A method and device for predicting the trajectory of nodes in an air, space, sea, and underwater unmanned swarm.

Abstract

The application discloses a kind of node trajectory prediction method and device of space sea submarine unmanned cluster, belong to mobile ad hoc network technical field, comprising: real-time data acquisition and pre-processing;Unmanned node trajectory prediction model based on gated recurrent unit, Transformer and width learning system is constructed, data is input into the parallel cascade module of gated recurrent unit and Transformer;The output result of module is weighted fusion, and the weight is dynamically adjusted using back propagation algorithm;Finally, the weighted fusion result and dynamic environment index are input into width learning system, and the output of width learning system is output through fully connected layer The final unmanned node trajectory prediction result, width learning system can also dynamically adjust network width according to data changes.The method provided in the application cascades gated recurrent unit and Transformer in parallel, realizes real-time response and long-term prediction combination, introduces width learning system to quickly adapt to data distribution changes through incremental expansion, and improves the environmental adaptability of prediction model.

Description

Technical Field

[0001] This invention belongs to the field of mobile ad hoc network technology, specifically relating to a trajectory prediction method for dynamic nodes such as drones, unmanned ships, and unmanned submarines, and is particularly applicable to ad hoc network environments composed of dynamic base stations and terminal nodes at sea. This invention relates to trajectory prediction technology in mobile ad hoc networks, and particularly to the application of a method and device for predicting the trajectory of nodes in an air-space-sea-submarine unmanned swarm. Background Technology

[0002] With the development of unmanned aerial, space, land, sea, and underwater systems, more and more nodes need to predict their trajectories based on real-time environment and status. The movement trajectory of a node is affected by a variety of factors, including natural environmental factors such as geographical location and weather, as well as the interaction between nodes. Existing trajectory prediction methods mostly use traditional machine learning and deep learning techniques such as Long Short-Term Memory (LSTM) networks and Recurrent Neural Networks (RNNs). However, these methods have problems such as poor real-time performance and large computational load when facing high-dimensional, complex environments and dynamically changing data.

[0003] At sea, nodes are affected by complex environments such as ocean currents and wind speeds. A single model cannot simultaneously model spatiotemporal dynamics. Traditional short-term prediction methods, such as Kalman filtering and single LSTM, can handle sudden movements, but they are insufficient in modeling long-term dependencies, resulting in serious accumulation of long-term prediction errors. Long-term prediction methods based on physical models are computationally complex and cannot respond to sudden obstacles in real time, failing to meet the millisecond-level requirements of sudden movements. GRU and Transformer, as powerful time-series modeling tools, can be effectively combined to handle the short-term and long-term dependencies of node trajectories. However, these models often face problems of computational speed and real-time performance when dealing with large-scale, dynamically changing data.

[0004] Gated Recurrent Units (GRUs) are an effective model for time series modeling, capable of capturing local temporal dependencies. However, GRUs struggle to capture long-range spatiotemporal dependencies, leading to significant accumulation of prediction errors in mobile ad hoc networks with dynamically changing topologies. Trajectory prediction in mobile ad hoc networks typically requires handling long-term spatial and temporal dependencies, making it difficult for a single GRU model to fully capture complex motion patterns.

[0005] Transformers excel at handling long-range dependencies and can capture global information through self-attention mechanisms. However, pure Transformer models are not sensitive enough to local temporal features and have high computational complexity, making them difficult to deploy on resource-constrained mobile nodes.

[0006] Combining GRU and Transformer enables real-time response and long-term prediction synchronization. In the face of sudden changes in motion state, GRU responds quickly, while Transformer maintains global consistency. The introduction of a Broad Learning System (BLS) further enhances and optimizes feature vectors. In the rapidly changing aerospace, sea, and diving scenarios, BLS dynamically adjusts the network width according to the frequency of data changes, improving the model's environmental adaptability. The combination of GRU, Transformer, and BLS improves the model's prediction accuracy and adapts it to experimental scenarios. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a method and apparatus for predicting the node trajectory of an unmanned swarm in air, space, sea, and underwater environments, in order to eliminate or improve one or more defects existing in the prior art. It solves the problems that a single model is difficult to model short-term and long-term dependencies simultaneously, has low real-time performance in handling large-scale dynamic data, and is difficult to model in complex environments in air, space, sea, and underwater scenarios. The present invention achieves hierarchical collaboration, combines the advantages of GRU and Transformer models, and realizes the combination of real-time response and long-term prediction. At the same time, it introduces BLS to dynamically adjust the network width, quickly process and learn the latest data, thereby improving the real-time performance of trajectory prediction. Through incremental expansion, it quickly adapts to changes in data distribution and improves the environmental adaptability of the prediction model.

