An unmanned ship trajectory tracking controller structure and control method based on meta learning

Abstract

The application provides a kind of unmanned ship trajectory tracking controller structure and control method based on meta learning, it is related to unmanned ship trajectory tracking control technical field, structure includes kinematics controller, dynamics controller, meta learning predictor, learned base function and c environmental condition data sets;The input end of the dynamics controller is connected with the output end of kinematics controller, meta learning predictor respectively, the output end of the dynamics controller is connected with the input end of unmanned ship model and c environmental condition data sets respectively;The input end of the meta learning predictor is connected with the output end of learned base function and unmanned ship model respectively, and the output end of the meta learning predictor is connected with the input end of the dynamics controller;The application accurately depicts complex dynamic characteristics in the process of unmanned ship and environment interaction using meta learning technology, forms effective ship model characteristics and marine environment characteristics knowledge expression, effectively improves the adaptability of unmanned ship in different environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned surface vessel (USV) trajectory tracking and control technology, and more particularly to a USV trajectory tracking controller structure and control method based on meta-learning. Background Technology

[0002] Trajectory tracking control is a fundamental problem in the motion control of unmanned surface vessels (USVs). Its control objective is to track a time-varying reference trajectory and reach a designated mission location within a specified time, subject to strict time constraints. Unlike path tracking control, trajectory tracking control needs to satisfy both time and space constraints simultaneously, and is often used in time-critical scenarios or missions. Due to the complex and variable marine environment in which USVs navigate, and the presence of numerous uncertainties, trajectory tracking control presents a significant challenge. In recent years, researchers have proposed methods including fuzzy control, adaptive control, and active disturbance rejection control to reduce and eliminate the impact of uncertainties on the performance of USV trajectory tracking control.

[0003] Existing methods for unmanned surface vessel (USV) trajectory tracking and control suffer from the following problems: First, existing neural network-based USV trajectory tracking and control methods primarily use shallow neural networks to approximate uncertainties. Due to their limited feature representation capabilities, they struggle to accurately characterize the complex dynamic characteristics of the USV's interaction with its environment. Second, existing deep neural network-based USV trajectory tracking and control methods suffer from high resource overhead during deep network training, making them unsuitable for practical applications in real-time control systems. Third, existing predictor-based USV trajectory tracking and control methods cannot accurately characterize uncertainties caused by different sea states, hindering the accumulation of knowledge across diverse marine environments. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a meta-learning-based unmanned surface vessel (USV) trajectory tracking controller structure and control method. This invention utilizes meta-learning technology to accurately characterize the complex dynamic features of the USV's interaction with the environment, forming an effective knowledge representation of ship model features and marine environment features, thereby effectively improving the adaptability of the USV in different environments.

[0005] The technical means employed in this invention are as follows:

[0006] An unmanned surface vessel trajectory tracking controller structure based on meta-learning includes a kinematic controller, a dynamic controller, a meta-learning predictor, learned basis functions, and c environmental condition datasets;

[0007] The input terminal of the kinematic controller is connected to a given reference input and the output terminal of the unmanned surface vessel model, respectively, and the output terminal of the kinematic controller is connected to the input terminal of the dynamic controller; the kinematic controller is used to calculate the guidance signal.

[0008] The input terminal of the dynamic controller is connected to the kinematic controller, the meta-learning predictor, the learned basis function, and the output terminal of the unmanned surface vessel model, respectively. The output terminal of the dynamic controller is connected to the input terminal of the unmanned surface vessel model and c environmental condition datasets, respectively. The dynamic controller is used to calculate the control input.

[0009] The input of the meta-learning predictor is connected to the learned basis function and the output of the unmanned surface vessel model, respectively, and the output of the meta-learning predictor is connected to the input of the dynamics controller; the meta-learning predictor is used to calculate the estimated value of the modeling error and the estimated value of the weights of the last layer of the deep network.

