High-sea-condition unmanned ship dynamic anti-interference control method based on body intelligence

Abstract

The invention relates to a high sea condition unmanned ship dynamic anti-interference control method based on intelligent body.The method comprises the following steps that information is collected, multi-source data are integrated through a data fusion algorithm, and multi-mode sensing data are obtained; constructing a body-equipped intelligent model, taking the multi-modal sensing data as input and the unmanned ship control instruction as output, and performing offline training; constructing an interference prediction model to predict the change trend of interference factors under the high sea condition, and performing adaptive interference compensation according to an interference prediction result; a hierarchical control structure is designed, a task planning layer makes a global navigation plan based on risk assessment, a motion control layer adopts a model prediction control and adaptive sliding mode control combined algorithm to convert instructions, and an execution mechanism layer drives an execution mechanism; meanwhile, online self-adaptive optimization is carried out, and the intelligent model with the body is updated and optimized according to feedback information. Compared with the prior art, the dynamic anti-interference capability, the operation precision and the reliability of the unmanned surface vehicle under the high sea condition are remarkably improved, the environment adaptability and the autonomous operation level of the unmanned surface vehicle are enhanced, and the unmanned surface vehicle is suitable for various high sea condition ocean operation scenes.

Description

Technical Field

[0001] The present invention relates to unmanned boat control technology, and in particular to a dynamic anti-disturbance control method for an unmanned boat in high sea conditions based on embodied intelligence. Background Art

[0002] Unmanned aerial vehicles (UAVs) are playing an increasingly important role in a wide range of fields, including marine resource development, maritime search and rescue, and environmental monitoring, due to their flexibility, efficiency, and the absence of direct human involvement. However, UAVs face unprecedented challenges in high sea conditions (Sea State Level 5 and above). In these conditions, the interaction of high ocean waves, high winds, and complex currents creates an extremely harsh and dynamic environment.

[0003] Large waves can exert tremendous impact and force on unmanned boats, causing them to shake and pitch violently. This intense mechanical vibration not only affects the proper functioning of the various precision equipment onboard, but can also damage the boat's structure, shortening its service life. Furthermore, the wave and current forces generated by large waves can cause rapid and drastic changes in the boat's attitude, making it difficult for it to maintain a stable navigational position.

[0004] High wind speeds create strong aerodynamic forces against the sides and tops of unmanned vehicles, further disrupting their trajectory. Strong winds can cause the unmanned vehicle to deviate from its intended route, making it more difficult to maintain track. During search and rescue missions, even the slightest deviation can result in missing the target, delaying precious rescue time.

[0005] The presence of complex ocean currents further exacerbates the complexity of high sea conditions. The thrust and pull generated by these currents are uncertain, and their direction and intensity can change at any moment. Unmanned vehicles (UAVs) navigate these currents like an invisible "maze," making it difficult to accurately predict their motion, posing significant challenges to path planning and control.

[0006] Existing control methods for unmanned aerial vehicles (UAVs) currently have numerous shortcomings when dealing with high sea conditions. Traditional model-based control methods often rely on precise mathematical models to describe the UAV's motion and environmental disturbances. However, in high sea conditions, the high complexity and uncertainty of the environment make it difficult to establish an accurate mathematical model, significantly reducing control effectiveness. While control methods based on sensor feedback can obtain real-time information about the UAV's status, simple sensor feedback alone cannot quickly and effectively compensate for and adjust to the complex and variable disturbances in high sea conditions. With the development of artificial intelligence, some intelligent control methods have also been proposed, including those based on machine learning. For example, CN116952235A discloses a method for generating track for an unmanned vehicle based on graph neural networks and deep reinforcement learning. The method first constructs a track generation scenario; then sets the environment, actions, rewards, and the environment after the unmanned vehicle performs the action as parameters of the deep reinforcement learning method; then constructs a memory pool with a fixed capacity to store experience; then constructs and trains an unmanned vehicle track generation model. The unmanned vehicle track generation model uses a dual deep Q-network model structure, in which the experience in the memory pool serves as training data, the current value network and the target value network use a graph convolutional neural network, and the deviation between the current value network and the target value network serves as the loss function; finally, the environmental information collected during the navigation of the unmanned vehicle is input into the model to obtain the action to be performed by the unmanned vehicle. Although these intelligent control algorithms can handle complex environmental information, such as obstacle information, to a certain extent, they still cannot achieve stable and reliable control in the face of extreme high sea conditions due to the lack of in-depth understanding of the interaction between the physical characteristics of the unmanned vehicle and the environment.

[0007] The increasing demand for marine operations places higher demands on the operational capabilities of unmanned vehicles (UAVs) in challenging sea conditions. Improving the dynamic anti-interference capabilities of UAVs in these conditions, ensuring they can accurately perform tasks such as track keeping and target tracking, has become a pressing technical challenge. Summary of the Invention

[0008] The purpose of this invention is to provide a dynamic anti-disturbance control method for unmanned boats in high sea conditions based on embodied intelligence. By combining the physical entity characteristics of the unmanned boat with intelligent algorithms, it can effectively cope with the complex and changeable interference factors in high sea conditions, improve the track keeping accuracy, attitude stability and operation reliability of the unmanned boat in harsh marine environments, and ensure that the unmanned boat can successfully complete various operation tasks in high sea conditions.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A dynamic anti-disturbance control method for an unmanned boat in high sea conditions based on embodied intelligence includes the following steps:

[0011] Multimodal information perception: Inertial measurement units, lidar, millimeter-wave radar, visual sensors, and meteorological sensors are deployed on the unmanned boat to collect information on the boat's motion status, surrounding environment, and meteorological conditions. Multi-source data is then integrated through data fusion algorithms to obtain multimodal perception data.

[0012] Embodied intelligence model establishment and offline training: Based on the physical structure and dynamic characteristics of the UAV, a deep learning algorithm is used to build an embodied intelligence model. Multimodal perception data is used as input, and UAV control instructions are used as output. The model is trained offline through simulation experiments and actual sea trial data, enabling it to learn the UAV's motion patterns and its interaction with the environment.

[0013] Dynamic interference prediction and compensation: Build an interference prediction model to predict the changing trend of interference factors under high sea conditions, and perform adaptive interference compensation based on the interference prediction results;

[0014] Implementing a hierarchical control strategy: Designing a hierarchical control structure, including a mission planning layer, a motion control layer, and an actuator layer. The mission planning layer formulates a global navigation plan based on the task and environmental information. The motion control layer converts the global navigation plan into motion control instructions based on the embodied intelligence model and interference prediction results. The actuator layer drives the actuator of the unmanned vehicle to achieve movement according to the motion control instructions.

[0015] Online adaptive optimization: Real-time collection of multimodal perception data and control effect feedback information, online update and optimization of the embodied intelligent model, and adjustment of model parameters and control strategies to adapt to dynamic changes in high sea conditions.

[0016] The data fusion algorithm adopts an adaptive volumetric federated Kalman filter algorithm to update the state vector of the unmanned boat, and adopts an attention mechanism to calculate the sensor reliability evaluation index according to the variance, update frequency and historical error of each sensor data to dynamically adjust the data fusion weight.

[0017] The embodied intelligent model adopts a hybrid architecture that integrates spatiotemporal Transformer and graph neural network, including a spatiotemporal Transformer module, a graph neural network module and a fusion layer, wherein the spatiotemporal Transformer module is used to process the time series and spatial features of multimodal data, divide the multimodal data into sequences according to time windows, and use a multi-head attention mechanism to calculate the dependencies between data at different times and in different modalities, and extract the spatiotemporal features of the data by stacking multiple layers of Transformer blocks; the graph neural network module is used to model the interactive topological relationship between the unmanned boat and the environment, abstracting the unmanned boat and its surrounding environment into a graph structure, in which the nodes include the unmanned boat's own state nodes and environment nodes, and the edges represent the interactive relationships between nodes. A graph convolutional network is used for feature propagation, and environmental interaction features are extracted through multiple layers of graph convolutional networks; the fusion layer uses a gated fusion mechanism to fuse the spatiotemporal features output by the spatiotemporal Transformer module and the environmental interaction features output by the graph neural network module, and outputs the control instructions for the unmanned boat.

[0018] The offline training adopts a training strategy that combines hierarchical reinforcement learning with imitation learning. Imitation learning pre-training is performed by collecting expert data, and then fine-tuning is performed in combination with hierarchical reinforcement learning. The hierarchical reinforcement learning divides the unmanned boat control task into high-level strategies and low-level strategies. The high-level strategy is responsible for task-level decision-making, and the low-level strategy generates specific control instructions based on high-level instructions and real-time environmental information. Deep Q networks are constructed for the high-level and low-level strategies respectively, and the network parameters are updated through hierarchical training. Among them, the reward function R of reinforcement learning comprehensively considers the task completion, track keeping accuracy and attitude stability and is defined as: R = R task +αR track +βR attitude ,

[0019] Among them, R task Reward for completing the task; R track Track keeping reward, based on the distance d between the current position of the unmanned boat and the planned track track calculate: d max is the maximum allowed track deviation; R attitude It is the attitude stability reward, based on the roll angle Calculation of the absolute value of the pitch angle θ: and θ max is the maximum allowed attitude angle; α and β are weight coefficients.

