Unmanned ship trajectory tracking controller and control method based on online deep learning
Patent Information
- Application Number
- CN202311083403.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-25
- Publication Date
- 2026-09-15
- Estimated Expiration
- 2043-08-25
AI Technical Summary
[0005]有鉴于此，本发明的目的在于提出一种基于在线深度学习的无人船轨迹跟踪控制器及控制方法，以解决现有无人船运动控制方法缺乏对控制输入的快速主动响应的技术问题
[0033] First, compared with existing control methods based on model uncertainty and external disturbances in DNN modeling systems, the online deep learning control method designed in this invention uses ESO to estimate the approximation error of DNN, ensuring the accuracy of network input information during DNN learning and enabling DNN to quickly capture model features.
Smart Images
Figure CN117031957B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned vessel control technology, and more particularly to an unmanned vessel trajectory tracking controller and control method based on online deep learning. Background Technology
[0002] Unmanned surface vessels (USVs), as a type of autonomous maritime vehicle, have broad application prospects in marine exploration, development, and protection due to their small size, flexible operation, and strong scalability. In recent years, the motion control problem of USVs has received widespread attention from scholars both domestically and internationally. Trajectory tracking, as a fundamental scenario in USV motion control, is frequently used for time-sensitive tasks. The control performance of USV trajectory tracking depends on the accuracy of its mathematical model. To capture the dynamic model of USVs in complex marine environments, numerous neural network control methods have been proposed, including feedforward neural networks, radial basis function neural networks, single-hidden-layer neural networks, and echo-state neural networks. However, these neural network methods are all shallow networks, with limited representational capabilities, making it difficult to achieve high-precision real-time learning of ship dynamic models in complex marine environments.
[0003] Over the past decade, with the continuous development of artificial intelligence technology and the continuous improvement of computing power, deep learning technology has become a key research direction in scientific research, commercial applications, and civilian use. Besides its remarkable achievements in image recognition and natural language processing, deep learning technology has also achieved research results far exceeding those of previous machine learning techniques in areas such as robot control, autonomous driving, and flight control. Meanwhile, deep neural networks (DNNs), as the foundational model of deep learning, have become an ideal choice for modeling complex dynamic systems due to their powerful nonlinear representation capabilities and feature memory characteristics.
[0004] Currently, integrating DNNs into the motion controllers of unmanned surface vessels (USVs) faces several key challenges, including: First, existing control methods based on DNNs for modeling system uncertainties and external disturbances suffer from network approximation errors, making them unsuitable for high-precision control scenarios. Second, the network training process of these methods often employs traditional offline learning methods, which cannot update the dynamic model of the controlled object in real time in the face of constantly changing and complex environments. Third, existing control methods based on Extended State Observers (ESOs) for estimating USV system uncertainties and external disturbances lack the ability to learn and memorize features, resulting in a lack of rapid and proactive responses to control inputs. Summary of the Invention
[0005] In view of this, the purpose of this invention is to propose an unmanned vessel trajectory tracking controller and control method based on online deep learning, so as to solve the technical problem that existing unmanned vessel motion control methods lack a fast and active response to control inputs.
[0006] The technical means employed in this invention are as follows:
[0007] An unmanned vessel trajectory tracking controller based on online deep learning includes a dynamic controller, an extended state observer, and an online deep learning vessel dynamic model;
[0008] The online deep learning ship dynamic model receives data collected from the dynamic dataset at the current moment and outputs the dynamic model learned by the DNN at the current moment to the extended state observer and dynamic controller;
[0009] The extended state observer receives the dynamic model learned by the DNN at the current moment and the actual speed information of the unmanned vessel, and outputs the estimated value of the DNN approximation error to the dynamic controller.
[0010] The dynamic controller receives the estimated value of the DNN approximation error, the given reference trajectory position information, the actual position information of the unmanned vessel, the actual speed information of the unmanned vessel, and the dynamic model learned by the DNN at the current moment, and outputs the derivative of the control input vector to the mathematical model of the controlled unmanned vessel.
[0011] Furthermore, the mathematical model of the controlled unmanned vessel is as follows:
[0012]
[0013] Where η=[p T ,ψ] T ∈R 3 p = [x, y] T ∈R 2 Let ν represent the position coordinates of the unmanned surface vessel (USV), and ψ∈(-π,π] represent the bow angle of the USV; ν=[u,v,r] T ∈R 3 For the unmanned surface vessel, the longitudinal velocity, lateral velocity, and bow roll rate are given; τ = [τ u ,τ v ,τ r ] T ∈R 3 For the control input of the unmanned vessel; M∈R 3×3 Let M be the inertial matrix of the unmanned vessel. * ∈R 3×3 The calibration value of M; R(ψ)∈R 3×3 Let f(ν,τ,c)∈R be the rotation matrix. 3 This is an unknown dynamic.
