Ship bridge crossing track prediction method

Abstract

The invention discloses a ship bridge-crossing track prediction method, which comprises the following steps of: converting a time sequence of a track of each ship into an RGB (Red, Green, Blue) image by using a Grubrum angle field, then forming a training sample set by using the RGB image, and introducing a convolutional neural network and an attention mechanism into a current mature GAN (Graphics Area Network) structure to construct a ship bridge-crossing track prediction model. The convolutional neural network enables the ship bridge-crossing track prediction model to automatically learn features of an RGB image representing a ship track, and the attention mechanism enables the ship bridge-crossing track prediction model to fully utilize structural information in the RGB image, so that the ship bridge-crossing track prediction model can pay attention to an important long-distance dependency relationship in the RGB image. Key features in the RGB image are extracted more accurately; the method has the advantages that the prediction precision is high, and the requirements for model design and optimization are low.

Description

Technical Field

[0001] The invention relates to a track prediction method, in particular to a ship bridge crossing track prediction method. Background Art

[0002] With the continuous growth of global trade, the increase in water transport traffic has gradually increased the risk of ship-bridge collisions, especially in narrow or busy waterways. Ship-bridge collision accidents are a more serious type of maritime traffic accidents, which may not only cause significant casualties and environmental pollution, but also cause expensive infrastructure losses and long-term traffic disruptions. With the increasing global shipping traffic, the risk of ships crossing complex waters such as bridge areas has increased accordingly. Traditional navigation safety warning systems are often insufficient when dealing with complex and dynamic scenarios, and the risk management and prevention of ship-bridge collisions have become particularly important. Due to the complex navigation environment and large traffic flow in the bridge area, the traffic control department needs to understand the future bridge crossing track of the ship in advance through the Vessel Traffic Services (VTS) to identify potential risks in advance, have ample time for early warning and prevention, and improve decision-making support capabilities. In this context, the long-term prediction of the ship's bridge crossing track is of great significance for the traffic control department to perceive the risk of ship collision in advance and plan traffic flow.

[0003] The rapid development of artificial intelligence technology has promoted the widespread application of deep learning technology in the field of ship track prediction. Among them, representative ship track prediction methods are based on long short-term memory networks, gated recurrent units, etc. However, these methods are still unable to fully explore the motion characteristics between ship tracks and are only applicable to track prediction scenarios within a few minutes or shorter time range, such as ship collision avoidance. With the further development of deep learning technology and machine vision, generative adversarial networks (GANs), as an emerging technology in deep learning methods, are gradually being applied to ship track prediction. Generative adversarial networks optimize between the generative network and the discriminative network through an adversarial process. This adversarial training process enables the generator to continuously improve the quality of generated data, thereby improving the accuracy of long-term ship track prediction.

[0004] At present, Wang et al. published a GAN with attention module and interaction module to predict the trajectories of multiple ships in the paper "A prediction model of vessel trajectory based on generative adversarial network" to improve the accuracy of ship trajectory prediction in complex situations; Jia et al. proposed a ship warning area track prediction model based on conditional temporal GAN ​​in the paper "Conditional temporal GAN ​​for intent-aware vessel trajectory prediction in the precautionary area" to solve the uncertainty of navigation intentions in different route branches. However, the above two methods still have the following problems when predicting ship tracks: (1) When using the time series of ship tracks as training samples of the GAN model, it is difficult to capture the complex association between the temporal features in the time series and their related (invariant) attributes, which makes the GAN model unable to fully explore the motion features in the time series of ship tracks, and the prediction accuracy is difficult to break through; (2) There are long-term dependencies in the time series of ship tracks. The GAN model may face the challenge of difficulty in capturing long-term memory and long-term trends when dealing with long-term dependency modeling, which puts higher requirements on the design and optimization of the model. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting the track of a ship passing a bridge with high prediction accuracy and low requirements for the design and optimization of the model.

[0006] The technical solution adopted by the present invention to solve the above technical problem is: a method for predicting the track of a ship passing a bridge, comprising the following steps:

[0007] Step 1: Based on the AIS data of ships in the target bridge area, the time series of the tracks of several ships from entering the channel where the target bridge area is located to leaving the channel where the target bridge area is located are obtained, and the time series of the track of each ship is converted into an RGB image using the Gramian Angular Field (GAF), and the RGB image is called a track RGB image. At this time, several track RGB images are obtained by conversion, and the obtained several track RGB images are used to form a training sample set, wherein each track RGB image is used as a training sample in the training sample set;

[0008] Step 2: Introduce convolutional neural network and attention mechanism into the generator and discriminator of the GAN model to obtain an improved GAN model, and use the improved GAN model as a ship bridge track prediction model;

[0009] Step 3: Use the training sample set to train the ship bridge track prediction model until the prediction accuracy of the ship bridge track prediction model meets the preset accuracy requirements, thereby obtaining a trained ship bridge track prediction model;

[0010] Step 4: When predicting the track of a ship passing the bridge, the time series of the historical track of the ship from entering the channel where the target bridge area is located to the current prediction time is converted into an RGB image, which is the track RGB image of the ship. Then, the track RGB image of the ship is used as the input data of the trained ship crossing the bridge track prediction model, and the trained ship crossing the bridge track prediction model is used to predict the track RGB image of the ship, and the predicted track RGB image output of the ship is obtained;

[0011] Step 5: The predicted track RGB image of the ship is reversely converted into a time series. The time series is the predicted result of the ship's bridge crossing track, thus completing the prediction of the ship's bridge crossing track.

[0012] Compared with the prior art, the advantages of the present invention are that the time series of each ship's track is converted into RGB images by using Gramian Angular Field (GAF) and then the RGB images are used to form a training sample set. This conversion can help the bridge track prediction model capture the dynamic and periodic characteristics in the time series of each ship's track, laying a foundation for subsequently improving the efficiency of the ship bridge track prediction model in extracting navigation features in the ship track; the convolutional neural network (Convolutional neural network) is introduced into the current mature GAN network structure. The ship crossing bridge track prediction model is constructed by using convolutional neural networks (CNN) and attention mechanism. The convolutional neural network enables the ship crossing bridge track prediction model to automatically learn the features of the RGB image that characterizes the ship track, and can make full use of the structural information in the RGB image, laying the foundation for fully mining the ship track features. The attention module calculates the correlation between the various parts in the RGB image and assigns different weights, so that the ship crossing bridge track prediction model can pay attention to the important long-distance dependencies in the RGB image, and this weight assignment helps to more accurately extract the key features in the RGB image, thereby improving the details and overall quality of the predicted RGB image generated by the ship crossing bridge track prediction model and improving the prediction accuracy of the ship crossing the bridge track. Therefore, the prediction accuracy of the present invention is high and the requirements for model design and optimization are low.

