Cross - domain Maritime Search and Rescue Object Trajectory Compensation Method, System, Device and Storage Medium

Abstract

The present invention discloses a cross-domain maritime search and rescue object trajectory compensation method, system, device and storage medium. The method includes: obtaining historical trajectory data including drift trajectory coordinate information with time tags and ocean current trajectory status information; selecting historical trajectory data for a set time period to generate a training set; using the training set to train a spatio-temporal graph neural network model to fit the mapping relationship between the drift trajectory coordinate information at historical moments and the drift trajectory coordinate information at future moments, and the spatio-temporal graph neural network model is provided with an attention mechanism; inputting the drift trajectory coordinate information and ocean current trajectory status information of the search and rescue object at historical moments into the spatio-temporal graph neural network model to obtain the prediction result of the drift trajectory coordinate information of the search and rescue object at future moments. By means of the prior historical trajectory data information, the present invention trains a spatio-temporal graph neural network model based on the attention mechanism, and can predict and compensate for the missing trajectory information of the search and rescue object.

Description

Technical Field

[0001] The present invention relates to the technical field of maritime search and rescue, and particularly to a method, system, device and storage medium for compensating the trajectory of a cross-domain maritime search and rescue object. Background Art

[0002] With the advancement of the global integration process and the continuous expansion of the scale of maritime transportation and ocean development, maritime activities are becoming increasingly frequent, and the frequency of maritime accidents is also increasing. Maritime accidents will bring huge disasters to a country at various levels such as politics, economy, and military. While causing casualties and property losses, they will also have an adverse impact on social development. Therefore, maritime search and rescue work has been increasingly valued by coastal countries. For industries such as the rapidly developing maritime transportation industry, efficient maritime search and rescue operations can provide an irreplaceable guarantee for the safety of personnel and property.

[0003] Currently, when it is difficult to accurately locate the moving position of a person who has fallen into the water and lost their signal, the maritime search and rescue command system mainly relies on satellite positioning to determine the drifting trajectory of the search and rescue object. However, in areas where the satellite positioning system signal is weak or satellite positioning cannot be carried out, the search and rescue work can only rely on the intuitive experience of search and rescue commanders to predict the drifting trajectory of the search and rescue object, resulting in the inability to accurately predict the drifting trajectory of the search and rescue object, seriously affecting the efficiency of search and rescue command and coordination work and missing the best search and rescue opportunity. Summary of the Invention

[0004] To solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a method, system, device and storage medium for compensating the trajectory of a cross-domain maritime search and rescue object.

[0005] The technical solution of the present invention is as follows:

[0006] In a first aspect, a method for compensating the trajectory of a cross-domain maritime search and rescue object is provided, and the method includes:

[0007] Obtaining historical trajectory data including drifting trajectory coordinate information with time tags and ocean current trajectory state information;

[0008] Selecting historical trajectory data for a set time period to generate a training set;

[0009] Training a spatio-temporal graph neural network model using the training set to establish an adjacency relationship between the drifting trajectory and the ocean current trajectory, and fitting a mapping relationship between the drifting trajectory coordinate information at historical moments and the drifting trajectory coordinate information at future moments, wherein the spatio-temporal graph neural network model is provided with an attention mechanism;

[0010] Input the historical moment drift trajectory coordinate information and ocean current trajectory status information of the search and rescue target into the spatio-temporal graph neural network model to obtain the prediction result of the future moment drift trajectory coordinate information of the search and rescue target.

[0011] In some possible implementation manners, the spatio-temporal graph neural network model includes a trajectory encoding sub-module, a graph feature extraction layer, a graph structure generation sub-module, a data processing sub-module, and a trajectory decoding sub-module;

[0012] The trajectory encoding sub-module is used to model the motion patterns of all trajectories included in the historical trajectory data to obtain a trajectory encoding matrix;

[0013] The graph feature extraction layer adopts a node classification layer based on an attention mechanism. The input and output of the graph feature extraction layer are respectively connected to the trajectory encoding sub-module and the graph structure generation sub-module. The graph feature extraction layer is used to fuse and update the trajectory encoding matrix of the trajectory nodes according to the attention weights of each trajectory node on the adjacent trajectory nodes;

[0014] The graph structure generation sub-module is used to model the temporal correlation between different trajectories according to the trajectory encoding matrix of the trajectory nodes output by the graph feature extraction layer;

[0015] The data processing sub-module is used to perform activation processing, connection processing, and noise addition processing on the trajectory encoding matrix output by the trajectory encoding sub-module and the output matrix of the graph structure generation sub-module;

[0016] The trajectory decoding sub-module is used to obtain a trajectory matrix including historical moments and future moments according to the processed trajectory encoding matrix output by the trajectory encoding sub-module and the output matrix of the graph structure generation sub-module.

