An underwater target intention recognition method based on trajectory sequence modeling

Abstract

The application discloses a method and device for underwater target intention recognition based on trajectory sequence modeling, equipment, medium and program, and relates to the field of artificial intelligence. The method comprises the following steps: constructing a multi-behavior intention task scene model of an underwater target, establishing a multi-intention rule library comprising convergence, inspection, avoidance and departure, and extracting trajectory features; constructing a Markov decision process model based on reinforcement learning, designing a differentiated reward function to generate multi-type trajectory data; normalizing and dividing the trajectory data to construct a sample set; constructing an LSTM model with a fusion attention mechanism to train and realize intention classification; and inputting real-time trajectory segments into the model by using a sliding window mechanism, and outputting real-time recognition results. The application effectively alleviates the difficulty of obtaining real underwater data, improves the generalization performance of the model, realizes effective extraction and real-time processing of key features of the trajectory sequence, and improves the recognition accuracy and real-time performance in a complex underwater environment.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and underwater target perception technology, specifically to an underwater target intent recognition method, device, electronic device, storage medium, and program based on trajectory sequence modeling. Background Technology

[0002] With the widespread application of autonomous underwater vehicles (AUVs) in marine exploration, resource development, environmental monitoring, and underwater operations, underwater target behavior analysis and intelligent decision-making have gradually become one of the core issues of intelligent underwater systems. The underwater environment is characterized by high complexity, limited perception conditions, low communication bandwidth, and high latency, posing challenges to traditional target recognition and tracking methods based on visual or sonar signals at the behavioral understanding level. Underwater target intent recognition, as a high-level cognitive task that infers the future behavior or purpose of a target from observational data, is of great significance for improving the intelligence level of autonomous systems.

[0003] In multi-agent cooperative tasks, intent recognition is particularly crucial. Different underwater vehicles need to cooperate to complete searches, tracking, and information sharing under conditions of limited communication and environmental uncertainty. Accurately predicting the behavioral intent of companions can significantly improve cooperation efficiency and task completion rate. Existing research has proposed that by integrating deep learning and cognitive reasoning models, effective inference of the behavioral intent of multiple agents can be achieved under limited data exchange conditions.

[0004] Furthermore, incorporating uncertainty reasoning, information reasoning, and potential outcome frameworks into intent recognition can enhance the understanding of target behavioral motivations, thereby improving the prediction accuracy of multi-objective tasks. The core process of tactical intent recognition consists of two key steps:

[0005] (1) Select and extract intent-related features from real-time updated underwater situational awareness and multi-source sensing data.

[0006] (2) The feature information is fused and analyzed to determine the target’s movement intention.

[0007] The technical process of underwater intent recognition typically includes behavioral feature extraction, state-action mapping establishment, and intent inference. However, due to the difficulty in obtaining trajectory data, the presence of noise, and the complex temporal correlations among multiple intent behaviors, traditional rule-based or shallow learning methods have limited performance in handling complex trajectories and dynamic behavioral patterns. To address the data acquisition challenge, reinforcement learning and simulation methods are widely used to generate controllable trajectory data.

[0008] By employing differentiated reward design and multi-agent policy optimization, diverse and multi-distributed target trajectory data can be generated, providing rich training samples for intent recognition models. In recent years, deep learning sequence models have received widespread attention in underwater intent recognition. Long Short-Term Memory (LSTM) networks can be used to enhance the discrimination ability of trajectories with strong temporal correlation and indistinct local behavioral features. By weighting key trajectory segments, the accuracy of intent prediction can be improved.

[0009] Currently, underwater target behavior and intent recognition is a research hotspot in the field of situation estimation and a core basis for decision-making systems to formulate task plans. In complex mission contexts, the complexity and dynamism of the underwater environment are significantly enhanced, making it difficult for decision-making systems to efficiently process massive amounts of environmental information and accurately determine the behavior and intent of underwater targets within a limited timeframe. This creates an urgent need for intelligent command systems—requiring intelligent means to assist decision-making systems in shortening response cycles and improving the scientific nature of decision-making.

