CN-122022038-A - Ship track prediction method based on behavior feature embedding

CN122022038ACN 122022038 ACN122022038 ACN 122022038ACN-122022038-A

Abstract

The invention provides a ship track prediction method based on behavior feature embedding, and belongs to the technical field of ship track prediction. The method comprises the steps of constructing a training sample set based on original AIS track data, constructing a ship track prediction model, training the ship track prediction model by using the training sample set, collecting AIS track data of a ship in a real scene, constructing input data, inputting the trained ship track prediction model, outputting displacement increment of the ship in a plurality of time steps in the future, accumulating and restoring the displacement increment to obtain absolute positions corresponding to the time step predictions, and connecting the absolute positions of the predictions according to a time sequence to form a prediction result of the future track of the ship. The prediction accuracy and stability under complex sailing behavior are improved.

Inventors

XIAO YUPING
YANG ZE
GUAN ZHENBO
Qiao Zejia
GE XIANG
WU KAI
LIANG DONG

Assignees

中国电子科技集团公司第五十四研究所

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (8)

1. The ship track prediction method based on behavior feature embedding is characterized by comprising the following steps of: Step 1, constructing a training sample set based on original AIS track data, wherein each sample data in the training sample set comprises an input sample and an output label, the input sample comprises input feature vectors of all time steps in a historical observation section, and the output label comprises displacement increment of all time steps in a prediction target section; Step 2, constructing a ship track prediction model, wherein the ship track prediction model comprises a double-layer time sequence coding module, a behavior self-coding embedding module and a behavior guiding prediction network based on behavior embedding driving, the double-layer time sequence coding module is composed of a first-stage time sequence coding network and a second-stage time sequence coding network and is used for carrying out time sequence feature modeling on input samples and outputting behavior characterization vectors of track fragments; inputting an input sample in the training sample set into a ship track prediction model to obtain a predicted displacement increment, and iteratively optimizing parameters of the ship track prediction model by taking an output label corresponding to the input sample as a displacement increment true value in a mode of minimizing an error loss function until the error loss function is converged to be within a threshold value to obtain a trained ship track prediction model; And 4, acquiring AIS track data of the ship in the real scene, wherein the AIS track data comprises time stamps, longitudes, latitudes, speeds and course information, constructing input data according to the AIS track data, inputting a trained ship track prediction model, outputting displacement increment of the ship in a plurality of time steps in the future, accumulating and restoring the displacement increment to obtain absolute positions predicted by corresponding time steps, and connecting the predicted absolute positions according to time sequences to form a predicted result of the future track of the ship.
2. The ship track prediction method based on behavior feature embedding of claim 1, wherein the specific manner of step 1 is as follows: acquiring original AIS track data, wherein the original AIS track data comprises a time stamp, longitude, latitude, speed and course information; Sorting original AIS track data according to time stamps, performing de-duplication treatment on track points with repeated time stamps, removing abnormal points based on a set speed threshold and a set position mutation threshold, performing time alignment on track data at different time intervals, and adopting smoothing treatment to reduce noise influence; Dividing the track into track segments with the same length by adopting a sliding window with a fixed time length, and allowing overlapping between adjacent track segments; For each track segment, dividing the track segment into a history observation segment and a prediction target segment according to a time sequence, wherein the history observation segment comprises a front part Longitude, latitude, speed and heading data of each time step, and the predicted target segment contains the post Longitude, latitude, speed, and heading data for each time step; the input feature vector of the construction history observation section at the time step t is as follows: Wherein, the 、 Representing the longitude and latitude respectively, The speed of the voyage is indicated, The input feature vectors of each time step of the history observation section are arranged in time sequence to form a feature matrix as an input sample : Defining longitude and latitude differences between adjacent time steps in the prediction target segment as displacement increments, and then the displacement increment of the time step t is as follows: Arranging the displacement increment of each time step in the predicted target segment in time sequence to form a matrix as an output label : And forming a training sample set by all the input samples and the corresponding output labels.
3. The ship track prediction method based on behavior feature embedding of claim 1, wherein in the double-layer time sequence coding module, a first-stage time sequence coding network adopts a single-layer long-short-period memory network LSTM to model the speed, the course and the change information of a track fragment for extracting a preliminary time sequence coding vector for describing the overall motion trend of the track, and a second-stage time sequence coding network adopts a two-way long-short-period memory network BiLSTM for further capturing behavior change features on a local time scale on the basis of the preliminary time sequence coding vector to obtain a secondary time sequence coding vector; the behavior characterization vector is formed by splicing the primary time sequence coding vector and the secondary time sequence coding vector in the same feature space along feature dimensions.
4. The ship track prediction method based on behavior feature embedding of claim 1, wherein the behavior self-encoding embedding module takes a behavior characterization vector output by the double-layer time sequence encoding module as input, performs nonlinear feature transformation and dimension compression on the behavior characterization vector through at least one layer of feedforward fully-connected neural network, and outputs a low-dimensional behavior embedding vector.
5. The ship track prediction method based on behavior feature embedding according to claim 1, wherein the behavior guiding prediction network based on behavior embedding driving comprises a two-way long-short-term memory network and a regression output layer, and the working mode is as follows: Splicing the behavior embedded vector with the input characteristic vector of each time step of the history observation section in the characteristic dimension to form a behavior enhancement input sequence; and inputting the behavior enhancement input sequence into the two-way long-short-term memory network, extracting time-dependent characteristics related to the future position change of the ship, and outputting displacement increment of the ship in a plurality of time steps in the future through a regression output layer.
6. The ship track prediction method based on behavior feature embedding according to claim 1, wherein the error loss function in the step 3 is a regression loss function based on displacement increment, and the calculation mode is that the longitude displacement increment and the latitude displacement increment of a plurality of time steps obtained through prediction are calculated with the actual displacement increment of the corresponding time step by time step, and the errors of the time steps are weighted and summed to obtain the overall prediction error.
7. The ship track prediction method based on behavior feature embedding of claim 1, wherein the AIS track data of the ship in the real scene in step 4 is real-time or quasi-real-time acquired AIS track data, and the manner of constructing the historical observation section data according to the AIS track data is as follows: Sorting AIS track data according to time stamp, de-duplicating track points with repeated time stamp, eliminating abnormal points based on set speed threshold and position abrupt change threshold, time aligning track data with different time intervals, and smoothing to reduce noise effect; Dividing the track into track segments with the same length by adopting a sliding window with fixed time length, and allowing overlapping between adjacent track segments, wherein each track segment comprises longitude, latitude, speed and course data of corresponding time steps; and constructing input feature vectors of all time steps according to the longitude, latitude, speed and heading data, and arranging the input feature vectors of all time steps in time sequence to form a feature matrix, namely the input data.
8. The ship track prediction method based on behavior feature embedding of claim 1, wherein in step 4, the specific way of forming the prediction result of the future track of the ship is as follows: Performing time sequence accumulation on longitude displacement increment and latitude displacement increment of a plurality of time steps in the future output by the ship track prediction model to obtain a corresponding predicted absolute position sequence; taking the last real position of the ship in the real scene as an initial position, gradually accumulating the displacement increment of each time step, and recovering to form a predicted position sequence of the ship in a plurality of time steps in the future; And carrying out smoothing or interpolation on the predicted position sequence to weaken high-frequency fluctuation or local abnormality generated in the displacement accumulation process, thereby obtaining a continuous future predicted track result of the ship.