[0013] Furthermore, in step 1, based on the AIS data of the ships in the target bridge area, a time series of the tracks of several ships from entering the channel where the target bridge area is located to leaving the channel where the target bridge area is located is obtained. The specific process of using GAF to convert the time series of the track of each ship into an RGB image is as follows:

[0014] Step A1, obtain AIS data of M ships in the target bridge area, where M is determined according to actual prediction needs. Theoretically, the larger the value, the better. The M ships are randomly numbered from 1 to M, and the ship numbered i is called the i-th ship, i = 1, 2, ..., M;

[0015] The longitude, latitude, heading, speed and timestamp in the AIS data of each ship are used to describe the navigation characteristics of each track point of each ship. Specifically, each track point in the AIS data of each ship is called the original track point, and the i-th ship is called the j-th track point. i The ship navigation characteristics of the original track point are recorded as Among them, j i =1,2,...,n i , n i represents the number of original track points passed by the i-th ship, As shown in formula (1):

[0016]

[0017] in, It means that the i-th ship is in its j-th i The longitude of the original track point, It means that the i-th ship is in its j-th i The latitude of the original track point, It means that the i-th ship is in its j-th i The heading of the original track point, It means that the i-th ship is in its j-th i The speed of the original track point, It means that the i-th ship is in its j-th i Timestamp of the original track points;

[0018] The original navigation behavior sequence of the i-th ship is recorded as S i ,

[0019] Step A2: Use Lagrange interpolation method to calculate the original navigation behavior sequence S of the i-th ship i Interpolation is performed to obtain the interpolated navigation behavior sequence of the i-th ship, which is recorded as C i , Among them, m iRepresents the original navigation behavior sequence S of the i-th ship i The number of track points passed by the i-th ship after interpolation will affect the original navigation behavior sequence S of the i-th ship. i After interpolation, the track point that the i-th ship passes through is called the interpolation track point, that is, the i-th ship passes through m i interpolated track points, The i-th ship is in its k-th i The ship navigation characteristics of k interpolated track points, i =1,2,...,m i , The i-th ship is in its k-th i The longitude of the interpolated track points, The i-th ship is in its k-th i The latitude of the interpolated track points, The i-th ship is in its k-th i The heading of the interpolated track points, The i-th ship is in its k-th i The speed of the interpolated track points, The i-th ship is in its k-th i The timestamps of the interpolated track points;

[0020] Step A3: Set the track point number threshold to 60, and determine whether the number of interpolated track points passed by each ship is less than 60. If it is less than 60, delete the interpolated navigation behavior sequence of the ship, otherwise keep it;

[0021] The number of interpolated navigation behavior sequences of all ships retained at this time is counted and recorded as W, that is, the interpolated navigation behavior sequences of W ships are retained at this time, and the W ships are called retained ships, that is, W retained ships;

[0022] The interpolated navigation behavior sequence of the wth reserved ship is recorded as D w , s w represents the number of interpolated track points passed by the w-th reserved ship, w = 1, 2, ..., W, It means that Article W reserves the ship in its Article E w The ship navigation characteristics of the interpolated track points, e w =1,2,...,s w ,s w represents the number of interpolated track points of the wth retained ship, It means that Article W reserves the ship in its Article E w The longitude of the interpolated track points, It means that Article W reserves the ship in its Article Ew The latitude of the interpolated track points, It means that Article W reserves the ship in its Article E w The heading of the interpolated track points, It means that Article W reserves the ship in its Article E w The speed of the interpolated track points, It means that Article W reserves the ship in its Article E w The timestamps of the interpolated track points;

[0023] Step A4: Interpolation navigation behavior sequence D of the retained ship based on the wth item w Construct 5 subsequences for the wth reserved ship. The first subsequence is the longitude subsequence, the second subsequence is the latitude subsequence, the third subsequence is the heading subsequence, the fourth subsequence is the speed subsequence, and the fifth subsequence is the timestamp subsequence. The longitude subsequence of the wth reserved ship is recorded as The latitude subsequence is recorded as The heading subsequence is recorded as The speed subsequence is recorded as The timestamp subsequence is recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as

[0024] Step A5: Use formula (2) to calculate the αth subsequence of the wth reserved ship. The e w Data Scaling is performed, where α=1,2,...,5, and the αth subsequence of the wth reserved ship is obtained. The e w Data Scaling data

[0025]

[0026] in, express The maximum data in express The minimum data in ;

[0027] Use the obtained to The αth scaled subsequence that constitutes the wth reserved ship is the αth scaled subsequence The e w Data:

[0028] Step A6: Use formula (3) to calculate the αth scaled subsequence of the wth reserved ship: The e w Data Transform to obtain the αth scaled subsequence of the wth reserved ship The e w Data The angle of the polar coordinates

[0029]

[0030] Among them, arccos is the symbol of inverse cosine;

[0031] Use the obtained to The αth scaled subsequence that constitutes the wth reserved ship The angle sequence of polar coordinates for The e w data, That is, the αth time series of the track of the wth reserved ship;

[0032] Step A7: Use the Gramian Angular Summation Fields (GASF) or Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) to convert the time series of each retained ship track into an RGB image, i.e., a track RGB image, and W track RGB images are obtained.

[0033] Furthermore, in the step A7, the specific process of converting the time series of the wth retained ship into the track RGB image using the Gramian Angular Summation Fields (GASF) of the Gramian Angular Field (GAF) is as follows:

[0034] Step B1: Use formula (4) to calculate the RGB image pixel matrix of the αth time series of the wth reserved ship track:

[0035]

[0036] Step B2: splice the RGB image pixel matrices of the five time series of the w-th retained ship track using formula (5) to obtain the RGB image pixel matrix of the w-th retained ship track:

[0037]

[0038] Step B3: Generate an RGB image of the wth retained ship track based on the RGB image pixel matrix of the wth retained ship track.

[0039] Furthermore, in the step A7, the specific process of converting the time series of the wth retained ship into the track RGB image using the Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) is as follows:

[0040] Step C1, using formula (6) to calculate the RGB image pixel matrix of the αth time series of the wth reserved ship track:

[0041]

[0042] Step C2: splice the RGB image pixel matrices of the five time series of the w-th retained ship track using formula (7) to obtain the RGB image pixel matrix of the w-th retained ship track:

[0043]

[0044] Step C3: generating an RGB image of the wth retained ship track based on the RGB image pixel matrix of the wth retained ship track.

