A deep neural network-based intelligent control optimization method for ship hybrid propulsion

Abstract

This invention discloses an intelligent control optimization method for ship hybrid propulsion based on deep neural networks. By setting a generalized force control target, constructing a supervised training dataset, and establishing a deep neural network model with an encoder and decoder structure, the encoder maps the target generalized force to the thrust and angle commands of the propeller. The decoder outputs the generalized force inversely based on the control commands. During training, a multi-objective loss function with six sub-items is introduced to construct the physical constraints related to the propeller into the form of a loss function. The optimal propeller command that satisfies various constraints is found. The model receives real-time control command input and can quickly output propeller control commands that meet the constraints, realizing optimized thrust allocation. While ensuring that the physical constraints of the propeller are met, the method balances thrust allocation accuracy and energy consumption, accurately tracks while reducing energy consumption and excessive use of actuators, solves the ship thrust allocation problem under complex constraints, and efficiently completes the ship trajectory tracking task.

Description

Technical Field

[0001] This invention belongs to the field of ship motion control technology, specifically relating to an intelligent control optimization method for ship hybrid propulsion based on deep neural networks. Background Technology

[0002] With the rapid development of computer technology, artificial intelligence, big data processing and automation technology, research on ship control systems has ushered in new technological opportunities. In particular, autonomous navigation systems, represented by intelligent ships, have become one of the key technologies in modern marine operations. Ships with dynamic positioning capabilities, as pioneers in the application of automated navigation and precise positioning technologies, have attracted widespread attention from the academic community and become an important research direction in the field of marine engineering.

[0003] Compared to land-based operations, the marine environment is far more complex and variable. In this context, achieving precise trajectory tracking and control of the vessel presents a greater challenge to the control system. The thrust distribution module, as the core of the ship's entire dynamic positioning system, primarily functions to rapidly calculate the thrust and angle of each propeller within a specified control cycle, thereby meeting the force and torque distribution requirements of the upper-level controller. Therefore, the thrust distribution process, to a certain extent, determines the success or failure of the ship's mission.

[0004] Currently, most ships' propulsion systems are typically equipped with a single type of propeller, such as an azimuth thruster or a waterjet thruster. This configuration limits the ship's maneuverability to some extent, especially during turning maneuvers, often resulting in a large turning radius. To overcome this deficiency, this invention proposes a hybrid propulsion control optimization scheme that combines two azimuth thrusters with two waterjet thrusters, thereby significantly improving the ship's maneuverability during navigation, especially exhibiting superior flexibility and handling performance in complex navigation and turning operations.

[0005] Deep Neural Networks (DNNs) are an effective means of ensuring precise motion control of ships through thrust distribution, providing a new design approach for thrust distribution optimization under multiple constraints. Currently, DNN methods have been successfully applied to various control systems. In recent years, this technology has been increasingly applied to the marine field. Therefore, DNNs can be used to provide an efficient evaluation of the mapping between motion controller requests and executable thruster commands. By constructing a large amount of high-quality thruster data to train a neural network similar to an autoencoder, the overall motion objectives and thruster constraints can be satisfied. Summary of the Invention

[0006] The technical problem that the present invention aims to solve is that there are multiple constraints in the process of ship thrust distribution, which greatly complicates the optimization solution and causes large distribution errors.

[0007] The present invention provides a method for intelligent control optimization of ship hybrid propulsion based on deep neural networks, the method comprising the following steps: The motion controller sets the ship's navigation trajectory information based on the control objectives required by the ship under different states, and obtains multiple target generalized forces; Multiple target generalized forces are fed into a deep neural network model with the optimization objectives of minimizing energy consumption and allocation error for control allocation operations. The optimization problem of multiple nonlinear constraints of each target generalized force is solved through the forward propagation process, and the thruster control command that satisfies the constraints is output. The thruster then performs the corresponding action, ultimately achieving the ship's trajectory tracking.

[0008] Preferably, the constraints of the deep neural network model are the thrust and azimuth angle limits for each thruster.

[0009] Preferably, the physical constraints in the thrust allocation process are considered to guide the neural network to learn the allocation strategy, a comprehensive loss function is constructed, and the model is trained based on the comprehensive loss function to obtain the deep neural network model to be adjusted.

[0010] Preferably, based on the physical constraints in the thrust allocation process, the allocation accuracy and energy consumption are considered to guide the neural network to learn the allocation strategy and construct a comprehensive loss function.

