Marine target route prediction method

Abstract

The invention discloses a marine target route prediction method which comprises the following steps: (1) collecting and preprocessing historical AIS data, calculating and introducing new features, and considering the influence of a ship in a complex scene; dividing a training set, a verification set and a test set; (2) extracting local space-time correlation characteristics by using a convolutional neural network; (3) modeling the time sequence dependence of the target by using a bidirectional long-short-term memory network; (4) weighting and extracting key moments or features by using an attention mechanism; (5) independently training each model, and evaluating the performance of the fused model on the test set; (6) extracting and outputting fixed-length data by using a moving window technology, and completing prediction from time periods to time periods; according to the invention, the future course of the target ship can be accurately predicted, powerful decision support is provided for ship drivers, and the safety of marine navigation is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of route prediction, and in particular to a method for predicting a target route at sea. Background Art

[0002] Accurate ship trajectory prediction is of great significance to maritime traffic management. It can effectively prevent ship collisions, optimize navigation routes and improve navigation efficiency. Compared with the relative fixed routes of vehicles on land, the navigation of ships in waters is more flexible and changeable, which undoubtedly increases the difficulty of predicting navigation trajectories. In view of this, the International Maritime Organization actively advocates the use of a variety of navigation systems, such as the bridge navigation watch alarm system, the Automatic Identification System (AIS) and the electronic chart display and information system. With the continuous advancement of maritime traffic monitoring technology and communication equipment, most ships are currently equipped with AIS systems to record location and other related information. The system can exchange information with the coastal AIS base station, thereby realizing the identification, positioning and real-time monitoring of ships. my country has built a complete coastal AIS base station network and accumulated a large amount of port ship-related information, which covers static information of ships (such as Maritime Mobile Service Identity (MMSI), size and type), dynamic information (such as latitude and longitude, ground speed and heading of the ship) and navigation-related information (such as ship draft and destination). Through deep processing of this information, we can extract and analyze the behavioral characteristics of the ship, and then achieve accurate prediction of the ship's navigation trajectory. However, it is worth noting that the AIS system usually sends a message every 2 seconds to 3 minutes, which leads to the huge size of the AIS data set. Therefore, how to efficiently process these AIS data to achieve accurate prediction of ship routes has become a hot and difficult issue in current research. Although several ship route prediction technologies have been launched, they all face certain limitations. In response to challenges such as missing ship trajectory data, the existing technology relies on the improved Kalman filter algorithm to construct a polynomial Kalman filter model. However, this type of physical modeling-based method is difficult to construct mathematical equations and is difficult to cope with complex and changing environmental factors. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a method for predicting the route of a target at sea, so as to solve the problem of trajectory prediction in complex scenarios.

[0004] Technical solution: A method for predicting a target route at sea according to the present invention comprises the following steps: (1) Collect and preprocess historical AIS data, calculate and introduce new features, consider the impact of the ship in complex scenarios; divide the data into training set, validation set and test set; (2) Use convolutional neural networks to extract local spatiotemporal correlation features; (3) Use bidirectional long short-term memory networks to model the temporal dependencies of targets; (4) Use the attention mechanism to weight the extraction of key moments or features; (5) Train each model independently and evaluate the performance of the fused model on the test set; (6) Use moving window technology to extract and output fixed-length data to complete predictions from time period to time period.

[0005] Furthermore, in step (1), the automatic identification system AIS historical data is collected, and the required features including time, longitude, latitude, speed over the ground, course over the ground and MMSI are extracted; Furthermore, in step (1), the AIS data is preprocessed, including abnormal data processing, default value processing and normalization; based on the dynamic information of the ship, the relative position, relative heading, relative speed and relative mass of the ship and the target ship are calculated and added as new features to obtain the distance between ships in complex scenarios, the overlap of navigation trajectories and the influence of ship types on ship trajectories.

[0006] Furthermore, step (2) is as follows: Use the convolutional neural network (CNN) to extract local spatiotemporal correlation features from the input flight trajectory: After the input data is normalized, it first enters the convolution layer to perform a convolution operation to extract data features; then enters the pooling layer to reduce the output dimension, where the pooling layer selects the average pooling layer; the pooling layer is connected to the Dropout layer; where the Dropout layer allows each neuron to have a probability Stop working, that is, each dimension of the output of the pooling layer is composed of probabilities becomes 0.

