A target submarine path prediction and tracking method for unmanned boats

Abstract

The present invention provides a target submarine path prediction and tracking method for an unmanned boat, and belongs to the field of unmanned boat dynamic tracking technology. The trajectory data is collected in various sea areas by the sonar radar array sensor of the underwater unmanned boat to ensure the comprehensiveness and robustness of the data. The core of the present invention is a target submarine trajectory prediction model that combines a multi-head attention mechanism and an LSTM network, which can accurately capture trajectory characteristics and predict the future movement trajectory of the target submarine. In particular, an environmental perception module is introduced to consider the impact of obstacles around the target submarine on the target submarine trajectory prediction, thereby improving the prediction accuracy. During the training process, a model interaction strategy is used to optimize the model performance and robustness. Finally, the system was successfully deployed on the unmanned boat, collecting data in real time and predicting the target trajectory, guiding the unmanned boat to accurately track, and providing strong technical support for fields such as ocean exploration, resource development, and security maintenance.

Description

Technical Field

[0001] The invention belongs to the technical field of unmanned boat dynamic tracking, and in particular relates to a target submarine path prediction and tracking method for an unmanned boat. Background Art

[0002] The importance and background significance of unmanned boats in predicting and tracking the trajectory of target submarines in the ocean are extremely significant. With the intensification of ocean competition, submarines, as invisible forces in underwater operations, have brought huge challenges to maritime security due to the concealment and uncertainty of their activity trajectories. As an intelligent surface platform, unmanned boats have become an important tool to solve this problem with their advanced perception technology, powerful data processing capabilities and precise tracking algorithms. The trajectory prediction of target submarines by unmanned boats can reveal the potential action path of submarines in advance and provide valuable early warning information for the maritime defense system. In addition, unmanned boats have also shown great potential in scientific research and environmental monitoring. By tracking the trajectory of submarines, unmanned boats can deeply understand the characteristics of the marine environment, provide valuable data for marine scientific research, and promote human cognition and protection of the ocean. The trajectory prediction and tracking of target submarines by unmanned boats is not only a key measure to enhance maritime security and defense capabilities, but also an important force to promote marine scientific research, resource development and environmental protection. Some existing target submarine trajectory prediction methods are as follows:

[0003] Prediction method based on physical model: The method of trajectory prediction using physical model is usually based on the principles of Newtonian mechanics or fluid mechanics. It predicts the future trajectory of the target submarine by considering its dynamic characteristics, environmental factors (such as wind, water current) and possible interaction forces. The advantage of this method is that it can more accurately reflect the real situation of the physical world, especially suitable for scenarios where the physical laws are known and the environmental parameters can be measured. However, its disadvantages are also obvious, that is, detailed environmental data and accurate physical models are required, and the computational complexity is high. For target submarines with complex and changeable marine environments or unknown physical characteristics, the prediction accuracy is relatively limited.

[0004] Prediction method based on classic machine learning model: Using classic machine learning model for trajectory prediction mainly collects historical trajectory data as training set, and uses algorithms such as decision tree, random forest, support vector machine, etc. to learn the mapping relationship between feature patterns in trajectory data and future positions. This method can capture the complexity and nonlinear characteristics of trajectory data, thereby improving the accuracy of prediction, but it depends on a large amount of high-quality training data and reasonable feature engineering. In addition, the existing machine models have poor feature extraction capabilities. In real usage scenarios, the error of trajectory prediction is relatively large, and the success rate of trajectory prediction cannot be guaranteed to be at a high level.

[0005] Therefore, the prediction method based on physical models relies too much on accurate mathematical and physical models, which leads to excessively high costs for mathematical and physical modeling of the real environment when using this method for trajectory prediction, which also means that the generalization ability of this method is relatively weak. The trajectory prediction method based on classical machine learning models has limited efficiency and ability to extract environmental features, and cannot guarantee the accuracy of trajectory prediction. Summary of the invention

[0006] In view of the above problems, the present invention provides a target submarine path prediction and tracking method for an unmanned boat, comprising the following steps:

[0007] S1, obtain the real-time three-dimensional position coordinate information of the target submarine through the sonar sensor, and obtain the time series data of the target submarine position trajectory in the time period a before this moment, and correspondingly obtain the three-dimensional coordinate time series data of the environmental obstacles in the time period a;

[0008] S2, input the data obtained in S1 into the trained target submarine trajectory prediction model, and output the target submarine position trajectory sequence prediction data within the b time period after this moment;

[0009] The target submarine trajectory prediction model includes an environment perception module and a target submarine trajectory prediction backbone network; the environment perception module analyzes and processes the three-dimensional coordinate time series data of environmental obstacles by constructing a graph neural network to obtain the target submarine environment interaction characteristics; the target submarine trajectory prediction backbone network includes a self-attention mechanism, a multi-head attention mechanism, an LSTM feature extraction module and a trajectory output module; the self-attention mechanism enables the model to automatically fuse the characteristics of the target submarine position trajectory time series data and the characteristics of the three-dimensional coordinate time series data of the environmental obstacles to improve the receptive field of the features; the multi-head attention mechanism performs feature extraction at different levels on the fusion features output by the self-attention mechanism, and then performs weighted summation to obtain multi-level fusion features; the bidirectional LSTM feature extraction unit of the LSTM feature extraction module simultaneously considers the forward and backward information of the input features, comprehensively captures the time series characteristics in the target submarine motion trajectory, and obtains bidirectional time series extraction features; then the trajectory output module outputs the prediction result;

[0010] S3, based on the output target submarine position trajectory sequence prediction data, dynamically adjusts the moving direction and speed of its own unmanned boat through the control system in real time, and always predicts and tracks the target submarine trajectory predicted by the trajectory prediction model in advance.