[0045] Furthermore, the ship bridge track prediction model includes a generator and a discriminator. The generator includes a one-dimensional convolutional neural network layer, an attention layer and a fully connected layer. The one-dimensional convolutional neural network layer is called the first one-dimensional convolutional neural network layer, the attention layer is called the first attention layer, and the fully connected layer is called the first fully connected layer. The first one-dimensional convolutional neural network layer is used to access a pre-constructed random noise vector that satisfies the normal distribution, and perform preliminary feature extraction and dimensional expansion on it, construct potential features containing information required to generate an image, and output it to the first attention layer. The first attention layer is used to screen and weight the potential features output by the first one-dimensional convolutional neural network layer to focus on key features and integrate feature information. The first fully connected layer is used to integrate and map the weighted image features output by the first attention layer to the image space, generate a final RGB image, that is, a predicted track RGB image, and output it; the discriminator includes a one-dimensional convolutional neural network layer, an attention layer, and a fully connected layer, wherein the one-dimensional convolutional neural network layer is referred to as the second one-dimensional convolutional neural network layer, the attention layer is referred to as the second attention layer, and the fully connected layer is referred to as the second fully connected layer; the second one-dimensional convolutional neural network layer is used to access the predicted RGB image and the real RGB image output by the first fully connected layer, wherein the real RGB image is a Gramian angle field (Gramian angle field) Angular Field (GAF) converts the time series of the ship's track into an RGB image of the track, and performs preliminary image feature extraction and feature difference comparison on the two, and obtains preliminary image features to output to the second attention layer. The second attention layer is used to further focus on the features that are more valuable for authenticity judgment in the preliminary image features output by the second one-dimensional convolutional neural network, and screens and weights the preliminary image features output by the second one-dimensional convolutional neural network to obtain screened weighted image features to output to the second fully connected layer. The second fully connected layer is used to perform a comprehensive analysis on the screened weighted image features output by the second attention layer to obtain a judgment result, which is a probability estimate of whether the two types of RGB images, the predicted RGB image and the real RGB image connected to the second one-dimensional convolutional neural network layer, are real images or generated images.

[0046] Furthermore, when predicting the bridge crossing track of a certain ship in step 4, the specific process of converting the time series of the historical track of the ship from entering the channel where the target bridge area is located to the current prediction time into an RGB image is as follows: firstly, the AIS data of the ship from entering the channel where the target bridge area is located to the current prediction time is obtained, and the AIS data of the ship is processed according to steps A1 to A2 to obtain the interpolated navigation behavior sequence of the ship, and then the ship is directly used as a reserved ship, and steps A4 to A7 are used for processing to obtain the track RGB image of the ship.

[0047] Furthermore, the specific process of the step 5 of inversely converting the predicted track RGB image of the ship into a time series is as follows:

[0048] Step D1: The RGB pixel matrix corresponding to the RGB image of the predicted track of the ship is recorded as GAF. GAF is composed of the RGB image pixel matrices of the five time series of the predicted track of the ship. The RGB image pixel matrix of the αth time series of the predicted track of the ship is recorded as GAF. α , GAF α The number of rows and columns of GAF is denoted as d. α is a d-order square matrix, and the number of interpolated track points of the predicted ship's track is d; GAF α The pixel value of the bth row and bth column in is recorded as (GAF α ) b,b , b=1,2,...,d;

[0049] Step D2: Use formula (8) to convert the RGB image pixel matrix GAF of the αth time series of the predicted track of the ship into α The pixel value of the bth row and the bth column in (GAF α ) b,b Perform reverse transformation to obtain the corresponding time data

[0050]

[0051] Step D3: After denormalization, we get || is the symbol for taking absolute value;

[0052] Step D4: Use to The αth time subsequence of the predicted ship is in, is the predicted time data of the αth time subsequence of the ship;

[0053] Step D5: Sequentially convert the first time subsequence To the 5th time subsequence The time series is formed by splicing, and the time series is the prediction result of the ship's bridge crossing track, thus completing the prediction of the ship's bridge crossing track. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A prediction flow chart of the ship bridge crossing track prediction method of the present invention;

[0055] Figure 2 The ship trajectory diagram in the northbound traffic flow extracted from the waters near the Jintang Bridge in the second embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of an RGB image obtained by converting the time series of a ship's track using GASF in Example 2 of the present invention;

[0057] Figure 4 A schematic diagram of an RGB image generated by a trained ship track prediction model (hereinafter referred to as GASF-GAN) in Example 2 of the present invention, in which the time series of the ship's track is converted using GASF, and the obtained RGB image is used as a training sample;

[0058] Figure 5 A schematic diagram showing a comparison between the ship's bridge crossing track predicted by the GASF-GAN model in the second embodiment of the present invention and the ship's bridge crossing track predicted by the ship's bridge crossing track prediction model trained using a simple time series as a training sample (hereinafter, the ship's bridge crossing track prediction model is referred to as TM-GAN) and the ship's actual bridge crossing track;

[0059] Figure 6 This is a schematic diagram of an RGB image obtained by converting the time series of the ship's track using GADF in Example 3 of the present invention;

[0060] Figure 7 A schematic diagram of an RGB image generated by using GADF to transform the time series of the ship's track as a training sample and trained ship bridge track prediction (hereinafter referred to as GADF-GAN) in Example 3 of the present invention;

[0061] Figure 8 This is a schematic diagram comparing the bridge crossing track of a ship predicted by the GADF-GAN model and the bridge crossing track of a ship predicted by the TM-GAN model in Example 3 of the present invention with the actual track of the ship. DETAILED DESCRIPTION

[0062] The present invention is further described in detail below with reference to the accompanying drawings.

