A low-cost power envelope inference method based on neural networks

Abstract

This invention belongs to the field of commercial aerospace and solid rocket engine performance optimization, and discloses a low-cost dynamic envelope inference method based on neural networks. The method includes: establishing a geometric model characterized by feature parameters for each propellant type; constructing a training sample dataset for each propellant type based on the feature parameters at a first scale, and calculating the burn surface area sequence and free volume sequence of each training sample as labels using the burn surface shift step size; training a first neural network and a second neural network for each propellant type using the training samples; constructing a prediction sample dataset for each propellant type at a second scale, and obtaining the result for each prediction sample using the corresponding first and second neural networks; constructing an engine thrust calculation model by combining zero-dimensional internal ballistic equations and one-dimensional steady isentropic flow theory, and calculating the engine thrust for each prediction sample, thereby obtaining the rocket engine dynamic envelope corresponding to each propellant type.

Description

Technical Field

[0001] This invention belongs to the field of commercial aerospace and solid rocket engine performance optimization, specifically involving a low-cost dynamic envelope inference method based on neural networks. Background Technology

[0002] Commercial space launch vehicles have received widespread attention in recent years due to their low cost, support from private and commercial policies, and reusability. However, commercial space launch vehicles typically use solid rocket motors, and their thrust calculations rely on accurate numerical simulations of the propellant combustion and propulsion process. However, the internal combustion dynamics of the engine during operation involve complex flow and geometric propulsion evolution, and establishing a high-precision propellant propulsion model would drastically increase the cost and iteration cycle of disciplinary analysis and calculations during the overall feasibility study phase.

[0003] To rapidly evaluate the thrust performance of a low-cost engine casing under length and diameter constraints during the overall design phase, a dynamic envelope inference method is needed that can quickly calculate an accurate family of thrust curves based on the input set of propellant grain characteristic parameters. This would provide a feasible design range reference for subsequent selection of the engine casing's length and diameter. However, existing methods typically rely on the chief engineer's experience or time-consuming manual iteration, which cannot quickly generate a quantified dynamic envelope range under given low-cost constraints on the engine casing's length and diameter. Summary of the Invention

[0004] The purpose of this invention is to provide a low-cost dynamic envelope inference method based on neural networks, which solves the problem of low efficiency in existing technologies such as manual iteration or empirical methods.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A low-cost dynamic envelope inference method based on neural networks includes:

[0007] A geometric model characterized by feature parameters is established for each charge type; a training sample dataset for each charge type is constructed based on the feature parameters according to the first scale, and the burning surface area sequence and free volume sequence of each training sample are calculated as labels in combination with the burning surface shift step size; the first neural network and the second neural network for each charge type are trained using the training samples;

[0008] A prediction sample dataset for each propellant type is constructed at a second scale. For each prediction sample, the predicted burn area sequence and the predicted free volume sequence are obtained using the corresponding first neural network and second neural network. An engine thrust calculation model is constructed by combining the zero-dimensional internal ballistic equation and the one-dimensional steady isentropic flow theory, and the engine thrust of each prediction sample is calculated, thereby obtaining the rocket engine power envelope corresponding to each propellant type.

[0009] Furthermore, the charge types include round-hole propellant charges, star-shaped propellant charges, and wing-shaped propellant charges;

[0010] The geometric model of a circular orifice propellant grain is characterized by the following characteristic parameters: engine length. Engine radius , ratio of inner diameter of round hole The ratio of the inner diameter of the circular hole according to Convert to inner diameter of round hole ;

[0011] The geometric model of the star-shaped propellant grain is characterized by the following characteristic parameters: engine length. Engine radius Feature length ratio Number of star apertures , star-edge angle Star angle coefficient ; where the feature length ratio according to Convert to star aperture length :

[0012] The geometric model of the wing-pole propellant grain is characterized by the following characteristic parameters: engine length. Engine radius , ratio of inner diameter of round hole Number of wing pillars wing column inclination angle wing pillar height ratio Wing tip length ratio Wing thickness ratio The ratio of the inner diameter of the circular hole according to Convert to inner diameter of round hole ; wing pillar height ratio according to Converted to wing column height Wing tip length ratio according to Converted to wingtip length Wing thickness ratio according to Converted to wing column thickness .