[0011] Preferably, the target submarine trajectory prediction model is trained based on a collected training data set, and the training data set is constructed by:

[0012] The target sea area where the unmanned boat performs trajectory prediction and tracking tasks is used as the target area, and other sea areas used for data collection are used as the source area, and the sonar radar sensor array carried by the unmanned boat is used to collect data;

[0013] Acquisition time step The target submarine position data at each time point is combined to obtain the complete submarine historical trajectory sequence data ;

[0014]

[0015] in, Indicates the collected Submarine historical trajectory sequence data, Indicates The specific position data of the submarine at that moment, specifically:

[0016]

[0017] Indicates The submarine in the collected data is at time The corresponding specific three-dimensional coordinate information, where the coordinate system is established with the unmanned boat as the origin of the coordinate system, and the north, east, and down directions are established respectively. x， y, z axis;

[0018] Collect the three-dimensional coordinate time series data of the target submarine environmental obstacles ;

[0019]

[0020]

[0021] in, Indicates The environmental obstacle position data The position sequence data of obstacles, and , Indicates the total number of obstacles around the target submarine; Indicates The environmental obstacle position data Obstacles at time The specific three-dimensional position coordinates corresponding to Item Target Submarine Environmental Obstacle Position Sequence Data for:

[0022]

[0023] Collect the time step of the target submarine The three-dimensional position coordinate sequence data obtained by combining them is used as the target submarine prediction trajectory sequence data :

[0024]

[0025] The collected historical trajectory sequence data of each target submarine Combine to get the target submarine historical trajectory input data set , for each collected target submarine environmental obstacle position sequence data Combine to get the target submarine environmental obstacle location data set , for each predicted trajectory sequence data corresponding to the collected target submarine Combine to get the target submarine predicted trajectory output data set ;

[0026] Among them, the target submarine historical trajectory dataset, the target submarine environmental obstacle location dataset, and the target submarine predicted trajectory output dataset collected in the target domain are defined as , and The target submarine historical trajectory dataset, target submarine environmental obstacle location dataset, and target submarine predicted trajectory output dataset collected in the source domain are defined as , and .

[0027] Preferably, the input end of the target submarine trajectory prediction backbone network includes a feature vector encoding layer, which collects the submarine historical movement trajectory data and encodes each sequence data Vector encoding is performed through the fully connected layer to obtain the encoded feature data .

[0028] Preferably, the processing process of the environment perception module is:

[0029] S11, based on graph neural network to extract the interaction features between the target submarine and the surrounding obstacles, construct a graph :

[0030]

[0031] in, Representatives select the total number of targets including submarines and obstacles +1 target as the set of nodes in the graph, for each node , , select As target submarine and obstacle node features respectively ;

[0032] represents the edge connecting each obstacle and the target submarine, E is the total number of obstacles, so there are a total of E edges connecting the target submarine, and the relative distance between the target submarine and the obstacle As the feature of this edge;

[0033]

[0034] in, Indicates the three-dimensional coordinate position information of the target submarine, Represents the three-dimensional coordinate position information of the obstacle;

[0035] S12, calculate the obstacle attention coefficient; design the obstacle weight calculation function, and add the corresponding obstacle weight to each edge according to this function :

[0036]

[0037] in, represents the normalization operation function;

[0038] S13, extract features from the graph neural network layer; use the Sigmoid activation function to activate each node feature to obtain the activation features of each node , based on the activation feature , and use two sequential fully connected layers and combine the Relu activation function to activate the features of each edge to obtain the features of each edge :

[0039]

[0040]

[0041] in, express Activation function, express Activation function, Represents the fully connected layer operation function;

[0042] Then, the features of each edge After addition, it is normalized by Softmax and then combined with the target submarine node feature After splicing, we get the output of the neural network layer of this graph ;

[0043]

[0044] The operation function of the graph neural network layer is defined as ;

[0045] S14, target submarine environment interaction feature calculation; design three interconnected graph neural network layers, and finally obtain the target submarine interaction features through a fully connected layer , and send the obtained obstacles into the target submarine trajectory prediction backbone network to perform target submarine trajectory prediction to enhance the trajectory prediction accuracy:

[0046]

[0047] in, Represents the fully connected layer operation function.

[0048] Preferably, the multi-head attention mechanism module uses eight attention heads to realize the self-attention mechanism fusion feature Extract features at different levels, then perform weighted processing on the eight levels of features to obtain multi-level fusion features ; The specific process is as follows:

[0049] Feature matrix calculation: Fusion features Through matrix transformation, they are mapped to On three different feature matrices:

[0050]

[0051] in, Respectively represent matrices The matrix mapping function, Indicates Group The mapping parameters used when mapping the feature matrix, ; Represents the fusion features The mapped Group Feature matrix;

[0052] Attention coefficient calculation: Based on the obtained Two feature matrices, calculate their attention coefficients :

[0053]

[0054] in, express Activation function, represents the scaling factor, Indicates The attention coefficient corresponding to the group attention head;

[0055] Multi-head attention mechanism output feature vector calculation: The attention coefficient corresponding to the group attention head and Group Multiplying the feature matrix can get Features output by the group attention head , the output features of the 8 attention heads After weighted summation, we get the multi-level fusion features after the multi-head attention mechanism. .

[0056] Preferably, the LSTM feature extraction module comprises a forward LSTM network feature extraction unit and a reverse LSTM network extraction unit;

[0057] In the forward LSTM network feature extraction unit, the multi-level fusion features are The corresponding forward time features are arranged in order from front to back in the time dimension , and input into the forward LSTM network feature extraction unit:

[0058]

[0059] in, represents the forward time feature after processing, Represents the operation function of the forward LSTM network feature extraction unit, and the forward LSTM network feature extraction unit includes 4 LSTM network layers connected in sequence;

[0060] In the reverse LSTM network feature extraction unit, the multi-level fusion features The corresponding reverse time features are arranged in order from back to front in the time dimension , and input into the reverse LSTM network feature extraction unit:

[0061]

[0062] in, represents the reverse time feature after processing, Represents the reverse LSTM network feature extraction unit operation function, and the reverse LSTM network feature extraction unit includes 4 LSTM network layers connected in sequence;

[0063] Then the processed forward time features And the reverse time characteristics after processing Merge features to get bidirectional temporal features ;

[0064]

[0065] in, Represents the feature merging operation function.

[0066] Preferably, the trajectory output module sends the obtained intermediate features to the ARMA module for prediction to obtain a preliminary trajectory prediction result. ; Then the preliminary trajectory prediction result is sent to two fully connected layers for further adjustment to obtain the final predicted trajectory prediction result ;

[0067]

[0068]

[0069] in, represent Operation function, Represents the fully connected layer operation function.

[0070] Preferably, the target submarine trajectory prediction model defines a loss function during the training process: for:

[0071]

[0072]

[0073] in, and Respectively represent The data of the group In each time step, the real three-dimensional coordinate data of the target submarine and the three-dimensional coordinate data of the target submarine predicted by the model are ; To calculate the mean absolute error function, To calculate the mean square error function;

[0074] At the same time, two identical target submarine trajectory prediction models are designed for training, one of which is used as the main model After training, it is used as the target submarine trajectory prediction model for actual use, and the other model is used as the secondary model , helping the main model update parameters during training and further improving model performance.

