Missile guidance and control integrated method based on neural network mechanism modeling

Abstract

The application discloses a missile guidance control integration method based on a neural network mechanism modeling, which comprises the following steps: in a continuous time system, a vertical plane missile guidance control integration model is established; according to the dynamic equation of the vertical plane missile guidance control integration, a state transition equation in a discrete time system is established; a data set for fitting the dynamic model of a neural network is established, and a neural network parameter model obtained through neural network regression calculation is used as a non-mechanism fitting model of a missile control system; and according to the neural network parameter model, a guidance control integration mechanism model used in a control cycle is obtained. The method uses the input and output data of a control process to perform self-modeling of a time-varying control system, meanwhile, the model is equivalent to a mechanism method, and the modeling result can be used for the control of a discrete time system, and the modeling principle has interpretability and reliability.

Description

Technical Field

[0001] This invention relates to missile guidance and control technology, specifically to an integrated missile guidance and control method based on neural network-based mechanism modeling. Background Technology

[0002] The missile-target relationship changes drastically at the terminal stage of interception, especially when the target is maneuverable, leading to significant changes in missile aerodynamics. This makes it difficult for the control system to meet the rapid response requirements of the guidance system. Designing the guidance and control systems simultaneously can minimize the coupling between them, improving the missile's agile response and reducing the miss distance. However, integrated guidance and control design faces uncertainties in aerodynamic and target characteristics, introducing complex nonlinear effects to the control system. Furthermore, the rapid changes in system state at the terminal stage and the time-varying and uncertain system model parameters negatively impact the control system design.

[0003] The difficulty in controlling time-varying complex control systems mainly stems from the challenge of tracking the control system model. General mechanistic modeling methods, based on physical laws, can only quantify the controlled object and environment to a limited extent. Excessive pursuit of modeling accuracy significantly increases model complexity, posing difficulties for control law design. In practical engineering control, discrete-time systems are typically used for object control. If the current system model can be tracked as closely as possible, feedforward compensation methods can be used to achieve high-precision control of the system, while also providing good robustness to external disturbances. Summary of the Invention

[0004] The technical problem to be solved by this invention is: the purpose of this invention is to provide a method for online dynamic self-modeling of a time-varying system such as missile guidance and control integration, to achieve isomorphic expression between the fitting model and the mechanism model, and to ensure that the guidance and control process has sufficient credibility and interpretability.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an integrated missile guidance and control method based on neural network mechanism modeling, comprising:

[0006] In a continuous-time system, an integrated model of vertical plane missile guidance and control is established under the vertical plane missile-target line-of-sight coordinate system.

[0007] Based on the dynamic equations of the vertical plane missile guidance and control integration, the state transition equations of the discrete-time system are established.

[0008] A dataset for fitting a dynamic model using a neural network is established. The neural network parameter model obtained through neural network regression calculation is used as a non-mechanistic fitting model for the missile control system.

[0009] Based on the neural network parameter model, a guidance and control integrated mechanism model used within the control cycle is obtained.

[0010] The missile guidance and control integration method based on neural network mechanism modeling further includes: using the guidance and control integration mechanism model to design the missile's control loop in discrete space, forming a missile guidance and control closed loop.

[0011] Furthermore, the vertical plane missile guidance and control integrated model in the vertical plane missile-eye line-of-sight coordinate system is as follows:

[0012]

[0013] Among them, the rate of change of line of sight tilt angle ε is the line-of-sight tilt angle; α is the missile angle of attack. Let u be the missile's elevation angle, and u = δ z , c 12 =-y α / R,c 22 =-y α / V,c 32 =m α , V is the missile velocity, R = [R0 0] T For relative position, δ z For the missile's pitch control deflection angle, ω z y is the pitch rate; α m is the lift coefficient relative to the angle of attack. α , and These are the moment coefficients relative to the angle of attack, pitch rate, and pitch deflection, respectively, Δ α and Δ ω All are unmodeled aerodynamic uncertainties, Δ ε Uncertainty caused by target maneuvering.

[0014] Furthermore, establishing the state transition equations in the discrete-time system includes:

[0015] By using the state feedback linearization method, the state and input are transformed to obtain a new controllable linear system. The new controllable linear system is then solved, and after discretization, the state transition equations are obtained.

