A ship intelligent anti-rolling fin control system and method based on multi-source information fusion

Abstract

The application discloses a ship intelligent anti-rolling fin control system and method based on multi-source information fusion. The system comprises: a data perception layer, used for collecting multi-source data in the ship navigation process, the multi-source data comprising ship body motion data, sea state environment data, liquid tank sloshing data and ship loading data; a data fusion layer, used for receiving the multi-source data, and performing time-space alignment and feature fusion on the multi-source data to generate dynamic characteristic parameters representing the ship-liquid tank coupling and the sea wave excitation effect; an intelligent decision-making layer, comprising a controller based on deep reinforcement learning, used for taking the dynamic characteristic parameters as input, calculating and outputting expected action instructions of the anti-rolling fin; and an execution feedback layer, used for receiving the expected action instructions, and performing PID control on the anti-rolling fin based on the expected action instructions and an actual fin angle feedback signal. The application can realize real-time prediction and adaptive suppression of ship roll.

Description

Technical Field

[0001] This invention belongs to the field of ship motion control technology, and more specifically, relates to a ship intelligent anti-roll fin control system and method based on multi-source information fusion. Background Technology

[0002] In the field of marine engineering, anti-roll fins, as an active roll reduction device, function primarily to suppress the ship's rolling motion in waves by generating a hydrodynamic torque opposite to the wave disturbance torque. Since its introduction in the early 20th century, anti-roll fin technology has evolved from mechanical to hydraulic, and from analog to digital control. Currently, mainstream anti-roll fin control systems generally employ a closed-loop control strategy based on sensor feedback. By monitoring motion parameters such as the ship's roll angle and roll rate in real time, they calculate fin angle commands using classical control theory (such as PID control) or modern control theory (such as LQG optimal control), driving the hydraulic actuator to adjust the fin angle and thus generate the required lift.

[0003] While existing technologies can meet the requirements for ship roll reduction to some extent, significant technical bottlenecks remain in practical applications. First, traditional control systems are typically designed based on linear models, treating the hull as a rigid body and neglecting the strong coupling effect between the hull's elastic deformation, liquid sloshing within tanks, and hull motion under complex sea conditions. With the increasing use of green energy sources such as LNG and ammonia fuel on ships, onboard fuel tanks and ballast tanks are more common under certain loading conditions. Liquid sloshing alters the ship's center of gravity and stability height, affecting the dynamic characteristics of roll motion and leading to decreased accuracy of traditional control models, or even control failure. Second, existing systems have limited environmental perception capabilities, relying mostly on passive feedback after hull roll, lacking active detection and prediction of the forward wave field. Control actions are lag-dependent and struggle to cope with sudden wave disturbances. Furthermore, the parameters of traditional control strategies are usually tuned based on specific loading conditions. When parameters such as ship displacement and center of gravity change due to loading variations, the controller struggles to adaptively adjust, resulting in decreased roll reduction performance and weak generalization ability.

[0004] In recent years, with the development of sensor technology and artificial intelligence algorithms, some studies have attempted to introduce intelligent algorithms such as neural networks and fuzzy control into the field of anti-roll fin control. However, these methods still have shortcomings in the effective fusion of multi-source heterogeneous information (such as hull motion, sea state radar, and liquid tank monitoring data), real-time performance assurance, and engineering applications. A complete technical solution that can integrate environmental perception, hull status, and liquid tank sloshing information has not yet been formed. For example, how to combine wave precursor information with liquid tank sloshing state to construct a forward-looking control decision model remains a technical challenge that urgently needs to be solved.

[0005] Therefore, there is an urgent need for an intelligent anti-roll fin control system that can comprehensively perceive the multi-dimensional state of the ship, environment, and liquid tank and has adaptive decision-making capabilities, in order to overcome the defects caused by model simplification, information silos, and control lag in the existing technology, and improve the anti-roll performance and navigation safety of ships in complex sea conditions. Summary of the Invention

[0006] The purpose of this invention is to propose a ship intelligent anti-roll fin control system and method based on multi-source information fusion, which solves the problems of existing anti-roll fin control systems lacking feedforward prediction capability, not considering the influence of coupled resonance between liquid tank sloshing and ship roll, and insufficient adaptability. It realizes swell feedforward prediction and suppression based on multi-source information fusion, active avoidance of ship-liquid tank coupled resonance, and adaptive optimal anti-roll control under different sea states and loading conditions.

[0007] To achieve the above objectives, in a first aspect, the present invention proposes a ship intelligent anti-roll fin control system based on multi-source information fusion, comprising: Data perception layer, data fusion layer, intelligent decision-making layer, and execution feedback layer; The data perception layer is used to collect multi-source data during the ship's navigation process. The multi-source data includes ship motion data, sea state environment data, liquid tank sloshing data, and ship loading data. The data fusion layer is used to receive multi-source data collected by the data perception layer, and to perform spatiotemporal alignment and feature fusion on the multi-source data to generate dynamic feature parameters characterizing the ship-tank coupling and wave excitation. The intelligent decision layer includes a controller based on deep reinforcement learning, which calculates and outputs the desired action command for the anti-roll fins using the dynamic feature parameters as input. The execution feedback layer is used to receive the desired action command and perform PID control on the anti-roll fin based on the desired action command and the actual fin angle feedback signal.

[0008] Optionally, the data sensing layer includes: The hull motion monitoring unit, which includes a fiber optic compass and an inertial measurement unit, is used to monitor the ship's current roll angle, roll rate, and roll acceleration in real time. The marine environment sensing unit includes a wave radar, which is used to detect the main wave direction, significant wave height, and average period of waves in front of the ship. The ship condition monitoring unit includes a liquid tank sloshing monitoring subsystem and a ship loading instrument. The liquid tank sloshing monitoring subsystem includes a radar level gauge array arranged on the top of the liquid tank and a pressure sensor arranged on the bottom of the tank, which is used to collect liquid level data and bottom pressure data at multiple points in the liquid tank. The ship loading instrument is used to obtain the ship's current displacement and center of gravity height.