[0008] This invention provides a method for predicting the node trajectories of an unmanned aerial vehicle (UAV) swarm, comprising the following steps:

[0009] S1. Real-time collection of motion data, environmental perception data, meteorological and terrain data collected by satellite, and interaction information between nodes of the unmanned cluster in the air, space, sea, and underwater. The motion data of the unmanned nodes includes velocity, acceleration, heading angle, and position coordinates. Data preprocessing is performed on the collected data, including data cleaning, normalization, and time alignment.

[0010] S2. Construct a prediction model based on the fusion of GRU gated recurrent unit, Transformer and BLS width learning system. The model adopts hierarchical time series processing. First, GRU and Transformer are cascaded in parallel. The data is preprocessed and input into the parallel cascaded module of GRU and Transformer.

[0011] S3. Perform weighted fusion of the feature vectors output by GRU and Transformer. Based on the mean square error feedback between the trajectory prediction result and the true value, use the backpropagation algorithm to dynamically adjust the weights of the weighted fusion of the outputs of GRU and Transformer.

[0012] S4. The weighted and fused feature vector processed by the gated recurrent unit and Transformer, real-time environmental data, and dynamic environmental indicators are concatenated to obtain a multi-source fusion feature vector. The multi-source fusion feature vector is then input into the width learning system. The width learning system further enhances and optimizes the multi-source fusion feature vector through its feature mapping and the generation of enhancement nodes. Finally, the output of the width learning system is output as the final unmanned node trajectory prediction result through a fully connected layer. In addition, the width learning system dynamically adjusts the network width based on changes in preprocessed data, computational load, and prediction requirements.

[0013] Furthermore, the parameter selection for the GRU and Transformer modules of the prediction model follows the principles of scenario-driven parameter selection and module collaboration.

[0014] Furthermore, node expansion or pruning is performed based on dynamically adjusted trigger conditions, which include data complexity metrics, model performance feedback, and computing resource status.

[0015] The beneficial effects of the present invention are at least as follows:

[0016] The trajectory prediction method and device for unmanned aerial, space, sea and submarine clusters of the present invention achieves efficient and robust trajectory prediction and network collaborative optimization in complex and ever-changing aerospace, space, sea and submarine environments by integrating a hierarchical temporal processing architecture and dynamic adaptive mechanism of GRU, Transformer and BLS.

[0017] By collaborating with GRU and Transformer, prediction accuracy is improved. GRU focuses on capturing short-term dependencies in time-series data and handling short-term state changes of unmanned nodes to achieve rapid response in sudden scenarios. Transformer excels at mining global dependencies and spatial features of data, analyzing long-term relationships such as relative positions and collaborative tasks among unmanned nodes, and maintaining long-term global consistency. The two processes in parallel and then perform weighted fusion of the outputs. The fusion result provides comprehensive feature information for subsequent BLS processing. This collaborative mechanism enables the model to learn more complex mobile ad hoc network patterns, significantly improving trajectory prediction accuracy.

[0018] BLS receives the fusion results of GRU and Transformer, performs deep feature processing, and mines potential complex data patterns through feature mapping and enhanced node generation. In addition, under the ever-changing environment of air, space, sea and submarine, BLS automatically expands the number of nodes according to the complexity of input data, responds to environmental changes by generating feature nodes, and adjusts computing resources by generating enhanced nodes, so that the model can adapt to data changes and avoid overfitting or underfitting.

[0019] Unmanned aerial, space, sea, and underwater scenarios are diverse and uncertain, with significant differences in conditions across different regions and missions. The combined model of GRU, Transformer, and BLS can learn richer and more comprehensive feature patterns, is less prone to overfitting, and has stronger generalization ability. It can maintain good performance in various complex and ever-changing aerial, space, sea, and underwater scenarios, and adapt to different mission requirements and environmental changes.

[0020] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0021] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0023] Figure 1 A flowchart of the node trajectory prediction method and device for an unmanned aerial, space, sea, and underwater swarm provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the overall structure of the unmanned node trajectory prediction model provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention.

