A Ship Track Prediction Method Based on the Fusion of Marine Environmental Characteristics

Abstract

The present invention discloses a ship trajectory prediction method based on the fusion of marine environmental characteristics, including: obtaining the ship dynamic sequence data, ship position sequence data, and marine environmental information at the current moment, and inputting them into the ship trajectory prediction model; performing One-hot encoding on the ship dynamic and position sequence data, and performing spatio-temporal grid position association decoding on the marine environmental information; performing Encoder encoding on the One-hot encoded data of the obtained ship dynamic sequence and ship position sequence, as well as the spatio-temporal grid decoded data of the marine environmental information, to obtain the ship time series characteristics and marine environmental time series characteristics; performing cross-modal alignment fusion on these two time series characteristics and then performing Decoder decoding to obtain the ship dynamic information and ship position information at the next moment, and outputting them after One-hot decoding. This method uses marine environmental information to supplement ship trajectory modeling and improves the accuracy of the ship trajectory prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of ship track prediction, and more specifically, to a ship track prediction method based on the fusion of marine environmental characteristics. Background Technique

[0002] With the development of deep learning and big data processing technologies, ship trajectory prediction research has evolved from traditional statistical analysis to prediction based on deep learning models. Ship trajectory prediction is of great significance for maritime route planning and scheduling, maritime traffic safety, marine environmental supervision and protection, maritime military operations, etc. The Automatic Identification System (AIS) provides detailed time-series data on ship navigation trajectories and motion states, including detailed information such as the ship's heading, speed, course, position, cargo type, ship length and width, etc. The accumulated massive ship trajectory information and marine environmental information can be used for modeling and analysis by deep learning models.

[0003] When a ship sails at sea, its motion attitude and trajectory are also deeply affected by the marine environment, including meteorology and hydrology. Moreover, with the progress of marine environmental monitoring technologies, a large amount of marine environmental data has been accumulated for modeling and analysis by deep learning models. Existing deep learning methods consider the ship's position and ship dynamic information, and the accuracy of ship trajectory prediction is limited. How to process the fusion of cross-modal marine environmental information and ship trajectory information and improve the accuracy of the ship trajectory model while comprehensively considering the marine environmental information in which the ship sails is a technical challenge.

[0004] The deficiencies of existing ship trajectory prediction methods mainly include:

[0005] (1) Traditional statistical learning methods based on AIS data and single-modal deep learning methods in ship trajectory prediction problems only consider single-modal AIS data and do not fully utilize the marine environmental information in which the ship sails, including marine meteorology, hydrology, etc. The accuracy of the model is limited, and there are large deviations in the predicted ship trajectories.

[0006] (2) The method based on Transformer with a single encoder-decoder architecture belongs to the seq2seq method, which only supports the input and output of data in the same modality, does not support the input of cross-modal data, and does not support the fusion of cross-modal different-dimensional information.

[0007] (3) Existing deep learning models use the mean squared error and its variant loss functions based on numerical regression modeling to model and predict continuous value outputs, and the predicted outputs deviate greatly from the true values.

[0008] Therefore, how to use marine meteorological and hydrological information to supplement ship trajectory modeling and improve the accuracy of ship trajectory prediction models is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] In view of the above problems, the present invention provides a ship trajectory prediction method based on the fusion of marine environmental characteristics to at least solve some of the technical problems mentioned in the above background technology.

[0010] To achieve the above object, the present invention adopts the following technical solutions:

[0011] The present invention provides a ship trajectory prediction method based on the fusion of marine environmental characteristics, including:

[0012] Obtain the ship dynamic sequence data and ship position sequence data at the current moment;

[0013] Based on the ship position sequence data, obtain marine environmental information;

[0014] Input the obtained ship dynamic sequence data, ship position sequence data, and marine environmental information into a ship trajectory prediction model based on the fusion of marine environmental characteristics;

[0015] In the ship trajectory prediction model based on the fusion of marine environmental characteristics:

[0016] Through the ship sequence information One-hot encoding module, perform One-hot encoding on the ship dynamic sequence data and ship position sequence data respectively to obtain the corresponding ship dynamic sequence One-hot encoded data and ship position sequence One-hot encoded data;