[0017] In some possible implementation manners, the trajectory encoding sub-module includes at least one layer of bidirectional recurrent neural network. The bidirectional recurrent neural network includes a bidirectional long short-term memory unit, or a bidirectional gated recurrent unit, or a combined network of a bidirectional long short-term memory unit and a bidirectional gated recurrent unit. The trajectory encoding sub-module models the motion patterns of all trajectories included in the historical trajectory data through the bidirectional recurrent neural network.

[0018] In some possible implementation manners, modeling the motion patterns of all trajectories included in the historical trajectory data to obtain a trajectory encoding matrix includes:

[0019] Performing a difference operation on the positions of adjacent moments at each moment by using the following formula one to obtain the relative correlation position at each moment;

[0020]

[0021] Among them, represents the longitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t, represents the longitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t - 1, represents the latitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t, represents the latitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t - 1, represents the relative correlation position of the i-th ocean current trajectory node or drift trajectory node at time t;

[0022] Embed the relative correlation position at each moment into a fixed-length vector using the following formula two;

[0023]

[0024] Among them, represents the correlation matrix obtained by mapping the relative correlation position information of the i-th ocean current trajectory node or drift trajectory node at time t and the embedding weight through the embedding function, φ represents the embedding function, and W ee represents the embedding weight;

[0025] Merge the ocean current trajectory nodes or drift trajectory nodes with a preset length as a fixed-length sequence, and use the fixed-length sequence as the input of a bidirectional recurrent neural network to obtain a trajectory encoding matrix.

[0026] In some possible implementation manners, fuse and update the trajectory encoding matrix of the trajectory nodes according to the attention weights of each trajectory node on the adjacent trajectory nodes, including:

[0027] Perform self-attention weighting processing on the set of trajectory encoding matrices of the input trajectory nodes using the following formula six;

[0028] e ij = A(Wm i ||Wm j ) Formula six

[0029] Among them, e ij represents the trajectory encoding matrix after splicing and weighting of the i-th trajectory node and the j-th trajectory node, A represents a mapping, W represents the weight matrix, m i represents the trajectory encoding matrix of the i-th trajectory node, and m j represents the trajectory encoding matrix of the j-th trajectory node, represents the set of real numbers, F' represents the dimension of the new feature vector, and the trajectory nodes include ocean current trajectory nodes or drift trajectory nodes;

[0030] Based on the masked attention mechanism, the attention is distributed to the set of adjacent trajectory nodes of the trajectory node by using the following formula seven;

[0031]

[0032] where, α ij represents the weight of the trajectory encoding matrix of the i-th trajectory node on all adjacent trajectory nodes, and softmax j represents the activation function used to normalize the data, exp represents the exponential function with the natural constant e as the base, M = [1, 2,..., N], and N represents the number of trajectory nodes in the set of trajectory encoding matrices of the trajectory nodes;

[0033] The new trajectory encoding matrix of the trajectory node is obtained by using the following formula nine;

[0034]

[0035] where, m′ i represents the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer, σ represents the sigmod activation function used to map the output trajectory encoding matrix to between 0 and 1, and M i represents the set of adjacent trajectory nodes of the i-th trajectory node.

[0036] In some possible implementation manners, the graph structure generation sub-module models the temporal correlation between different trajectories according to the trajectory encoding matrix of the trajectory node output by the graph feature extraction layer by using the following formula thirteen;

[0037]

[0038] where, represents the output matrix of the following formula structure generation sub-module at time t + 1, G_GRU represents modeling the temporal correlation between different trajectories, represents the output matrix of the following formula structure generation sub-module at time t, W g represents the temporal weight matrix of different trajectories, and m′ i,t represents the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer at time t.

[0039] In some possible implementation manners, the trajectory decoding sub-module obtains the trajectory matrix including historical moments and future moments by using formula seventeen and formula eighteen according to the trajectory encoding matrix output by the processed trajectory encoding sub-module and the output matrix of the graph structure generation sub-module;

[0040]

[0041]

[0042] Among them, represents T obs At the time of T+1, the merged trajectory encoding matrix, the output matrix of the graph structure generation sub-module, and white noise are combined. D_GRU represents decoding and modeling of the matrix. represents the relative correlation position information of the i-th trajectory node at time T obs The correlation matrix obtained by mapping the embedding weight through the embedding function at the time, and W d represents the weight parameter of the trajectory decoding sub-module. represents predicting the longitude value of the coordinate of the i-th trajectory node at time T obs +1. represents predicting the latitude value of the coordinate of the i-th trajectory node at time T obs +1.

[0043] In a second aspect, a cross-domain maritime search and rescue object trajectory compensation system is provided. The system includes:

[0044] An information acquisition module, configured to acquire historical trajectory data including drift trajectory coordinate information with time tags and ocean current trajectory state information;

[0045] A data generation module, configured to select historical trajectory data in a set time period to generate a training set;

[0046] A spatio-temporal graph neural network model training module, configured to use the training set to train the spatio-temporal graph neural network model to establish the adjacency relationship between the drift trajectory and the ocean current trajectory, and fit the mapping relationship between the drift trajectory coordinate information at historical moments and the drift trajectory coordinate information at future moments;

[0047] A prediction module, configured to input the drift trajectory coordinate information and ocean current trajectory state information of the search and rescue object at historical moments into the spatio-temporal graph neural network model to obtain the prediction result of the drift trajectory coordinate information of the search and rescue object at future moments.