[0026] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0027] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0028] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0029] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings, in which the same reference numerals represent the same or similar parts or the same or similar steps.

[0030] The air-space-sea-submarine network in this invention includes space-based, air-based, sea-based, and submarine-based networks. The space-based network consists of high-orbit high-throughput satellites. The air-based network includes medium-altitude long-endurance UAVs and high-altitude long-endurance UAVs. The sea-based network includes unmanned surface vessels (USVs) and civilian vessels equipped with high-throughput satellite terminals, underwater acoustic receivers, and shipborne base stations. The submarine-based network consists of unmanned submarines. USVs, USVs, and unmanned submarines access civilian vessels through unmanned, random access to construct a maritime local area network. Based on motion trajectory prediction, a cooperative neighbor discovery algorithm is formed.

[0031] This invention provides a method for predicting the node trajectories of an unmanned aerial vehicle (UAV) swarm, such as... Figure 1 As shown, the method includes the following steps S1 to S4:

[0032] S1. Real-time collection of motion data, environmental perception data, meteorological and terrain data collected by satellite, and interaction information between nodes of the unmanned cluster in the air, space, sea, and underwater. The motion data of the unmanned nodes includes velocity, acceleration, heading angle, and position coordinates. Data preprocessing is performed on the collected data, including data cleaning, normalization, and time alignment.

[0033] S2. Construct a prediction model based on the fusion of GRU gated recurrent unit, Transformer and BLS width learning system. The model adopts hierarchical time series processing. First, GRU and Transformer are cascaded in parallel. The data is preprocessed and input into the parallel cascaded module of GRU and Transformer.

[0034] S3. Perform weighted fusion of the feature vectors output by GRU and Transformer. Based on the mean square error feedback between the trajectory prediction result and the true value, use the backpropagation algorithm to dynamically adjust the weights of the weighted fusion of the outputs of GRU and Transformer.

[0035] S4. The weighted fusion feature vector processed by GRU and Transformer, real-time environmental data, and dynamic environmental indicators are input into the width learning system. The width learning system further enhances and optimizes the fused features through its feature mapping and the generation of enhanced nodes. Finally, the output of the width learning system is output through a fully connected layer to produce the final unmanned node trajectory prediction result, including the node's position coordinates and motion state. In addition, BLS dynamically adjusts the network width based on changes in preprocessed data, computational load, and prediction requirements.

[0036] By predicting the movement trajectories of unmanned nodes, prior information is provided for neighbor discovery in ad hoc networks, thereby optimizing network topology maintenance, reducing communication overhead, and improving network stability in dynamic environments.

[0037] The steps described above will be explained in detail below.

[0038] In step S1 above, data is collected in real time from each node of the unmanned swarm, including drones, unmanned ships, and unmanned submarines. The data includes the node's position, velocity, acceleration, and heading angle. In addition to local node data, the system can also collect information from other nodes, such as their relative positions and velocities, environmental perception data, and meteorological and terrain data collected by satellites.

[0039] The collected data undergoes outlier detection and processing, with outliers corrected or deleted. Filtering algorithms are used to remove noise interference from the data. Data from different dimensions is normalized to ensure they fall within the same scale range. For location coordinate data, a maximum-minimum normalization method is used to map longitude and latitude values ​​to the [0,1] interval; the same method is applied to velocity data. Based on the timestamps of each node, the system aligns the data across all nodes to ensure that data from different nodes can be processed within the same time window.

[0040] In step S2 above, the prediction model based on the fusion of GRU gated recurrent unit, Transformer, and BLS width learning system is referenced. Figure 2 As shown, the prediction model combines GRU and Transformer in a parallel cascade manner, and the results processed by GRU and Transformer are weighted and fused. The fused feature vector, real-time environmental data and dynamic environmental indicators are input into the wide learning system.

[0041] The gated loop unit (GRU) and the Transformer are connected in parallel, and their outputs are weighted and fused to achieve an efficient combination of local and global feature extraction. The GRU achieves millisecond-level response, mitigating the impact of signal transmission delay. The Transformer is suitable for handling long-term navigation across media, maintaining global path consistency, and performing global modeling and denoising of complex data subject to interference. The GRU handles short-term state changes of each unmanned node, responds quickly to sudden movements, updates short-term predictions, and helps nodes adjust their neighbor lists in a timely manner. The Transformer extracts long-term dependency features by analyzing the relative positions of each unmanned node, cooperative task allocation relationships, formation evolution, and cluster communication topology evolution, predicting the future clustering areas of nodes and helping to optimize the global neighbor discovery strategy.