[0007] Furthermore, step (3) is specifically as follows: a bidirectional long short-term memory network Bi-LSTM is used to construct the temporal dependency of the target; wherein the Bi-LSTM neural network is composed of two layers of LSTM, and the output values ​​of the forward and backward LSTM units of the same input are concatenated to obtain the output vector of the Bi-LSTM network.

[0008] Furthermore, step (4) is as follows: using the similarity between the hidden state and the target state, the attention weight is calculated through the scoring function to measure the importance of each hidden state; Function, normalize the weights to probability distribution; use the normalized attention weights to weighted sum the hidden states; the flattening layer converts the input data into one-dimensional data, and finally outputs the final result through a fully connected layer; the scoring function calculates the similarity between the features of each time step in the input sequence and the target state of the current input, and the formula is as follows: ; in, is the first The hidden state at the time step, is the query input, Is a hidden state and Score the similarity between them.

[0009] Furthermore, step (5) is as follows: on the basis of constructing the convolutional neural network, the bidirectional long short-term memory network and the attention mechanism, the three models CNN, Bi-LSTM and the attention mechanism are combined in series; the loss function is defined, and the evaluation indicators including MSE, MAE, accuracy and recall are used for evaluation; among them, the loss function mean square error and root mean square error are used to evaluate the prediction performance of the model.

[0010] Furthermore, step (6) is as follows: using the moving window technology, extract a fixed-length data chain and input it into the trained model, and output the latitude, longitude, heading and speed of the next moment; then move the window by a unit length, use the result just output as part of the input data, remove the data of the oldest moment, and make a prediction for the next moment; until the trajectory result of the last moment of the target window is output, the output results are combined together and output as the trajectory prediction result of the target time period.

[0011] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, any one of the methods for predicting a maritime target route is implemented.

[0012] A storage medium described in the present invention stores a computer program, and when the computer program is executed by a processor, it implements any one of the methods for predicting a maritime target route.

[0013] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: by integrating convolutional neural networks, long short-term memory networks and attention mechanisms, the advantages of each model are fully utilized, and the accuracy of prediction is significantly enhanced. In addition, in actual complex navigation scenarios, in view of the mutual influence between the trajectories of different ships and the types of ships, the present invention innovatively introduces relative position, heading, speed and relative mass as new feature parameters. These features are not only in line with reality, but also greatly broaden the application prospects. In order to improve the speed and accuracy of the model and enhance its practicality, the present invention adheres to the concept of moving windows and realizes prediction from one time period to another, effectively compensating for the defect of reduced prediction accuracy of the model for long time periods, and the problem of slow prediction speed caused by the increase of feature parameters. The present invention can accurately predict the future route of the target ship and provide powerful decision-making support for ship drivers, thereby effectively improving the safety of maritime navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is the overall flow chart of the present invention; Figure 2 It is a basic structure diagram of the convolutional neural network of the present invention; Figure 3 It is the Bi-LSTM network structure diagram of the present invention; Figure 4 It is a schematic diagram of the attention mechanism principle of the present invention; Figure 5 It is a schematic diagram of the principle of the moving window technology of the present invention. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0016] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting a maritime target route, comprising the following steps: (1) Collect and preprocess historical AIS data, calculate and introduce new features, and consider the impact of the ship in complex scenarios; divide the data into training sets, validation sets, and test sets; including the following steps: (11) Extract historical AIS data, retain the required feature fields in the data set, and resample; the feature fields include: time, longitude, latitude, ground speed, ground course, and MMSI features; (12) Processing of erroneous data, noisy data and missing values ​​through filtering and interpolation; (13) Calculate and introduce relative heading, relative speed, relative position and relative mass as new features. The formula is as follows: The relative heading formula is as follows: △C(t)=C(t)-C(t-1) Among them, △C(t) represents the heading at two adjacent moments; C(t) represents the heading at moment t; C(t-1) represents the heading at moment t-1; The relative speed formula is as follows: △V(t)=V(t)-V(t-1) Among them, △V(t) represents the change of ship speed between two adjacent moments; The relative position formula is as follows: △E(t)=E(t)-E(t-1); △N(t)=N(t)-N(t-1) △E(t) and △N(t) represent the changes of longitude and latitude between two adjacent moments respectively; (2) Use convolutional neural networks to extract local spatiotemporal correlation features; e.g. Figure 2 As shown, the following steps are included: (21) After normalization, the input data first enters the convolution layer for convolution operation to extract data features; (22) Then enter the pooling layer to reduce the output dimension; (23) The pooling layer is connected to the Dropout layer. The essence of the Dropout layer is to make each neuron Stop working, that is, each dimension of the output of the pooling layer is composed of probabilities becomes , which can reduce the complex co-adaptation relationship between neurons and improve the generalization ability of the model.