[0017] Through the marine environmental information spatio-temporal grid position association decoding module, based on the ship position sequence data, decode the marine environmental information to obtain marine environmental information spatio-temporal grid decoded data;

[0018] Through N Encoder encoding modules, perform Encoder encoding on the ship dynamic sequence One-hot encoded data and the ship position sequence One-hot encoded data to obtain ship time series features;

[0019] Through the N Encoder encoding modules, perform Encoder encoding on the marine environmental information spatio-temporal grid decoded data to obtain marine environmental time series features;

[0020] Through the cross-modal alignment and fusion module, perform cross-modal alignment and fusion on the ship time series features and the marine environmental time series features to obtain ship time series features based on the fusion of marine environmental characteristics;

[0021] Through N Decoder decoding modules, combined with the ship information output sequence, perform Decoder decoding processing on the ship time-series features based on the fusion of marine environmental characteristics to obtain the predicted ship dynamic information and ship position information at the next moment;

[0022] Through the ship sequence information One-hot decoding module, perform One-hot decoding processing on the predicted ship dynamic information and ship position information at the next moment and then output.

[0023] Further, the obtaining of the ship dynamic sequence data and ship position sequence data at the current moment specifically includes:

[0024] Obtain the AIS ship trajectory data fields at the current moment from the Automatic Identification System for Ships; the AIS ship trajectory data fields include AIS ship dynamic data fields and AIS ship position data fields;

[0025] Perform preprocessing on the AIS ship trajectory data fields;

[0026] Sort the preprocessed AIS ship trajectory data fields according to the MMSI identifier in time sequence, and filter out all data with the same MMSI identifier, which are all the navigation sequence information of the same ship;

[0027] Segment the navigation sequence information with a time interval of more than half an hour for the same MMSI identifier, and each segment of data is the navigation trajectory sequence information of the same ship;

[0028] Sample the segmented data at a preset time interval, and use the interpolation algorithm to complete the data without point positions. Finally, obtain the ship trajectory sequence data; the ship trajectory sequence data includes ship dynamic sequence data and ship position sequence data.

[0029] Further, the preprocessing of the AIS ship trajectory data; specifically includes:

[0030] Clean the missing trajectory data in the AIS ship trajectory data fields;

[0031] Eliminate the incorrect trajectory data in the AIS ship trajectory data fields;

[0032] Remove the duplicate trajectory data in the AIS ship trajectory data fields.

[0033] Further, based on the ship position sequence data, obtain marine environmental information; specifically includes:

[0034] In the target sea area, obtain the ocean absolute geostrophic current data and sea surface wind field data associated with the ship position sequence data;

[0035] According to the preset longitude and latitude intervals, perform spatio-temporal grid processing on the ocean absolute geostrophic current data and the sea surface wind field data respectively to obtain the corresponding absolute geostrophic current grid data and sea surface wind field grid data;

[0036] Perform normalization processing on the absolute geostrophic current grid data and the sea surface wind field grid data, and use the normalized absolute geostrophic current grid data and sea surface wind field grid data as the ocean environment information for ship trajectory prediction.

[0037] Further, the absolute geostrophic current grid data geo_flow is selected by an absolute geostrophic current selection function; expressed as:

[0038] geo_flow u,v = g(lon, lat, time, depth)

[0039] where g(·) represents the absolute geostrophic current selection function; lon represents the longitude of the target sea area; lat represents the latitude of the target sea area; time represents the current moment; depth represents the depth of the ocean absolute geostrophic current data; geo_flow u,v represents the flow velocities of the ocean absolute geostrophic current in the east-west direction u and the north-south direction v.

[0040] Further, the sea surface wind field grid data wind is selected by an ocean wind field selection function; expressed as:

[0041] wind u,v = w(lon, lat, time, height)

[0042] where w(·) represents the ocean wind field selection function; lon represents the longitude of the target sea area; lat represents the latitude of the target sea area; time represents the current moment; height represents the height of the sea surface wind field data; wind u,v represents the wind speeds of the sea surface wind field in the east-west direction u and the north-south direction v.

[0043] Further, the cross-modal alignment and fusion module includes: multiple feed-forward neural network modules based on residual connections, 1×1 convolution modules, and fusion modules; where:

[0044] Each feed-forward neural network module based on residual connections consists of Linear and RELU, and is used to perform further deep learning on the ocean environment time series features to obtain ocean environment deep time series features;

[0045] Through the 1×1 convolution module, an alignment operation is performed on the ocean environment depth time series features and the ship time series features;

[0046] Through the fusion module, the aligned ocean environment depth time series features and ship time series features are added and fused to obtain ship time series features based on the fusion of ocean environment features.