[0010] The input of the learned basis function is connected to the output of c environmental condition datasets, and the output of the learned basis function is connected to the input of the dynamic controller and the meta-learning predictor, respectively; the learned basis function is used to calculate the learned basis function.

[0011] The input terminals of the c environmental condition datasets are respectively connected to the output terminals of the dynamic controller and the unmanned surface vessel model, and the output terminals of the c environmental condition datasets are connected to the input terminals of the learned basis functions; the c environmental condition datasets are used to calculate the training dataset.

[0012] Furthermore, the formula for the unmanned surface vessel model is:

[0013]

[0014] Where η = [x, y, ψ] T ∈R 3 For the unmanned surface vessel's position and heading information, [·] T This represents the transpose of a vector; ν = [u, v, r] T ∈R 3 The velocity information of the unmanned surface vessel is given by Y(ν,τ); Y(ν,τ) is a basis function constructed from a deep network. Here, n represents the weights of the last layer of the deep network, and n is the number of neurons in the last layer; M ∈ R 3×3 Let M be the inertial matrix of the unmanned surface vessel. * ∈R 3×3 The calibration value of M; τ = [τ u ,τ v ,τ r ] T ∈R 3 The input to the unmanned surface vessel is ε; ε∈R is the modeling error; R(ψ)∈R 3×3 It is a rotation matrix.

[0015] Furthermore, the input signal of the kinematic controller is a given reference trajectory signal η. rThe kinematic controller outputs a guidance signal v, along with the actual position information η of the unmanned surface vessel. g ;

[0016] The formula for the kinematic controller is:

[0017]

[0018] Among them, e η K represents the kinematic tracking error. 1c ∈R 3×3 It is a diagonal gain matrix; κ1∈R + This is the design value.

[0019] Furthermore, the input signal of the dynamic controller is the guidance signal v. g The unmanned surface vessel's actual speed information ν, the learned basis function Y(ν,τ), and the estimated modeling error. And the estimated values ​​of the weights of the last layer of the deep network The output signal of the dynamic controller is the control input τ;

[0020] The formula for the dynamic controller is:

[0021]

[0022] Among them, e v This is for dynamic tracking error; It is a diagonal gain matrix; κ2∈R + This is the design value.

[0023] Furthermore, the input signal of the c environmental condition datasets is the unmanned surface vessel speed information ν collected under c different environmental conditions. c and control input τ c The output signals of the c environmental condition datasets are the training dataset D. c ;

[0024] Drive the unmanned surface vessel to perform random actions and collect training input [ν] c ,τ c ] Calculate the corresponding tag value Define the training dataset D c for:

[0025]

[0026] Furthermore, the input signal to the learned basis function is the training dataset D. c The output signal is the learned basis function Y(ν,τ);

[0027] A meta-learning method is used to train a deep network. The two-layer optimization problem is defined as follows:

[0028]

[0029] The internal optimization problem is solved using the least squares method; the external optimization problem is solved using the stochastic gradient descent method.

[0030] The internal optimization problem is:

[0031]

[0032] The external optimization problem is:

[0033]

[0034] Where, θ c This represents the optimal weight in the c-th environment.

[0035] Furthermore, the input signal of the meta-learning predictor is the actual speed information ν of the unmanned surface vessel and the learned basis function Y(ν,τ), and the output signal of the meta-learning predictor is the estimated value of the modeling error. And the estimated values ​​of the weights of the last layer of the deep network

[0036] The formula for the meta-learning predictor is as follows:

[0037]

[0038] in, This is an estimate of the actual speed ν of the unmanned surface vessel; For unmanned surface vessel speed estimation bias; λ1∈R, Γ ε ∈R、Γ θ ∈R represents the design parameters; Proj(·) represents the projection operator.