[0063] Embodiment 1: Figure 1 As shown, a method for predicting a ship's bridge track includes the following steps:

[0064] Step 1: Based on the AIS data of ships in the target bridge area, the time series of the tracks of several ships from entering the channel where the target bridge area is located to leaving the channel where the target bridge area is located are obtained, and the time series of the track of each ship is converted into an RGB image using the Gramian Angular Field (GAF), and the RGB image is called a track RGB image. At this time, several track RGB images are obtained by conversion, and the obtained several track RGB images are used to form a training sample set, wherein each track RGB image is used as a training sample in the training sample set;

[0065] Step 2: Introduce convolutional neural network and attention mechanism into the generator and discriminator of the GAN model to obtain an improved GAN model, and use the improved GAN model as a ship bridge track prediction model;

[0066] Step 3: Use the training sample set to train the ship bridge track prediction model until the prediction accuracy of the ship bridge track prediction model meets the preset accuracy requirements, thereby obtaining a trained ship bridge track prediction model;

[0067] Step 4: When predicting the track of a ship passing the bridge, the time series of the historical track of the ship from entering the channel where the target bridge area is located to the current prediction time is converted into an RGB image, which is the track RGB image of the ship. Then, the track RGB image of the ship is used as the input data of the trained ship crossing the bridge track prediction model, and the trained ship crossing the bridge track prediction model is used to predict the track RGB image of the ship, and the predicted track RGB image output of the ship is obtained;

[0068] Step 5: The predicted track RGB image of the ship is reversely converted into a time series. The time series is the predicted result of the ship's bridge crossing track, thus completing the prediction of the ship's bridge crossing track.

[0069] In this embodiment, the time series of each ship's track is converted into RGB images by using Gramian Angular Field (GAF), and then the RGB images are used to form a training sample set. This conversion can help the bridge track prediction model capture the dynamic and periodic characteristics of each ship's track in the time series, laying a foundation for improving the efficiency of the ship bridge track prediction model in extracting navigation features in the ship track. In the current mature GAN network structure, a convolutional neural network is introduced. The ship crossing bridge track prediction model is constructed by using convolutional neural networks (CNN) and attention mechanism. The convolutional neural network enables the ship crossing bridge track prediction model to automatically learn the features of the RGB image that characterizes the ship track, and can make full use of the structural information in the RGB image, laying the foundation for fully mining the ship track features. The attention module calculates the correlation between the various parts in the RGB image and assigns different weights, so that the ship crossing bridge track prediction model can pay attention to the important long-distance dependencies in the RGB image, and this weight assignment helps to more accurately extract the key features in the RGB image, thereby improving the details and overall quality of the predicted RGB image generated by the ship crossing bridge track prediction model and improving the prediction accuracy of the ship crossing the bridge track. Therefore, the prediction accuracy of the present invention is high and the requirements for model design and optimization are low.

[0070] Example 2: This example is basically the same as Example 1, except that: in this example, based on the AIS data of ships in the target bridge area waters in step 1, a time series of the tracks of several ships from entering the channel where the target bridge area is located to leaving the channel where the target bridge area is located is obtained, and the specific process of using GAF to convert the time series of the track of each ship into an RGB image is as follows:

[0071] Step A1, obtain AIS data of M ships in the target bridge area, where M is determined according to actual prediction needs. Theoretically, the larger the value, the better. The M ships are randomly numbered from 1 to M, and the ship numbered i is called the i-th ship, i = 1, 2, ..., M;

[0072] The longitude, latitude, heading, speed and timestamp in the AIS data of each ship are used to describe the navigation characteristics of each track point of each ship. Specifically, each track point in the AIS data of each ship is called the original track point, and the i-th ship is called the j-th track point. i The ship navigation characteristics of the original track point are recorded as Among them, j i =1,2,...,n i , n i represents the number of original track points passed by the i-th ship, As shown in formula (1):

[0073]

[0074] in, It means that the i-th ship is in its j-th i The longitude of the original track point, It means that the i-th ship is in its j-th i The latitude of the original track point, It means that the i-th ship is in its j-th i The heading of the original track point, It means that the i-th ship is in its j-th i The speed of the original track point, It means that the i-th ship is in its j-th i Timestamp of the original track points;

[0075] The original navigation behavior sequence of the i-th ship is recorded as S i ,

[0076] Step A2: Use Lagrange interpolation method to calculate the original navigation behavior sequence S of the i-th ship i Interpolation is performed to obtain the interpolated navigation behavior sequence of the i-th ship, which is recorded as C i , Among them, m i Represents the original navigation behavior sequence S of the i-th ship i The number of track points passed by the i-th ship after interpolation will affect the original navigation behavior sequence S of the i-th ship. i After interpolation, the track point that the i-th ship passes through is called the interpolation track point, that is, the i-th ship passes through m i interpolated track points, The i-th ship is in its k-th i The ship navigation characteristics of k interpolated track points, i =1,2,...,m i , The i-th ship is in its k-th i The longitude of the interpolated track points, The i-th ship is in its k-th i The latitude of the interpolated track points, The i-th ship is in its k-th i The heading of the interpolated track points, The i-th ship is in its k-th i The speed of the interpolated track points, The i-th ship is in its k-th i The timestamps of the interpolated track points;

[0077] Step A3: Set the track point number threshold to 60, and determine whether the number of interpolated track points passed by each ship is less than 60. If it is less than 60, delete the interpolated navigation behavior sequence of the ship, otherwise keep it;

[0078] The number of interpolated navigation behavior sequences of all ships retained at this time is counted and recorded as W, that is, the interpolated navigation behavior sequences of W ships are retained at this time, and the W ships are called retained ships, that is, W retained ships;

[0079] The interpolated navigation behavior sequence of the wth reserved ship is recorded as D w , s w represents the number of interpolated track points passed by the w-th reserved ship, w = 1, 2, ..., W, It means that Article W reserves the ship in its Article E w The ship navigation characteristics of the interpolated track points, e w =1,2,...,s w ,s w represents the number of interpolated track points of the wth retained ship, It means that Article W reserves the ship in its Article E w The longitude of the interpolated track points, It means that Article W reserves the ship in its Article E w The latitude of the interpolated track points, It means that Article W reserves the ship in its Article E w The heading of the interpolated track points, It means that Article W reserves the ship in its Article E w The speed of the interpolated track points, It means that Article W reserves the ship in its Article E w The timestamps of the interpolated track points;

[0080] Step A4: Interpolation navigation behavior sequence D of the retained ship based on the wth item w Construct 5 subsequences for the wth reserved ship. The first subsequence is the longitude subsequence, the second subsequence is the latitude subsequence, the third subsequence is the heading subsequence, the fourth subsequence is the speed subsequence, and the fifth subsequence is the timestamp subsequence. The longitude subsequence of the wth reserved ship is recorded as The latitude subsequence is recorded as The heading subsequence is recorded as The speed subsequence is recorded as The timestamp subsequence is recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as

[0081] Step A5: Use formula (2) to calculate the αth subsequence of the wth reserved ship. The e w Data Scaling is performed, where α=1,2,...,5, and the αth subsequence of the wth reserved ship is obtained. The e w Data Scaling data

[0082]

[0083] in, express The maximum data in express The minimum data in ;

[0084] Use the obtained to The αth scaled subsequence that constitutes the wth reserved ship is the αth scaled subsequence The e w Data:

[0085] Step A6: Use formula (3) to calculate the αth scaled subsequence of the wth reserved ship: The e w Data Transform to obtain the αth scaled subsequence of the wth reserved ship The e w Data The angle of the polar coordinates

[0086]

[0087] Where arccos is the inverse cosine symbol; using the obtained to The αth scaled subsequence that constitutes the wth reserved ship The angle sequence of polar coordinates for The e w data, That is, the αth time series of the track of the wth reserved ship;

[0088] Step A7: Use the Gramian Angular Summation Fields (GASF) or Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) to convert the time series of each retained ship track into an RGB image, i.e., a track RGB image, and W track RGB images are obtained.

