Neural network correction-based near-submarine robot interference observation controller and method

Abstract

The invention discloses a near-submarine robot disturbance observation controller and method based on neural network correction, and the controller consists of three parts: 1) a disturbance observer with RBF neural network correction, which is used for monitoring and compensating external disturbance in real time; the double-closed-loop integral sliding mode controller comprises a position ring and a speed ring and generates control torque according to expected and actual positions to guide navigation of the near-submarine robot; and 3) the comparator combines the outputs of the two, realizes accurate interference compensation by comparing the control torque with the disturbance estimation, and finally outputs the control torque containing the interference compensation. According to the method, under the condition that speed information is unobservable, the control precision and the anti-interference capability are improved by intelligently estimating modeling errors and external disturbance, and high-precision heading, pitch angle and depth control is achieved.

Description

Technical Field

[0001] The invention belongs to the field of automatic control and relates to a near-seabottom robot interference observation controller and method based on neural network correction. Background Art

[0002] The underwater motion of a near-seabot is a highly complex, multi-input, multi-output, nonlinear system. This inherent uncertainty makes it difficult to accurately model the system's dynamics, making control extremely complex. Internally, due to inaccurate measurement and modeling, the parameters of the dynamic model are difficult to accurately determine. Externally, the system is also subject to load fluctuations and various unpredictable disturbances. Therefore, appropriate control schemes are necessary to ensure the robustness of the near-seabot's control system in response to these parameter uncertainties and external disturbances.

[0003] To improve the tracking accuracy of near-seabottom robot control systems in the presence of external disturbances and parameter uncertainty, currently, the main approaches include adaptive control, disturbance observers, and robust control. Both adaptive and robust control methods require high accuracy in the near-seabottom robot's dynamic model. The basic idea behind disturbance observers is to construct a new dynamic system to observe or estimate the uncertainties in the original control system. The observer's estimated output is then used to offset the effects of these uncertainties, thereby improving the control performance of the existing controller. Therefore, disturbance observers effectively suppress unpredictable or random external disturbances, significantly enhancing the robustness of the system. However, due to the relatively precise model required, disturbance observers often present significant challenges in practical applications. Summary of the Invention

[0004] The purpose of the embodiments of the present invention is to provide a near-seabed robot interference observation controller and method based on neural network correction, so as to solve the problem that the existing near-seabed robot interference observation controller cannot take into account external unknown disturbances, resulting in low anti-interference ability and low control accuracy of the interference observation controller.

[0005] The technical solution adopted in the embodiment of the present invention is: a near-seabottom robot interference observation controller based on neural network correction, comprising:

[0006] A disturbance observer with RBF neural network correction is used to observe external disturbances;

[0007] A dual closed-loop integral sliding mode controller is used to provide a control torque for the near-seabottom robot's navigation process based on an input desired position and a current position fed back by the near-seabottom robot. The dual closed-loop integral sliding mode controller includes a position loop and a velocity loop.

[0008] The comparator is used to compare the control torque output by the dual closed-loop integral sliding mode controller and the disturbance estimate output by the disturbance observer with RBF neural network correction as negative feedback, perform disturbance observation compensation, and output the control torque with the disturbance observation compensation term.

[0009] Furthermore, the integral sliding mode controller of the position loop is shown as follows:

[0010]

[0011] e η =η-η d ;

[0012] Among them, s η represents the position sliding surface, e η represents the position error, η represents the position of the near-sea robot in the world coordinate system, and η d represents the desired position of the position loop, e η (δ) represents the function of position error over time δ, k1 represents the adjustable parameter of the position sliding surface; t is the time;

[0013] The integral sliding mode controller of the speed loop is shown in the following equation:

[0014]

[0015] e v =v b -v d ;

[0016]

[0017] Among them, s v represents the velocity loop sliding surface, e v represents the speed error, k2 represents the adjustable parameter of the speed loop sliding surface, e v (δ) represents the function of velocity error over time δ; v b represents the velocity of the near-sea robot in the body coordinate system, v d is the expected speed of the speed loop; is the first-order derivative of the expected position matrix of the near-sea robot in the world coordinate system, J -1 (η) represents the inverse matrix of the rotation matrix J(η) from the body coordinate system to the world coordinate system.

[0018] Furthermore, the control torque τ output by the comparator is expressed as follows:

[0019]

[0020] Among them, τ1 is the control torque output by the integral sliding mode controller of the speed loop, is the disturbance observer compensation term output by the disturbance observer with RBF neural network correction; M a is the inertia matrix corresponding to the vertical control force, s v represents the velocity loop sliding surface, e v represents the speed error, and k2 represents the adjustable parameter of the speed loop sliding surface.