[0042] Input of GRU module For motion time series features, it represents the linear velocity of the unmanned node in three dimensions within the corresponding time series length. angular velocity and acceleration One-dimensional heading angle and three-dimensional position coordinates The output is the encoded hidden motion state and short-term trajectory feature vector. Input to the Transformer module This represents multi-dimensional spatiotemporal characteristics, indicating the three-dimensional linear velocity, angular velocity, and acceleration of the unmanned node within the corresponding time series length, the one-dimensional heading angle, the three-dimensional coordinates of the position, and the wind speed. Meteorological data consisting of wave height and precipitation intensity, and ocean current speed. Ocean data consisting of water depth and topographic data, including the relative positions between multiple nodes. Relative velocity The data includes multi-node interaction data from the cluster topology, satellite-collected meteorological remote sensing data, and terrain augmentation data. The output includes long-term dependent feature vectors. The cross-attention matrix represents the collaborative relationships between nodes.

[0043] GRU effectively captures short-term dependencies in time series data by controlling the flow of information through update and reset gates. GRU consists of two gates: the update gate and the reset gate. Their computation involves the input data, the hidden state from the previous time step, and the corresponding weight matrix and bias term. Then, the candidate hidden states are calculated, and combined with the information from the reset gate, the current hidden state is finally updated. Its core computational steps are as follows:

[0044] The reset gate controls the degree to which information from the previous hidden state is retained; the reset gate outputs a vector. ,in, It is the sigmoid activation function. and These are the weight matrix and bias term of the reset gate, respectively. It was in a hidden state a moment ago. It is the current input vector;

[0045] The update gate determines the degree of update of the hidden state, and the update gate output vector. ;

[0046] Candidate states and candidate hidden states are generated by combining the reset gate and the current input. , where ⊙ denotes element-wise multiplication, and tanh denotes the hyperbolic tangent activation function;

[0047] By updating the door and merging the old and new states, the final hidden state is achieved. .

[0048] The Transformer primarily consists of a self-attention mechanism and a feed-forward network. The self-attention mechanism aggregates information by calculating the weights of the query, key, and value. Multi-head attention parallelizes this process multiple times to capture information from different subspaces. Additionally, positional encoding is used to inject positional information into the sequence, as the Transformer itself cannot process sequence order. The specific components are described below:

[0049] The self-attention mechanism captures global dependencies within a sequence through query, key, and value matrices, first generating Q, K, and V matrices. , , Where X represents the input sequence, , and It is a learnable projection matrix, scaling dot product attention. ,in This represents the dimension of the key vector, and the scaling factor prevents gradient vanishing.

[0050] Multi-head attention mechanisms involve using multiple attention heads in parallel to capture features from different subspaces. The calculation for each head is as follows: ,in , and These are the independent projection matrices for each head. To output the projection matrix;

[0051] The feedforward neural network performs a nonlinear transformation on the attention output. ,in and It is a weight matrix. and It is a bias term;

[0052] Position encoding is performed by injecting sequence position information. , , where pos is the sequence position and i is the dimension index.

[0053] The following section explains the parameter selection for the GRU and Transformer modules in the prediction model:

[0054] The parameter selection for the GRU and Transformer modules of the prediction model follows the principles of scenario-driven parameter selection and module collaboration.

[0055] The required latency varies depending on the mission scenario of drones, unmanned ships, and unmanned submarines, and is dynamically adjusted according to motion patterns, environmental complexity, and computational resource constraints.

[0056] Drone swarms require millisecond-level response and high-precision prediction. Gated loop units need short time steps, reducing the number of hidden layers in the gated loop units to no more than 2. The number of hidden units is dynamically adjusted based on real-time computing resources. When resources are sufficient, the number is increased to improve accuracy, and when resources are limited, the number is reduced to reduce latency. The number of attention heads and encoder / decoder layers in the Transformer are dynamically set according to the current task scenario and computing resources. The number of attention heads is no more than 4, and the number of encoder / decoder layers is no more than 3. The feedforward network dimension is adapted to the sensor data dimension. Local attention is used to reduce computational overhead. The network parameter weights inside the gated loop units and Transformer are updated through online learning to adapt to sudden changes in motion state.

