Sparse data driven guidance method and device integrating physical information neural network

Abstract

The present invention discloses a sparse data driven guidance method and device integrating physical information neural network, which relates to the field of computer systems and is used to reduce the dependence on label data. The present invention converts the nonlinear multi-constraint dynamic trajectory optimization guidance problem into a convex trajectory optimization optimal control problem by convexification, and linearizes the nonlinear dynamic equation constraints into the form of an equation of the guidance control variable with respect to the state variable, and then introduces it into the physical information loss of the online guidance model. The label data set is obtained by solving the convex trajectory optimization optimal control problem to train the online guidance model, and the trained online guidance model is used for online real-time guidance. The present invention realizes the hybrid driven training of "data + model", effectively reduces the dependence of the deep neural network model on data, and improves the generalization performance of the model.

Description

Technical Field

[0001] The present invention relates to the field of computer systems, and in particular to a sparse data driven guidance method and device integrating physical information neural network. Background Art

[0002] Launch vehicles are the foundation of the development of aerospace technology, and their intelligent online guidance technology is a hot issue that urgently needs to be broken through. At present, the online guidance models and methods based on traditional deep learning neural network models are highly dependent on the quantity and quality of label data (i.e., sample data). For example, the "time-space constraint intelligent collaborative guidance method based on BP neural network" proposed in the Chinese patent document with publication number CN117824441A, its primary task is to use the single-missile guidance model to simulate the proportional guidance of a given single missile for maneuvering targets to construct a training set. Faced with the problems of emergency launch of launch vehicles and orbital rescue with very limited engineering data and complex models, the existing deep learning neural network online guidance methods are difficult to apply.

[0003] Embedded physical information neural network is a method that combines physical laws with deep learning. It has excellent performance in solving the forward and inverse problems of partial differential equations and ordinary differential equations. The online guidance of launch vehicles belongs to a type of nonlinear time-varying optimization problem with dynamic ordinary differential equation constraints. The dynamic differential equation describes the constraint relationship between the derivative of the rocket state variable with respect to time and the guidance control variable. The optimization solution process needs to determine the control variable parameters according to the degree of satisfaction of the state variable performance index. There are significant differences between the online guidance problem of launch vehicles and the forward and inverse problems of differential equations, so it is impossible to directly apply embedded physical information neural network to solve it. Summary of the invention

[0004] The purpose of the present invention is to provide a sparse data driven guidance method and device integrating physical information neural network to solve the online guidance problem of launch vehicles and reduce the dependence on label data in order to solve all or part of the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A sparse data driven guidance method integrating physical information neural network, comprising:

[0007] Construct the first trajectory planning guidance model with nonlinear dynamic differential equation constraints;

[0008] Convexifying the first trajectory planning and guidance model to obtain a second trajectory planning and guidance model containing linear dynamic differential equation constraints;

[0009] Obtaining a label data set for offline guidance according to the second trajectory planning guidance model, wherein the label data includes state variables and control variables that match each other;

[0010] Using the label data set to train an online guidance model integrating a physical information neural network, the online guidance model takes state variables as input and control variables as output;

[0011] The real-time control variables of the guidance are obtained from the real-time state variables using the trained online guidance model.

[0012] To solve the above problems, the present invention also provides a sparse data-driven guidance device integrating physical information neural network, comprising a processor and a storage medium, wherein the storage medium stores a computer program. When the processor runs the computer program in the storage medium, the sparse data-driven guidance method integrating physical information neural network can be executed.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0014] The present invention performs model convexification and linearization transformation on the nonlinear dynamic differential equation constraints of the launch vehicle trajectory planning and guidance, and converts the nonlinear dynamic differential equation constraints into linear dynamic differential equation constraints and the equation form of the guidance control variables with respect to the state variables, so that the trajectory planning and guidance problem can be converted into a mathematical form suitable for an embedded physical information neural network, thereby introducing the embedded physical information neural network into the field of intelligent guidance. By adding the dynamic differential equation constraints to the loss function of the deep neural network model, an online guidance model integrating the embedded physical information neural network is designed, which realizes the hybrid driven training of "data + model", effectively reduces the dependence of the deep neural network model on data, and improves the generalization performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The present invention will now be described by way of example with reference to the accompanying drawings, in which:

[0016] Figure 1 It is a flow chart of the sparse data driven guidance method of the fusion physical information neural network provided in the embodiment of the present application.

[0017] Figure 2 It is a training flow chart of the online guidance model provided in the embodiment of the present application.

