Manned submersible control method based on neural network disturbance observer

Abstract

The present application relates to a manned submersible control method based on a neural network disturbance observer, comprising the following steps: step 1, constructing a kinematics and dynamics model of the manned submersible; step 2, processing sea trial data of the manned submersible, calculating input and output data of the neural network, and performing normalization processing on the input and output data to generate training data; step 3, training an LSTM neural network using the training data, and using the trained neural network as an observer to estimate disturbances; and step 4, constructing a model predictive controller to achieve motion control of the manned submersible. The LSTM proposed in the present application can accurately estimate the disturbance received by the manned submersible. The control method of the present application can effectively suppress various disturbances received by the manned submersible, and does not require a control distribution module, thereby simplifying the control process and avoiding the problems of propeller thrust and thrust rate saturation, thereby ensuring high control accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the motion control field of manned submersible, and particularly relates to a manned submersible control method based on a neural network disturbance observer. BACKGROUND

[0002] The manned submersible plays an important role in the development and utilization of the ocean, and can carry technical personnel and scientists to the deep sea destination for direct observation and analysis, and fully exert the wisdom of human beings. In the complex environment, the control system has a high control performance requirement.

[0003] The existing manned submersible usually adopts the PID control technology. However, when the system has strong nonlinearity and disturbance, the single PID control precision is poor. In addition, the existing anti-disturbance control algorithm mostly only considers the influence of the sea current on the manned submersible. In addition to the sea current, the model parameter perturbation of the manned submersible and the inaccurate dynamics model of the thruster also affect the control precision of the manned submersible. Accurate estimation of these disturbances can improve the control precision of the manned submersible. Moreover, the control method of most manned submersibles uses the pseudo-inverse matrix as the control distribution module to solve the overdrive problem of the manned submersible. The control system process of this kind of control system is complex, and the thruster thrust obtained may have the problems of thruster thrust and thruster speed saturation. SUMMARY

[0004] The present application aims to provide a manned submersible control method based on a neural network disturbance observer, which effectively solves the influence of the sea current and the system parameter time-varying nonlinearity, model error uncertainty and other disturbances on the control precision, and solves the thruster thrust saturation problem without a control distribution module.

[0005] The technical solution adopted by the present application to achieve the above-mentioned purpose is as follows: a manned submersible control method based on a neural network disturbance observer, comprising the following steps:

[0006] 1) constructing a model predictive controller according to the dynamics model of the manned submersible;

[0007] 2) obtaining the sea trial data of the manned submersible through a sensor, and processing the data into training data for training the neural network;

[0008] 3) training the neural network using the training data, and using the trained neural network as an observer to estimate the disturbance;

[0009] 4) obtaining the thruster thrust through the model predictive controller according to the estimated disturbance, and realizing the motion control of the manned submersible.

[0010] The dynamics model of the manned submersible in step 1) is:

[0011]

[0012]

[0013] Where: x1 represents the velocity of the manned submersible in the carrier coordinate system; x2 represents the position and attitude of the manned submersible in the geodetic coordinate system; in addition, J(x2), M, C(x1), D(x1), and G(x2) represent the rotation matrix, inertia matrix, Coriolis and centripetal matrix, damping matrix, and force and torque caused by gravity, respectively; F represents the thrust of the manned submersible; d represents disturbance.

[0014] F = Bτ, where B represents the thrust configuration matrix and τ represents the thrust generated by the thruster.

[0015] Step 2) Acquire sea trial data of the manned submersible through sensors and process it into training data for training the neural network, including the following steps:

[0016] A1) Calculate the thrust τ:

[0017] τ=ρn 2 D 4 K T

[0018] Where ρ represents seawater density, n represents propeller speed, D represents propeller diameter, and K... T Indicates the thrust coefficient;

[0019] A2) Perform interpolation on x2 measured by the sensor;

[0020] A3) Calculate the interference d:

[0021]

[0022] A4) Perform linear normalization on τ, x2 and d respectively, and then fuse the normalized τ and x2 as the input of the neural network, and the normalized d as the output of the neural network.