[0057] For long-term cruise and steady-state operation of unmanned surface vessels (USVs), resistance to environmental interference is required. The time step and number of hidden layers of the gated loop unit are dynamically set according to the task scenario and computing resources. The number of hidden layers is set to 2 to 4, and the number of hidden units balances short-term response and noise resistance. The number of attention heads and encoder and decoder layers of the Transformer are adapted to the environmental data. The number of attention heads is set to 4 to 8, and the number of encoder and decoder layers is no more than 4. The dimension of the feedforward network is positively correlated with the number of interaction nodes. The ability to model long sequences is enhanced by position encoding. The attention window covers a complete motion change cycle of the USV during stable operation in a regular task, such as the average time to cruise from one port to another and complete a full-area scan.

[0058] The unmanned submarine swarm requires low power consumption and long-term prediction. The gated loop unit extends the timing window to capture slow-changing motion patterns. The number of hidden units and layers is dynamically compressed according to the power budget. After compression, the number of hidden layers is no more than 2. The number of attention heads is reduced by the Transformer, which is no more than 2. The number of encoder and decoder layers and the dimension of the feedforward network are increased to improve the representation capability. The number of encoder stacked layers is no less than 4. Sparse attention is used to reduce the computational load.

[0059] From a system perspective, considering the synergistic effect between modules, the GRU and Transformer are time-aligned. The GRU time step is T, and the Transformer attention window is set to kT. k is adjusted according to the task stage. During sudden motion, k takes a smaller value, and during steady-state navigation, k takes a larger value. The smaller value is no greater than 2, and the larger value is no less than 3. The outputs of GRU and Transformer are aligned to the same dimension through linear projection to ensure the consistency of the fusion module input.

[0060] In step S3 above, GRU and Transformer respectively generate short-term trajectory feature vectors. and long-term dependent feature vectors By merging their outputs through weighted summation, and assigning weights to the outputs of GRU and Transformer respectively, the weighted sum output will be... It can be represented as ,in and These are learnable weights.

[0061] Dynamically adjusting the weighted fusion weights of GRU and Transformer using the backpropagation algorithm includes the following steps:

[0062] Define the forward propagation structure of the model, which includes two parallel prediction modules, GRU and Transformer, and a weighted summation layer to fuse the outputs of the two modules;

[0063] Choosing the mean squared error as the loss function, the loss function L is expressed as: ;

[0064] Calculate the loss between the weighted fused output and the true label using the defined loss function; starting from the output layer, compute the gradient of the loss function with respect to each parameter in reverse order. The gradient is Calculate the pair using the chain rule and The gradients are respectively and ;

[0065] Update the weights using gradient descent.

[0066] Repeat the steps of calculating loss, calculating gradient, and updating weights until the decrease in the loss function value is lower than the preset threshold or the preset number of iterations is reached.

[0067] In step S4 above, the input of BLS includes the weighted fusion feature vector processed by GRU and Transformer, real-time environmental data, and dynamic environmental indicators. The dynamic environmental indicators include the rate of change of statistical features of the input data and the status of computing resource usage. It quickly fuses multi-source data including satellite positioning and environmental factors to provide real-time updated prediction results and supports dynamic adjustment of neighbor discovery parameters. The width learning system generates high-dimensional feature representations through feature nodes and enhancement nodes, maps them to the target dimension through a fully connected layer, and outputs the trajectory prediction results of unmanned nodes in the future, including the position coordinates and motion state of the nodes. In addition, under the condition of changing air, space, sea and submarine environments, BLS automatically expands the number of nodes according to the complexity of input data, generates feature nodes to respond to environmental changes, and generates enhancement nodes to adjust computing resources, so that the model can adapt to data changes and avoid overfitting or underfitting.

[0068] BLS fuses multi-source data, utilizing a randomly generated weight matrix through its feature mapping layer. The fused input data is mapped to a high-dimensional feature space to obtain feature nodes Z, represented as: Where X is the fused input data, It is an activation function. This is the bias vector.

[0069] Enhanced nodes are generated based on feature nodes through nonlinear transformations using random matrices. The feature nodes are transformed to generate the enhanced node Y, represented as: ,in This is the bias vector.