[0010] Although some progress has been made in the research on underwater target intent recognition in recent years, under the dual constraints of complex task scenarios and practical application needs, there are still several key issues in this field that need to be addressed. These issues are mainly reflected in the following aspects: (1) Underwater target intent has strong concealment and diversity, making it difficult to model in a unified manner; (2) It is difficult to obtain real underwater multi-intent trajectory data, which restricts the effective application of data-driven methods; (3) Underwater target intent recognition needs to take into account both recognition accuracy and real-time performance, and existing methods are difficult to balance the two.

[0011] In summary, given the scarcity of underwater environment datasets and the complexity of the environment, this invention proposes an underwater target intent recognition method based on trajectory sequence modeling. The aim is to identify intent by simulating underwater intent trajectory data, providing a solution and technical reserves for underwater intent recognition tasks. Summary of the Invention

[0012] This application discloses an underwater target intent recognition method, device, electronic device, storage medium, and program based on trajectory sequence modeling, aiming to solve the technical problems of strong concealment of target behavior, complex maneuvering patterns, intertwined multiple intent features, low recognition accuracy, and insufficient real-time performance of traditional methods in complex underwater environments.

[0013] To achieve the above objectives, the first aspect of this application provides a method for underwater target intent recognition based on trajectory sequence modeling, comprising: Construct a multi-behavior intent task scenario model for underwater targets, establish a multi-intent rule base including convergence, inspection, avoidance, and departure, and extract trajectory features corresponding to various intents based on the target kinematic model; A Markov decision process model driven by intent is constructed based on reinforcement learning methods. A differentiated reward function is designed according to the multi-intent rule base, and multi-type underwater target motion trajectory data is generated by combining the proximal policy optimization algorithm. The generated trajectory data is normalized and sequenced to construct a sample dataset for model training. A long short-term memory network model integrating an attention mechanism is constructed and trained using the sample dataset. The model performs feature extraction and temporal modeling on trajectory sequences to achieve target intent classification. A sliding window mechanism is used to segment the real-time acquired target motion trajectory and input it into the trained model, outputting the real-time recognition result of the underwater target's intent.

[0014] Optionally, trajectory features are extracted based on the target kinematic model, and the relative distance, relative radial velocity, and relative azimuth difference between the target position vector and the observation node position are defined. For convergence intentions, the rule is to continuously approach the observation node and align the course, reduce the relative spacing and make the track angle close; The rule for the inspection intention is to move within the pre-defined circular corridor, which increases the variance of the turning rate and the path curvature. For evasive intent, the rule is that the heading deviation and the frequency of course changes increase, which manifests as a sudden change in trajectory; The rule for the intention to leave is to continuously move away from the center of gravity of the area, and the relative distance scale continuously increases.

[0015] Optionally, a state space containing position, speed, and heading information is constructed, and an action space for heading changes and speed adjustments is defined; The differentiated reward function includes: for convergence intention, reinforcing approach behavior, penalizing departure behavior, and encouraging heading toward the target direction; for inspection intention, rewarding movement within the loop area and encouraging circular circling around the target; for avoidance intention, providing incentives based on a threshold range of turning amplitude; and for departure intention, rewarding increasing relative distance and constraining heading deviation.

[0016] Optionally, the generated trajectory data is stored in CSV format, including time step number, position coordinates, motion speed, heading angle, rate of change of velocity, and rate of change of angular velocity; The trajectory data is normalized and divided into sequences according to the time step order to generate a fixed-length time-series sample dataset.

[0017] Optionally, the long short-term memory network model that integrates attention mechanisms uses the long short-term memory network as the base network to extract the temporal features of the trajectory sequence; In the attention mechanism, the features of the previous state at the current moment are used as the query vector. The relevance of the key vectors in the historical state sequence is measured to obtain the attention weight. The value vectors of the historical state sequence are weighted and summed to obtain the global context representation. The intention is then classified by combining the hidden state of the Long Short-Term Memory network.

[0018] Optionally, a sliding window is initialized, the continuous trajectory sequence is divided into windows of fixed length, a state feature sequence within the window is constructed, and normalization is performed. The sliding window content is updated at each sampling time. When the window length meets the preset requirements, the normalized state feature sequence is input into the trained intent recognition model for forward inference, and the intent category and its probability result of the corresponding window are output, so as to realize real-time recognition with dynamic updates.

