A data-driven recursive neural network method for ship motion models

Abstract

The present invention discloses a data-driven recursive neural network ship motion model method, which includes: obtaining the low-frequency motion ship position, heading and propeller operation information of the ship through ship navigation tests; constructing the ship motion recursive neural network model architecture, and constructing a recursive neural network training data set and a test data set based on the obtained ship low-frequency motion information and propulsion system operation information; based on the training data set, constructing a recursive neural network ship motion model using different activation functions and optimization algorithms and performing iterative training; based on the test data set, testing and optimizing the established ship motion neural network model to determine the final ship motion recursive neural network model. Based on the data-driven principle, the present invention establishes a data-driven recursive neural network ship motion model only based on the measured ship position and heading, and the current ship propulsion system information, realizes the direct prediction of future ship position and heading, without the need to identify system parameters, and can realize the data-driven control of intelligent ships, thereby constructing a ship intelligent control system.

Description

Technical Field

[0001] The present invention relates to the field of constructing a ship motion neural network model, and particularly to a method for a ship motion model based on a data-driven recurrent neural network. Background Art

[0002] With the rapid development of ocean exploitation, ship automatic control has received increasing attention, such as dynamic positioning for offshore operations and unmanned control of surface ships. Generally, these operations require very high positioning and course control accuracy. Therefore, an accurate ship motion model needs to be established to obtain better ship control performance.

[0003] The earliest ship motion mathematical models are mechanism models based on hydrodynamics and Newton's laws of motion, including the Abkovitz model and the MMG model. For the above mechanism models, their accuracy depends to a large extent on the hydrodynamic coefficients in the models. There are several methods for determining these hydrodynamic parameters, including empirical formulas, constrained model tests, and system identification, etc. The results of hydrodynamic coefficients obtained from constrained model tests are relatively reliable, but the process is difficult, costly, and time-consuming. Although it is convenient to obtain ship hydrodynamic coefficients using empirical formulas, it is not applicable to all types of ships and is not accurate. Thus, methods for identifying ship hydrodynamic coefficients based on system input and output are proposed, such as extended Kalman filtering, recursive least squares, support vector machines, etc.

[0004] In recent years, in order to bypass the complex process of solving ship hydrodynamic parameters, model-free data-driven models based on system input and output have been widely studied. The data-driven ship motion model is a black-box model, which can directly predict the future ship motion state based on the input and output data of the ship without identifying the hydrodynamic parameters of the ship. Since neural networks have strong generalization ability and can simulate complex nonlinear systems, it has strong superiority in establishing a data-driven ship motion model.

[0005] For surface moving ships, ship position information is often obtained through a satellite navigation system and a position reference system, and ship course information is obtained through a compass system. Therefore, the ship position and speed in the earth-fixed coordinate system can be measured, while the longitudinal / lateral speed and yaw angular velocity in the body-fixed coordinate system cannot be directly measured. However, in the currently established ship motion neural network models, in addition to using ship position and course as neural network inputs, longitudinal, lateral speeds, and yaw angular velocity are often used as inputs, which is unreasonable.

[0006] Therefore, based on the data-driven principle, a ship motion model based on a neural network can be established only based on the measured ship position and speed, as well as the current ship propulsion system information, to directly predict the future ship position and heading without the need to identify system parameters. This will facilitate the data-driven control of intelligent ships and thus construct a ship intelligent control system. Summary of the Invention

[0007] (I) Technical Problem to be Solved

[0008] To solve the above problems of the prior art, considering that the longitudinal / lateral speed and yaw angular velocity of a ship cannot be directly measured, and a data-driven ship motion model is established only based on the input of the system propulsion system and the measured ship position and heading output data, the present invention proposes a method for a ship motion model based on a data-driven recurrent neural network.

