Autonomous ship navigation induction method and device

Abstract

The application discloses an autonomous ship navigation induction method, which predicts the behavior of the autonomous ship and surrounding ships through a trained trajectory space-time prediction model, thereby inducing the navigation of the autonomous ship. The establishment process of the prediction model comprises the following steps: obtaining the voyage time sequence data of the ships around the autonomous ship, constructing a ship navigation situation awareness graph G θ of the autonomous ship θ , which reflects the navigation situation of the autonomous ship and surrounding ships, wherein the ship navigation posture comprises ship coordinates, a time stamp, a ground speed and a ground heading; through entity embedding of the ship navigation situation awareness graph G θ , a ship trajectory vector is obtained, and the spatial features between ships, the local time features of the ships and the global time features of the ships are captured by machine learning; the obtained features are decoded by deep learning, so as to train the trajectory space-time prediction model.

Description

Technical Field

[0001] This invention relates to the field of maritime navigation safety, specifically to intelligent ship driving technology, and particularly to a method and device for guiding autonomous ship navigation. Background Technology

[0002] Improving maritime navigation safety is a critical issue in sea areas and ports with high traffic density and complex conditions. To enhance navigation safety, it is necessary to provide intelligent navigation systems with trajectory prediction and hazard warning functions. However, due to the complex and ever-changing marine environment, especially in congested port waters, it is not easy for unmanned or autonomous vessels to predict the surrounding navigation environment, particularly other moving targets. Summary of the Invention

[0003] This invention provides a method for autonomous ship navigation guidance based on a spatiotemporal dynamic graph neural network. The technical solution of this method includes:

[0004] It not only considers the impact of the ship's own properties on its sea track, but also incorporates the interaction between the ship and surrounding vessels into the modeling. It also includes a unified framework that can capture both temporal and spatial dependencies, and can effectively prevent gradient vanishing and exploding when capturing long-range sequences.

[0005] In this embodiment of the invention, a spatial graph neural network is first used to learn a ship navigation situational awareness map to capture the spatial dependencies of entities. Within the spatial graph neural network, a knowledge-aware graph attention mechanism is also designed to calculate the influence of adjacent entities and cooperative relationships on the target entity.

[0006] Then, the learned entity embeddings are encoded using a GRU network to obtain the local temporal dependencies of the entities; next, the learned entity embeddings are encoded using a Transformer to obtain the global temporal dependencies of the entities; finally, the learned entity embeddings are decoded using a GRU network to train the model and make reasonable predictions, i.e., navigation guidance.

[0007] Experiments show that the method provided in this invention can effectively solve the limitations of spatial and temporal dependencies in the interaction modeling between ships at sea, and effectively aggregate information from neighboring ships to provide more reasonable and accurate navigation guidance for autonomous ships. Attached Figure Description

[0008] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein:

[0009] Figure 1 This is a flowchart of an autonomous ship navigation guidance method based on a spatiotemporal dynamic graph neural network, which is one embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of a ship navigation situational awareness map, one of the embodiments of the present invention.

[0011] Figure 3 This is a dynamic spatial graph neural network structure diagram, one of the embodiments of the present invention.

[0012] Figure 4 This is a schematic diagram of ship information flow based on a GRU-gated neural unit, which is one embodiment of the present invention.

[0013] Figure 5 This is an overall architecture diagram of a GRU gated neural unit, one of the embodiments of the present invention.

[0014] Figure 6 This is a network structure diagram of the Transformer layer in one embodiment of the present invention. Detailed Implementation

[0015] To facilitate understanding of the technical solutions disclosed herein, the following provides a detailed explanation of the technical terms used in this disclosure.

[0016] A ship navigation situational awareness map is a spatial map generated by the collaborative interaction of multiple ships in a local sea area, which can reflect the mutual influence between multiple ships.

[0017] Spatial graph neural networks are a type of computational method specifically designed for learning spatially dependent neural networks.