[0070] The output of BLS is fed into a fully connected layer, where the final node trajectory prediction result is calculated using a weight matrix. The calculation formula is as follows: ,in It is the weight matrix of the fully connected layer. It is a bias vector. This is the output of BLS; in this way, the BLS module plays a key role in the unmanned node trajectory prediction model. It not only integrates the advantages of GRU and Transformer, but also improves the model's adaptability and prediction performance through dynamic adjustment and optimization mechanisms.

[0071] BLS automatically adjusts the network width based on the system's real-time prediction demand and computing resource status to optimize the use of computing resources. When real-time prediction demand is high and computing resources are sufficient, the number of nodes in the feature mapping layer and enhancement node layer is increased to enhance the model's expressive power. When prediction demand is low or computing resources are scarce, the number of nodes is reduced to avoid resource waste. The specific steps are as follows:

[0072] Define the dynamic adjustment trigger conditions, including data complexity measurement, model performance feedback, and computing resource status;

[0073] Data complexity metrics include entropy monitoring and gradient rate of change. The Shannon entropy of the input data within the current window is calculated as follows: When the Shannon entropy increases by more than 20% compared to the Shannon entropy of the previous time window, node expansion is triggered, and the mean absolute value of the gradient within the time window is calculated for motion state data, environmental perception data, and node interaction data. ,like This triggers an adjustment.

[0074] The performance feedback metric for the model is the model prediction residual. The mean square error (MSE) between the predicted trajectory and the actual trajectory is calculated in real time. If the residual increases beyond the threshold for k consecutive steps, an adjustment is triggered.

[0075] Computing resource status includes hardware resource utilization, hardware resource idleness allowing for node expansion, and overload triggering pruning;

[0076] Adjusting network width includes node expansion and node pruning;

[0077] When an increase in data complexity or a decrease in model performance is detected, and resources are sufficient, node expansion is performed. Nodes include feature nodes and enhancement nodes, and this is done through mutual information. Feature screening is performed to select features that are highly correlated with the current mutation. Based on the characteristics, the weights and biases of new feature nodes are randomly generated, the new nodes are integrated into the network, and nonlinear transformations are performed on the newly added feature nodes to generate enhanced nodes.

[0078] When data complexity decreases or resources become scarce, node pruning is performed. The importance of each node is evaluated by calculating the L1 norm of the node weights, and nodes with importance scores below a preset threshold are removed from the network.

[0079] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0080] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0081] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the node trajectory of an unmanned aerial vehicle (UAV) swarm, characterized in that, The space-air-sea-submarine network comprises space-based, air-based, sea-based, and submersible-based networks. The space-based network consists of high-orbit, high-throughput satellites; the air-based network consists of unmanned aerial vehicles (UAVs); the sea-based network includes unmanned surface vessels (USVs) and civilian vessels equipped with high-throughput satellite terminals, underwater acoustic receivers, and shipborne base stations; and the submersible network consists of unmanned submarines. UAVs, USVs, and unmanned submarines connect to civilian vessels via unmanned, on-demand access to construct a maritime local area network. Based on motion trajectory prediction, a cooperative neighbor discovery algorithm is developed. The trajectory prediction includes the following steps: S1. Real-time collection of motion data, environmental perception data, meteorological and terrain data collected by satellite, and interaction information between unmanned nodes in the air-space-sea-submarine unmanned cluster; and data preprocessing of the collected data. S2. Construct a prediction model based on the fusion of gated recurrent unit, Transformer and width learning system. The model adopts hierarchical time-series processing. First, the gated recurrent unit and Transformer are cascaded in parallel. The real-time collected data is preprocessed and then input into the parallel cascaded module of gated recurrent unit and Transformer. S3. Perform weighted fusion on the feature vectors output by the gated loop unit and the Transformer. Based on the mean square error feedback between the trajectory prediction result and the true value, dynamically adjust the weights of the weighted fusion using the backpropagation algorithm. S4. The weighted and fused feature vector processed by the gated recurrent unit and Transformer, real-time environmental data, and dynamic environmental indicators are concatenated to obtain a multi-source fusion feature vector. The multi-source fusion feature vector is then input into the width learning system. The width learning system further enhances and optimizes the multi-source fusion feature vector through its feature mapping and the generation of enhancement nodes. Finally, the output of the width learning system is output as the final unmanned node trajectory prediction result through a fully connected layer. In addition, the width learning system dynamically adjusts the network width based on changes in preprocessed data, computational load, and prediction requirements. By predicting the movement trajectories of unmanned nodes, prior information is provided for neighbor discovery in ad hoc networks, thereby optimizing network topology maintenance, reducing communication overhead, and improving network stability in dynamic environments.