[0009] Optionally, in the liquid tank sloshing monitoring subsystem, no fewer than three radar level gauges are arranged on the top of each liquid tank in an array; the sampling frequency of both the radar level gauges and the pressure sensor is no less than 10Hz.

[0010] Optionally, the data fusion layer includes: The data preprocessing module is used to preprocess the multi-source data, and the preprocessing includes time synchronization, spatial alignment and noise filtering. The multi-source information fusion module is used to fuse the preprocessed multi-source data using an extended Kalman filter algorithm to generate the dynamic feature parameters; The dynamic characteristic parameters are a 10-dimensional vector, including: current roll angle, roll angular velocity, roll angular acceleration, main wave direction of the preceding wave, significant wave height, average period, liquid level characteristic value of the tank, bilge pressure characteristic value, current displacement, and center of gravity height.

[0011] Optionally, the controller employs a deep deterministic strategy gradient algorithm, with its state space consisting of the dynamic feature parameters output by the data fusion layer, and its action space consisting of the desired fin angle and desired angular velocity of the anti-roll fins on both sides of the ship.

[0012] Optionally, the depth deterministic policy gradient algorithm includes: Actor network and Critic network, both of which are 3-layer fully connected neural networks; The desired fin angle ranges from -25° to +25°, and the desired angular velocity ranges from -15° / s to +15° / s.

[0013] Optionally, the controller is pre-trained in an offline phase using a digital twin environment built upon a six-degree-of-freedom hydrodynamic model of the ship and a CFD simulation model of tank sloshing. The reward function for the pre-training is designed as follows: R=-(α·| |+β·|δ|+γ·| |) in δ is the roll angle, and δ is the fin angle. α represents the rate of change of fin angle, and β and γ are weighting coefficients. After training converges, the network parameters are stored in the controller for online forward computation. The network parameters include the connection weights and biases of each layer of the Actor network and the Critic network.

[0014] Optionally, the execution feedback layer includes: Variable domain fuzzy PID controller, electro-hydraulic servo system, anti-roll fin and fin angle feedback encoder; The variable universe fuzzy PID controller is used to receive the desired action command output by the intelligent decision layer and the actual fin angle feedback signal sent by the fin angle feedback encoder. It adjusts the PID parameters online and dynamically expands and shrinks the universe of discourse through fuzzy rules to drive the electro-hydraulic servo system so that the anti-roll fin tracks the desired action command.

[0015] Optionally, the variable universe of discourse fuzzy PID controller dynamically adjusts the scaling factor of the PID parameters according to the actual deviation magnitude, and dynamically expands and contracts the universe of discourse according to the deviation variation range.

[0016] Secondly, this invention proposes a ship intelligent anti-roll fin control method based on multi-source information fusion, applied to the ship intelligent anti-roll fin control system based on multi-source information fusion as described in the first aspect, comprising the following steps: The data perception layer collects multi-source data during the ship's navigation process, including ship motion data, sea state and environment data, liquid tank sloshing data, and ship loading data. The multi-source data is spatiotemporally aligned and feature-fused through a data fusion layer to generate dynamic feature parameters characterizing the ship-tank coupling and wave excitation. The dynamic feature parameters are input into the trained deep reinforcement learning controller through the intelligent decision layer to calculate the expected action command of the anti-roll fin. The desired action command is received by the execution feedback layer, and the anti-roll fin is subjected to PID control based on the desired action command and the actual fin angle feedback signal.

[0017] The beneficial effects of this invention are as follows: By constructing a control architecture consisting of a data perception layer, a data fusion layer, an intelligent decision-making layer, and an execution feedback layer, the data perception layer simultaneously collects ship motion data, sea state environment data, tank sloshing data, and ship loading data. The data fusion layer performs spatiotemporal alignment and feature fusion on multi-source data to generate dynamic feature parameters characterizing ship-tank coupling and wave excitation. The intelligent decision-making layer uses a deep reinforcement learning controller to output the optimal fin angle command in real time. The execution feedback layer combines the actual fin angle signal to perform PID closed-loop tracking, achieving predictive feedforward suppression before the arrival of swells and effectively avoiding the lag of traditional feedback control. At the same time, by fusing tank sloshing data and loading status, it can actively identify and suppress ship-tank coupling resonance, significantly improving the roll reduction effect under some loading conditions. In addition, the deep reinforcement learning controller, using the fused dynamic features as input, has the ability to adaptively adjust the control strategy under different sea states and loading conditions, overcoming the poor generalization of traditional fixed parameter control, and ultimately greatly improving the ship's roll reduction performance and navigation safety in complex sea states.

[0018] The system of the present invention has other features and advantages that will be apparent from or will be set forth in detail in the accompanying drawings and following detailed description, which together serve to explain the particular principles of the invention. Attached Figure Description

[0019] The above and other objects, features and advantages of the present invention will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0020] Figure 1 A schematic diagram of a ship intelligent anti-roll fin control system based on multi-source information fusion according to Embodiment 1 of the present invention is shown. Detailed Implementation

[0021] The invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0022] Example 1

[0023] like Figure 1 As shown, a ship intelligent anti-roll fin control system based on multi-source information fusion includes: Data perception layer, data fusion layer, intelligent decision-making layer, and execution feedback layer; The data perception layer is used to collect multi-source data during the ship's navigation process. The multi-source data includes ship motion data, sea state and environment data, liquid tank sloshing data, and ship loading data. The data fusion layer is used to receive multi-source data collected by the data perception layer, and to perform spatiotemporal alignment and feature fusion on the multi-source data to generate dynamic feature parameters that characterize the ship-tank coupling and wave excitation. The intelligent decision layer includes a deep reinforcement learning-based controller that takes dynamic feature parameters as input, calculates and outputs the desired action command for the anti-roll fin; The execution feedback layer is used to receive the desired action command and perform PID control on the anti-roll fin based on the desired action command and the actual fin angle feedback signal.