[0018] In this disclosure, autonomous vessel can refer to an unmanned ship or a ship with automatic driving capabilities. However, it is obvious that the technical solutions of this disclosure can also be used for intelligent auxiliary judgment of manned ships. Therefore, the autonomous vessel of this disclosure can also be understood as the description of Maritime Autonomous Surface Ships (MASS) published by the International Maritime Organization (IMO).

[0019] In the navigation guidance of unmanned and autonomous ships, planning the ship's route should ensure the ship's safety and avoid dangers. However, ship motion is influenced by various factors, such as wave loads, structural stress, wind forces, and even steering forces. Therefore, ship motion can be considered a complex nonlinear behavior. Predicting ship trajectories using machine learning algorithms requires establishing ship kinematic equations. However, complex environmental factors significantly impact ship motion, leading to complex trajectories and making it difficult to establish accurate mathematical models in real time.

[0020] Deep recurrent neural networks (RNNs) and their successor gated recurrent unit (GRU) networks can be widely used to learn complex schemes from large amounts of sequential information. However, maritime traffic data can be represented as a time series on a route network, with interactions between ships established based on spatial proximity at sea. Therefore, while RNN models can treat trajectory sequences from different ships as independent data streams, they cannot utilize the spatial information within the traffic data.

[0021] With the development of graph neural networks, graph modeling of spatiotemporal data has become a research hotspot. Many methods have achieved impressive results in terms of prediction accuracy. Therefore, graph neural network (GNN) models have been proposed to capture spatial and temporal information, and are often combined with recurrent neural networks to model the spatiotemporal relationships in ship navigation trajectory data.

[0022] While graph structures have achieved significant improvements in incorporating spatiotemporal data into prediction models, their application in navigation guidance for autonomous vessels remains unsatisfactory. This is not only because navigation guidance tasks themselves place high demands on the model's encoding and decoding capabilities, but also because several objective factors currently limit their development. These can be summarized as follows:

[0023] (1) Most current studies on ship navigation guidance consider the influence of the ship's own attributes on its sea track, and often ignore the interaction between the ship and the surrounding ships.

[0024] (2) Current modeling methods for ship navigation data are mostly inadequate in acquiring local and global temporal correlations. For example, they only capture local temporal correlations while ignoring global information. When missing data occurs, they only learn local noise. When capturing long-range sequences, they may also suffer from gradient vanishing or exploding.

[0025] (3) Currently, there is a lack of methods that can simultaneously consider local and global time dependencies and incorporate the spatiality of entities into the modeling process.

[0026] To address the challenges in autonomous ship navigation guidance and improve the accuracy of spatiotemporal network data prediction, this disclosure proposes an autonomous ship navigation guidance method based on spatiotemporal dynamic graph neural networks.

[0027] According to one or more embodiments, an autonomous vessel navigation guidance method predicts the behavior of the autonomous vessel and surrounding vessels using a trained spatiotemporal trajectory prediction model, thereby guiding the autonomous vessel's navigation.

[0028] The process of establishing the prediction model includes,

[0029] By obtaining the time-series route data of vessels surrounding the autonomous vessel (including AIS data), a vessel navigation situational awareness map G is constructed for the autonomous vessel. θ The ship's navigation situational awareness map G θ This reflects the navigation status of the autonomous vessel and surrounding vessels, where vessel navigation status includes vessel coordinates, timestamp, speed over land, and heading over land.

[0030] By analyzing the ship navigation situational awareness map G θ The entity embedding is used to obtain the ship trajectory vector, and machine learning is used to capture the spatial features between ships, the local temporal features of ships, and the global temporal features of ships.

[0031] The features obtained by decoding through deep learning are used to train the spatiotemporal prediction model of the trajectory.

[0032] According to one or more embodiments, an autonomous ship navigation guidance method based on a spatiotemporal dynamic graph neural network includes the following steps:

[0033] Step 1: Constructing the ship's navigation situational awareness map using two geographical nodes e i and e j Relationship strength The ship's navigation situational awareness map (G) is constructed by combining the ship's own attributes (including COG heading to ground and SOG speed to ground). θ ;

[0034] Step 2: Model the spatial dependencies of ship entities and use a spatial graph neural network to capture the spatial relationships between ships;

[0035] Step 3: Model local temporal dependencies and use GRU gated units to continuously capture the local temporal dependencies of entities;

[0036] Step 4: Model global time dependencies and use Transformer to directly capture global time dependencies in the sequence;

[0037] Step 5, navigation guidance, uses a GRU network to decode the learned entity embeddings to train the model for reasonable predictions.

