Underwater unmanned vehicle state reconstruction method under incomplete information

Abstract

The invention relates to an underwater unmanned vehicle state reconstruction method under incomplete information. The method comprises the following steps: obtaining multi-source sensor data of an underwater unmanned vehicle, and carrying out space-time alignment and data fusion; constructing a state prediction model, taking environment characteristics and a sensor sequence as input, introducing physical constraint loss and a dynamic forgetting mechanism based on a marine environment, and predicting a motion state; state information missing detection and classification are carried out by using a multi-source evidence fusion missing detection model and a Bayesian network-based missing reason classification method in combination with a dynamic threshold adaptive mechanism; constructing an adaptive state space model of the marine environment, adopting an adaptive reconstruction strategy according to a missing type, and realizing state reconstruction under incomplete information in combination with a layered reconstruction framework and marine dynamics constraint optimization; and carrying out reliability verification on the reconstruction result, and if the verification is not passed, carrying out feedback optimization through reinforcement learning. Compared with the prior art, the state reconstruction problem of the UUV in a complex marine environment and under the condition of information loss is solved, and the reconstruction precision, robustness and adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of underwater unmanned vehicle state reconstruction technology, and in particular to a method for underwater unmanned vehicle state reconstruction under incomplete information. Background Technology

[0002] Unmanned underwater vehicles (UUVs) are increasingly important as key equipment in modern ocean exploration, resource development, and military applications. In actual operation, UUVs need to accurately acquire their own position, attitude, velocity, and other status information to ensure the completion of complex tasks such as seabed topography mapping, target search, and hydrological data collection. However, due to the complex marine environment and limitations of their own systems, UUV status information is frequently lost intermittently.

[0003] The marine environment is extremely complex. The strong corrosiveness of seawater, varying water pressure at different depths, complex currents, and variable hydrological conditions can all interfere with the normal operation of various sensors on UUVs. For example, in deep-sea areas, water pressure can cause deformation of the sensor housing, affecting its measurement accuracy; while in areas with strong currents, UUVs may be subjected to additional impact forces, causing fluctuations or interruptions in the sensor output signal. Simultaneously, the electromagnetic environment in the ocean is also complex, with various natural and man-made electromagnetic fields. For sensors that rely on electromagnetic signals to transmit data, this can easily cause signal attenuation, distortion, or even loss, resulting in incomplete UUV status information.

[0004] From the perspective of the UUV's own system, it carries a wide variety of sensors, including an inertial measurement unit (IMU), a global positioning system (GPS), a Doppler velocimeter (DVL), and a depth sensor. Different sensors have their own measurement principles and accuracy characteristics, and may face problems such as communication bandwidth limitations and incompatible data transmission protocols during data transmission. For example, GPS cannot be used directly underwater; when the UUV surfaces to receive GPS signals, signal interruptions may occur due to obstructions from waves, clouds, or fog. While the IMU can provide attitude and acceleration information in real time, cumulative errors accumulate over time. The accuracy of the DVL is affected in shallow waters or areas with complex seabed topography. Furthermore, if the data processing unit inside the UUV has insufficient processing power when faced with large amounts of complex sensor data, some status information may be lost during processing.

[0005] In previous research and applications, traditional methods for addressing the problem of UUV status information loss mainly relied on redundant backups of single sensors or simple data interpolation. For example, multiple sensors of the same type were used, and when one sensor failed or lost data, other sensors were switched on to continue operating. However, this approach not only increased hardware costs and system complexity, but also failed to effectively solve the information loss problem when multiple sensors were simultaneously affected by environmental interference. Simple data interpolation methods, such as linear interpolation, only estimated based on known data from adjacent time points, resulting in low reconstruction accuracy. Chinese patent CN118094870A discloses a UUV collaborative information reconstruction system and method based on minimum KL divergence. This method improves position accuracy by establishing a kinematic model, using Lagrange interpolation to smooth position prediction, and employing the minimum KL divergence criterion for high-precision fusion of position prediction and measurement information. However, this method does not fully consider the dynamic characteristics of UUV movement and the influence of complex marine environments, resulting in low reconstruction accuracy and failing to meet the needs of high-precision UUV status perception in practical applications.

[0006] With the continuous expansion of UUV application scenarios and the increasing demands of tasks, it is urgent to develop a method that can accurately reconstruct the state of UUVs under incomplete information. Summary of the Invention

[0007] The purpose of this invention is to provide a state reconstruction method for underwater unmanned vehicles (UUVs) under incomplete information, which solves the problem of state reconstruction when UUVs continuously lose state information such as position, attitude, and velocity due to various factors in complex marine environments, improves the accuracy and reliability of UUV state perception, and ensures its safety and efficiency during mission execution.

[0008] The objective of this invention can be achieved through the following technical solutions: A method for reconstructing the state of an underwater unmanned vehicle under incomplete information includes the following steps: Multi-source data fusion preprocessing: Acquire multi-source sensor data from underwater unmanned vehicles and perform spatiotemporal alignment and high-precision data fusion; State prediction model construction: A state prediction model integrating marine environmental features is constructed, with environmental features and sensor sequences as inputs. The environmental features include ocean current velocity, water pressure, and seabed topography slope. Physical constraint loss and a dynamic forgetting mechanism based on the marine environment are introduced to achieve high-precision prediction of the motion state of underwater unmanned vehicles. State information missing detection and classification: Using a multi-source evidence fusion missing detection model and a Bayesian network-based missing cause classification method, combined with a dynamic threshold adaptive mechanism, state information missing detection and classification is performed. Missing types include hardware fault missing, communication interruption missing, and data processing error missing. State Reconstruction: Construct an adaptive state-space model for the marine environment, adopt an adaptive reconstruction strategy based on the missing type, and combine a hierarchical reconstruction framework with marine dynamics constraint optimization to achieve state reconstruction under incomplete information; Reconstruction result verification and feedback optimization: The reliability of the reconstruction result is verified. If the verification fails, feedback optimization is performed through reinforcement learning.

[0009] The loss function of the state prediction model is: in, The total loss of the state prediction model. For mean square error loss, For physical constraint loss, For incompressible constraint loss, , These are the weighting coefficients. For the sample size, For the first i The predicted state of each sample is a vector consisting of position, attitude, velocity, and angular velocity. For the first i The true state of each sample The equations for UUV dynamics are as follows: For the first i The derivative of the predicted state of each sample. For the first i Control input variables for each sample , , For the first i Partial derivatives of the velocity components along the x, y, and z axes for each sample.

[0010] The dynamic forgetting mechanism based on the marine environment specifically includes: Forgotten Gate weight update: in, For dynamic forget gate output, It is the sigmoid activation function. , Let be the weight matrix of the hidden state at time t and time t-1 for the forget gate. Let be the hidden state at time t. Let be the weight matrix of the cell state at time t-1 to the forgetting gate. The cell state at time t-1. This is the weight matrix of the high-dimensional environmental feature vectors to the forgetting gate. For high-dimensional environmental feature vectors, For the bias term of the forget gate; Define environmental complexity index : in, The ocean current velocity gradient, The rate of change of water pressure The slope of the terrain; when At the same time, increase the update frequency of the forget gate to accelerate the forgetting of old states: in, This represents the environmental complexity threshold. This is the adjustment coefficient.