[0019] To achieve the above objectives, a second aspect of this application also provides an underwater target intent recognition device based on trajectory sequence modeling, comprising: The multi-intent modeling module is used to construct underwater target multi-behavior intent task scenario models, establish a multi-intent rule base including convergence, inspection, avoidance and departure, and extract trajectory features corresponding to various intents based on the target kinematic model; The trajectory generation module is used to construct an intent-driven Markov decision process model based on reinforcement learning methods, design a differentiated reward function according to the multi-intent rule base, and generate multi-type underwater target motion trajectory data in combination with the proximal policy optimization algorithm. The data preprocessing module is used to normalize and segment the generated trajectory data to construct a sample dataset for model training. The model building module is used to build a long short-term memory network model that integrates an attention mechanism. The model is trained using the sample dataset. The model performs feature extraction and temporal modeling on trajectory sequences to achieve target intent classification. The real-time recognition module uses a sliding window mechanism to segment the real-time acquired target motion trajectory and input it into the trained model, outputting the real-time recognition result of the underwater target's intent.

[0020] To achieve the above objectives, a third aspect of this application also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when executed by the processor, the computer program implements an underwater target intent recognition method based on trajectory sequence modeling as provided above.

[0021] To achieve the above objectives, a fourth aspect of this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the underwater target intent recognition method based on trajectory sequence modeling as provided above.

[0022] To achieve the above objectives, the fifth aspect of this application also provides a computer program product, including a computer program that, when executed by a processor, implements an underwater target intent recognition method based on trajectory sequence modeling as provided above.

[0023] This application discloses an underwater target intent recognition method, device, electronic device, storage medium, and program based on trajectory sequence modeling. The method includes: constructing an underwater target multi-behavior intent task scenario model, establishing a multi-intent rule base including convergence, inspection, avoidance, and departure, and extracting trajectory features corresponding to various intents based on the target kinematic model; constructing an intent-driven Markov decision process model based on reinforcement learning, designing a differentiated reward function according to the multi-intent rule base, and generating multi-type underwater target motion trajectory data by combining a proximal policy optimization algorithm; normalizing and segmenting the generated trajectory data to construct a sample dataset for model training; constructing a long short-term memory network model with an attention mechanism, training it using the sample dataset, wherein the model performs feature extraction and temporal modeling on the trajectory sequence to achieve target intent classification; and using a sliding window mechanism to segment the real-time acquired target motion trajectories into the trained model and outputting the real-time recognition result of the underwater target intent. This application constructs a unified modeling framework for multiple intentions, including convergence, inspection, avoidance, and departure, thereby achieving a standardized expression of complex underwater behaviors and their key trajectory features and enhancing the descriptive capabilities in multi-intention scenarios. It introduces a trajectory generation method based on reinforcement learning, which can automatically generate multi-type and multi-distributed target motion data, effectively alleviating the difficulty of acquiring real underwater data and improving the data support capability and generalization performance of model training. Furthermore, it employs a long short-term memory network model with an attention mechanism, combined with a sliding window strategy, to achieve effective extraction and real-time processing of key features of trajectory sequences, thus achieving high recognition accuracy and good real-time performance in complex underwater environments. Attached Figure Description

[0024] Figure 1 Schematic diagram provided for embodiments of this application; Figure 2 A schematic diagram of a task scenario modeling provided in an embodiment of this application; Figure 3 A schematic diagram of a trajectory generation network provided in an embodiment of this application; Figure 4 A schematic diagram of an intent recognition network provided in an embodiment of this application; Figure 5 The first simulation result diagram provided for the embodiments of this application; Figure 6 The second simulation result diagram provided for the embodiments of this application; Figure 7 A method flowchart is provided for the embodiments of this application.

[0025] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] Traditional underwater target intent recognition techniques typically include behavioral feature extraction, state-action mapping establishment, and intent inference. However, due to the difficulty in obtaining trajectory data and its presence of noise, as well as the complex temporal correlations among multiple intent behaviors, traditional rule-based or shallow learning methods have limited performance in handling complex trajectories and dynamic behavioral patterns. Furthermore, the difficulty in acquiring real underwater multi-intent trajectory data restricts the effective application of data-driven methods; and underwater target intent recognition needs to balance recognition accuracy and real-time performance, which existing methods struggle to achieve.