[0038] That is, the ship navigation situational awareness map is learned through a graph neural network to obtain the spatial features between ships, and then a deep learning algorithm is used to further capture the local and global time dependencies of the ships. The essence of capturing or learning local and global time dependencies here is to continue to reconstruct our ship feature vectors so that the reconstructed ship feature vectors contain local and global time dependencies.

[0039] Among them, the ship navigation situational awareness map G in step 1 θ G θ =(ε θ ,R θ ),in, Is it Figure G? θ The set of vertices (N ships) of graph G; R is the graph G. θ The edge set in the array is denoted as The two nodes e i and e j Relationship strength The adjacency matrix is ​​typically used to measure this.

[0040] In step 2, during the process of learning the ship's navigation situational awareness map using a spatial graph neural network,

[0041] First, the multi-hop neighbors of the target entity The embedding learns the target entity e by inward aggregation in the spatial graph. v The higher-order representation of ;

[0042] Then, a weighted representation of the entity is generated through a knowledge-aware graph attention mechanism to calculate the magnitude of the influence of neighboring entities and cooperative relationships on the target entity;

[0043] Finally, semantics and attributes from different neighbors and cooperative relationships are aggregated through an information aggregation operation to refine the embedding of the target entity.

[0044] In step 3, the GRU gated unit (recurrent unit) is a recurrent neural network structure that takes sequence data as input and recursively follows the evolution direction of the sequence. It can learn temporal correlations and handle short-term prediction problems. The local temporal dependency of the entity refers to the similarity of the motion patterns exhibited by the ship in adjacent time intervals θ-1, θ, and θ+1, i.e., the entity embeddings... They are similar.

[0045] In step 4, the Transformer is a deep learning model that uses a self-attention mechanism. This mechanism can assign different weights according to the different importance of each part of the input data, and is designed to process sequential data such as natural language. The global temporal dependency refers to the fact that the sequential data within a complete time period contains clear motion characteristics of the ship. The global temporal dependency is the key to connecting ship motion patterns.

[0046] In step 5, the navigation guidance refers to using a model to predict and generate multiple reasonable ship navigation trajectories. Here, a simple decoding layer is used, with 0 = θ. obs Starting from time +1, predict the future position of the ship. Specifically, embedding data from the Transformer... As input to the decoding layer, it is connected to random Gaussian noise to generate various future predictions.

[0047] The predicted relative position is given by the following formula:

[0048]

[0049] Where z represents the added noise; Indicates the hidden state of the decoder GRU; W d τ represents the decoder weights; F is a linear layer. DC This refers to the GRU network, and Pv refers to a simple label for the ship's position coordinates.

[0050] Finally, the number of time steps θ will be obtained. obs The +1 prediction result is added to the historical sequence for the next prediction to gradually complete the navigation guidance of autonomous ships.

[0051] According to one or more embodiments, such as Figure 1 As shown, a ship navigation guidance method based on a spatiotemporal dynamic graph neural network includes the following steps:

[0052] Step S11: Construct a ship navigation situational awareness map, utilizing two geographical nodes e i and e j Relationship strength Constructing a ship navigation situational awareness map G based on the ship's own attributes. θ ;

[0053] Step S12, for graph G θ The initial embedding of entities in the middle is used to capture the spatial relationships between ships using a spatial graph neural network;

[0054] Step S13: Model local temporal dependencies and use GRU gated units to continuously capture the local temporal dependencies of entities;

[0055] Step S14: Model global time dependencies and use Transformer to directly capture global time dependencies in the sequence;

[0056] Step S15, navigation guidance, uses a GRU network to decode the learned entity embeddings to train the model for reasonable predictions.