[0016] X k+1 =A k+1,k X k +B k+1,k U k +H k+1,k D k ,

[0017] Among them, Xk and X k+1 The system states at time k and time k+1 are respectively; A k+1,k U is the state transition matrix from time k to time k+1; k B is the control variable at time k; k+1,k D is the control coefficient matrix from time k to time k+1. k H is a measure of system uncertainty. k+1,k Uncertainty coefficient matrix.

[0018] Furthermore, the dataset used to establish the neural network fitting dynamic model, through neural network regression calculation, converges to obtain a neural network parameter model as a non-mechanistic fitting model for the missile control system, including:

[0019] Starting from the current time k, sample the current state quantity x. k And by tracing back n control cycles of the control process, the pre-control state data x is obtained. k-n ,x k-n+1 ,...,x k-1 Control data u k-n ,u k-n+1 ,...,u k-1 and the state variable data x after control k-n+1 ,x k-n+2 ,...,x k A dataset is formed to fit the dynamic model of the neural network. The neural network regression calculation is performed. The neural network structure is adjusted according to the specific physical process characteristics. The neural network parameter model obtained by calculation convergence is the non-mechanistic fitting model of the guidance and control system.

[0020] Furthermore, the neural network input-output relationship used in the neural network regression calculation, based on a four-layer neural network structure, is as follows:

[0021] O = f o (W 2o f h2 (W 12 f1(W i1 I+B h1 )+B h2 )+B o ),

[0022] Where I is the input vector and O is the output vector; f h1 (·), f h2 (·), f o (·) represent the activation functions of neurons in two hidden layers and one output layer, respectively, and the connection weight matrix between neurons in each layer is W. i1 W 12 W 2oThe bias vectors of each layer of neurons are B. h1 B h2 B o .

[0023] Furthermore, the integrated guidance and control mechanism model obtained based on the neural network parameter model during the control cycle includes:

[0024] S6.1, Linearize the neuron activation function:

[0025] By differentiating the neuron activation function at the neuron input value, the neuron activation function is linearly transformed into the form f(x) = s0 + s1x, where x is the input variable of the activation function, and s0 and s1 have corresponding values ​​depending on the type of activation function.

[0026] S6.2, reconstruct the input-output relationship of the neural network forward propagation using the linearized activation function f(x);

[0027] S6.3, comparing the neural network input-output relationship expression with the discrete state transition equation expression, we obtain:

[0028] X k+1 =A k+1,k X k +B k+1,k U k +H k+1,k D k ,

[0029] when O = X k+1 Then, the integrated guidance and control mechanism model equation is expressed as:

[0030]

[0031] in, To fit the obtained system state matrix, The control coefficient matrix obtained by fitting is... This is the model uncertainty compensation term.

[0032] Furthermore, the input-output relationship of the neural network forward propagation is reconstructed using the linearized activation function f(x) as follows:

[0033]

[0034] in,

[0035] D = S o,0 +S o,1 .*(W 2o (S 2,0+S 2,1 .*(W 12 (S 1,0 +S 1,1 .*B h1 )+B h2 ))+B o ),

[0036] S 1,0 S 2,0 S o,0 S is a vector consisting of the linearization coefficients s0 of the activation functions on the two hidden layers and one output layer, respectively; 1,1 S 2,1 S o,1 This is a vector consisting of the linearization coefficients s1 of the activation functions on the two hidden layers and one output layer, respectively; * indicates element-wise multiplication of the matrix.

[0037] A control system based on the missile guidance and control integration method modeled according to the aforementioned neural network mechanism includes:

[0038] The first module is used to establish an integrated model of vertical plane missile guidance and control in a vertical plane missile-eye line-of-sight coordinate system in a continuous time system.

[0039] The second module is used to establish the state transition equations in the discrete-time system based on the dynamic equations of the vertical plane missile guidance and control integration.

[0040] The third module is used to establish a dataset for fitting the dynamic model of the neural network. The neural network parameter model obtained by neural network regression calculation is used as a non-mechanistic fitting model of the missile control system.

[0041] The fourth module is used to obtain the integrated guidance and control mechanism model used within the control cycle based on the neural network parameter model.