[0021] Furthermore, the interference observer with RBF neural network correction includes:

[0022] RBF neural network is used to fit the direct error between the prior underwater dynamics model and the actual underwater dynamics model of the near-sea robot;

[0023] The disturbance observer is used to estimate the external disturbance torque and correct the estimated external disturbance torque based on the fitting result of the RBF neural network.

[0024] Furthermore, the RBF neural network inputs attitude angle information directly related to the frontal surface, the acceleration integral term, and the thruster output torque information. The RBF neural network is connected through weights to output the acceleration of the prior underwater dynamics model, i.e., the expected acceleration, the angular velocity of the prior underwater dynamics model, and the difference between the actual acceleration and the actual angular velocity;

[0025] The loss function of the RBF neural network is obtained by the difference between the actual acceleration information that can be measured and the acceleration information obtained through the prior underwater dynamics model, and the mean square loss function is adopted.

[0026] Furthermore, the output of the RBF neural network is:

[0027]

[0028] in, represents the torque effect due to modeling error and thrust derating, W * is the ideal fitting parameter of RBF neural network, W *T It's W * The transpose of , S(Z) represents the hidden layer of the RBF neural network, W *T S(Z) is the value of the RBF neural network output layer under the ideal fitting parameters, and ε is the value of the RBF neural network output layer under the ideal fitting parameters W. * The error below.

[0029] Furthermore, the output of the disturbance observer is:

[0030]

[0031] in, is the disturbance torque estimated by the disturbance observer with RBF neural network correction, β is the defined intermediate variable, is the first-order derivative of β, K0 is an adjustable parameter; v b represents the speed of the near-sea robot in the body coordinate system, M a Represents the inertia matrix corresponding to the vertical control force, C a (v b ) represents the Coriolis force matrix corresponding to the heeling control force, D a (v b ) represents the damping term corresponding to the pitch control force, g a (η) represents the yaw control force, i.e., the restoring force matrix, G ca represents the Golgi force matrix generated by the rotation of the near-seafloor robot thrusters; yes The transpose of represents the actual fitting parameters of the RBF neural network obtained through training, Represents the value of the RBF neural network output layer under the actual fitting parameters; τ a is the control force output obtained according to the prior underwater dynamics model.

[0032] The interference observation control method of the near-seabottom robot based on neural network correction uses the above-mentioned interference observation controller of the near-seabottom robot based on neural network correction for control.

[0033] Further, Represents the ideal fitting parameter W of the RBF neural network for nonlinear functions * and the actual fitting parameters The error, is the actual estimated value of the RBF neural network, Represents the ideal fitting parameter W of the RBF neural network * and the actual fitting parameters The error, yes The transpose of , there exists a positive real number σ such that

[0034] in, is the error of the disturbance observer with RBF neural network correction, τ D is the disturbance torque of the near-seabottom robot, is the disturbance observation compensation term output by the disturbance observer with RBF neural network correction, Taking the derivative we get

[0035]

[0036] Wherein, τ is the control torque output by the comparator;

[0037] definition V3 represents half of the square of the external interference observation error, for The transpose of V3 is obtained by taking the derivative of V3.

[0038]

[0039] Among them, the intermediate variable is τ D The first-order derivative of , the intermediate variable μ∈(0,1), D is the upper limit of the rate of change of the external interference force;

[0040] Establish the global Lyapunov function V:

[0041]

[0042] Among them, V1 represents the energy function of position error, V2 represents the energy function of velocity error, and V3 represents the energy function of the error between the disturbance observation force of the disturbance observer with RBF neural network correction and the actual external disturbance force; represent The transpose of s η represents the position sliding surface, For s η The transpose of s v represents the sliding surface of the velocity loop, For s v The transpose of

[0043] Taking the derivative of the global Lyapunov function, we get:

[0044]

[0045] in, For s η The first derivative of For s v The first-order derivative of , we get:

[0046]

[0047] Among them, e v (t) represents the function of velocity error with time t, for e v (t) is the first-order derivative with respect to time, k2 represents the adjustable parameter of the velocity loop sliding surface; K0 represents a 4×4 diagonal matrix, and the elements on the diagonal of K0 correspond to the coefficients multiplied by the four degrees of freedom of the near-sea robot four-degree-of-freedom model; λ min (K0) represents the minimum value in the main diagonal of the diagonal matrix K0, λmax (K0) represents the maximum value in the main diagonal of the diagonal matrix K0; is the first-order derivative of the current body velocity, is the first derivative of the desired velocity; M a -1 Inertia matrix M corresponding to the vertical control force a The inverse of

[0048] make have to:

[0049]

[0050] Among them, Γ v is the intermediate parameter,

[0051] Select K0, k2 to ensure 1-k2M a -1 、 are all positive real numbers;

[0052] The position error, speed error and external interference error The combined total energy function satisfies the following formula:

[0053]

[0054] Among them, V(t) represents the change function of the total energy function V with time t, and V(0) represents the total energy function value at the initial moment, that is, time 0.