[0011] Preferably, the weights of the comprehensive loss function are adjusted through experiments, and then the deep neural network model to be adjusted is obtained. When in use, multiple target generalized forces and real-time control commands are input into the deep neural network model, and the thruster control commands that meet the constraints are output.

[0012] Preferably, the physical constraints in the thrust distribution process are calculated based on the thruster control commands, using the force and torque commands from the upper controller as control inputs, to calculate the thrust magnitude and azimuth angle of each thruster, thereby generating longitudinal force, lateral force, and rotational torque commands.

[0013] Preferably, the comprehensive loss function includes a generalized force matching loss term, an allocation error minimization loss term, a thruster control command amplitude limitation loss term, a thruster command change rate limitation loss term, a power loss term, and an azimuth angle exclusion zone penalty term.

[0014] Preferably, the deep neural network model adopts an "encoder-decoder" structure. The encoder realizes the nonlinear mapping from the target generalized force to the thruster control command through a multi-layer neural network. The decoder reconstructs the generalized force according to the control command and compares it with the original target, thus forming the supervision signal and physical consistency constraint during the training phase.

[0015] Preferably, the supervision dataset of the deep neural network model is constructed by combining the thrust magnitude and azimuth angle of each thruster with the target generalized force and the corresponding optimal thruster command.

[0016] Preferably, the target generalized force is the force and torque commanded by the dynamic positioning control system.

[0017] This invention provides a method for intelligent control optimization of ship hybrid propulsion based on deep neural networks. By setting a generalized force control target, constructing a supervised training dataset, and establishing a deep neural network model with an encoder and decoder structure, the encoder maps the target generalized force to thrust and angle commands of the propeller. The decoder outputs the generalized force based on the control commands. During training, a multi-objective loss function with six sub-items is introduced, constructing the propeller-related physical constraints into a loss function form to find the optimal propeller command that satisfies various constraints. The model receives real-time control command input and can quickly output propeller control commands that meet the constraints, achieving optimized thrust allocation. While ensuring that the propeller's physical constraints are met, it balances thrust allocation accuracy and energy consumption, accurately tracking while reducing energy consumption and excessive actuator use. This solves the ship thrust allocation problem under complex constraints and efficiently completes ship trajectory tracking tasks. Attached Figure Description

[0018] Figure 1 A schematic diagram of the control system structure of a ship hybrid propulsion intelligent control optimization method based on deep neural networks, provided for an embodiment of the present invention; Figure 2 A deep neural network architecture diagram provided for embodiments of the present invention; Figure 3 A schematic diagram of the northeast coordinate system and the ship's hull coordinate system provided for an embodiment of the present invention. Detailed Implementation

[0019] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0020] This invention provides a ship hybrid propulsion intelligent control optimization method based on deep neural networks, applicable to the ship thrust distribution problem under complex multi-constraint environments. Especially in ship dynamic positioning systems facing challenges such as control accuracy, physical constraints, and energy efficiency optimization, it can efficiently achieve real-time control. The method includes the following steps: like Figure 1 As shown, firstly, the motion controller obtains the system's control input based on the pre-set ship navigation trajectory information. Then, the target generalized force vector is input into the constructed deep neural network model for control allocation. Through the forward propagation process, the control command vector for the thruster is output, and the thruster performs corresponding actions, ultimately achieving ship trajectory tracking.

[0021] In this process, the neural network constructed in this invention adopts an "encoder-decoder" structure, and the architecture of this network is as follows: Figure 2 As shown, the encoder uses a multi-layer neural network to achieve a nonlinear mapping from the target generalized force to the thruster control commands, outputting control vectors including the thrust magnitude and azimuth angle of each thruster. The decoder reconstructs the generalized force based on the control commands and compares it with the original target, forming the supervision signal and physical consistency constraints during the training phase. The specific implementation steps are as follows: The first step is to use any point on the Earth's surface as the origin of the coordinate system. Establish the northeast coordinate system Then, taking the ship's center of gravity as the origin of the coordinate system... Establish the ship's coordinate system ,like Figure 3 As shown, a kinematic and dynamic model of the ship is established based on the northeast coordinate system and the hull coordinate system:

[0022]

[0023] in, This indicates the ship's actual position and heading angle in the northeast coordinate system; This represents the linear velocity and angular velocity of the ship in the ship's coordinate system. Let be the transformation matrix between the northeast coordinate system and the ship's coordinate system, and satisfy . ; Represents the system's inertia matrix; Represents the ship's hydrodynamic damping coefficient matrix; Indicates the ship's control inputs; This indicates interference from external deep-sea environmental factors such as wind, waves, and currents.

