CN-115238375-B - Deep learning method for multi-step prediction of navigation states of two ships

CN115238375BCN 115238375 BCN115238375 BCN 115238375BCN-115238375-B

Abstract

The invention discloses a deep learning method for multi-step prediction of two ship navigation states, which comprises the steps of rapidly extracting, screening and interpolating effective ship data from massive data according to AIS data characteristics, acquiring a ship sequence data pair with the largest potential risk in the same space-time range according to a ship potential risk formula, constructing a ship navigation state prediction model, effectively connecting an encoder and a decoder through an interaction module, realizing intercommunication among ship historical data information, and carrying out self-feedback adjustment according to calculated loss values when the model is trained so as to help the model to quickly converge. The invention realizes multi-step prediction of two ship navigation states by combining the ship potential risk field and the long-short-period memory network model, overcomes the limitation that the prediction of a single ship model ignores the influence of surrounding ships, and provides a theoretical basis for the selection of multi-ship model data.

Inventors

LIU TAO
XU XIANG
Lei Zhengling
HUO YUCHI
MENG WEI
GAO JIN

Assignees

上海海事大学

Dates

Publication Date: 20260512
Application Date: 20220705

Claims (7)

1. A deep learning method for multi-step prediction of two ship navigation states is characterized by comprising the following steps: Step S1, acquiring original data of an automatic ship identification system, and preprocessing the original data of the automatic ship identification system to obtain preprocessed data; step S2, calculating the sum of weighted potential risk values of the data sequence pairs according to the preprocessed data and the time-matched ship automatic identification system data sequence pairs, and reserving the ship sequence data pair with the largest current potential risk to obtain a ship sequence data pair set; extracting a ship data matrix set meeting time conditions based on the preprocessed data, and matching corresponding data matrixes according to the ship automatic identification system data sequence pair; the limiting conditions of the time t in the ship data matrix set which is extracted to meet the time conditions are as follows: wherein: in order to predict the length of time of the required history data, In order to predict the starting moment of time, To predict a length of time of future data; S3, constructing a ship navigation state prediction network model according to the ship sequence data pair set so as to predict the future state of the ship; the ship navigation state prediction network model consists of an encoder, an interaction module and a decoder, The encoders are two sub-encoders with the same LSTM unit and are used for acquiring ship historical data information; the interaction module is used for fusing motion information among ships; the decoder is two sub-decoders with identical LSTM units independent of the encoder for predicting the future state of the ship; S4, calculating an error between an output value and a true value of the ship navigation state prediction network model according to the ship navigation state prediction network model, and dynamically adjusting the learning rate of the ship navigation state prediction model to obtain a prediction result; And S5, realizing data visualization of the prediction result.
2. The two-ship sailing state multi-step prediction oriented deep learning method according to claim 1, wherein in the step S1, preprocessing the ship automatic identification system raw data includes: S1.1, for the obtained original data of the ship automatic identification system, determining the data type and adjusting the data format of the original data, and primarily extracting the data through a binary search algorithm; S1.2, further refining the data range of the preliminary extracted data and eliminating abnormal data according to channel constraint boundaries and Laida criteria to obtain missing data; and S1.3, complementing the missing data through segmentation cubic spline interpolation to obtain the preprocessed data.
3. The two-ship sailing state multi-step prediction oriented deep learning method according to claim 2, wherein in the step S1.1, the preliminary data extraction by the binary search algorithm specifically includes extracting ship sailing state sequence data in a geographic range according to the geographic range related to the ship sailing state prediction from the original data of the ship automatic identification system.
4. The method for deep learning of multi-step prediction for two-vessel sailing state as claimed in claim 1, wherein in step S2, calculating the sum of weighted risk potential values of the pair of data sequences includes extracting longitude and latitude information, sailing information and captain information of the pair of data sequences of the automatic ship identification system, and calculating the slave of two vessels From moment to moment A risk potential value for the moment of time.
5. The two-ship sailing state oriented multi-step prediction deep learning method according to claim 4, wherein in the step S2, the pair of ship sequence data with the greatest reserved current potential risk is specifically: if the two vessels are from From moment to moment If the potential risk value at the moment is not greater than the current preset maximum value, eliminating the current corresponding ship sequence data pair; if the two vessels are from From moment to moment And if the potential risk value at the moment is larger than the current preset maximum value, updating the maximum value, and reserving the current corresponding ship sequence data pair.
6. The two-ship sailing state multi-step prediction oriented deep learning method according to claim 1, wherein in step S4, calculating an error between the ship sailing state prediction network model output value and the true value includes: step S4.1, mapping any ship sequence data in the ship sequence data pair set to the ship sequence data through maximum and minimum normalization Within the range; S4.2, acquiring training data according to the ship sequence data pair set, inputting the training data to the ship navigation state prediction network model, calculating to obtain an output value of the ship navigation state prediction network model and a true value of the ship navigation state prediction network model, and calculating an error between the output value and the true value; And S4.3, judging the error result, and dynamically adjusting the learning rate of the ship navigation state prediction network model according to the judging result to obtain the prediction result.
7. The two-ship sailing state oriented multi-step prediction deep learning method according to claim 1, wherein in the step S5, the prediction result is visualized by displaying the prediction result through python' S matplotlib kit to realize the visualization of the prediction result.