[0013] Furthermore, when constructing the first neural network and the second neural network for each charge type, the burning surface shift step size for each charge type is first constructed:

[0014] ;

[0015] in, , and These are the burning surface shift steps for round-hole propellant grains, star-shaped propellant grains, and wing-shaped propellant grains, respectively. Indicates the number of steps the burning surface is moved;

[0016] For each charge type, a corresponding training sample dataset is constructed. First, the value space of each feature parameter corresponding to the charge type is defined. Then, the Latin hypercube sampling method is used to sample the value space of all feature parameters at the first scale. Each set of feature parameter values ​​obtained from each sampling is used as a training sample.

[0017] For each training sample, the burning surface area and free volume of the circular hole propellant, star hole propellant, and wing-shaped propellant at each burning surface shift step are calculated using geometric modeling, thereby obtaining the corresponding burning surface area sequence and free volume sequence as the label of the training sample.

[0018] Furthermore, the training sample dataset is preprocessed, including filtering the surface area sequence and free volume sequence corresponding to each training sample using a Savitzky-Golay filter; the maximum order of the Savitzky-Golay filter is set to 2, and the single-side size of the moving window is set to 3.

[0019] Furthermore, the first and second neural networks for each type of charge are fully connected neural networks, including an input layer, a hidden layer, and an output layer. In the embodiments of the present invention, the activation function of the hidden layer of the first and second neural networks is selected as the hyperbolic tangent function; the activation function of the output layer is selected as a linear function; the first and second neural networks are trained using the Bayesian regularization algorithm, and the prediction accuracy of the first and second neural networks is measured using the complex correlation coefficient until the preset requirements are met and training stops.

[0020] Furthermore, a prediction sample dataset for predicting the dynamic envelope is constructed for each charge type; wherein the construction form of each prediction sample is consistent with that of the training samples, the difference being that the second size is larger than the first size;

[0021] For a predicted sample of each charge type, it is input into the first neural network and the second neural network trained for that charge type to obtain the corresponding predicted burning surface area sequence and predicted free volume sequence, and the engine thrust curve of the predicted sample is determined based on the engine thrust calculation model.

[0022] Furthermore, an engine thrust calculation model is constructed by combining the zero-dimensional internal ballistic equations with the one-dimensional steady isentropic flow theory, including:

[0023] For each predicted sample, the predicted burning area sequence is obtained. And predicting free volume sequences Based on this, calculate the average density of the gas. Combustion chamber pressure :

[0024] ;

[0025] ;

[0026] In the formula, For differential operators, and To predict the burning surface area sequence And predicting free volume sequences exist The interpolated combustion surface area and free volume at each moment and These are the average gas density and the combustion chamber pressure, respectively. Density of the drug column This represents the area of ​​the rocket engine nozzle throat. Characteristic velocity, The gas constant is... The operating temperature of the combustion chamber, and its thermal coefficient. To the specific heat ratio of gases Related single-valued functions; , and These are the surface burning rate, burning rate coefficient, and burning rate pressure index, respectively:

[0027] ;

[0028] ;

[0029] The average density of the gas was obtained by solving the problem. Combustion chamber pressure Then, combined with the nozzle expansion ratio The nozzle exit pressure is calculated using the following formula:

[0030] ;

[0031] In the formula, The nozzle expansion ratio, This refers to the nozzle exit pressure.

[0032] Nozzle exit gas velocity Represented as:

[0033] ;

[0034] Nozzle mass flow rate Represented as:

[0035] ;

[0036] Based on nozzle mass flow rate Nozzle exit gas velocity With nozzle exit pressure Obtain engine thrust for:

[0037] ;

[0038] In the formula, External environmental pressure. This represents the nozzle exit area.

[0039] Furthermore, by using the engine thrust calculation model, the engine thrust at different times is calculated using each predicted sample, thereby constructing the engine thrust curve;

[0040] The engine thrust curves corresponding to all predicted samples for each type of propellant form the thrust envelope for that propellant type, which is the rocket engine power envelope.

[0041] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the low-cost dynamic envelope inference method based on neural networks.