[0047] Furthermore, a cross-attention mechanism is introduced into the Decoder decoding module.

[0048] Furthermore, it further includes: for the ship trajectory prediction model based on the fusion of ocean environment features, calculating the loss through a multi-granularity fusion loss function based on One-hot encoding;

[0049] The multi-granularity fusion loss function L based on One-hot encoding multi_grain is expressed as:

[0050]

[0051] wherein, L CE represents the cross-entropy loss function; y fine represents the fine-grained true value One-hot representation; represents the fine-grained predicted output One-hot representation; y med represents the medium-grained true value One-hot representation; represents the medium-grained predicted output One-hot representation; y coarse represents the coarse-grained true value One-hot representation; represents the coarse-grained predicted output One-hot representation; α represents the weight coefficient of the medium-grained; β represents the weight coefficient of the coarse-grained.

[0052] Through the above technical solutions, compared with the prior art, the present invention discloses a ship track prediction method based on the fusion of ocean environment features, having the following beneficial effects:

[0053] The present invention uses a multi-encoding - single-decoding module architecture to effectively process the ocean environment meteorological and hydrological information features, and uses the ocean environment information to supplement the ship trajectory modeling, thereby improving the accuracy of trajectory prediction.

[0054] The present invention uses a cross-modal alignment and fusion module to fuse and process the ocean environment meteorological and hydrological information, including information such as absolute geostrophic flow and sea surface wind field. Through spatio-temporal grid processing and modal alignment and fusion, the modeling accuracy is improved.

[0055] The present invention designs a multi-granularity fusion loss function based on one-hot encoding, constructs a trajectory prediction model for data quantization classification problems, effectively limits the output value range of the ship trajectory model, enables the trajectory prediction to limit the area and the value range of the limited state, and further improves the accuracy and effectiveness of model prediction.

[0056] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained according to the provided accompanying drawings.

[0058] Figure 1 Schematic diagram of a ship trajectory prediction method based on marine environment feature fusion provided by an embodiment of the present invention.

[0059] Figure 2 Schematic diagram of the process for obtaining ship trajectory sequence data provided by an embodiment of the present invention.

[0060] Figure 3 Schematic diagram of the process for obtaining marine environment information provided by an embodiment of the present invention.

[0061] Figure 4 Schematic diagram of the multi-granularity one-hot encoding representation provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0063] An embodiment of the present invention discloses a ship trajectory prediction method based on marine environment feature fusion. Refer to Figure 1 as shown, which specifically includes the following steps:

[0064] Obtain the ship dynamic sequence data and ship position sequence data at the current moment;

[0065] Based on the ship position sequence data, obtain marine environment information;

[0066] Input the obtained ship dynamic sequence data, ship position sequence data, and marine environment information into the ship trajectory prediction model based on marine environment feature fusion;

[0067] Perform the following operations in the ship trajectory prediction model based on marine environment feature fusion:

[0068] Through the ship sequence information One - hot encoding module, perform One - hot encoding on the ship dynamic sequence data and the ship position sequence data respectively to obtain the corresponding ship dynamic sequence One - hot encoded data and ship position sequence One - hot encoded data;

[0069] Through the marine environment information spatio - temporal grid position association decoding module, based on the ship position sequence data, perform decoding processing on the marine environment information to obtain the marine environment information spatio - temporal grid decoded data;

[0070] Through N Encoder encoding modules, perform Encoder encoding processing on the ship dynamic sequence One - hot encoded data and the ship position sequence One - hot encoded data to obtain the ship time - series features;

[0071] Through N Encoder encoding modules, perform Encoder encoding processing on the marine environment information spatio - temporal grid decoded data to obtain the marine environment time - series features;

[0072] Through the cross - modal alignment and fusion module, perform cross - modal alignment and fusion processing on the ship time - series features and the marine environment time - series features to obtain the ship time - series features based on marine environment feature fusion;

[0073] Through N Decoder decoding modules, combined with the ship information output sequence, perform Decoder decoding processing on the ship time - series features based on marine environment feature fusion to obtain the predicted ship dynamic information and ship position information at the next moment;

[0074] Through the ship sequence information One - hot decoding module, perform One - hot decoding processing on the predicted ship dynamic information and ship position information at the next moment and then output.