[0011] In the missing state information detection and classification, a missing detection model using multi-source evidence fusion is employed for missing information detection, specifically including: Define a time continuity index for sensor data sequences. for: in, For the previous valid data time, For the current moment, The nominal sampling period of the sensor. This is the adjustment coefficient; Constructing a spatial consistency index by utilizing the physical constraints between sensor data : in, The predicted velocity is calculated using position difference. Let be the position at time t. For time intervals, For DVL speed measurement, For the variance of velocity measurement; Based on the combined temporal and spatial detection results, a joint detection index is defined. : Among them, weight Dynamically adjust based on sensor type; When satisfied , and If any of the following conditions are met, it is determined that information is missing. The threshold for time continuity. This is the spatial consistency threshold. This is the threshold for joint detection.

[0012] In the state information missing detection and classification, the dynamic threshold adaptive mechanism is specifically as follows: The time continuity threshold is adaptively adjusted as follows: in, As a time continuity benchmark threshold, p The current water pressure, For reference water pressure, This is the water pressure influence coefficient. The electromagnetic interference influence coefficient. I Electromagnetic interference intensity, The adaptively adjusted time continuity threshold; The spatial consistency threshold is adaptively adjusted as follows: in, As a spatial consistency benchmark threshold, The influence coefficient of ocean currents. For ocean current speed, The adaptively adjusted spatial consistency threshold; The joint detection threshold is adaptively adjusted as follows: The sensor health index is defined by dynamically calibrating threshold parameters based on the sensor's historical health status. : in, For historical data windows, Let be the standard deviation of the k-th measurement. This is the value measured by the sensor for the kth time. Measure the average value of the sensor; When the sensor health index is less than the sensor health index threshold ,in, As a joint detection benchmark threshold, The adaptively adjusted joint detection threshold, For data augmentation coefficients, , The threshold for the sensor health index. These are preset coefficients.

[0013] In the state reconstruction, the hierarchical reconstruction framework includes a physical model layer, a data-driven layer, and a decision fusion layer. The physical model layer uses a UUV dynamics model to extrapolate the state and predict short-term missing values. The data-driven layer uses a trained Bi-Attention LSTM model to reconstruct long-term missing values. The decision fusion layer calculates the confidence scores of the reconstruction results from the physical model layer and the data-driven layer to determine the fusion weights. The reconstruction results from the two layers are then fused to obtain the final reconstructed state.

[0014] In the state reconstruction, ocean dynamics constraint optimization is performed based on a manifold optimization objective function, wherein the manifold optimization objective function is: in, Let k be the UUV state vector at time k. This represents the final reconstructed state after fusion. These are the weighting coefficients. Represents attitude manifold constraints, , Indicates velocity boundary constraints, , The time window length, Let be the attitude angle of the UUV at time k. The attitude angle safety threshold, The linear velocity vector of the UUV. This represents the maximum linear velocity of the UUV.

[0015] In the state reconstruction, the adaptive reconstruction strategy based on the missing type is specifically as follows: For hardware failures, DVL speed integral is used to assist in reconstruction: At the same time, the weight of IMU data fusion is reduced, while the weight of DVL and depth sensor fusion is increased. For the reconstructed location estimation, This refers to the position at the previous moment. Size of the integration window For DVL speed, For time intervals; For communication interruptions, terrain matching is used to assist in reconstruction: the terrain height map of the current area is obtained from a pre-stored seabed terrain database. Based on depth sensor data With UUV draft Calculate possible locations Position estimation is optimized through particle filtering. : For data processing errors and missing information, extrapolation using a dynamic model and historical states is employed: in, For the reconstruction speed estimation, For UUV speed, To control the input, The Coriolis force matrix, Here is the damping matrix. This is the UUV quality matrix.

[0016] In the reconstruction result verification and feedback optimization, the reliability verification includes physical domain verification and data domain verification. The physical domain verification includes dynamic consistency testing and kinematic loop closure testing, while the data domain verification includes sensor fusion consistency verification and deep learning model confidence evaluation. The dynamic consistency check is based on the reconfiguration state and calculates the theoretical acceleration using the dynamic equations. Based on theoretical acceleration and IMU measurement of acceleration Calculate the dynamic consistency deviation index : in, To prevent parameters with a denominator of zero, the verification is deemed unsuccessful when the dynamic consistency deviation index exceeds the dynamic deviation threshold. The kinematic closed-loop test calculates the kinematic deviation index between the velocity integral-reconstructed position and the directly reconstructed position. : in, To reconstruct the position by integrating the velocity, , To directly reconstruct the location, For the integration window, For the maximum speed of UUV, For time intervals, The speed of reconstruction; when the kinematic deviation index is greater than the kinematic deviation threshold, the verification is deemed unsuccessful; The sensor fusion consistency verification utilizes cross-validation with complete sensor data: the sensor fusion consistency index is calculated by comparing the velocity measured by DVL with the velocity obtained by integrating the acceleration through IMU. : in, Accelerating the IMU The variance of historical speed measurements is used; if the sensor fusion consistency index is lower than the fusion consistency threshold, the verification is deemed unsuccessful. The confidence evaluation of the deep learning model utilizes the output distribution entropy of the state prediction model. As a confidence index, if the confidence index is less than the confidence threshold, the verification is deemed unsuccessful.

[0017] In the reconstruction result verification and feedback optimization, a reinforcement learning method is adopted to model the reconstruction verification process as a reinforcement learning problem, and the reconstruction parameters are dynamically adjusted by the agent: The state space is defined as follows: ,in, One-hot encoding for missing types, Environmental complexity is defined as: in, , , These are ocean current velocity, water pressure, and topographic slope, respectively. , , These are reference values ​​for ocean current velocity, water pressure, and terrain slope, respectively. Action space is defined as: , These represent the adjustment amounts for the physical model weights, the dynamic constraint weights in manifold optimization, the kinematic constraint weights, and the data augmentation coefficients, respectively. The reward function is designed as follows: ,in, These are the weighting coefficients.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses multi-source data fusion preprocessing to perform high-precision alignment and fusion of sensor data from different sampling rates and spatiotemporal references, providing a high-quality data foundation for state reconstruction. More importantly, the state prediction model innovatively incorporates real-time marine environmental features such as ocean current velocity, water pressure, and seabed topography slope, and introduces physical constraint loss, so that the prediction results not only rely on historical data, but also strictly follow the dynamic laws of underwater vehicles and their physical environment. This data-driven and physical model-deeply integrated approach fundamentally solves the problem of insufficient accuracy caused by neglecting the dynamic characteristics of the environment and physical constraints in existing technologies, and significantly improves the accuracy and reliability of state estimation.

[0019] (2) This invention enables state prediction to respond in real time to the influence of external water flow by directly using ocean current velocity as model input; it overcomes the cumulative error of sensors by using water pressure data to assist depth calibration; and it introduces seabed topographic slope information to provide key topographic reference for underwater positioning in GPS-free environments. This design of actively sensing and adapting to the environment greatly improves the robustness of state reconstruction of underwater unmanned vehicles under harsh conditions such as strong currents and complex seabed topography.

[0020] (3) This invention first uses a multi-source evidence fusion model for missing data detection. Then, it uses a Bayesian network to classify the causes of missing data, accurately distinguishing different types such as hardware failure, communication interruption, or data processing error. Finally, it triggers corresponding adaptive reconstruction strategies based on the diagnostic results. For example, in the case of hardware failure, DVL integration is used and sensor fusion weights are adjusted; in the case of communication interruption, pre-stored seabed maps are used for terrain matching and positioning, thereby achieving accurate and efficient recovery of missing data.