[0024] Specifically, the intelligent anti-roll fin control system for ships based on multi-source information fusion in this embodiment includes a data perception layer, a data fusion layer, an intelligent decision-making layer, and an execution feedback layer. The data perception layer is responsible for collecting multi-source data during the ship's navigation process. This multi-source data covers hull motion data, sea state environment data, liquid tank sloshing data, and ship loading data, thus overcoming the information limitations of traditional systems that rely solely on hull roll feedback and providing a comprehensive state perception foundation for control decisions. The data fusion layer receives various multi-source data from the data perception layer, performs spatiotemporal alignment and feature fusion processing on this data, eliminates differences in time references and spatial coordinates between different sensors, and extracts dynamic feature parameters that comprehensively characterize the coupling effect between the ship and the liquid tank, as well as the excitation effect of waves. This effectively solves the problem of strong coupling modeling of hull elastic deformation, liquid sloshing, and ship motion, providing a high-precision state representation for intelligent decision-making. The intelligent decision-making layer embeds a controller based on deep reinforcement learning. This controller takes the dynamic feature parameters output from the data fusion layer as input and, through autonomous learning and strategy optimization, calculates and outputs the desired action commands (including desired fin angle and angular velocity) that the anti-roll fin should execute in real time. This allows it to adaptively adjust the fin angle commands according to real-time sea conditions and ship status, overcoming the problems of poor adaptability to operating conditions and delayed response to sudden disturbances caused by fixed parameters in traditional control models. The execution feedback layer receives the desired action commands from the intelligent decision-making layer and simultaneously acquires the current actual fin angle feedback signal of the anti-roll fin. Based on the deviation between the desired and actual values, a PID control algorithm is used to drive the anti-roll fin to execute the corresponding actions, forming a closed-loop control that ensures the accuracy and stability of the anti-roll fin's actions. This approach retains the reliability of classical control while incorporating the adaptive advantages of intelligent algorithms. Through the collaborative work of the above four layers, the system realizes closed-loop control of the entire process from multi-source data acquisition, information fusion, intelligent decision-making to precise execution. It has completed the leap from passive response to active prediction and from single control to multi-dimensional fusion in ship roll reduction control. This enables the roll reduction fin system to have the ability to actively predict complex sea conditions, adapt to changes in ship loading, and respond quickly to sudden wave disturbances, which greatly improves the stability and safety of ship navigation. At the same time, it provides an expandable technical framework for intelligent ship control.

[0025] In this embodiment, the data awareness layer includes: The hull motion monitoring unit, which includes a fiber optic compass and an inertial measurement unit, is used to monitor the ship's current roll angle, roll rate, and roll acceleration in real time. The marine environment sensing unit includes a wave radar, which is used to detect the main wave direction, significant wave height, and average period of waves in front of the ship. The ship condition monitoring unit includes a liquid tank sloshing monitoring subsystem and a ship loading instrument. The liquid tank sloshing monitoring subsystem includes a radar level gauge array arranged on the top of the liquid tank and a pressure sensor arranged on the bottom of the tank, which is used to collect liquid level data and bottom pressure data at multiple points in the liquid tank. The ship loading instrument is used to obtain the ship's current displacement and center of gravity height.

[0026] Specifically, the data perception layer includes a hull motion monitoring unit, a sea state environment perception unit, and a ship status monitoring unit. The hull motion monitoring unit, composed of a fiber optic compass and an inertial measurement unit (IMU), can monitor the ship's current roll angle, roll rate, and roll acceleration in real time at a sampling frequency of no less than 100 Hz, providing the system with high-precision instantaneous hull motion status. The sea state environment perception unit uses wave radar to detect the main wave direction, significant wave height, and average period of waves within a 3-nautical-mile radius ahead of the ship at an update frequency of no less than 1 Hz, thereby obtaining forward-looking information on wave excitation. The ship condition monitoring unit comprises a liquid tank sloshing monitoring subsystem and a ship loading instrument. The liquid tank sloshing monitoring subsystem arranges radar level gauges in an array on the top of each liquid tank (including ballast tanks and fuel tanks), while pressure sensors are arranged on the bottom of the tanks. All sensors synchronously collect liquid level data and bottom pressure data at multiple points within the tanks at a frequency of no less than 10Hz, reflecting the dynamic characteristics of liquid sloshing in real time. The ship loading instrument obtains the ship's current displacement and center of gravity height, characterizing the ship's overall loading status. Through the coordinated operation of these three monitoring units, the data perception layer can comprehensively collect four types of multi-source data: ship motion, sea conditions ahead, liquid tank sloshing, and ship loading, providing rich and accurate raw information for subsequent information fusion.

[0027] In this embodiment, in the liquid tank sloshing monitoring subsystem, no fewer than three radar level gauges are arranged on the top of each liquid tank in an array; the sampling frequency of both the radar level gauges and the pressure sensor is no less than 10Hz.