[0020] The interference prediction model is a long short-term memory network model that integrates a multi-scale spatiotemporal attention mechanism. The corresponding interference prediction process includes the following steps:

[0021] Divide historical multimodal perception data into data tensors containing time dimension T and modality dimension M Each of these lines Represents multimodal data at time t;

[0022] The interference prediction model maps different modal data to a unified dimensional space through an embedding layer. Then, the long-term dependencies of time series data are captured through LSTM units. A multi-scale spatiotemporal attention mechanism is introduced in the time dimension to calculate the weights of different time steps to capture multi-scale time dependencies and the association between different modal data. In the modal dimension, the attention weights of different modal features are calculated to achieve inter-modal information fusion. Finally, the prediction results are output through a fully connected layer.

[0023] The adaptive interference compensation is specifically: calculating the corresponding force according to the predicted waves, strong winds, and ocean currents, and generating compensation force and compensation torque by adjusting the propeller thrust and rudder angle of the unmanned boat, specifically:

[0024] When wave disturbance is predicted, the wave height h is predicted by the disturbance prediction model. wave , wavelength λ wave and wave direction θ wave , calculate the force F exerted by waves on the unmanned boat wave :F wave =ρg∫ S n·ξdS, where ρ is the seawater density, g is the acceleration due to gravity, S is the wetted surface area of the unmanned boat, n is the surface normal vector, and ξ is the surface displacement vector caused by the waves. By adjusting the propeller thrust and rudder angle, a compensation force F is generated. comp1 and compensation torque M comp1 :F comp1 =k f1 ·F wave , M comp1 =k m1 ·F wave ×r, where k f1 、k m1 is the compensation coefficient, r is the position vector from the action point to the center of gravity of the unmanned boat;

[0025] For strong wind disturbances, the wind speed v predicted by the disturbance prediction model is wind and wind direction θ wind , calculate the aerodynamic force F of wind on the unmanned boat wind and aerodynamic moment M wind : Among them, ρ air is the air density, C F (θ wind ), C M (θ wind) are the aerodynamic coefficient matrix and aerodynamic moment coefficient matrix, S ref is the reference area, l ref is the reference length; by adjusting the heading and propulsion of the unmanned boat, a compensation force F is generated comp2 and compensation torque M comp2 :F comp2 =k f2 ·F wind , M comp2 =k m2 ·M wind , where k f2 、k m2 is the compensation coefficient;

[0026] For ocean current disturbance, the ocean current velocity v predicted by the disturbance prediction model is current , calculate the force F exerted by the ocean current on the unmanned boat current :F current =ρ water C D |v rel |v rel , where ρ water is the density of seawater, C D is the drag coefficient matrix, v rel =v current -v is the relative velocity vector, v is the linear velocity and angular velocity vector of the unmanned boat in the body coordinate system; by adjusting the propulsion and steering of the unmanned boat, a compensation force F is generated. comp3 and compensation torque M comp3 :F comp3 =k f3 ·F current , M comp3 =k m3 ·F current ×r, where k f3 、k m3 is the compensation coefficient.

[0027] The task planning layer adopts a dynamic path planning algorithm based on risk assessment, comprehensively considers wave height, wind speed, ocean current speed and obstacle density to assess environmental risks, obtains environmental risk assessment indicators, and adds the environmental risk assessment indicators and risk weight coefficients to the heuristic function of path planning after weighting, and considers task priority in path planning. For search and rescue missions, when the target location is detected, the path planning is dynamically adjusted to give priority to driving to the target point; for environmental monitoring tasks, according to the preset monitoring routes and stations, combined with real-time environmental information, the navigation sequence and stay time are optimized; the global navigation path generated by the task planning layer is sent to the motion control layer.

[0028] The motion control layer receives the global navigation path instructions of the mission planning layer and the dynamic interference prediction results of the embodied intelligent model, adopts the model predictive control algorithm to perform rolling optimization based on the unmanned boat dynamics model, and compensates for interference and model uncertainty through adaptive sliding mode control, generates motion control instructions and sends them to the actuator layer, wherein the adaptive sliding mode control law u in the adaptive sliding mode control is smc For: u smc =u mpc -Ksgn(s), where u mpc is the control variable output by the model predictive control, K is the sliding mode gain matrix adjusted in real time by the adaptive algorithm, sgn(·) is the sign function, and s(t) is the sliding surface function.

[0029] In the online adaptive optimization, a dual-time-scale online learning algorithm based on dynamic weight adjustment is adopted to assign dynamic weights to different types of real-time feedback information, and the model parameter update is divided into a fast update layer and a slow update layer for separate execution. The fast update layer adopts a stochastic gradient descent-momentum algorithm to update the parameters related to the current short-term environmental changes and control responses, and the slow update layer adopts an adaptive learning rate algorithm to update the parameters related to the long-term performance and stability of the model.

[0030] In the online adaptive optimization, the performance of the embodied intelligent model is evaluated by cross-validation. When the evaluated performance index exceeds a preset threshold, the model structure optimization mechanism is triggered to optimize the model structure by adding or deleting the embodied intelligent model neural network layer or adjusting the number of neurons.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] (1) Excellent dynamic anti-interference ability

[0033] The present invention integrates multiple sensor data to achieve multimodal data perception, and adopts an improved federated Kalman filter combined with an attention mechanism algorithm to accurately capture the dynamic changes of interference factors such as waves, strong winds, and ocean currents under high sea conditions. The embodied intelligent model is based on a spatiotemporal Transformer and graph neural network fusion architecture, deeply learns the interaction rules between the unmanned boat and the environment, and combines hierarchical reinforcement learning and imitation learning strategies to enable the model to quickly generate effective control strategies to deal with interference. The dynamic interference prediction model can accurately predict interference trends, and the compensation strategy combines the dynamic characteristics of the unmanned boat to adjust the control input in real time. The three work together to greatly improve the dynamic anti-interference capability of the unmanned boat under high sea conditions and effectively reduce the impact of interference on the motion state of the unmanned boat.

[0034] (2) High-precision job execution

[0035] The mission planning layer of the present invention uses a dynamic path planning algorithm based on risk assessment, fully considering the risks of high sea conditions and planning a safe and efficient navigation path. The motion control layer uses a composite algorithm that combines model predictive control with adaptive sliding mode control, combined with the prediction results of the embodied intelligent model, to achieve precise control of the unmanned vehicle's motion state. The actuator layer uses advanced drive and control technologies to ensure the precise execution of control instructions. The online adaptive optimization mechanism dynamically adjusts the model and control strategy based on real-time feedback information, further improving control accuracy and significantly reducing the unmanned vehicle's track keeping error and attitude control error in high sea conditions, meeting the needs of various high-precision marine operations.

[0036] (3) Strong environmental adaptability

[0037] The present invention is designed based on the complexity and dynamics of high sea conditions, from the special protective treatment of sensor selection and installation to the processing of complex data by data fusion algorithms; from the learning of changing environments by embodied intelligent models to the response of dynamic interference prediction and compensation strategies to different interference factors, and the real-time response to environmental changes by online adaptive optimization, so that the unmanned boat can quickly adapt to the harsh environment of sea condition level 5 and above, and can stably perform tasks regardless of large waves, high wind speeds or complex ocean currents, thereby expanding the operating range and environmental adaptability of the unmanned boat.

[0038] (4) High reliability and robustness

[0039] The hierarchical control strategy of the present invention decomposes complex control tasks into different levels, with clear division of labor and collaborative work at each level, reducing system complexity and failure risks. The fault diagnosis and fault tolerance module equipped at the actuator layer can promptly detect and handle actuator failures, ensuring that the unmanned boat can continue to operate even if some components fail. The online adaptive optimization mechanism continuously optimizes the embodied intelligent model and control strategy, enabling the system to maintain good performance during long-term operation, effectively improving the reliability and robustness of the unmanned boat's operations, reducing the probability of mission failure, and lowering maintenance costs.