[0021] (4) This invention designs a physical constraint loss term, taking the physical laws such as the dynamic equations and attitude representations of UUVs as optimization objectives, forcing the reconstruction results to conform to physical reality, and avoiding the generation of mathematically correct but physically impossible states. In addition, a dynamic forgetting mechanism is introduced, which can adjust the model's memory cycle according to environmental characteristics, enabling it to forget irrelevant historical information more quickly in complex environments. This fusion ensures that the reconstruction results not only fit the data, but also follow natural laws.

[0022] (5) For each reconstruction result, the present invention performs reliability verification. If the verification fails, a reinforcement learning-based feedback optimization mechanism is triggered to automatically adjust the model parameters or reconstruction strategy. This makes the system no longer a static algorithm, but an intelligent agent with online learning and continuous improvement capabilities. Through long-term operation, it can continuously learn from successful and unsuccessful experiences, thus becoming more and more accurate for specific sea areas or specific aircraft platforms, demonstrating a powerful self-evolution capability.

[0023] (6) The robust handling strategies of this invention for different types of missing data ensure that the system can maintain basic functions even when partial failures occur, greatly enhancing its survivability and mission success rate in actual tasks. In addition, this technical framework does not depend on a specific type of vehicle, has good versatility and scalability, and is easy to integrate and promote on different types of autonomous underwater vehicles, with broad military, scientific research and commercial prospects. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0026] This embodiment provides a method for reconstructing the state of an underwater unmanned vehicle under incomplete information, such as... Figure 1 As shown, it includes the following steps: S1, Multi-source data fusion preprocessing: Acquire multi-source sensor data from underwater unmanned vehicles and perform spatiotemporal alignment and high-precision data fusion.

[0027] In the UUV hardware system design, standardized interfaces are used to connect the sensors and the data processing unit. The inertial measurement unit (IMU) is connected to the data processing unit via the SPI (Serial Peripheral Interface) to ensure stable transmission of high-frequency data (100-1000Hz); the Doppler velocimeter (DVL) uses the RS-422 serial port protocol to transmit velocity data at a frequency of 10-50Hz; the depth sensor connects via I... 2 The UUV uses a C-bus connection to transmit depth information at approximately 50Hz; the GPS module provides location data via a UART serial port at 1Hz when the UUV surfaces. The data processing unit employs a high-performance embedded processor (such as the NVIDIA Jetson AGX Xavier), possessing sufficient computing power and multi-interface support, enabling it to simultaneously receive and process data from multiple sensor sources.

[0028] To address the issues of UUV sensor data susceptibility to interference, transmission delays, and time asynchrony in complex marine environments, this embodiment achieves high-precision preprocessing of multi-source heterogeneous data by constructing an adaptive weighted fusion model and a spatiotemporal calibration mechanism. Specific technical details are as follows: 1.1 Multi-sensor spatiotemporal alignment mechanism.

[0029] The sampling frequency of the IMU on a UUV is typically 100-1000Hz, DVL is 10-50Hz, GPS is only effective on the water surface and has an update rate of 1Hz, and the sampling frequency of the depth sensor is about 50Hz. Clock drift and transmission delay of different sensors will cause data timestamp deviations, and a dynamic calibration model needs to be established.

[0030] A high-precision clock chip (such as a DS3231 chip with a crystal oscillator frequency of 10MHz) is installed inside the UUV as a reference clock, with a frequency stability of ±2ppm. When each sensor collects data, it timestamps the data according to the synchronization signal of the reference clock.

[0031] First, hardware clock drift compensation is performed. This embodiment uses a second-order polynomial to fit the clock skew model: in, Due to clock skew, For polynomial coefficients, The noise is Gaussian white noise. In practice, clock skew data is collected every 100ms, and the least squares method is used to analyze the polynomial coefficients. Update.

[0032] By synchronously acquiring reference clock and sensor clock signals, and using the least squares method to estimate coefficients in real time, the calibration accuracy can reach ±10μs.

[0033] Next, dynamic modeling of transmission delay is performed. For DVL data transmitted via RS-422 serial port, the transmission delay... ARIMA prediction model is established based on baud rate and data frame length: Where p and q are the model orders, For model parameters, This represents the prediction error. In this embodiment, the model order... Model parameters are determined using the maximum likelihood estimation method. By updating model parameters in real time through a sliding window (window size set to 50), the delay prediction error is controlled within ±0.5ms, ensuring accurate temporal alignment of data from different sensors.

[0034] 1.2 Adaptive weighted fusion model.

[0035] Considering that sensor reliability in the marine environment changes dynamically with factors such as depth and ocean currents, an adaptive weight allocation algorithm based on fuzzy logic is designed.

[0036] 1.2.1 Definition of sensor reliability indicators.

[0037] For IMU accelerometer data, define reliability metrics. for: in, The current acceleration vector, This is the average of the past 10 sampling points. This is the adjustment coefficient. When a UUV encounters a strong ocean current, a sudden change in acceleration can lead to... As the weight decreases, the weight automatically decreases.

[0038] The reliability index of DVL (Doppler velocimeter) can be defined as: in, The speed measured by the Doppler velocimeter. For the smoothed historical speed, This is the sensitivity coefficient for velocity changes.

[0039] For the reliability indicators of other sensors, such as the reliability indicator of depth sensor, it can be determined based on pressure change rate and noise level, and the reliability indicator of magnetometer can be determined based on magnetic field interference and heading jitter. The specific calculation formulas will not be elaborated here, and how they are defined will not affect the achievement of the purpose of this invention.

[0040] 1.2.2 Fuzzy logic weight decision.

[0041] Taking DVL speed measurement as an example, the input is the measured value. With IMU acceleration The deviation of the velocity obtained by integration and ocean current speed The output is the fusion weight. .

[0042] The fuzzy rule design is as follows: If Small and Small, then Large (weight 0.8-1.0); if large and Large, then Small (weight 0.2-0.4). The fuzzy membership function uses a Gaussian form: , Where c is the center value, The width is determined through offline training. Among them, for deviation central value ,width Regarding ocean current speed central value ,width .

[0043] Weighted fusion formula: fused velocity vector for: in, n The number of sensors (such as IMU, DVL, GPS) participating in the fusion. Output weights for fuzzy logic. For sensor reliability indicators, meet This ensures that the fused velocity vector can comprehensively utilize the effective information from each sensor, thereby improving data accuracy.

[0044] 1.3 Sensor deviation compensation under deep-sea high-pressure environment.

[0045] When a UUV descends to a depth of 1000 meters, the water pressure can reach 10 MPa, causing deformation of the IMU's casing and resulting in acceleration deviation. Establish a pressure-deviation mapping model: Where p is the water pressure (unit: MPa) measured by the depth sensor. This is the pressure sensitivity matrix. This is a zero-bias vector. Calibration was performed in a high-pressure test chamber (pressure range 0-15 MPa, step size 1 MPa), yielding the following: By measuring water pressure in real time and inputting it into the model, acceleration deviation can be compensated to ±0.03 m / s². 2 Within this range, it meets the needs of deep-sea navigation.