[0027] refer to Figure 1 and Figure 7 The first embodiment of this application provides an underwater target intent recognition method based on trajectory sequence modeling to solve the technical problems mentioned in the background art, such as the strong concealment of target behavior, complex maneuvering patterns, intertwining of multiple intent features, low recognition accuracy and insufficient real-time performance of traditional methods in complex underwater environments. This method can be executed by a processor, which can be set in a terminal or server. The execution process of an underwater target intent recognition method based on trajectory sequence modeling can be as follows: Step S101: Construct an underwater target multi-behavior intent task scenario model, establish a multi-intent rule base including convergence, inspection, avoidance and departure, and extract trajectory features corresponding to various intents based on the target kinematic model.

[0028] Specifically, a typical underwater mission scenario was constructed for underwater target intent recognition, combined with... Figure 2 The scene consists of a central observation node, a preset warning boundary, and underwater targets. The central observation node is located at the center of the scene and serves as the origin of the global coordinate system. It is responsible for monitoring and evaluating the motion status of underwater targets. The preset warning area, centered on the central observation node, represents the observation boundary and is used to delineate the safe and areas of concern.

[0029] The establishment of a multi-intent rule base aims to formalize expert experience and physical laws, serving as the foundational framework for intent determination and providing effective reward signals for subsequent reinforcement learning trajectory generation. Based on the target's kinematic model, this rule base utilizes mechanisms such as speed control, heading control, and pre-defined warning zone constraints to achieve differentiated maneuver trajectories under different intents. The agent's two-dimensional kinematic model is as follows:

[0030] To establish a unified underwater target intent recognition model, the kinematics of the target are first defined and characterized. Let the target position vector be , the position of the central observation node be , and the relative distance, relative radial velocity, and relative azimuth difference be defined as:

[0031]

[0032]

[0033] The following provides continuous and systematic mathematical rule descriptions for several types of intents, so that the rule base can be used as a criterion for intent recognition algorithms.

[0034] In one embodiment of this application, trajectory features are extracted based on the target kinematic model, and the relative distance, relative radial velocity, and relative azimuth difference between the target position vector and the observation node position are defined. For convergence intentions, the rule is to continuously approach the observation node and align the course, reduce the relative spacing and make the track angle close; The rule for the inspection intention is to move within the pre-defined circular corridor, which increases the variance of the turning rate and the path curvature. For evasive intent, the rule is that the heading deviation and the frequency of course changes increase, which manifests as a sudden change in trajectory; The rule for the intention to leave is to continuously move away from the center of gravity of the area, and the relative distance scale continuously increases.

[0035] Furthermore, when constructing the convergence intention rule, its essence is to continuously approach and point towards the central observation node, which can be mathematically expressed as:

[0036] The specific description of the inspection intent rule is as follows: The inspection target moves within a predefined ring, with increased angular velocity variance and trajectory curvature, exhibiting circling or localized maneuvers rather than continuous approach or departure. Its mathematical rule can be expressed as:

[0037] The main characteristics of constructing rules for evasion intentions are behavioral mutation and high uncertainty. The mathematical rules can be expressed as follows:

[0038] When constructing the departure intention rule, its main characteristic is that it continuously moves away from the central observation node. The mathematical rule can be expressed as:

[0039] Step S102: Construct an intent-driven Markov decision process model based on reinforcement learning methods, design a differentiated reward function according to a multi-intent rule base, and generate multi-type underwater target motion trajectory data by combining a proximal policy optimization algorithm.

[0040] Specifically, reinforcement learning is used to generate motion trajectories with specific intent characteristics to simulate typical target behavior trajectories driven by different intent rules. The main approach starts from the fundamental theories of reinforcement learning, constructs a trajectory generation model oriented towards underwater intent, and designs a constraint-compliant trajectory generation strategy based on a proximal policy optimization algorithm. This chapter first introduces the basic framework of reinforcement learning, Markov decision processes, and related policy optimization methods.

[0041] In reinforcement learning, the reward function determines the behavioral tendency of the agent and is the core mechanism that guides it to form a specific intention trajectory. In order to enable the agent to learn different types of maneuvering patterns in a unified environment model, this invention designs a directional and discriminative reward structure for different intention types, so that intention preferences are clearly reflected through reward signals during the training process.