[0075] Next, each of the above steps will be described separately.

[0076] Step 1: Obtain the ship dynamic sequence data and ship position sequence data at the current moment, as Figure 2 shown.

[0077] (1) Obtain the AIS ship trajectory data fields at the current moment from the Automatic Identification System (AIS); the AIS ship trajectory data fields include AIS ship dynamic data fields and AIS ship position data fields; specifically:

[0078] The AIS ship dynamic data fields include: Maritime Mobile Service Identity (MMSI) field, speed over ground (SOG), course over ground (COG), heading angle (Heading), and accumulated navigation distance (acc_dist);

[0079] The AIS ship position data fields include: message timestamp (BaseDateTime), latitude coordinate (LAT), longitude coordinate (LON), and other fields;

[0080] (2) Preprocess the AIS ship trajectory data fields; specifically include:

[0081] Clean the missing trajectory data in the AIS ship trajectory data fields; that is: ① Remove the missing data with unique identifiers, including the Maritime Mobile Service Identity (MMSI) field being empty; ② Remove the missing data of dynamic information, including cases where fields such as message timestamp (BaseDateTime), latitude coordinate (LAT), longitude coordinate (LON), speed over ground (SOG), course over ground (COG), and heading angle (Heading) are empty;

[0082] Eliminate the incorrect trajectory data in the AIS ship trajectory data fields; common incorrect trajectory data includes the following situations: ① The MMSI field is 0; ② The speed over ground (SOG) is greater than the reasonable value. The speed of aircraft carriers and guided missile destroyers is generally about 30 knots, the speed of small ships such as missile boats and torpedo boats is 40 - 60 knots, the speed of ordinary ships is generally below 15 knots, and the speed of a few large cargo ships is close to 20 knots. Set the navigation speed to be greater than 70 knots;

[0083] Remove the duplicate trajectory data in the AIS ship trajectory data fields; that is, when the timestamp information and longitude and latitude information of multiple AIS data are the same, delete the duplicate data and keep one item;

[0084] (3) Sort the preprocessed AIS ship trajectory data fields according to the MMSI identifier in time order, and filter out all data with the same MMSI identifier, which is all the navigation sequence information of the same ship;

[0085] Since the ship trajectory data reported by the AIS system mixes all the ship position and status information at the current time point in this area, through the MMSI identification, the data of the same ship can be selected, which is convenient for subsequent processing;

[0086] (4) Segment the navigation sequence information with the same MMSI identification and a time interval exceeding half an hour. Each segment of data is the navigation trajectory sequence information of the same ship;

[0087] (5) Sample the segmented data at a preset time interval, and use the interpolation algorithm to complete the data without position points. Finally, the ship trajectory sequence data is obtained; the ship trajectory sequence data includes ship dynamic sequence data and ship position sequence data.

[0088] Step 2: Based on the ship position sequence data, obtain the marine environment information; as Figure 3 shown.

[0089] For the current ship trajectory prediction model, considering only the ship position information for trajectory prediction, without considering the influence of the ship and the surrounding marine environment on the ship trajectory and ship attitude, the prediction accuracy is limited. The present invention uses marine environment data including ocean absolute geostrophic current data and sea surface wind field data to improve the ship trajectory and attitude prediction accuracy. Specifically:

[0090] (1) In the target sea area, obtain the ocean absolute geostrophic current data and sea surface wind field data associated with the ship position sequence data;

[0091] Since a ship is affected by the geostrophic current when sailing on the sea surface and needs to change its heading and speed to adapt to the geostrophic current, and the maximum draft of modern ships is about 10 m, modeling the ship trajectory based on the geostrophic current data associated with the ship position helps to improve the prediction accuracy of the ship speed and heading;

[0092] Since the sea surface wind speed will affect the sailing attitude and speed of the ship, based on the sea surface wind field data associated with the ship position, the ship speed and heading prediction ability can be improved;

[0093] (2) According to the preset longitude and latitude intervals, perform spatio-temporal grid processing on the ocean absolute geostrophic current data and sea surface wind field data respectively to obtain the corresponding absolute geostrophic current grid data and sea surface wind field grid data; specifically:

[0094] Select the marine environment information related to the analysis of the target sea area. The longitude of the target sea area is lon ∈ [lon min , lon max , and the latitude is lat ∈ [lat min , lat max .