[0042] Furthermore, the fourth module also includes: using an integrated guidance and control mechanism model to design a control loop for the missile in discrete space, forming a closed-loop guidance and control loop for the missile.

[0043] The advantages of this invention compared to the prior art are:

[0044] (1) The present invention has high state tracking accuracy, and can fit and model and correct system parameters and external disturbances online, and has strong robustness.

[0045] This invention considers the continuous state tracking or change of the guidance and control system under control action. It uses the control action and system state as learning samples, and performs neural network fitting modeling of the guidance and control system through real-time online learning and training. The parameters of the neural network are used to invert the mechanism model of the controlled object and dynamically identify and track the parameters. During the control process, the overall control strategy remains unchanged, and the system model is updated in each control cycle. Aerodynamic characteristics and target state changes that are difficult to model are modeled in real time, achieving adaptive system modeling and ensuring control accuracy.

[0046] (2) The neural network fitting mechanism model of the present invention is isomorphic to the mechanism model in form, and the control process is interpretable, which improves the reliability of the control.

[0047] This invention establishes a mathematical isomorphism between the parameterized model of the neural network and the discrete-time state transition equation by fitting the integrated guidance and control process to a neural network. This overcomes the limitation of general neural networks, which can only optimize some parameters without physical interpretation when applied to control processes. This method utilizes equivalent mathematical expressions to obtain the evolution of the dynamic system itself and the state transition laws under controlled action, thus completing the modeling of the mechanism / quasi-mechanism model of the controlled state transition process.

[0048] (3) The method of the present invention is simple, effective and highly adaptable.

[0049] The neural network-based mechanism self-modeling method of the present invention is determined by the system control input and output, and can track changes in system parameters within a short number of control cycles. For the same guidance and control integrated system, it can adapt to large changes in system parameters without the need for additional structural adjustments. Attached Figure Description

[0050] Figure 1 A schematic diagram of a neural network-based mechanism modeling method for integrated guidance and control;

[0051] Figure 2 Diagram showing the relationship between the projectile and target coordinate systems;

[0052] Figure 3 This is a schematic diagram of neural network connections. Detailed Implementation

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] like Figure 1 As shown, this embodiment provides a neural network-based mechanism modeling method for integrated missile guidance and control, including:

[0055] I. In a continuous-time system, establish the dynamic equations for the integration of guidance and control.

[0056] In three-dimensional space, construct a motion model of the missile and the interceptor target, such as... Figure 2 As shown, o-xyz and o-x1y1z1 represent the reference inertial coordinate system and the line-of-sight coordinate system, respectively, and a r V is the relative acceleration between the projectile and the target. r R is the relative velocity, R is the relative position, and Ω is the angular velocity of coordinate system o-x1y1z1 relative to o-xyz. × Let be the diagonal matrix spanned by vector Ω.

[0057] According to the projectile-eye kinematics relationship,

[0058]

[0059] In the line-of-sight coordinate system, R = [R 0 0] T Angular velocity Ω is

[0060]

[0061] Where ε represents the line-of-sight tilt angle and η represents the line-of-sight deflection angle, therefore the relative speed is...

[0062]

[0063] Among them, V rx V ry V rz Relative velocities V r The three-axis components of the coordinate system.

[0064] After rearranging equations (1), (2), and (3), the relative acceleration a is obtained. r

[0065]

[0066] Where a rx a ry a rz The relative accelerations a and a are respectively r The three-axis components in the line-of-sight coordinate system.

[0067] Let a represent the accelerations of the missile and the target in the line-of-sight coordinate system. m =[a mx1 a my1 a mz1 ] T and a t =[a tx1 a ty1 a tz1 ] T a mx1 ,a my1 ,amz1 a m The three-axis components, a tx1 ,a ty1 ,a tz1 a t The three-axis components, at which point the projectile-object relative dynamic model can be expressed as:

[0068]

[0069] Because the integrated guidance and control design considers both target maneuvering and missile aerodynamic models, the model exhibits complex nonlinearities. Guidance and control in the vertical and horizontal planes can be decomposed. This invention focuses on the pitch motion in the vertical plane, and the relative motion equations between the missile and target are as follows:

[0070]

[0071] Among them, the rate of change of line of sight tilt angle a θ =a mθ θ represents the direction of the missile's velocity.