[0017] (3) Use bidirectional long short-term memory networks to model the temporal dependencies of targets; e.g. Figure 3 As shown, the following steps are included: (31) Build a Bi-LSTM model and use the output of step (2) as the input of the model; (32) The output values ​​of the forward and backward LSTM units of the same input are concatenated to obtain the output vector of the Bi-LSTM network; the specific process is as follows: Bidirectional long short-term memory network (Bi-LSTM): a recursive neural network that processes data from the forward and backward directions of the time series through two independent LSTM networks, and combines the outputs of the two for better feature representation and modeling; Forward LTSM / Backward LTSM: Forward LTSM is a standard LSTM network that processes the input sequence in chronological order; it receives input information from the beginning (earliest time step) to the end (latest time step) of the sequence; Backward LSTM is also an LSTM network, but it processes the input sequence in reverse chronological order; that is, it reads input information from the end (latest time step) of the sequence forward (earliest time step); the formula is as follows: Assume the input sequence is ,in is the length of the sequence; the forward LSTM starts from arrive Process the input sequentially, generating a sequence of hidden states , the backward LSTM starts from arrive Process the input sequentially, generating a sequence of hidden states ; The output vector concatenation method is as follows: ;

[0018] in, represents the output vector, represents the concatenation function, Represents the hidden state vector at time t generated by the forward and backward LSTM.

[0019] (4) Using the attention mechanism to extract key moments or features; e.g. Figure 4 As shown in the figure, the details are as follows: using the similarity between the hidden state and the target state, the attention weight is calculated through the scoring function to measure the importance of each hidden state; function, normalizes these weights into probability distribution; uses the normalized attention weights to perform weighted summation on the hidden states, and the result of the weighted summation is called the context vector; the flattening layer converts the multidimensional data into one-dimensional data, and finally outputs the final result through a fully connected layer; Among them, the scoring function: calculates the similarity between the features of each time step in the input sequence and the target state of the current input. The definition of the scoring function can be written as the following formula:

[0020] in, is the first The hidden state at the time step, is the query input; Flatten layer: In a neural network, the function of the Flatten layer is to flatten the input multi-dimensional tensor into a one-dimensional vector so that it can be passed to the fully connected layer (Dense layer) for further processing; Fully connected layer: The most common type of layer in a neural network. Its main feature is that each neuron is connected to all neurons in the previous layer. The function of the fully connected layer is to map the input data to the output space through linear transformation and nonlinear activation function.

[0021] (5) Train each model independently and evaluate the performance of the fused model on the test set; specifically, after constructing the convolutional neural network, bidirectional long short-term memory network and attention mechanism, the three models CNN, Bi-LSTM and attention mechanism are combined in series; define the loss function and use evaluation indicators including MSE, MAE, accuracy and recall for evaluation; among them, the loss function mean square error and root mean square error are used to evaluate the prediction performance of the model.

[0022] (6) Use the moving window technique to extract and output fixed-length data and complete the prediction from time period to time period. Figure 5 As shown, the following steps are included: (61) According to the determined moving window length, a fixed-length data link is selected as input; (62) Inputting the input data chain into the fused model in step (5); (63) Add the output result of the next moment to the end of the input data chain, remove the data of the earliest moment in the original input data chain, and input the data chain into the model obtained in step (5) again to obtain the prediction result of the next moment; (64) Repeat step (63) until the prediction results for all moments in the target time period are obtained, and combine the results into the ship trajectory prediction result output for the target time period.