[0057] The specific step-by-step process is as follows.

[0058] In step S11, the ship navigation situational awareness map G θ G θ =(ε θ ,R θ ),in, Is it Figure G? θ The set of vertices (N ships); R represents graph G. θ An adjacency matrix is ​​a set of edges that connect different vertices in a given network. in, Representing two nodes e i and e j The strength of the relationship between two ships is usually measured by their geographical proximity.

[0059] Specifically, in this embodiment, the relationship strength The mathematical formula for calculation is as follows:

[0060]

[0061] in, α represents the position coordinates of the entities (vertices) in the two graphs at time θ; α and β represent the longitude and latitude of the ship, respectively. This is the Euclidean distance. Note that if two geographically located nodes are not closely related, then... A well-constructed ship navigation situational awareness map is as follows: Figure 2 As shown.

[0062] In step S12, Figure G θ The initial embedding of entities in the middle is performed, and spatial graph neural networks are used to capture the spatial relationships between ships. Figure 3 Configure its structure diagram. For example... Figure 3 As shown, the specific steps are as follows.

[0063] In this embodiment, the target entity's multi-hop neighbors are... The embedding learns the target entity e by inward aggregation in the spatial graph. v The higher-order representation of . That is,

[0064] Using vectors To map each ship entity and its corresponding attributes (e.g., ground speed (SOG) and ground heading (COG)), where d v This is an appropriate embedding size. Then, at time θ, the embedding of each node v and edge (v,k) can be expressed as:

[0065]

[0066] in, It is e v The initial embedding at the propagation layer where l=0 is followed by refinement through a layer-by-layer propagation process, a v This indicates the ship's basic attribute information. n f r and f p It is a neural network (using a single layer of Conv1D for all f*), which encodes nodes, edges, and pairwise relationships respectively. Target entity e v and its neighboring entity e k The cooperative relationship between them is defined as r v,k Knowledge comes from neighboring entity e k The target entity e propagates to layer l v The definition can be expressed as:

[0067]

[0068] in, It is the attention score for each propagation on the edge, representing the value in r. v,k Under the given relationship, how much information is transmitted from k to v;

[0069] The trainable weight matrix represents the propagation layer l;

[0070] The neighbor e is generated by layer (l-1). k Embedding;

[0071] It is a collaborative relationship r v,k Embedding, d r It refers to the embedding size.

[0072] Based on the above method, this embodiment provides a knowledge-aware graph attention mechanism to calculate the influence of adjacent entities and cooperative relationships on the target entity. Specifically, in the ship navigation guidance problem, if two ships are close to each other, their navigation awareness and trajectories are more likely to influence each other. Various ships at sea possess different attributes, such as speed, heading, and position, which can be used to guide the information flow in the spatial graph. The specific definitions are as follows:

[0073]

[0074] Where f(x) is the ReLU nonlinear activation function; σ is the Sigmoid activation function; the connection operation is represented by ||; W and b are the trainable weight matrix and bias term, respectively, and Z2 is an intermediate variable.

[0075] Furthermore, the Softmax function is used to standardize the weight coefficients of each neighbor.

[0076]

[0077] Step S12 also involves an information aggregation method. The information aggregation operation aggregates semantics and attributes from different neighbors and cooperative relationships to refine the embedding of the target entity. Specifically, given... Update e at propagation layer l v The embedding is expressed as:

[0078]

[0079] Where σ is the activation function of ReLU. Note that... This item retains e v It contains historical embedded information, and as a residual connection, it can effectively alleviate the oversmoothing problem.

[0080] After iterative propagation through layer L, e v It can aggregate semantics and attributes from its higher-order (i.e., L-order) neighbors to enrich its embedding.

[0081] Furthermore, a layer aggregation mechanism is employed to concatenate the representations from each step into a single vector, e. v The final entity embedding is expressed as:

[0082]

[0083] in, || is the output entity embedding of the last propagation layer L; || is the join operation.

[0084] In step S13, local temporal dependencies are modeled, and GRU gating units are used to continuously capture the local temporal dependencies of entities. In this embodiment, the temporal pattern is revealed through changes between snapshots in the temporal dimension.