[0055] The beneficial effects of the embodiments of the present invention are as follows: a near-sea robot interference observation controller and method based on neural network correction are disclosed, which can use an RBF neural network to estimate the torque change of the near-sea robot caused by modeling errors and thrust depreciation when speed information is unobservable. By combining the RBF neural network and the interference observer to form an interference observer with RBF neural network correction, external disturbances can be estimated when speed information is unobservable, and the accuracy and convergence time of the interference observer can be improved; based on the interference observer with RBF neural network correction, combined with a dual closed-loop integral sliding mode controller with a position loop and a velocity loop, a heading control force, a pitch control force and a depth control force are output, and ultimately high-precision control of the heading, pitch angle and depth during navigation is achieved; by utilizing the correction effect of the neural network on the interference observer, unknown external disturbances can be estimated when the ship speed and external disturbances are unknown, thereby improving the anti-interference capability of the overall control system; and solves the problem that the existing near-sea robot interference observation controller cannot consider unknown external disturbances, resulting in low anti-interference capability and low control accuracy of the interference observation controller. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 4 is a flowchart of the neural network-based estimation algorithm of this embodiment.

[0058] Figure 2 This is a structural block diagram of the near-seabottom robot interference observation controller based on neural network correction in this embodiment.

[0059] Figure 3 It is a control curve diagram for straight flight using different pitch angles. Figure 3 (a) to (c) correspond to the curves of pitch angle, depth and heading angle changing with time under the control of different controllers. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0061] This embodiment provides a neural network-corrected near-sea robot interference observation and control method, comprising the following steps:

[0062] Based on prior knowledge, the prior underwater dynamics model of the near-seafloor robot is established as shown in the following formula:

[0063]

[0064] Among them, M RB is the rigid body inertia matrix, C RB (v b ) is the Coriolis force matrix;

[0065]

[0066] Where m represents the mass of the machine, z G 、y G 、x G Represents the three-dimensional position of the robot's center of gravity relative to the robot's centroid. x , I y , I z , I xy , Ixz , I yz Represents the moment of inertia of the robot. For example, I x Represents the robot's moment of inertia around the x-axis, I xy Represents the product of the robot's x and y moments of inertia. The physical meaning is that when the robot rotates around the x-axis, it will cause the y-axis to rotate.

[0067]

[0068] Among them, v b is the speed of the robot near the seabed, is the acceleration of the near-seabottom robot, τ is the control input force of the near-seabottom robot (that is, the controller output control force), τ D is the external interference force and torque on the near-sea robot, τ H are the hydrodynamic forces and moments:

[0069]

[0070] Among them, M A is the added mass,

[0071] in, The coefficient representing the coupling between the robot's x-axis rotational motion and the robot's y-axis translational motion in the added mass, Represents the additional inertia of the robot when it rotates around the x-axis, C A (v b ) is the Coriolis force matrix of the added mass;

[0072]

[0073] in: Represents the additional mass coefficient when the robot moves linearly along the robot's z, y, and x axes. represents the coupling coefficient of the robot's y-axis translation and rotation along the x-axis, Represents the coupling coefficient of the robot's x-axis translation and y-axis rotation, represents the additional inertia of the robot rotating along the z-axis, represents the coupling coefficient of the robot's y-axis rotation and x-axis translation, represents the additional inertia of the robot rotating along the y-axis, D(v b ) is the hydrodynamic damping term:

[0074] D(v b )=-diag{X u ,Y v ,Z w ,K p ,M q ,Nr}

[0075] -diag{X |u|u |u|,Y |v|v |v|,Z |w|w |w|,K |p|p |p|,M |q|q |q|,N |r|r |r|}

[0076] Among them, X u 、Y v , Z w represents the linear friction coefficient caused by the laminar boundary layer when the robot translates along the x, y, and z axes, K p 、M q 、N r represents the linear friction coefficient caused by the laminar boundary layer when the robot rotates along the x, y, and z axes. |u|u 、Y |v|v , Z |w|w represents the secondary friction coefficient caused by the turbulent boundary layer when the robot translates along the x, y, and z axes, K |p|p 、M |q|q 、N |r|r represents the secondary friction coefficient caused by the turbulent boundary layer when the robot rotates along the x, y, and z axes. |u|, |v|, |w|, |p|, |q|, and |r| represent the absolute values ​​of u, v, w, p, q, and r. g(η) is the Gökçen force matrix for the position η of the near-seabottom robot in the world coordinate system:

[0077]

[0078] Among them, W = mg represents the weight of the robot, B = ρg▽ represents the buoyancy of the robot underwater, [x B ,y B ,z B ] is the buoyancy center coordinate; η = [xy zφθψ], where x, y, and z correspond to the x, y, and z axis positions of the near-sea robot in the world coordinate system, and φ, θ, and ψ correspond to the roll angle, pitch angle, and heading angle; the velocity v of the near-sea robot in the body coordinate system is b =[uv wp qr], where u, v, and w correspond to the velocities of the near-seabot in the x, y, and z axes of the body coordinate system, and p, q, and r correspond to the angular velocities of the near-seabot in the x, y, and z axes of the body coordinate system;

[0079] The corresponding relationship between the world coordinate system and the body coordinate system is:

[0080]

[0081] in, represents the first-order derivative of the current pose (position and attitude) matrix of the near-sea robot in the world coordinate system, and J(η) represents the rotation matrix from the body coordinate system to the world coordinate system;

[0082] According to the correspondence between the world coordinate system and the body coordinate system, the integral sliding mode controller of the position loop is designed as shown in the following formula:

[0083]

[0084] Among them, s η represents the sliding surface of the position loop, e η represents the error between the current position and the desired position of the near-sea robot in the world coordinate system, e η =η-η d , η d represents the desired position of the near-sea robot in the world coordinate system, e η (δ) represents the function of position error over time δ, where δ represents the time differential in [0, t] and k1 represents the adjustable parameter of the position loop sliding surface;

[0085] The virtual speed control quantity can be obtained by using the integral sliding mode controller of the position loop. On this basis, the control design of the speed loop can be carried out. According to the correspondence between the world coordinate system and the body coordinate system, namely η and v b The corresponding relationship between , the desired speed control amount can be obtained as:

[0086]

[0087] Where, v d is the expected speed of the speed loop in the body coordinate system, is the first-order derivative of the expected position matrix of the near-sea robot in the world coordinate system, J -1 (η) is the inverse matrix of the rotation matrix J(η);

[0088] The desired velocity is selected based on the established position loop Lyapunov function Taking the derivative of V2, where V2 represents the energy function of the velocity error, we can get:

[0089]

[0090] in, The sliding surface s of the position loop η The transpose of

[0091] Transforming the expected speed into Then we get:

[0092]

[0093] That can guarantee The rationality of the selection of the expected speed is proved to be able to ensure e η =η-η d Converging to zero ensures that the near-sea robot can guarantee the convergence of position tracking according to the expected speed, that is, the position can converge to the expected position.

[0094] Considering the near-sea robot model and environmental interference, a velocity loop tracking controller (velocity sliding mode controller) is designed to control the near-sea robot to track the desired velocity:

[0095] Combining formulas (1) and (2) yields:

[0096]

[0097] Where, τ=[00f z f p f q f r ] T Represents the body output control force. Since the robot cannot output the force of the body coordinate system x, y, it is zero. z ,f p ,f q ,f r Represents the output force along the z-axis of the body coordinate system and the rotational torque along the x, y, and z-axes of the body coordinate system.

[0098] Due to the limitation of the robot's output force dimension, we can make the following simplifications, ignoring the external interference force τ D Under the premise of , only the dimension in which the robot can output control force is considered in the dynamic model, and the following four-degree-of-freedom dynamic model of the near-sea robot can be obtained

[0099]

[0100] Where M a Represents the inertia matrix corresponding to the vertical control force, C a (v b ) represents the Coriolis force matrix corresponding to the heeling control force, D a (v b ) represents the damping term corresponding to the pitch control force, g a (η) represents the yaw control force, G ca represents the Golgi force matrix generated by the rotation of the propeller of the near-sea robot;

[0101]

[0102] Among them, O 2×6 Represents the two degrees of freedom that cannot be actively controlled, namely the x and y axes;

[0103]

[0104] D a (v b )=-diag{Z w ,K p ,M q ,N r}

[0105] -diag{Z |w|w |w|,K |p|p |p|,M |q|q |q|,N |r|r |r|}

[0106]

[0107] i represents the i-th propeller, J represents the propeller's moment of inertia, Ω represents the vector of the fuselage's angular velocity, and Ω = [p, q, r] T ,ω i represents the angular velocity of the propeller rotation, Represents the direction of the x, y, and z axes of the body coordinate system in the world coordinate system; τ a is the control force output obtained according to the prior underwater dynamics model, τ a =[f z f p f q f r ] T .