[0089] In this embodiment, in step A7, the specific process of converting the time series of the wth retained ship into the track RGB image using the Gramian Angular Summation Fields (GASF) of the Gramian Angular Field (GAF) is as follows:

[0090] Step B1: Use formula (4) to calculate the RGB image pixel matrix of the αth time series of the wth reserved ship track:

[0091]

[0092] Step B2: splice the RGB image pixel matrices of the five time series of the w-th retained ship track using formula (5) to obtain the RGB image pixel matrix of the w-th retained ship track:

[0093]

[0094] Step B3: Generate an RGB image of the wth retained ship track based on the RGB image pixel matrix of the wth retained ship track.

[0095] In this embodiment, GASF is used to convert the time series of ship tracks into RGB images. This conversion can help capture the dynamic and periodic characteristics in the time series data, laying a foundation for subsequently improving the extraction efficiency of navigation features in ship tracks.

[0096] In this embodiment, the ship bridge track prediction model includes a generator and a discriminator. The generator includes a one-dimensional convolutional neural network layer, an attention layer and a fully connected layer. Its one-dimensional convolutional neural network layer is called the first one-dimensional convolutional neural network layer, its attention layer is called the first attention layer, and its fully connected layer is called the first fully connected layer; the first one-dimensional convolutional neural network layer is used to access a pre-constructed random noise vector that satisfies the normal distribution, and perform preliminary feature extraction and dimensional expansion on it, construct potential features containing information required to generate an image, and output it to the first attention layer. The first attention layer is used to screen and weight the potential features output by the first one-dimensional convolutional neural network layer to focus on key features and integrate feature information. , the image features after feature weighting are output to the first fully connected layer, the first fully connected layer is used to integrate and map the image features after feature weighting outputted by the first attention layer to the image space, generate the final RGB image, that is, the predicted track RGB image, and output it; the discriminator includes a one-dimensional convolutional neural network layer, an attention layer and a fully connected layer, the one-dimensional convolutional neural network layer is called the second one-dimensional convolutional neural network layer, the attention layer is called the second attention layer, and the fully connected layer is called the second fully connected layer; the second one-dimensional convolutional neural network layer is used to access the predicted RGB image and the real RGB image outputted by the first fully connected layer, wherein the real RGB image is a Gramian angle field (Gramian Angular Field (GAF) converts the time series of the ship's track into an RGB image of the track, and performs preliminary image feature extraction and feature difference comparison on the two, and obtains preliminary image features to output to the second attention layer. The second attention layer is used to further focus on the features that are more valuable for authenticity judgment in the preliminary image features output by the second one-dimensional convolutional neural network, and screens and weights the preliminary image features output by the second one-dimensional convolutional neural network to obtain screened weighted image features to output to the second fully connected layer. The second fully connected layer is used to perform a comprehensive analysis on the screened weighted image features output by the second attention layer to obtain a judgment result, which is a probability estimate of whether the two types of RGB images, the predicted RGB image and the real RGB image connected to the second one-dimensional convolutional neural network layer, are real images or generated images.

[0097] In this embodiment, the ship bridge track prediction model incorporates a one-dimensional convolutional neural network into the original structure of GAN (generator and discriminator), so that the one-dimensional convolutional neural network can automatically learn the characteristics of the ship track RGB image, and make full use of the structural information in the image, laying the foundation for the ship bridge track prediction model to fully explore the ship track characteristics. In addition, the ship bridge track prediction model introduces an attention mechanism into the original structure of GAN (generator and discriminator), so that the correlation between the various parts of the RGB image can be calculated through the attention module, and different weights are assigned, so that the ship bridge track prediction model can pay attention to the important long-distance dependencies in the RGB image through the attention module, and this weight distribution method helps to more accurately extract the key features in the RGB image, thereby improving the details and overall quality of the generated RGB image, and improving the accuracy of the long-term prediction of the ship bridge track.

[0098] In this embodiment, when predicting the bridge crossing track of a certain ship in step 4, the specific process of converting the time series of the historical track of the ship from entering the channel where the target bridge area is located to the current prediction time into the RGB image is as follows: firstly, the AIS data of the ship from entering the channel where the target bridge area is located to the current prediction time is obtained, and the AIS data of the ship is processed according to steps A1 to A2 to obtain the interpolated navigation behavior sequence of the ship, and then the ship is directly used as a reserved ship, and steps A4 to A7 are used to process it to obtain the track RGB image of the ship.

[0099] In this embodiment, the specific process of step 5 inversely converting the predicted track RGB image of the ship into a time series is as follows:

[0100] Step D1: The RGB pixel matrix corresponding to the RGB image of the predicted track of the ship is recorded as GAF. GAF is composed of the RGB image pixel matrices of the five time series of the predicted track of the ship. The RGB image pixel matrix of the αth time series of the predicted track of the ship is recorded as GAF. α , GAF α The number of rows and columns of GAF is denoted as d. α is a d-order square matrix, and the number of interpolated track points of the predicted ship's track is d; GAF α The pixel value of the bth row and bth column in is recorded as (GAF α ) b,b , b=1,2,...,d;

[0101] Step D2: Use formula (8) to convert the RGB image pixel matrix GAF of the αth time series of the predicted track of the ship into α The pixel value of the bth row and the bth column in (GAFα ) b,b Perform reverse transformation to obtain the corresponding time data

[0102]

[0103] Step D3: After denormalization, we get || is the absolute value symbol;

[0104] Step D4: Use to The αth time subsequence of the predicted ship is in, is the predicted time data of the αth time subsequence of the ship;

[0105] Step D5: Sequentially convert the first time subsequence To the 5th time subsequence The time series is formed by splicing, and the time series is the prediction result of the ship's bridge crossing track, thus completing the prediction of the ship's bridge crossing track.