[0046] S2, State Prediction Model Construction: Construct a state prediction model that integrates marine environmental characteristics. Using environmental characteristics and sensor sequences as inputs, introduce physical constraint loss and a dynamic forgetting mechanism based on the marine environment to achieve high-precision prediction of the motion state of underwater unmanned vehicles.

[0047] 2.1 State prediction model embedded with marine environmental characteristics.

[0048] Network Structure Design: The movement of UUVs in the ocean is affected by environmental factors such as ocean currents, water pressure, and topography. Traditional LSTMs struggle to capture the coupling relationship between environment and motion. This model employs a bidirectional LSTM (Bi-LSTM) combined with an environmental feature embedding layer to construct a state prediction model, including: Environmental characteristics input: including ocean current speed Water pressure p, seabed topography slope It is mapped to a high-dimensional feature vector through the embedding layer: in, The weight matrix ( (for the embedded dimension) For bias vectors, This represents a high-dimensional environmental feature vector.

[0049] Sensor sequence input: including IMU acceleration angular velocity DVL speed The depth d constitutes the input sequence. ,in, For the time step, local features are extracted through convolutional layers: in, For convolution kernel, For bias, It is the ReLU activation function. This refers to the local features extracted from the sensor input sequence, i.e., the hidden state at time t.

[0050] Bidirectional LSTM layer: Forward LSTM and backward LSTM process forward and reverse timing information respectively (e.g. (etc.), capturing the causal relationships of motion states. Capturing the causal relationships of motion states, ultimately outputting... .

[0051] 2.2 Fusion of dynamic attention mechanism and physical constraints.

[0052] 2.2.1 To establish the correlation between environmental features and motion states, an attention mechanism is designed to calculate the influence weights of environmental features on the state at each time step, as follows: Attention score calculation: in, , , This is the weight matrix. For bias, This represents the attention score.

[0053] Normalized weights: in, To influence the weight; Context vector: Finally, connect the fully connected output layer using the formula. Output the state prediction results.

[0054] 2.2.2 Loss function.

[0055] Physical constraint loss function: Traditional MSE loss cannot guarantee that the prediction results conform to UUV dynamic constraints. A physical constraint term is introduced, resulting in dynamic consistency loss: in, For physical constraint loss, For the sample size, For the predicted state (representing position, attitude, velocity, angular velocity). For location, Euler angles, Linear velocity, Angular velocity, For the first i The derivative of the predicted state of each sample. For the first i Control input variables for each sample The UUV dynamic equations are specifically expressed as follows: in, Here is the attitude rotation matrix. Let be the attitude differential matrix. For the quality matrix, The Coriolis force matrix, Here is the damping matrix. To control the input, As a distractor, It is an obliquely symmetric matrix for angular velocity.

[0056] Incompressible constraint loss (for velocity fields): in, For incompressible constraint loss, , , For the first i Partial derivatives of the velocity components along the x, y, and z axes for each sample.

[0057] Assuming seawater is approximately incompressible and the velocity field divergence should be close to zero, the spatial derivative is calculated using the finite difference method. Total loss function: in, The total loss of the state prediction model. For mean square error loss, , , These are the weighting coefficients. For the first i The predicted state of each sample is a vector consisting of position, attitude, velocity, and angular velocity. For the first i The true state of each sample.

[0058] This embodiment uses the Adam optimizer, sets the initial learning rate to 0.001, and employs early stopping. When the loss function value on the validation set no longer decreases for 5 consecutive rounds, training is stopped and the optimal model parameters are saved.

[0059] Model evaluation and optimization. The trained model is evaluated using a test set, and the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination are calculated on the test set. The model's predictive performance is comprehensively evaluated using metrics such as [insert metrics here]. Based on the evaluation results, the model's prediction error in different scenarios is analyzed. If the model is found to perform poorly in certain specific environments (such as areas with strong ocean currents), the model structure or hyperparameters are adjusted, such as increasing the number of neurons in the LSTM layer or adjusting the parameters of the attention mechanism. The model is then retrained and evaluated until its performance meets the design requirements.

[0060] 2.3 Dynamic forgetting mechanism based on marine environment.

[0061] UUVs exhibit significantly different motion patterns in different marine environments (such as deep-sea laminar currents versus shallow-sea turbulent currents), making it difficult for traditional LSTM with fixed forgetting gate parameters to adapt. Therefore, this embodiment designs an environment-aware dynamic forgetting gate, including: 1) Forgotten Gate Weight Update: in, For dynamic forget gate output, It is the sigmoid activation function. , Let be the weight matrix of the hidden state at time t and time t-1 for the forget gate. Let be the hidden state at time t. Let be the weight matrix of the cell state at time t-1 to the forgetting gate. The cell state at time t-1. This is the weight matrix of the high-dimensional environmental feature vectors to the forgetting gate. For high-dimensional environmental feature vectors, This is the bias term for the forget gate.

[0062] 2) Adaptive adjustment of the forgetting factor: Define environmental complexity index : in, The ocean current velocity gradient, The rate of change of water pressure For terrain slope. When (With the threshold set to 0.8), increasing the forget gate update frequency accelerates the forgetting of old states: in, This represents the environmental complexity threshold. This is the adjustment coefficient.

[0063] 2.4 Model Training and Ocean Scene Transfer Learning.

[0064] Data Augmentation and Scene Segmentation: Data on UUV operations in different marine environments was collected, including raw data from sensors and corresponding real-time UUV status information (obtained through high-precision positioning equipment and inertial navigation systems). The data was arranged chronologically, with N = 100 consecutive time steps considered as a sample. Scene segmentation included: 1) Calm sea areas (ocean current velocity) ); 2) Moderate sea state ( ); 3) Areas with strong ocean currents ( To enhance scarce strong current data, virtual samples were generated through temporal stretching, adding Gaussian noise, and dynamic time warping (DTW), ensuring a 1:1:1 ratio of data from various scenarios. The sample set was divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. During the partitioning process, it was ensured that the data in each set could cover the motion states of UUVs in different scenarios, such as calm seas, moderate sea states, and strong current areas, and possess good representativeness. The data was standardized, particularly for sensor data sequences. Using formula Normalization is performed, where and These are the mean and standard deviation of the data, respectively, to improve the training efficiency and generalization ability of the model.

[0065] Pre-training phase: Use data from calm sea areas to train the basic model and learn the basic motion patterns of UUVs.

[0066] Fine-tuning stage: Environmental features are added to the strong ocean current scene data, and the parameters of the environment embedding layer and attention layer are adjusted through gradient backpropagation to preserve the basic motion features.

[0067] S3, State Information Missing Detection and Classification: Using a multi-source evidence fusion missing detection model and a Bayesian network-based missing cause classification method, combined with a dynamic threshold adaptive mechanism, state information missing detection and classification is performed.

[0068] To address the issue of intermittent data loss from UUV sensor data in complex marine environments, this embodiment proposes a missing information detection and classification method that integrates multi-source evidence reasoning and dynamic threshold adaptation. By constructing a sensor reliability Bayesian network, accurate detection and classification of missing information are achieved. Specific technical details are as follows: 3.1 Missing evidence fusion model.