[0072] The equation of motion for missile pitch is:

[0073]

[0074] Where m is the missile mass, V is the missile velocity, Y is the lift, and Y' is the missile angle of attack α and the missile pitch deflection angle δ. z The function is given by Δ, where Δ represents the missile thrust and gravitational acceleration on the missile, P is the missile thrust, and g is the gravitational acceleration. Let ω be the missile's elevation angle. z J is the pitch rate. z Let M be the moment of inertia of the missile in the pitch direction, and M be the pitching moment.

[0075] Since the pitch rudder has a relatively small effect on lift, a linearized expression can be obtained.

[0076]

[0077] Among them, y α Let Δ be the lift coefficient for the angle of attack. Y For the unmodeled effects of lift, m α , and These are the moment coefficients for angle of attack, pitch rate, and pitch deflection, respectively, Δ M This represents the unmodeled effect of pitch moment.

[0078] By combining the projectile motion equation (6) and the missile dynamics equation (7), an integrated guidance and control model can be obtained.

[0079]

[0080] Among them, u=δ z , c 12 =-y α / R,c 22 =-y α / V,c 32 =m α , u is a variable, c 11 ,c 12 ,c 22 ,c 32 ,c 33 Both 'b' and 'b' are coefficients.

[0081] The model uncertainty term is

[0082]

[0083] Where, Δ α and Δ ω All are unmodeled aerodynamic uncertainties, Δ ε Uncertainty caused by target maneuvering.

[0084] II. Establish the state transition equations for the discrete-time system.

[0085] The guidance and control integrated equations are transformed into state-space form, and then the state-space equations are solved.

[0086] Based on the state feedback linearization method, the following state transformation is performed on equation (8).

[0087]

[0088] Input transformation

[0089]

[0090] Among them, the angle parameter Ψ(ω ε ,α,ω z ) for ω ε ,α,ω z The function; to obtain a new controllable linear system.

[0091]

[0092] Where x1, x2, and x3 are state variables, and ν is a control variable;

[0093] Solve this linear system and discretize it to obtain the state transition equation.

[0094] X k+1 =A k+1,k X k +B k+1,k U k +H k+1,k D k (12)

[0095] in

[0096]

[0097] Where τ is the control period.

[0098] Among them, X k and X k+1 The system states at time k and time k+1 are respectively; A k+1,k U is the state transition matrix from time k to time k+1; k B is the control variable at time k; k+1,k D is the control coefficient matrix from time k to time k+1. k H is a measure of system uncertainty. k+1,k Uncertainty coefficient matrix.

[0099] Third, control the period at the current time k and use a neural network to fit the integrated parameterized model of guidance and control.

[0100] For the current time k, sample the current state variable x. k And by tracing back n control cycles of the control process, the pre-control state data x is obtained. k-n ,x k-n+1 ,...,x k-1 Control data u k-n ,u k-n+1 ,...,u k-1 and the state variable data x after control k-n+1 ,x k-n+2 ,...,x k A dataset is formed to fit the dynamic model of the neural network, and neural network regression calculation is performed; k is a positive integer.

[0101] The neural network is designed as a feedforward fully connected neural network. For example... Figure 3 As shown, without loss of generality, a typical four-layer neural network structure is adopted, namely one input layer, two hidden layers, and one output layer. The input vector I consists of n inputs, and the output vector O consists of r outputs. Hidden layers 1 and 2 have p and q neurons respectively, where n is a positive integer.

[0102] The inputs to each layer of neurons are as follows:

[0103] H 1,in =[h 11,in ,h12,in ,…,h 1p,in ] T H 2,in =[h 21,in ,h 22,in ,…,h 2q,in ] T ,

[0104] Y in =[y 1,in ,y 2,in ,…,y r,in ] T ;

[0105] The outputs of neurons in each layer are as follows:

[0106] H 1,out =[h 11,out ,h 12,out ,…,h 1p,out ] T H 2,out =[h 21,out ,h 22,out ,…,h 2q,out ] T ,

[0107] Y = [y 1,out ,y 2,out ,…,y r,out ] T ;

[0108] The connection weights between neurons in each layer are respectively

[0109] w i1,ij (i=1,2,…,n; j=1,2,…,p), w 12,jk (j=1,2,…,p; k=1,2,…,q),

[0110] w 2o,kl (k=1,2,…,q; l=1,2,…,r);