[0018] Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 The following are comparison diagrams of the embodiment of the present application and the baseline algorithm in terms of altitude trajectory, range trajectory, speed trajectory, heading angle trajectory and rocket mass trajectory. DETAILED DESCRIPTION

[0019] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0020] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other alternative features that are equivalent or have similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0021] The online guidance model based on the traditional deep learning neural network model is highly dependent on the quantity and quality of label data. Although the embedded physical information neural network can achieve high-performance solution of the forward and inverse problems of partial differential equations and ordinary differential equations by relying on sparse data, it cannot be directly applied to the solution of the online guidance problem of the launch vehicle. The embodiment of the present application provides a sparse data-driven guidance method and device integrating the physical information neural network, aiming to apply the embedded physical information neural network to the solution of the online guidance problem of the launch vehicle, realize sparse data-driven guidance, reduce the dependence on label data, and improve the generalization ability of the guidance model.

[0022] In some embodiments, Figure 1 As shown, the embodiment of the present application provides a sparse data driven guidance method integrating physical information neural network, which includes the following process:

[0023] S1. Construct the first trajectory planning guidance model with nonlinear dynamic differential equation constraints.

[0024] The first trajectory planning and guidance model of the launch vehicle, in some embodiments, includes nonlinear dynamic differential equation constraints, state variable constraints, control variable constraints and performance indicator functions.

[0025] 1) Nonlinear dimensionless dynamic differential equation constraints.

[0026] ;

[0027] In the formula, , , , , They are , , , , The equation, is the rocket position vector, For speed, is the heading angle, is the mass of the rocket, together forming the first state variable . Control variable components and Represent thrust and angle of attack respectively, and together constitute the first control variable . To express the atmospheric resistance, we have:

[0028] , is the air density, is the reference area, is the drag coefficient, is the sea level atmospheric density, is the density scale, is the rocket's flight altitude, is the radius of the Earth, is the gravitational acceleration on the earth’s surface; is the rocket specific impulse.

[0029] 2) State variable boundary constraints.

[0030] The nonlinear dynamic differential equation constraints are expressed as ,in Indicates time, for The nonlinear dynamic differential equation constraint describes the constraint relationship between the rocket state variables (derivatives of) with respect to time and the control variables.

[0031] ;

[0032] In the formula, represents the maximum landing speed, represents the maximum landing angle of attack, Indicates the weight of the rocket structure; Indicates the moment when the rocket begins to land. Indicates the final landing moment of the rocket. express The first state variable at time , . They respectively represent , , , , , exist The constraints on the heading angle and angle of attack are to ensure that the rocket finally lands vertically.

[0033] 3) Control variable constraints.

[0034] ;

[0035] in, and are the (allowable) minimum and maximum thrusts, and are the (allowed) minimum and maximum angles of attack, respectively.

[0036] 4) Performance indicator function.

[0037] definition is the performance index of the objective function. In the first trajectory planning guidance model, The goal of model optimization is to minimize the performance metric.

[0038] The above constraints belong to a type of nonlinear multi-constrained dynamic trajectory optimization guidance problem, that is, the optimization of the first trajectory planning guidance model can be summarized as: design control variables , so that:

[0039] .

[0040] S2. Convexify the first trajectory planning and guidance model to obtain a second trajectory planning and guidance model containing linear dynamic differential equation constraints, so as to transform the trajectory planning and guidance problem into a convex trajectory optimization optimal control problem.

[0041] In some embodiments, the first trajectory planning guidance model is convexified using a linearization method and a relaxation technique.

[0042] As a feasible approach, the method of convexifying the first trajectory planning guidance model includes:

[0043] 1) Convert the nonlinear dimensionless dynamic differential equation constraints into linearized dimensionless dynamic differential equation constraints.

[0044] For the first control variable Dimension expansion processing to obtain the second control variable ,in ; The first state variable Convert to the second state variable ,in The constraints of the nonlinear dynamic differential equation after variable transformation are expressed as .

[0045] At a certain moment, the second state variable of the label data and the corresponding second control variable As a reference trajectory, Convert to linear dynamic differential equation constraints:

[0046] ;

[0047] in, ,

[0048] ,

[0049] .

[0050] 2) State variable boundary constraints.

[0051] ;

[0052] in, and Respectively Moment and The second state variable at time , express Moment .

[0053] 3) Control variable constraints.

[0054] ;

[0055] in, .

[0056] 4) Constraints on the second state variable and the second control variable.

[0057] ;

[0058] in, and Respectively The second state variable and the second control variable at time , and Respectively The second state variable and the second control variable of the moment label data, and represent the trust region constraint parameters of the second state variable and the second control variable respectively.

[0059] 5) Performance indicator function.

[0060] In the second trajectory planning guidance model, .