[0095] Grid the marine environmental information according to a preset longitude and latitude interval (for example, an interval of 0.5 degrees of longitude and latitude) to obtain marine environmental grid data Sea_grid lat，lon , expressed as:

[0096] Sea_grid lat，lon ={wind, geo_flow}

[0097] In this formula:

[0098] ① The absolute geostrophic flow grid data geo_flow is selected by the absolute geostrophic flow selection function; expressed as:

[0099] geo_flow u,v =g(lon, lat, time, depth)

[0100] Among them, g(·) represents the absolute geostrophic flow selection function; lon represents the longitude of the target sea surface area; lat represents the latitude of the target sea surface area; time represents the current moment; depth represents the depth of the marine absolute geostrophic flow data (which can be 1m, 5m, 10m, and 20m); geo_flow u,v represents the flow velocity of the marine absolute geostrophic flow in the east-west direction u and the north-south direction v.

[0101] ② The sea surface wind field grid data wind is selected by the ocean wind field selection function; expressed as:

[0102] wind u,v =w(lon, lat, time, height)

[0103] Among them, w(·) represents the ocean wind field selection function; lon represents the longitude of the target sea surface area; lat represents the latitude of the target sea surface area; time represents the current moment; height represents the height of the sea surface wind field data (which can be 10m); wind u,v represents the wind speed of the sea surface wind field in the east-west direction u and the north-south direction v.

[0104] (3) Normalize the absolute geostrophic flow grid data and the sea surface wind field grid data, and the normalized value range is [0, 1]; expressed as:

[0105]

[0106] Among them, e represents the geostrophic flow grid data geo_flow or the ocean wind field grid data wind; i represents the east-west direction u or the north-south direction v; e min represents the minimum value of the geostrophic flow grid data or the ocean wind field grid data; e max represents the maximum value of the geostrophic flow grid data or the ocean wind field grid data.

[0107] The normalized absolute geostrophic current grid data and sea surface wind field grid data are used as ocean environment information for ship trajectory prediction.

[0108] Step 3: Input the obtained ship dynamic sequence data, ship position sequence data, and ocean environment information into the ship trajectory prediction model based on ocean environment feature fusion;

[0109] The ship trajectory prediction model based on ocean environment feature fusion is designed for cross-modal fusion based on the typical Transformer architecture; it specifically includes a ship sequence information One-hot encoding module, an ocean environment information spatio-temporal grid decoding module, N Encoder encoding modules, a cross-modal alignment and fusion module, N Decoder decoding modules, and a ship sequence information One-hot decoding module.

[0110] Step 4: Through the ship sequence information One-hot encoding module, perform One-hot encoding on the ship dynamic sequence data and the ship position sequence data respectively to obtain the corresponding ship dynamic sequence One-hot encoded data and ship position sequence One-hot encoded data; specifically:

[0111] (1) Construct a ship sequence information One-hot encoding table

[0112] First, separately count the maximum and minimum values of the ship dynamic sequence data and the ship position sequence data; determine the value range [x min , x max of each type of data, where x is a specific field in the ship dynamic sequence data or the ship position sequence data, that is, x ∈ {lon, lat, SOG, COG, Heading, acc_dist};

[0113] Second, set the dimension D of the One-hot encoding one-hot Construct an encoding table; this encoding table contains two fields: id and x one-hot ; where id is the serial number of the encoding table, and id ∈ [0, D one-hot - 1]; x one-hot is the encoding value, and there is a one-to-one mapping relationship between the id and x one -hot in the encoding table, supporting reverse lookup.

[0114] (2) Ship sequence information One-hot encoding

[0115] Quantitatively map the corresponding x real value to D one-hot One-hot encodings through the following formula to obtain the x one-hot encoding representation:

[0116]

[0117] Step 5: Through the spatio-temporal grid position association decoding module of marine environmental information, based on the ship position sequence data, decode the marine environmental information to obtain the spatio-temporal grid decoding data of marine environmental information;

[0118] Specifically, based on the ship position sequence data, associate the marine environmental grid point data (including absolute geostrophic current grid point data and sea surface wind field grid point data) corresponding to the spatio-temporal coordinates to achieve the decoding of marine environmental information.