[0048] In a third aspect, a cross-domain maritime search and rescue object trajectory compensation device is provided. The device includes: a memory, a processor, and a communication interface;

[0049] The memory is configured to store instructions;

[0050] The processor is configured to load and execute the instructions in the memory to execute the above method;

[0051] The communication interface is configured to perform communication.

[0052] In a fourth aspect, a computer-readable storage medium is provided. The storage medium stores at least one instruction, and the instruction is loaded and executed by a processor to execute the above method.

[0053] The main advantages of the technical solution of the present invention are as follows:

[0054] The method, system, device and storage medium for compensating the trajectory of cross-domain maritime search and rescue objects of the present invention can predict and compensate for the missing trajectory information by training a spatio-temporal graph neural network model based on the attention mechanism with the help of prior historical trajectory data information, so that the possible position area of the search and rescue object can be quickly and accurately predicted after the signal of the search and rescue object is lost, significantly reducing the search range and saving the search time. Description of the Drawings

[0055] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, form a part of the present invention, and the illustrative embodiments and descriptions thereof are used to explain the present invention without unduly limiting the present invention. In the drawings:

[0056] Figure 1 is a flowchart of the method for compensating the trajectory of cross-domain maritime search and rescue objects according to an embodiment of the present invention;

[0057] Figure 2 is a schematic diagram of the trajectory acquisition method in the method for compensating the trajectory of cross-domain maritime search and rescue objects according to an embodiment of the present invention;

[0058] Figure 3 is a schematic diagram of the structure of the spatio-temporal graph neural network model in the method for compensating the trajectory of cross-domain maritime search and rescue objects according to an embodiment of the present invention;

[0059] Figure 4 is a schematic diagram of the processing principle of the masked attention mechanism in the method for compensating the trajectory of cross-domain maritime search and rescue objects according to an embodiment of the present invention;

[0060] Figure 5 is a schematic diagram of the processing principle of the multi-head self-attention mechanism in the method for compensating the trajectory of cross-domain maritime search and rescue objects according to an embodiment of the present invention;

[0061] Figure 6 is a schematic diagram of the structure of the cross-domain maritime search and rescue object trajectory compensation system according to an embodiment of the present invention;

[0062] Figure 7 is a schematic diagram of the structure of the cross-domain maritime search and rescue object trajectory compensation device according to an embodiment of the present invention. Detailed Embodiments

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and the corresponding drawings. 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 shall fall within the scope of protection of the present invention.

[0064] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the drawings.

[0065] In a first aspect, referring to Figures 1-3 , an embodiment of the present invention provides a method for compensating the trajectory of a cross-domain maritime search and rescue object, the method including the following steps:

[0066] Step S100, obtaining historical trajectory data including drift trajectory coordinate information with time tags and ocean current trajectory status information;

[0067] Step S200, selecting historical trajectory data for a set time period to generate a training set;

[0068] Step S300, training a spatio-temporal graph neural network model using the training set to establish the adjacency relationship between the drift trajectory and the ocean current trajectory, and fitting the mapping relationship between the drift trajectory coordinate information at historical moments and the drift trajectory coordinate information at future moments, wherein the spatio-temporal graph neural network model is provided with an attention mechanism;

[0069] Step S400, inputting the drift trajectory coordinate information and ocean current trajectory status information of the search and rescue object at historical moments into the spatio-temporal graph neural network model to obtain the prediction result of the drift trajectory coordinate information of the search and rescue object at future moments.

[0070] The steps and principles of the method for compensating the trajectory of a cross-domain maritime search and rescue object provided by an embodiment of the present invention will be specifically described below.

[0071] Step S100, obtaining historical trajectory data including drift trajectory coordinate information with time tags and ocean current trajectory status information.

[0072] In an embodiment of the present invention, the historical trajectory data may be real trajectory data or simulated trajectory data.

[0073] Optionally, when the real trajectory data samples are few, the following simulation method can be used to obtain historical trajectory data including drift trajectory coordinate information with time tags and ocean current trajectory status information:

[0074] Select a set sea area, and release a simulated human model with a positioning system in the selected set sea area. The positioning system is, for example, a GPS positioning system, a Beidou satellite positioning system, or a positioning system that can achieve positioning through low-earth orbit satellites. Collect the drift trajectory coordinate information and ocean current trajectory state information of the simulated human model that change over time to obtain corresponding simulated trajectory data, and use the simulated trajectory data as historical trajectory data.

[0075] Among them, the set sea area may include at least one sea area where shipwrecks frequently occur in practice.

[0076] By using the simulated human model to collect simulated trajectory data, it is possible to make up for the deficiencies of real trajectory data, improve the prediction accuracy, and save the prediction cost at the same time.