[0106] In this example, the AIS data of the waters near the Jintang Bridge in Ningbo-Zhoushan Port from June 1, 2020 to December 31, 2020 was selected as the research object. The ship trajectory diagram in the northbound traffic flow after extraction in the waters near the Jintang Bridge is shown in Figure 2. Figure 2 shown. Figure 2 In the figure, point A and point B are the locations of the two piers of Jintang Bridge, with specific coordinates of (30°03′39.16″N, 121°48′5.08″E) and (30°03′39.7″N, 121°48′27.95″E) respectively. There is a lane separation system between the two piers of Jintang Bridge, separating the traffic flows in two opposite directions (the left side of the channel is the southbound traffic flow, and the right side of the channel is the northbound traffic flow). In order to better analyze the changes in ship traffic flow, only the northbound traffic flow is selected for analysis, and the AIS tracks (i.e., ship AIS data) of 1091 ships in the waters are extracted, as shown in Figure 1. Figure 2 The local enlarged image is shown in Figure 1. These AIS tracks are processed using steps A1 to A4. The processed AIS track data is converted into RGB images using GADF, where the RGB image converted from a track data is shown in Figure 1. Figure 3 As shown in FIG. 1 , these RGB images are divided into a training set and a validation set. The training set is sent to the ship bridge track prediction model proposed in the present invention to train the ship bridge track prediction model (the trained ship bridge track prediction model is referred to as GASF-GAN). Six RGB images generated by the validation set are randomly selected for visualization, as shown in FIG. Figure 4 As shown in the figure. The initial score (IS), FrechetInception distance (FID) score and maximum mean discrepancy (MMD) of the ship bridge track prediction model in the validation set were recorded as the error indicators of the ship bridge track prediction model, which were 2.3016, 5.0318 and 1.4734 respectively. Subsequently, these RGB images were inversely converted into time series.

[0107] In order to verify the prediction effect of using GASF for graph transformation and then training the ship crossing bridge track prediction model, the ship crossing bridge track prediction model was trained using a simple time series (the trained ship crossing bridge track prediction model is referred to as the TM-GAN model) and the GASF-GAN model for ship track prediction. The prediction results of 6 navigation tracks were randomly selected for comparison. Figure 5 As shown. Figure 5 It can be seen that compared with the TM-GAN model, the predicted trajectory of the GASF-GAN model proposed in the present invention has a higher fit with the actual trajectory and a better prediction effect. In order to further quantify the prediction accuracy of the two track prediction models, the TM-GAN model and the GASF-GAN model, the mean square error (MSE), final displacement error (FDE), Flamenche distance (FD) and Euclidean distance (Average Euclidean Distance, AED) are selected as error indicators, and the four average error indicators of the predicted trajectory are calculated, as shown in Table 1:

[0108] Table 1 Error indicators of trajectories predicted using TM-GAN model and GASF-GAN model

[0109]

[0110] By analyzing the data in Table 1, it can be seen that compared with the TM-GAN model, the prediction effect of the GASF-GAN model proposed in the present invention is significantly improved.

[0111] Embodiment 3: This embodiment is basically the same as Embodiment 2, except that: in this embodiment, in step A7, instead of using the Gramian Angular Summation Fields (GASF) of the Gramian Angular Field (GAF) to convert the time series of the w-th retained ship into a track RGB image, the Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) is used to convert the time series of the w-th retained ship into a track RGB image.

[0112] In this embodiment, in step 7, the specific process of converting the time series of the wth retained ship into the track RGB image using the Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) is as follows:

[0113] Step C1, using formula (6) to calculate the RGB image pixel matrix of the αth time series of the wth reserved ship track:

[0114]

[0115] Step C2: splice the RGB image pixel matrices of the five time series of the w-th retained ship track using formula (7) to obtain the RGB image pixel matrix of the w-th retained ship track:

[0116]

[0117] Step C3: generating an RGB image of the wth retained ship track based on the RGB image pixel matrix of the wth retained ship track.

[0118] In this embodiment, GADF is used to convert the time series of ship tracks into RGB images. This conversion can help capture the dynamic and periodic characteristics in the time series data, laying a foundation for subsequently improving the efficiency of extracting navigation features in ship tracks.

[0119] In this example, the AIS data of the waters near the Jintang Bridge in Ningbo-Zhoushan Port from June 1, 2020 to December 31, 2020 was selected as the research object. The ship trajectory diagram in the northbound traffic flow after extraction in the waters near the Jintang Bridge is shown in the figure below: Figure 2 shown. Figure 2In the figure, point A and point B are the locations of the two piers of Jintang Bridge, with specific coordinates of (30°03′39.16″N, 121°48′5.08″E) and (30°03′39.7″N, 121°48′27.95″E) respectively. There is a lane separation system between the two piers of Jintang Bridge, separating the traffic flows in two opposite directions (the left side of the channel is the southbound traffic flow, and the right side of the channel is the northbound traffic flow). In order to better analyze the changes in ship traffic flow, only the northbound traffic flow is selected for analysis, and the AIS tracks (i.e., ship AIS data) of 1091 ships in the waters are extracted, as shown in Figure 1. Figure 2 The AIS tracks are processed using steps A1 to A4.

[0120] Use GADF to convert the processed trajectory data into RGB images. The RGB image converted from a certain trajectory data is as follows: Figure 6 As shown in FIG. 1 , these RGB images are divided into a training set and a validation set. The training set is sent to the ship bridge track prediction model proposed in the present invention to train the ship bridge track prediction model (the trained ship bridge track prediction model is referred to as the GADF-GAN model). Six RGB images generated by the validation set are randomly selected for visualization, as shown in FIG. Figure 7 As shown. The initial score (IS), Frechet Inception distance (FID) score and maximum mean discrepancy (MMD) of the ship bridge track prediction model in the validation set are recorded as the error indicators of the ship bridge track prediction model, which are 0.4324, 16.8178 and 2.0120 respectively. And compare it with the relevant error indicators of the prediction results of the GASF-GAN model in Example 2, as shown in Table 2:

[0121] Table 2 Index performance of GASF-GAN model and GADF-GAN model

[0122]

[0123] Analyzing Table 2, it can be seen that the prediction error of the GADF-GAN model is larger than that of the GASF-GAN model. Subsequently, these RGB images are inversely converted into time series.