[0028] Specifically, the configuration of the liquid tank sloshing monitoring subsystem is as follows: For each liquid tank on the ship (including ballast tanks, fuel tanks, etc.), at least three radar level gauges are arranged on its top. These radar level gauges are distributed in an array (e.g., uniformly arranged along the longitudinal or transverse direction of the liquid tank) to simultaneously acquire liquid level information at different locations within the tank. By arranging multi-point array radar level gauges on the top, the spatial distribution of the liquid surface during sloshing within the tank can be more comprehensively reflected, avoiding the representativeness issues that may arise from single-point measurements. Simultaneously, pressure sensors are arranged at the bottom of each liquid tank to measure the hydrostatic and dynamic pressure of the liquid at the bottom. The sampling frequency of all sensors (including radar level gauges and pressure sensors) is no less than 10Hz to ensure rapid capture of the dynamic process of liquid sloshing and to acquire high temporal resolution data on liquid level and pressure changes. The collected multi-point liquid level data and bottom pressure data are used as the output of the liquid tank sloshing monitoring subsystem and provided to the data fusion layer for subsequent spatiotemporal alignment and feature fusion processing.

[0029] In this embodiment, the data fusion layer includes: The data preprocessing module is used to preprocess multi-source data, including time synchronization, spatial alignment, and noise filtering. The multi-source information fusion module is used to fuse preprocessed multi-source data using the extended Kalman filter algorithm to generate dynamic feature parameters; The dynamic characteristic parameters are a 10-dimensional vector, including: current roll angle, roll rate, roll acceleration, main wave direction of the wave ahead, significant wave height, average period, liquid level characteristic value of the tank, bilge pressure characteristic value, current displacement, and center of gravity height.

[0030] Specifically, the data fusion layer includes a data preprocessing module and a multi-source information fusion module. The data preprocessing module preprocesses the multi-source data collected by the data sensing layer, including time synchronization, spatial alignment, and noise filtering. Time synchronization involves aligning the timestamps of asynchronous data generated by different sensors due to their different sampling frequencies (e.g., 100Hz for the inertial measurement unit, 1Hz for the wave radar, and 10Hz for the level gauge and pressure sensor), unifying the data from each sensor to the same time reference through interpolation or extrapolation. Spatial alignment involves converting the coordinate differences caused by the different installation positions of various sensors (such as radar level gauges, pressure sensors, IMUs, and wave radars), unifying all data into the ship's coordinate system. Noise filtering uses methods such as low-pass filtering or median filtering to remove outliers and environmental noise from the original signal, improving data quality. The preprocessed multi-source data is then fed into the multi-source information fusion module.

[0031] The multi-source information fusion module employs the Extended Kalman Filter (EKF) algorithm to fuse preprocessed multi-source data. EKF uses a six-degree-of-freedom nonlinear motion model of the ship as the state equation and sensor observation data as the measurement equation. Through a two-step recursive prediction-update process, it estimates the optimal state of the system in the sense of minimum variance, thereby achieving feature-level fusion of asynchronous, heterogeneous multi-source data. The resulting dynamic feature parameters are a 10-dimensional vector, specifically including: current roll angle, roll rate, roll acceleration, main wave direction of the preceding wave, significant wave height, average period, tank level characteristic value, bilge pressure characteristic value, current displacement, and center of gravity height. This 10-dimensional vector comprehensively characterizes the ship-tank coupling dynamics and wave excitation, providing a complete and accurate state description for the intelligent decision-making layer.

[0032] In this embodiment, the controller adopts a deep deterministic strategy gradient algorithm. Its state space is the dynamic feature parameters output by the data fusion layer, and its action space is the expected fin angle and expected angular velocity of the anti-roll fins on the port and starboard sides of the ship.

[0033] Specifically, the controller within the intelligent decision-making layer employs the Deep Deterministic Policy Gradient (DDPG) algorithm. The controller's state space is directly derived from the dynamic feature parameters output by the data fusion layer, namely the aforementioned 10-dimensional vector, including the current roll angle, roll rate, roll acceleration, main wave direction of the preceding wave, significant wave height, average period, tank level characteristic value, bilge pressure characteristic value, current displacement, and center of gravity height. This comprehensively characterizes the current state of the ship-wave-tank coupling system. The controller's action space is defined as the desired fin angle and desired angular velocity of the anti-roll fins on both sides of the ship, forming a 4-dimensional action vector. Specifically, each anti-roll fin outputs a desired fin angle (range -25° to +25°) and a desired angular velocity (range -15° / s to +15° / s). The controller takes the dynamic feature vector output by the data fusion layer every 0.1 seconds as input, and calculates forward through the Actor network in the DDPG algorithm to output the optimal expected fin angle and expected angular velocity, which are then passed to the execution feedback layer as expected action commands, thereby achieving adaptive optimal control of the anti-roll fin.

[0034] In this embodiment, the deep deterministic policy gradient algorithm includes: Both the Actor network and the Critic network are 3-layer fully connected neural networks. The desired fin angle ranges from -25° to +25°, and the desired angular velocity ranges from -15° / s to +15° / s.

[0035] In this embodiment, the controller is pre-trained in a digital twin environment built on a six-degree-of-freedom hydrodynamic model of the ship and a CFD simulation model of liquid tank sloshing during the offline phase. The reward function for the pre-training is designed as follows: R=-(α·| |+β·|δ|+γ·| |) in δ is the roll angle, and δ is the fin angle. α represents the rate of change of fin angle, and β and γ are weighting coefficients. After training converges, the network parameters are stored in the controller for online forward computation. The network parameters include the connection weights and biases of each layer of the Actor and Critic networks.