[0042] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the low-cost dynamic envelope inference method based on a neural network.

[0043] Compared with the prior art, the present invention has the following technical features:

[0044] This invention establishes solid rocket motor propellant models for circular, star-shaped, and wing-pole configurations, enabling automatic calculation of burner surface area and free volume under different propellant parameter input conditions. It innovatively proposes a neural network-assisted rapid thrust calculation method. This method trains the neural network using a small training sample dataset and then uses a large prediction sample dataset to replace the traditional, time-consuming process of propellant surface geometric shifting with the neural network's prediction results. Based on the input propellant parameters, it predicts the burner surface area and free volume sequences, and combines this with internal ballistic calculations to obtain engine thrust. This achieves thrust calculations in seconds and obtains the dynamic envelopes of various engine configurations under different engine length and diameter conditions, providing a thrust limit range reference for the low-cost design of solid rocket motors in commercial space launch vehicles. This method has been applied to the overall design of the propulsion system of low-cost commercial space launch vehicles, reducing thrust calculation time to 2 seconds and rapidly generating dynamic envelopes for different propellant configurations under engine length and diameter constraints. Attached Figure Description

[0045] Figure 1 This is a schematic flowchart of the method of the present invention;

[0046] Figure 2 The diagram shows the geometric models of three types of propellant grains for solid rocket motors; (a) and (b) are the right and front views of a circular hole propellant grain; (c) and (d) are the right and front views of a star hole propellant grain; (e) and (f) are the right and front views of a wing-shaped propellant grain.

[0047] Figure 3 The prediction accuracy of the first neural network and the second neural network for the circular hole drug cartridge in this embodiment of the invention;

[0048] Figure 4 The prediction accuracy of the first neural network and the second neural network of the star-shaped drug column in the embodiments of the present invention;

[0049] Figure 5 The prediction accuracy of the first neural network and the second neural network of the wing-shaped propellant column in the embodiments of the present invention;

[0050] Figure 6 This refers to the rocket engine power envelope of the circular propellant grain in this embodiment of the invention;

[0051] Figure 7 This refers to the rocket engine power envelope of the star-shaped propellant grain in this embodiment of the invention;

[0052] Figure 8 This refers to the rocket engine power envelope of the wing-shaped propellant grain in this embodiment of the invention. Detailed Implementation

[0053] To address the problems existing in the prior art, it is necessary to introduce neural network approximation into thrust calculation; this invention provides a low-cost dynamic envelope inference method based on neural networks, see [link to related document]. Figure 1 For each charge type, a geometric model characterized by feature parameters is established; based on the feature parameters, a training sample dataset for each charge type is constructed according to a first scale, and combined with the corresponding burning surface shift step size, the burning surface area sequence and free volume sequence corresponding to each training sample are calculated as labels; for each charge type, the first neural network and the second neural network are trained using the training samples respectively; the inputs of the first neural network and the second neural network are both feature parameters, and the outputs are the predicted burning surface area sequence and the predicted free volume sequence, respectively;

[0054] A prediction sample dataset for each propellant type is constructed at a second scale. For each prediction sample, a predicted burn area sequence and a predicted free volume sequence are obtained using the corresponding first and second neural networks. An engine thrust calculation model is constructed by combining the zero-dimensional internal ballistic equations and the one-dimensional steady isentropic flow theory, and the engine thrust for each prediction sample is calculated. This yields the rocket engine power envelope corresponding to each propellant type, supporting the overall design of commercial aerospace rocket engines. The specific implementation process of this invention will be further described in detail below with reference to the accompanying drawings.

[0055] Step 1: Establish geometric models for various charge types.

[0056] The types of propellant charges described in this plan include round-hole propellant charges, star-hole propellant charges, and wing-type propellant charges; such as Figure 2 As shown.

[0057] Step 1.1, Geometric model of the circular orifice propellant column.

[0058] A round-hole propellant charge refers to a propellant charge with an internal axially oriented round hole, such as... Figure 2 As shown in (a) and (b).

[0059] The geometric model of a circular orifice propellant grain is characterized by the following characteristic parameters: engine length. Engine radius , ratio of inner diameter of round hole Among them, the ratio of the inner diameter of the circular hole is defined. This is to avoid unusual or interference phenomena in the round-hole propellant cartridge; the ratio of the inner diameter of the round hole... It needs to be further converted into the inner diameter of the circular hole according to the constraint conditions of equation (1). .