[0039] This invention also provides a control method for an unmanned surface vessel (USV) trajectory tracking controller based on meta-learning, implemented based on the structure of the USV trajectory tracking controller based on meta-learning as described in any one of claims 1-7, characterized by comprising the following steps:

[0040] Collect c control inputs τ of the unmanned surface vessel under different environmental conditions c and speed information ν c Construct training dataset D c ;

[0041] Using training dataset D c The deep network is trained to obtain the learned basis function Y(ν,τ);

[0042] The meta-learning predictor calculates an estimate of the modeling error based on the learned basis function Y(ν,τ) and the velocity information ν. And the estimated values ​​of the weights of the last layer of the deep network Output to the dynamics controller;

[0043] Given a reference trajectory signal η r The kinematic controller is based on the reference trajectory signal η r The deviation from the actual position information η of the unmanned surface vessel is used to calculate the guidance signal v. g ;

[0044] The dynamic controller receives the guidance signal v from the kinematic controller. g The learning basis function Y(ν,τ), the actual velocity information ν of the unmanned surface vessel, and the estimated modeling error output by the meta-learning predictor. And the estimated values ​​of the weights of the last layer of the deep network To calculate the control input τ;

[0045] The control input τ is input to the unmanned surface vessel model to generate the position and velocity information at the next moment, thereby controlling the unmanned surface vessel.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] First, compared with existing unmanned surface vessel trajectory tracking and control methods based on neural networks, this invention adopts meta-learning technology and uses deep neural networks to model uncertainties, overcoming the shortcomings of shallow networks in terms of feature representation ability, and realizing the accurate characterization of complex dynamic characteristics in the interaction process between unmanned surface vessels and the environment.

[0048] Second, compared with existing unmanned surface vessel trajectory tracking and control methods based on deep neural networks, this invention constructs a meta-learning predictor to dynamically update the outermost weight parameters of the deep network, thus solving the problem that deep learning cannot effectively meet the real-time requirements of the control system.

[0049] Third, compared with existing unmanned surface vessel trajectory tracking control methods based on predictors, the meta-learning predictor designed in this invention achieves accurate online estimation of deep network modeling errors, providing the control system with accurate model information, disturbance information, and modeling error information simultaneously. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1This is a schematic diagram of the unmanned surface vessel trajectory tracking controller based on meta-learning of the present invention.

[0052] Figure 2 This is a graph of the loss function of the meta-learning method in the training set.

[0053] Figure 3 This is a graph of the loss function of the meta-learning of this invention on the validation set.

[0054] Figure 4 This is a schematic diagram of the motion trajectory of the unmanned surface vessel based on meta-learning according to the present invention.

[0055] Figure 5 This is a graph showing the longitudinal, lateral, and heading tracking error of the unmanned surface vessel of this invention.

[0056] Figure 6 This is a graph showing the longitudinal velocity, lateral velocity, and bow roll rate of the unmanned surface vessel of this invention.

[0057] Figure 7 This is a diagram showing the effect of observing the deep network modeling error of this invention.

[0058] Figure 8 This is the control input curve diagram of the unmanned surface vessel of the present invention. Detailed Implementation

[0059] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0060] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0061] like Figure 1As shown, this invention provides a meta-learning-based unmanned surface vessel trajectory tracking controller structure, including a kinematic controller, a dynamic controller, a meta-learning predictor, learned basis functions, and c environmental condition datasets;

[0062] The input terminal of the kinematic controller is connected to the given reference input and the output terminal of the unmanned surface vessel model, respectively, and the output terminal is connected to the input terminal of the dynamic controller.

[0063] The input terminal of the dynamic controller is connected to the kinematic controller, the meta-learning predictor, the learned basis function, and the output terminal of the unmanned surface vessel model, respectively. The output terminal is connected to the input terminal of the unmanned surface vessel model and the c environmental condition datasets, respectively.

[0064] The input of the meta-learning predictor is connected to the output of the learned basis function and the unmanned surface vessel model, respectively, and the output is connected to the input of the dynamics controller.