[0009] (II) Technical Solution

[0010] To achieve the above object, the main technical solutions adopted by the present invention include:

[0011] 101. Conduct a ship navigation test to make the ship travel in a certain maneuvering state, obtain the ship motion state information through a satellite navigation system, a position reference system, and a compass system, and record the real-time operation information of the propulsion system propeller during ship navigation;

[0012] 102. Construct the recurrent neural network model structure of ship motion, determine the neuron nodes of the input layer, hidden layer, and output layer of the neural network, and determine the hyperparameters of the neural network;

[0013] 103. Based on the obtained low-frequency ship motion information and propulsion system operation information, construct a recurrent neural network training data set and a neural network test data set;

[0014] 104. Based on the recurrent neural network training data set, with the goal of minimizing the loss function, iteratively train the ship motion neural network model to obtain the weight coefficients in the neuron nodes of each layer of the neural network;

[0015] 105. Based on the neural network test data set, test and optimize the established ship motion neural network model to determine the final ship motion recurrent neural network model;

[0016] The ship motion state and propulsion system propeller operation information obtained from the ship navigation test include: obtaining the ship position and heading information through the sensor system, filtering out the high-frequency wave motion components and measurement noise in the measured ship position and heading information through the filtering system to obtain the low-frequency ship motion state information at each sampling moment, denoted as η t =[x t ,yt , ψ t T , where: η represents the low-frequency motion vector of the ship, and x, y, ψ are the north coordinate, east coordinate, and heading angle of the ship in the earth coordinate system respectively. Obtain the operation information of the ship propulsion system, including the rotational speed of each propeller and the rudder angle, etc., denoted as u t = [n 1t , n 2t ,..., α 1t , α 2t ,..., δ 1t , δ 2t ,...] T , where u represents the operation state vector of the thruster, n i , α i , δ i , i = 1, 2,... are the rotational speed of the propeller, azimuth angle, and rudder angle of the rudder respectively; the subscript t represents the current sampling moment.

[0017] The recursive neural network model structure of the ship motion includes determining the nodes of its input layer, hidden layer, and output layer. The input data set is labeled as {x0, x1,..., x t , x t+1 ,...}, the output set of the hidden layer is denoted as {φ0, φ1,..., φ t , φ t+1 ,...}, and the output data set is denoted as {y0, y1,..., y t , y t+1 ,...};

[0018] For the recursive neural network model of the ship motion, the neuron signal of the input layer introduces the delay system information to improve the expression ability of the neural network model for the dynamic changes of the time-varying system. Taking the ship thrust at the current moment and the historical ship position and heading as the input of the ship motion neural network model, the input layer is described as: X t = [τ t , η t , η t-1 , η t-2 T , where τ t = [τ x,t , τ y,t , τ N,t T represents the propulsion force vector corresponding to the sampling moment t, where τ x,t represents the longitudinal thrust at the moment t, τ y,t represents the lateral thrust at the moment t, and τ N,t represents the turning moment at the moment t; ​​​They are the low-frequency motion information of the ship after filtering at the current sampling moment t, the moment t-1 (i.e., the previous sampling moment), and the moment t-2 (i.e., the sampling moment before the previous one), respectively.

[0019] For the recurrent neural network model of the ship motion, the expression of the hidden layer is: where N is the number of neurons in the input layer, φ is the activation function, w ij , b ij are the weight and bias term passed from the input layer X i to the neuron j in the hidden layer, respectively.

[0020] The recurrent neural network model of the ship motion predicts the future ship motion state, i.e., the ship position and heading at the next moment, based on the historical state of the ship motion and the propeller thrust information. Therefore, the output layer is expressed as: That is, the low-frequency motion state of the ship at the next sampling moment t+1 starting from the current moment t.

[0021] For the recurrent neural network model of the ship motion, the appropriate number of hidden layers, the number of neurons in the hidden layer, the appropriate activation function, and the optimization algorithm are selected through the cross-validation method to minimize the loss function.