[0024] The second step describes the thrust distribution process of the ship's hybrid propulsion system. Thrust distribution is the core of the ship's motion control system. The core task of the thrust distribution problem is to calculate the thrust magnitude and azimuth angle of each thruster based on the controller's commands, using the force and torque commands from the upper-level controller as control inputs to the module. This generates the longitudinal force, lateral force, and slewing torque commands required by the control system. Therefore, the mathematical description of the thrust distribution problem is:

[0025] in, Indicates the force and torque commanded by the dynamic positioning control system; This indicates the magnitude of thrust generated by the four propellers on the ship; and The thrust generated by the two azimuth thrusters; and These are the thrust generated by the two water jet propulsion units; Indicates the azimuth angle of the four propellers on the ship; and These are the azimuth angles of the two azimuth thrusters; and These are the azimuth angles of the two waterjet propulsion units; The configuration matrix of the thrusters and has .matrix For one A matrix, wherein each column of the matrix has the following specific form:

[0026] in, ; These represent the longitudinal and lateral positions of the i-th propeller relative to the ship's center of gravity, respectively.

[0027] The third step is to construct a supervised dataset for training this deep neural network. First, a large number of target generalized forces are defined. This is used to simulate the control objectives required by a ship under different states. Then, for each target generalized force, an optimization problem with multiple nonlinear constraints is solved to obtain a unique, physically feasible thruster command vector *n*. This command vector includes the thrust magnitude and azimuth angle of each thruster, satisfying the actual physical constraints such as the thruster's operating range and angle limitations. Finally, the target generalized forces and their corresponding optimal thruster commands form training samples, creating a high-quality dataset suitable for supervised learning. The specific physical constraints related to the thrusters are shown below:

[0028] Among them, the thruster command vector ; and These represent the minimum and maximum thrust of the i-th thruster, respectively. and These represent the minimum and maximum values ​​of the azimuth angle of the i-th thruster, respectively. This represents the thrust amplitude of the i-th thruster at the previous moment; and These represent the minimum and maximum values ​​of the thrust change per unit time for the i-th thruster, respectively. This represents the azimuth amplitude of the i-th thruster at the previous moment; and Let represent the minimum and maximum values ​​of the azimuth angle change per unit time for the i-th thruster, respectively. The range of rudder angle change for the two azimuth thrusters is: The range of rudder angle variation for the two waterjet propulsion units is as follows: The range of the rudder angle variation rate for the four thrusters is: .

[0029] The fourth step involves constructing a comprehensive loss function for training the neural network. Multiple loss functions are designed to achieve different control objectives and physical constraints. These loss functions comprehensively consider not only the physical constraints in the thrust allocation process but also the control allocation accuracy and system energy consumption, guiding the neural network to learn the allocation strategy. To achieve accurate thrust allocation by the deep neural network under multiple physical constraints, this invention constructs a comprehensive loss function system composed of multiple loss terms, covering objectives such as physical consistency, task accuracy, energy consumption constraints, and operational feasibility.

[0030] The mathematical expression for the comprehensive loss function is as follows:

[0031] Among them, the generalized force matching loss term The control commands used to constrain the output of the intermediate layers of the network can reproduce the effect consistent with the target generalized force, enhancing the network's physical modeling capabilities; and the loss term for minimizing the assigned error is also included. The primary task loss measures the deviation between the overall network output and the expected generalized force; the thruster control command amplitude limitation loss term... Used to limit the amplitude of thrust and angle commands to their physical boundaries; thruster command change rate limitation loss term Used to control the rate of change of thruster commands over time, avoiding control instability caused by abrupt changes; power loss term The network tends to generate a thrust allocation strategy with lower energy consumption; azimuth penalty term. The system explicitly penalizes thrusters that enter a pre-defined restricted area. This hierarchical, multi-objective loss design ensures that the model not only accurately achieves its mission objectives but also demonstrates good feasibility and engineering constraint satisfaction in practical applications.

[0032] In the formula:

[0033] in, These are the weighting factors for each loss function; For index variables; To calculate the mean squared error function of the data; The system control force and torque are calculated using a formula after being output from the encoder. ; The final output of this neural network is the system control force and torque; The thrust and azimuth amplitude of the thruster at the previous moment; and The boundary value of the azimuth angle restricted area for the first azimuth thruster; and This is the boundary value of the azimuth angle restricted area for the second azimuth thruster.