[0065] The input of the learning basis function is connected to the output of the c environmental condition datasets, and the output is connected to the input of the dynamic controller and the meta-learning predictor, respectively.

[0066] The input terminals of the c environmental condition datasets are respectively connected to the output terminals of the dynamic controller and the unmanned surface vessel model, and the output terminals are connected to the input terminals of the learned basis functions.

[0067] The mathematical model of the unmanned surface vessel is as follows:

[0068]

[0069] Where η = [x, y, ψ] T ∈R 3 For the unmanned surface vessel's position and heading information, [·] T This represents the transpose of a vector; ν = [u, v, r] T ∈R 3 The velocity information of the unmanned surface vessel is given by Y(ν,τ); Y(ν,τ) is a basis function constructed from a deep network. Here, n represents the weights of the last layer of the deep network, and n is the number of neurons in the last layer; M ∈ R 3×3 Let M be the inertial matrix of the unmanned surface vessel. * ∈R 3×3 The calibration value of M; τ = [τ u ,τ v ,τ r ] T ∈R 3 The input to the unmanned surface vessel is ε; ε∈R is the modeling error; R(ψ)∈R 3×3 Let be a rotation matrix, represented as:

[0070]

[0071] A. Kinematic controller design

[0072] The input signal to the kinematic controller is the given reference trajectory signal η. r The actual position information η of the unmanned surface vessel is used to output a guidance signal v. g .

[0073] The kinematic controller is designed as follows:

[0074]

[0075] Among them, e η K represents the kinematic tracking error. 1c ∈R 3×3 It is a diagonal gain matrix; κ1∈R + This is the design value.

[0076] B. Dynamics Controller Design

[0077] The input signal to the dynamic controller is the guidance signal v. g The unmanned surface vessel's actual speed information ν, the learned basis function Y(ν,τ), and the estimated modeling error. And the estimated values ​​of the weights of the last layer of the deep network The output signal is the control input τ.

[0078] The formula for the dynamic controller is as follows:

[0079]

[0080] Among them, e v For dynamic tracking error; K 2c ∈R 3×3 It is a diagonal gain matrix; κ2∈R + This is the design value.

[0081] Design of C environmental condition datasets

[0082] The input signals for the c environmental condition datasets are the unmanned surface vessel (USV) velocity information ν and control input τ collected under c different environmental conditions, and the output signal is the training dataset D. c .

[0083] The c environmental condition datasets are designed as follows:

[0084] Drive the unmanned surface vessel to perform random actions and collect training input [ν] c ,τ c ] Calculate the corresponding tag value Define the training dataset D c for

[0085]

[0086] D. Design of learning basis functions

[0087] The input signal to the learned basis function is the training dataset D. c The output signal is the learned basis function Y(ν,τ).

[0088] The basis functions for learning are designed as follows:

[0089] A meta-learning method is used to train a deep network. The two-layer optimization problem is defined as follows:

[0090]

[0091] Internal optimization problem Solve the external optimization problem using the least squares method. Solve using the stochastic gradient descent method, θ c This represents the optimal weight in the c-th environment.

[0092] E. Meta-learning predictor design

[0093] The meta-learning predictor takes the actual velocity information ν of the unmanned surface vessel (USV) and the learned basis function Y(ν,τ) as input signals, and outputs an estimate of the modeling error. And the estimated values ​​of the weights of the last layer of the deep network

[0094] The formula for the meta-learning predictor is as follows:

[0095]

[0096] in, This is an estimate of the actual speed ν of the unmanned surface vessel; For unmanned surface vessel speed estimation bias; λ1∈R, Γ ε ∈R、Γ θ ∈R represents the design parameters; Proj(·) represents the projection operator.