[0108] The following are all the forces and moments of the near-sea robot, in addition to the information obtained from the prior underwater dynamics model, including the effects of modeling errors, environmental disturbances, and thrust reduction. The four-degree-of-freedom model of the near-sea robot is rewritten as follows:

[0109]

[0110] in, represents the torque effect due to modeling errors and thrust derating;

[0111] The design of the sliding mode controller of the speed loop is shown as follows:

[0112]

[0113] Among them, s v represents the velocity sliding surface, e v represents the speed error, e v =v b -v d , k2 represents the adjustable parameter of the velocity sliding surface, ev (δ) represents the function of velocity error over time δ, which is obtained by adding time to the error when calculating the error;

[0114] The control torque τ set by the integral sliding mode controller of the speed loop is as follows:

[0115]

[0116] Where τ1 is the control torque obtained by the speed loop sliding mode controller, Represents the disturbance observer compensation term of the disturbance observer output.

[0117] The logic of estimating the torque due to modeling error and thrust reduction based on RBF neural network is as follows: Figure 1 As shown in the figure, the motion state information is input into the RBF neural network model. The motion state information includes the attitude angle information directly related to the oncoming surface (roll angle φ, pitch angle θ and heading angle ψ), the acceleration integral term (the acceleration integral of the world coordinate system x, y, and z axes), and the thruster output torque information, i.e., τ a =[f z f p f q f r ] T ,

[0118] Right now,

[0119] l x Indicates the distance of the propeller on the x-axis of the body coordinate system, l y Represents the distance of the propeller on the y-axis of the body coordinate system. RBF neural networks map low-dimensional composite information to high-dimensional acceleration information through weighted connections. This process involves selecting appropriate basis function center points (used to perform nonlinear mapping on the input data, mapping low-dimensional composite information to high-dimensional acceleration information), calculating basis function variances (appropriate variance values ​​can balance the model's fitting and generalization capabilities, avoiding overfitting or underfitting), activating calculations, calculating basis function outputs, and combining them using weights to obtain the final result. The specific steps are as follows:

[0120] Center calculation: Use a clustering algorithm (such as k-means) to determine a set of center points. These center points represent specific locations of the original data in the feature space. Usually, the clustering algorithm automatically selects the center point that best represents the data distribution.

[0121] Basis function variance calculation: For each center point, calculate the variance of the corresponding basis function;

[0122] Activation calculation: For each sample in the input data, calculate the distance between it and each center point, and define the distance function r = || XX i ||, Supplement: X represents the center point, X i Represents the data point, represents the distance between sample points, and inputs the distance as the weight into the basis function. The basis function maps the input data to the high dimension in the feature space. The basis function is the Gaussian kernel radial basis function. The specific expression of the Gaussian kernel function is: Among them: φ(r) represents the Gaussian kernel function, ε represents the hyperparameter of the Gaussian kernel function, r represents the distance, and the activation function is expressed as follows:

[0123]

[0124] Among them, Z represents the output of the neural network, s i (Z) is the abbreviation of each neuron, i represents the serial number of the neuron, which is used to distinguish different neurons; c i is the center vector, b i is the base width parameter;

[0125] Weight calculation: The output value of each basis function is multiplied by the corresponding weight, and then the output values ​​of all basis functions are weighted and summed to obtain the final output acceleration value. The weight is usually learned through a training process, using the stochastic gradient descent method for training. The loss function is obtained by the difference between the measurable acceleration information and the acceleration information obtained by the prior underwater dynamics model. The loss function adopts the mean square loss function. Represents the mean of the sum of squares of the differences between the predicted value Γ and the target value Z, and n represents the sample size. The structure of the neural network has 150 hidden layers and 13 inputs, namely: depth, roll angle φ, pitch angle θ, heading angle ψ, speed w of the z-axis of the body coordinate system, angular velocities p, q, r of the x, y, and z-axes of the body coordinate system, speeds of the four propellers, and acceleration of the x-axis in the world coordinate system. The input is based on depth, roll angle φ, pitch angle θ, heading angle ψ, speed w of the z-axis of the body coordinate system, angular velocities p, q, r of the x, y, and z-axes of the body coordinate system, speeds of the four propellers, and acceleration of the x-axis in the world coordinate system to obtain the current motion state of the robot, and then based on the current motion state, the robot can be predicted according to the following equations: Get acceleration information Then, the acceleration difference is obtained by comparing the actual robot's acceleration information, and the motion state information and the acceleration difference are input into the RBF neural network. The final training goal is to minimize the acceleration difference, so that the RBF neural network can be used to characterize the torque effect caused by modeling error and thrust reduction.