[0042] In one embodiment of this application, a state space containing position, speed, and heading information is constructed, and an action space for heading changes and speed adjustments is defined; The differentiated reward functions include: for convergence intentions, reinforcing approach behavior, penalizing departure behavior, and encouraging heading toward the target direction; for inspection intentions, rewarding movement within the loop area and encouraging circular circling around the target; for avoidance intentions, providing incentives based on the threshold range of the turning amplitude; and for departure intentions, rewarding increasing the relative distance and constraining heading deviation.

[0043] Furthermore, the core objective of convergence intent is to continuously move the agent closer to the central observation node. Therefore, the design principle of the reward function is to reinforce approaching behavior, punish distancing behavior, and encourage maneuvers towards the target. The core logic of the reward is:

[0044] Among them, This indicates the Euclidean distance between the enemy target and our side. This indicates the distance of the enemy target from the red line. This indicates the clamp formed by the line connecting the enemy target's direction of movement and our target.

[0045] The reward function for the inspection intention should emphasize maintaining distance stability and the movement trend along the boundary. The core logic of the reward is:

[0046] in This indicates that entering the red line is subject to penalty based on the agent. This means that when the agent goes outside the ring, a penalty is imposed, and the distance term in the reward encourages the agent to stay within the ring area, including the heading term. Encourage intelligent agents to circle around the target.

[0047] The reward function for evasion intent should emphasize the magnitude of course changes, maneuver frequency, and trajectory complexity. The core logic of the reward is:

[0048] in, This indicates the change in the current steering angle; its absolute value represents the steering amplitude. This represents the minimum threshold for steering angle. This indicates the maximum threshold for steering angle. This represents the median value for moderate steering angle. The scaling factor represents the intensity of the excitation.

[0049] The key characteristic of the intention to leave is that the target's movement behavior takes continuously increasing the relative distance from the central observation node as its core objective, and the core logic of the reward is:

[0050] in, The changing trend of the distance between the target and our side is the core criterion for escape behavior; the heading deviation term Stability used to constrain the escape direction; safety boundary penalty term It is used to reflect the severity of the situation when a target enters a high-risk area.

[0051] After designing a differentiated reward function, a trajectory generation network was designed in conjunction with the Actor-Critic structure, as illustrated in the diagram below. Figure 3 As shown.

[0052] Step S103: Normalize and segment the generated trajectory data to construct a sample dataset for model training.

[0053] In one embodiment of this application, the generated trajectory data is stored in CSV format, including time step number, position coordinates, motion speed, heading angle, velocity change rate, and angular velocity change rate; The trajectory data is normalized and divided into sequences according to the time step order to generate a fixed-length time series sample dataset.

[0054] Specifically, the dataset used in this invention takes underwater target motion trajectory data as the research object and stores it in CSV format. The specific format includes: time step number, position coordinates, motion speed, heading angle, velocity change rate, and angular velocity change rate.

[0055] Step S104: Construct a long short-term memory network model that integrates attention mechanism, train it using sample dataset, and perform feature extraction and temporal modeling on trajectory sequences to achieve target intent classification.

[0056] Specifically, the original trajectory data undergoes systematic preprocessing, and an attention mechanism is incorporated to optimize the LSTM intent recognition network. In underwater target intent recognition tasks, simply relying on traditional time-series models to recursively model historical information is insufficient to fully characterize the differentiated contributions of different behavioral stages to intent judgment. To enhance the model's ability to focus on key behavioral segments and discriminative features, this invention introduces an attention mechanism into the LSTM framework. By explicitly weighting the historical trajectory states, the network is guided to adaptively focus on the moments and feature dimensions that have a significant impact on intent inference.

[0057] In one embodiment of this application, the long short-term memory network model that integrates attention mechanism uses the long short-term memory network as the base network to extract the temporal features of the trajectory sequence; In the attention mechanism, the features of the previous state at the current moment are used as the query vector. The relevance of the key vectors in the historical state sequence is measured to obtain the attention weight. The value vectors of the historical state sequence are weighted and summed to obtain the global context representation. The intention is then classified by combining the hidden state of the Long Short-Term Memory network.

[0058] Furthermore, this mechanism no longer relies solely on the recursive results of the LSTM hidden states. Instead, it uses the features of the previous state at the current time step as the query to measure the relevance of the keys in the historical state sequence, thereby adaptively allocating different attention weights. The specific network structure is as follows: Figure 4 As shown.