[0075] Preferably, the training process of the target submarine trajectory prediction model is:

[0076] S21, using the dataset collected in the target domain For the main model Perform pre-training and use the loss function The calculated loss value Combine the gradient descent method to optimize the model parameters and iterate the training Second-rate;

[0077] S22, main model Iterative training After that, the dataset is obtained by collecting data from the source domain. The data in the sub-model Train and use the loss function Calculate the training loss Combine the gradient descent method to optimize the model parameters;

[0078] S23, this uses the main model and the secondary model for mixed iterative training, and continues to use the data For the main model Train and optimize, and define the main model network parameters obtained in the first round of hybrid iterative training as ; Using the dataset Pair Model Train and optimize, and define the sub-model network parameters obtained in the first round of iterative hybrid training as ;

[0079] At the same time, based on the main model network parameters obtained And the sub-model network parameters The main model network parameters are optimized and adjusted. The updated main model network parameters are: :

[0080]

[0081] in, represents the model parameter weight;

[0082] S24, dynamically update the model parameter weights; according to the loss value calculated in the hybrid iterative training And the loss value Adjust the model tuning weights:

[0083]

[0084] in, represents the parameter weights before the model is updated, represents the parameter weight after the model is updated, express Normalization function;

[0085] S25, repeat steps S23 to S24, a total of iterations After completing the number of iterations, the main model The network model parameters The network model parameters finally used by the target submarine trajectory prediction model are saved to obtain the optimized target submarine trajectory prediction model.

[0086] Compared with the prior art, the present invention has the following beneficial effects:

[0087] (1) Improving the refinement and automation of trajectory prediction: The target submarine trajectory prediction model based on LSTM with multi-head attention mechanism designed by this method can significantly improve the automation of target submarine trajectory prediction. Once the model is trained, there is no need to re-model the target submarine trajectory prediction in different actual scenarios. The trajectory prediction can be directly performed after the sensor collects data. Compared with other methods, the automation of trajectory prediction is further improved. In addition, the environmental perception module based on graph neural network designed by the present invention takes into account the impact of obstacles around the target submarine on the target submarine's moving trajectory, so as to improve the refinement of the trajectory prediction of the target submarine. Finally, the model performance in complex underwater scenarios is further improved through the model reinforcement training mechanism, the refinement of the trajectory prediction of the target submarine in various marine environments is enhanced, and the ability of unmanned boats to predict the trajectory of the target submarine in complex scenarios is improved, so as to ensure the smooth progress of tracking operations;

[0088] (2) Enhance the intelligence level of tracking operations: The present invention integrates the advantages of multi-head attention mechanism and LSTM network in the target submarine trajectory prediction model. The system can deeply explore the intrinsic characteristics of trajectory data and accurately predict the future movement trajectory of the target submarine even in a complex and changeable marine environment. In addition, the sonar radar array sensor and the target submarine trajectory prediction model are seamlessly connected on the unmanned boat to enhance the ability to respond to the target submarine immediately. This enables the unmanned boat to respond to various emergencies more flexibly during the tracking process and improves the overall level of intelligent decision-making;

[0089] (3) Enhance the accuracy of trajectory prediction in complex scenarios: Based on the environmental perception module constructed by the present invention, the impact of obstacles around the target submarine on the target submarine's moving trajectory is taken into account. When the obstacles around the target submarine are relatively complex, the trajectory of the target submarine can still be accurately predicted. In addition, through the model interaction training strategy constructed in the present invention, the model's learning ability for complex scenarios can be improved. When predicting the trajectory of the target submarine, it is ensured that the unmanned boat can still accurately predict the moving trajectory of the target submarine in complex scenarios, thereby enhancing the anti-interference ability of the unmanned boat's trajectory prediction and tracking operations;

[0090] (4) Improving the generalization and robustness of trajectory prediction: In the data collection stage, in order to avoid the influence of environmental characteristics brought by a single data collection environment, the present invention conducts multi-scenario data collection to ensure the diversity of training data, which can improve the generalization and robustness of the model. In addition, the multi-head attention mechanism module in the trajectory prediction model and the parameter fusion operation in the training strategy can enhance the generalization and robustness of the model during prediction. Therefore, when the unmanned boat actually predicts the trajectory of the target submarine, the generalization and robustness of its trajectory prediction are improved, thereby further improving the mission success rate of the unmanned boat in performing tracking tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 This is the overall logic block diagram of the target submarine path prediction and tracking method of the present invention.

[0092] Figure 2 This is a schematic diagram of the target submarine trajectory prediction model structure of the present invention.

[0093] Figure 3 This is a schematic diagram of the network structure of the LSTM feature extraction module of the present invention.

[0094] Figure 4 This is a flow chart of the target submarine trajectory prediction model training of the present invention.

[0095] Figure 5 Deploy an actual workflow diagram for the model.

[0096] Figure 6 This is a system structure diagram based on the target submarine trajectory prediction model.

[0097] Figure 7 This is a graph showing the experimental results of algorithm model comparison. DETAILED DESCRIPTION

[0098] The invention will be further described below in conjunction with specific embodiments.

[0099] The overall logic of the present invention is as follows Figure 1As shown, the sonar radar array sensor of the underwater unmanned boat collects trajectory data in various sea areas to ensure the comprehensiveness and robustness of the data. The core of the present invention is to combine the multi-head attention mechanism and the target submarine trajectory prediction model of the LSTM network, which can accurately capture the trajectory characteristics and predict the future movement trajectory of the target submarine. In particular, the environmental perception module is introduced to consider the impact of obstacles around the target submarine on the target submarine trajectory prediction, thereby improving the prediction accuracy. During the training process, the model interaction strategy is used to optimize the model performance and robustness. Finally, the system was successfully deployed on the unmanned boat, collecting data in real time and predicting the target trajectory, guiding the unmanned boat to accurately track, and providing strong technical support for the fields of ocean exploration, resource development, and safety maintenance, demonstrating its great potential and value in practical applications.