[0111] The bias of neurons in each layer is as follows

[0112] B h1 =[b h1,1 ,b h1,2 ,…b h1,p ] T B h2 =[b h2,1 ,b h2,2 ,…b h2,q ] T B o =[b o,1 ,b o,2 ,…b o,r] T ;

[0113] The activation functions of the hidden layer and the output layer neurons are f, respectively. h1 (·), f h2 (·), f o (·); The connection weight matrix between neurons in each layer is W. i1 W 12 W 2o The bias vectors of each layer of neurons are B. h1 B h2 B o .

[0114] The input and output of a single neuron in a neural network are as follows:

[0115]

[0116] h out,j =f(h in,j ), j = 1, 2, ..., p

[0117] For hidden layer 1, we have

[0118] H 1,in =[h 11,in h 12,in …h 1p,in ] T =W i1 X+B h1

[0119] H 1,out =[h 11,out h 12,out …h 1p,out ] T =f h1 (H 1,in )

[0120] For hidden layer 2, we have

[0121] H 2,in =[h 21,in h 22,in …h 2q,in ] T =W h2 H 1,out +B h2

[0122] H 2,out =[h 21,out h 22,out …h 2q,out ] T =f h2 (H 2,in )

[0123] The output layer uses pureline neurons. The output expression of each neuron is as follows:

[0124]

[0125] Then there is

[0126] O = [y1 y2…y r ] T =f o (O in ) = O in =W 2o H 2,out +B o

[0127] From the forward propagation algorithm, it is easy to see that the input-output relationship of the entire neural network is as follows:

[0128] O = f o (W 2o *f h2 (W 12 *f1(W i1 *I+B h1 )+B h2 )+B o (13)

[0129] The structure of the neural network can be adjusted according to specific circumstances. Within this control cycle, a dataset is composed of historical state variables and control variables. The neural network regression calculation is used to obtain a parameterized neural network model. Since the network model is completely fitted by the input and output data, it is a non-mechanistic modeling process, and thus becomes a non-mechanistic model of integrated guidance and control.

[0130] Fourth, the parameterized non-mechanistic model of the neural network is transformed into a discrete state transition mechanistic model through mathematical isomorphism and equivalence.

[0131] To achieve an isomorphic equivalent transformation from a neural network to a mechanistic model, the neuron activation function is differentiated at the neuron's input value, and linearly transformed into the following form.

[0132] f(x)=s0+s1x (14)

[0133] Where f(x) is the linear excitation function, x is the input variable of the excitation function, and s0 and s1 are linearization coefficients.

[0134] 1) If the neuron activation function is the Sigmoid function:

[0135] If we transform it into the form f(x) = s0 + s1x through equivalent dynamic linearization, then we have

[0136]

[0137]

[0138] 2) If the neuron activation function is the tanh function:

[0139] If we also perform equivalent dynamic linearization on it, we can transform it into the form f(x) = s0 + s1x.

[0140] Then there is

[0141]

[0142]

[0143] 3) If the neuron activation function is the Purelin function: p(x) = x

[0144] It is already a linear function, corresponding to the linearization coefficients.

[0145] s0 = 0, s1 = 1

[0146] Similarly, the dynamic linearization coefficients of activation functions for other types of neurons can be obtained.

[0147] Substitute equation (14) into the forward propagation process of the neural network.

[0148] Hidden layer 1 output (p×1 dimension):

[0149] H 1,out =f h1 (H 1,in ) = S 1,0 +S 1,1 .*H 1,in =S 1,0 +S 1,1 .*(W i1 X+B h1 )

[0150] Hidden layer 2 output (q×1 dimension):

[0151] H 2,out =f h2 (H 2,in ) = S 2,0 +S 2,1 .*H 2,in =S 2,0 +S 2,1 .*(W 12 H 2,in +B h2 )

[0152] Output layer output (r×1 dimension):

[0153] O = f o (W 2o H 2,out +B o ) = S o,1 .*W 2o S 2,1 .*W 12 S 1,1 .*W i1 X+S o,0 +S o,1 .*(W 2o (S 2,0 +S 2,1 .*(W 12 (S 1,0 +S 1,1 .*B h1 )+B h2 ))+B o )

[0154] Among them, S 1,0 S 2,0 S o,0 S is a vector consisting of the linearized coefficients s0 of the activation functions on the two hidden layers and one output layer, respectively. 1,1 S 2,1 S o,1 The vector consists of the linearization coefficients s1 of the activation functions on the two hidden layers and one output layer, respectively. * indicates element-wise multiplication of the matrix.