2. The node trajectory prediction method for an unmanned aerial, space, sea, and underwater swarm according to claim 1, characterized in that, In step S2, the roles of the gated recurrent unit, Transformer, and width learning system module in the prediction model, as well as their inputs and outputs, are as follows: The gated loop unit and the Transformer are connected in parallel. The outputs of the two are weighted and fused to achieve an efficient combination of local and global feature extraction. The gated loop unit achieves millisecond-level response, which alleviates the impact of signal transmission delay. The Transformer is suitable for handling long-term navigation across media, maintaining global path consistency, and performing global modeling and denoising of complex data subject to interference. The gated loop unit handles the short-term state changes of each unmanned node, responds quickly to sudden movements, updates short-term predictions, and helps nodes adjust their neighbor lists in a timely manner. Transformer analyzes the relative positions of each unmanned node, the cooperative task allocation relationship, the formation evolution, and the cluster communication topology evolution to extract long-term dependency features, predict the future clustering area of ​​nodes, and help optimize the global neighbor discovery strategy. Input of the gated loop unit module For motion time series features, it represents the linear velocity of the unmanned node in three dimensions within the corresponding time series length. angular velocity and acceleration One-dimensional heading angle and three-dimensional position coordinates The output is the encoded hidden motion state and short-term trajectory feature vector. Input to the Transformer module This represents multi-dimensional spatiotemporal characteristics, indicating the three-dimensional linear velocity, angular velocity, and acceleration of the unmanned node within the corresponding time series length, the one-dimensional heading angle, the three-dimensional coordinates of the position, and the wind speed. Meteorological data consisting of wave height and precipitation intensity, and ocean current speed. Ocean data consisting of water depth and topographic data, including the relative positions between multiple nodes. Relative velocity The data includes multi-node interaction data from the cluster topology, satellite-collected meteorological remote sensing data, and terrain augmentation data. The output includes long-term dependent feature vectors. The cross-attention matrix represents the collaborative relationships between nodes; The input to the width learning system includes a weighted and fused feature vector processed by a gated recurrent unit and a Transformer, real-time environmental data, and dynamic environmental indicators. The dynamic environmental indicators include the rate of change of statistical features of the input data and the status of computational resource usage. It quickly fuses multi-source data, provides real-time updated prediction results, and supports dynamic adjustment of neighbor discovery parameters. The width learning system generates high-dimensional feature representations through feature nodes and augmentation nodes, maps them to the target dimension through a fully connected layer, and outputs the trajectory prediction results of unmanned nodes over a period of time, including the node's position coordinates and motion state. In addition, under the condition of changing air, space, sea, and underwater environments, BLS automatically expands the number of nodes according to the complexity of the input data, generates feature nodes to respond to environmental changes, and generates augmentation nodes to adjust computational resources, allowing the model to adapt to data changes and avoid overfitting or underfitting.