[0100] Construction of unmanned submarine tracking data set: In order to eliminate the interference of environmental features on data collection, the present invention performs data collection in the target domain and source domain respectively. In the process of data collection, the historical movement trajectory data and predicted trajectory data of the target submarine are sampled, and the data of the surrounding environment of the target submarine is also sampled to provide data support for the subsequent introduction of the target submarine environmental interaction features into the target submarine trajectory prediction model;

[0101] Design of target submarine trajectory prediction model: The target submarine trajectory prediction model designed by the present invention mainly includes a trajectory prediction backbone network and an environmental perception module. The trajectory prediction backbone network mainly includes four parts: self-attention mechanism, multi-head attention mechanism, LSTM feature extraction module and trajectory output module. The environmental perception module extracts features from environmental data by establishing a graph neural network to obtain the target submarine environmental interaction features;

[0102] Design of environmental perception module: When predicting the target submarine trajectory, obstacles around the target submarine will also affect the trajectory of the target submarine. In this invention, an environmental perception module is constructed, which can model the obstacles around the target object as target submarine environmental interaction features and introduce them into the target submarine trajectory prediction backbone network to further improve the accuracy of trajectory prediction;

[0103] Design of target submarine trajectory prediction model training strategy: In the present invention, a model interaction training strategy is introduced when training the target submarine trajectory prediction model. A main model and a sub-model are designed in the model training, and the two models are trained at the same time. The parameters of the main model are updated using the results of the two model training. This strategy can improve the overall performance of the model and enhance the robustness of the model when predicting the trajectory of the target submarine. This allows the unmanned boat to achieve accurate trajectory prediction in different water conditions and under different target submarine movement conditions when predicting the trajectory of the target submarine.

[0104] Actual deployment and use of the model: The trained target submarine trajectory prediction model is deployed on the actual unmanned boat, and the actual sonar radar sensor array is deployed on the unmanned boat to collect the historical trajectory data of the target submarine. The collected target sequence data is sent to the target submarine trajectory prediction model to obtain the target submarine trajectory data at the future moment. And according to the predicted target submarine trajectory data, the power system of the unmanned boat is adjusted to complete the tracking of the target submarine.

[0105] The invention will be further described below in conjunction with specific embodiments.

[0106] 1. Construction of training dataset

[0107] Data collection environment design: The method proposed in the present invention is a data-driven method, and the quality of data collection plays a crucial role in the performance of the method. Data collection in a single sea environment will inevitably introduce the impact of environmental characteristics on data collection operations. In order to eliminate the impact of environmental characteristics on data collection operations, the present invention uses a sonar radar sensor array deployed on an unmanned boat to collect data in the target domain and source domain respectively; wherein the target domain is the operating sea area of ​​the current target boat, and the source domain is other sea scenes.

[0108] Construction of historical movement trajectory of target submarine: In the present invention, the target submarine trajectory prediction model deployed on the unmanned boat terminal is used to analyze and process the past trajectory sequence data of the target submarine, so as to complete the prediction of the future movement trajectory of the target submarine by the unmanned boat. The purpose of the present invention is to provide a target submarine trajectory prediction method to assist the unmanned boat in achieving high-precision target submarine tracking. The collected data set does not need to consider the model of the target submarine during actual tracking. Therefore, the collection time step of the present invention is The target submarine position data is obtained by splicing the target submarine position data at each time point to obtain the complete target submarine historical trajectory sequence data :

[0109]

[0110] in, Indicates the collected The target submarine historical trajectory sequence data, Indicates The specific position data of the target submarine at a certain moment, specifically:

[0111]

[0112] Since the actual ocean environment is a three-dimensional environment, the submarine's movement space is also a three-dimensional movement space. Therefore, Indicates The submarine in the collected data is at time The corresponding specific three-dimensional coordinate information, where the coordinate system is established by taking the unmanned boat as the coordinate origin and establishing the x, y, and z axes in the north, east, and down directions respectively.

[0113] Target submarine environmental obstacle position data construction: In addition to using the historical movement trajectory of the target submarine, the present invention also considers the impact of the target submarine environmental obstacles on its movement trajectory when performing trajectory prediction. Therefore, the present invention collects the target submarine environmental obstacle three-dimensional coordinate time series data , so that in the subsequent trajectory prediction, the graph neural network can be used to extract The target submarine environment interaction features are extracted and sent to the trajectory prediction model to enhance the trajectory prediction accuracy;

[0114]

[0115]

[0116] in, Indicates The environmental obstacle position data The position sequence data of obstacles, and , Represents the total number of obstacles in the target submarine's surroundings; Indicates The environmental obstacle position data Obstacles at time The specific three-dimensional position coordinates corresponding to . Item Target Submarine Environmental Obstacle Position Sequence Data for:

[0117]

[0118] The coordinate system is established with the unmanned boat as the origin, and the x, y, and z axes are established in the north, east, and down directions respectively.

[0119] Construction of target submarine prediction trajectory sequence data: based on acquisition time step The submarine position data is used as the historical trajectory sequence data of the target submarine, and the time step of the target submarine is The trajectory sequence data obtained by combining the position data of the target submarine with the target submarine prediction trajectory sequence data of the present invention is used as the trajectory sequence data of the target submarine prediction trajectory sequence data of the present invention. ;

[0120]

[0121] The target submarine historical trajectory sequence data , Target submarine environmental obstacle location data As the input data of the target submarine trajectory prediction model, the output data of the target submarine trajectory prediction model is the target submarine prediction trajectory sequence data Therefore, for each target submarine historical trajectory sequence data collected Combine to get the target submarine historical trajectory input data set , for each target submarine environmental obstacle location data collected Combine to get the target submarine environmental obstacle location data set , for each predicted trajectory data corresponding to the target submarine collected Combine to get the target submarine predicted trajectory output data set :

[0122]

[0123]

[0124]

[0125] in, .

[0126] Dataset construction in source domain and target domain: Based on the above data collection process, data collection is performed in the target domain and source domain respectively. Among them, the target submarine historical trajectory dataset, target submarine environmental obstacle location dataset, and target submarine predicted trajectory output dataset collected in the target domain are defined as , and Similarly, the target submarine historical trajectory dataset, target submarine environmental obstacle location dataset, and target submarine predicted trajectory output dataset collected in the source domain are defined as , and .