[0022] For the recurrent neural network model of the ship motion, based on the obtained operating state information of the propulsion system: u t =[n 1t , n 2t ,..., α 1t , α 2t ,..., δ 1t , δ 2t ,...] T , according to the dynamic characteristics of the propeller, the equivalent thrust output by each propeller is solved, and then the longitudinal thrust, lateral thrust, and turning moment of the total output of the propulsion system are comprehensively obtained, that is, the propulsion force vector τ t =[τ x,t , τ y,t , τ N,t T , and then combined with the low-frequency motion state information of the ship after filtering: A recurrent neural network data set is constructed, including the input data set: {x0, x1,..., x t , x t+1 ,...}, where X t =[τ t , η t , η t-1 , η t-2 T ; and the output data set is {y0, y1,..., y t , y​​t+1 ,...}, where The established data set is divided into two parts, one part is the training data set and the other part is the test data set;

[0023] According to the input data set and the output data set, establish the mapping relationship of the ship motion neural network model: y t+1 = F(X t ), where F(·) is an implicit expression function;

[0024] Based on the established data set of the recurrent neural network ship motion model, use the selected activation function and optimization algorithm for iterative training until the training requirements are met. Further, based on the obtained recurrent neural network ship motion model through training, use the test data set to conduct model testing, verify the generalization performance of the established neural network model, and conduct model optimization to determine the final recurrent neural network ship motion model.

[0025] (III) Beneficial effects of the present invention:

[0026] The proposed data-driven recurrent neural network ship motion model only uses the input / measurement output data of the system, that is, takes the historical position and heading measurement values of the ship and the propeller thrust as the input layer to predict the ship's position and heading at future moments. The established recurrent neural network ship motion model is a non-parametric model. Different from the traditional model parameter identification method, the model established by this method does not need to identify the hydrodynamic parameters of the ship. Therefore, there is no model mismatch problem caused by ignoring highly nonlinear components, which not only reduces the modeling workload but also ensures the accuracy of model fitting. At the same time, it will be beneficial to realize the data-driven control of intelligent ships, thereby constructing a ship intelligent control system. Description of the Drawings

[0027] Figure 1 is the schematic diagram of the recurrent neural network for ship motion;

[0028] Figure 2 is the schematic diagram of the training of the recurrent neural network model for ship motion;

[0029] Figure 3 is the result of training the RNN using the Relu activation function;

[0030] Figure 4 is the result of training the RNN using the LeakyRelu activation function;

[0031] Figure 5 is the result of training the RNN using the Softplus activation function;

[0032] Figure 6 is the test result based on Relu;

[0033] Figure 7 Test results based on Softplus;

[0034] Figure 8 Test results based on LeakyRelu;

[0035] Figure 9 MSE value of the RNN based on leakyrelu;

[0036] Figure 10 Test results based on Adam; Detailed implementation manners

[0037] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0038] All the technical and scientific terms used in the embodiments of the present invention have the same meanings as those commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the embodiments of the present invention in the specification of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The term "and / or" used in the embodiments of the present invention includes any and all combinations of one or more of the related listed items.

[0039] Based on the data-driven ship motion modeling, considering that establishing an accurate data-driven ship motion model is the key to realizing data-driven control, the present invention proposes a data-driven recursive neural network ship motion modeling method for predicting the future position and heading of a ship. Different from the traditional model identification method that requires identifying hydrodynamic parameters, the recursive neural network model only uses the system input / output data, that is, the historical ship position and heading measurement values and the propeller thrust information as the input layer.

[0040] The implementation steps of the present invention are as follows:

[0041] 101. Conduct a ship navigation test to make the ship travel in a certain maneuvering state, obtain the ship motion state information through the satellite navigation system, position reference system and compass system, and record the real-time operation information of the propeller of the propulsion system during ship navigation;

[0042] 102. Build the recursive neural network model structure of ship motion, determine the neuron nodes of the input layer, hidden layer and output layer of the neural network, and determine the hyperparameters of the neural network;

[0043] 103. Based on the obtained low-frequency ship motion information and the operation information of the propulsion system, build a recursive neural network training data set and a neural network test data set;

[0044] 104. Based on the training dataset of the recurrent neural network, with the goal of minimizing the loss function, iteratively train the ship motion neural network model to obtain the weight coefficients in the neuron nodes of each layer of the neural network.

[0045] 105. Based on the neural network test dataset, test and optimize the established ship motion neural network model to determine the final ship motion recurrent neural network model.

[0046] Taking the CSII ship as the implementation object, this ship is a 1:70 ship model of a certain supply ship. This ship is equipped with a bow thruster, two main propellers and two rudders at the tail. The ship performance parameters are shown in Table 1.