[0061] S3. Obtain a label data set for offline guidance according to the second trajectory planning guidance model, where the label data includes state variables and control variables that match each other.

[0062] In some embodiments, a method for solving a second trajectory planning guidance model to obtain a label data set for offline guidance includes:

[0063] The second trajectory planning guidance model is discretized using the trapezoidal integration method. The so-called trapezoidal integration method, also known as the trapezoidal rule, uses a trapezoid to estimate the area under the curve, and the intersection of the upper and lower bases of the trapezoid with the curve is the sampling point for discretizing the curve. The denser the sampling, the closer the fitting area is to the required area.

[0064] Given the initial state variables, aerodynamic model parameters ( ) and atmospheric density model parameters ( ), the second trajectory planning guidance model is iteratively solved to obtain offline guidance label data under the initial state variables. Based on the same method, interference factors are introduced into the initial state variables, aerodynamic model parameters and atmospheric density model parameters, and the second trajectory planning guidance model is iteratively solved to obtain offline guidance label data under the corresponding state variables. All label data are aggregated to form a label data set.

[0065] In some feasible implementations, interference factors are introduced into the initial state variables, aerodynamic model parameters, and atmospheric density model parameters. (or other values, the same applies to the interference factors introduced later) Deviation range, aerodynamic model parameters Deviation range, atmospheric density model parameters The corresponding parameter values ​​are randomly determined in the uncertain interference factors such as the deviation range. The label data obtained by iteratively solving the second trajectory planning guidance model constitutes the offline guidance label data set for multi-source uncertain factors. .

[0066] S4. Labeled Data Collection An online guidance model integrating physical information neural network is trained, which takes state variables as input and control variables as output.

[0067] In some embodiments, the online guidance model includes data information loss and physical information loss , carrying the linear dynamic differential equation constraints mentioned above in the physical information loss.

[0068] As a feasible implementation method, the linear dynamic differential equation constraints introduced in the physical information loss are introduced into the physical information loss of the online guidance model after being linearly transformed into equation constraints of the control variables with respect to the state variables.

[0069] Specifically, there are:

[0070] 1) The linear dynamic differential equation constraints are expressed as:

[0071] .

[0072] 2) In the process of convexification of the first trajectory planning guidance model, the linear dynamic differential equation constraints are transformed as follows using linear transformation:

[0073] ,

[0074] ;

[0075] 3) ;

[0076] Then the equation constraints of the control variables on the state variables are constructed as:

[0077] .

[0078] In some embodiments, Figure 2 As shown, the online guidance model includes an input layer, several hidden layers (seven layers are used as an example in the embodiment of the present application) and an output layer. is the input feature, and the second control variable As output features. Each hidden layer consists of 64 neurons. Each hidden layer is connected to an activation function. In some specific implementations, the configured activation function selects a ReLU function that can avoid the gradient vanishing problem in most cases and has a fast convergence speed.

[0079] In the above embodiment, the online guidance model includes data information loss and physical information loss For example, in some embodiments, the two losses are calculated using the mean square error method, which is:

[0080] Model Loss ,

[0081] ;

[0082] In the formula, Represents a collection of labeled data The total number of label data in the subscript Indicates Data training, superscript Represented as label data, represents the input of the current online guidance model, represents the output of the current online guidance model, and Respectively represent The input and output of labeled data during training.

[0083] In addition, before training the online guidance model, in order to improve the model training efficiency and enable the online guidance model to train and converge faster and better, the label data set The state variables and control variables in are normalized and compressed to within the range.

[0084] As an optional method, during the iterative training of the online guidance model, the Adam optimizer is used to update the network parameters during the iteration process.

[0085] Through the design of the above embodiment, intelligent online guidance driven by sparse data is realized, the dependence on the amount of label data is reduced, and the generalization ability of the guidance model is improved.

[0086] S5. Obtaining guided real-time control variables from real-time state variables using the trained online guidance model.

[0087] Finally, the online guidance model trained in step S4 is loaded onto the launch vehicle, and the launch vehicle can obtain the corresponding control variables in real time according to the current state variables, thereby realizing online guidance. This provides a feasible solution to the problems of emergency launch of launch vehicles and orbital rescue.

[0088] An embodiment of the present application also provides a sparse data-driven guidance device integrating physical information neural network, including a processor and a storage medium, wherein the storage medium stores a computer program. When the processor runs the computer program in the storage medium, the sparse data-driven guidance method integrating physical information neural network of the above embodiment can be executed.