[0040] (5) Efficient learning and optimization capabilities

[0041] In this paper, offline training of the embodied intelligent model utilizes a data collection method that combines simulation with actual sea trials. A strategy combining hierarchical reinforcement learning and imitation learning is employed to accelerate model learning and improve learning outcomes. Online adaptive optimization, through dynamic weight adjustment and a dual-timescale online learning algorithm, enables the model to rapidly adapt to environmental changes and mission requirements, continuously optimizing control strategies without the need for frequent human intervention, thereby enhancing the autonomous operational capabilities and intelligence of the unmanned vehicle. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0043] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0044] This embodiment provides a dynamic anti-disturbance control method for unmanned boats in high sea conditions based on embodied intelligence, which is used to solve the problems faced by unmanned boats in high sea conditions, such as severe disturbances, unstable posture and path deviation. Figure 1 As shown, the method includes the following steps:

[0045] Step 1) Multimodal information perception: Deploy an inertial measurement unit, lidar, millimeter-wave radar, visual sensor, and meteorological sensor on the unmanned boat to collect information on the unmanned boat's motion status, surrounding environment, and meteorological conditions. The multi-source data is integrated through a data fusion algorithm to obtain multimodal perception data.

[0046] In high sea conditions, multimodal information perception must overcome the impact of high waves, strong winds, complex ocean currents, and other factors on sensor performance, while also accurately capturing the status and environmental information of the unmanned boat. To this end, this embodiment has been deeply designed from two aspects: sensor selection and installation, and data fusion algorithms.

[0047] 1. Sensor selection and installation

[0048] (1) Inertial Measurement Unit (IMU)

[0049] This embodiment uses a high-precision IMU based on MEMS technology and with temperature compensation function, such as the ADIS16505 model. This IMU can operate stably within the temperature range of -40℃-85℃. Its accelerometer measurement range is ±16g with an accuracy of 0.001g; its gyroscope measurement range is ±2000° / s with an accuracy of 0.01° / s. The IMU senses the angular motion and linear motion of the unmanned boat in three-dimensional space through a sensitive chip, and outputs raw data including the linear acceleration vector measured by the accelerometer. where a x 、a y 、a z are the accelerations along the x, y, and z axes of the UAV coordinate system; the angular velocity vector measured by the gyroscope is ω x 、ω y 、ω zare the angular velocities around the x, y, and z axes, respectively. To reduce the impact of hull vibration on measurement accuracy, a flexible shock-absorbing bracket was used to install the IMU near the center of gravity of the unmanned vehicle, keeping the installation error angle less than 0.5°, and a waterproof sealant was used to protect the interface.

[0050] (2) LiDAR

[0051] This embodiment adopts a combination of a 32-line mechanical rotating laser radar (such as Velodyne VLP-32C) and a solid-state blind spot laser radar (such as Livox Mid-40). The 32-line laser radar has a vertical field of view of -30°-10°, a horizontal field of view of 360°, and a maximum detection distance of 100 meters, which is used for long-range environmental scanning; the solid-state laser radar has a field of view of 120°×25° and a detection distance of 50 meters, which makes up for the short-range blind spot of the 32-line laser radar. The VLP-32C is installed in the center of the top of the unmanned boat to ensure unobstructed 360° scanning; the Mid-40 is installed under the bow of the boat to cover the short-range blind spot. During installation, a six-axis calibration device is required to jointly calibrate the two radars to establish a unified coordinate system with an error control within ±1cm.

[0052] LiDAR measures the target distance by emitting a laser beam and receiving the reflected signal. Its measurement principle is based on time of flight (ToF). The distance calculation formula is: Where d is the target distance, c is the speed of light (c = 299792458 m / s), and t is the time from laser emission to reception. In high sea conditions, violent wave fluctuations can cause unstable lidar echo signals. Therefore, an echo intensity threshold and a multi-sampling filtering algorithm are used to remove invalid echo data caused by wave foam and other factors.

[0053] (3) Millimeter-wave radar

[0054] This embodiment uses a millimeter-wave radar in the 77 GHz frequency band (such as TIIWR1642), which has a range resolution of 0.15 meters, a velocity resolution of 0.1 m / s, and an angle resolution of ±3°.

[0055] Millimeter wave radar uses the Doppler effect to measure the relative speed of the target. The speed calculation formula is: Where v is the target relative velocity, λ is the radar wavelength (the 77 GHz frequency band corresponds to a wavelength of about 3.9 mm), and f d To prevent wind noise from interfering with radar signals in strong winds, a windshield with a wave transmittance of over 95% is installed on the outside of the radar antenna, and an adaptive frequency domain filtering algorithm is used to suppress clutter signals generated by wind noise.

[0056] In this example, the TIIWR1642 millimeter-wave radar, operating in the 77 GHz frequency band, is installed high on either side of the hull and secured with metal brackets coated with absorbing material to reduce interference. After installation, distance and speed are calibrated using standard targets to ensure measurement errors within ±0.15 m (distance) and ±0.1 m / s (speed).

[0057] (4) Visual sensor

[0058] This embodiment is equipped with a dual-camera system, including a 12-megapixel RGB camera (such as Baslerac A2500-14gm) and a thermal imaging camera (such as FLIR Boson 640). The RGB camera is used for target recognition in visible light environments, and the thermal imaging camera can detect the thermal radiation signals of targets on the sea surface at night or in severe weather such as fog. The YOLOv7 algorithm based on deep learning is used for target detection. The algorithm extracts image features through a backbone network (such as CSPDarknet), the neck network (SPPF+PAN) fuses features of different scales, and the head network outputs the target category, confidence and location information (x min ,y min ,x max ,y max ), where (x min ,y min ) is the coordinate of the upper left corner of the target bounding box, (x max ,y max ) is the coordinate of the lower right corner.

[0059] In addition, in order to deal with the camera shake problem in high sea conditions, an image stabilization algorithm based on feature point matching is introduced to extract Harris corner points and calculate the affine transformation matrix between adjacent frames. Perform geometric correction on the image.

[0060] In this embodiment, the RGB camera is mounted above the bow, and the thermal imaging camera is mounted to its side, secured with a dedicated fixture. After installation, internal and external parameters are calibrated using Zhang's calibration method to obtain the camera's intrinsic parameters. A calibration plate is used to establish a conversion relationship with the UAV's coordinate system, with the calibration error controlled within ±0.5 pixels.

[0061] (5) Weather sensors

[0062] In this embodiment, the anemometer uses an ultrasonic sensor (such as Young81000) with a wind speed measurement range of 0-75m / s and an accuracy of ±0.3m / s; the wind direction measurement range is 0-360° and an accuracy of ±3°. The barometer uses a digital high-precision pressure sensor (such as MS5837) with a measurement range of 300-1200hPa and an accuracy of ±0.01hPa. The meteorological sensor is installed at the top of the mast of the unmanned boat and kept at a distance of more than 1 meter from other sensors to avoid mutual interference. In this embodiment, the two are set to be 1.5m apart. During installation, ensure that the anemometer and wind direction are horizontal and the barometer is vertical, and use a lightning protection device to protect the sensor circuit to prevent damage from lightning strikes in high sea conditions. At the same time, to prevent salt spray corrosion, the sensor surface is coated with a nano-level anti-corrosion coating, and the coating thickness is controlled at 5-10μm.

[0063] 2. Data fusion algorithm

[0064] In view of the nonlinear, time-varying and noise characteristics of sensor data under high sea conditions, in this embodiment, the data fusion algorithm adopts the Adaptive Cubic Federated Kalman Filter (ACKF) algorithm to update the state vector of the unmanned boat, and adopts the attention mechanism to calculate the sensor reliability evaluation index according to the variance, update frequency and historical error of each sensor data to dynamically adjust the data fusion weight.

[0065] Assume that the state vector of the unmanned boat is Where (x, y, z) is the position in the geographic coordinate system, in meters (m); They are roll angle, pitch angle and yaw angle respectively, with the unit being radian (rad); is the linear velocity in meters per second (m / s); is the angular velocity in radians per second (rad / s); a x 、a y 、a z The linear acceleration measured by the IMU, in meters per second squared (m / s 2 );ω x 、ω y 、ω z The angular velocity measured by the IMU, in radians per second (rad / s).

[0066] The state transition equation is: Where f is the nonlinear state transfer function, constructed by quaternion method and Newton's laws of motion; is the control input vector, which includes propeller speed and rudder angle commands; is the process noise vector, which obeys Gaussian distribution ), is the process noise covariance matrix; Forecast status.

[0067] For the IMU subsystem, the observation equation is: in, is the observation noise vector, is the IMU observation noise covariance matrix.

[0068] The lidar subsystem obtains the relative position of the unmanned boat and the obstacle through point cloud data processing Its observation equation is: Among them, h Lidar It is the lidar observation function, which is realized through coordinate system transformation and point cloud clustering algorithm; is the lidar observation noise vector, is the lidar observation noise covariance matrix.

[0069] The observation equation of the millimeter wave radar subsystem is: Among them, v rel is the target relative velocity, θ rel is the target relative azimuth; is the millimeter-wave radar observation noise vector, is the millimeter-wave radar observation noise covariance matrix.