[0069] A spatiotemporal joint detection mechanism needs to be constructed. Missing UUV sensor data may manifest as time-series interruptions or spatial inconsistencies, necessitating the establishment of a spatiotemporal joint detection model. This mainly includes: 1) Time series continuity detection: Detection of sensor data sequences Define time continuity index for: in, For the previous valid data time, For the current moment, This is the sensor's nominal sampling period (e.g., 10ms for an IMU). This is the adjustment coefficient. When ( When the time continuity threshold is set to 0.3, the time series is considered interrupted.

[0070] 2) Spatial consistency detection: Utilizing the physical constraints between sensor data (such as the differential relationship between velocity and position), a spatial consistency index is constructed. : in, The predicted velocity is calculated using position difference. Let be the position at time t. For time intervals, For DVL speed measurement, For the variance of velocity measurement; when ( When the spatial consistency threshold is set to 0.4, spatial dimension inconsistency is determined.

[0071] 3) Joint detection: Based on the combined temporal and spatial detection results, joint detection indicators are defined. : Among them, weight Dynamically adjust based on sensor type (e.g., IMUs prioritize time continuity). DVL focuses more on spatial consistency. ).when ( When the joint detection threshold is set to 0.5, it is determined that there is missing information.

[0072] In this embodiment, when the following conditions are met... , and If any of the following conditions are met, it is determined that information is missing.

[0073] 3.2 Classification of missing causes based on Bayesian networks.

[0074] 1) Bayesian Network Topology. A three-layer Bayesian network is constructed, comprising environment nodes, sensor nodes, and missing type nodes. The main node definitions are as follows: Environment Node: Ocean Current Velocity Water pressure P, electromagnetic interference intensity I; Sensor node: IMU health status DVL health status GPS health status Missing type node: Hardware failure , communication interruption Data processing errors .

[0075] 2) Construction of Conditional Probability Table (CPT).

[0076] Conditional probability of ocean current velocity affecting sensor hardware failure: This indicates that the probability of DVL hardware failure increases significantly under strong ocean currents.

[0077] The conditional probability of communication interruption due to electromagnetic interference: The electromagnetic interference intensity I is calculated by the deviation between the magnetometer measurement and the historical average. in, Standard deviation, This is the magnetometer measurement value at the current time t. This represents the average historical magnetic field strength.

[0078] 3) Posterior probability inference: When missing information is detected, Bayes' theorem is used to calculate the posterior probability of each missing type: in, For missing data types, E represents environmental evidence. S represents sensor state evidence ( The joint tree algorithm is used for efficient reasoning to determine the most likely cause of the missing information.

[0079] 3.3 Dynamic threshold adaptive mechanism.

[0080] Traditional fixed thresholds cannot meet the detection needs of different marine environments, so this embodiment designs an environment-aware dynamic threshold model.

[0081] 1) The time continuity threshold is adaptively adjusted as follows: in, As a time continuity benchmark threshold,p The current water pressure, For reference water pressure, This is the water pressure influence coefficient. The electromagnetic interference influence coefficient. I Electromagnetic interference intensity, This is the adaptively adjusted time continuity threshold. A larger time interval tolerance is allowed under deep-sea high pressure or strong electromagnetic interference conditions.

[0082] 2) The spatial consistency threshold is adaptively adjusted as follows: in, As a spatial consistency benchmark threshold, The influence coefficient of ocean currents. For ocean current speed, This is the adaptively adjusted spatial consistency threshold. During strong ocean currents, the uncertainty in velocity measurement increases, so the spatial consistency requirement should be appropriately reduced.

[0083] 3) The joint detection threshold is adaptively adjusted as follows: The sensor health index is defined by dynamically calibrating threshold parameters based on the sensor's historical health status. : in, For historical data windows, Let be the standard deviation of the k-th measurement. This is the value measured by the sensor for the kth time. Measure the average value of the sensor; When the sensor health index is less than the sensor health index threshold, that is Tighten the detection threshold: ,in, As a joint detection benchmark threshold, The adaptively adjusted joint detection threshold, For data augmentation coefficients, , This is the threshold for the sensor's health index.

[0084] S4, State Reconstruction: Construct an adaptive state-space model for the marine environment, adopt an adaptive reconstruction strategy based on the missing type, and combine a hierarchical reconstruction framework with marine dynamics constraint optimization to achieve state reconstruction under incomplete information.

[0085] To address the challenge of state estimation for UUVs in information-deficient scenarios, this embodiment proposes a hierarchical reconstruction algorithm that integrates physical constraints and deep learning. By constructing an adaptive state-space model for the marine environment and fusing multi-scale features guided by an attention mechanism, high-precision state reconstruction is achieved even with incomplete information. Specific technical details are as follows: 4.1 Adaptive state-space model for marine environment.

[0086] Time-varying parameter state-space modeling: The motion of UUVs in the ocean is affected by time-varying factors such as ocean currents and water pressure, which are difficult to accurately describe using traditional fixed parameter models. An environment-adaptive state-space model is established as follows.

[0087] Equations of state: Wherein, the state vector Control input (Thrust and torque), environmental disturbances Process noise , To control the input matrix, The environmental disturbance matrix. The time-varying state transition matrix. The expression is: in, This is the nominal transition matrix (based on rigid body kinematics). For the environmental disturbance matrix, This is the environmental impact coefficient matrix. This represents a high-dimensional environmental feature vector.

[0088] The corresponding observation equation is: Observation vector Includes valid sensor data, observation matrix Dynamically adjust based on sensor operating status (set to 0 for rows with missing sensors) to observe noise. .

[0089] Initialization of the marine environment adaptive state-space model: During the UUV startup phase, the basic state-space model is initialized based on its design parameters. Initial state vectors are set: Location Attitude determined by GPS positioning or pre-deployed coordinates on the water surface The default setting is horizontal, with a linear velocity. and angular velocity Initialize to 0. Initialize the nominal state transition matrix. Based on the rigid body kinematics equations, for example, in discrete time, the relationship between position and velocity is constructed as follows: in It is a 3×3 identity matrix. The sampling time interval is set to 0.01s, and the attitude and angular velocity transition matrices are constructed according to similar rigid body kinematics principles. Control input matrix. Environmental interference matrix Initialization is performed based on the characteristics of the UUV thruster and the environment. Environmental characteristic data, including ocean current velocity, is acquired in real time. Water pressure p, seabed topography slope And so on, calculated through the environmental feature embedding layer. And according to the formula: Update the state transition matrix to complete model initialization.

[0090] 4.2 Layered Restructuring Framework and Attention Integration.

[0091] This embodiment adopts a three-layer reconstruction architecture design: to address different types of missing information, a layered architecture is constructed that includes a physical model layer, a data-driven layer, and a decision fusion layer.

[0092] At the physical model layer (lower-level reconstruction), state extrapolation is performed using the UUV dynamics model, which is suitable for short-term missing data (<5s). State covariance: in, These are the extrapolated state values ​​for the physical model layer. This is the optimal estimate from the previous moment. This corresponds to the covariance.

[0093] When missing state information is detected, the physical model layer is first initiated for underlying reconstruction. This is based on the optimal estimated state from the previous time step. and control input Using state equations Perform state extrapolation. For example, if a UUV is in The position of the moment is The speed is Then the predicted position at time k It can be obtained through the following calculation: Meanwhile, according to the covariance propagation formula... Update the state covariance, where the process noise covariance is... The adjustment is made dynamically based on environmental complexity, such as appropriately increasing the size in strong ocean current environments. The values ​​of elements related to speed are used to reflect greater uncertainty.