[0060] (1);

[0061] Step 1.2, Geometric model of the star-shaped propellant column.

[0062] A star-shaped propellant grain refers to a propellant grain with a star-shaped hole in its internal axial direction. This star-shaped hole is formed by multiple rectangular holes arranged circumferentially, such as... Figure 2 As shown in (c) and (d).

[0063] The geometric model of the star-shaped propellant grain is characterized by the following characteristic parameters: engine length. Engine radius Feature length ratio Number of star apertures , star-edge angle Star angle coefficient ; combination Figure 2 (d), the included angle of the star's edge It refers to the included angle between adjacent star apertures; star angle coefficient This refers to the angle between the diagonal and the long side of the star aperture. Angle with the edge of the star The ratio, i.e. Define the feature length ratio. This is to avoid singularities and interference phenomena in the star-shaped propellant grain configuration; characteristic length ratio It needs to be converted into star aperture length according to the constraint conditions of equation (2). :

[0064] (2);

[0065] Step 1.3, Geometric model of wing-shaped propellant grain.

[0066] A wing-shaped propellant grain is formed by adding multiple radiating wing-shaped columns circumferentially at the rear end of a circular hole, based on a circular hole propellant grain. For example... Figure 2 As shown in (e) and (f), each wing column is an airfoil-shaped structure.

[0067] The geometric model of a wing-shaped propellant grain is characterized by the following characteristic parameters:

[0068] Engine length Engine radius , ratio of inner diameter of round hole Number of wing pillars wing column inclination angle wing pillar height ratio Wing tip length ratio Wing thickness ratio Number of wing pillars This refers to the number of wing post holes and the wing post inclination angle. This refers to the angle between the hypotenuse at the tip of the wing-pillar and the axis of the wing-pillar propellant grain; wing-pillar height ratio This refers to the height of the wing-shaped propellant grain on its cross-section. With engine radius The ratio; the wingtip length ratio This refers to the wingtip length of the wing pillar. With engine length The ratio; wing thickness ratio Refers to the thickness of the wing column With engine radius The ratio of .

[0069] Define the ratio of the inner diameter of the circular hole wing pillar height ratio Wing tip length ratio Wing thickness ratio This is to avoid strange or interference phenomena in the configuration of the wing-shaped propellant grain.

[0070] The ratio of the inner diameter of the round hole It needs to be converted into the inner diameter of the circular hole according to the constraint conditions of equation (3). :

[0071] (3);

[0072] wing pillar height ratio It needs to be converted into wing column height according to the constraint conditions of equation (4). :

[0073] (4);

[0074] Wing tip length ratio It needs to be converted into wingtip length according to the constraint conditions of equation (5). :

[0075] (5);

[0076] Wing thickness ratio The constraint conditions in equation (6) need to be converted into wing column thickness. :

[0077] (6);

[0078] Step 2, building and training the neural network.

[0079] Step 2.1, construct the combustion surface shift step for each charge type, as shown in equation (7):

[0080] (7);

[0081] in, , and These are the burning surface shift steps for round-hole propellant grains, star-shaped propellant grains, and wing-shaped propellant grains, respectively. This indicates the number of steps the burning surface is pushed forward; for example, in the example of this invention, .

[0082] Step 2.2: Construct a corresponding training sample dataset for each type of charge.

[0083] In this plan, for each type of explosive charge:

[0084] First, the value space of each feature parameter corresponding to the charge type is defined; then, for all feature parameter value spaces, the Latin hypercube sampling method is used to sample at a first scale, and the specific value of a set of feature parameters obtained each time is used as a training sample, thereby obtaining the training sample dataset of the charge type; in the embodiment of the present invention, the first scale is 100, that is, 100 training samples are generated by sampling.

[0085] The training samples for the circular orifice drug cartridges are represented as follows: The training samples for star-shaped drug grains are represented as follows: The training samples for wing-shaped propellant grains are represented as follows: .