[0059] Step S105: The real-time collected target motion trajectory is segmented and input into the trained model using a sliding window mechanism, and the real-time recognition result of the underwater target's intent is output.

[0060] Specifically, the sliding window technique is used to segment the preprocessed trajectory data, dividing the long-term sequence into several fixed-length subsequences, extracting the corresponding motion features within each subsequence, and realizing real-time recognition of trajectory intent.

[0061] In one embodiment of this application, a sliding window is initialized, the continuous trajectory sequence is divided into windows of fixed length, a state feature sequence within the window is constructed, and normalization processing is performed. The sliding window content is updated at each sampling time. When the window length meets the preset requirements, the normalized state feature sequence is input into the trained intent recognition model for forward inference, and the intent category and its probability result of the corresponding window are output, so as to realize real-time recognition with dynamic updates.

[0062] Furthermore, firstly, the sliding window is initialized, and a sequence of state features within the window is constructed:

[0063] right Normalization is performed:

[0064] Will Inputting the LSTM model for forward inference:

[0065] Determine the intent category of the current window:

[0066] The recognition result for output window k:

[0067] The algorithm is driven by time steps, updating the content of the sliding window at each sampling time. When the window length meets the preset requirements, a fixed-length state feature sequence is constructed and the trained LSTM intent recognition model is called to complete the inference, thereby outputting the intent category and probability result of the corresponding window. Subsequently, by inputting a trajectory sequence and setting a sliding window, the trajectory is identified in real time. The simulation results are shown in the figure below. Figure 5 and Figure 6 As shown.

[0068] To address the aforementioned technical problems, the second embodiment of this application provides an underwater target intent recognition device based on trajectory sequence modeling to solve the same technical problems as the method embodiment. The device may include the following modules: a multi-intent modeling module, a trajectory generation module, a data preprocessing module, a model building module, and a real-time recognition module.

[0069] Among them, the multi-intent modeling module is used to construct a multi-behavioral intent task scenario model for underwater targets, establish a multi-intent rule base including convergence, inspection, avoidance and departure, and extract trajectory features corresponding to various intents based on the target kinematic model; The trajectory generation module is used to construct an intent-driven Markov decision process model based on reinforcement learning methods, design a differentiated reward function according to a multi-intent rule base, and generate multi-type underwater target motion trajectory data by combining a proximal policy optimization algorithm. The data preprocessing module is used to normalize and segment the generated trajectory data to construct a sample dataset for model training. The model building module is used to build a long short-term memory network model that integrates an attention mechanism. It is trained using a sample dataset. The model performs feature extraction and temporal modeling on trajectory sequences to achieve target intent classification. The real-time recognition module uses a sliding window mechanism to segment the real-time acquired target motion trajectory and input it into the trained model, outputting the real-time recognition result of the underwater target's intent.

[0070] It is understood that the above-described device embodiments correspond to the method embodiments. Therefore, the device embodiments have all the beneficial effects of the method embodiments, and will not be repeated here.

[0071] To address the aforementioned technical problems, the third embodiment of this application also provides an electronic device, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements an underwater target intent recognition method based on trajectory sequence modeling as provided in the preceding embodiments.

[0072] To address the aforementioned technical problems, the fourth embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements an underwater target intent recognition method based on trajectory sequence modeling as provided in the preceding embodiments.

[0073] To address the aforementioned technical problems, the fifth embodiment of this application also provides a computer program product, including a computer program that, when executed by a processor, implements an underwater target intent recognition method based on trajectory sequence modeling as provided in the preceding embodiments.

[0074] It should be noted that in this article, relations such as first and second are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0075] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for underwater target intent recognition based on trajectory sequence modeling, characterized in that, include: Construct a multi-behavior intent task scenario model for underwater targets, establish a multi-intent rule base including convergence, inspection, avoidance, and departure, and extract trajectory features corresponding to various intents based on the target kinematic model; A Markov decision process model driven by intent is constructed based on reinforcement learning methods. A differentiated reward function is designed according to the multi-intent rule base, and multi-type underwater target motion trajectory data is generated by combining the proximal policy optimization algorithm. The generated trajectory data is normalized and sequenced to construct a sample dataset for model training. A long short-term memory network model integrating an attention mechanism is constructed and trained using the sample dataset. The model performs feature extraction and temporal modeling on trajectory sequences to achieve target intent classification. A sliding window mechanism is used to segment the real-time acquired target motion trajectory and input it into the trained model, outputting the real-time recognition result of the underwater target's intent.