[0077] Step S200, select historical trajectory data for a set time period to generate a training set.

[0078] Specifically, in an embodiment of the present invention, in order to improve the trajectory compensation prediction accuracy of the search and rescue target, the set time period can be determined according to the time period in which the search and rescue target is located and the sea area state in that time period. The sea area state can, for example, include sea area ocean current information, wind direction and wind speed information, and sea area geographical environment information, etc., to ensure that the sea area state information within the set time period is as similar as possible to the sea area state information in the time period when the search and rescue target is located.

[0079] Furthermore, in order to facilitate subsequent model training and data processing, in an embodiment of the present invention, the historical trajectory data including drift trajectory coordinate information and ocean current trajectory state information is set in matrix form. The historical trajectory data in matrix form can be specifically expressed as:

[0080]

[0081] Among them, represents the m-th trajectory at the n-th moment, t n represents the n-th moment, and represent the position coordinates. The m-th trajectory may be a drift trajectory or an ocean current trajectory.

[0082] In an embodiment of the present invention, the ocean current trajectory is used to constrain the drift trajectory and verify the prediction result of the spatio-temporal graph neural network model.

[0083] Step S300, use the training set to train the spatio-temporal graph neural network model to establish the adjacency relationship between the drift trajectory and the ocean current trajectory, and fit the mapping relationship between the drift trajectory coordinate information at historical moments and the drift trajectory coordinate information at future moments. Among them, the spatio-temporal graph neural network model is provided with an attention mechanism.

[0084] See Figure 3, in an embodiment of the present invention, the spatio-temporal graph neural network model includes: a trajectory encoding sub-module, a graph feature extraction layer, a graph structure generation sub-module, a data processing sub-module, and a trajectory decoding sub-module;

[0085] The trajectory encoding sub-module is used to model the motion patterns of all trajectories included in the historical trajectory data to obtain a trajectory encoding matrix;

[0086] The graph feature extraction layer adopts a node classification layer based on the attention mechanism. The input and output of the graph feature extraction layer are respectively connected to the trajectory encoding sub-module and the graph structure generation sub-module. The graph feature extraction layer is used to fuse and update the trajectory encoding matrix of each trajectory node according to the attention weights of the trajectory node on the adjacent trajectory nodes;

[0087] The graph structure generation sub-module is used to model the temporal correlation between different trajectories according to the trajectory encoding matrix of the trajectory nodes output by the graph feature extraction layer;

[0088] The data processing sub-module is used to perform activation processing, connection processing, and noise addition processing on the trajectory encoding matrix output by the trajectory encoding sub-module and the output matrix of the graph structure generation sub-module;

[0089] The trajectory decoding sub-module is used to obtain a trajectory matrix including historical moments and future moments according to the processed trajectory encoding matrix output by the trajectory encoding sub-module and the output matrix of the graph structure generation sub-module.

[0090] In an embodiment of the present invention, the trajectory encoding sub-module includes at least one layer of bidirectional recurrent neural network. The bidirectional recurrent neural network includes a bidirectional long short-term memory unit, or a bidirectional gated recurrent unit, or a combined network of a bidirectional long short-term memory unit and a bidirectional gated recurrent unit. The trajectory encoding sub-module models the motion patterns of all trajectories included in the historical trajectory data through the bidirectional recurrent neural network.

[0091] Specifically, when the trajectory encoding sub-module models the motion pattern of the trajectory to obtain the trajectory encoding matrix, the following steps S310-S312 are adopted.

[0092] Step S310, perform a difference operation on the positions of adjacent moments at each moment to obtain the relative correlation position at each moment;

[0093]

[0094] Where represents the longitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t, represents the longitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t-1, Denote the coordinate latitude value of the i-th ocean current trajectory node or drifting trajectory node at time t. Denote the coordinate latitude value of the i-th ocean current trajectory node or drifting trajectory node at time t-1. Denote the relative correlation position of the i-th ocean current trajectory node or drifting trajectory node at time t.

[0095] Step S311, embed the relative correlation position at each moment into a fixed-length vector by using the following formula two;

[0096]

[0097] Where, Denote the correlation matrix obtained by mapping the relative correlation position information of the i-th ocean current trajectory node or drifting trajectory node at time t and the embedding weight through the embedding function. φ represents the embedding function, and W ee Denote the embedding weight.

[0098] Step S312, merge the ocean current trajectory nodes or drifting trajectory nodes with a preset length as a fixed-length sequence, use the fixed-length sequence as the input of a bidirectional recurrent neural network, and obtain a trajectory encoding matrix.

[0099] Specifically, obtain the hidden state trajectory encoding matrix by using the following formula three;

[0100]

[0101] Where, Denote the hidden state trajectory encoding matrix of the i-th trajectory node at time t in the trajectory encoding sub-module. M_GRU represents the mapping function of the trajectory encoding sub-module, which is used to model the motion patterns of all trajectories. Denote the hidden state trajectory encoding matrix of the i-th trajectory node at time t-1 in the trajectory encoding sub-module. W m Denote the weight parameter of the trajectory encoding sub-module. The weight parameter W m Can be shared by all trajectories. The trajectory nodes include ocean current trajectory nodes or drifting trajectory nodes.