[0070] The adaptive cubature Kalman filter process is as follows:

[0071] (1) Time update

[0072] Calculate the Sigma point set: Where n is the dimension of the state vector, λ = α 2 (n+κ)-n, α is used to determine the distribution range of Sigma points (usually 0.001-1), and κ is the adjustment parameter (usually 0); Representation matrix The i-th column of .

[0073] Spread Sigma Points:

[0074] Compute the predicted state mean and covariance:

[0075]

[0076] in, and is the weight coefficient, β is used to utilize the second-order moment information of the state distribution (β = 2 under Gaussian distribution).

[0077] (2) Measurement update

[0078] Calculate the measurement Sigma point set:

[0079] Compute the measurement forecast mean and covariance:

[0080]

[0081] Compute the cross-covariance matrix:

[0082] Calculate the Kalman gain:

[0083] Update state estimates and covariances:

[0084] The attention mechanism is introduced to dynamically adjust the data fusion weight according to the reliability of different sensors in high sea conditions. The reliability evaluation index r is defined i (i corresponds to different sensors), calculated through parameters such as the variance, update frequency, and historical error of sensor data:

[0085] in, is the variance of the measurement data of sensor i, f i is the data update frequency of sensor i, e i is the historical average error of sensor i, and m is the total number of sensors.

[0086] The final fused state estimate is: in, is the state estimate of each sensor after ACKF filtering. Through this fusion algorithm, high-precision fusion of multimodal sensor data can be achieved in complex sea conditions, providing an accurate and reliable data basis for the dynamic anti-disturbance control of the unmanned vehicle.

[0087] When implementing the above data fusion algorithm, deploy the algorithm program in the unmanned boat control unit, set the data acquisition frequency to 100Hz, and perform the following steps:

[0088] S1, initialization

[0089] Define the initial state vector of the unmanned boat Contains information such as position, attitude, speed, etc.; initialization process noise covariance matrix and the observation noise covariance matrix Set the parameters according to the sensor accuracy.

[0090] S2, local filtering

[0091] Each sensor data is input into the corresponding subsystem for local Kalman filtering. Taking the IMU subsystem as an example, the local estimation value is calculated based on the state transfer equation and observation equation. and the local estimated covariance

[0092] S3, attention weight calculation

[0093] Calculate the reliability evaluation index r of each sensor based on the sensor data variance, update frequency and historical error i For example, for LiDAR, by counting the variance of point cloud data per unit time, the number of data updates, and the historical distance measurement error, we can substitute the formula Calculate weights.

[0094] S4, global fusion: according to the attention weight r i and local estimates, calculate the global estimate and the global estimated covariance P k|k , realizing multimodal data fusion.

[0095] Step 2) Establishment and offline training of the embodied intelligence model: Based on the physical structure and dynamic characteristics of the UAV, a deep learning algorithm is used to build an embodied intelligence model. Multimodal perception data is used as input and UAV control instructions are used as output. The model is trained offline through simulation experiments and actual sea trial data to enable it to learn the motion patterns of the UAV and its interaction with the environment.

[0096] The embodied intelligence model is designed to simulate the interaction of unmanned vehicles (UVs) with their environment in high sea conditions and learn optimal control strategies under complex disturbances. In response to the dynamics and uncertainty of high sea conditions and the physical characteristics of the UVs, innovative design approaches were implemented in three areas: model architecture, training data processing, and reinforcement learning algorithms.

[0097] 1. Model Architecture

[0098] In this embodiment, the embodied intelligent model adopts a hybrid architecture that integrates spatiotemporal Transformer and graph neural network (GNN), including a spatiotemporal Transformer module, a graph neural network module and a fusion layer. Taking into account the spatiotemporal characteristics of multimodal data under high sea conditions and the topological relationship between the interaction between the unmanned boat and the environment, this architecture can effectively extract data features and model dynamic interaction processes.

[0099] (1) Spatiotemporal Transformer Module

[0100] The spatiotemporal Transformer module is used to process the time series and spatial features of multimodal data.

[0101] Specifically, first, the multimodal data (IMU data, lidar point cloud, visual image, etc.) is divided into a sequence X = [x1, x2, ..., x T ], where x t is the multimodal data vector at time t, with dimension D.

[0102] Afterwards, the Multi-Head Attention (MHA) mechanism is used to calculate the dependencies between data at different times and in different modalities:

[0103]

[0104] MHA(X)=[head1;…;head h ]W O ,

[0105] Among them, Q, K, and V are query, key, and value matrices respectively, which are obtained by linear transformation of X; d k is the dimension of the key; h is the number of heads; W O is the output weight matrix.

[0106] After stacking multiple layers of Transformer blocks, the spatiotemporal features Z of the data are extracted st .

[0107] (2) Graph Neural Network Module

[0108] The graph neural network module is used to model the interactive topological relationship between the unmanned boat and the environment.

[0109] Specifically, the UAV and its surrounding environment are abstracted into a graph structure G = (V, E), where the node V contains the UAV's own state nodes (such as position, posture, and speed) and environmental nodes (such as wave parameters and obstacle locations), and the edges E represent the interaction between nodes. A graph convolutional network (GCN) is used for feature propagation: Among them, H (l) is the feature matrix of the l-th layer node; is the adjacency matrix with self-loop added; A represents the connection relationship between nodes; for The diagonal node degree matrix of W (l) is the learnable weight matrix; σ is the activation function.

[0110] Through the multi-layer graph convolutional network, the environmental interaction feature Z is extracted gnn .

[0111] (3) Fusion layer

[0112] The fusion layer uses a gated fusion mechanism to fuse the spatiotemporal features Z output by the spatiotemporal Transformer module. st and the environmental interaction feature Z output by the graph neural network module st , output the control command of the unmanned boat: Z=g⊙Z st +(1-g)⊙Z gnn , where g = σ(W g [Z st ; Z gnn ]) is the gate vector, W g is the weight matrix, and ⊙ represents element multiplication.

[0113] Finally, the fused feature Z is input to the fully connected layer, and the control command u=[u1,u2,…,u n ], where u i They correspond to control parameters such as propeller speed and rudder angle respectively.

[0114] In actual implementation, a model architecture based on the fusion of a spatiotemporal Transformer and a graph neural network (GNN) was built in the central processing unit of the unmanned boat. The network was constructed using the PyTorch framework. The spatiotemporal Transformer module consists of eight Transformer blocks, each with a 16-head attention mechanism. The GNN module uses a three-layer graph convolutional network with a node feature dimension of 256. The network weights were initialized using the Xavier initialization method, and the bias was set to 0. The initial learning rate was set to 0.001, and a cosine annealing learning rate adjustment strategy was used.

[0115] 2. Training data collection and processing

[0116] (1) Data collection

[0117] This embodiment collects data by combining simulation with actual sea trials. In terms of simulation, CFD (computational fluid dynamics) software (such as OpenFOAM) and a ship motion simulation platform (such as ShipX) are used to build a high-sea-condition virtual environment to simulate interference factors such as waves, wind, and ocean currents under sea conditions of level 5 and above. Different initial conditions (such as the initial position, heading, and load of the unmanned boat) and mission scenarios (such as straight-line navigation, target tracking, and obstacle avoidance) are set, and the state information (position, attitude, speed, etc.) of the unmanned boat at each time step, sensor data (IMU, lidar, vision, etc.), and corresponding control instructions are recorded. During the actual sea trial, a typical high-sea-condition sea area is selected, and an unmanned boat equipped with a multi-sensor system performs a variety of tasks, and the raw sensor data and control data are collected synchronously.

[0118] In this example, simulation data was generated using the CFD software OpenFOAM and the ship motion simulation platform ShipX. Sea conditions of levels 5-9 were simulated, and different mission scenarios (straight-line navigation, target tracking, and obstacle avoidance) were set. The UAV's status, sensor data, and control instructions were collected. Each simulation lasted 30 minutes, and a total of 1,000 simulations were performed, resulting in approximately 5 million data samples. Actual sea trial data was collected in a typical high-seas area. The UAV, equipped with the sensor system, performed various missions, totaling 100 hours of sea trials and collecting approximately 2 million data points.

[0119] (2) Data preprocessing

[0120] 1) Time synchronization

[0121] Since different sensors have different sampling frequencies, a linear interpolation method based on timestamp is used to synchronize the data. i The sampling time series is {t i,1 ,t i,2 ,…,t i,n}, whose measurement value sequence is {x i,1 ,x i,2 ,…,x i,n}, the target sampling time series is {t1,t2,…,t m}, then synchronize the sensor s i In t j The value of the moment Calculated by linear interpolation:

[0122]

[0123] Among them, t i,k ≤t j ≤t i,k+1 .

[0124] 2) Data normalization

[0125] Normalize the data of different modes to eliminate the dimension effect.

[0126] For numerical data (such as acceleration and velocity), minimum-maximum normalization is used: Among them, x is the original data, x min and x max are the minimum and maximum values of the data respectively.