[0097] This invention also provides a control method for an unmanned surface vessel (USV) trajectory tracking controller based on meta-learning. The implementation of the meta-learning-based USV trajectory tracking controller structure includes the following steps:

[0098] First, we collect the unmanned surface vessel control input τ and velocity information ν under c different environmental conditions to construct a training dataset D. c Using this dataset D c The deep network is trained to obtain the learned basis function Y(ν,τ). Given the reference trajectory signal η... r At that time, the kinematic controller uses the reference trajectory signal ηr The deviation from the actual position information η of the unmanned surface vessel is used to calculate the guidance signal v. g The dynamic controller receives the guidance signal v from the kinematic controller. g The learning basis function Y(ν,τ), the actual speed information ν of the unmanned surface vessel, and the estimated value of the modeling error. And the estimated values ​​of the weights of the last layer of the deep network The control input τ is calculated; the control input τ is input to the unmanned surface vessel model to generate the position and velocity information at the next moment, thereby controlling the unmanned surface vessel.

[0099] The control parameters for equations (3), (4), and (7) are selected as follows:

[0100] K 1c = diag{0.6,0.6,0.6}, K 2c =diag{5,5,2}, κ1=1, κ2=1, λ1=10, Γ ε =25,Γ θ =80, where diag represents a diagonal matrix.

[0101] Simulation results are as follows Figure 2-8 As shown. Figure 2 A graph of the loss function of the meta-learning method of this invention in the training set is given. Figure 3 The loss function graphs of the present invention's meta-learning are presented in the validation set under three wind disturbance conditions: level 0, level 3, and level 6. Figure 2 , 3 It can be seen that both loss values ​​tend to stabilize after 500 training rounds. Figure 4 A schematic diagram of the motion trajectory of the unmanned surface vessel based on meta-learning of this invention is given, which shows that the actual trajectory of the unmanned surface vessel tracks the reference trajectory very well. Figure 5 The longitudinal, lateral, and heading tracking error curves of the unmanned surface vessel of the present invention are given. It can be seen that the tracking errors in the three directions eventually converge to near zero after a transient process of about 30 seconds. Figure 6 The graphs show the longitudinal velocity, lateral velocity, and bow roll rate of the unmanned surface vessel of this invention. It can be seen that both the actual velocity information and the estimated velocity information achieve good tracking of the guidance signal. Figure 7 The estimation performance of the deep network modeling error of the present invention is presented. It can be seen that the designed meta-learning predictor can always accurately estimate the modeling error, regardless of whether it is a transient or steady-state process. Figure 8 The thrust in the longitudinal and lateral directions and the torque in the yaw direction of the unmanned surface vessel of the present invention are given. It can be seen that the control input in the three directions tends to stabilize after the transient process.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A structure for an unmanned surface vessel trajectory tracking controller based on meta-learning, characterized in that: This includes kinematic controllers, dynamic controllers, meta-learning predictors, and learned basis functions. A dataset of environmental conditions; The input terminal of the kinematic controller is connected to a given reference input and the output terminal of the unmanned surface vessel model, respectively, and the output terminal of the kinematic controller is connected to the input terminal of the dynamic controller; the kinematic controller is used to calculate the guidance signal. The input terminals of the dynamics controller are connected to the kinematics controller, the meta-learning predictor, the learned basis functions, and the output terminals of the unmanned surface vessel (USV) model, respectively. The output terminals of the dynamics controller are connected to the USV model and... The input terminals of several environmental condition datasets are connected; the dynamic controller is used to calculate the control input; The input of the meta-learning predictor is connected to the learned basis function and the output of the unmanned surface vessel model, respectively, and the output of the meta-learning predictor is connected to the input of the dynamics controller; the meta-learning predictor is used to calculate the estimated value of the modeling error and the estimated value of the weights of the last layer of the deep network. The input of the learned basis function is... The outputs of the environmental condition datasets are connected, and the outputs of the learned basis functions are connected to the inputs of the dynamic controller and the meta-learning predictor, respectively; the learned basis functions are used to calculate the learned basis functions. The A dataset of environmental conditions is used to compute the training dataset; The input signal to the meta-learning predictor is the actual speed information of the unmanned surface vessel. and the learning basis functions The output signal of the meta-learning predictor is an estimate of the modeling error. And the estimated values ​​of the weights of the last layer of the deep network ; The formula for the meta-learning predictor is as follows: in, For the actual speed information of the unmanned surface vessel The estimated value; The speed estimation error of the unmanned surface vessel; , , For design parameters; This is the projection operator.