[0102] Furthermore, in an embodiment of the present invention, the graph feature extraction layer is a node classification technology based on Graph Attention. The graph feature extraction layer adopts a node classification layer based on the attention mechanism. Its basic idea is to fuse and update the feature representation of the node according to the attention weight of each node on its neighboring nodes.

[0103] In an embodiment of the present invention, the input of each graph feature extraction layer is a trajectory encoding matrix of a trajectory node. The set of trajectory encoding matrices of trajectory nodes can be expressed as:

[0104]

[0105] Among them, m represents the set of trajectory encoding matrices of trajectory nodes, and m l represents the trajectory encoding matrix of the l-th trajectory node, N represents the number of trajectory nodes in the set of trajectory encoding matrices of trajectory nodes, represents the set of real numbers, and F represents the dimension of the feature vector.

[0106] The output of each graph feature extraction layer is a new set of trajectory encoding matrices of trajectory nodes obtained by fusing and updating the set of trajectory encoding matrices of trajectory nodes. The new set of trajectory encoding matrices of trajectory nodes can be expressed as:

[0107]

[0108] Among them, m′ represents the new set of trajectory encoding matrices of trajectory nodes, and m′ l represents the new trajectory encoding matrix of the l-th trajectory node, and F′ represents the new dimension of the feature vector.

[0109] Taking the fusion and update of the representation of the i-th trajectory node as an example, the graph feature extraction layer fuses and updates the trajectory encoding matrix of the trajectory node according to the attention weight of each trajectory node on the adjacent trajectory nodes, and performs the following steps S320 - S322.

[0110] Step S320, according to the set of trajectory encoding matrices of the input trajectory nodes, perform self-attention weighted processing using the following formula six;

[0111] e ij = A(Wm i ||Wm j ) Formula six

[0112] Among them, e ij represents the trajectory encoding matrix after splicing and weighting of the i-th trajectory node and the j-th trajectory node, A represents a mapping, W represents the weight matrix, m i represents the trajectory encoding matrix of the i-th trajectory node, m j represents the trajectory encoding matrix of the j-th trajectory node, m i , m j ∈ m, and the i-th trajectory node is one of the adjacent nodes of the j-th trajectory node.

[0113] In formula six, the weight matrix W can be shared by all trajectory nodes. The principle of formula six is to increase the dimension of the trajectory encoding matrices of the i-th trajectory node and the j-th trajectory node by using the linear mapping of a shared weight matrix W, and then use (Wm i ||Wmj ) The operation concatenates the trajectory encoding matrices of the i-th and j-th trajectory nodes after the dimensionality increase transformation, and finally maps the concatenated high-dimensional trajectory encoding matrix to a real number using the mapping A.

[0114] Based on the above, in one embodiment of the present invention, the correlation between the i-th and j-th trajectory nodes is learned through the learnable weight matrix parameter W and the mapping A.

[0115] Step S321, based on the masked attention mechanism, the attention is distributed to the set of neighboring trajectory nodes of the trajectory node using the following formula seven;

[0116]

[0117] where, α ij represents the weight of the trajectory encoding matrix of the i-th trajectory node on all adjacent trajectory nodes, softmax j represents the activation function, which is used to normalize the data, exp represents the exponential function with the natural constant e as the base, and ∑ k∈ M exp(e ik ) represents the sum of all trajectory encoding matrices of the i-th trajectory node, M = [1, 2,..., N], and e ik represents the trajectory encoding matrix after concatenation and weighting of the i-th and k-th trajectory nodes.

[0118] See Figure 4 , in one embodiment of the present invention, the masked attention mechanism is used, and through the softmax activation technique, the attention is only distributed to the set of neighboring trajectory nodes of the trajectory node, which can prevent the loss of structural information.

[0119] Furthermore, if the above mapping A is a single-layer feedforward neural network, formula seven can be expressed as formula eight;

[0120]

[0121] where, represents the transpose of the parameters of the single-layer feedforward neural network, LeakyReLu represents the activation function of the single-layer feedforward neural network, and M i represents the set of neighboring trajectory nodes of the i-th trajectory node.

[0122] Step S322, obtain the new trajectory encoding matrix of the trajectory node output by the graph feature extraction layer using the following formula nine;

[0123]

[0124] Among them, m′ i represents the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer, and σ represents the sigmod activation function, which is used to map the output trajectory encoding matrix to between 0 and 1.

[0125] For the t-th moment, the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer can be expressed as:

[0126]

[0127] Among them, m′ i,t represents the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer at the t-th moment, represents the weight of the trajectory encoding matrix of the i-th trajectory node at the t-th moment on all adjacent trajectory nodes.

[0128] The new trajectory encoding matrix of the trajectory node obtained based on the above method integrates neighborhood information.