[0023] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, any one of the methods for predicting a maritime target route is implemented.

[0024] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any one of the methods for predicting a maritime target route.

Claims

1. A method for predicting a maritime target route, characterized in that: The following steps are involved: (1) Collect and preprocess historical AIS data, calculate and introduce new features, consider the impact of the ship in complex scenarios; divide the data into training set, validation set and test set; (2) Use convolutional neural networks to extract local spatiotemporal correlation features; (3) Use bidirectional long short-term memory networks to model the temporal dependencies of targets; (4) Use the attention mechanism to weight the extraction of key moments or features; (5) Train each model independently and evaluate the performance of the fused model on the test set; (6) Use moving window technology to extract and output fixed-length data to complete predictions from time period to time period.

2. A method for predicting a maritime target route according to claim 1, characterized in that: In step (1), the automatic identification system AIS historical data is collected and the required features including time, longitude, latitude, ground speed, ground heading and MMSI are extracted.

3. A method for predicting a maritime target route according to claim 1, characterized in that: In step (1), the AIS data is preprocessed, including abnormal data processing, default value processing and normalization; based on the dynamic information of the ship, the relative position, relative heading, relative speed and relative mass of the ship and the target ship are calculated and added as new features to obtain the distance between ships in complex scenarios, the overlap of navigation trajectories and the influence of ship types on ship trajectories.

4. A method for predicting a maritime target route according to claim 1, characterized in that: Step (2) is as follows: Use the convolutional neural network (CNN) to extract local spatiotemporal correlation features from the input flight trajectory: After the input data is normalized, it first enters the convolution layer to perform a convolution operation to extract data features; Then enter the pooling layer, which is connected to the Dropout layer; the Dropout layer allows each neuron to Stop working, that is, each dimension of the output of the pooling layer is composed of probabilities becomes 0.

5. A method for predicting a maritime target route according to claim 1, characterized in that: Step (3) is as follows: Use a bidirectional long short-term memory network Bi-LSTM to construct the temporal dependency of the target; wherein the Bi-LSTM neural network is composed of two layers of LSTM, and the output values ​​of the forward and backward LSTM units of the same input are concatenated to obtain the output vector of the Bi-LSTM network.

6. A method for predicting a maritime target route according to claim 1, characterized in that: Step (4) is as follows: using the similarity between the hidden state and the target state, the attention weight is calculated through the scoring function to measure the importance of each hidden state; Function, normalize the weights to probability distribution; use the normalized attention weights to weighted sum the hidden states; the flattening layer converts the input data into one-dimensional data, and finally outputs the final result through a fully connected layer; the scoring function calculates the similarity between the features of each time step in the input sequence and the target state of the current input, and the formula is as follows: ； in, is the first The hidden state at the time step, is the query input, Is a hidden state and Score the similarity between them.

7. A method for predicting a maritime target route according to claim 1, characterized in that: Step (5) is as follows: on the basis of constructing the convolutional neural network, bidirectional long short-term memory network and attention mechanism, the three models CNN, Bi-LSTM and attention mechanism are combined in series; the loss function is defined, and the evaluation indicators including MSE, MAE, accuracy and recall are used for evaluation; among them, the loss function mean square error and root mean square error evaluate the prediction performance of the model.

8. A method for predicting a maritime target route according to claim 1, characterized in that: Step (6) is as follows: Use the moving window technique to extract a fixed-length data chain and input it into the trained model, outputting the latitude, longitude, heading and speed at the next moment; then move the window by a unit length, use the result just output as part of the input data, remove the data at the oldest moment, and make a prediction for the next moment; Until the trajectory result of the last moment of the target window is output, the output results are combined together and output as the trajectory prediction result of the target time period.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, a method for predicting a maritime target route is implemented according to any one of claims 1 to 8.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a method for predicting a maritime target route is implemented according to any one of claims 1-8.

Images