[0047] Table 1 CSII Ship Parameters

[0048]

[0049] Establish a CSII ship maneuvering motion model, conduct a navigation test, obtain the ship position and heading information at each sampling moment, and filter out the high-frequency wave motion components and measurement noise in the measured ship position and heading information through a filtering system to obtain the low-frequency motion state information of the ship at each sampling moment, denoted as η t =[x t ,y t ,ψ t T , where: η represents the low-frequency motion vector of the ship, and x, y, ψ are the north coordinate, east coordinate and heading angle of the ship in the earth coordinate system respectively.

[0050] Obtain the operation information of the ship propulsion system. Since the direction angles of the bow thruster and the tail propellers are constant, only record the rotational speeds of each propeller and the rudder angle information, denoted as u t =[n 1t ,n 2t ,n 3t ,δ 1t ,δ 2t T . Where u represents the propulsion system operation state vector, and n i , α i , δ i ,i = 1,2,... are the rotational speed, azimuth angle and rudder angle of the propeller respectively. The subscript t in the above parameters represents the current sampling moment.

[0051] Introduce the delayed system information into the neuron of the input layer of the ship motion recurrent neural network model to improve the expression ability of the neural network model for the dynamic changes of the time-varying system. Taking the current ship thrust and the historical ship position and heading as the input of the ship motion neural network model, the input layer is described as: X t =[τ t ,η​​t , η t-1 , η t-2 T , where τ t = [τ x,t , τ y,t , τ N,t T represents the propulsive force vector corresponding to the sampling time t, where τ x,t represents the longitudinal thrust at time t, τ y,t represents the lateral thrust at time t, τ N,t represents the yaw moment at time t; are respectively the low-frequency motion information of the ship after filtering at the current sampling time t, the previous sampling time t - 1, and the sampling time t - 2 before the previous one.

[0052] The expression of the hidden layer of the recurrent neural network ship motion model is: where N is the number of neurons in the input layer, φ is the activation function, w ij , b ij are respectively the weight and bias term passed from the input layer X i to the hidden layer neuron j.

[0053] Based on the historical state of the ship motion and the propeller thrust information, the future ship motion state, that is, the ship position and heading at the next moment, is predicted. Therefore, the output layer is expressed as: That is, the low-frequency motion state of the ship at the next sampling time t + 1 starting from the current time t.

[0054] The established recurrent neural network ship motion model is as Figure 1 shown.

[0055] Based on obtaining the operating state information of the propulsion system: u t = [n 1t , n 2t , n 3t , δ 1t , δ 2t T , according to the dynamic characteristics of the propeller, solve the equivalent thrust output by each propeller, and then comprehensively obtain the longitudinal thrust, lateral thrust, and yaw moment of the total output of the propulsion system, that is, obtain the propulsive force vector τ t = [τ x,t , τ y,t , τ N,t T , and then combine with the low-frequency motion state information of the ship after filtering: Construct a recurrent neural network data set, including the input data set: {x0, x1,..., x t , x​​​​t+1 ,...}, where X t = [τ t , η t , η t-1 , η t-2 T ; and the output data set {y0, y1,..., y t , y t+1 ,...}, where The established data set is divided into two parts, one part is the training data set, and the other part is the test data set. Taking 0.1 s as the time step, 5000 groups of input and output data are obtained, of which 500 groups are used as the training set of the recurrent neural network model, and the remaining data are used as the test set to test the generalization degree of the recurrent neural network model.

[0056] According to the input data set and the output data set, establish the mapping relationship of the ship motion neural network model: y t+1 = F(X t ), where F(·) is an implicit expression function.

[0057] For the recurrent neural network ship motion model, the number of hidden layers is selected as one layer and the number of hidden layer neuron nodes is 9 by the cross-test method. In order to compare the effects of different activation functions and optimization methods on the recurrent neural network model, based on the established training data set, the Relu, LeakyRelu and Softplus activation functions, as well as the gradient descent, Adam and momentum optimization methods are respectively used for simulation. The training principle of the recurrent neural network ship motion model is as Figure 2 shown.