[0086] Combining equations (1) to (7) and geometric models of various charge types, for each training sample, the burning surface area and free volume of the circular hole charge, star hole charge, and wing-shaped charge at each burning surface shift step are calculated using geometric modeling, thereby obtaining the corresponding burning surface area sequence and free volume sequence, which serve as the label for the training sample.

[0087] Step 2.3: For each charge type, preprocess the training sample dataset, specifically by filtering the burn area sequence and free volume sequence corresponding to each training sample using a Savitzky-Golay filter; the maximum order of the Savitzky-Golay filter... Set to 2 to move the window size on one side. Set it to 3.

[0088] Step 2.4: For each type of charge, construct the corresponding first neural network and second neural network.

[0089] In this scheme, the first and second neural networks for each type of charge are fully connected neural networks, including an input layer, a hidden layer, and an output layer. In the embodiments of the present invention, the activation function of the hidden layer of the first and second neural networks is selected as the hyperbolic tangent function; the activation function of the output layer is selected as a linear function.

[0090] The number of hidden layer nodes in the first and second neural networks of the circular-hole and star-shaped propellant grains is set to 20; while the number of hidden layer nodes in the first and second neural networks of the wing-shaped propellant grains is set to 40.

[0091] The input layer of the first neural network for each charge type is used to input training samples of that charge type, and the output layer is used to output the predicted burning surface area sequence; the input layer of the second neural network for each charge type is used to input training samples of that charge type, and the output layer is used to output the predicted free volume sequence.

[0092] Step 2.5: Based on the preprocessed training sample dataset for each charge type, use the Bayesian regularization algorithm to train the corresponding first neural network and second neural network.

[0093] The following explanation uses the training sample dataset after preprocessing the circular orifice propellant column as an example:

[0094] The preprocessed training samples Inputting the first neural network with the filtered burning surface area sequence as the label of the training samples, the output of the first neural network is the predicted burning surface area sequence; while... If the filtered free volume sequence is used as the label for the training samples, the output of the second neural network is the predicted free volume sequence.

[0095] During training, the multiple correlation coefficient between the predicted burning area sequence output by the first neural network and the burning area sequence in the training sample labels is... A correlation coefficient greater than 0.95 indicates that the prediction accuracy of the first neural network meets the requirements; similarly, the correlation coefficient between the predicted free volume sequence output by the second neural network and the free volume sequence in the training sample labels is also considered high. When the value is greater than 0.95, it indicates that the prediction accuracy of the second neural network meets the requirements; at this point, training is stopped, and the trained first and second neural networks are saved.

[0096] Taking the first neural network as an example, the multiple correlation coefficient The calculation is as follows:

[0097] (8);

[0098] in, , , The first The multiple correlation coefficient, total sum of squares, and residual sum of squares for each training sample; and For the first The training sample at the th ... The predicted and actual burning areas at each step length (i.e., the values ​​of the burning area sequence in the labels at each step length). For the first The average value of the real burning surface area sequence corresponding to each training sample. This represents the number of training samples.

[0099] In practical applications, for a new set of feature parameters of a circular orifice propellant column, inputting them into the trained first neural network and second neural network will yield the corresponding predicted burning surface area sequence and predicted free volume sequence.

[0100] The training process for star-shaped and wing-shaped propellant grains is the same as that for round-hole propellant grains, and will not be repeated here.

[0101] Step 3, Prediction Phase.

[0102] For each charge type, a prediction sample dataset is constructed for predicting the kinetic envelope. Each prediction sample in the dataset is constructed in the same way as the training samples, consisting of a set of feature parameters for that charge type. During the construction of the prediction sample dataset, Latin hypercube sampling is performed on a second scale, larger than the first scale, within the value space of each feature parameter. Each set of feature parameter values ​​obtained from each sampling is considered a prediction sample, thus obtaining the prediction sample dataset. In this embodiment, the second scale is 2000, meaning 2000 prediction samples are generated for each charge type to cover the sampling space as much as possible. This method effectively saves computational resources and improves computational efficiency by training the neural network with a smaller training sample dataset and performing predictions with a larger prediction sample dataset.