[0119] Step 6: Through N Encoder encoding modules, perform Encoder encoding on the ship dynamic sequence One-hot encoded data and the ship position sequence One-hot encoded data to obtain the ship time series features; and through the N Encoder encoding modules, perform Encoder encoding on the spatio-temporal grid decoding data of marine environmental information to obtain the marine environmental time series features;

[0120] N can take values between 3 and 12, and a comprehensive trade-off optimization of computational complexity, computational time, and computational accuracy is carried out in combination with the training process, and it is generally set to 6.

[0121] Step 7: Through the cross-modal alignment and fusion module, perform cross-modal alignment and fusion on the ship time series features and the marine environmental time series features to obtain the ship time series features fused based on marine environmental features;

[0122] The cross-modal alignment and fusion module includes multiple (for example, M ff pieces) of feed-forward neural network modules based on residual connections, 1×1 convolutional modules, and fusion modules; among them:

[0123] Each feed-forward neural network module based on residual connection consists of Linear and RELU, and is used to perform further deep learning on the marine environmental time series features to obtain the deep marine environmental time series features; the number M of feed-forward neural network modules ff can be 1 to 5 and can be adjusted according to the training effect.

[0124] Through the 1×1 convolutional module, perform alignment operations on the deep marine environmental time series features and the ship time series features to make their dimensions consistent;

[0125] Through the fusion module, add and fuse the aligned deep marine environmental time series features and the ship time series features to obtain the ship time series features fused based on marine environmental features;

[0126] In summary, in this cross-modal alignment and fusion module, the time-series features of the marine environment are used to perturb the encoded data of the ship time-series features, making the trajectory prediction model more comprehensive and accurate in modeling.

[0127] Step 8: Through N Decoder decoding modules, combined with the ship information output sequence, perform Decoder decoding on the ship time-series features fused based on the marine environment features to obtain the predicted ship dynamic information and ship position information at the next moment. At this time, the predicted ship dynamic information and ship position information are in the One-hot encoding mode.

[0128] The cross-attention mechanism is introduced in this Decoder decoding module, which can effectively enhance the feature expression ability. In addition, the number of this Decoder decoding modules should be the same as the number of the above Encoder encoding modules.

[0129] Step 9: Through the ship sequence information One-hot decoding module, perform One-hot decoding on the predicted ship dynamic information and ship position information at the next moment and then output.

[0130] Specifically, when outputting, the model output y one-hot value is inversely mapped to the serial number y id of the One-hot encoding by looking up the encoding table, and the predicted value y^ of the ship trajectory information is calculated through the following formula.

[0131]

[0132] Among them, y max and y min are the maximum value and the minimum value defined in the construction of the code table for this field data respectively, and D one-hot represents the dimension of the code table of this type of data.

[0133] In the embodiment of the present invention, the input ship dynamic sequence data and ship position sequence data are One-hot encoded, and One-hot decoding is performed during output to obtain the predicted ship dynamic information and ship position information; this can limit the input range of the model, and then train the model to output within the limited range, improving the training efficiency and accuracy.

[0134] Step 10: Multi-granularity fusion loss function based on One-hot encoding

[0135] In the prior art, calculating the loss for the output through a single fine-grained One-hot encoding easily leads to misclassification of similar predicted data, and makes the model prone to oscillation and non-convergence during training. To address this problem, the embodiments of the present invention design a loss function with multi-granularity fusion to reduce the problem of difficult convergence in training caused by a single fine-grained One-hot encoding. The multi-granularity One-hot encoding representation is as Figure 4 shown, listing three representations: coarse-grain, medium-grain, and fine-grain. It can be extended to more granularity representations according to training needs. The granularity of the coarse-grain representation is M c times that of the fine-grain, and the granularity of the medium-grain representation is M m times that of the fine-grain. M c and M m can be set to different values according to training, usually set to 5 and 3.