[0129] See Figure 5 , furthermore, in an embodiment of the present invention, the output of the graph feature extraction layer can also be weighted and integrated through a multi-head self-attention mechanism to improve the robustness of the trajectory compensation method.

[0130] Specifically, when weighting and integrating the output of the graph feature extraction layer through a multi-head self-attention mechanism, multiple self-attention calculations are used, and the results obtained from each calculation are concatenated or summed.

[0131] Extract K trajectory encoding matrices from the graph feature extraction layer, and the trajectory encoding matrix calculated by the concatenation method can be expressed as:

[0132]

[0133] Among them, || represents concatenation, represents the weight of the trajectory encoding matrix of the i-th trajectory node obtained by the k-th head on all adjacent trajectory nodes, and W k represents the weight matrix of the k-th head, and K represents the number of heads.

[0134] The trajectory encoding matrix of the output of the graph feature extraction layer calculated by the summation method can be expressed as:

[0135]

[0136] Since Therefore

[0137] Further, in an embodiment of the present invention, the graph structure generation sub-module includes at least one layer of recurrent neural network, and the graph structure generation sub-module models the temporal correlation between different trajectories through the recurrent neural network.

[0138] Specifically, in an embodiment of the present invention, the following formula XIII is used to model the temporal correlation between different trajectories;

[0139]

[0140] Among them, represents the output matrix of the formula structure generation sub-module at time t + 1, and G_GRU represents the modeling of the temporal correlation between different trajectories. represents the output matrix of the formula structure generation sub-module at time t, and W g represents the temporal weight matrix of different trajectories.

[0141] Further, in an embodiment of the present invention, the data processing sub-module activates the trajectory encoding matrix output by the trajectory encoding sub-module using the following formula XIV, activates the output matrix of the graph structure generation sub-module using the following formula XV, and connects and adds noise to the trajectory encoding matrix and the output matrix of the graph structure generation sub-module after activation processing using the following formula XVI;

[0142]

[0143]

[0144]

[0145] Among them, represents the activated trajectory encoding matrix of the i-th trajectory node. represents the trajectory encoding matrix of the i-th trajectory at the final time T obs of the historical trajectory. represents the output matrix of the i-th trajectory passing through the graph structure generation sub-module at the final time T obs of the historical trajectory. represents the output matrix of the i-th trajectory in the graph structure generation sub-module after activation. represents T obs At this moment, the combined trajectory encoding matrix, the output matrix of the graph structure generation sub-module, and white noise are represented. z represents a noise sampling vector subject to a normal distribution, and || represents matrix connection.

[0146] Further, in an embodiment of the present invention, the trajectory decoding sub-module generates a trajectory matrix including historical and future moments by using Formula XVII and Formula XVIII based on the trajectory encoding matrix output by the processed trajectory encoding sub-module and the output matrix of the graph structure generation sub-module;

[0147]

[0148]

[0149] Among them, represents the merged trajectory encoding matrix, the output matrix of the graph structure generation sub-module, and white noise at time T obs +1. D_GRU represents decoding and modeling of the matrix. represents the relative correlation position information of the i-th trajectory node at time T obs and the correlation matrix obtained by mapping the embedding weight through the embedding function. W d represents the weight parameter of the trajectory decoding sub-module. represents the predicted longitude value of the coordinate of the i-th trajectory node at time T obs +1. represents the predicted latitude value of the coordinate of the i-th trajectory node at time T obs +1.

[0150] In a second aspect, referring to Figure 6 , an embodiment of the present invention further provides a cross-domain maritime search and rescue object trajectory compensation system 500. The system 500 includes:

[0151] An information acquisition module 501, configured to acquire historical trajectory data including drift trajectory coordinate information with time tags and ocean current trajectory state information;

[0152] A data generation module 502, configured to select historical trajectory data for a set time period to generate a training set;

[0153] A spatio-temporal graph neural network model training module 503, configured to use the training set to train the spatio-temporal graph neural network model to establish the adjacency relationship between the drift trajectory and the ocean current trajectory, and fit the mapping relationship between the drift trajectory coordinate information at historical moments and the drift trajectory coordinate information at future moments;

[0154] A prediction module 504, configured to input the drift trajectory coordinate information and ocean current trajectory state information of the search and rescue object at historical moments into the spatio-temporal graph neural network model to obtain the prediction result of the drift trajectory coordinate information of the search and rescue object at future moments.

[0155] The cross-domain maritime search and rescue object trajectory compensation system provided by an embodiment of the present invention can use the above cross-domain maritime search and rescue object trajectory compensation method to realize the prediction of the drift trajectory coordinate information of the search and rescue object at a future moment and the compensation of the drift trajectory coordinate information at a specific moment.

[0156] In a third aspect, referring to Figure 7 , an embodiment of the present invention further provides a cross-domain maritime search and rescue object trajectory compensation device 600, which includes a memory 601, a processor 602, and a communication interface 603;

[0157] The memory 601 is used to store instructions;

[0158] The processor 602 is used to load and execute the instructions in the memory 601 to execute the above cross-domain maritime search and rescue object trajectory compensation method;

[0159] The communication interface 603 is used for communication.