[0124] In order to verify the prediction effect of the ship bridge track prediction model proposed in this invention after the graphics conversion using GADF, the TM-GAN model and the GADF-GAN model are used for track prediction respectively, and the prediction results of 16 navigation tracks are randomly selected for comparison. Figure 8 As shown, from Figure 8It can be seen that compared with the TM-GAN model, the predicted trajectory of the GADF-GAN model proposed in the present invention has a higher fit with the actual trajectory and a better prediction effect. In order to further quantify the prediction accuracy of the two track prediction models, the TM-GAN model and the GADF-GAN model, the mean square error (MSE), final displacement error (FDE), Flamenche distance (FD) and Euclidean distance (Average Euclidean Distance, AED) are selected as error indicators, and the four average error indicators of the predicted trajectory of the validation set are calculated, as shown in Table 3:

[0125] Table 3 Error indicators of trajectories predicted using TM-GAN model and GADF-GAN model

[0126]

[0127] From the analysis of Table 3, we can see that although the prediction accuracy of the GADF-GAN model is not as good as that of the GASF-GAN model, the prediction effect of the GADF-GAN model is still improved compared with the TM-GAN model.

[0128] In summary, the ship bridge track prediction method of the present invention has high prediction accuracy and low requirements for model design and optimization.

Claims

1. A method for predicting the track of a ship passing a bridge, characterized in that The following steps are involved: Step 1: Based on the AIS data of ships in the target bridge area, the time series of the tracks of several ships from entering the channel where the target bridge area is located to leaving the channel where the target bridge area is located are obtained, and the time series of the track of each ship is converted into an RGB image using the Gramian Angular Field (GAF), and the RGB image is called a track RGB image. At this time, several track RGB images are obtained by conversion, and the obtained several track RGB images are used to form a training sample set, wherein each track RGB image is used as a training sample in the training sample set; Step 2: Introduce convolutional neural network and attention mechanism into the generator and discriminator of the GAN model to obtain an improved GAN model, and use the improved GAN model as a ship bridge track prediction model; Step 3: Use the training sample set to train the ship bridge track prediction model until the prediction accuracy of the ship bridge track prediction model meets the preset accuracy requirements, thereby obtaining a trained ship bridge track prediction model; Step 4: When predicting the track of a ship passing the bridge, the time series of the historical track of the ship from entering the channel where the target bridge area is located to the current prediction time is converted into an RGB image, which is the track RGB image of the ship. Then, the track RGB image of the ship is used as the input data of the trained ship crossing the bridge track prediction model, and the trained ship crossing the bridge track prediction model is used to predict the track RGB image of the ship, and the predicted track RGB image output of the ship is obtained; Step 5: The predicted track RGB image of the ship is reversely converted into a time series. The time series is the predicted result of the ship's bridge crossing track, thus completing the prediction of the ship's bridge crossing track.

2. A method for predicting a ship's bridge track according to claim 1, characterized in that In step 1, based on the AIS data of ships in the target bridge area, a time series of the tracks of several ships from entering the channel where the target bridge area is located to leaving the channel where the target bridge area is located is obtained. The specific process of using GAF to convert the time series of the track of each ship into an RGB image is as follows: Step A1, obtain AIS data of M ships in the target bridge area, where M is determined according to actual prediction needs. Theoretically, the larger the value, the better. The M ships are randomly numbered from 1 to M, and the ship numbered i is called the i-th ship, i = 1, 2, ..., M; The longitude, latitude, heading, speed and timestamp in the AIS data of each ship are used to describe the navigation characteristics of each track point of each ship. Specifically, each track point in the AIS data of each ship is called the original track point, and the i-th ship is called the j-th track point. i The ship navigation characteristics of the original track point are recorded as N i ji , where j i =1,2,...,n i , n i represents the number of original track points passed by the i-th ship, As shown in formula (1): in, It means that the i-th ship is in its j-th i The longitude of the original track point, It means that the i-th ship is in its j-th i The latitude of the original track point, The i-th ship is in its j-th i The heading of the original track point, It means that the i-th ship is in its j-th i The speed of the original track point, The i-th ship is in its j-th i Timestamp of the original track points; The original navigation behavior sequence of the i-th ship is recorded as S i , Step A2: Use Lagrange interpolation method to calculate the original navigation behavior sequence S of the i-th ship i Interpolation is performed to obtain the interpolated navigation behavior sequence of the i-th ship, which is recorded as C i Among them, m i Represents the original navigation behavior sequence S of the i-th ship i The number of track points passed by the i-th ship after interpolation will affect the original navigation behavior sequence S of the i-th ship. i After interpolation, the track point that the i-th ship passes through is called the interpolation track point, that is, the i-th ship passes through m i interpolated track points, The i-th ship is in its k-th i The ship navigation characteristics of k interpolated track points, i =1,2,...,m i , The i-th ship is in its k-th i The longitude of the interpolated track points, The i-th ship is in its k-th i The latitude of the interpolated track points, The i-th ship is in its k-th i The heading of the interpolated track points, The i-th ship is in its k-th i The speed of the interpolated track points, The i-th ship is in its k-th i The timestamps of the interpolated track points; Step A3: Set the track point number threshold to 60, and determine whether the number of interpolated track points passed by each ship is less than 60. If it is less than 60, delete the interpolated navigation behavior sequence of the ship, otherwise keep it; The number of interpolated navigation behavior sequences of all ships retained at this time is counted and recorded as W, that is, the interpolated navigation behavior sequences of W ships are retained at this time, and the W ships are called retained ships, that is, W retained ships; The interpolated navigation behavior sequence of the wth reserved ship is recorded as D w , s w represents the number of interpolated track points passed by the w-th reserved ship, w = 1, 2, ..., W, It means that Article W reserves the ship in its Article E w The ship navigation characteristics of the interpolated track points, e w =1,2,...,s w ,s w represents the number of interpolated track points of the wth retained ship, It means that Article W reserves the ship in its Article E w The longitude of the interpolated track points, It means that Article W reserves the ship in its Article E w The latitude of the interpolated track points, It means that Article W reserves the ship in its Article E w The heading of the interpolated track points, It means that Article W reserves the ship in its Article E w The speed of the interpolated track points, It means that Article W reserves the ship in its Article E w The timestamps of the interpolated track points; Step A4: Interpolation navigation behavior sequence D of the retained ship based on the wth item w Construct 5 subsequences for the wth reserved ship. The first subsequence is the longitude subsequence, the second subsequence is the latitude subsequence, the third subsequence is the heading subsequence, the fourth subsequence is the speed subsequence, and the fifth subsequence is the timestamp subsequence. The longitude subsequence of the wth reserved ship is recorded as The latitude subsequence is recorded as The heading subsequence is recorded as The speed subsequence is recorded as The timestamp subsequence is recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as for The e w data, recorded as Step A5: Use formula (2) to calculate the αth subsequence of the wth reserved ship. The e w Data Scaling is performed, where α=1,2,...,5, and the αth subsequence of the wth reserved ship is obtained. The e w Data Scaling data in, express The maximum data in express The minimum data in ; Use the obtained to The αth scaled subsequence that constitutes the wth reserved ship is the αth scaled subsequence The e w Data: Step A6: Use formula (3) to calculate the αth scaled subsequence of the wth reserved ship: The e w Data Transform to obtain the αth scaled subsequence of the wth reserved ship The e w Data The angle of the polar coordinates Among them, arccos is the symbol of inverse cosine; Use the obtained to The αth scaled subsequence that constitutes the wth reserved ship The angle sequence of polar coordinates for The e w data, That is, the αth time series of the track of the wth reserved ship; Step A7: Use the Gramian Angular Summation Fields (GASF) or Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) to convert the time series of each retained ship track into an RGB image, i.e., a track RGB image, and W track RGB images are obtained.