[0036] Specifically, the deep deterministic policy gradient algorithm comprises an Actor network (policy network) and a Critic network (value network), both of which are 3-layer fully connected neural networks. The input layer contains 128 nodes, the hidden layer contains 64 nodes, and the output layer contains 32 nodes. The activation function is ReLU. In the controller's output desired action commands, the desired fin angle range for the port and starboard anti-roll fins is -25° to +25°, and the desired angular velocity range is -15° / s to +15° / s. During the offline pre-training phase, a digital twin environment is built based on a six-DOF hydrodynamic model of the ship and a liquid tank sloshing CFD simulation model. This generates a large amount of sample data under different sea states and loading conditions, and the deep deterministic policy gradient algorithm is used for training. The pre-training reward function is designed as R = -(α·| | + β·|δ| + γ·| |), where δ is the roll angle, and δ is the fin angle. Let α, β, and γ represent the rate of change of the fin angle, and α, β, and γ be the weighting coefficients. This reward function minimizes the roll angle while also considering the control cost of the fin angle and its rate of change. After training convergence, the network parameters (including the connection weights and biases of each layer in the Actor and Critic networks) are embedded in the shipboard controller for online forward computation. In online applications, the controller receives dynamic feature parameters output from the data fusion layer every 0.1 seconds and directly calculates the desired fin angle and desired angular velocity through the Actor network, eliminating the need for online updates to the network parameters and thus meeting the real-time requirements of ship control.

[0037] In this embodiment, the execution feedback layer includes: Variable domain fuzzy PID controller, electro-hydraulic servo system, anti-roll fin and fin angle feedback encoder; The variable universe of discourse fuzzy PID controller is used to receive the desired action command output by the intelligent decision layer and the actual fin angle feedback signal sent by the fin angle feedback encoder. It adjusts the PID parameters online and dynamically expands and shrinks the universe of discourse through fuzzy rules to drive the electro-hydraulic servo system so that the anti-roll fin tracks the desired action command.

[0038] Specifically, the execution feedback layer includes a variable universe of discourse fuzzy PID controller, an electro-hydraulic servo system, anti-roll fins, and a fin angle feedback encoder. The variable universe of discourse fuzzy PID controller receives the desired action commands (i.e., desired fin angle and desired angular velocity) output from the intelligent decision layer, and simultaneously receives the actual fin angle feedback signal sent in real time by the fin angle feedback encoder. It compares the desired value with the actual value to obtain the deviation and the rate of change of deviation as input quantities. The controller integrates a fuzzy inference mechanism, which can adjust the proportional, integral, and derivative coefficients (i.e., proportional factors) of the PID controller online according to the magnitude and trend of the current deviation, thereby achieving real-time optimization of control parameters. Furthermore, the variable universe of discourse fuzzy PID controller also has a dynamic expansion and contraction of the universe of discourse function: when the deviation is large, it automatically expands the universe of discourse range to improve response speed; when the deviation decreases, it automatically shrinks the universe of discourse range to enhance control accuracy. By combining online adjustment of PID parameters using fuzzy rules with dynamic expansion and contraction of the universe of discourse, the controller can maintain excellent dynamic response and steady-state accuracy under different operating conditions. The control commands output by the variable domain fuzzy PID controller drive the electro-hydraulic servo system, causing the anti-roll fin to rotate to the desired angle. Simultaneously, the fin angle feedback encoder transmits the actual fin angle back to the controller in real time, forming a position closed-loop feedback. This execution feedback layer ensures that the anti-roll fin can quickly and accurately track the desired action commands given by the intelligent decision-making layer, thereby effectively suppressing the ship's roll.

[0039] In this embodiment, the variable universe of discourse fuzzy PID controller dynamically adjusts the scaling factor of the PID parameters according to the actual deviation magnitude, and dynamically expands and contracts the universe of discourse according to the deviation variation range.

[0040] Specifically, the variable universe of discourse fuzzy PID controller dynamically adjusts the proportional factor of the PID parameters according to the actual deviation magnitude and dynamically expands and contracts the universe of discourse according to the deviation variation range. Specifically, the controller uses the deviation *e* between the desired fin angle and the actual fin angle and its rate of change *ec* as input variables, and performs fuzzy inference through a preset fuzzy rule table. When the deviation *e* is large, the controller automatically increases the proportional factor *Kp* to enhance the system's rapid response capability, causing the anti-roll fin to quickly approach the desired angle; when the deviation *e* decreases to a smaller range, the controller appropriately decreases the proportional factor *Kp* and increases the integral factor *Ki* to eliminate steady-state error and avoid overshoot. Simultaneously, the controller dynamically adjusts the universe of discourse of the input variables (*e* and *ec*) according to the variation range of the deviation *e*: when the absolute value of the deviation *e* is large, the controller automatically expands the range of the universe of discourse (e.g., expanding the basic universe of discourse of *e* from [-6,6] to [-12,12]), allowing the fuzzy partitioning to cover a larger input space and avoiding control saturation; when the deviation *e* gradually converges to near zero, the controller dynamically shrinks the universe of discourse to improve the resolution of small deviations, thereby achieving fine adjustment. Through the coordinated operation of online adjustment of the scaling factor and dynamic expansion of the domain of discourse, the variable domain of discourse fuzzy PID controller can adaptively balance response speed and control accuracy under different operating conditions, effectively overcoming the shortcomings of fixed parameters and invariant domain of discourse of traditional PID controllers, and ensuring high-precision tracking of the anti-roll fins to the desired action commands.

[0041] Example 2

[0042] This embodiment provides a ship intelligent anti-roll fin control system based on multi-source information fusion, including a data perception layer, a data fusion layer, an intelligent decision-making layer, and an execution feedback layer.

[0043] The data sensing layer comprises a hull motion monitoring unit, a sea state environment sensing unit, and a ship condition monitoring unit. The hull motion monitoring unit, consisting of a fiber optic compass and an inertial measurement unit (IMU), monitors the ship's current roll angle, roll rate, and roll acceleration in real time at a sampling frequency of at least 100 Hz. The sea state environment sensing unit uses wave radar to detect the main wave direction, significant wave height, and average period of waves within a 3-nautical-mile radius ahead of the ship at an update frequency of at least 1 Hz. The ship condition monitoring unit includes a tank sloshing monitoring subsystem and a ship loading instrument: the tank sloshing monitoring subsystem arranges at least three radar level gauges in an array on top of each tank (including ballast tanks and fuel tanks), and pressure sensors are arranged on the bilge. All sensors collect level data and bilge pressure data from multiple points within the tanks at a frequency of at least 10 Hz; the ship loading instrument obtains the ship's current displacement and center of gravity height. Through these units, the data sensing layer collects multi-source data including hull motion data, sea state environment data, tank sloshing data, and ship loading data.