[0103] For a predicted sample of each charge type, it is input into the first neural network and the second neural network trained for that charge type to obtain the corresponding predicted burning surface area sequence and predicted free volume sequence: and ;in, To predict the first in the fire surface area sequence The area of ​​the burning surface corresponding to each step of the burning surface movement. To predict the first free volume sequence The free volume corresponding to each step of the burning surface movement.

[0104] This scheme does not directly use the characteristic parameters of the input propellant type to map the final thrust output through a neural network. The reason is that if the thrust is directly output, the nonlinear numerical characteristics of the internal ballistic physics equations will lead to a decrease in accuracy. Therefore, this scheme uses a neural network to input the characteristic parameters to obtain the predicted burner surface area sequence and the predicted free volume sequence. Then, the engine thrust is calculated using the predicted burner surface area sequence and the predicted free volume sequence. This achieves accurate thrust calculation while ensuring a short calculation time.

[0105] Specifically, an engine thrust calculation model is constructed by combining the zero-dimensional internal ballistic equations and the one-dimensional steady isentropic flow theory, as follows:

[0106] Step 3.1, zero-dimensional internal ballistic calculation of rocket engine.

[0107] For each predicted sample, the predicted burning area sequence is obtained. And predicting free volume sequences Based on this, the average density of the gas is calculated using equations (9) and (10). Combustion chamber pressure :

[0108] (9);

[0109] (10);

[0110] In the formula, For differential operators, and To predict the burning surface area sequence And predicting free volume sequences exist The interpolated combustion surface area and free volume at each moment and These are the average gas density and the combustion chamber pressure, respectively. Density of the drug column This represents the area of ​​the rocket engine nozzle throat. Characteristic velocity, The gas constant is... The operating temperature of the combustion chamber, and its thermal coefficient. To the specific heat ratio of gases The relevant single-valued functions are shown in equation (12); , and These are the surface burning rate, the burning rate coefficient, and the burning rate pressure index, respectively, and their specific relationships are shown in equation (13):

[0111] (11);

[0112] (12);

[0113] Step 3.2, calculation of one-dimensional steady isentropic flow in the nozzle.

[0114] For each predicted sample, the average gas density is obtained by solving the problem. Combustion chamber pressure Then, combined with the nozzle expansion ratio The nozzle exit pressure is calculated using the following formula:

[0115] (13);

[0116] In the formula, The nozzle expansion ratio, This is the nozzle outlet pressure.

[0117] Nozzle exit gas velocity It can be represented as:

[0118] (14);

[0119] Nozzle mass flow rate It can be represented as:

[0120] (15);

[0121] Based on nozzle mass flow rate Nozzle exit gas velocity With nozzle exit pressure Engine thrust can be obtained for:

[0122] (16);

[0123] In the formula, External environmental pressure. This represents the nozzle exit area.

[0124] Using the engine thrust calculation model described above, the thrust can be calculated based on each predicted sample. The engine thrust at each moment is measured to obtain engine thrust curves at different times.

[0125] The thrust envelope formed by the engine thrust curves corresponding to all predicted samples for each type of propellant is the rocket engine power envelope used for subsequent overall rocket design reference. Based on the power envelopes of different propellant types and the performance requirements of the rocket engine, the propellant type is selected and the optimal characteristic parameters are determined.

[0126] Example

[0127] In one embodiment of the present invention, the characteristic parameter value ranges of the three charge types—circular hole charge, star hole charge, and wing-shaped charge—when constructing training samples and prediction samples are shown in Table 1.

[0128] Table 1: Range of characteristic parameters.

[0129]

[0130] In this embodiment, the training sample dataset has a size of 400, while the prediction sample dataset has a size of 2000.

[0131] The internal ballistic parameters used in the engine thrust calculation are shown in Table 2.

[0132] Table 2: Values ​​of internal ballistic parameters.

[0133]

[0134] like Figures 3 to 5 The figure shows the prediction accuracy of the first and second neural networks for each charge type in this embodiment; the multiple correlation coefficients in the figure are used to illustrate this. It can be seen that the prediction accuracy of the first and second neural networks corresponding to each charge type is relatively high, which can well meet the actual use requirements.