[0160] The memory 601, the processor 602, and the communication interface 603 are interconnected through a bus 604. The bus 604 can be a peripheral component interconnect standard (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 604 can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, Figure 7 only a thick line is shown in

[0161] but it does not mean that there is only one bus or one type of bus.

[0162] The above memory 601 can be a random access memory (RAM), a flash memory, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a register, a hard disk, a removable hard disk, a CD-ROM, or any other form of storage medium known to those skilled in the art.

[0163] The above communication interface 603 can be, for example, an interface card, etc., and can be an Ethernet interface or an Asynchronous Transfer Mode (ATM) interface.

[0164] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium storing at least one instruction that is loaded and executed by a processor to perform the above cross-domain maritime search and rescue object trajectory compensation method.

[0165] The cross-domain maritime search and rescue object trajectory compensation method, system, device, and storage medium provided by an embodiment of the present invention can predict and compensate for missing trajectory information by training a spatio-temporal graph neural network model based on an attention mechanism with the aid of prior historical trajectory data information, so as to quickly and accurately predict the possible location area of the search and rescue object after the signal of the search and rescue object is lost, significantly narrowing the search scope and saving the search time.

[0166] It should be noted that in this article, the terms "including", "comprising", or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device including a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or device.

[0167] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated herein.

[0168] In several embodiments provided in the present application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation. For example, multiple modules can be combined, or some features can be ignored or not executed.

[0169] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above embodiments can be completed by hardware, or can be completed by a program instructing relevant hardware. The program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a disk, or an optical disc, etc.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for compensating the trajectory of cross - domain maritime search and rescue objects, characterized in that, Including: Obtaining historical trajectory data including drifting trajectory coordinate information with time tags and ocean current trajectory status information; Selecting historical trajectory data within a set time period to generate a training set; Training a spatio-temporal graph neural network model using the training set to establish the adjacency relationship between the drifting trajectory and the ocean current trajectory, and fitting the mapping relationship between the drifting trajectory coordinate information at historical moments and the drifting trajectory coordinate information at future moments, wherein the spatio-temporal graph neural network model is provided with an attention mechanism; Inputting the drifting trajectory coordinate information and ocean current trajectory status information of the search and rescue object at historical moments into the spatio-temporal graph neural network model to obtain the prediction result of the drifting trajectory coordinate information of the search and rescue object at future moments.

2. The method for compensating the trajectory of cross - domain maritime search and rescue objects according to claim 1, characterized in that, The spatio-temporal graph neural network model includes a trajectory encoding sub-module, a graph feature extraction layer, a graph structure generation sub-module, a data processing sub-module, and a trajectory decoding sub-module; The trajectory encoding sub-module is used to model the motion patterns of all trajectories included in the historical trajectory data to obtain a trajectory encoding matrix; The graph feature extraction layer adopts a node classification layer based on the attention mechanism. The input and output of the graph feature extraction layer are respectively connected to the trajectory encoding sub-module and the graph structure generation sub-module. The graph feature extraction layer is used to fuse and update the trajectory encoding matrix of the trajectory nodes according to the attention weights of each trajectory node on the adjacent trajectory nodes; The graph structure generation sub-module is used to model the temporal correlation between different trajectories according to the trajectory encoding matrix of the trajectory nodes output by the graph feature extraction layer; The data processing sub-module is used to perform activation processing, connection processing, and noise addition processing on the trajectory encoding matrix output by the trajectory encoding sub-module and the output matrix of the graph structure generation sub-module; The trajectory decoding sub-module is used to obtain a trajectory matrix including historical moments and future moments according to the processed trajectory encoding matrix output by the trajectory encoding sub-module and the output matrix of the graph structure generation sub-module.

3. The method for compensating the trajectory of cross - domain maritime search and rescue objects according to claim 2, characterized in that, The trajectory encoding sub-module includes at least one layer of bidirectional recurrent neural network. The bidirectional recurrent neural network includes a bidirectional long short-term memory unit, or a bidirectional gated recurrent unit, or a combined network of a bidirectional long short-term memory unit and a bidirectional gated recurrent unit. The trajectory encoding sub-module models the motion patterns of all trajectories included in the historical trajectory data through the bidirectional recurrent neural network.

4. The method for compensating the trajectory of cross - domain maritime search and rescue objects according to claim 3, characterized in that, Modeling the motion patterns of all trajectories included in the historical trajectory data to obtain a trajectory encoding matrix, including: Performing a difference operation on the positions of adjacent moments at each moment using the following formula one to obtain the relative correlation position at each moment; Among them, represents the longitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t, represents the longitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t-1, represents the latitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t, represents the latitude value of the coordinate of the i-th ocean current trajectory node or drift trajectory node at time t-1, represents the relative correlation position of the i-th ocean current trajectory node or drift trajectory node at time t; Embedding the relative correlation position at each moment into a fixed-length vector using the following formula two; Among them, represents the correlation matrix obtained by mapping the relative correlation position information and embedding weight of the i-th ocean current trajectory node or drift trajectory node at time t through an embedding function, φ represents the embedding function, and W ee represents the embedding weight; Merging ocean current trajectory nodes or drifting trajectory nodes of a preset length as a fixed-length sequence, using the fixed-length sequence as the input of the bidirectional recurrent neural network to obtain a trajectory encoding matrix.