[0136] Based on the three types of encoding representations, the multi-granularity fusion loss function L multi_grain based on the One-hot encoding is expressed as:

[0137]

[0138] where L CE represents the cross-entropy loss function; y fine represents the true value One-hot representation of the fine-grain; represents the predicted output One-hot representation of the fine-grain; y med represents the true value One-hot representation of the medium-grain; represents the predicted output One-hot representation of the medium-grain; y coarse represents the true value One-hot representation of the coarse-grain; represents the predicted output One-hot representation of the coarse-grain; α represents the weight coefficient of the medium-grain; β represents the weight coefficient of the coarse-grain. α and β are usually set to 0.3, and different hyperparameters can be set to train multiple models for optimization.

[0139] Based on this loss function, in the embodiments of the present invention, during the training process of the ship trajectory prediction model based on ocean environment feature fusion, the general transformer architecture model backpropagation training method is adopted. According to the multi-granularity fusion strategy, multiple granularity inputs and outputs are designed, and the corresponding loss function is used for backpropagation training. By setting different hyperparameters, the optimal model is selected for ship trajectory prediction.

[0140] The above steps 1-step 10 are only for convenience of description and do not limit the specific implementation order of each step.

[0141] In summary, in view of the problems existing in the existing ship trajectory prediction, such as insufficient modeling accuracy of single-modal data, difficult cross-modal information fusion processing, and large prediction output deviation of the training method based on numerical regression, the present invention provides a ship trajectory prediction method based on the fusion of marine environmental characteristics, designs a multi-encoding - single-decoding module architecture and a cross-modal alignment and fusion module for the fusion of marine environmental characteristics, solves the problem of fusion modeling of marine environmental characteristics and trajectory characteristics, and improves the modeling accuracy; designs a ship sequence One-hot encoding and multi-granularity fusion loss, constructs a trajectory prediction model as a data quantization classification problem, and effectively improves the effectiveness and accuracy of the model prediction output. The method provided by the present invention can be applied in the fields of maritime route planning and scheduling, maritime traffic safety, marine environmental supervision and protection, maritime military operations, etc., to improve the effectiveness and accuracy of ship trajectory prediction and help improve the efficiency of supervision in related fields.

[0142] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. For the same or similar parts among the various embodiments, reference may be made to each other.

[0143] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A ship track prediction method based on the fusion of marine environmental characteristics, characterized in that, Including: Obtain the ship dynamic sequence data and ship position sequence data at the current moment; Based on the ship position sequence data, obtain the marine environment information; Input the obtained ship dynamic sequence data, ship position sequence data, and marine environment information into the ship trajectory prediction model based on marine environment feature fusion; In the ship trajectory prediction model based on marine environment feature fusion: Through the ship sequence information One-hot encoding module, perform One-hot encoding on the ship dynamic sequence data and the ship position sequence data respectively to obtain the corresponding ship dynamic sequence One-hot encoding data and ship position sequence One-hot encoding data; Through the marine environment information spatio-temporal grid position association decoding module, based on the ship position sequence data, perform decoding processing on the marine environment information to obtain the marine environment information spatio-temporal grid decoding data; Through N Encoder encoding modules, perform Encoder encoding processing on the ship dynamic sequence One-hot encoding data and the ship position sequence One-hot encoding data to obtain ship time series features; Through the N Encoder encoding modules, perform Encoder encoding processing on the marine environment information spatio-temporal grid decoding data to obtain marine environment time series features; Through the cross-modal alignment and fusion module, perform cross-modal alignment and fusion processing on the ship time series features and the marine environment time series features to obtain ship time series features based on marine environment feature fusion; Through N Decoder decoding modules, combined with the ship information output sequence, perform Decoder decoding processing on the ship time series features based on marine environment feature fusion to obtain the predicted ship dynamic information and ship position information at the next moment; Through the ship sequence information One-hot decoding module, perform One-hot decoding processing on the predicted ship dynamic information and ship position information at the next moment and then output.

2. The ship track prediction method based on the fusion of marine environmental characteristics according to claim 1, characterized in that The obtaining of the ship dynamic sequence data and ship position sequence data at the current moment specifically includes: Obtain the AIS ship trajectory data fields at the current moment from the Automatic Identification System for Ships; the AIS ship trajectory data fields include AIS ship dynamic data fields and AIS ship position data fields; Perform preprocessing on the AIS ship trajectory data fields; Sort the preprocessed AIS ship trajectory data fields according to time based on the MMSI identifier, and filter out all data with the same MMSI identifier, which are all the navigation sequence information of the same ship; Segment the navigation sequence information with a time interval of more than half an hour for the same MMSI identifier, and each segment of data is the navigation trajectory sequence information of the same ship; Sample the segmented data at a preset time interval, and use the interpolation algorithm to complete the data without point positions, and finally obtain the ship trajectory sequence data; the ship trajectory sequence data includes ship dynamic sequence data and ship position sequence data.