[0127] 2. Target Submarine Trajectory Prediction Model

[0128] In the present invention, the target submarine trajectory prediction model constructed is as follows Figure 2As shown in the figure, it mainly consists of two parts: the environment perception module and the target submarine trajectory prediction backbone network. The environment perception module analyzes and processes the three-dimensional coordinate time series data of obstacles around the target submarine by constructing a graph neural network to obtain the target submarine environmental interaction characteristics, and sends this feature to the target submarine trajectory prediction backbone network for fusion processing; the target submarine trajectory prediction backbone network includes a self-attention mechanism, a multi-head attention mechanism, an LSTM feature extraction module and a trajectory output module; the self-attention mechanism enables the model to automatically fuse the characteristics of the target submarine position trajectory time series data and the characteristics of the three-dimensional coordinate time series data of the surrounding obstacles to improve the receptive field of the features; the multi-head attention mechanism performs feature extraction at different levels on the fusion features output by the self-attention mechanism, and then performs weighted summation of the different levels of features extracted by multiple attention heads to obtain multi-level fusion features; then, the multi-level fusion features are input into the LSTM feature extraction module, and the bidirectional LSTM feature extraction unit of the LSTM feature extraction module can simultaneously consider the forward and backward information of the input features, thereby comprehensively capturing the time series characteristics in the target submarine motion trajectory and obtaining bidirectional time series extraction features; then the trajectory output module uses the ARMA module to perform preliminary trajectory prediction on the obtained bidirectional time series extraction features, and further adjusts them using the fully connected layer to obtain the final target submarine prediction output trajectory. Based on the above module design, a trajectory prediction model with more accurate prediction accuracy is finally realized. When it is deployed in unmanned boats, accurate trajectory prediction of target submarines can be performed to enhance the tracking capability of target submarines in complex scenarios.

[0129] 1. Environmental perception module

[0130] In the marine environment, the trajectory prediction of target submarines is significantly affected by obstacles. Obstacles not only change the trajectory by direct contact, but the eddies generated by their movement also indirectly affect the path. To this end, the present invention innovatively adopts graph neural network technology to accurately capture the complex interaction characteristics of the target submarine and surrounding obstacles. By constructing a graph structure and learning information transfer between nodes, the model can deeply understand the comprehensive impact of obstacles on the movement of target submarines, including direct collisions and indirect fluid effects. This innovative method significantly improves the accuracy and robustness of trajectory prediction, and enhances the environmental adaptability of trajectory prediction for target submarines in actual scenarios. Compared with the method that relies solely on historical mobile trajectory characteristics for prediction, the accuracy and environmental adaptability of trajectory prediction are greatly improved. It provides strong support for unmanned boats to achieve accurate tracking of target submarines. Specifically, the following steps are included:

[0131] Graph neural network construction: When predicting the trajectory of a target submarine, there are various types of obstacles in the marine environment. The existence of these obstacles will affect the trajectory of the target submarine. Some obstacles will directly come into physical contact with the target submarine, causing the target submarine's trajectory to shift. Some obstacles do not directly collide with the target submarine, but the eddies and vortices generated by their movement in the water will indirectly affect the target submarine's movement trajectory. Therefore, the present invention extracts the interactive features between the target submarine and the surrounding obstacles based on the graph neural network to further enhance the accuracy of trajectory prediction for the target submarine. The constructed graph is ;

[0132]

[0133] (1) Among them, Representatives select the total number of targets including submarines and obstacles +1 target as the node set in the constructed graph. In the present invention, for each node , , select As target submarine and obstacle node features respectively ;

[0134] (2) Represents the edge connecting each obstacle and the target submarine. Since there are E obstacles in the constructed graph, there are a total of E edges connecting the target submarine. And the relative distance between the target submarine and the obstacle As the feature of this edge;

[0135]

[0136] in, Indicates the three-dimensional coordinate position information of the target submarine, Represents the three-dimensional coordinate position information of the obstacle;

[0137] Obstacle attention coefficient design: based on the created graph representation , where each edge corresponds to the relationship between the obstacle and the target submarine. For the E obstacles, each obstacle has a different impact on the trajectory of the target submarine. When the obstacle is closer to the target submarine, its impact on the trajectory of the target submarine is naturally greater. Therefore, the obstacle weight calculation function is designed in the present invention, and the obstacle weight corresponding to each edge is added according to this function. :

[0138]

[0139] in, represents the normalization operation function;

[0140] Design of graph neural network layer feature extraction unit: For the graph neural network constructed by the present invention, the Sigmoid activation function is used to activate each node feature to obtain the activation feature of each node Based on the activation feature , and use two sequential fully connected layers and combine the Relu activation function to activate the features of each edge to obtain the features of each edge :

[0141]

[0142]

[0143] in, express Activation function, express Activation function, Represents the fully connected layer operation function;

[0144] Then, the features of each edge After addition, it is normalized by Softmax and then combined with the target submarine node feature After splicing, we get the output of the neural network layer of this graph ;

[0145]

[0146] Therefore, the operation function of the graph neural network layer can be defined as :

[0147]

[0148] in, Represents the graph neural network layer operation function;

[0149] Calculation of target submarine environment interaction features: In this invention, in order to further improve the accuracy of extracting target submarine environment interaction features, a total of three interconnected graph neural network layers are designed, and the target submarine interaction features are finally obtained through a fully connected layer. , and send the obtained obstacles into the target submarine trajectory prediction backbone network to perform target submarine trajectory prediction to enhance the trajectory prediction accuracy:

[0150]

[0151] in, represents the fully connected layer operation function, and then the target submarine environment interaction features Integrate it into the target submarine trajectory prediction backbone network.

[0152] 2. Target submarine trajectory prediction backbone network

[0153] (1) Feature vector encoding layer: for the collected submarine historical movement trajectory data , for each sequence data , N is the size of the data set; the sequence data is connected through the fully connected layer Perform vector encoding to obtain encoded feature data ;

[0154]

[0155] in, Represents the fully connected layer operation function.

[0156] (2) Self-attention mechanism module design: In the real ocean environment, there are various obstacles around the target submarine, which will affect the trajectory of the target submarine. Therefore, the target submarine environment obstacle features extracted by the target submarine environment perception module are Introduced into this target submarine trajectory prediction model. When predicting the target submarine trajectory, it is necessary to consider the encoded submarine historical trajectory feature data at the same time And environmental obstacle feature data The present invention designs a self-attention mechanism module to and Features Perform adaptive fusion to obtain fused features , in order to utilize this fusion feature to improve the accuracy of target submarine trajectory prediction;

[0157]

[0158] in, Represents the self-attention mechanism operation function.

[0159] (3) Multi-head attention mechanism module design: In order to further improve the model’s ability to extract data features, a multi-head attention mechanism module is designed in this paper to Perform multi-level feature extraction. In the multi-head attention mechanism module, each attention head represents a level of feature extraction. The present invention uses eight attention heads to realize the fusion feature extraction. Extract features at different levels, then perform weighted processing on the eight levels of features to obtain multi-level fusion features The specific process is as follows:

[0160] 1) Feature matrix calculation: Fusion features Through matrix transformation, they are mapped to On three different feature matrices:

[0161]

[0162] in, Respectively represent matrices The matrix mapping function, Indicates Group The mapping parameters used in feature matrix mapping. In order to improve the global feature extraction capability, a total of 8 sets of feature matrices are designed in this invention. . Represents the fusion features The mapped Group Feature matrix.