[0094] In the data-driven layer (high-level reconstruction), a pre-trained Bi-Attention LSTM model is used to reconstruct long-term missing data (>5s): in, For the state reconstruction values ​​of the data-driven layer, The sensor sequence before the loss. Based on the current environmental characteristics, the model output includes state predictions such as position and orientation. It is a Bi-Attention LSTM model.

[0095] Specifically, the sensor sequence before the loss ( and current environmental characteristics The input model, through its internal bidirectional LSTM layers, attention fusion layers, and other structures, outputs the predicted state. For example, if location information is missing, the model will predict the location at the time of the missing information based on historical IMU data, DVL data, and current environmental information such as ocean currents and water pressure. .

[0096] In the attention fusion layer, the confidence weights of the reconstruction results from the two layers are calculated to achieve adaptive fusion. The confidence of the physical model is: Trace of covariance matrix The larger the value, the higher the uncertainty in the physical model's predictions. The smaller the value.

[0097] The confidence level of the data model is: in, This represents the prediction error of the data model before the data is missing. This is the actual value. The model's prediction for time k is based on the data before the data loss. The smaller the error, the better. The larger the value.

[0098] Then, calculate the fusion weights. The final reconstruction result is: .

[0099] 4.3 Reconstruction and optimization of ocean dynamic constraints.

[0100] First, state correction is based on manifold optimization. The UUV state space has physical constraints (such as attitude angle singularities and velocity boundaries). A manifold optimization model is designed to correct the reconstruction results, including: 1) Attitude manifold constraints: Euler angle reconstruction results To satisfy the non-singular constraint, construct the loss function: in, Represents attitude manifold constraints, The time window length, Let be the attitude angle of the UUV at time k. This is the attitude angle safety threshold. When... hour, This indicates that the orientation is close to a singular surface and requires correction. Quaternion interpolation is used to correct this, converting the Euler angles into quaternion representations. The corrected quaternions are obtained by interpolation calculations in the quaternion space. Then convert back to Euler angles to ensure the attitude angles satisfy the non-singular constraint. This indicates that the surface is close to a singular surface and needs to be corrected by quaternion interpolation.

[0101] 2) Velocity boundary constraints: Check the reconstructed linear velocity Does it exceed the maximum speed of UUV? ,like Then according to the projection operator The velocity is corrected to satisfy the boundary conditions.

[0102] Therefore, the corresponding loss function is obtained: in, Indicates velocity boundary constraints, The linear velocity vector of the UUV. This represents the maximum linear velocity of the UUV.

[0103] 3) Construct the manifold optimization objective function: in, Let k be the UUV state vector at time k. This represents the final reconstructed state after fusion. These are the weighting coefficients.

[0104] The objective function is solved using gradient descent, and the objective function with respect to the state vector is calculated. gradient: , And update the state vector based on the gradient: The learning rate The process is iteratively updated until the objective function converges, resulting in an optimized reconstructed state that satisfies the dynamic constraints.

[0105] 4.4 Adaptive Reconstruction Strategy Guided by Missing Patterns.

[0106] Differentiated processing based on missing type: Different reconstruction strategies are adopted according to the missing classification results.

[0107] Hardware failure missing ( If an IMU hardware failure causes a loss of acceleration, DVL velocity integral is used to assist in reconstruction. At the same time, the weight of IMU data fusion is reduced (e.g., multiplied by 0.3), while the weight of DVL and depth sensor fusion is increased (multiplied by 1.5). For the reconstructed location estimation, This refers to the position at the previous moment. Size of the integration window For DVL speed, For time intervals.

[0108] Communication interruption missing ( If GPS communication is interrupted, resulting in location loss, terrain matching is used to assist in reconstruction: the terrain elevation map of the current area is obtained from a pre-stored seabed terrain database. Based on depth sensor data With UUV draft Calculate possible locations Position estimation is optimized through particle filtering. : Initialize a set of particles The weight of each particle is calculated based on the likelihood function of terrain matching. The optimal estimated position is obtained through iterative calculations via resampling, prediction, and update steps. .

[0109] Data processing error missing ( If velocity data is missing due to errors in processing, extrapolate using the dynamic model and historical states: in, For the reconstruction speed estimation, For UUV speed, To control the input, The Coriolis force matrix, Here is the damping matrix. This is the UUV quality matrix.

[0110] 4.5 Real-time compensation mechanism for ocean current disturbances.

[0111] When a UUV moves in a strong ocean current, the reconstruction error mainly comes from the disturbance of the ocean current. A real-time compensation model is designed.

[0112] 1) Ocean current velocity field estimation: Using Doppler frequency shift data from DVL, the three-dimensional ocean current velocity is estimated using the least squares method. in, For DVL number i Measurement speed of each beam The velocity of the UUV relative to the seawater. m The number of beams is used. To improve estimation accuracy, a sliding window (window size set to 20) is used to process historical DVL data and update the ocean current velocity estimate in real time.

[0113] 2) State reconstruction with ocean current compensation: Explicitly introducing the influence of ocean currents into the state equations: For position updates, according to the formula When reconstructing the location, the offset caused by ocean currents should be considered. For example, if the ocean current speed is... The speed of the UUV itself is attitude rotation matrix The actual rate of change of position is For velocity reconstruction, the ocean current velocity estimate is used. Reconstructed value of the relative seawater velocity of the UUV Add them together to get the final reconstructed velocity value. This ensures that the reconstruction results better reflect the actual motion state of UUVs in ocean current environments.

[0114] Through ocean current compensation, strong ocean currents ( The position reconstruction error decreased from 4.2m to 1.3m.

[0115] S5, Reconstruction Result Verification and Feedback Optimization: Verify the reliability of the reconstruction result. If the verification fails, use reinforcement learning to optimize the feedback.

[0116] To address the need for reliability verification and continuous optimization of UUV state reconstruction results, this embodiment proposes a closed-loop verification and optimization framework that integrates multi-source heterogeneous evidence and reinforcement learning. By constructing a physical consistency verification model of marine environmental constraints and an adaptive feedback adjustment mechanism, the reconstruction accuracy is dynamically improved. Specific technical details are as follows: 5.1 Consistency verification model for multi-source heterogeneous evidence.

[0117] This embodiment features a physical-data dual-domain verification architecture. The UUV state reconstruction results must simultaneously satisfy physical constraints and data-driven consistency. The dual-domain joint verification model is established as follows.

[0118] 1) Physical domain verification metrics.

[0119] 1.1) Dynamic consistency check: based on the reconfigured state Theoretical acceleration is calculated using the dynamic equations. Based on theoretical acceleration and IMU measurement of acceleration Calculate the dynamic consistency deviation index : in, To prevent parameters with a denominator of zero, when the dynamic consistency deviation index exceeds the dynamic deviation threshold... When the value is set to 0.3, the dynamics are deemed inconsistent, and the verification fails.

[0120] 1.2) Kinematic closed-loop test: Calculate the kinematic deviation index between the velocity integral reconstructed position and the direct reconstructed position. : in, To reconstruct the position by integrating the velocity, , To directly reconstruct the location, For the integration window, For the maximum speed of UUV, For time intervals, The speed of reconstruction; when the kinematic deviation index is greater than the kinematic deviation threshold. If the verification fails, the verification is deemed unsuccessful.

[0121] 2) Data domain validation metrics.