[0127] For image data, standardization is used: Among them, μ and σ are the mean and standard deviation of the image pixel values, respectively.

[0128] 3) Data augmentation

[0129] In view of the limited data collection in high sea conditions, the visual image data were enhanced by random rotation (±15°), random scaling (0.8-1.2 times), and adding Gaussian noise (mean 0, variance 0.05) to expand the training dataset.

[0130] 3. Offline training of deep reinforcement learning

[0131] In this embodiment, offline training adopts a training strategy that combines hierarchical reinforcement learning with imitation learning. By collecting expert data for imitation learning pre-training, and then combining hierarchical reinforcement learning for fine-tuning, the model's learning efficiency and generalization ability are improved under complex tasks in high sea conditions.

[0132] (1) Imitation learning

[0133] Collect expert (experienced UAV operator or optimal control trajectory in simulation) data {(s i ,a i )}, use the behavior cloning algorithm for pre-training. Define the imitation learning loss function:

[0134]

[0135] Among them, N is the number of expert data samples, π(s i ;θ) is the model according to the state s i The predicted action, θ is the model parameter. Through pre-training, the model is equipped with certain prior knowledge to accelerate the reinforcement learning process.

[0136] In this embodiment, the collected expert data (about 500,000) are input into the model, and the behavioral cloning algorithm is used for pre-training. The model is trained for 100 epochs so that the model has a preliminary control strategy.

[0137] (2) Hierarchical reinforcement learning

[0138] The control task of the unmanned boat is divided into high-level strategies and low-level strategies. The high-level strategy is responsible for task-level decision-making, such as selecting a navigation path based on the target location; the low-level strategy generates specific control instructions based on high-level instructions and real-time environmental information. Define the high-level state space S h Contains task objectives, global environment information, etc., action space A h is the path planning instruction; the low-level state space S l Contains the local state of the unmanned boat and the real-time data of the sensor, the action space A l are control parameters such as propeller speed and rudder angle. A deep Q network (DQN) is constructed for high-level and low-level strategies respectively:

[0139]

[0140] Among them, θh and θ l are high-level and low-level network parameters, respectively. and are the network weights and biases, and σ is the activation function. Through layered training, the complexity of strategy learning is reduced.

[0141] Based on the imitation learning pre-training, reinforcement learning fine-tuning is performed. The reinforcement learning reward function R is defined to comprehensively consider factors such as task completion, track keeping accuracy, and attitude stability: R = R task +αR track +βR attitude , where R task Reward for completing the task (e.g. 100 points for reaching the target point, -50 points for not completing it); R track Track keeping reward, based on the distance d between the current position of the unmanned boat and the planned track track calculate: d max is the maximum allowed track deviation; R attitude It is the attitude stability reward, based on the roll angle Calculation of the absolute value of the pitch angle θ: and θ max is the maximum allowed attitude angle; α and β are weight coefficients, which are determined through experimental optimization.

[0142] In this embodiment, both the high-level policy network and the low-level policy network adopt a fully connected neural network structure. The high-level network has an input dimension of 128 and an output dimension of 8; the low-level network has an input dimension of 256 and an output dimension of 4. The discount factor γ is set to 0.99, and the experience replay buffer capacity is 100,000 data items.

[0143] The Dual Deep Q Network (DDQN) algorithm is used for training, and the network parameters θ are updated to maximize the long-term cumulative reward:

[0144]

[0145] Among them, y i =r i +γmax a′ Q′(s i+1 ,a′;θ′),r i is the reward at time i, γ is the discount factor, Q′ and θ′ are the target networks, and N is the number of samples. This embodied intelligence modeling method effectively learns the dynamic anti-disturbance control strategy for unmanned vehicles in high sea conditions, enabling autonomous and stable operation in complex environments.

[0146] In this example, the number of training episodes is set to 5000, with each episode lasting 10 minutes. The network parameters are updated using the Dual Deep Q Network (DDQN) algorithm, with the target network parameters updated every 100 steps.

[0147] Step 3) Dynamic interference prediction and compensation: Construct an interference prediction model to predict the changing trends of interference factors such as waves, strong winds, and ocean currents under high sea conditions, and perform adaptive interference compensation based on the interference prediction results.

[0148] In high sea conditions, interference factors such as waves, strong winds, and ocean currents are highly dynamic and uncertain, posing a serious threat to the safety of the UAV and its mission execution. This step effectively addresses interference by building an advanced interference prediction model and developing a precise compensation strategy.

[0149] 1. Construction of interference prediction model

[0150] Based on the spatiotemporal correlation and complex nonlinear characteristics of interference factors in high sea conditions, a long short-term memory network (MST-LSTM) model integrating a multi-scale spatiotemporal attention mechanism is constructed as an interference prediction model to achieve accurate prediction of interference factors such as waves, wind speed, and ocean currents.

[0151] Divide historical multimodal perception data into data tensors containing time dimension T and modality dimension M Each of these lines Represents multimodal data at time t; including attitude and acceleration information measured by IMU, wave shape data detected by lidar, ocean current speed and direction obtained by millimeter-wave radar, and wind speed and direction collected by meteorological sensors.

[0152] In the MST-LSTM model, the different modal data are first mapped to a unified dimensional space through the embedding layer: E t =Embedding(x t ), where E t is the embedding vector at time t.

[0153] The LSTM unit is used to capture the long-term dependencies of time series data. Its core calculation process is as follows: t =σ(W f ·[E t ,h t-1 ]+b f ),

[0154] i t =σ(W i ·[E t ,h t-1 ]+b i ),

[0155]

[0156] o t =σ(W o ·[E t ,h t-1 ]+b o ),

[0157] h t =o t ⊙tanh(C t ),

[0158] Among them, f t 、i t 、o t They are forget gate, input gate and output gate respectively; C t is the cell state; h t is the hidden state; σ is the Sigmoid activation function; W f 、W i 、W C 、W o is the weight matrix; b f 、b i 、b C 、b o is the bias vector; ⊙ represents element-wise multiplication.

[0159] In order to capture the multi-scale temporal dependencies and the associations between different modal data, a multi-scale spatiotemporal attention mechanism is introduced. In the time dimension, a multi-head attention mechanism is defined to calculate the weights of different time steps:

[0160]

[0161] MHA(H)=[head1;…;head h ]W O ,

[0162] Where H=[h1,h2,…,h T ] is the hidden state sequence output by LSTM; Q, K, and V are query, key, and value matrices, respectively, which are obtained by linear transformation of H; d k is the dimension of the key; h is the number of heads; W O is the output weight matrix.

[0163] In the modality dimension, by calculating the attention weights of different modal features, inter-modal information fusion is achieved:

[0164]

[0165] in, is the hidden state corresponding to the mth mode at time t; w is the learnable weight vector; Pooling is the pooling operation; α m is the attention weight of the m-th modality; is the hidden state after fusion.

[0166] Finally, the prediction result is output through the fully connected layer: in, is the interference factor data for prediction, such as wave height, wind speed, ocean current speed, etc. in the future period; W fc is the weight matrix of the fully connected layer; b fc is the bias vector.

[0167] The model training uses the mean square error (MSE) loss function: Where N is the number of samples; Y n is the real interference factor data; Predict data for the model.

[0168] In practice, preprocessed historical multimodal data is divided into 10-minute time windows and fed into the model for training. The training batch size is set to 64, the training cycle is 200 epochs, and the Adam optimizer is used with the mean squared error (MSE) loss function. The latest 10 minutes of multimodal data is collected in real time and fed into the trained model to predict changes in disturbance factors such as wave height, wind speed, and ocean current speed over the next 5 minutes.

[0169] 2. Adaptive interference compensation

[0170] Combined with the dynamic characteristics of the unmanned boat, the corresponding forces are calculated according to the predicted waves, strong winds, and ocean current interference. The compensation force and compensation torque are generated by adjusting the propeller thrust and rudder angle of the unmanned boat. Specifically:

[0171] The dynamic equation of the unmanned boat is defined as:

[0172]

[0173] in, is the position and attitude vector of the unmanned vehicle in the inertial coordinate system, x, y, z are the three-dimensional position coordinates, φ is the roll angle, θ is the pitch angle, and ψ is the yaw angle; v = [u, v, w, p, q, r] T are the linear velocity and angular velocity vectors of the unmanned vehicle in the body coordinate system, u, v, w are the linear velocities along the x, y, and z axes of the fixed coordinate system, respectively, and p, q, and r are the angular velocities around the x, y, and z axes of the fixed coordinate system, respectively; J(η) is the transformation matrix; M is the inertia matrix; C(v) is the Coriolis force and centripetal force matrix; D(v) is the damping matrix; G(η) is the gravity and buoyancy matrix; τ is the control input vector; τ disturbare the disturbance force and torque vectors.