Description

Ship track prediction method based on behavior feature embedding Technical Field The invention relates to the technical field of ship track prediction, in particular to a ship track prediction method based on behavior feature embedding. Background The automatic ship identification system (AIS) can record dynamic data such as position information, navigational speed, heading and the like of the ship in real time, and provides important data support for track analysis, offshore traffic management and navigational safety. In practical application, the future motion trail of the ship is accurately predicted based on AIS data, and the method has important significance for navigation safety early warning, port scheduling optimization and anti-collision detection. The existing track prediction method mainly comprises a physical modeling method and a data driving method. The physical modeling method generally constructs a mathematical model based on kinematic parameters such as ship speed, heading and the like to conduct track calculation, and generally assumes that the ship moves along a uniform straight line or turns simply, and the calculation process is relatively simple. However, in a practical complex and changeable sailing mode, it is difficult to accurately capture the nonlinear behavior of the ship, so that the prediction accuracy is limited. The data driving method carries out unified modeling on the historical track data through the deep learning network, predicts the change rule of the learning track, and can mine potential features in the data. However, due to the significant difference between different navigation modes, the unified modeling is easy to ignore the diversity of the navigation modes, and in practical application, the problems of insufficient robustness of abnormal behaviors, difficulty in labeling training samples, scarcity of data and the like are also faced, so that the reliability and the accuracy of prediction are affected. Therefore, the prior art still has obvious defects in the aspects of processing the complex sailing behavior of the ship, the multi-category track mode and the abnormal track prediction, and needs to be further improved and optimized. Disclosure of Invention The invention provides a ship track prediction method based on behavior feature embedding, which solves the problems of neglecting navigation mode difference and lacking robustness in the existing method. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: a ship track prediction method based on behavior feature embedding comprises the following steps: Step 1, constructing a training sample set based on original AIS track data, wherein each sample data in the training sample set comprises an input sample and an output label, the input sample comprises input feature vectors of all time steps in a historical observation section, and the output label comprises displacement increment of all time steps in a prediction target section; Step 2, constructing a ship track prediction model, wherein the ship track prediction model comprises a double-layer time sequence coding module, a behavior self-coding embedding module and a behavior guiding prediction network based on behavior embedding driving, the double-layer time sequence coding module is composed of a first-stage time sequence coding network and a second-stage time sequence coding network and is used for carrying out time sequence feature modeling on input samples and outputting behavior characterization vectors of track fragments; inputting an input sample in the training sample set into a ship track prediction model to obtain a predicted displacement increment, and iteratively optimizing parameters of the ship track prediction model by taking an output label corresponding to the input sample as a displacement increment true value in a mode of minimizing an error loss function until the error loss function is converged to be within a threshold value to obtain a trained ship track prediction model; And 4, acquiring AIS track data of the ship in the real scene, wherein the AIS track data comprises time stamps, longitudes, latitudes, speeds and course information, constructing input data according to the AIS track data, inputting a trained ship track prediction model, outputting displacement increment of the ship in a plurality of time steps in the future, accumulating and restoring the displacement increment to obtain absolute positions predicted by corresponding time steps, and connecting the predicted absolute positions according to time sequences to form a predicted result of the future track of the ship. Further, the specific mode of the step 1 is as follows: acquiring original AIS track data, wherein the original AIS track data comprises a time stamp, longitude, latitude, speed and course information; Sorting original AIS track data according to time stamps, performing de-duplication treatment on track points w