3. A ship track prediction method based on the fusion of marine environmental characteristics according to claim 2, characterized in that, The performing of the preprocessing on the AIS ship trajectory data; specifically includes: Clean the missing trajectory data in the AIS ship trajectory data field; Eliminate the incorrect trajectory data in the AIS ship trajectory data field; Remove the duplicate trajectory data in the AIS ship trajectory data field.

4. A ship track prediction method based on the fusion of marine environmental characteristics according to claim 1, characterized in that Based on the ship position sequence data, obtain ocean environment information; specifically including: In the target sea area, obtain the ocean absolute geostrophic current data and sea surface wind field data associated with the ship position sequence data; According to the preset longitude and latitude intervals, perform spatio-temporal gridding processing on the ocean absolute geostrophic current data and sea surface wind field data respectively to obtain the corresponding absolute geostrophic current grid data and sea surface wind field grid data; Perform normalization processing on the absolute geostrophic current grid data and the sea surface wind field grid data, and use the normalized absolute geostrophic current grid data and sea surface wind field grid data as the ocean environment information for ship trajectory prediction.

5. A ship track prediction method based on the fusion of marine environmental characteristics according to claim 4, characterized in that, The absolute geostrophic current grid data geo_flow is selected by the absolute geostrophic current selection function; expressed as: geo_flow u,v = g(lon, lat, time, depth) Among them, g(·) represents the absolute geostrophic flow selection function; lon represents the longitude of the target sea area; lat represents the latitude of the target sea area; time represents the current moment; depth represents the depth of the ocean absolute geostrophic flow data; geo_flow u,v represents the flow velocities of the ocean absolute geostrophic flow in the east-west direction u and the north-south direction v.

6. The ship track prediction method based on the fusion of marine environmental characteristics according to claim 4, characterized in that, The sea surface wind field grid data wind is selected by the ocean wind field selection function; expressed as: wind u,v = w(lon, lat, time, height) Among them, w(·) represents the ocean wind field selection function; lon represents the longitude of the target sea area; lat represents the latitude of the target sea area; time represents the current moment; height represents the height of the sea surface wind field data; wind u,v represents the wind speeds of the sea surface wind field in the east-west direction u and the north-south direction v.

7. A ship track prediction method based on the fusion of marine environmental characteristics according to claim 1, characterized in that, The cross-modal alignment and fusion module includes: multiple feed-forward neural network modules based on residual connections, 1×1 convolution modules, and fusion modules; where: Each feed-forward neural network module based on residual connections consists of Linear and RELU, and is used to perform further deep learning on the ocean environment time series features to obtain the ocean environment deep time series features; Through the 1×1 convolution module, perform alignment operations on the ocean environment deep time series features and the ship time series features; Through the fusion module, add and fuse the aligned ocean environment deep time series features and ship time series features to obtain the ship time series features based on ocean environment feature fusion.

8. A ship track prediction method based on the fusion of marine environmental characteristics according to claim 1, characterized in that The cross-attention mechanism is introduced into the Decoder decoding module.

9. A ship trajectory prediction method based on the fusion of marine environmental characteristics according to claim 1, characterized in that, It also includes: For the ship trajectory prediction model based on ocean environment feature fusion, calculate the loss through the multi-granularity fusion loss function based on One-hot encoding; The multi-granularity fusion loss function L based on One-hot encoding multi_grain is expressed as: Among them, L CE represents the cross-entropy loss function; y fine represents the fine-grained true value One-hot representation; represents the fine-grained predicted output One-hot representation; y med represents the medium-grained true value One-hot representation; represents the medium-grained predicted output One-hot representation; y coarse represents the coarse-grained true value One-hot representation; represents the coarse-grained predicted output One-hot representation; α represents the weight coefficient of the medium grain size; β represents the weight coefficient of the coarse grain size.

Images