[0089] The present application embodiment also compares the performance of the above method (OURS) with the guidance using the baseline algorithm DNN (Deep Nueral Network) and SCP (Sequential Convex Optimization). Figure 3 The following is a comparison of the height trajectories of each model. Figure 4 The following is a comparison of the flight trajectories of each model. Figure 5 The speed trajectory comparison diagram of each model is shown. Figure 6 The following is a comparison of the heading angle trajectories of each model. Figure 7 The figure shows the comparison of rocket mass trajectories of various models. Figure 3-Figure 7 It can be seen that the method proposed in this application has better performance than ordinary deep neural networks.

[0090] The present invention is not limited to the above-mentioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A sparse data driven guidance method integrating physical information neural network, characterized in that: include: Construct the first trajectory planning guidance model with nonlinear dynamic differential equation constraints; Convexifying the first trajectory planning and guidance model to obtain a second trajectory planning and guidance model containing linear dynamic differential equation constraints; Obtaining a label data set for offline guidance according to the second trajectory planning guidance model, wherein the label data includes state variables and control variables that match each other; Using the label data set to train an online guidance model integrating a physical information neural network, the online guidance model takes state variables as input and control variables as output; The real-time control variables of the guidance are obtained from the real-time state variables using the trained online guidance model.

2. The sparse data driven guidance method of integrating physical information neural network according to claim 1, characterized in that: Before training the online guidance model with the label data set, each state variable and control variable in the label data set is first normalized.

3. The sparse data driven guidance method of integrating physical information neural network according to claim 1 or 2, characterized in that: The first trajectory planning and guidance model is constructed as: Satisfy the constraints: ， In the formula, ， in, is the performance index of the objective function. In the first trajectory planning guidance model, ; is the nonlinear dynamic differential equation constraint, Indicates time; , , , , , They are , , , , , Derivatives with respect to time; is the rocket position vector, For speed, is the heading angle, is the mass of the rocket, together forming the first state variable ; and Represent thrust and angle of attack respectively, and together constitute the first control variable ; represents the atmospheric resistance, , is the air density, is the reference area, is the drag coefficient, is the sea level atmospheric density, is the density scale, is the rocket's flight altitude, is the distance from the rocket to the center of the earth, is the radius of the Earth, is the gravitational acceleration on the earth’s surface; is the rocket specific impulse; represents the maximum landing speed, represents the maximum landing angle of attack, Indicates the weight of the rocket structure; Indicates the moment when the rocket begins to land. Indicates the final landing moment of the rocket. and are the minimum thrust and the maximum thrust respectively, and are the minimum angle of attack and the maximum angle of attack respectively; express The first state variable at time , .

4. The sparse data driven guidance method of the fusion physical information neural network as claimed in claim 3, characterized in that: The first trajectory planning guidance model is convexified using a linearization method and a relaxation technique.

5. The sparse data driven guidance method of integrating physical information neural network as claimed in claim 4, characterized in that: The method for convexifying the first trajectory planning and guidance model includes: For the first control variable Dimension expansion processing to obtain the second control variable ,in ; The first state variable Convert to the second state variable ,in ;get ; Tag data state variables and the corresponding control variables As a reference trajectory, Convert to linear dynamics differential equation constraints ; Where: ， ， ； The constraints are: ； The performance index of the objective function in the second trajectory planning guidance model is: ; in, express Moment , , and Respectively The second state variable and the second control variable at time , and Respectively The second state variable and the second control variable of the label data at the time, and represent the trust region constraint parameters of the second state variable and the second control variable respectively.

6. The sparse data driven guidance method of integrating physical information neural network according to claim 1, characterized in that: Obtaining a label data set for offline guidance according to the second trajectory planning guidance model includes: Discretizing the second trajectory planning and guidance model using a trapezoidal integration method; Based on the given initial state variables, aerodynamic model parameters and atmospheric density model parameters, and based on the introduction of interference factors into the initial state variables, aerodynamic model parameters and atmospheric density model parameters, the corresponding label data are respectively solved to form a label data set.

7. The sparse data driven guidance method of integrating physical information neural network as claimed in claim 5, characterized in that: The physical information loss of the online guidance model carries the linear dynamic differential equation constraints.

8. The sparse data driven guidance method of integrating physical information neural network according to claim 7, characterized in that: The linear dynamic differential equation constraints are introduced into the physical information loss of the online guidance model after being linearly transformed into equation constraints of the control variables with respect to the state variables.

9. The sparse data driven guidance method of integrating physical information neural network as claimed in claim 8, characterized in that: The equation constraints of the control variables with respect to the state variables are constructed as: 。 10. A sparse data driven guidance device integrating physical information neural network, comprising a processor and a storage medium, characterized in that: The storage medium stores a computer program, and when the processor runs the computer program in the storage medium, it can execute the sparse data-driven guidance method of the fusion physical information neural network as described in any one of claims 1-9.

Images