajectory prediction,”AdHocNetworks,vol.117,no.February,p.102476,Jun.2021,doi:

[0140] 10.1016 / j.adhoc.2021.102476.

[0141]

[16] R.W.Liu,J.Nie,S.Garg,Z.Xiong,Y.Zhang,and M.S.Hossain,“Data-DrivenTrajectory Quality Improvement for Promoting Intelligent VesselTrafficServices in 6G-Enabled Maritime IoT Systems,”IEEE Internet Things J,vol.8,

[0142] no.7,pp.5374–5385,Apr.2021,doi:10.1109 / JIOT.2020.3028743.

[0143]

[17] M.Uney,L.M.Millefiori,and P.Braca,“Data Driven VesselTrajectoryForecasting Using Stochastic Generative Models,”in ICASSP 2019-2019IEEE International Conference on Acoustics,Speech and Signal Processing(ICASSP),May 2019,pp.8459–8463,doi:10.1109 / ICASSP.2019.8683444.

[0144]

[18] L.P.Perera,“Navigation vector based ship maneuvering prediction,”Ocean Eng,vol.138,no.November 2016,pp.151–160,Jul.2017,doi:10.1016 / j.oceaneng.2017.04.017.

[0145]

[19] L.Sang,X.Yan,A.Wall,J.Wang,and Z.Mao,“CPA Calculation Methodbased on AIS Position Prediction,”J Navig,vol.69,no.6,pp.1409–1426,Nov.2016,doi:10.1017 / S0373463316000229.

[0146]

[20] Z.Xiao,L.Zhang,X.Fu,W.Zhang,J.T.Zhou,and R.S.M.Goh,“ConcurrentProcessing Cluster Design to Empower Simultaneous Prediction for Hundreds ofVessels’Trajectories in Near Real-Time,”IEEE Trans Syst Man,Cybern Syst,vol.51,no.3,pp.1–14,2020,doi:10.1109 / TSMC.2019.2906381.

[0147]

[21] D.Alizadeh,A.A.Alesheikh,and M.Sharif,“Vessel TrajectoryPrediction Using Historical Automatic Identification System Data,”J Navig,vol.74,no.1,pp.156–174,Jan.2021,doi:10.1017 / S0373463320000442.

[0148]

[22] Z.Xiao et al.,“Big Data Driven Vessel Trajectory and NavigatingState Prediction With Adaptive Learning,Motion Modeling and ParticleFiltering Techniques,”IEEE Trans Intell Transp Syst,vol.23,no.4,pp.3696–3709,Apr.2022,doi:10.1109 / TITS.2020.3040268.

[0149]

[23] S.Hochreiter and J.Schmidhuber,“Long Short-Term Memory,”NeuralComput,vol.9,no.8,pp.1735–1780,Nov.1997,doi:10.1162 / neco.1997.9.8.1735.

[0150]

[24] J.Chung,C.Gulcehre,K.Cho,and Y.Bengio,“Empirical Evaluation ofGated Recurrent Neural Networks on Sequence Modeling,”arXiv,pp.1–9,Dec.2014,[Online].Available: http: / / arxiv.org / abs / 1412.3555 .

[0151]

[25] X.Song et al.,“Pedestrian Trajectory Prediction Based on DeepConvolutional LSTM Network,”IEEE Trans Intell Transp Syst,vol.22,no.6,pp.3285–3302,Jun.2021,doi:10.1109 / TITS.2020.2981118.