[0023] The specific steps for constructing the model prediction controller are as follows:

[0024] B1) Establish a state-space model for the manned submersible;

[0025] The dynamic model of a manned submersible is represented as follows:

[0026]

[0027] Where, x = [x1x2] T For the manned submersible, in the known reference trajectory (x r ,τ r ,dr Perform a Taylor expansion at () and discretize it to obtain the state-space expression of the manned submersible:

[0028] Δx(k+1)=AΔx(k)+B1Δτ(k)+B d Δd(k)

[0029] y(k)=CΔx(k)+y(k-1)

[0030] Δx(k+1) = x(k+1) - x(k)

[0031] Δx(k)=x(k)-x(k-1)≈T s f(x(k-1),τ(k-1),d(k-1))

[0032] Δτ(k)=τ(k)-τ(k-1)

[0033] Δd(k)=d(k)-d(k-1)

[0034]

[0035]

[0036]

[0037] Among them, A, B1, B d C represents the state matrix, input matrix, disturbance matrix, and output matrix, respectively; k represents the discrete time step; I represents the identity matrix; T S It represents the sampling time; Δ represents the amount of change;

[0038] B2) Obtain the output prediction equation based on the output matrix C;

[0039] In model predictive control, the control time domain is denoted as Nc, the prediction time domain as Np, and the output prediction equation is:

[0040] Y p (k+1|k)=S x Δx(k)+S y y(k)+S u ΔU(k)+S d Δd(k)

[0041] in,

[0042]

[0043]

[0044]

[0045] Among them, S x S y S u S d The intermediate matrix used to derive the prediction equation; i is the intermediate variable used in the summation formula;

[0046] B3) Set the optimization objective for the predictive controller;

[0047] J = ||Q y (Y p (k+1|k)-R(k+1))|| 2 +||Q u ΔU(k)|| 2

[0048] Where J represents the cost function, Y p For system output, R is the desired pose of the manned submersible, and Q is... y and Q u Here is the weight coefficient matrix; ΔU(k) ​​represents the input sequence;

[0049] According to the following constraints:

[0050]

[0051]

[0052] Where, τ min and τ max Δτ represents the thrust constraint that the thruster can provide. min and Δτ max Constraints representing the rate of change of the thruster;

[0053] The above optimization can be transformed into a quadratic programming problem as follows:

[0054]

[0055] st.C1ΔU(k)≤b

[0056] Among them, H, f1 T C1 and b are the transformed parameter matrices;

[0057] Take the first term Δτ(k|k) of the obtained ΔU, and apply τ=τ(k-1)+Δτ(k) to the thruster of the manned submersible to realize the model predictive control of the manned submersible.

[0058] The present invention has the following beneficial effects and advantages:

[0059] 1. This invention uses an LSTM neural network to estimate disturbances, which provides more accurate estimation of disturbances such as ocean current disturbances and model parameter perturbations.

[0060] 2. The model predictive controller based on disturbance design in this invention has higher accuracy than traditional PID control.

[0061] 3. This paper proposes an optimization target design that can effectively solve the overdrive problem without requiring an additional control allocation module. This not only simplifies the control system process but also addresses the thrust saturation problem of manned submersible thrusters, thereby improving control performance. Attached Figure Description

[0062] Figure 1 The control block diagram of the manned submersible based on the neural network interference observer of the present invention;

[0063] Figure 2 The geodetic coordinate system and carrier coordinate system of the manned submersible in the embodiments of the present invention;

[0064] Figure 3 Side view of the thruster distribution of a manned submersible in an embodiment of the present invention;

[0065] Figure 4 A bottom view of the thruster distribution of a manned submersible in an embodiment of the present invention;

[0066] Figure 5 Rear view of the thruster distribution of a manned submersible in an embodiment of the present invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0068] like Figure 1 As shown, this invention discloses a manned submersible control method based on a neural network interference observer, comprising the following steps:

[0069] 1) Construct a kinematic and dynamic model of a manned submersible;

[0070] 2) Process the sea trial data of the manned submersible, calculate the input and output data, and normalize the input and output data to generate training data;

[0071] 3) Train an LSTM neural network using the training data, and use the trained neural network as an observer to estimate the disturbance;

[0072] 4) Construct a model predictive controller to realize motion control of the manned submersible.

[0073] Establish Figure 2 The two coordinate systems shown, and the dynamic and kinematic models of the manned submersible in step 1) are as follows:

[0074]

[0075]

[0076] For manned submersibles, pitch and roll are primarily adjusted via the mercury system and ballast water system. Furthermore, the pitch and roll angles remain almost constant during submersible motion; therefore, the motion of these two degrees of freedom (pitch and roll) can be neglected. Thus, in the above equation: x1=[ξ η ζ ψ] T ξ represents the manned submersible's northward position, η represents its eastward position, ζ represents its diving depth, and ψ represents its heading angle; x2 = [uvwr] T Let u represent the forward and backward velocities of the manned submersible, v represent the lateral velocity of the manned submersible, and w represent the vertical velocity of the manned submersible. Furthermore, J(x2), M, C(x1), D(x1), and G(x2) represent the rotation matrix, inertia matrix, Coriolis and centripetal matrix, damping matrix, and the forces and torques caused by gravity, respectively. F represents the thrust from the thrusters acting on the manned submersible, and d represents the disturbance.