[0155] S 1,0 =[a 10,1 a 10,2 …a 10,p ] T S 1,1 =[a 11,1 a 11,2 …a 11,p ] T

[0156] S 2,0 =[a 20,1 a 20,2 …a 20,q ] T S 2,1 =[a 21,1 a 21,2 …a 21,q ] T

[0157] S o,0 =[a o0,1 a o0,2 …a o0,r ]T S o,1 =[a o1,1 a o1,2 …a o1,r ] T

[0158] After a transformation, the neural network output equation (13) can be expressed as:

[0159]

[0160] in

[0161]

[0162] D = S o,0 +S o,1 .*(W 2o (S 2,0 +S 2,1 .*(W 12 (S 1,0 +S 1,1 .*B h1 )+B h2 ))+B o )

[0163] Compare equation (15) with the mechanistic model equation (12) in the form of discrete state transition equations.

[0164] X k+1 =A k+1,k X k +B k+1,k U k +H k+1,k D k

[0165] when O = X k+1 Then the mechanism model equation can be expressed as

[0166]

[0167] in, To fit the obtained system state matrix, This is the control coefficient matrix obtained through fitting. This is the model uncertainty compensation term.

[0168] In step four, the non-mechanistic model of the neural network is transformed into an equivalent mathematical expression. By local and dynamic linearization of the neuron activation function, the neural network parameters are transformed into discrete equations in the state transition form that are consistent with the mechanistic model. That is, the guidance and control system self-modeling within the current control cycle is completed by fitting the neural network.

[0169] Fifth, by using an integrated guidance and control mechanism model to replace modeling based on physical principles, the missile's control loop design in discrete space is formed, creating a self-generating and self-updating control closed loop.

[0170] A control system based on the above-mentioned missile guidance and control integration method modeled on neural network mechanisms includes:

[0171] The first module is used to establish an integrated model of vertical plane missile guidance and control in a vertical plane missile-eye line-of-sight coordinate system in a continuous time system.

[0172] The second module is used to establish the state transition equations in the discrete-time system based on the dynamic equations of the vertical plane missile guidance and control integration.

[0173] The third module is used to establish a dataset for fitting the dynamic model of the neural network. The neural network parameter model obtained by neural network regression calculation is used as a non-mechanistic fitting model of the missile control system.

[0174] The fourth module is used to obtain the integrated guidance and control mechanism model used within the control cycle based on the neural network parameter model.

[0175] Furthermore, the fourth module also includes: using an integrated guidance and control mechanism model to design a control loop for the missile in discrete space, forming a closed-loop guidance and control loop for the missile.

[0176] Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by using the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A missile guidance and control integrated method based on neural network mechanism modeling, characterized in that, include: In a continuous-time system, an integrated model of vertical plane missile guidance and control is established under the vertical plane missile-target line-of-sight coordinate system. Based on the dynamic equations of the vertical plane missile guidance and control integration, the state transition equations of the discrete-time system are established. A dataset for fitting a dynamic model using a neural network is established. The neural network parameter model obtained through neural network regression calculation is used as a non-mechanistic fitting model for the missile control system. Based on the neural network parameter model, a guidance and control integrated mechanism model used within the control cycle is obtained, including: S6.1, Linearize the neuron activation function: By differentiating the neuron activation function at the neuron's input value, the neuron activation function is linearly transformed into... In the form of, x Input variables for the activation function, s 0 and s 1. Different types of excitation functions have corresponding values; S6.2, using a linearized excitation function Reconstruct the input-output relationship of the forward propagation of the neural network; S6.3, comparing the neural network input-output relationship expression with the discrete state transition equation expression, we obtain: ， in, and They are respectively k Time and k The system state at time +1; for k Time's up k The state transition matrix at time +1; for k Control the quantity at all times; for k Time's up k The control action coefficient matrix at time +1 As a measure of system uncertainty, Uncertainty coefficient matrix; when , Then, the integrated guidance and control mechanism model equation is expressed as: ， in, To fit the obtained system state matrix, The control coefficient matrix obtained by fitting is... This is the model uncertainty compensation term.