[0126] Since RBF neural network has strong fitting ability for nonlinear functions, the following reasonable assumptions are made for the designed estimation algorithm based on neural network RBF: Represents the torque effect caused by modeling error and thrust derating. The RBF neural network estimates the torque effect caused by modeling error and thrust derating as follows:

[0127]

[0128] Among them, W * It is the best fitting parameter of RBF neural network for nonlinear function, that is, the ideal fitting parameter. The activation function of neurons in RBF neural network is c, b are the center vector and base width parameters; the fitting parameters are the fitting parameters c and b of each neuron, W *T It's W * The transpose of , S(Z) represents the hidden layer of the RBF neural network, W *T S(Z) is the value of the output layer of the RBF neural network under the ideal fitting parameters, and ε is the value of the output layer of the RBF neural network under the ideal fitting parameters W. * There is an estimation error under

[0129] definition Represents the ideal fitting parameter W of the RBF neural network for nonlinear functions * and the actual fitting parameters Due to the universal approximation ability of RBF neural network, as long as the appropriate loss function and gradient descent algorithm are selected, the following reasonable assumptions can be made:

[0130]

[0131] Where P is the number of training times, it can be obtained that when P continues to increase, there exists a positive real number σ such that yes The transpose of is the actual estimated value of the RBF neural network. The designed RBF neural network model can be used to calculate the above unknown interference torque. Performing online estimation and compensating into the designed controller can improve the control effect of the controller.

[0132] The disturbance observer with RBF neural network correction designed under the consideration of modeling error (estimating the disturbance torque of the environment during the movement of the near-sea robot) is defined as follows:

[0133]

[0134] in, represents the disturbance torque estimated by the disturbance observer with RBF neural network correction; β is the defined intermediate variable, and its first-order derivative is In order to facilitate the subsequent proof of the determined intermediate variable, the initial value of β is 0; K0 is an adjustable parameter.

[0135] The structure of the near-seabottom robot interference observation controller based on RBF neural network correction is as follows: Figure 2 As shown in the figure, its design primarily considers the complex disturbances encountered during direct navigation, utilizing a dual closed-loop integral sliding mode controller to improve the robustness of the overall controller. In actual situations, modeling errors and unknown speeds exist. The strong fitting properties of RBF neural networks are used to fit the direct error between the prior underwater dynamics model (underwater dynamics model) and the actual underwater dynamics model of the near-sea robot. The loss function of the RBF neural network is constructed using observable acceleration information, and a disturbance observer with RBF neural network correction is designed to compensate for unknown disturbances encountered during the actual process.

[0136] definition is the error of the disturbance observer with RBF neural network correction, Taking the derivative we get

[0137]

[0138] definition The transpose of V3 represents half of the square of the external interference observation error. Taking the derivative of V3, we get based on Will as well as The calculation formula is substituted into:

[0139]

[0140] Among them, the intermediate variable Intermediate variable μ∈(0,1), D is the upper limit of the rate of change of the external interference force, and different upper limits D can be set according to different situations.

[0141] Lyapunov theory can effectively prove whether the controller converges and verify whether the control error can eventually converge. According to the different components, the global Lyapunov function is established:

[0142]

[0143] Wherein, V1 represents the energy function of the position error, V2 represents the energy function of the velocity error, and V3 represents the energy function of the error between the disturbance observation force of the disturbance observer with RBF neural network correction in this embodiment and the actual external disturbance force; represents the error between the disturbance observation force of the disturbance observer with RBF neural network correction and the actual external disturbance force, represent The transpose of .

[0144] Taking the derivative of the global Lyapunov function, we get:

[0145]

[0146] We can get:

[0147]

[0148] in, Represents e v (t) is the first-order derivative with respect to time; K0 represents a 4×4 diagonal matrix. Since only the control model with four degrees of freedom is considered, the elements on the diagonal of K0 represent the coefficients multiplied by each degree of freedom, λ min (K0) represents the minimum value in the main diagonal of the diagonal matrix K0, λ max (K0) represents the maximum value in the main diagonal of the diagonal matrix K0; is the first-order derivative of the current body velocity, is the first derivative of the desired velocity. In this embodiment, the symbol is the derivative symbol, for example It is v d The first derivative of is the first derivative of V.

[0149] make The above formula can be rewritten as:

[0150]

[0151] Choosing appropriate K0 and k2 can ensure 1-k2M a -1 、 are all positive real numbers, Γ v As an intermediate parameter, the total energy function V composed of position error, velocity error and external interference error satisfies the following formula:

[0152]

[0153] Where V(t) represents the total energy function V over time, and V(0) represents the total energy function value at the initial time, i.e., time 0. The total energy function V represents a function composed of errors. When the energy function V is less than a certain value, it means that the control error can converge to a certain value.

[0154] It can be seen from the above formula that the final overall error is ultimately uniformly bounded. By reasonably setting the coefficients K0 and k2, the control error can converge to any bounded region. Since the convergence of the energy function is independent of the size of k1, it is only necessary to ensure that k1 is a positive number.