[0014] Furthermore, the steps for establishing the online deep learning ship dynamic model are as follows:
[0015] The DNN model is trained using an online learning method, and the loss function is defined as follows:
[0016]
[0017] in, denoted as , where p is the true distribution of the actual data; and P is the unknown distribution of the training data.
[0018] Based on the dynamic dataset {ν in the current iteration t ,τ t ,c t} Computing the dynamic model of DNN learning And calculate the loss value based on the above loss function. Based on this loss value The dynamic model learned by the DNN is optimized using the Adam optimizer. Update.
[0019] Furthermore, the formula for the extended state observer is:
[0020]
[0021] in, This is an estimate of ν; It is an estimate of f0; ω is the control gain; ω0 is the bandwidth of the extended state observer.
[0022] Furthermore, the formula for the dynamic controller is:
[0023]
[0024] Where θ represents the weights of the DNN, θ∈Θ, Θ is the parameter space; t is the number of iterations; γ is the quality gain coefficient, and w c To control the gain.
[0025] This invention also provides an unmanned vessel trajectory tracking control method based on online deep learning, implemented based on any of the above-mentioned unmanned vessel trajectory tracking controllers based on online deep learning, comprising the following steps:
[0026] Collect the current unmanned vessel control input τ t Wind disturbance information c t And unmanned surface vessel speed information ν t Constructing a dynamic dataset {ν t ,τ t ,c t};
[0027] Based on dynamic datasets, compute dynamic models learned by DNNs.
[0028] Extended state observer based on dynamic model The approximation error of the DNN is estimated, and the estimated value is obtained.
[0029] Reference trajectory information η r The unmanned surface vessel's actual position information η, its actual speed information ν, and the dynamic model learned by the DNN. and the estimated value of the DNN approximation error The result is transmitted to the dynamic controller to obtain the derivative of the control input vector.
[0030] The derivative of the control input vector The mathematical model input to the controlled unmanned vessel generates position and velocity information for the next moment, thereby controlling the controlled unmanned vessel.
[0031] Furthermore, a loss function is defined, and the loss value is calculated based on the loss function. Based on this loss value The dynamic model learned by the DNN is optimized using the Adam optimizer. Update.
[0032] Compared with the prior art, the present invention has the following advantages:
[0033] First, compared with existing control methods based on model uncertainty and external disturbances in DNN modeling systems, the online deep learning control method designed in this invention uses ESO to estimate the approximation error of DNN, ensuring the accuracy of network input information during DNN learning and enabling DNN to quickly capture model features.
[0034] Second, compared with existing control methods based on DNN modeling system model uncertainty and external disturbances, the online deep learning control method designed in this invention can update the learned unmanned vessel dynamic model in real time, which can effectively reduce the computational cost of repeated training.
[0035] Third, compared with existing control methods based on ESO to estimate the uncertainty of unmanned vessel system models and external disturbances, the online deep learning control method designed in this invention can utilize the memory characteristics of DNN to retain the learned dynamic model features of the unmanned vessel, improve the system's response speed to control inputs, and achieve high-performance tracking control. Attached Figure Description
[0036] 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.
[0037] Figure 1 This is a schematic diagram of the unmanned vessel trajectory tracking controller based on online deep learning according to the present invention.
[0038] Figure 2 This is a schematic diagram of the motion trajectory of the unmanned vessel using three different control methods of the present invention.
[0039] Figure 3 This is a diagram showing the trajectory tracking error of the unmanned vessel using three different control methods of this invention.
[0040] Figure 4 This is a graph showing the mean square error metric for DNN training in this invention.
[0041] Figure 5 This is a graph showing the longitudinal velocity, lateral velocity, and bow roll rate of the unmanned surface vessel of this invention.
[0042] Figure 6 This is an observational diagram of the DNN approximation error of this invention.
[0043] Figure 7 This is a control input diagram for the unmanned vessel using three different control methods of the present invention. Detailed Implementation
[0044] 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.
[0045] 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.
[0046] like Figure 1 As shown, the present invention provides an unmanned vessel trajectory tracking controller based on online deep learning, including a dynamic controller, an extended state observer, and an online deep learning vessel dynamic model.
[0047] The input terminals of the dynamic controller are connected to the output terminals of the extended state observer, the online deep learning ship dynamic model, and the unmanned vessel, respectively, and the output terminals are connected to the input terminals of the unmanned vessel and the dynamic dataset, respectively.
[0048] The input terminal of the extended state observer is connected to the output terminal of the online deep learning ship dynamic model and the unmanned vessel, respectively, and the output terminal is connected to the input terminal of the dynamic controller.
[0049] The input of the online deep learning ship dynamic model is connected to the output of the dynamic dataset, and the output is connected to the input of the extended state observer and the dynamic controller, respectively.