2. The missile guidance and control integrated method based on neural network mechanism modeling according to claim 1, characterized in that, Also includes: By utilizing an integrated guidance and control mechanism model, a control loop for the missile in discrete space is designed to form a closed-loop guidance and control loop for the missile.

3. The missile guidance and control integrated method based on neural network mechanism modeling according to claim 1, characterized in that: The vertical plane missile guidance and control integrated model in the vertical plane missile-eye line-of-sight coordinate system is as follows: ； Among them, the rate of change of line of sight tilt angle , Angle of view; α For the missile angle of attack, The missile's elevation angle, , V For missile speed, Relative position δ z For the missile's pitch control deflection angle, ω z The pitch rate; The lift coefficient is relative to the angle of attack. , and These are the moment coefficients relative to the angle of attack, pitch rate, and pitch deflection, respectively. and All of these are unmodeled aerodynamic uncertainties. Uncertainty caused by target maneuvering.

4. The missile guidance and control integrated method based on neural network mechanism modeling according to claim 1, characterized in that: The establishment of the state transition equations in the discrete-time system includes: By using the state feedback linearization method, the state and input are transformed to obtain a new controllable linear system. The new controllable linear system is then solved, and after discretization, the state transition equations are obtained. ， in, and They are respectively k Time and k The system state at time +1; for k Time's up k The state transition matrix at time +1; for k Control the quantity at all times; for k Time's up k The control action coefficient matrix at time +1 As a measure of system uncertainty, Uncertainty coefficient matrix.

5. The missile guidance and control integrated method based on neural network mechanism modeling according to claim 1, characterized in that: The dataset used to establish the neural network fitting dynamic model, through neural network regression calculation, converges to obtain a neural network parameter model as a non-mechanistic fitting model for the missile control system, including: From the present k At any given time, sample the current state quantity. And trace back the control process n One control cycle is used to obtain the pre-control state data. Control data and the state variable data after control A dataset is formed to fit the dynamic model of the neural network. The neural network regression calculation is performed. The neural network structure is adjusted according to the specific physical process characteristics. The neural network parameter model obtained by calculation convergence is the non-mechanistic fitting model of the guidance and control system.

6. The missile guidance and control integrated method based on neural network mechanism modeling according to claim 5, characterized in that: The neural network input-output relationship used in the neural network regression calculation, based on a four-layer neural network structure, is as follows: ， in, I For the input vector, O This is the output vector; , , The activation functions for neurons in two hidden layers and one output layer are given, and the connection weight matrix between neurons in each layer is given. , , The bias vectors of each layer of neurons are as follows: , , .

7. The missile guidance and control integrated method based on neural network mechanism modeling according to claim 6, characterized in that, The linearized excitation function The input-output relationship of the reconstructed neural network forward propagation is as follows: ， in, , ， , , The linearization coefficients of the activation functions on the two hidden layers and one output layer are respectively. s A vector consisting of 0s; , , The linearization coefficients of the activation functions on the two hidden layers and one output layer are respectively. s A vector consisting of 1s; This indicates the element-wise multiplication operation of matrices.

8. A control system for a missile guidance and control integrated method based on neural network mechanism modeling according to any one of claims 1 to 7, characterized in that, include: The first module is used to establish an integrated model of vertical plane missile guidance and control in a vertical plane missile-eye line-of-sight coordinate system in a continuous time system. The second module is used to establish the state transition equations in the discrete-time system based on the dynamic equations of the vertical plane missile guidance and control integration. The third module is used to establish a dataset for fitting the dynamic model of the neural network. The neural network parameter model obtained by neural network regression calculation is used as a non-mechanistic fitting model of the missile control system. The fourth module is used to obtain the integrated guidance and control mechanism model used within the control cycle based on the neural network parameter model.

9. The control system according to claim 8, characterized in that, The fourth module also includes: using an integrated guidance and control mechanism model to design a control loop for the missile in discrete space, forming a closed-loop missile guidance and control loop.

Images