[0174] When wave disturbance is predicted, the wave height h is predicted by the disturbance prediction model. wave , wavelength λ wave and wave direction θ wave , calculate the force F exerted by waves on the unmanned boat wave :F wave =ρg∫ S n·ξdS, where ρ is the seawater density, g is the acceleration of gravity, S is the wetted surface area of the unmanned boat, n is the surface normal vector, and ξ is the surface displacement vector caused by the waves. By adjusting the propeller thrust and rudder angle, a compensation force F is generated. comp1 and compensation torque M comp1 :F comp1 =k f1 ·F wave , M comp1 =k m1 ·F wave ×r, where k f1 、k m1 is the compensation coefficient, and r is the position vector from the action point to the center of gravity of the unmanned boat.

[0175] For strong wind disturbances, the wind speed v predicted by the disturbance prediction model is wind and wind direction θ wind , calculate the aerodynamic force F of wind on the unmanned boat wind and aerodynamic moment M wind : Among them, ρ air is the air density, C F (θ wind ), C M (θ wind ) are the aerodynamic coefficient matrix and aerodynamic moment coefficient matrix, S ref is the reference area, l ref is the reference length; by adjusting the heading and propulsion of the unmanned boat, a compensation force F is generated comp2 and compensation torque M comp2 :F comp2 =k f2 ·F wind , M comp2 =k m2 ·M wind , where k f2 、k m2 is the compensation coefficient.

[0176] For ocean current disturbance, the ocean current velocity v predicted by the disturbance prediction model is current , calculate the force F exerted by the ocean current on the unmanned boat current :F current =ρwater C D |v rel |v rel , where ρ water is the density of seawater, C D is the drag coefficient matrix, v rel =v current -v is the relative velocity vector, v is the linear velocity and angular velocity vector of the unmanned boat in the body coordinate system; by adjusting the propulsion and steering of the unmanned boat, a compensation force F is generated. comp3 and compensation torque M comp3 :F comp3 =k f3 ·F current , M comp3 =k m3 ·F current ×r, where k f3 、k m3 is the compensation coefficient.

[0177] The final control input τ final is the task demand control quantity τ task And the sum of each interference compensation amount: τ final =τ task +F comp1 +M comp1 +F comp2 +M comp2 +F comp3 +M comp3 .

[0178] Through the above-mentioned dynamic interference prediction and compensation method, it is possible to perceive interference in advance and implement precise compensation in complex environments with high sea conditions, effectively improving the dynamic anti-interference performance and operational stability of the unmanned boat.

[0179] In the specific implementation process, when the predicted wave height h wave Exceed the threshold, calculate the wave force F wave By adjusting the propeller thrust (increase 10%-30%) and the rudder angle (adjust ±15°), a compensation force F is generated. comp1 and compensation torque M comp1 According to the predicted wind speed v wind and wind direction θ wind , calculate the aerodynamic force F wind and aerodynamic moment M wind , adjust the unmanned boat's heading (deflection ±20°) and propulsion force (increase or decrease 15%-25%) to compensate. current , calculate the force F current By adjusting the propeller speed and rudder angle combination, a compensation force F is generated. comp3 and compensation torque M comp3 .

[0180] Step 4) Implement a hierarchical control strategy: Design a hierarchical control structure, including a task planning layer, a motion control layer, and an actuator layer. The task planning layer formulates a global navigation plan based on the task and environmental information. The motion control layer converts the global navigation plan into motion control instructions based on the embodied intelligence model and interference prediction results. The actuator layer drives the UAV actuator to achieve movement according to the motion control instructions.

[0181] The hierarchical control strategy divides the control tasks of the unmanned vehicle in high sea conditions into a task planning layer, a motion control layer, and an actuator layer. These layers work together to achieve efficient and stable operation in complex marine environments through information exchange and coordination. This strategy fully considers the complex and changing interference factors in high sea conditions and the high precision requirements for mission execution, and adopts a variety of innovative technologies and algorithms to improve control performance.

[0182] 1. Task planning layer

[0183] The mission planning layer is primarily responsible for developing a global navigation plan based on mission requirements and current environmental information. In high sea conditions, traditional path planning algorithms struggle to cope with complex and changing ocean conditions, so a dynamic path planning algorithm based on risk assessment is employed.

[0184] Define environmental risk assessment indicator R env , taking into account the wave height h wave , wind speed v wind , ocean current speed v current , obstacle density ρ obs Other factors:

[0185]

[0186] Among them, h max 、v max 、v c-max are the safety thresholds for wave height, wind speed, and ocean current speed, respectively; α1, α2, α3, and α4 are weight coefficients, which are determined through a large number of simulations and optimization of actual sea trial data, and are used to adjust the degree of influence of each factor on risk assessment.

[0187] The improved A algorithm is used for path planning. On the basis of the traditional A algorithm, the environmental risk assessment index R env Incorporate the heuristic function f(n): f(n)=g(n)+h(n)+βR env (n), where g(n) is the actual cost from the starting node to the current node n; h(n) is the estimated cost from the current node n to the target node; β is the risk weight coefficient, which is used to control the impact of environmental risk on path selection; R env(n) is the environmental risk assessment value at node n. This improvement allows the planned path to not only consider the shortest path but also avoid high-risk areas, improving the safety of the unmanned vehicle in high sea conditions. When the mission or environment changes, the path is replanned every 30 seconds.

[0188] At the same time, the mission planning layer also needs to consider mission priorities. For search and rescue missions, when the target location is detected, the path planning is dynamically adjusted to prioritize driving to the target point; for environmental monitoring missions, according to the preset monitoring routes and stations, combined with real-time environmental information, the navigation sequence and stay time are optimized. The mission planning layer generates a global navigation path P = [p1, p2, ..., p N ], where p i =(x i ,y i ,z i ) are the node coordinates on the path and are sent to the motion control layer.

[0189] 2. Motion control layer

[0190] The motion control layer receives the global navigation path instructions from the mission planning layer and the dynamic interference prediction results of the embodied intelligent model. It uses the model predictive control (MPC) algorithm to perform rolling optimization based on the unmanned boat dynamics model, and compensates for interference and model uncertainty through adaptive sliding mode control. It generates motion control instructions and sends them to the actuator layer.

[0191] Based on the nonlinear dynamic model of the unmanned boat, a prediction model is established. In the discrete time domain, the state equation of the unmanned boat can be expressed as: k+1 =f(x k ,u k )+w k ,in, is the state vector of the unmanned boat at time k, including position and attitude vector η k and the linear and angular velocity vectors v k ;u k is the control input vector at time k, including propeller speed and rudder angle, etc.; f is the nonlinear state transfer function; w k is the process noise vector, which obeys Gaussian distribution w k ~N(0,Q k ), Q k is the process noise covariance matrix.

[0192] The core of model predictive control is to predict the system output in the future based on the current state at each sampling moment. The prediction time domain is defined as N p , the control time domain is N c (N c ≤N p). At time k, the following objective function is solved by rolling optimization:

[0193]

[0194] Among them, x k+i|k is the state at time k+i predicted at time k; x r,k+i is the state of the reference trajectory at time k+i; Q is the state weight matrix; R is the control input weight matrix. By solving the objective function, the optimal control sequence is obtained And the first control quantity u k|k Applied to unmanned boats.

[0195] In this embodiment, the prediction time domain N is set p =10, control time domain N c =5, rolling optimization objective function, calculate the optimal control sequence, and update the control instructions every 100ms.

[0196] Due to the strong interference and model uncertainty in high sea conditions, adaptive sliding mode control is introduced to compensate. The sliding mode surface function s(t) is defined as: Where, e(t)=x(t)-x r (t) is the state error vector; λ is a positive definite diagonal matrix.

[0197] Design of adaptive sliding mode control law u smc :u smc =u mpc -Ksgn(s), where u mpc is the control variable output by the model predictive control; K is the sliding mode gain matrix; and sgn(·) is the sign function. Simultaneously, an adaptive algorithm is used to adjust the sliding mode gain K in real time to adapt to disturbances and model changes. For example, based on Lyapunov stability theory, an adaptive law is designed: Where Γ is the adaptive gain matrix.

[0198] The motion control layer will generate the final motion control instruction u ctrl =[u propeller ,δ rudder ] T , where u propeller is the propeller speed, δ rudder is the rudder angle, which is sent to the actuator layer.

[0199] 3. Executive agency layer

[0200] The actuator layer drives the propellers, steering gear, and other actuators of the unmanned boat according to instructions from the motion control layer. In high sea conditions, the actuators must have fast response and high reliability to cope with complex and changing control requirements.

[0201] The propeller is driven by a permanent magnet synchronous motor, which uses vector control technology to achieve precise speed regulation. The dynamic model of the motor can be expressed as:

[0202]

[0203] Among them, ω m is the motor rotor angular velocity; J is the moment of inertia; T e is the electromagnetic torque; T l is the load torque; B is the viscous friction coefficient; ψ s is the stator flux; ω s is the synchronous angular velocity; L s is the stator inductance; i s is the stator current; n p is the number of motor pole pairs; Im(·) represents the imaginary part. By adjusting the stator current i s The amplitude and phase of the propeller speed u propeller precise control.