2. The unmanned surface vessel trajectory tracking controller structure based on meta-learning according to claim 1, characterized in that, The formula for the unmanned surface vessel model is: in, For the position and heading information of the unmanned surface vessel, Represents the transpose of a vector; For the speed information of the unmanned surface vessel; These are the basis functions constructed from deep networks; For the weights of the last layer of the deep network, This refers to the number of neurons in the last layer. For the inertial matrix of the unmanned surface vessel, for The calibration value; For the control input of the unmanned surface vessel; This represents modeling error; It is a rotation matrix.

3. The unmanned surface vessel trajectory tracking controller structure based on meta-learning according to claim 1, characterized in that, The input signal of the kinematic controller is a given reference trajectory signal. and the actual location information of unmanned surface vessels The output signal of the kinematic controller is a guidance signal. ; The formula for the kinematic controller is: in, For kinematic tracking error; It is a diagonal gain matrix; This is the design value.

4. The unmanned surface vessel trajectory tracking controller structure based on meta-learning according to claim 1, characterized in that, The input signal of the dynamic controller is a guidance signal. Actual speed information of unmanned surface vessels Basis functions for learning Estimated value of modeling error And the estimated values ​​of the weights of the last layer of the deep network The output signal of the dynamic controller is the control input. ; The formula for the dynamic controller is: in, This is for dynamic tracking error; It is a diagonal gain matrix; This is the design value.

5. The unmanned surface vessel trajectory tracking controller structure based on meta-learning according to claim 1, characterized in that, The The input signal for each environmental condition dataset is... Unmanned surface vessel speed information collected under different environments and control input The The output signal of each environmental condition dataset is the training dataset. ; Drive the unmanned surface vessel to perform random actions and collect training input Calculate the corresponding tag value Define the training dataset for: 。 6. The unmanned surface vessel trajectory tracking controller structure based on meta-learning according to claim 1, characterized in that, The input signal to the learned basis function is the training dataset. The output signal is the learned basis function. ; Deep networks are trained using meta-learning methods, and the two-layer optimization problem is defined as follows: The internal optimization problem is solved using the least squares method; the external optimization problem is solved using the stochastic gradient descent method. The internal optimization problem is: The external optimization problem is: in, For the first The optimal weights under each environment.

7. A control method for an unmanned surface vessel (USV) trajectory tracking controller based on meta-learning, implemented based on the USV trajectory tracking controller structure based on meta-learning as described in any one of claims 1-6, characterized in that, Includes the following steps: collect Unmanned surface vessel control inputs under different environmental conditions and speed information Building a training dataset ; Using the training dataset Training a deep network yields the learned basis functions. ; Meta-learning predictors are based on learned basis functions. Speed ​​information Calculate the estimated value of the modeling error And the estimated values ​​of the weights of the last layer of the deep network Output to the dynamics controller; Given a reference trajectory signal The kinematic controller is based on the reference trajectory signal With the actual location information of the unmanned surface vessel Deviation calculation guidance signal ; The dynamic controller receives guidance signals from the kinematic controller. Basis functions for learning Actual speed information of unmanned surface vessels The estimated modeling error output by the meta-learning predictor And the estimated values ​​of the weights of the last layer of the deep network To calculate control input ; Control input The data is input into the unmanned surface vessel (USV) model to generate the position and velocity information for the next moment, thereby enabling control of the USV.

Images