[0163] 2) Attention coefficient calculation: Based on the obtained Two feature matrices, calculate their attention coefficients :

[0164]

[0165] in, express Activation function, represents the scaling factor, Indicates The attention coefficient corresponding to the group of attention heads.

[0166] 3) Calculation of output feature vector of multi-head attention mechanism: The attention coefficient corresponding to the group attention head and Group Multiplying the feature matrix can get Features output by the group attention head , because in the present invention, a total of 8 sets of feature matrices are designed, that is, there are eight attention heads. In the present invention, the output features of the 8 attention heads are After weighted summation, we get the multi-level fusion features after the multi-head attention mechanism. :

[0167]

[0168]

[0169] (4) LSTM feature extraction module: Its structure is as follows Figure 3 As shown, the multi-level fusion features output by the multi-head attention mechanism The present invention constructs a feature extraction module based on LSTM. In this module, in order to further improve the model's ability to understand time series data and enhance the accuracy of the final prediction of the target submarine trajectory. The designed LSTM feature extraction module includes a forward LSTM network feature extraction unit and a reverse LSTM network extraction unit.

[0170] In the forward LSTM network feature extraction unit, the multi-level fusion features are The corresponding forward time features are arranged in order from front to back in the time dimension , and input into the forward LSTM network feature extraction unit:

[0171]

[0172] in, represents the forward time feature after processing, Represents the operation function of the forward LSTM network feature extraction unit, and the forward LSTM network feature extraction unit includes 4 LSTM network layers connected in sequence;

[0173] In the reverse LSTM network feature extraction unit, the multi-level fusion features The corresponding reverse time features are arranged in order from back to front in the time dimension , and input into the reverse LSTM network feature extraction unit:

[0174]

[0175] in, represents the reverse time feature after processing, Represents the reverse LSTM network feature extraction unit operation function, and the reverse LSTM network feature extraction unit includes 4 LSTM network layers connected in sequence;

[0176] Then the processed forward time features And the reverse time characteristics after processing Merge features to get bidirectional temporal features ;

[0177]

[0178] in, Represents the feature merging operation function.

[0179] (5) Trajectory output module: based on the collected intermediate features , combined with the trajectory prediction module designed by the present invention, the final trajectory prediction result of the target submarine can be generated. In the final trajectory prediction module, the idea of ​​combining neural network and ARMA algorithm is adopted. ARMA is an effective trajectory prediction method. After the intermediate features are sent to the ARMA module for prediction, the preliminary trajectory prediction result is obtained. ; Then the preliminary trajectory prediction result is sent to two fully connected layers for further adjustment to obtain the final predicted trajectory prediction result ;

[0180]

[0181]

[0182] in, represent Operation function, Represents the fully connected layer operation function.

[0183] 3. Target Submarine Trajectory Prediction Model Training

[0184] In the present invention, a model interactive learning training strategy is used to train the target submarine trajectory prediction model. The use of the model interactive learning training strategy significantly improves the flexibility and accuracy of the model in submarine trajectory prediction. This strategy not only effectively balances the model's learning ability for different trajectory types, but also promotes the generalization ability of the model in complex environments. When the model is actually used, it can improve the unmanned boat's ability to predict the trajectories of various target submarines in various underwater scenarios. While enhancing the model's learning effect on complex trajectories, the present invention also improves the accuracy and stability of the overall trajectory prediction, providing a more reliable and intelligent model for unmanned boats to predict the trajectory of target submarines, and enhancing the unmanned boat's overall perception of target submarines. Its flow chart is as follows Figure 4 shown.

[0185] Trajectory prediction loss function design, the goal of this invention is to predict the target submarine in the future time step The trajectory position, therefore, defines the loss function for:

[0186]

[0187]

[0188] in, and Respectively represent The data of the group In each time step, the real three-dimensional coordinate data of the target submarine and the three-dimensional coordinate data of the target submarine predicted by the model are To calculate the mean absolute error function, To calculate the mean square error function.

[0189] When the model is trained in the present invention, two identical trajectory prediction models are trained based on the designed target submarine trajectory prediction model, and one of the models is used as the main model. After training, it is used as the target submarine trajectory prediction model for actual use, and the other model is used as the secondary model Help the main model update parameters during training to further improve model performance;

[0190] S1, using the dataset collected in the target domain For the main model Perform pre-training and use the loss function The calculated loss value Combine the gradient descent method to optimize the model parameters and iterate the training times, iterative training at this time After the test, the prediction ability of the main model has stabilized. To further enhance the model performance, the main model parameters are further optimized in combination with the secondary model:

[0191] S2, main model Iterative training After that, in order to further improve the full-scenario capability of target submarine trajectory prediction, the dataset collected in the source domain is used The data in the sub-model Train and use the loss function Calculate the training loss Combine the gradient descent method to optimize the model parameters;

[0192] S3, at this time, the main model and the secondary model are used for mixed iterative training, and the data is continued to be used. For the main model Train and optimize, and define the main model network parameters obtained in the first round of hybrid iterative training as ; Using the dataset Pair Model Train and optimize, and define the sub-model network parameters obtained in the first round of iterative hybrid training as ;

[0193] At the same time, based on the main model network parameters obtained And the sub-model network parameters The main model network parameters are optimized and adjusted. The updated main model network parameters are: :

[0194]

[0195] in, represents the model parameter weight;

[0196] S4, dynamically update the model parameter weights according to the loss value calculated in the hybrid iterative training And the loss value Adjust the model adjustment weights so that the adjustment weights are better when the next round of model main parameter adjustment is carried out:

[0197]

[0198] in, represents the parameter weights before the model is updated, represents the parameter weight after the model is updated, express Normalization function;

[0199] S5, repeat steps S,3 to S4, a total of iterations After completing the number of iterations, the main model The network model parameters The network model parameters finally used by the target submarine trajectory prediction model are saved to obtain the optimized target submarine trajectory prediction model.