[0058] Figures 3 to 5 They are respectively the results of training the recurrent neural network model using the Relu, LeakyRelu and Softplus activation functions. In addition, each figure contains three subgraphs, which respectively represent the output values x, y and ψ from top to bottom. The four lines in each subgraph respectively represent the expected values and output values trained by the three optimization methods of Gradient descent, Adam and Momentum.

[0059] Comparative analysis results: In Figure 3 , the output values obtained by Gradient descent and Momentum have less loss compared with the expected values, while the fluctuation between the output value obtained by Adam and the reference value is relatively small. In Figure 4 ​Among them, the output values of the network optimized by Gradient Descent and Momentum fluctuate greatly compared with the expected values, especially the fitting error of ψ. The fitting values obtained by Adam are relatively ideal. Generally speaking, under the same optimization method, choosing Softplus as the activation function has a better effect than Relu. In Figure 5 Among them, under the action of LeakyRelu, the error of the output value obtained by Momentum is significantly larger. Especially, the fluctuation of the output value of y is the most obvious compared with the reference value, while the results obtained by the gradient descent method and the Adam method have smaller fluctuations and are relatively ideal.

[0060] Furthermore, the reliability and generalization performance of the training effect of the recurrent neural network are tested through the test set. The test set consists of 4500 groups of data, which belong to the same distribution range as the data in the training set but are not related to each other. Figures 6 - 8 (a) shows the test results using Relu, LeakyRelu, and Softplus activation functions respectively. At the same time, for the convenience of comparison, in Figures 6 - 8 (b), the error between the predicted value and the reference value is shown. Table 2 gives the predicted average error rates of the recurrent neural network based on different activation functions and optimization methods.

[0061] Table 2 Average Errors Based on Different Activation Functions and Optimization Methods

[0062]

[0063] In Figure 6 Among them, based on using Relu as the activation function, the x and y errors between the output value of the Gradient Descent optimization mode and the reference value are relatively large, while the values obtained by Adam and Momentum are relatively close to the reference value. In addition, the fitting effect of Adam on ψ is very ideal, but the results obtained by Gradient Descent and Momentum are exactly the opposite. In Figure 7 Among them, the fitting error of the Momentum optimization method on x is very obvious. On the contrary, the Gradient Descent and Adam optimization methods can better fit x. For y and ψ, the values of Gradient Descent and Momentum are very different from the ideal values. Although the result of the Adam optimization method also has a certain error, it is smaller than the previous two. In Figure 8 Among them, the fitting effects of the three optimization methods on x are relatively ideal. For the output of y, the result of the momentum method is unsatisfactory, which is significantly different from the reference value. Through comprehensive comparison, the ψ values obtained by Gradient Descent and Momentum deviate significantly from the ideal value, while the performance of Adam is relatively perfect.

[0064] Figure 9 The root mean square error results of the recurrent neural network based on leakyrelu are given. From Figures 4 - 10 the comparison, it is not difficult to see that when combining different activation functions, the comprehensive fitting effect of Adam is better than that of Gradient descent and Momentum. In addition, it can also be seen from Table 2 that the error rate of the results obtained by Adam is relatively small. In addition, Figure 9 it shows that Adam has a faster convergence speed. Therefore, in the following simulations, the Adam method is selected to verify the performance when using different activation functions.

[0065] Figure 10 are the test results of applying Adam to three different activation functions. It can be seen that there is no difference in the fitting effect of the three activation functions on x, which is close to the reference value. For the fitting of y, the LeakyRelu activation function obtains better results. For ψ, obviously, there is a large gap between the value obtained by Softplus and the reference value. Although the value obtained by Relu has little difference from the reference value, it is not very ideal compared with LeakyRelu.

[0066] Through the verification of the test set, LeakyRelu is finally selected as the activation function, and Adam is selected as the optimization algorithm.