[0044] The data fusion layer comprises a data preprocessing module and a multi-source information fusion module. The data preprocessing module performs time synchronization, spatial alignment, and noise filtering on the multi-source data collected by the data sensing layer: time synchronization uses interpolation to align sensor data from different sampling frequencies to the same time reference; spatial alignment transforms the measurements from each sensor to a unified ship coordinate system; noise filtering uses low-pass or median filtering to remove outliers and high-frequency noise. The preprocessed data is then fed into the multi-source information fusion module, which uses the Extended Kalman Filter (EKF) algorithm for fusion. EKF uses a six-degree-of-freedom nonlinear motion model of the ship as the state equation and the observation data from each sensor as the measurement equation. Through a two-step recursive estimation of prediction and update, it estimates the optimal state of the system, generating a 10-dimensional dynamic characteristic parameter vector, specifically including: current roll angle, roll rate, roll acceleration, main wave direction of the wave ahead, significant wave height, average period, tank level characteristic value, bilge pressure characteristic value, current displacement, and center of gravity height. This 10-dimensional vector comprehensively characterizes the ship-tank coupling dynamics and the excitation effect of waves.

[0045] The intelligent decision-making layer comprises a deep reinforcement learning-based controller employing the Deep Deterministic Policy Gradient (DDPG) algorithm. The DDPG algorithm consists of an Actor network (policy network) and a Critic network (value network), both being 3-layer fully connected neural networks with 128 nodes in the input layer, 64 nodes in the hidden layer, and 32 nodes in the output layer, using ReLU as the activation function. The controller's state space is a 10-dimensional dynamic feature vector output from the data fusion layer, and its action space consists of the expected fin angles and angular velocities of the anti-roll fins on both sides of the ship, forming a 4-dimensional action vector. The expected fin angle ranges from -25° to +25°, and the expected angular velocity ranges from -15° / s to +15° / s. During the offline phase, the controller is pre-trained in a digital twin environment built upon a six-DOF hydrodynamic model of the ship and a CFD simulation model of tank sloshing, generating a large amount of sample data under different sea states and loading conditions. The pre-training reward function is designed as R = -(α·| |+β·|δ|+γ·| |), where δ is the roll angle, and δ is the fin angle. Let α be the rate of change of fin angle, and β be the weighting coefficients. After training convergence, the connection weights and biases of each layer of the Actor network and Critic network are fixed in the shipboard controller. In online application, the controller receives a 10-dimensional state vector output from the data fusion layer every 0.1 seconds, calculates the desired fin angle and desired angular velocity through the Actor network, and outputs them as the desired action command to the execution feedback layer.

[0046] The execution feedback layer comprises a variable domain fuzzy PID controller, an electro-hydraulic servo system, a roll stabilizer fin, and a fin angle feedback encoder. The variable domain fuzzy PID controller receives the desired action command (desired fin angle and desired angular velocity) output from the intelligent decision layer, and simultaneously receives the actual fin angle feedback signal sent in real time by the fin angle feedback encoder. The controller uses the deviation between the desired and actual values, and the rate of change of the deviation, as input. It adjusts the proportional factor of the PID parameters online using fuzzy rules and dynamically expands and contracts the domain according to the deviation range: when the deviation is large, the domain is expanded to improve response speed; when the deviation is small, the domain is contracted to improve control accuracy. The control command output from the variable domain fuzzy PID controller drives the electro-hydraulic servo system, enabling the roll stabilizer fin to accurately track the desired angle, and the fin angle feedback encoder forms a closed-loop position feedback. Through the coordinated operation of these four layers, the system achieves closed-loop control throughout the entire process, from multi-source data acquisition, information fusion, intelligent decision-making to precise execution.

[0047] The system's workflow is as follows: First, the ship motion monitoring unit (fiber optic compass and inertial measurement unit) in the data perception layer collects the ship's current roll angle, roll rate, and roll acceleration in real time at a sampling frequency of no less than 100Hz; the sea state environment perception unit (wave radar) detects the main wave direction, significant wave height, and average period of waves within a 3-nautical-mile radius ahead of the ship at an update frequency of no less than 1Hz; the liquid tank sloshing monitoring subsystem arranges no less than three radar level gauges in an array on the top of each liquid tank and arranges pressure sensors on the bottom of the tank, simultaneously collecting multi-point liquid level data and bottom pressure data at a frequency of no less than 10Hz; the ship loading instrument obtains the current displacement and center of gravity height. The above multi-source data (including ship motion data, sea state environment data, liquid tank sloshing data, and ship loading data) are transmitted to the data fusion layer in real time.

[0048] The data fusion layer first performs time synchronization (aligning data from different sampling frequencies to the same time base), spatial alignment (converting data from various sensors to a unified ship coordinate system), and noise filtering (removing outliers and high-frequency interference) on the multi-source data through a data preprocessing module. The preprocessed data is then fed into the multi-source information fusion module, which employs the Extended Kalman Filter (EKF) algorithm. Using the ship's six-degree-of-freedom nonlinear motion model as the state equation and the observations from each sensor as the measurement equation, it generates a 10-dimensional dynamic feature parameter vector through a two-step recursive estimation process of prediction and update. This vector includes: current roll angle, roll rate, roll acceleration, main wave direction of the wave ahead, significant wave height, average period, tank level characteristic value, bilge pressure characteristic value, current displacement, and center of gravity height. This 10-dimensional vector comprehensively represents the ship-tank coupling dynamics and wave excitation effects, and is output to the intelligent decision-making layer.