[0200] 4. Model Deployment and Implementation

[0201] Based on the designed target submarine trajectory prediction model and the described target submarine trajectory prediction model training strategy, the optimized target submarine trajectory prediction model is finally obtained. And this optimized target submarine trajectory prediction model is deployed to the unmanned boat to help the unmanned boat predict the trajectory of the target submarine and achieve stable tracking of it. The actual deployment flow chart of the target submarine path prediction model and tracking method for the unmanned boat is shown in the figure below. Figure 5 As shown:

[0202] Hardware deployment: In the present invention, it is necessary to collect information about the target submarine and its surrounding obstacles to obtain the position sequence data of the target submarine and its surrounding obstacles. Therefore, the present invention deploys a sonar sensor array on the unmanned boat to meet the need for information collection of the target submarine and its surrounding environmental obstacles;

[0203] Software deployment: Deploy the target submarine trajectory prediction model to the computing terminal of the unmanned boat to process and analyze the information data collected by the sensor;

[0204] Trajectory prediction information analysis and processing: acquisition time The target submarine trajectory sequence data and the position sequence data of the obstacles around the target submarine are sent to the target submarine trajectory prediction model for prediction processing to obtain the predicted time The three-dimensional coordinate trajectory of the target submarine is predicted, and the unmanned boat autonomously tracks the target submarine according to the three-dimensional coordinate trajectory of the target submarine obtained by real-time prediction;

[0205] Autonomous tracking of unmanned boats: The predicted future trajectory of the target submarine contains its three-dimensional position data at the future moment. Since the three-dimensional coordinate system is constructed based on the coordinate origin set by the unmanned boat, the unmanned boat can adjust the speed and direction of movement in real time according to the predicted position, approach the predicted trajectory of the target submarine, and finally realize autonomous tracking of the target submarine.

[0206] To implement the method of the present invention, the present invention designs a set of underwater target submarine trajectory prediction system, the system structure is as follows Figure 6 As shown. In the present invention, a sonar radar sensor array is deployed in an unmanned boat, and the array is used to collect characteristic information of the target submarine. The historical trajectory information of the target submarine collected is sent to a dynamic memory for storage, and the memory is continuously refreshed as the prediction process proceeds. The submarine historical movement trajectory data stored in the memory is sent to the trajectory prediction model together with the real-time obstacle situation characteristics around the target submarine collected. The predicted trajectory information is then analyzed and processed by the information processing unit, and the analysis results are sent to the unmanned boat control system, and the navigation speed and navigation direction of the unmanned boat are continuously adjusted to achieve accurate tracking of the target submarine. It should be noted that the hardware circuits involved are all powered by a unified power supply module.

[0207] Subsequently, the algorithm performance comparison experiment was carried out using the method proposed in the present invention, the trajectory prediction algorithm Social-Lstm, and the algorithm GAT. The experimental results are as follows: Figure 7 The prediction error is defined as follows: the absolute distance difference between the actual target submarine position and the target position predicted by the model is defined as Then, the absolute distance differences of all prediction points in the prediction process are added together to obtain the total absolute distance difference of the prediction process. . Use this absolute distance total difference The prediction error is obtained by dividing the predicted trajectory distance. The present invention performs trajectory prediction for four different distance segments: 50m, 150m, 500m, and 1000m. The results show that the method proposed by the present invention has good performance in all distance segments.

[0208] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0209] Although the above describes the specific implementation methods of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A target submarine path prediction and tracking method for an unmanned boat, characterized in that: The following steps are involved: S1, obtain the real-time three-dimensional position coordinate information of the target submarine through the sonar sensor, and obtain the time series data of the target submarine position trajectory in the time period a before this moment, and correspondingly obtain the three-dimensional coordinate time series data of the environmental obstacles in the time period a; S2, input the data obtained in S1 into the trained target submarine trajectory prediction model, and output the target submarine position trajectory sequence prediction data within the b time period after this moment; The target submarine trajectory prediction model includes an environment perception module and a target submarine trajectory prediction backbone network; the environment perception module analyzes and processes the three-dimensional coordinate time series data of environmental obstacles by constructing a graph neural network to obtain the target submarine environment interaction characteristics; the target submarine trajectory prediction backbone network includes a self-attention mechanism, a multi-head attention mechanism, an LSTM feature extraction module and a trajectory output module; the self-attention mechanism enables the model to automatically fuse the characteristics of the target submarine position trajectory time series data and the characteristics of the three-dimensional coordinate time series data of the environmental obstacles to improve the receptive field of the features; the multi-head attention mechanism performs feature extraction at different levels on the fusion features output by the self-attention mechanism, and then performs weighted summation to obtain multi-level fusion features; the bidirectional LSTM feature extraction unit of the LSTM feature extraction module simultaneously considers the forward and backward information of the input features, comprehensively captures the time series characteristics in the target submarine motion trajectory, and obtains bidirectional time series extraction features; Then the trajectory output module outputs the prediction results; The target submarine trajectory prediction model is trained based on the collected training data set, and the training data set is constructed by: The target sea area where the unmanned boat performs trajectory prediction and tracking tasks is used as the target area, and other sea areas used for data collection are used as the source area, and the sonar radar sensor array carried by the unmanned boat is used to collect data; Acquisition time step The target submarine position data at each time point is combined to obtain the complete submarine historical trajectory sequence data ; in, Indicates the collected Submarine historical trajectory sequence data, Indicates The specific position data of the submarine at that moment, specifically: Indicates The submarine in the collected data is at time The corresponding specific three-dimensional coordinate information, where the coordinate system is established with the unmanned boat as the origin of the coordinate system, and the north, east, and down directions are established respectively. x, y, z axis; Collect the three-dimensional coordinate time series data of the target submarine environmental obstacles ; in, Indicates The environmental obstacle position data The position sequence data of obstacles, and , Indicates the total number of obstacles around the target submarine; Indicates The environmental obstacle position data Obstacles at time The specific three-dimensional position coordinates corresponding to Item Target Submarine Environmental Obstacle Position Sequence Data for: Collect the time step of the target submarine The three-dimensional position coordinate sequence data obtained by combining them is used as the target submarine prediction trajectory sequence data : The collected historical trajectory sequence data of each target submarine Combine to get the target submarine historical trajectory input data set , for each collected target submarine environmental obstacle position sequence data Combine to get the target submarine environmental obstacle location data set , for each predicted trajectory sequence data corresponding to the collected target submarine Combine to get the target submarine predicted trajectory output data set ; Among them, the target submarine historical trajectory dataset, the target submarine environmental obstacle location dataset, and the target submarine predicted trajectory output dataset collected in the target domain are defined as , and The target submarine historical trajectory dataset, target submarine environmental obstacle location dataset, and target submarine predicted trajectory output dataset collected in the source domain are defined as , and ; The training process of the target submarine trajectory prediction model is: S21, using the dataset collected in the target domain For the main model Perform pre-training and use the loss function The calculated loss value Combine the gradient descent method to optimize the model parameters and iterate the training Second-rate; S22, main model Iterative training After that, the dataset is obtained by collecting data from the source domain. The data in the sub-model Train and use the loss function Calculate the training loss Combine the gradient descent method to optimize the model parameters; S23, this uses the main model and the secondary model for mixed iterative training, and continues to use the data For the main model Train and optimize, and define the main model network parameters obtained in the first round of hybrid iterative training as ; Using the dataset Pair Model Train and optimize, and define the sub-model network parameters obtained in the first round of iterative hybrid training as ; At the same time, based on the main model network parameters obtained And the sub-model network parameters The main model network parameters are optimized and adjusted. The updated main model network parameters are: : in, represents the model parameter weight; S24, dynamically update the model parameter weights; according to the loss value calculated in the hybrid iterative training And the loss value Adjust the model tuning weights: in, represents the parameter weights before the model is updated, represents the parameter weight after the model is updated, express Normalization function; S25, repeat steps S23 to S24, a total of iterations After completing the number of iterations, the main model The network model parameters The network model parameters finally used by the target submarine trajectory prediction model are saved to obtain the optimized target submarine trajectory prediction model; S3, based on the output target submarine position trajectory sequence prediction data, dynamically adjusts the moving direction and speed of its own unmanned boat through the control system in real time, and always predicts and tracks the target submarine trajectory predicted by the trajectory prediction model in advance.