2. The underwater target intent recognition method based on trajectory sequence modeling as described in claim 1, characterized in that, Trajectory features are extracted based on the target kinematic model, and the relative distance, relative radial velocity, and relative azimuth difference between the target position vector and our position are defined. For convergence intentions, the rule is to continuously approach the observation node and align the course, reduce the relative spacing and make the track angle close; The rule for the inspection intention is to move within the pre-defined circular corridor, which increases the variance of the turning rate and the path curvature. For evasive intent, the rule is that the heading deviation and the frequency of course changes increase, which manifests as a sudden change in trajectory; The rule for the intention to leave is to continuously move away from the center of gravity of the area, and the relative distance scale continuously increases.

3. The underwater target intent recognition method based on trajectory sequence modeling as described in claim 1, characterized in that, Construct a state space containing position, velocity, and heading information, and define the action space for heading changes and velocity adjustments; The differentiated reward function includes: for attack intent, reinforcing approach behavior, punishing away behavior, and encouraging heading towards the target; for reconnaissance intent, rewarding movement within the loop area and encouraging circling around the target; for deception intent, providing incentives based on a threshold range of turning amplitude; and for escape intent, rewarding increasing relative distance and constraining heading deviation.

4. The underwater target intent recognition method based on trajectory sequence modeling as described in claim 1, characterized in that, The generated trajectory data is stored in CSV format, including time step number, position coordinates, motion speed, heading angle, rate of change of velocity, and rate of change of angular velocity. The trajectory data is normalized and divided into sequences according to the time step order to generate a fixed-length time-series sample dataset.

5. The underwater target intent recognition method based on trajectory sequence modeling as described in claim 1, characterized in that, The long short-term memory network model that integrates attention mechanisms uses a long short-term memory network as the base network to extract temporal features of trajectory sequences; In the attention mechanism, the features of the previous state at the current moment are used as the query vector. The relevance of the key vectors in the historical state sequence is measured to obtain the attention weight. The value vectors of the historical state sequence are weighted and summed to obtain the global context representation. The intention is then classified by combining the hidden state of the Long Short-Term Memory network.

6. The underwater target intent recognition method based on trajectory sequence modeling as described in claim 1, characterized in that, Initialize the sliding window, divide the continuous trajectory sequence into fixed-length windows, construct the state feature sequence within the window and perform normalization processing; The sliding window content is updated at each sampling time. When the window length meets the preset requirements, the normalized state feature sequence is input into the trained intent recognition model for forward inference, and the intent category and its probability result of the corresponding window are output, so as to realize real-time recognition with dynamic updates.

7. An underwater target intent recognition device based on trajectory sequence modeling, characterized in that, include: The multi-intent modeling module is used to construct underwater target multi-behavior intent task scenario models, establish a multi-intent rule base including convergence, inspection, avoidance and departure, and extract trajectory features corresponding to various intents based on the target kinematic model; The trajectory generation module is used to construct an intent-driven Markov decision process model based on reinforcement learning methods, design a differentiated reward function according to the multi-intent rule base, and generate multi-type underwater target motion trajectory data in combination with the proximal policy optimization algorithm. The data preprocessing module is used to normalize and segment the generated trajectory data to construct a sample dataset for model training. The model building module is used to build a long short-term memory network model that integrates an attention mechanism. The model is trained using the sample dataset. The model performs feature extraction and temporal modeling on trajectory sequences to achieve target intent classification. The real-time recognition module uses a sliding window mechanism to segment the real-time acquired target motion trajectory and input it into the trained model, outputting the real-time recognition result of the underwater target's intent.

8. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the underwater target intent recognition method based on trajectory sequence modeling as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the underwater target intent recognition method based on trajectory sequence modeling as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The method includes a computer program that, when executed by a processor, implements the underwater target intent recognition method based on trajectory sequence modeling as described in any one of claims 1 to 6.

Images