[0050] The mathematical model of the controlled unmanned vessel is expressed by the following formula:
[0051]
[0052] Where η=[p T ,ψ] T ∈R 3 p = [x, y] T ∈R 2 Let ν represent the position coordinates of the unmanned surface vessel (USV), and ψ∈(-π,π] represent the bow angle of the USV; ν=[u,v,r] T ∈R 3 For the unmanned surface vessel, the longitudinal velocity, lateral velocity, and bow roll rate are given; τ = [τ u ,τ v ,τ r ] T ∈R 3 For the control input of the unmanned vessel; M∈R3×3 For the inertial matrix of the unmanned vessel,
[0053] M * ∈R 3×3 The calibration value of M; R(ψ)∈R 3×3 Let be the rotation matrix, represented as:
[0054]
[0055] f(ν,τ,c)∈R 3 For unknown dynamics, it is represented as:
[0056] f(ν,τ,c)=M -1 (-C(ν)vD(ν)ν+τ w )+(M -1 -(M * ) -1 )τ (3)
[0057] Where C(ν)∈R 3×3 Here is the matrix of centripetal and Coriolis force coefficients; D(ν)∈R 3×3 Let c be the damping matrix, and c = [V w,t ,γ w,t ] T ∈R 2 Information about wind disturbance;
[0058] τ w =[τ wu ,τ wv ,τ wr ] T ∈R 3 For unknown ocean disturbances caused by wind, waves, and currents, we mainly consider wind disturbances, and the model is as follows:
[0059]
[0060] Among them, V w ∈R, γ w ∈R represent relative wind speed and relative wind direction, respectively; C X C Y C N ∈R represents the wind force coefficient; ρ a ∈R + air density; A Fw A Lw ∈R + These are the front and side projection wind zones, respectively; L oa ∈R + The total length of the unmanned vessel.
[0061] A. Dynamic Controller Design
[0062] The input signal to the dynamic controller is the given reference trajectory position information η. r =[x r ,y r ,ψ r ] T The unmanned surface vessel's actual position information η, its actual velocity information ν, and the estimated value of the DNN approximation error. Dynamic model of DNN learning in the t-th iteration The output signal is the derivative of the control input vector.
[0063]
[0064] The dynamic controller is designed as follows:
[0065]
[0066] Where the subscript θ represents the weights of the DNN, θ∈Θ, Θ is the parameter space; t is the number of iterations; γ is the quality gain coefficient, and w c To control the gain.
[0067] B. Design of Extended State Observer
[0068] The input signal to the extended state observer is the dynamic model learned by the DNN in the t-th iteration. The actual speed information ν of the unmanned surface vessel and the output signal are estimates of the DNN approximation error.
[0069] The extended state observer is designed as follows:
[0070]
[0071] in, This is an estimate of ν; It is an estimate of f0; ω0 represents the control gain; ω0 represents the ESO bandwidth.
[0072] C. Online deep learning ship dynamic model
[0073] The input signal for the online deep learning ship dynamics model is the data {ν} collected from the dynamic dataset in the t-th iteration. t ,τ t ,c t The output signal is the dynamic model learned by the DNN in the t-th iteration.
[0074]
[0075] The design of the online deep learning ship dynamics model is as follows:
[0076] The DNN model is trained using an online learning method. The loss function is defined as follows:
[0077]
[0078] in, denoted as , where p is the true distribution of the actual data, and P is the unknown distribution of the training data.
[0079] Based on the dynamic dataset {ν in the current iteration t ,τ t ,c t} Computing the dynamic model of DNN learning And calculate the loss value based on the above loss function. Based on this loss value The dynamic model learned by the DNN is optimized using the Adam optimizer. Update.
[0080] This invention also provides an unmanned surface vessel (USV) trajectory tracking and control method based on online deep learning, implemented using an USV trajectory tracking controller based on online deep learning, with the following steps:
[0081] In the t-th iteration, the unmanned vessel control input τ is collected. t Wind disturbance information c t And unmanned surface vessel speed information ν t Constructing a dynamic dataset {ν t ,τ t ,c t}, based on this dynamic dataset {ν t ,τ t ,c t Train a DNN online to obtain a dynamically learned model. Meanwhile, ESO is based on this dynamic model Estimate the DNN approximation error. Use the reference trajectory information η r The unmanned surface vessel's actual position information η, its actual speed information ν, and the dynamic model learned by the DNN. and the estimated value of the DNN approximation error The data is transmitted to the dynamic controller to control the unmanned vessel.
[0082] The control parameters for equations (5) and (6) are selected as follows:
[0083] γ = diag{50, 50, 30}, w c = diag{0.6,0.6,0.2}, K1=10, K2=25, where diag represents a diagonal matrix.