[0204] The steering gear adopts an electric hydraulic steering gear system, and its control model is:

[0205]

[0206] Among them, δ is the actual value of the rudder angle; u δ is the rudder angle command output by the motion control layer; τ is the servo time constant. To improve the servo's response speed and control accuracy, a feedforward compensation and feedback correction control strategy are introduced. Feedforward compensation pre-adjusts the servo input based on the motion control command, accelerating response speed. Feedback correction uses an angle sensor to monitor the rudder angle in real time and compares it with the command rudder angle to perform error correction.

[0207] The actuator layer is also equipped with a fault diagnosis and fault tolerance module. By real-time monitoring of actuator parameters such as current, voltage, and temperature, deep learning-based fault diagnosis models (such as convolutional neural network-long short-term memory network, CNN-LSTM) are used to determine whether an actuator has failed. Once a fault is detected, fault tolerance mechanisms are immediately activated, such as switching to a backup propeller or adopting a redundant servo control strategy, to ensure that the unmanned boat can continue to perform its mission even if some actuators fail.

[0208] Through the above-mentioned hierarchical control strategy, from task planning to coordinated control of actuators, the control performance and mission execution capability of the unmanned boat in high sea conditions are effectively improved.

[0209] In this embodiment, the permanent magnet synchronous motor utilizes vector control technology. Based on motion control layer commands, the propeller speed is controlled by adjusting the stator current amplitude and phase, achieving a control accuracy of ±1 rpm. The electro-hydraulic steering system receives rudder angle commands and, through a combination of feedforward compensation and feedback correction, maintains a rudder angle tracking error of less than ±0.5°. Furthermore, a fault diagnosis module monitors the actuator status every 50 ms and immediately activates fault tolerance mechanisms if a fault is detected.

[0210] Step 5) Online adaptive optimization: Collect multimodal perception data and control effect feedback information in real time, update and optimize the embodied intelligent model online, and adjust model parameters and control strategies to adapt to the dynamic changes in high sea conditions.

[0211] Online adaptive optimization is a key component in ensuring the continued stability of unmanned vehicles in complex and dynamically changing sea conditions. This component collects real-time feedback and dynamically adjusts and optimizes the embodied intelligent model based on it, enabling the unmanned vehicle to quickly adapt to environmental changes and continuously improve its anti-interference performance and control accuracy.

[0212] 1. Feedback information collection

[0213] During the operation of the unmanned boat, comprehensive and real-time multi-dimensional feedback information is collected, including the following categories:

[0214] (1) Status feedback information

[0215] Use high-precision inertial measurement unit (IMU) to obtain the actual attitude angle (roll angle) of the unmanned boat in real time Pitch angle θ actual , yaw angle ψ actual ), linear acceleration and angular velocity At the same time, the actual position of the unmanned boat in the geographic coordinate system (x actual ,y actual ,z actual ) and linear velocity (v x-actual ,v y-actual ,v z-actual ). These status information are used to reflect the difference between the actual motion state of the unmanned boat and the expected state.

[0216] (2) Environmental feedback information

[0217] The laser radar, millimeter wave radar and visual sensor continuously perceive the changes in the surrounding environment. The laser radar collects the wave height h in real time. wave-current , wavelength λ wave-current and obstacle distance information; millimeter-wave radar monitors ocean current speed and wind direction and speed (vwind-actual ,θ wind-actual ); visual sensors are used to identify the location and state changes of targets on the sea surface. This environmental information helps the model promptly understand the impact of dynamic changes in the external environment on the unmanned vehicle.

[0218] (3) Control effect feedback information

[0219] Calculate the track deviation e track , by comparing the actual position of the unmanned boat (x actual ,y actual ) and the corresponding point (x ref ,y ref ) is obtained by the Euclidean distance, that is, Attitude error e attitude The attitude difference calculation is performed using quaternion, and the expected attitude quaternion is q ref , the actual attitude quaternion is q actual ,but in Represents quaternion multiplication. Control instruction execution deviation e control Calculate the actual propeller speed u propeller-actual , rudder angle δ rudder-actual The command value u output by the motion control layer propeller-ref , δ rudder-ref The difference between

[0220] All feedback information is collected at a fixed sampling period Δt and transmitted to the central processing unit via a high-speed data bus, providing a data basis for the online optimization of the model.

[0221] In the actual implementation process, the information collection program is set in the unmanned boat data collection module:

[0222] (1) Status information acquisition: IMU, GNSS, and DVL collect the status information of the UAV at a frequency of 100 Hz and transmit it to the central processing unit through the CAN bus.

[0223] (2) Environmental information collection: LiDAR, millimeter-wave radar, visual sensors, and meteorological sensors collect data at their respective frequencies and transmit them via Ethernet.

[0224] (3) Control effect information collection: Real-time calculation of track deviation, attitude error and control instruction execution deviation, updated every 100ms.

[0225] 2. Model update and optimization

[0226] Based on the collected feedback information, a dual-time-scale online learning algorithm based on dynamic weight adjustment is used to update and optimize the embodied intelligence model to achieve rapid convergence of model parameters and adaptive adjustment of control strategies.

[0227] (1) Dynamic weight adjustment

[0228] Assign dynamic weights to different types of feedback information to highlight the information that is more critical to model optimization in the current environment. Define the weight vector where w state 、w env 、w control are the weights of state feedback information, environment feedback information, and control effect feedback information respectively. The weight adjustment is based on the complexity of the current sea conditions and task requirements, and is achieved through the fuzzy logic system. Taking the complexity of the sea conditions as an example, according to the wave height h wave-current , wind speed v wind-actual , ocean current speed The system calculates the fuzzy value of the complexity of the sea state through a membership function, and then uses fuzzy inference rules to determine the adjustment amount of each weight. For example, when the sea state is complex and the target is being tracked, the weights of environmental feedback and control effect feedback are increased to quickly adapt to environmental changes and improve tracking accuracy.

[0229] In this embodiment, the weight of the feedback information is adjusted every 5 seconds through the fuzzy logic system according to the sea conditions and mission requirements.

[0230] (2) Dual-time-scale online learning

[0231] The model parameter update is divided into a fast update layer and a slow update layer. The fast update layer is mainly for parameters that are closely related to the current short-term environmental changes and control responses, such as some connection weights W in the neural network. fast and bias b fast The slow update layer is used to adjust the parameters that have a greater impact on the long-term performance and stability of the model, such as the hyperparameters in the network structure and some deep network weights W slow , bias b slow .

[0232] For the fast update layer, the stochastic gradient descent-momentum (SGD-Momentum) algorithm is used to update the parameters. Let the loss function L be about the parameter W fast The gradient of The parameter update formula is:

[0233]

[0234] Where k represents the number of iterations; α fast is the fast learning rate; β fastis the momentum factor; is the momentum term.

[0235] For the slow update layer, an adaptive learning rate algorithm (such as Adagrad) is used to update the parameters. Let the loss function L be about the parameter W slow The gradient of The parameter update formula is:

[0236]

[0237] Among them, α slow is a slow learning rate; ∈ is a very small constant to prevent the denominator from being zero.

[0238] In this embodiment, the fast update layer updates parameters every 10 iterations; the slow update layer updates parameters every 100 iterations.

[0239] After each model update, the model performance is evaluated by cross-validation. The mean square error (MSE) on the validation set is calculated. validation :

[0240]

[0241] Among them, N validation is the number of samples in the validation set; is the model prediction value; y i is the true value.

[0242] In this example, the validation set is used to evaluate the model performance every 10 minutes. validation If the preset threshold is exceeded or the model performance shows a downward trend, the model structure optimization mechanism is triggered. The model structure is optimized by adding or deleting neural network layers, adjusting the number of neurons, etc., to ensure that the model always maintains good performance in a high sea state dynamic environment.

[0243] Through the above-mentioned online adaptive optimization method, the embodied intelligent model of the unmanned boat can dynamically adjust the model parameters and control strategies according to the real-time changes in the high sea conditions and the mission execution status, effectively improving the adaptability and operational reliability of the unmanned boat in complex marine environments.

[0244] Through the above detailed specific implementation methods, a dynamic anti-disturbance control method for unmanned boats in high sea conditions based on embodied intelligence can be realized to ensure stable and efficient operation of the unmanned boats in high sea conditions.