3. The node trajectory prediction method for an unmanned aerial, space, sea, and underwater swarm according to claim 1, characterized in that, The selection of the gated recurrent unit and Transformer parameters in step S2 is as follows: Scene-driven parameter selection dynamically adjusts based on motion mode, environmental complexity, and computational resource constraints. The required latency varies depending on the mission scenario of drones, unmanned ships, and unmanned submarines. Drone swarms require millisecond-level response and high-precision prediction. Gated loop units need short time steps, reducing the number of hidden layers in the gated loop units to no more than 2. The number of hidden units is dynamically adjusted based on real-time computing resources. When resources are sufficient, the number is increased to improve accuracy, and when resources are limited, the number is reduced to reduce latency. The number of attention heads and encoder / decoder layers in the Transformer are dynamically set according to the current task scenario and computing resources. The number of attention heads is no more than 4, and the number of encoder / decoder layers is no more than 3. The feedforward network dimension is adapted to the sensor data dimension. Local attention is used to reduce computational overhead. The network parameter weights inside the gated loop units and Transformer are updated through online learning to adapt to sudden changes in motion state. For long-term cruise and steady-state operation of unmanned surface vessels (USVs), resistance to environmental interference is required. The time step and number of hidden layers of the gated loop unit are dynamically set according to the task scenario and computing resources. The number of hidden layers is set to 2 to 4, and the number of hidden units balances short-term response and noise resistance. The number of attention heads and encoder and decoder layers of the Transformer are adapted to the environmental data. The number of attention heads is set to 4 to 8, and the number of encoder and decoder layers is no more than 4. The dimension of the feedforward network is positively correlated with the number of interaction nodes. The ability to model long sequences is enhanced by position encoding. The attention window covers a complete motion change cycle of the USV during stable operation in a regular task, such as the average time to cruise from one port to another and complete a full-area scan. The unmanned submarine swarm requires low power consumption and long-term prediction. The gated loop unit extends the timing window to capture slow-changing motion patterns. The number of hidden units and layers is dynamically compressed according to the power budget. After compression, the number of hidden layers is no more than 2. The number of attention heads is reduced by the Transformer, which is no more than 2. The number of encoder and decoder layers and the dimension of the feedforward network are increased to improve the representation capability. The encoder stacking layers are no less than 4. Sparse attention is used to reduce the computational load. To ensure collaborative operation between modules, the gated loop unit and the Transformer are time-aligned. The time step of the gated loop unit is T, and the attention window of the Transformer is set to kT. k is adjusted according to the task stage, taking a smaller value during sudden motion and a larger value during steady-state navigation. The smaller value is no greater than 2, and the larger value is no less than 3. The outputs of the gated loop unit and the Transformer are aligned to the same dimension through linear projection to ensure the consistency of the fusion module input.

4. The node trajectory prediction method for an unmanned aerial, space, sea, and underwater swarm according to claim 1, characterized in that, The step S3, which involves dynamically adjusting the gated recurrent unit and the Transformer weighted fusion weights using the backpropagation algorithm, is as follows: S31. Define the forward propagation structure of the model, including two parallel prediction modules: a gated recurrent unit and a Transformer, and a weighted summation layer that fuses the outputs of the two modules. Assign weights to the outputs of the gated recurrent unit and the Transformer respectively. and The output of the gated loop unit is The output of Transformer is Then the output after weighted summation It can be represented as: ; S32. Choose the mean squared error as the loss function, and the loss function L is expressed as: The loss between the weighted fusion output and the true label is calculated using the defined loss function; S33. Starting from the output layer, calculate the gradient of the loss function with respect to each parameter in reverse order. The gradient is Calculate the pair using the chain rule and The gradients are respectively and Update the weights using gradient descent. S34. Repeat the above steps of calculating loss, calculating gradient and updating weights until the decrease in the loss function value is lower than the preset threshold or the preset number of iterations is reached.

5. The node trajectory prediction method for an unmanned aerial, space, sea, and underwater swarm according to claim 1, characterized in that, In step S4, the width learning system dynamically adjusts the network width based on real-time predicted demand and computing resource availability. S41. Define dynamic adjustment trigger conditions: data complexity measurement, model performance feedback, and computing resource status; Data complexity metrics include entropy monitoring and gradient rate of change. The Shannon entropy of the input data within the current window is calculated as follows: When the Shannon entropy increases by more than 20% compared to the Shannon entropy of the previous time window, node expansion is triggered; the mean absolute value of the gradient within the time window is calculated for motion state data, environmental perception data, and node interaction data. ,like This triggers an adjustment. The performance feedback metric for the model is the model prediction residual. The mean square error (MSE) between the predicted trajectory and the actual trajectory is calculated in real time. If the residual increases beyond the threshold for k consecutive steps, an adjustment is triggered. Computing resource status includes hardware resource utilization, hardware resource idleness allowing for node expansion, and overload triggering pruning; S42. Dynamically adjust the specific process: node expansion, node pruning; When an increase in data complexity or a decrease in model performance is detected, and resources are sufficient, node expansion is performed. Nodes include feature nodes and enhancement nodes, and this is done through mutual information. Feature screening is performed to select features that are highly correlated with the current mutation. Based on the characteristics, the weights and biases of new feature nodes are randomly generated, the new nodes are integrated into the network, and nonlinear transformations are performed on the newly added feature nodes to generate enhanced nodes. When data complexity decreases or resources become scarce, node pruning is performed. The importance of each node is evaluated by calculating the L1 norm of the node weights, and nodes with importance scores below a preset threshold are removed from the network.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 5.

Images