[0085] The GRU recurrent unit is a recurrent neural network structure that takes sequence data as input and recursively follows the evolution direction of the sequence. It can learn temporal correlations and handle short-term prediction problems. The local temporal dependency of the entity refers to the similarity of the motion patterns exhibited by the ship in adjacent time intervals θ-1, θ, and θ+1, i.e., the entity embeddings... They are similar.

[0086] Furthermore, a series of embeddings generated from graph snapshots, namely the spatial representation of the ship, are incorporated. The data is sorted by timestamp and treated as continuous data, and then time patterns are learned through a sequence decoder. Figure 4This is a schematic diagram of ship information flow based on GRU neural network encoding. Furthermore, the input and output structures of GRU are the same as those of a typical RNN network. Figure 5 This is a schematic diagram of the specific architecture of the GRU neural network. Specifically,

[0087] Based on the input value x of the node at the current time step (θ) (θ=1,2,…,θ obs ) and the hidden layer state h passed down from the previous time step (θ-1) This yields two gating states: reset gating r and update gating z. The formula is as follows:

[0088]

[0089] Where σ is σ(z)=1 / (1+e -z The data can be converted into a 0-1 range to act as a gating signal; W is the weight matrix to be trained. Among them, the update gate z determines how much information from the previous time step and the current time step needs to be transmitted; the reset gate r determines how much historical information needs to be forgotten or discarded, thereby reducing the risk of gradient vanishing by controlling the amount of historical information transmitted.

[0090] After receiving the gating signal, the hidden layer state h at the current time step ′(θ) The updated calculation formula is as follows:

[0091]

[0092] in, This indicates that the previously hidden state information is selectively "forgotten"; This indicates selectively "remembering" information from the current time step. It's important to note that the gating signal z ranges from 0 to 1. The closer the gating signal is to 1, the more information is "remembered"; conversely, the closer it is to 0, the more information is "forgotten." Forgotten information... It incorporates certain dimensional information from the current node's input and adds certain dimensional information from the current node's input to maintain a constant state.

[0093] In step S14, global temporal dependence is modeled by directly capturing the global temporal dependencies in the sequence using a Transformer. The Transformer is a deep learning model employing a self-attention mechanism, which assigns different weights based on the importance of different parts of the input data, aiming to process sequence data such as natural language. The global temporal dependence refers to the presence of clear motion characteristics of the ship within a complete time period of sequence data; global temporal dependence is key to connecting ship motion patterns. Specifically,

[0094] In the problem of navigation guidance for ships at sea, temporal information is not only sequentially related, but also indirectly affects the ship's future navigation planning due to dynamic maritime scene information. Mastering global temporal information is key to improving the accuracy of ship navigation guidance. This paper leverages the advantages of Transformer in resolving globally dependent temporal problems to enhance the model's learning of maritime scene information and ship interaction information, thereby further improving the model's prediction accuracy.

[0095] Similar to the GRU layer, the Transformer layer is also applied to each node separately. Figure 6 This is a schematic diagram of the Transformer network structure. Specifically,

[0096] A Transformer layer primarily consists of a multi-head attention layer, a shared feedforward neural working layer, and a batch normalization layer between them. The multi-head attention mechanism of the Transformer network makes the model pay more attention to the global temporal dependencies of the ship when extracting interaction information features and historical trajectory features of the target ship. Furthermore, the multi-head attention layer is based on a dot-product attention mechanism, as shown in the following formula derivation:

[0097]

[0098] in, It is short for Query, representing the query vector of the target entity; It is short for Key, and is the representation vector of other entities; It is short for Value, representing the specific value of an entity; where |θ obs | represents the length of the entire sequence; d k It represents the dimensions of the query vector and the key vector; d v The dimension representing the value.

[0099] Furthermore, the sequences generated by the GRU layer are first arranged into a matrix, that is, the sequences are arranged into a matrix. The layers are stacked and used as input to the Transformer layer. Secondly, the resulting matrix after stacking is represented as... Where v represents the corresponding node.