[0049] The deep reinforcement learning controller in the intelligent decision-making layer (using the Deep Deterministic Policy Gradient (DDPG) algorithm) receives a 10-dimensional dynamic feature vector output from the data fusion layer every 0.1 seconds as input to the state space. The controller internally contains pre-trained network parameters (connection weights and biases of each layer in the Actor and Critic networks). The pre-training phase utilizes a digital twin environment built based on a six-DOF hydrodynamic model of a ship and a CFD simulation model of liquid tank sloshing. The reward function is R = -(α·| | + β·|δ| + γ·| |). In online applications, the controller performs forward calculations through the Actor network and directly outputs the desired fin angle (range -25° to +25°) and desired angular velocity (range -15° / s to +15° / s) of the anti-roll fins on both sides, which are then transmitted to the execution feedback layer as desired action commands.

[0050] In the feedback layer, the variable universe of discourse fuzzy PID controller receives the desired action command and simultaneously receives the actual fin angle feedback signal sent in real time by the fin angle feedback encoder. The controller calculates the deviation between the desired and actual values ​​and the rate of change of the deviation, adjusts the proportional factor of the PID parameters online using fuzzy rules, and dynamically expands and contracts the universe of discourse according to the range of deviation changes (expanding the universe of discourse to improve response speed when the deviation is large, and shrinking the universe of discourse to improve control accuracy when the deviation is small). The controller's output command drives the electro-hydraulic servo system, causing the anti-roll fin to rotate to the desired angle. The fin angle feedback encoder transmits the actual fin angle back in real time, forming a position closed loop. Thus, the system completes a full control cycle and continuously repeats the above process, achieving real-time prediction and adaptive suppression of ship roll.

[0051] Example 3

[0052] A ship intelligent anti-roll fin control method based on multi-source information fusion, applied to the ship intelligent anti-roll fin control system based on multi-source information fusion described in Example 1, includes the following steps: The data perception layer collects multi-source data during the ship's navigation process, including ship motion data, sea state and environment data, liquid tank sloshing data, and ship loading data. The multi-source data is spatiotemporally aligned and feature-fused through a data fusion layer to generate dynamic feature parameters characterizing the ship-tank coupling and wave excitation. The dynamic feature parameters are input into the trained deep reinforcement learning controller through the intelligent decision layer to calculate the expected action command of the anti-roll fin. The desired action command is received by the execution feedback layer, and the anti-roll fin is subjected to PID control based on the desired action command and the actual fin angle feedback signal.

[0053] Specifically, firstly, multi-source data is collected during the ship's navigation process through the data perception layer. This multi-source data includes hull motion data (such as roll angle, roll rate, and roll acceleration), sea state data (such as the main wave direction, significant wave height, and average period of the waves ahead), tank sloshing data (such as liquid levels at multiple points within the tanks and bilge pressure), and ship loading data (such as current displacement and center of gravity height). The data perception layer synchronously acquires the above data at different sampling frequencies using equipment such as fiber optic compasses, inertial measurement units, wave radar, radar level gauge arrays, pressure sensors, and ship loading instruments, providing the system with comprehensive state awareness information.

[0054] Secondly, the multi-source data is spatiotemporally aligned and feature-fused through a data fusion layer to generate dynamic characteristic parameters characterizing the ship-tank coupling and wave excitation. Since the sampling frequencies, installation positions, and coordinate systems of each sensor are different, the data fusion layer first preprocesses the raw data, including time synchronization (interpolating asynchronous data to the same time base), spatial alignment (converting all data to a unified ship coordinate system), and noise filtering (removing outliers and high-frequency interference). Subsequently, an extended Kalman filter algorithm is used to perform feature-level fusion on the preprocessed data. Using the ship's six-degree-of-freedom nonlinear motion model as the state equation and the observations from each sensor as the measurement equation, the optimal state of the system is recursively estimated, outputting a 10-dimensional dynamic characteristic parameter vector. This vector encompasses the current roll angle, roll rate, roll acceleration, main wave direction of the preceding wave, significant wave height, average period, tank level characteristic value, bilge pressure characteristic value, current displacement, and center of gravity height, thus comprehensively reflecting the ship-tank coupling dynamics and external wave excitation.

[0055] Next, the dynamic feature parameters are input into the trained deep reinforcement learning controller through the intelligent decision layer to calculate the expected action command for the anti-roll fins. The controller adopts the Deep Deterministic Policy Gradient (DDPG) algorithm, with its state space being a 10-dimensional dynamic feature vector output by the data fusion layer, and its action space being the expected fin angle and expected angular velocity of the anti-roll fins on both the port and starboard sides. The controller has been pre-trained offline in a digital twin environment built based on a six-degree-of-freedom hydrodynamic model of the ship and a CFD simulation model of tank sloshing. The reward function is designed to minimize the cost of changes in roll angle and fin angle. After training convergence, the network parameters are fixed in the controller. In online application, the controller receives the current dynamic feature parameters every 0.1 seconds and directly outputs the expected fin angle (range -25° to +25°) and expected angular velocity (range -15° / s to +15° / s) as the expected action command through forward calculation via the Actor network.