Claims

1. A method for modeling a ship motion model based on a data-driven recurrent neural network, characterized in that Including the following steps:

101. Conduct a ship navigation test to make the ship travel in a certain maneuvering state. Obtain the ship's motion state information through the satellite navigation system, position reference system, and compass system, and record the real-time operation information of the propellers of the propulsion system during ship navigation, including: obtain the ship's position and heading information through the sensor system, filter out the high-frequency wave motion components and measurement noise in the measured ship position and heading information through the filtering system, and obtain the low-frequency motion state information of the ship at each sampling moment, denoted as η t =[x t ,y t ,ψ t T , where: η represents the low-frequency motion vector of the ship, and x, y, and ψ are the northward coordinate, eastward coordinate, and heading angle of the ship in the earth coordinate system respectively; obtain the operation information of the ship's propulsion system, including the rotational speed of each propeller and the rudder angle information, denoted as u t =[n 1t ,n 2t ,...,α 1t ,α 2t ,...,δ 1t ,δ 2t T , where u represents the operation state vector of the propeller, n i , α i , δ i , i = 1, 2,... are the rotational speed of the propeller, azimuth angle, and rudder angle of the steering gear respectively; the subscript t represents the current sampling moment;​​ 102. Construct the recursive neural network model structure of ship motion, determine the neuron nodes of the input layer and output layer of the neural network, and determine the number of hidden layers and the number of neuron nodes in the hidden layer by the cross-test method; 103. Based on the obtained low-frequency motion information of the ship and the operation information of the propulsion system, construct a recursive neural network training dataset and a neural network test dataset; the input dataset is marked as {x0, x1,..., x t , x t+1 ,...}, the output set of the hidden layer is denoted as {φ0, φ1,..., φ t , φ t+1 ,...}, and the output dataset is denoted as {y0, y1,..., y t , y t+1 ,...}; establish the mapping relationship of the ship motion neural network model: y t+1 = F(X t ), where F(·) is an implicit expression function; 104. Based on the recursive neural network training data set, construct a recursive neural network ship motion model by using different activation functions and optimization algorithms. With the goal of minimizing the loss function, select appropriate activation functions and optimization algorithms to iteratively train the ship motion neural network model, and obtain the weight coefficients in the neuron nodes of each layer of the neural network; 105. Based on the neural network test data set, conduct model testing on the established ship motion neural network model, verify the generalization performance of the established neural network model, and perform model optimization to determine the final recursive neural network model of ship motion.

2. The method for modeling a recursive neural network ship motion model based on data driving according to claim 1, wherein: For the recursive neural network model of ship motion, the neuron signals of the input layer introduce the information of the delay system to improve the expression ability of the neural network model for the dynamic changes of the time-varying system. The current ship thrust, historical ship position and heading are used as the inputs of the ship motion neural network model. The input layer is described as: X t =[τ t ,η t ,η t-1 ,η t-2 T , where τ t =[τ x,t ,τ y,t ,τ N,t T represents the propulsive force vector corresponding to the sampling time t, where τ x,t represents the longitudinal thrust at time t, τ y,t represents the lateral thrust at time t, and τ N,t represents the yaw moment at time t; are the low-frequency motion information of the ship processed by filtering at the current sampling time t, the previous sampling time t - 1, and the sampling time t - 2, respectively;​​ The expression of the hidden layer is: where N is the number of neurons in the input layer, φ is the activation function, and w ij , b ij are the weight and bias term respectively passed from the input layer X i to the j-th neuron in the hidden layer; Based on the historical state of the ship's motion and the propeller thrust information, the future ship motion state, that is, the ship position and heading at the next moment, is predicted. Therefore, the output layer is expressed as: That is, the low-frequency motion state of the ship at the next sampling moment t + 1 starting from the current moment t; Based on the obtained operating state information of the propulsion system: u t = [n 1t , n 2t ,..., α 1t , α 2t ,..., δ 1t , δ 2t ,...] T , according to the dynamic characteristics of the thrusters, solve the equivalent thrust output by each thruster, and then comprehensively obtain the longitudinal thrust, lateral thrust and yaw moment of the total output of the propulsion system, that is, obtain the propulsion force vector τ corresponding to each moment t t = [τ x,t , τ y,t , τ N,t T , and then combine with the filtered low-frequency motion state information of the ship: Construct a recurrent neural network data set, including the input data set: {x0, x1,..., x t , x t+1 ,...}, where X t = [τ t , η t , η t-1 , η t-2 T ;​​ and the output data set is {y0, y1,..., y t , y t+1 ,...}, where the established data set is divided into two parts, one part is the training data set and the other part is the test data set.

Images