[0155] In order to verify the performance of the disturbance observation controller based on RBF neural network correction, a simulation comparison experiment was set up. In the simulation comparison experiment, 15% uncertainty was added to the near-sea robot model parameters to simulate the actual modeling error of the hydrodynamic parameters. The disturbance term was set as:

[0156] τ D =[10sin(t) 0.8sin(t) 0.8sin(t) 0.8cos(t)+1] T ;

[0157] Four thruster saturation parameters |F i |≤37N,i∈{1,2,3,4};

[0158] The controller parameters are as follows:

[0159]

[0160] The simulation process is designed as follows: the underwater near-seabottom robot uses different pitch angles such as 35 degrees and 50 degrees to navigate straight, and compares the attitude control effect of the traditional PID controller and the interference observation controller based on RBF neural network correction and the interference observer without neural network correction during the straight navigation process. The simulation experiment compares the attitude and depth control capabilities of the near-seabottom robot when using different pitch angles such as 35 degrees and 50 degrees for straight navigation. The final control effect is as follows Figure 3 As shown. Figure 3 It can be seen that the interference observation controller of the near-sea robot based on neural network correction proposed in this embodiment can realize the near-sea robot to sail straight with a large pitch angle through the estimation of the model by the neural network and the design of the double closed-loop integral sliding mode controller based on the interference observer without obtaining the speed. From the simulation comparison results, it can be seen that by comparing PID, no RBF neural network correction and with RBF neural network correction, the final control effect is: the maximum pitch angle error after stabilization corresponds to 5.84 degrees, 0.79 degrees, and 0.18 degrees; the maximum depth error after stabilization corresponds to 0.074m, 0.015m, and 0.0015m; the maximum heading angle error after stabilization corresponds to 3.94 degrees, 0.24 degrees, and 0.05 degrees (using the prior model Design a disturbance observer and use neural network to estimate get Then, the control effect is obtained by using the interference observer based on the RBF neural network correction), which shows that the near-sea robot interference observation controller based on the neural network correction proposed in this embodiment can significantly improve the attitude and depth control effects during the straight navigation process.

[0161] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A near-seabottom robot interference observation controller based on neural network correction, characterized by: include: A disturbance observer with RBF neural network correction is used to observe external disturbances; A dual closed-loop integral sliding mode controller is used to provide a control torque for the near-seabottom robot's navigation process based on an input desired position and a current position fed back by the near-seabottom robot. The dual closed-loop integral sliding mode controller includes a position loop and a velocity loop. The comparator is used to compare the control torque output by the dual closed-loop integral sliding mode controller and the disturbance estimate output by the disturbance observer with RBF neural network correction as negative feedback, perform disturbance observation compensation, and output the control torque with the disturbance observation compensation term.

2. The near-sea robot interference observation controller based on neural network correction according to claim 1 is characterized in that: The integral sliding mode controller of the position loop is shown in the following equation: e η =th-th d ; Among them, s η represents the position sliding surface, e η represents the position error, η represents the position of the near-sea robot in the world coordinate system, and η d represents the desired position of the position loop, e η (δ) represents the function of position error over time δ, k1 represents the adjustable parameter of the position sliding surface; t is the time; The integral sliding mode controller of the speed loop is shown in the following equation: yes v =v b -v d ; Among them, s v represents the velocity loop sliding surface, e v represents the speed error, k2 represents the adjustable parameter of the speed loop sliding surface, e v (δ) represents the function of velocity error over time δ; v b represents the velocity of the near-sea robot in the body coordinate system, v d is the expected speed of the speed loop; is the first-order derivative of the expected position matrix of the near-sea robot in the world coordinate system, J -1 (η) represents the inverse matrix of the rotation matrix J(η) from the body coordinate system to the world coordinate system.

3. The near-sea robot interference observation controller based on neural network correction according to claim 1 is characterized in that: The control torque τ output by the comparator is shown as follows: Among them, τ1 is the control torque output by the integral sliding mode controller of the speed loop, is the disturbance observer compensation term output by the disturbance observer with RBF neural network correction; M a is the inertia matrix corresponding to the vertical control force, s v represents the velocity loop sliding surface, e v represents the speed error, and k2 represents the adjustable parameter of the speed loop sliding surface.

4. The near-sea robot interference observation controller based on neural network correction according to any one of claims 1 to 3, characterized in that: The interference observer with RBF neural network correction includes: RBF neural network is used to fit the direct error between the prior underwater dynamics model and the actual underwater dynamics model of the near-sea robot; The disturbance observer is used to estimate the external disturbance torque and correct the estimated external disturbance torque based on the fitting result of the RBF neural network.