3. A method for predicting a ship's bridge track according to claim 2, characterized in that In the step A7, the specific process of converting the time series of the wth retained ship into the track RGB image using the Gramian Angular Sum Fields (GASF) of the Gramian Angular Field (GAF) is as follows: Step B1: Use formula (4) to calculate the RGB image pixel matrix of the αth time series of the wth reserved ship track: Step B2: splice the RGB image pixel matrices of the five time series of the w-th retained ship track using formula (5) to obtain the RGB image pixel matrix of the w-th retained ship track: Step B3: Generate an RGB image of the wth retained ship track based on the RGB image pixel matrix of the wth retained ship track.

4. A method for predicting a ship's bridge track according to claim 2, characterized in that In the step A7, the specific process of converting the time series of the wth reserved ship into the track RGB image using the Gramian Angular Difference Fields (GADF) of the Gramian Angular Field (GAF) is as follows: Step C1, using formula (6) to calculate the RGB image pixel matrix of the αth time series of the wth reserved ship track: Step C2: splice the RGB image pixel matrices of the five time series of the w-th retained ship track using formula (7) to obtain the RGB image pixel matrix of the w-th retained ship track: Step C3: generating an RGB image of the wth retained ship track based on the RGB image pixel matrix of the wth retained ship track.

5. A method for predicting a ship's bridge track according to claim 1, characterized in that The ship bridge track prediction model includes a generator and a discriminator. The generator includes a one-dimensional convolutional neural network layer, an attention layer and a fully connected layer. The one-dimensional convolutional neural network layer is called the first one-dimensional convolutional neural network layer, the attention layer is called the first attention layer, and the fully connected layer is called the first fully connected layer; the first one-dimensional convolutional neural network layer is used to access a pre-constructed random noise vector that satisfies the normal distribution, and perform preliminary feature extraction and dimensional expansion on it, construct potential features containing information required to generate an image, and output it to the first attention layer. The first attention layer is used to screen and weight the potential features output by the first one-dimensional convolutional neural network layer to focus on key features and integrate feature information, so as to obtain The image features after feature weighting are output to the first fully connected layer, and the first fully connected layer is used to integrate and map the image features after feature weighting outputted by the first attention layer to the image space, generate the final RGB image, that is, the predicted track RGB image, and output it; the discriminator includes a one-dimensional convolutional neural network layer, an attention layer and a fully connected layer, and the one-dimensional convolutional neural network layer is called the second one-dimensional convolutional neural network layer, the attention layer is called the second attention layer, and the fully connected layer is called the second fully connected layer; the second one-dimensional convolutional neural network layer is used to access the predicted RGB image and the real RGB image outputted by the first fully connected layer, wherein the real RGB image is a Gramian angle field (Gramian Angular Field (GAF) converts the time series of the ship's track into an RGB image of the track, and performs preliminary image feature extraction and feature difference comparison on the two, and obtains preliminary image features to output to the second attention layer. The second attention layer is used to further focus on the features that are more valuable for authenticity judgment in the preliminary image features output by the second one-dimensional convolutional neural network, and screens and weights the preliminary image features output by the second one-dimensional convolutional neural network to obtain screened weighted image features to output to the second fully connected layer. The second fully connected layer is used to perform a comprehensive analysis on the screened weighted image features output by the second attention layer to obtain a judgment result, which is a probability estimate of whether the two types of RGB images, the predicted RGB image and the real RGB image connected to the second one-dimensional convolutional neural network layer, are real images or generated images.

6. A method for predicting a ship's bridge track according to claim 1, characterized in that When predicting the bridge crossing track of a certain ship in step 4, the specific process of converting the time series of the historical track of the ship from entering the channel where the target bridge area is located to the current prediction time into the RGB image is as follows: firstly, the AIS data of the ship from entering the channel where the target bridge area is located to the current prediction time is obtained, and the AIS data of the ship is processed according to steps A1 to A2 to obtain the interpolated navigation behavior sequence of the ship, and then the ship is directly used as a reserved ship, and steps A4 to A7 are used for processing to obtain the track RGB image of the ship.

7. A method for predicting a ship's bridge track according to claim 6, characterized in that The specific process of step 5 to reversely convert the predicted track RGB image of the ship into a time series is as follows: Step D1: The RGB pixel matrix corresponding to the RGB image of the predicted track of the ship is recorded as GAF. GAF is composed of the RGB image pixel matrices of the five time series of the predicted track of the ship. The RGB image pixel matrix of the αth time series of the predicted track of the ship is recorded as GAF. α , GAF α The number of rows and columns of GAF is denoted as d. α is a d-order square matrix, and the number of interpolated track points of the predicted ship's track is d; GAF α The pixel value of the bth row and bth column in is recorded as (GAF α ) b,b , b=1,2,...,d; Step D2: Use formula (8) to convert the RGB image pixel matrix GAF of the αth time series of the predicted track of the ship into α The pixel value of the bth row and the bth column in (GAF α ) b,b Perform reverse transformation to obtain the corresponding time data Step D3: After denormalization, we get || is the absolute value symbol; Step D4: Use to The αth time subsequence of the predicted ship is in, is the predicted time data of the αth time subsequence of the ship; Step D5: Sequentially convert the first time subsequence To the 5th time subsequence The time series is formed by splicing, and the time series is the prediction result of the ship's bridge crossing track, thus completing the prediction of the ship's bridge crossing track.

Images