Description

Deep learning method for multi-step prediction of navigation states of two ships Technical Field The invention relates to the technical field of ship state prediction, in particular to a two-ship navigation state multi-step prediction oriented deep learning method. Background In maritime safety research, ship collision prevention is always a difficult task to be solved urgently, and as the number of shipping ships increases, the demand of shipping business increases, and the task to be solved becomes more important and urgent. Under the traditional ship driving environment, ship collision prevention is comprehensively completed by ship operators according to ship instrument information, navigation environment and driving experience. For the research of ship state prediction, two methods based on a motion model and a learning model are mainly available. Because of the different properties of ships, the existing motion model method is not suitable for big data analysis. The kalman filtering method in the existing learning model method is difficult to select a proper kernel function and needs to determine model parameters by other methods to help convergence. The Gaussian process regression method in the learning method estimates regression model parameters according to posterior distribution assumptions and sample data, and the distribution characteristics of the regression model parameters enable the method to be widely applicable, but are not beneficial to expansion and calculation when facing to a massive data set, and have the limitation that the influence of surrounding ships is neglected in single ship model prediction. Disclosure of Invention The invention aims to provide a deep learning method for multi-step prediction of two ship navigation states. The method aims to solve the problem that the existing ship state prediction method is not suitable for big data analysis and the limitation that the influence of surrounding ships is ignored in single ship model prediction. In order to achieve the above purpose, the invention is realized by the following technical scheme: The invention provides a deep learning method for multi-step prediction of two ship navigation states, which comprises the following steps: Step S1, acquiring original data of an automatic ship identification system, and preprocessing the original data of the automatic ship identification system to obtain preprocessed data; step S2, calculating the sum of weighted potential risk values of the data sequence pairs according to the preprocessed data and the time-matched ship automatic identification system data sequence pairs, and reserving the ship sequence data pair with the largest current potential risk to obtain a ship sequence data pair set; S3, constructing a ship navigation state prediction network model according to the ship sequence data pair set so as to predict the future state of the ship; S4, calculating an error between an output value and a true value of the ship navigation state prediction network model according to the ship navigation state prediction network model, and dynamically adjusting the learning rate of the ship navigation state prediction model to obtain a prediction result; And S5, realizing data visualization of the prediction result. Preferably, in the step S1, preprocessing the raw data of the automatic ship identification system includes: S1.1, for the obtained original data of the ship automatic identification system, determining the data type and adjusting the data format of the original data, and primarily extracting the data through a binary search algorithm; S1.2, further refining the data range of the preliminary extracted data and eliminating abnormal data according to channel constraint boundaries and Laida criteria to obtain missing data; and S1.3, complementing the missing data through segmentation cubic spline interpolation to obtain the preprocessed data. Preferably, in the step S1.1, the preliminary extraction of data by the binary search algorithm specifically includes extracting, from the raw data of the automatic ship identification system, ship navigation state sequence data within a geographic range involved in prediction of the ship navigation state. Preferably, in the step S2, the matching the data sequence pair of the automatic ship identification system according to time includes extracting a set of ship data matrices meeting a time condition based on the preprocessed data, and matching a corresponding data matrix according to the data sequence pair of the automatic ship identification system. Preferably, the limiting conditions of the time t in the ship data matrix set which meet the time conditions are as follows: to-tm≤t≤to+tn-1 Where t m is the length of time for predicting the required history data, T o is the predicted starting time of the process, T n is the length of time for predicting future data. Preferably, in the step S2, calculating the sum of weighted risk potential value