[0245] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A dynamic anti-disturbance control method for an unmanned boat in high sea conditions based on embodied intelligence, characterized in that: The following steps are involved: Multimodal information perception: Inertial measurement units, lidar, millimeter-wave radar, visual sensors, and meteorological sensors are deployed on the unmanned boat to collect information on the boat's motion status, surrounding environment, and meteorological conditions. Multi-source data is then integrated through data fusion algorithms to obtain multimodal perception data. Embodied intelligence model establishment and offline training: Based on the physical structure and dynamic characteristics of the UAV, a deep learning algorithm is used to build an embodied intelligence model. Multimodal perception data is used as input, and UAV control instructions are used as output. The model is trained offline through simulation experiments and actual sea trial data, enabling it to learn the UAV's motion patterns and its interaction with the environment. Dynamic interference prediction and compensation: Build an interference prediction model to predict the changing trend of interference factors under high sea conditions, and perform adaptive interference compensation based on the interference prediction results; Implementing a hierarchical control strategy: Designing a hierarchical control structure, including a mission planning layer, a motion control layer, and an actuator layer. The mission planning layer formulates a global navigation plan based on the task and environmental information. The motion control layer converts the global navigation plan into motion control instructions based on the embodied intelligence model and interference prediction results. The actuator layer drives the actuator of the unmanned vehicle to achieve movement according to the motion control instructions. Online adaptive optimization: Real-time collection of multimodal perception data and control effect feedback information, online update and optimization of the embodied intelligent model, and adjustment of model parameters and control strategies to adapt to dynamic changes in high sea conditions.

2. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The data fusion algorithm adopts an adaptive volumetric federated Kalman filter algorithm to update the state vector of the unmanned boat, and adopts an attention mechanism to calculate the sensor reliability evaluation index according to the variance, update frequency and historical error of each sensor data to dynamically adjust the data fusion weight.

3. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The embodied intelligent model adopts a hybrid architecture that integrates spatiotemporal Transformer and graph neural network, including a spatiotemporal Transformer module, a graph neural network module and a fusion layer, wherein the spatiotemporal Transformer module is used to process the time series and spatial features of multimodal data, divide the multimodal data into sequences according to time windows, and use a multi-head attention mechanism to calculate the dependencies between data at different times and in different modalities, and extract the spatiotemporal features of the data by stacking multiple layers of Transformer blocks; the graph neural network module is used to model the interactive topological relationship between the unmanned boat and the environment, abstracting the unmanned boat and its surrounding environment into a graph structure, in which the nodes include the unmanned boat's own state nodes and environment nodes, and the edges represent the interactive relationships between nodes. A graph convolutional network is used for feature propagation, and environmental interaction features are extracted through multiple layers of graph convolutional networks; the fusion layer uses a gated fusion mechanism to fuse the spatiotemporal features output by the spatiotemporal Transformer module and the environmental interaction features output by the graph neural network module, and outputs the control instructions for the unmanned boat.

4. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The offline training adopts a training strategy that combines hierarchical reinforcement learning with imitation learning. Imitation learning pre-training is performed by collecting expert data, and then fine-tuning is performed in combination with hierarchical reinforcement learning. The hierarchical reinforcement learning divides the unmanned boat control task into high-level strategies and low-level strategies. The high-level strategy is responsible for task-level decision-making, and the low-level strategy generates specific control instructions based on high-level instructions and real-time environmental information. Deep Q networks are constructed for the high-level and low-level strategies respectively, and the network parameters are updated through hierarchical training. Among them, the reward function R of reinforcement learning comprehensively considers the task completion, track keeping accuracy and attitude stability and is defined as: R = R task +αR track +βR attitude , where R task Reward for completing the task; R track Track keeping reward, based on the distance d between the current position of the unmanned boat and the planned track track calculate: d max is the maximum allowed track deviation; R attitude It is the attitude stability reward, based on the roll angle Calculation of the absolute value of the pitch angle θ: and θ max is the maximum allowed attitude angle; α and β are weight coefficients.

5. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The interference prediction model is a long short-term memory network model that integrates a multi-scale spatiotemporal attention mechanism. The corresponding interference prediction process includes the following steps: Divide historical multimodal perception data into data tensors containing time dimension T and modality dimension M Each of these lines Represents multimodal data at time t; The interference prediction model maps different modal data to a unified dimensional space through an embedding layer. Then, the long-term dependencies of time series data are captured through LSTM units. A multi-scale spatiotemporal attention mechanism is introduced in the time dimension to calculate the weights of different time steps to capture multi-scale time dependencies and the association between different modal data. In the modal dimension, the attention weights of different modal features are calculated to achieve inter-modal information fusion. Finally, the prediction results are output through a fully connected layer.

6. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The adaptive interference compensation is specifically: calculating the corresponding force according to the predicted waves, strong winds, and ocean currents, and generating compensation force and compensation torque by adjusting the propeller thrust and rudder angle of the unmanned boat, specifically: When wave disturbance is predicted, the wave height h is predicted by the disturbance prediction model. wave , wavelength λ wave and wave direction θ wave , calculate the force F exerted by waves on the unmanned boat wave :F wave =ρg∫ S n·ξdS, where ρ is the seawater density, g is the acceleration due to gravity, S is the wetted surface area of the unmanned boat, n is the surface normal vector, and ξ is the surface displacement vector caused by the waves. By adjusting the propeller thrust and rudder angle, a compensation force F is generated. comp1 and compensation torque M comp1 :F comp1 =k f1 ·F wave , M comp1 =k m1 ·F wave ×r, where k f1 、k m1 is the compensation coefficient, r is the position vector from the action point to the center of gravity of the unmanned boat; For strong wind disturbances, the wind speed v predicted by the disturbance prediction model is wind and wind direction θ wind , calculate the aerodynamic force F of wind on the unmanned boat wind and aerodynamic moment M wind : Among them, ρ air is the air density, C F (θ wind ), C M (θ wind ) are the aerodynamic coefficient matrix and aerodynamic moment coefficient matrix, S ref is the reference area, l ref is the reference length; by adjusting the heading and propulsion of the unmanned boat, a compensation force F is generated comp2 and compensation torque M comp2 :F comp2 =k f2 ·F wind , M comp2 =k m2 ·M wind , where k f2 、k m2 is the compensation coefficient; For ocean current disturbance, the ocean current velocity v predicted by the disturbance prediction model is current , calculate the force F exerted by the ocean current on the unmanned boat current :F current =ρ water C D |v rel |v rel , where ρ water is the density of seawater, C D is the drag coefficient matrix, v rel =v current -v is the relative velocity vector, v is the linear velocity and angular velocity vector of the unmanned boat in the body coordinate system; by adjusting the propulsion and steering of the unmanned boat, a compensation force F is generated. comp3 and compensation torque M comp3 :F comp3 =k f3 ·F current , M comp3 =k m3 ·F current ×r, where k f3 、k m3 is the compensation coefficient.

7. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The task planning layer adopts a dynamic path planning algorithm based on risk assessment, comprehensively considers wave height, wind speed, ocean current speed and obstacle density to assess environmental risks, obtains environmental risk assessment indicators, and adds the environmental risk assessment indicators and risk weight coefficients to the heuristic function of path planning after weighting, and considers task priority in path planning. For search and rescue missions, when the target location is detected, the path planning is dynamically adjusted to give priority to driving to the target point; for environmental monitoring tasks, according to the preset monitoring routes and stations, combined with real-time environmental information, the navigation sequence and stay time are optimized; the global navigation path generated by the task planning layer is sent to the motion control layer.

8. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: The motion control layer receives the global navigation path instructions of the mission planning layer and the dynamic interference prediction results of the embodied intelligent model, adopts the model predictive control algorithm to perform rolling optimization based on the unmanned boat dynamics model, and compensates for interference and model uncertainty through adaptive sliding mode control, generates motion control instructions and sends them to the actuator layer, wherein the adaptive sliding mode control law u in the adaptive sliding mode control is smc For: u smc =u mpc -Ksgn(s), where u mpc is the control variable output by the model predictive control, K is the sliding mode gain matrix adjusted in real time by the adaptive algorithm, sgn(·) is the sign function, and s(t) is the sliding surface function.

9. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: In the online adaptive optimization, a dual-time-scale online learning algorithm based on dynamic weight adjustment is adopted to assign dynamic weights to different types of real-time feedback information, and the model parameter update is divided into a fast update layer and a slow update layer for separate execution. The fast update layer adopts a stochastic gradient descent-momentum algorithm to update the parameters related to the current short-term environmental changes and control responses, and the slow update layer adopts an adaptive learning rate algorithm to update the parameters related to the long-term performance and stability of the model.

10. The method for dynamic anti-disturbance control of an unmanned boat in high sea conditions based on embodied intelligence according to claim 1, characterized in that: In the online adaptive optimization, the performance of the embodied intelligent model is evaluated by cross-validation. When the evaluated performance index exceeds a preset threshold, the model structure optimization mechanism is triggered to optimize the model structure by adding or deleting the embodied intelligent model neural network layer or adjusting the number of neurons.

Images