5. The near-sea robot interference observation controller based on neural network correction according to claim 4 is characterized in that: The RBF neural network inputs attitude angle information directly related to the oncoming surface, the acceleration integral term, and the thruster output torque information. The RBF neural network is connected through weights to output the acceleration of the prior underwater dynamics model, that is, the expected acceleration, the angular velocity of the prior underwater dynamics model, and the difference between the actual acceleration and the actual angular velocity; The loss function of the RBF neural network is obtained by the difference between the actual acceleration information that can be measured and the acceleration information obtained through the prior underwater dynamics model, and the mean square loss function is adopted.

6. The near-sea robot interference observation controller based on neural network correction according to claim 4 is characterized in that: The output of the RBF neural network is: in, represents the torque effect due to modeling error and thrust derating, W * is the ideal fitting parameter of RBF neural network, W *T It's W * The transpose of , S(Z) represents the hidden layer of the RBF neural network, W *T S(Z) is the value of the RBF neural network output layer under the ideal fitting parameters, and ε is the value of the RBF neural network output layer under the ideal fitting parameters W. * The error below.

7. The near-sea robot interference observation controller based on neural network correction according to claim 4 is characterized in that: The output of the disturbance observer is: in, is the disturbance torque estimated by the disturbance observer with RBF neural network correction, β is the defined intermediate variable, is the first-order derivative of β, K0 is an adjustable parameter; v b represents the speed of the near-sea robot in the body coordinate system, M a Represents the inertia matrix corresponding to the vertical control force, C a (v b ) represents the Coriolis force matrix corresponding to the heeling control force, D a (v b ) represents the damping term corresponding to the pitch control force, g a (η) represents the yaw control force, i.e., the restoring force matrix, G ca represents the Golgi force matrix generated by the rotation of the near-seafloor robot thrusters; yes The transpose of represents the actual fitting parameters of the RBF neural network obtained through training, Represents the value of the RBF neural network output layer under the actual fitting parameters; τ a is the control force output obtained according to the prior underwater dynamics model.

8. A method for interference observation and control of a near-sea robot based on neural network correction, characterized in that: The control is performed using the near-seabottom robot interference observation controller based on neural network correction as described in any one of claims 1 to 3 and 5 to 7.

9. The interference observation and control method for a near-sea robot based on neural network correction according to claim 8 is characterized in that: Represents the ideal fitting parameter W of the RBF neural network for nonlinear functions * and the actual fitting parameters The error, is the actual estimated value of the RBF neural network, Represents the ideal fitting parameter W of the RBF neural network * and the actual fitting parameters The error, yes The transpose of , there exists a positive real number σ such that 10. The interference observation and control method for a near-sea robot based on neural network correction according to claim 9 is characterized in that: is the error of the disturbance observer with RBF neural network correction, τ D is the disturbance torque of the near-seabottom robot, is the disturbance observation compensation term output by the disturbance observer with RBF neural network correction, and for τ D Taking the derivative we get Wherein, τ is the control torque output by the comparator; definition V3 represents half of the square of the external interference observation error, for The transpose of V3 is obtained by taking the derivative of V3. Among them, the intermediate variable is τ D The first-order derivative of , the intermediate variable μ∈(0,1), D is the upper limit of the rate of change of the external interference force; Establish the global Lyapunov function V: Among them, V1 represents the energy function of position error, V2 represents the energy function of velocity error, and V3 represents the energy function of the error between the disturbance observation force of the disturbance observer with RBF neural network correction and the actual external disturbance force; represent The transpose of s η represents the position sliding surface, For s η The transpose of s v represents the sliding surface of the velocity loop, For s v The transpose of Taking the derivative of the global Lyapunov function, we get: in, For s η The first derivative of For s v The first-order derivative of , we get: Among them, e v (t) represents the function of velocity error with time t, for e v (t) is the first-order derivative with respect to time, k2 represents the adjustable parameter of the velocity loop sliding surface; K0 represents a 4×4 diagonal matrix, and the elements on the diagonal of K0 correspond to the coefficients multiplied by the four degrees of freedom of the near-sea robot four-degree-of-freedom model; λ min (K0) represents the minimum value in the main diagonal of the diagonal matrix K0, λ max (K0) represents the maximum value in the main diagonal of the diagonal matrix K0; is the first-order derivative of the current body velocity, is the first derivative of the desired velocity; M a -1 Inertia matrix M corresponding to the vertical control force a The inverse of make have to: Among them, Γ v is the intermediate parameter, Select K0, k2 to ensure 1-k2M a -1 、 are all positive real numbers; The position error, speed error and external interference error The combined total energy function satisfies the following formula: Among them, V(t) represents the change function of the total energy function V with time t, and V(0) represents the total energy function value at the initial moment, that is, time 0.

Images