2. A target submarine path prediction and tracking method for an unmanned boat as claimed in claim 1, characterized in that: The input end of the target submarine trajectory prediction backbone network includes a feature vector encoding layer, which collects the submarine historical movement trajectory data and encodes each sequence data Vector encoding is performed through the fully connected layer to obtain the encoded feature data .

3. A target submarine path prediction and tracking method for an unmanned boat as claimed in claim 1, characterized in that: The processing process of the environment perception module is as follows: S11, based on graph neural network to extract the interaction features between the target submarine and the surrounding obstacles, construct a graph : in, Representatives select the total number of targets including submarines and obstacles +1 target as the set of nodes in the graph, for each node , , select As target submarine and obstacle node features respectively ; represents the edge connecting each obstacle and the target submarine, E is the total number of obstacles, so there are a total of E edges connecting the target submarine, and the relative distance between the target submarine and the obstacle As the feature of this edge; in, Represents the three-dimensional coordinate position information of the target submarine, Represents the three-dimensional coordinate position information of the obstacle; S12, calculate the obstacle attention coefficient; design the obstacle weight calculation function, and add the corresponding obstacle weight to each edge according to this function : in, represents the normalization operation function; S13, extract features from the graph neural network layer; use the Sigmoid activation function to activate each node feature to obtain the activation features of each node , based on the activation feature , and use two sequential fully connected layers and combine the Relu activation function to activate the features of each edge to obtain the features of each edge : in, express Activation function, express Activation function, Represents the fully connected layer operation function; Then, the features of each edge After addition, it is normalized by Softmax and then combined with the target submarine node feature After splicing, we get the output of the neural network layer of this graph ; The operation function of the graph neural network layer is defined as ; S14, target submarine environment interaction feature calculation; design three interconnected graph neural network layers, and finally obtain the target submarine interaction features through a fully connected layer , and send the obtained obstacles into the target submarine trajectory prediction backbone network to perform target submarine trajectory prediction to enhance the trajectory prediction accuracy: in, Represents the fully connected layer operation function.

4. A target submarine path prediction and tracking method for an unmanned boat as claimed in claim 1, characterized in that: The multi-head attention mechanism module uses eight attention heads to realize the fusion feature of the self-attention mechanism Feature extraction at different levels, followed by weighted processing of the eight levels of features obtained to obtain multi-level fusion features ; The specific process is as follows: Feature matrix calculation: Fusion features Through matrix transformation, they are mapped to On three different feature matrices: in, Respectively represent matrices The matrix mapping function, Indicates Group The mapping parameters used when mapping the feature matrix, ; Represents the fusion features The mapped Group Feature matrix; Attention coefficient calculation: Based on the obtained Two feature matrices, calculate their attention coefficients : in, express Activation function, represents the scaling factor, Indicates The attention coefficient corresponding to the group attention head; Multi-head attention mechanism output feature vector calculation: The attention coefficient corresponding to the group attention head and Group Multiplying the feature matrix can get Features output by the group attention head , the output features of the 8 attention heads After weighted summation, we get the multi-level fusion features after the multi-head attention mechanism. .

5. A target submarine path prediction and tracking method for an unmanned boat as claimed in claim 1, characterized in that: The LSTM feature extraction module includes a forward LSTM network feature extraction unit and a reverse LSTM network extraction unit; In the forward LSTM network feature extraction unit, the multi-level fusion features are The corresponding forward time features are arranged in order from front to back in the time dimension , and input into the forward LSTM network feature extraction unit: in, represents the processed forward time feature, Represents the operation function of the forward LSTM network feature extraction unit, and the forward LSTM network feature extraction unit includes 4 LSTM network layers connected in sequence; In the reverse LSTM network feature extraction unit, the multi-level fusion features The corresponding reverse time features are arranged in order from back to front in the time dimension , and input into the reverse LSTM network feature extraction unit: in, represents the reverse time feature after processing, Represents the reverse LSTM network feature extraction unit operation function, and the reverse LSTM network feature extraction unit includes 4 LSTM network layers connected in sequence; Then the processed forward time features And the reverse time characteristics after processing Merge features to get bidirectional time features ; in, Represents the feature merging operation function.

6. A target submarine path prediction and tracking method for an unmanned boat as claimed in claim 1, characterized in that: The trajectory output module sends the intermediate features obtained to the ARMA module for prediction to obtain the preliminary trajectory prediction results. ; Then the preliminary trajectory prediction result is sent to two fully connected layers for further adjustment to obtain the final predicted trajectory prediction result ; in, represent Operation function, Represents the fully connected layer operation function.

7. A target submarine path prediction and tracking method for an unmanned boat as claimed in claim 1, characterized in that: The target submarine trajectory prediction model defines a loss function during the training process for: in, and Respectively represent The data set is in In each time step, the real three-dimensional coordinate data of the target submarine and the three-dimensional coordinate data of the target submarine predicted by the model are ; To calculate the mean absolute error function, To calculate the mean square error function; At the same time, two identical target submarine trajectory prediction models are designed for training, one of which is used as the main model After training, it is used as the target submarine trajectory prediction model for actual use, and the other model is used as the secondary model , helping the main model update parameters during training and further improving model performance.

Images