[0034] The fifth step involves adjusting the weights of the loss function to obtain the adjusted trained neural network model. When in use, real-time control commands are input to the adjusted trained neural network model, which then outputs thruster control commands that satisfy the constraints. In this invention, the comprehensive loss function consists of six loss sub-functions, whose weight factors... to Using manually set fixed weights, the weight coefficients of each loss term These are pre-set parameters. To prevent any one loss term from dominating the training process, reasonable weight ratios are manually set through experimentation, aiming to balance the gradient contributions of each loss term during training.

[0035] This invention addresses the control optimization problem in ship motion control and proposes a control allocation method based on deep neural networks, which has the following advantages compared to existing technologies: (1) Traditional thrust allocation methods based on nonlinear optimization are computationally complex and difficult to meet the real-time control requirements in dynamic environments. This invention transforms thrust allocation from optimization solution to forward inference by training a deep neural network, which greatly reduces the computation time and is suitable for high-frequency control scenarios.

[0036] (2) According to current research, existing optimization algorithms suffer from computational complexity and low allocation accuracy when solving thrust allocation problems with multiple constraints. To address this issue, this invention proposes a novel method for handling multiple physical constraints. During the training phase, this invention embeds multiple physical constraints, such as thrust amplitude, angle range, rate of change, energy consumption, and angle restricted areas, into a multi-objective loss function. This ensures that the thruster commands output by the neural network have good executability and engineering feasibility, avoiding the problems of complex constraint handling or easy failure in traditional methods.

[0037] (3) While ensuring the accuracy of generalized force, by introducing an energy consumption-related loss function, the model is guided to learn an energy-saving thrust allocation strategy, achieving a comprehensive balance of accuracy, stability and energy efficiency, which is superior to the traditional method that only pursues thrust allocation accuracy.

Claims

1. A method for intelligent control optimization of ship hybrid propulsion based on deep neural networks, characterized in that, The method includes the following steps: The motion controller sets the ship's navigation trajectory information based on the control objectives required by the ship under different states, and obtains multiple target generalized forces; Multiple target generalized forces are fed into a deep neural network model with the optimization objectives of minimizing energy consumption and allocation error for control allocation operations. The optimization problem of multiple nonlinear constraints of each target generalized force is solved through the forward propagation process, and the thruster control command that satisfies the constraints is output. The thruster then performs the corresponding action, ultimately achieving the ship's trajectory tracking.

2. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 1, characterized in that, The constraints of the deep neural network model are the thrust and azimuth angle limitations for each thruster.

3. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 1, characterized in that, The physical constraints in the thrust allocation process are considered to guide the neural network to learn the allocation strategy. A comprehensive loss function is constructed, and the model is trained based on the comprehensive loss function to obtain the deep neural network model to be adjusted.

4. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 3, characterized in that, Based on the physical constraints in the thrust allocation process, and taking into account the control of allocation accuracy and energy consumption, the neural network is guided to learn the allocation strategy and a comprehensive loss function is constructed.

5. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 3, characterized in that, The weights of the comprehensive loss function are adjusted through experiments, and then the deep neural network model to be adjusted is obtained. When in use, multiple target generalized forces and real-time control commands are input into the deep neural network model, and the thruster control commands that meet the constraints are output.

6. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 3, characterized in that, The physical constraints in the thrust distribution process are based on the thruster control commands, using the force and torque commands from the upper controller as control inputs, to calculate the thrust magnitude and azimuth angle of each thruster, thereby generating longitudinal force, lateral force, and rotational torque commands.

7. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 3, characterized in that, The comprehensive loss function includes a generalized force matching loss term, an allocation error minimization loss term, a thruster control command amplitude limitation loss term, a thruster command change rate limitation loss term, a power loss term, and an azimuth angle restricted area penalty term.

8. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 1, characterized in that, The deep neural network model adopts an "encoder-decoder" structure. The encoder realizes the nonlinear mapping from the target generalized force to the thruster control command through a multi-layer neural network. The decoder reconstructs the generalized force according to the control command and compares it with the original target, thus forming the supervision signal and physical consistency constraint during the training phase.

9. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 1, characterized in that, Based on the thrust magnitude and azimuth angle of each thruster, and combined with the target generalized force, a supervised dataset for the deep neural network model is constructed using the target generalized force and the corresponding optimal thruster command.

10. The intelligent control optimization method for ship hybrid propulsion based on deep neural networks as described in claim 1, characterized in that, The target generalized force refers to the force and torque commanded by the dynamic positioning control system.

Images