[0056] Finally, the desired action command is received by the execution feedback layer, and PID control is performed on the anti-roll fin based on the desired action command and the actual fin angle feedback signal. The execution feedback layer includes a variable universe of discourse fuzzy PID controller, an electro-hydraulic servo system, the anti-roll fin, and a fin angle feedback encoder. The variable universe of discourse fuzzy PID controller compares the desired fin angle with the actual fin angle (feedback from the encoder in real time) to obtain the deviation and the rate of change of the deviation. The proportional factor of the PID parameters is adjusted online through fuzzy rules, and the universe of discourse is dynamically expanded and contracted according to the range of deviation changes to optimize the response speed and control accuracy. The command output by the controller drives the electro-hydraulic servo system, enabling the anti-roll fin to accurately track the desired angle. At the same time, the fin angle feedback encoder forms a position closed loop to ensure the accuracy of the executed action. Through the cyclic execution of the above four steps, the system achieves real-time prediction and adaptive suppression of ship roll.

[0057] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A ship intelligent anti-roll fin control system based on multi-source information fusion, characterized in that, include: Data perception layer, data fusion layer, intelligent decision-making layer, and execution feedback layer; The data perception layer is used to collect multi-source data during the ship's navigation process. The multi-source data includes ship motion data, sea state environment data, liquid tank sloshing data, and ship loading data. The data fusion layer is used to receive multi-source data collected by the data perception layer, and to perform spatiotemporal alignment and feature fusion on the multi-source data to generate dynamic feature parameters characterizing the ship-tank coupling and wave excitation. The intelligent decision layer includes a controller based on deep reinforcement learning, which calculates and outputs the desired action command for the anti-roll fins using the dynamic feature parameters as input. The execution feedback layer is used to receive the desired action command and perform PID control on the anti-roll fin based on the desired action command and the actual fin angle feedback signal.

2. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 1, characterized in that, The data perception layer includes: The hull motion monitoring unit, which includes a fiber optic compass and an inertial measurement unit, is used to monitor the ship's current roll angle, roll rate, and roll acceleration in real time. The marine environment sensing unit includes a wave radar, which is used to detect the main wave direction, significant wave height, and average period of waves in front of the ship. The ship condition monitoring unit includes a liquid tank sloshing monitoring subsystem and a ship loading instrument. The liquid tank sloshing monitoring subsystem includes a radar level gauge array arranged on the top of the liquid tank and a pressure sensor arranged on the bottom of the tank, which is used to collect liquid level data and bottom pressure data at multiple points in the liquid tank. The ship loading instrument is used to obtain the ship's current displacement and center of gravity height.

3. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 2, characterized in that, In the liquid tank sloshing monitoring subsystem, no fewer than three radar level gauges are arranged in an array on the top of each liquid tank; the sampling frequency of both the radar level gauges and the pressure sensor is no less than 10Hz.

4. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 1, characterized in that, The data fusion layer includes: The data preprocessing module is used to preprocess the multi-source data, and the preprocessing includes time synchronization, spatial alignment and noise filtering. The multi-source information fusion module is used to fuse the preprocessed multi-source data using an extended Kalman filter algorithm to generate the dynamic feature parameters; The dynamic characteristic parameters are a 10-dimensional vector, including: current roll angle, roll angular velocity, roll angular acceleration, main wave direction of the preceding wave, significant wave height, average period, liquid level characteristic value of the tank, bilge pressure characteristic value, current displacement, and center of gravity height.

5. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 1, characterized in that, The controller employs a deep deterministic strategy gradient algorithm. Its state space consists of the dynamic feature parameters output by the data fusion layer, and its action space consists of the desired fin angle and desired angular velocity of the anti-roll fins on both sides of the ship.

6. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 5, characterized in that, The depth-deterministic policy gradient algorithm includes: Actor network and Critic network, both of which are 3-layer fully connected neural networks; The desired fin angle ranges from -25° to +25°, and the desired angular velocity ranges from -15° / s to +15° / s.

7. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 6, characterized in that, The controller is pre-trained in an offline phase using a digital twin environment built upon a six-degree-of-freedom hydrodynamic model of the ship and a CFD simulation model of tank sloshing. The reward function for the pre-training is designed as follows: R=-(α·| |+b·|d|+c·| |) in δ is the roll angle, and δ is the fin angle. α represents the rate of change of fin angle, and β and γ are weighting coefficients. After training converges, the network parameters are stored in the controller for online forward computation. The network parameters include the connection weights and biases of each layer of the Actor network and the Critic network.

8. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 1, characterized in that, The execution feedback layer includes: Variable domain fuzzy PID controller, electro-hydraulic servo system, anti-roll fin and fin angle feedback encoder; The variable universe fuzzy PID controller is used to receive the desired action command output by the intelligent decision layer and the actual fin angle feedback signal sent by the fin angle feedback encoder. It adjusts the PID parameters online and dynamically expands and shrinks the universe of discourse through fuzzy rules to drive the electro-hydraulic servo system so that the anti-roll fin tracks the desired action command.

9. The intelligent ship roll stabilization fin control system based on multi-source information fusion according to claim 8, characterized in that, The variable universe of discourse fuzzy PID controller dynamically adjusts the scaling factor of the PID parameters according to the actual deviation magnitude, and dynamically expands and contracts the universe of discourse according to the deviation variation range.

10. A ship intelligent anti-roll fin control method based on multi-source information fusion, applied to the ship intelligent anti-roll fin control system based on multi-source information fusion as described in any one of claims 1 to 9, characterized in that, Includes the following steps: The data perception layer collects multi-source data during the ship's navigation process, including ship motion data, sea state and environment data, liquid tank sloshing data, and ship loading data. The multi-source data is spatiotemporally aligned and feature-fused through a data fusion layer to generate dynamic feature parameters characterizing the ship-tank coupling and wave excitation. The dynamic feature parameters are input into the trained deep reinforcement learning controller through the intelligent decision layer to calculate the expected action command of the anti-roll fin. The desired action command is received by the execution feedback layer, and the anti-roll fin is subjected to PID control based on the desired action command and the actual fin angle feedback signal.

Images