[0084] Simulation results are as follows Figure 2-7 As shown. Figure 2 The diagrams illustrate the motion trajectories of unmanned surface vessels using three different control methods, including the online deep learning control method (DNN+ESO) designed in this invention, the actual trajectory based on ESO, the actual trajectory based on DNN, and the reference trajectory. Although all three control methods achieve circular trajectory tracking with similar steady-state performance, it can be seen from the figures that the convergence speed of the online deep learning control method designed in this invention is significantly faster than the other two methods. Figure 3 The tracking errors of the unmanned surface vessel (USV) under three control methods are presented. It can be seen that the tracking errors in all three directions eventually converge to near zero. Figure 4 The mean squared error metric curve for DNN training is given. Figure 5 The curves of the longitudinal velocity, lateral velocity, and bow roll rate of the unmanned surface vessel are presented. Figure 6 The observed performance curves of the DNN approximation error are presented. It can be seen that the DNN approximation error and its estimated value converge to near zero in steady state, which indirectly demonstrates the learning and representation capabilities of the DNN in continuous training. Figure 7 The thrust in the longitudinal and lateral directions and the torque in the yaw direction of the unmanned surface vessel are given for three control methods.
[0085] 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. An online deep learning based unmanned ship trajectory tracking controller, characterized in that: This includes dynamic controllers, extended state observers, and online deep learning ship dynamic models; The online deep learning ship dynamic model receives data collected from the dynamic dataset at the current moment and outputs the dynamic model learned by the DNN at the current moment to the extended state observer and dynamic controller; The extended state observer receives the dynamic model learned by the DNN at the current moment and the actual speed information of the unmanned vessel, and outputs the estimated value of the DNN approximation error to the dynamic controller. The dynamic controller receives the estimated value of the DNN approximation error, the given reference trajectory position information, the actual position information of the unmanned vessel, the actual speed information of the unmanned vessel, and the dynamic model learned by the DNN at the current moment, and outputs the derivative of the control input vector to the mathematical model of the controlled unmanned vessel. The formula for the extended state observer is: in, for The estimated value; yes The estimated value; This represents the DNN approximation error. To control the gain; To expand the bandwidth of the state observer; For the unmanned vessel, the longitudinal velocity, lateral velocity, and bow roll rate are given. For the control input of the unmanned vessel; For the inertial matrix of the unmanned vessel, for The calibration value, For the first Dynamic model of DNN learning in each iteration; The formula for the dynamic controller is: in, For the weights of the DNN, , For parameter space; This represents the number of iterations. This is the mass gain coefficient. To control the gain, It is a rotation matrix; For unknown dynamics, Given the reference trajectory position information, This refers to the actual location information of the unmanned vessel.
2. The unmanned vessel trajectory tracking controller based on online deep learning according to claim 1, characterized in that, The mathematical model of the controlled unmanned vessel is as follows: in, , The coordinates of the unmanned vessel's position. For the bow angle of the unmanned vessel; For the unmanned vessel, the longitudinal velocity, lateral velocity, and bow roll rate are given. For the control input of the unmanned vessel; For the inertial matrix of the unmanned vessel, for The calibration value; It is a rotation matrix; This is an unknown dynamic.
3. The unmanned vessel trajectory tracking controller based on online deep learning according to claim 1, characterized in that, The steps for establishing the online deep learning ship dynamic model are as follows: The DNN model is trained using an online learning method, and the loss function is defined as follows: in, Expected loss; This represents the true distribution of actual data; The unknown distribution of the training data; Based on the dynamic dataset in the current iteration Computing the dynamic model of DNN learning And calculate the loss value based on the above loss function. Based on this loss value The Adam optimizer is used to optimize the dynamic model learned by the DNN. Update.
4. A method for trajectory tracking and control of an unmanned vessel based on online deep learning, implemented based on the unmanned vessel trajectory tracking controller based on online deep learning as described in any one of claims 1-3, characterized in that, The steps include the following: Collect current unmanned surface vessel control input Wind disturbance information And unmanned vessel speed information Building dynamic datasets ; Based on dynamic datasets, compute dynamic models learned by DNNs. , Extended State Observer Based on Dynamic Model The approximation error of the DNN is estimated, and the estimated value is obtained. ; Reference trajectory information Actual location information of unmanned vessels Actual speed information of unmanned vessels Dynamic models for DNN learning and the estimated value of the DNN approximation error The result is transmitted to the dynamic controller to obtain the derivative of the control input vector. ; The derivative of the control input vector The mathematical model input to the controlled unmanned vessel generates position and velocity information for the next moment, thereby controlling the controlled unmanned vessel.
5. The unmanned vessel trajectory tracking and control method based on online deep learning according to claim 4, characterized in that, Also includes: Define a loss function and calculate the loss value based on the loss function. Based on this loss value The Adam optimizer is used to optimize the dynamic model learned by the DNN. Update.