[0077] Furthermore, F = Bτ, where B represents the thrust configuration matrix and τ represents the thrust generated by the thruster.

[0078] Step 2) of processing manned submersible sea trial data includes the following steps:

[0079] Calculate the thrust τ:

[0080] τ=ρn 2 D 4 K T (3)

[0081] Where ρ represents seawater density, n represents propeller speed, D represents propeller diameter, and K... T This represents the thrust coefficient.

[0082] Since the sampling times of the sensors are different, the interpolation function interp1 in MATLAB is used to interpolate the x2 variables measured by the sensors.

[0083] Calculate the interference d:

[0084]

[0085] We perform linear normalization on τ, x2, and d respectively, and then fuse the normalized τ and x2 as the input of the neural network, and the normalized d as the output of the neural network.

[0086] Step 3) The specific steps for training the LSTM neural network are as follows:

[0087] Build an LSTM neural network model based on the Tensorflow architecture;

[0088] The input and output dimensions of the LSTM neural network are set. In this invention, the input dimension is 1×8 and the output dimension is 1×4.

[0089] The number of nodes in the LSTM neural network is set. This invention uses 100 LSTM units;

[0090] Configure the optimizer and loss function. This method uses the Adam algorithm optimizer and selects the mean squared error function (MSE) as the loss function.

[0091] Set the number of iterations and batch size; in this invention, the number of iterations is set to 10 and the batch size is set to 36.

[0092] The trained LSTM neural network can accurately predict interference d as an interference observer, and the controller of the manned submersible can be designed based on the interference information.

[0093] Step 4) The specific steps for constructing the model predictive controller are as follows:

[0094] Establish a state-space model for a manned submersible:

[0095] Due to the nonlinearity of the motion model of the manned submersible, it can be written as:

[0096]

[0097] Where, x = [x1x2] T To avoid complex nonlinear optimization solutions, the system needs to be linearized, and the equation needs to be applied to the known reference trajectory (x). r ,τ r ,d r Performing a Taylor expansion at () and retaining the first-order terms while ignoring the higher-order terms, we obtain:

[0098]

[0099] Discretizing the above equation using the Euler method yields:

[0100]

[0101] Where T S It is the sampling time, if (x r ,τ r ,d r If we take the trajectory (x(k-1), τ(k-1), d(k-1)) at time k-1, then the above formula can be written as:

[0102]

[0103] Simplifying, we get:

[0104] Δx(k+1)=AΔx(k)+B1Δτ(k)+B d Δd(k) (9)

[0105] in:

[0106]

[0107] Therefore, the state-space expression of a manned submersible can be written as:

[0108]

[0109] The state matrix A can be represented as:

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] Where: m represents the mass of the manned submersible, I ZZ Let represent the moment of inertia, ρ represent the density of seawater, and L represent the length of the manned submersible. In addition to the parameters defined above, the remaining parameters... X uu ,X vr ,X vv ,X ww ,X rr ,X vr , Y r ,Y0,Y v ,Y v|r| ,Y vw ,Y wr ,Y v|v| , Z0,Z w Z |w| Z vr Z w|w| Z rr , N r ,N0,Nv N v|r| N v|v| N vw N wr N r|r| This represents the hydrodynamic coefficient.

[0119] according to Figures 3-5 As shown, Tsru, Tsrd, Tmrb, and Tmrf represent the right stern horizontal thruster, right stern vector thruster, right stern vertical thruster, and right bow vertical thruster, respectively; Th represents the bow thruster; and xsd represents the distance from the midpoint of the stern vector thruster to the yz plane. Figure 4 In the diagram, Tsld and Tmlf represent the left stern vector thruster and left bow vertical thruster, respectively; ysd represents the distance from the midpoint of the stern vector thruster to the xz plane; xh represents the distance from the midpoint of the bow thruster to the yz plane; xmb and ymb represent the distances from the midpoint of the stern vertical thruster to the yz and xz planes, respectively; xmf and ymf ​​represent the distances from the midpoint of the bow vertical thruster to the yz and xz planes, respectively; α represents the angle between the axis of the stern vector thruster and the x-axis.

[0120] Figure 5 In the diagram, Tslu and Tsld represent the left stern horizontal thruster and the left stern vector thruster, respectively, and zsu represents the distance from the midpoint of the stern horizontal thruster to the xy plane.