[0122] 2.1) Sensor fusion consistency verification: Cross-validation is performed using data from sensors with no missing data, such as comparing the velocity measured by DVL with the velocity obtained by integrating the acceleration through IMU to calculate the sensor fusion consistency index. : in, Accelerating the IMU The historical speed measurement variance is used; if the sensor fusion consistency index is lower than the fusion consistency threshold, the verification is deemed unsuccessful.

[0123] 2.2) Confidence of deep learning models: Using the output distribution entropy of the Bi-Attention LSTM as a confidence metric: in, The entropy value represents the probability of the state components output by the model; the smaller the entropy value, the higher the confidence level.

[0124] 5.2 Dynamic verification threshold under time-varying environment.

[0125] Traditional fixed thresholds cannot adapt to different marine environments. Therefore, a dynamic threshold model for environmental awareness is designed, including: 1) Adjustment of dynamic deviation threshold: in, As the baseline threshold, Let p be the ocean current velocity gradient, and p be the water pressure. , This is the environmental impact coefficient. Under strong ocean currents or deep-sea high pressure, a greater tolerance for dynamic deviations is allowed.

[0126] 2) Kinematic deviation threshold: in, As the baseline threshold, This represents the influence coefficient of ocean currents. As ocean current velocity increases... This will increase to adapt to the greater uncertainty in the kinematics of UUVs in strong ocean current environments.

[0127] 5.3 Strengthen the feedback optimization mechanism for learning guidance.

[0128] First, define the environmental complexity index: define environmental complexity by comprehensively considering ocean currents, water pressure, and topographical factors. : in, , , These are ocean current velocity, water pressure, and topographic slope, respectively. This is a reference value used for normalization.

[0129] The reconstruction verification process is modeled as a reinforcement learning problem, whereby the agent dynamically adjusts the reconstruction parameters. The state space is defined as follows: ,in, One-hot encoding for missing types (3D); Action space is defined as: , These represent the adjustments to the physical model weights, the dynamic constraint weights in manifold optimization, the kinematic constraint weights, and the data augmentation coefficients, respectively; each action component has a set value range, such as... .

[0130] The reward function is designed as follows: ,in, The weighting coefficients balance the improvement of validation metrics with motion smoothness. This reward function encourages the reduction of dynamic and kinematic errors and improves sensor fusion consistency, while penalizing excessive motion adjustments.

[0131] Policy Network Architecture: An Actor-Critic architecture is built using TensorFlow or PyTorch frameworks. The Actor network outputs the action probability distribution, and the Critic network evaluates the state-action values. in, Let be the probability distribution for choosing action a in state s. , , This is the weight matrix. , , For bias vectors, This is a state-action value function.

[0132] The algorithm is trained using Proximal Policy Optimization (PPO), and a certain number of state-action-reward sequences are collected in each training cycle. Calculate the advantage function ,in The state-value function (which can be approximated by a Critic network) is defined by maximizing the objective function. Update Actor network parameters Simultaneously, the mean squared error loss function is used to update the Critic network parameters. After a certain number of training iterations (e.g., 100 iterations), the updated strategy is applied to the UUV state reconstruction process. In this embodiment, the pruning parameters... .

[0133] 5.4 Anomaly tracing in multi-scale time series.

[0134] Error source tracing in Long Short-Term Memory (LSTM) fusion: When the reconstruction error exceeds a threshold, the source of the error is traced through multi-scale time series analysis. 1) Short-term error source tracing: Calculate the mutual information between sensor data and reconstructed state using a sliding window (window size 50): in, For mutual information between sensor data and reconstructed state, For the measurement data of the i-th sensor, The reconstructed value for the j-th state. For joint probability distribution, The marginal probability distribution of sensor data, The marginal probability distribution of the reconstructed state.

[0135] Low mutual information values ​​are a major source of error in sensors. If the mutual information value between a sensor's data and the reconstructed state is below a threshold (e.g., 0.1), then the sensor data is considered to be the main cause of short-term errors, and the sensor's operating status and data validity should be further examined.

[0136] 2) Long-term error source tracing: First, investigate the reconstructed state error. Perform a stationarity test (e.g., using the ADF test). If it is not stationary, perform differencing to make it stationary. Then determine the parameters p and q (autoregressive order and moving average order) of the ARIMA model, and determine the optimal parameters using a minimization information criterion (e.g., AIC, BIC). According to the formula... Establish a model, in which and For model parameters, The noise is white noise. By analyzing the statistical characteristics of the model residuals, such as variance and autocorrelation function, the error trend is determined. If the residual variance suddenly increases, it indicates potential problems such as sudden environmental changes or sensor drift. Based on this, the state reconstruction strategy should be adjusted or the sensor should be inspected and maintained.

[0137] 5.5 Incremental model update strategy.

[0138] To adapt to the time-varying characteristics of UUV system parameters, an incremental model update mechanism is designed: Data buffer management: Maintain a fixed-size (1000 groups) sliding window buffer to store verified (data). Status-sensor data pairs When the buffer is full, old data is replaced according to the first-in, first-out (FIFO) principle to ensure that the buffer always contains the latest and most reliable data samples.

[0139] Incremental update of model parameters: For the Bi-Attention LSTM model, online gradient descent is used for updating. Among them, learning rate , t is the update step number.

[0140] Model Validation: After each update, 20% of the samples are randomly selected from the buffer as the validation set, and the validation error is calculated. Model Validation and Rollback: If the error of the updated model on the validation set increases by more than 15%, the model is considered overfitted or unsuitable for the new data. The model is immediately rolled back to the effective parameters before the last update, and the current parameter update process is stopped. Simultaneously, the system checks for anomalies in the new data or significant changes in the environment, adjusts the data preprocessing method or model structure, and performs incremental updates again to ensure the model maintains good performance throughout long-term UUV operations.

[0141] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for reconstructing the state of an underwater unmanned vehicle under incomplete information, characterized in that, Includes the following steps: Multi-source data fusion preprocessing: Acquire multi-source sensor data from underwater unmanned vehicles and perform spatiotemporal alignment and high-precision data fusion; State prediction model construction: A state prediction model integrating marine environmental features is constructed, with environmental features and sensor sequences as inputs. The environmental features include ocean current velocity, water pressure, and seabed topography slope. Physical constraint loss and a dynamic forgetting mechanism based on the marine environment are introduced to achieve high-precision prediction of the motion state of underwater unmanned vehicles. State information missing detection and classification: Using a multi-source evidence fusion missing detection model and a Bayesian network-based missing cause classification method, combined with a dynamic threshold adaptive mechanism, state information missing detection and classification is performed. Missing types include hardware fault missing, communication interruption missing, and data processing error missing. State Reconstruction: Construct an adaptive state-space model for the marine environment, adopt an adaptive reconstruction strategy based on the missing type, and combine a hierarchical reconstruction framework with marine dynamics constraint optimization to achieve state reconstruction under incomplete information; Reconstruction result verification and feedback optimization: The reliability of the reconstruction result is verified. If the verification fails, feedback optimization is performed through reinforcement learning.

2. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information as described in claim 1, characterized in that, The loss function of the state prediction model is: in, The total loss of the state prediction model. For mean square error loss, For physical constraint loss, For incompressible constraint loss, , These are the weighting coefficients. For the sample size, For the first i The predicted state of each sample is a vector consisting of position, attitude, velocity, and angular velocity. For the first i The true state of each sample The equations for UUV dynamics are as follows: For the first i The derivative of the predicted state of each sample. For the first i Control input variables for each sample , , For the first i Partial derivatives of the velocity components along the x, y, and z axes for each sample.

3. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information as described in claim 1, characterized in that, The dynamic forgetting mechanism based on the marine environment specifically includes: Forgotten Gate weight update: in, For dynamic forget gate output, It is the sigmoid activation function. , Let be the weight matrix of the hidden state at time t and time t-1 for the forget gate. Let be the hidden state at time t. Let be the weight matrix of the cell state at time t-1 to the forgetting gate. The cell state at time t-1. This is the weight matrix of the high-dimensional environmental feature vectors to the forgetting gate. For high-dimensional environmental feature vectors, For the bias term of the forget gate; Define environmental complexity index : in, The ocean current velocity gradient, The rate of change of water pressure The slope of the terrain; when At the same time, increase the update frequency of the forget gate to accelerate the forgetting of old states: in, This represents the environmental complexity threshold. This is the adjustment coefficient.

4. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information as described in claim 1, characterized in that, In the missing state information detection and classification, a missing detection model using multi-source evidence fusion is employed for missing information detection, specifically including: Define a time continuity index for sensor data sequences. for: in, For the previous valid data time, For the current moment, The nominal sampling period of the sensor. This is the adjustment coefficient; Constructing a spatial consistency index by utilizing the physical constraints between sensor data : in, The predicted velocity is calculated using position difference. Let t be the position at time t. For time intervals, For DVL speed measurement, Variance of velocity measurement; Based on the combined temporal and spatial detection results, a joint detection index is defined. : Among them, weight Adjust dynamically based on sensor type; When satisfied , and If any of the following conditions are met, it is determined that information is missing. The threshold for time continuity. This is the spatial consistency threshold. This is the threshold for joint detection.

5. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information as described in claim 4, characterized in that, In the state information missing detection and classification, the dynamic threshold adaptive mechanism is specifically as follows: The time continuity threshold is adaptively adjusted as follows: in, As a time continuity benchmark threshold, p The current water pressure, For reference water pressure, This is the water pressure influence coefficient. The electromagnetic interference influence coefficient. I Electromagnetic interference intensity, The adaptively adjusted time continuity threshold; The spatial consistency threshold is adaptively adjusted as follows: in, As a spatial consistency benchmark threshold, The influence coefficient of ocean currents. For ocean current speed, The adaptively adjusted spatial consistency threshold; The joint detection threshold is adaptively adjusted as follows: The sensor health index is defined by dynamically calibrating threshold parameters based on the sensor's historical health status. : in, For historical data windows, Let be the standard deviation of the k-th measurement. This is the value measured by the sensor for the kth time. Measure the average value of the sensor; When the sensor health index is less than the sensor health index threshold ,in, As the joint detection benchmark threshold, The adaptively adjusted joint detection threshold, For data augmentation coefficients, , The threshold for the sensor health index. These are preset coefficients.

6. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information as described in claim 1, characterized in that, In the state reconstruction, the hierarchical reconstruction framework includes a physical model layer, a data-driven layer, and a decision fusion layer. The physical model layer uses a UUV dynamics model to extrapolate the state and predict short-term missing values. The data-driven layer uses a trained Bi-Attention LSTM model to reconstruct long-term missing values. The decision fusion layer calculates the confidence scores of the reconstruction results from the physical model layer and the data-driven layer to determine the fusion weights. The reconstruction results from the two layers are then fused to obtain the final reconstructed state.

7. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information according to claim 1, characterized in that, In the state reconstruction, ocean dynamics constraint optimization is performed based on a manifold optimization objective function, wherein the manifold optimization objective function is: in, Let k be the UUV state vector at time k. This represents the final reconstructed state after fusion. These are the weighting coefficients. Represents attitude manifold constraints, , Indicates velocity boundary constraints, , The time window length, Let be the attitude angle of the UUV at time k. The attitude angle safety threshold, The linear velocity vector of the UUV. This represents the maximum linear velocity of the UUV.

8. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information as described in claim 1, characterized in that, In the state reconstruction, the adaptive reconstruction strategy based on the missing type is specifically as follows: For hardware failures, DVL speed integral is used to assist in reconstruction: At the same time, the weight of IMU data fusion is reduced, while the weight of DVL and depth sensor fusion is increased. For the reconstructed location estimation, This refers to the position at the previous moment. Size of the integration window For DVL speed, For time intervals; For communication interruptions, terrain matching is used to assist in reconstruction: the terrain height map of the current area is obtained from a pre-stored seabed terrain database. Based on depth sensor data With UUV draft Calculate possible locations Position estimation is optimized through particle filtering. : For data processing errors and missing information, extrapolation using a dynamic model and historical states is employed: in, For the reconstruction speed estimation, For UUV speed, To control the input, The Coriolis force matrix, Here is the damping matrix. This is the UUV quality matrix.

9. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information according to claim 1, characterized in that, In the reconstruction result verification and feedback optimization, the reliability verification includes physical domain verification and data domain verification. The physical domain verification includes dynamic consistency testing and kinematic loop closure testing, while the data domain verification includes sensor fusion consistency verification and deep learning model confidence evaluation. The dynamic consistency check is based on the reconfiguration state and calculates the theoretical acceleration using the dynamic equations. Based on theoretical acceleration and IMU measurement of acceleration Calculate the dynamic consistency deviation index : in, To prevent parameters with a denominator of zero, the verification is deemed unsuccessful when the dynamic consistency deviation index exceeds the dynamic deviation threshold. The kinematic closed-loop test calculates the kinematic deviation index between the velocity integral-reconstructed position and the directly reconstructed position. : in, To reconstruct the position by integrating the velocity, , To directly reconstruct the location, For the integration window, For the maximum speed of UUV, For time intervals, The speed of reconstruction; when the kinematic deviation index is greater than the kinematic deviation threshold, the verification is deemed unsuccessful; The sensor fusion consistency verification utilizes cross-validation with complete sensor data: the sensor fusion consistency index is calculated by comparing the velocity measured by DVL with the velocity obtained by integrating the acceleration through IMU. : in, Accelerating the IMU The variance of historical speed measurements is used; if the sensor fusion consistency index is lower than the fusion consistency threshold, the verification is deemed unsuccessful. The confidence evaluation of the deep learning model utilizes the output distribution entropy of the state prediction model. As a confidence index, if the confidence index is less than the confidence threshold, the verification is deemed unsuccessful.

10. The method for reconstructing the state of an underwater unmanned vehicle under incomplete information according to claim 9, characterized in that, In the reconstruction result verification and feedback optimization, a reinforcement learning method is adopted to model the reconstruction verification process as a reinforcement learning problem, and the reconstruction parameters are dynamically adjusted by the agent: The state space is defined as: ,in, One-hot encoding for missing types, Environmental complexity is defined as follows: in, , , These are ocean current velocity, water pressure, and topographic slope, respectively. , , These are reference values ​​for ocean current velocity, water pressure, and terrain slope, respectively. Action space is defined as: , These represent the adjustment amounts for the physical model weights, the dynamic constraint weights in manifold optimization, the kinematic constraint weights, and the data augmentation coefficients, respectively. The reward function is designed as follows: ,in, These are the weighting coefficients.

Images