Claims

1. A ship trajectory prediction method, characterized in that: The following steps are involved: S1. Obtain the first observation sequence of length g from the ground dataset, input the first observation sequence into the first machine learning model, and obtain the intermediate prediction sequence of length l Intermediate prediction sequence The specific acquisition process includes: The observation sequence in the ground dataset The elements of are sequentially input into the first machine learning model, where t represents the position in the current trajectory sequence; Update the output h of the jth hidden layer of the first machine learning model according to the following formula j :h j =σ j (h j-1 w j +b j ;θ), where the hidden state h1 of the first hidden layer is σ is the activation function of the first hidden layer, n is the total number of inputs, b t is the bias of the first hidden layer, w t is the weight of the connection layer, σ j is the nonlinear activation function of the jth hidden layer with learnable parameters θ, b j is the bias of the jth hidden layer, w tj is the weight of the j-th connection layer; The vector composed of the output of all hidden layers is input into the fully connected layer to obtain an intermediate prediction sequence of length l Intermediate prediction sequence The specific acquisition process includes: The observation sequence in the satellite dataset The elements of are sequentially input into the first machine learning model, where t represents the position in the current trajectory sequence; Obtain the forward hidden sequence of the second machine learning model according to the following formula and the backward hidden sequence in is the output set of the fully connected layer that represents the hidden state connected to the satellite sub-block model. The hidden layer state dimension size is q, which is the same as the dimension size of the model input data. The trajectory points of the sequence input at the current moment, i.e., the input sample points of the second machine learning model; Represents the parameter set in the forward layer mapping process of the second machine learning model: represents the i-th time point in the trajectory sequence of length g input to the forward layer of the second machine learning model, Represents the parameter set in the backward layer mapping process of the second machine learning model: represents the i-th time point in the trajectory sequence of length g input to the backward layer, L represents the quantization error, N represents the total number of training samples, represents the i-th time point in the mapping sample of length h, and θ represents the parameter value in each input-to-output mapping process; Splice the forward hidden sequence and the backward hidden sequence to obtain a spliced ​​state sequence, input the spliced ​​state sequence into the fully connected layer to obtain an intermediate prediction sequence of length l A second observation sequence of length g is observed from the satellite dataset, and the second observation sequence is input into the second machine learning model to obtain an intermediate prediction sequence of length l S2, select the third observation sequence of length g of the same trajectory segment from the ground-satellite mixed data set, and compare the third observation sequence with the intermediate prediction sequence Intermediate prediction sequence Fusion is performed to form a new fusion input data of length g+2l The ground-satellite hybrid dataset is a mixture of a ground dataset and a satellite dataset; S3, the new fusion input data As input to the third machine learning model, the prediction results are obtained; The first machine learning model is a multi-layer perceptron; the second machine learning model and the third machine learning model are both BiGRU neural networks.

2. The ship trajectory prediction method according to claim 1, characterized in that: The acquisition process of the first machine learning model, the second machine learning model and the third machine includes: The collected ship time series trajectory data is used as the ground AIS sample set, and the collected ship time series trajectory data is used as the satellite AIS sample set. The ground AIS sample set and the satellite AIS sample set are mixed, and the mixed data set is divided into a training set and a test set; Using the training set as input to a ground sub-model and a satellite sub-model, training the ground sub-model and the satellite sub-model to obtain a first machine learning model and a second machine learning model; The intermediate outputs of the optimal ground sub-model and satellite sub-model obtained during the training process are combined with the trajectories of the same ship in the mixed data set into a new fusion training set, the new fusion training set is input into the fusion prediction block, the fusion prediction block is trained, and the third machine learning model is obtained.

3. The ship trajectory prediction method according to claim 1, characterized in that: In step S3, the prediction result of the jth trajectory sequence point Expressed as: Among them, u j =BiGRU(μ t ,u j-1 ;θ), θ represents the parameter value in the mapping process of the third machine learning model, W y and b y Is to output the third machine learning model Map to the next predicted position The trainable parameters of the third machine learning model, is the output set representing the hidden state connected to the third machine learning model. The hidden layer state dimension size is q, 2q represents the total data dimension size in the third machine learning model, d j is the Cartesian plane distance between the jth trajectory sequence point and the next trajectory sequence point j+1, that is, the adjacent distance, d j The calculation formula is as follows: lat j+1 Indicates the latitude of the j+1th trajectory sequence point, lon j+1 Indicates the longitude of the j+1th trajectory sequence point.

4. A terminal device, characterized in that: include: one or more processors; A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the method according to any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that The device stores a computer program, which implements the steps of the method according to any one of claims 1 to 3 when executed by a processor.

Images