[0121] The input matrix B1 is represented as:

[0122]

[0123] Interference matrix B d Represented as:

[0124]

[0125] The output matrix C is represented as:

[0126]

[0127] Derive the output prediction equation;

[0128] In model predictive control, the control time domain is denoted as Nc, and the prediction time domain is denoted as Np. To derive the prediction equation, the following assumptions are made:

[0129] Assumption 1: Outside the control time domain, the control quantity remains unchanged;

[0130] Assumption 2: The observation disturbance remains unchanged after time k;

[0131] The output prediction equation is:

[0132] Y p (k+1|k)=Sx Δx(k)+S y y(k)+S u ΔU(k)+S d Δd(k) (23)

[0133] in,

[0134]

[0135]

[0136]

[0137] Give the optimization objective of the predictive controller;

[0138] J = ||Q y (Y p (k+1|k)-R(k+1))|| 2 +||Q u ΔU(k)|| 2 (27)

[0139] Where J represents the cost function, Y p For system output, R is the desired pose of the manned submersible, and Q is... y and Q u This is the weighting coefficient matrix. The first term on the right-hand side of the equation aims to make the control output as close as possible to the desired output; the second term aims to minimize the thrust variable.

[0140] The thrust that a manned submersible's thrusters can provide has an upper limit, and in order to protect the motors, there is also an upper limit to the changes between two adjacent control variables. Therefore, the following constraints are imposed:

[0141]

[0142] Where τ min and τ max Δτ represents the thrust constraint that the thruster can provide. min and Δτ max This represents the constraint on the rate of change of the thruster.

[0143] The above optimization can be transformed into the following quadratic programming problem:

[0144]

[0145] in,

[0146]

[0147]

[0148]

[0149]

[0150] Among them, U max U min Let represent the constraint matrix and nu represent the number of input variables. Solve the above optimization problem, take the first term Δτ(k|k) of the obtained ΔU, and apply τ=τ(k-1)+Δτ(k) to the system to realize the model predictive control of the manned submersible.

Claims

1. A manned submersible control method based on a neural network interference observer, characterized in that, Includes the following steps: 1) Construct a model predictive controller based on the dynamic model of the manned submersible; 2) Acquire sea trial data of the manned submersible through sensors and process it into training data for training the neural network; 3) Train the neural network using the training data, and use the trained neural network as an observer to estimate the disturbance; 4) Based on the estimated disturbance, the thrust is obtained through the model predictive controller to realize the motion control of the manned submersible; The dynamic model of the manned submersible mentioned in step 1) is as follows: ； ； in: This represents the velocity of the manned submersible in the carrier coordinate system; This indicates the position and attitude of the manned submersible in the geodetic coordinate system; furthermore... J(x 2 M, C(x) 1 ), D(x 1 ) and G(x 2 ) These represent the rotation matrix, inertia matrix, Coriolis and centripetal matrix, damping matrix, and the force and torque caused by gravity, respectively. F This indicates the thrust experienced by the manned submersible. d Indicates interference; ,in, B Represents the thrust configuration matrix. This indicates the thrust generated by the propeller; Step 2) Acquire sea trial data of the manned submersible through sensors and process it into training data for training the neural network, including the following steps: A1) Calculate thrust : ； in, ρ Indicates the density of seawater. n Indicates the propeller speed. D Indicates the propeller diameter. K T Indicates the thrust coefficient; A2) Perform interpolation on x2 measured by the sensor; A3) Calculate interference d : ； A4) respectively for ，x 2 and d Perform linear normalization, then fuse the normalized results. and x 2 As input to the neural network, the normalized d As the output of the neural network.

2. The manned submersible control method based on a neural network interference observer according to claim 1, characterized in that, The specific steps for constructing the model prediction controller are as follows: B1) Establish a state-space model of the manned submersible; The dynamic model of a manned submersible is represented as follows: ； in, For manned submersibles, in a known reference trajectory Perform a Taylor expansion and discretize the expression to obtain the state-space representation of the manned submersible: ； ； Among them, A, B1, B d C represents the state matrix, input matrix, disturbance matrix, and output matrix, respectively; k represents the discrete time step; I represents the identity matrix; T S It is the sampling time; Indicates the amount of change; B2) Obtain the output prediction equation based on the output matrix C; In model predictive control, the control time domain is denoted as... Nc The prediction time domain is denoted as Np The output prediction equation is: ； in, ； ； ； in, , , , The intermediate matrix used to derive the prediction equation; i is the intermediate variable used in the summation formula; B3) Set the optimization objective for the predictive controller; ； in, J Represents the cost function, Y p For system output, R The desired attitude of the manned submersible. Q y and Q u This is the weight coefficient matrix; Represents the input sequence; According to the following constraints: ； in, τ min and τ max This indicates the thrust constraint that the thruster can provide. τ min and τ max Constraints representing the rate of change of the thruster; The above optimization can be transformed into a quadratic programming problem as follows: ； in, H, f 1 T , C 1 and b It is the transformed parameter matrix; Take the solution U The first item ,Will The thrusters acting on the manned submersible enable model predictive control of the manned submersible.

Images