5. The method for compensating the trajectory of cross - domain maritime search and rescue objects according to claim 4, characterized in that, Fusing and updating the trajectory encoding matrix of the trajectory nodes according to the attention weights of each trajectory node on the adjacent trajectory nodes, including: Performing self-attention weighting processing using the following formula six according to the set of trajectory encoding matrices of the input trajectory nodes; e ij = A(Wm i ||Wm j ) Equation 6 Among them, e ij represents the trajectory encoding matrix after splicing and weighting the i-th trajectory node and the j-th trajectory node. A represents a mapping, W represents the weight matrix, and m i represents the trajectory encoding matrix of the i-th trajectory node, and m j represents the trajectory encoding matrix of the j-th trajectory node. represents the set of real numbers, F′ represents the new feature vector dimension, and the trajectory nodes include ocean current trajectory nodes or drift trajectory nodes; Based on the masked attention mechanism, the attention is distributed to the set of adjacent trajectory nodes of the trajectory node by using the following formula seven; where α ij represents the weight of the trajectory encoding matrix of the i-th trajectory node on all adjacent trajectory nodes, and softmax j represents the activation function used to normalize the data, exp represents the exponential function with the natural constant e as the base, M = [1, 2,..., N], and N represents the number of trajectory nodes in the trajectory encoding matrix set of the trajectory nodes; The new trajectory encoding matrix of the trajectory node is obtained by using the following formula nine; Among them, m′ i represents the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer, and σ represents the sigmod activation function, which is used to map the output trajectory encoding matrix to the range of 0 to 1. M i represents the set of adjacent trajectory nodes of the i-th trajectory node.

6. The method for compensating the trajectory of cross - domain maritime search and rescue objects according to claim 5, characterized in that, The graph structure generation sub-module models the temporal correlation between different trajectories by using the following formula thirteen according to the trajectory encoding matrix of the trajectory node output by the graph feature extraction layer; Among them, represents the output matrix of the following formula structure generation sub-module at time t+1. G_GRU represents modeling the temporal correlation between different trajectories. represents the output matrix of the following formula structure generation sub-module at time t. W g represents the temporal weight matrix of different trajectories. m′ i,t represents the new trajectory encoding matrix of the i-th trajectory node output by the graph feature extraction layer at time t.

7. The cross-domain maritime search and rescue object trajectory compensation method according to claim 6, characterized in that, The trajectory decoding sub-module obtains the trajectory matrix including historical moments and future moments by using formula seventeen and formula eighteen according to the trajectory encoding matrix output by the processed trajectory encoding sub-module and the output matrix of the graph structure generation sub-module; Among them, represents T obs At the moment of T + 1, the merged trajectory encoding matrix, the output matrix of the graph structure generation sub-module, and white noise are combined. D_GRU represents decoding and modeling of the matrix. represents the correlation matrix obtained by mapping the relative correlation position information and embedding weight of the i-th trajectory node at the moment of T obs through the embedding function, and W d represents the weight parameter of the trajectory decoding sub-module. represents the predicted longitude value of the coordinate of the i-th trajectory node at the moment of T obs + 1. represents the predicted latitude value of the coordinate of the i-th trajectory node at the moment of T obs + 1.

8. A cross-domain maritime search and rescue object trajectory compensation system, characterized in that, The system includes: An information acquisition module, configured to acquire historical trajectory data including drifting trajectory coordinate information with time tags and ocean current trajectory status information; A data generation module, configured to select historical trajectory data in a set time period to generate a training set; A spatio-temporal graph neural network model training module, configured to train the spatio-temporal graph neural network model by using the training set to establish the adjacency relationship between the drifting trajectory and the ocean current trajectory, and fit the mapping relationship between the drifting trajectory coordinate information at historical moments and the drifting trajectory coordinate information at future moments; A prediction module, configured to input the drifting trajectory coordinate information and ocean current trajectory status information of the search and rescue object at historical moments into the spatio-temporal graph neural network model to obtain the prediction result of the drifting trajectory coordinate information of the search and rescue object at future moments.

9. A cross-domain maritime search and rescue object trajectory compensation device, the device comprising: A memory, a processor and a communication interface; The memory is used to store instructions; The processor is configured to load and execute the instructions in the memory to execute the method according to any one of claims 1 to 7; The communication interface is used for communication.

10. A computer-readable storage medium, characterized in that, The storage medium stores at least one instruction, and the instruction is loaded and executed by the processor to execute the method according to any one of claims 1 to 7.

Images