[0100] The Query vector, Key vector, and Value corresponding to the v node can be calculated using the following formula:

[0101] Q v =H v W Q ,K v =H v W K V=Hv W V (14)

[0102] in, and These are all projection matrices that need to be learned and are shared by all nodes.

[0103] Furthermore, the attention function can be calculated using the following formula:

[0104]

[0105] In multi-head attention, a total of S sets of projection matrices are used to project H. v Projecting onto different K groups of queries, keys, and values, i.e., by concatenating multiple independent attention modules, a multi-head attention mechanism is obtained. Specifically, the formula for the multi-head attention mechanism is expressed as follows:

[0106] MultiHead(H v = Concat(hrad1,…,hrad) S W O (16)

[0107] Among them, W O It is another linear output projection matrix; head s This represents a specific attention function. Specifically, head... s The mathematical formula is as follows:

[0108]

[0109] in, and It is the projection matrix of the s-th attention head.

[0110] Furthermore, the Transformer layer also involves multiple forward propagation layers, including: a forward propagation layer consisting of a two-layer fully connected layer, where the first layer uses ReLU activation and the second layer does not use an activation function. The Transformer layer also involves two residual connection & regularization layers, consisting of residual connections and regularization components, with the specific calculation formula as follows:

[0111] LayerNorm(X+MultiHead(X)) (18)

[0112] LayerNorm(X+FeedForward(X)) (19)

[0113] Here, X represents the input value of this layer; LayerNorm() represents the function of the residual connection and regularization layer; FeedForward() represents the function of the forward propagation layer. It should be noted that residual connections are typically used to solve the training problems of multi-layer networks; regularization performs a regularization operation on the input of each neuron in each layer, making the model converge faster.

[0114] Based on the above method, the Transformer layer lacks the ability to capture sequential sequences; that is, the attention function ignores the relative positions within the sequence, resulting in similar outcomes regardless of how the sequence structure is shuffled. Here, position embedding is introduced to make the Transformer layer aware of the relative positions within the entire sequence. Specifically, the position embedding is a feature vector of length l, which is aggregated into the original encoding vector H. v In this process, new representations are generated. The specific formula is as follows:

[0115]

[0116] Among them, e pos This represents the positional encoding vector; the specific calculation formula is as follows:

[0117]

[0118] Where pos represents the position of an element in the sequence, and i represents the dimension of the element.

[0119] Furthermore, sequences with positional encoding The result, calculated using formula (21), is ultimately used as the input to the Transformer layer. After encoding, n matrices can be obtained. Its dimensions are the same as the input, used to describe the global temporal dependencies of the ships.

[0120] In step S15, navigation guidance is performed by using a GRU network to decode the learned entity embeddings to train the model for reasonable predictions. From real-world trajectory datasets, it is necessary to learn the ship's motion patterns, i.e., the ship's response to environmental changes. Due to the uncertainty of ship motion, it is desirable for the model to generate multiple reasonable and realistic trajectories for implementing ship navigation guidance. Specifically,

[0121] Using a simple decoding layer, θ = θ from the Transformer layer obs The embedding is taken as input and concatenated with random Gaussian noise to generate various future predictions. The relative positions of the predictions are given by the following formula:

[0122]

[0123] Where z represents the added noise; Indicates the hidden state of the decoder GRU; W d τ represents the weights of the decoder; τ is a linear layer.

[0124] The time step θ will be obtained. obs The +1 prediction result is added to the historical sequence for the next prediction.

[0125] The loss function formula for the entire model is as follows:

[0126]

[0127] Here, δ is a hyperparameter used to control the sensitivity of the squared error loss.

[0128] According to one or more embodiments, an electronic device includes a memory and a processor. The memory stores a configuration program that can run on the processor. When the configuration program is executed by the processor, it can implement the autonomous ship navigation guidance method described in the above embodiments of the present invention. This autonomous ship navigation guidance method predicts the behavior of the autonomous ship and surrounding vessels using a trained spatiotemporal prediction model, thereby guiding the autonomous ship's navigation. The process of establishing the prediction model includes obtaining the route time-series data of the autonomous ship and surrounding vessels (here, AIS data is used; AIS stands for Automatic Identification System) to construct a ship navigation situational awareness map G for the autonomous ship. θ The ship's navigation situational awareness map G θ It reflects the navigation status of the autonomous vessel and surrounding vessels; through the vessel navigation status awareness map G θ The system embeds entities to obtain ship trajectory vectors, and uses deep learning to capture spatial features between ships, local temporal features of ships, and global temporal features of ships. The features obtained are then decoded by deep learning to train a spatiotemporal prediction model for the ship's trajectory. Here, "embedding" refers to a mapping from raw data to a graph. The resulting graph is a set of encoded vectors or matrices. By encoding the raw data, a ship navigation situational awareness map is formed. The purpose of constructing the graph is to learn the spatial features (vectors) of the autonomous ship and surrounding ships. This is achieved using graph neural networks.

[0129] It should be understood that in the embodiments of the present invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0130] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, apparatuses, or units, or they may be electrical, mechanical, or other forms of connection.

[0131] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0132] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for guiding the navigation of an autonomous vessel, characterized in that, By using a trained spatiotemporal prediction model of the trajectory, the behavior of the autonomous vessel and surrounding vessels is predicted, thereby guiding the autonomous vessel's navigation. The process of establishing the prediction model includes... By obtaining time-series data on the routes of vessels surrounding the autonomous vessel, a ship navigation situational awareness map G is constructed for the autonomous vessel. θ The ship's navigation situational awareness map G θ It reflects the navigation status of the autonomous vessel and surrounding vessels. The vessel navigation status here includes vessel coordinates, timestamp, speed above ground, and heading above ground. By using the ship navigation situational awareness map G θ The entity embedding is used to obtain the ship trajectory vector, and machine learning is used to capture the spatial features between ships, the local temporal features of ships, and the global temporal features of ships. Features obtained through deep learning decoding are used to train a spatiotemporal prediction model for flight paths. Ship navigation situational awareness map G θ Among them, G θ =(ε θ ,R θ ), Is it Figure G? θ The set of vertices, where N is the number of ships; R represents graph G θ An adjacency matrix is ​​a set of edges that connect different vertices in a given network. Using the geographical nodes of two ships e i and e j Relationship strength Used to measure the geographical proximity of the two ships. Learning the ship navigation situational awareness map G using a spatial graph neural network model θ Capture the spatial relationship characteristics between ships. The learning process using a spatial graph neural network model includes: Multi-hop neighbors of the target entity The embedding is achieved through the ship navigation situational awareness map G θ Inward aggregation to learn the target entity e v The higher-order representation of ; The impact of adjacent entities and cooperative relationships on the target entity is calculated by using entity weighted representation; Information aggregation operations are used to aggregate semantics and attributes from different neighbors and cooperative relationships to refine the embedding of the target entity.

2. The autonomous ship navigation guidance method according to claim 1, characterized in that, The local temporal dependency features of entities are captured using the GRU gated cell model. The local temporal dependency of the entities refers to the similarity in the motion patterns exhibited by the ships in adjacent time intervals θ-1, θ, and θ+1, i.e., the entities they represent are embedded in each other. They are similar, and 'v' refers to ships.

3. The autonomous vessel navigation guidance method according to claim 1, characterized in that, The Transformer model is used to capture the global temporal dependency features of the ship entity. The global time-dependent feature is a sequence of data within a complete time period that includes ship motion features, representing a series of ship motion patterns.

4. The autonomous ship navigation guidance method according to claim 1, characterized in that, The GRU network model is used to decode entity embeddings to train a spatiotemporal prediction model for ship trajectories, thereby predicting ship trajectories.

5. A navigation guidance device for autonomous vessels, characterized in that, The device includes a memory; and A processor coupled to the memory, the processor being configured to execute a program stored in the memory, which, when executed by the processor, implements the method as described in any one of claims 1 to 4.

6. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 4.

Images