[0135] Figures 6 to 8 This refers to the rocket engine dynamic envelope corresponding to each propellant type in this embodiment. Through these rocket engine dynamic envelopes, designers can quickly and intuitively find feasible thrust curves directly under the constraints of rocket engine size. At the same time, the envelope can also be used as input for overall and ballistic design to quickly perform center of mass assessment, aerodynamic trim, and ballistic optimization.

[0136] Compared with existing methods that directly use numerical simulation to calculate thrust, the calculation time statistics are shown in Table 3, which further demonstrates that the calculation efficiency of the method of the present invention is effectively improved compared with the prior art.

[0137] Table 3: Comparison of calculation efficiency of rocket engine power envelope.

[0138]

[0139] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A low-cost power envelope inference method based on neural networks, characterized in that, include: A geometric model characterized by feature parameters is established for each type of charge; Based on the feature parameters, a training sample dataset for each charge type is constructed according to the first scale. The burning surface area sequence and free volume sequence of each training sample are calculated as labels by combining the burning surface shift step size. The first neural network and the second neural network for each type of charge were trained using training samples; A prediction sample dataset for each propellant type is constructed at the second scale. For each prediction sample, the predicted burn area sequence and the predicted free volume sequence are obtained using the corresponding first neural network and second neural network. An engine thrust calculation model is constructed by combining the zero-dimensional internal ballistic equation and the one-dimensional steady isentropic flow theory, and the engine thrust of each prediction sample is calculated, thereby obtaining the rocket engine power envelope corresponding to each propellant type. The charge types include round-hole propellant grains, star-hole propellant grains, and wing-shaped propellant grains; The geometric model of the round-bore propellant grain is characterized by the following characteristic parameters: engine length , engine radius , round-bore inner diameter ratio ; wherein the round-bore inner diameter ratio is converted into a round-bore inner diameter according to ; The geometric model of the star-hole propellant grain is characterized by the following characteristic parameters: engine length , engine radius , characteristic length ratio , number of star-holes , star-edge included angle , star-angle coefficient ; wherein the characteristic length ratio is converted into the star-hole length according to : The geometric model of the wing-pole propellant grain is characterized by the following characteristic parameters: engine length. Engine radius , ratio of inner diameter of round hole Number of wing pillars wing column inclination angle wing pillar height ratio Wing tip length ratio Wing thickness ratio The ratio of the inner diameter of the circular hole according to Convert to inner diameter of round hole ; wing pillar height ratio according to Converted to wing column height Wing tip length ratio according to Converted to wingtip length Wing thickness ratio according to Converted to wing column thickness ; When constructing the first neural network and the second neural network for each charge type, the burning surface shift step size for each charge type is first constructed: ； in, , and These are the burning surface shift steps for round-hole propellant grains, star-shaped propellant grains, and wing-shaped propellant grains, respectively. Indicates the number of steps the burning surface is moved; For each charge type, a corresponding training sample dataset is constructed. First, the value space of each feature parameter corresponding to the charge type is defined. Then, the Latin hypercube sampling method is used to sample the value space of all feature parameters at the first scale. Each set of feature parameter values ​​obtained from each sampling is used as a training sample. For each training sample, the burning surface area and free volume of the circular hole propellant, star hole propellant, and wing-shaped propellant at each burning surface shift step are calculated using geometric modeling, thereby obtaining the corresponding burning surface area sequence and free volume sequence as the label of the training sample.

2. The low-cost dynamic envelope inference method based on neural networks according to claim 1, characterized in that, The training sample dataset is preprocessed, including filtering the surface area sequence and free volume sequence corresponding to each training sample using a Savitzky-Golay filter; the maximum order of the Savitzky-Golay filter is set to 2, and the single-side size of the moving window is set to 3.

3. The low-cost dynamic envelope inference method based on neural networks according to claim 1, characterized in that, For each type of charge, the first and second neural networks are fully connected neural networks, including an input layer, a hidden layer, and an output layer. The activation function of the hidden layer of the first and second neural networks is the hyperbolic tangent function, and the activation function of the output layer is a linear function. The first and second neural networks are trained using a Bayesian regularization algorithm, and the prediction accuracy of the first and second neural networks is measured using the multiple correlation coefficient until the preset requirements are met, at which point training stops.

4. The low-cost dynamic envelope inference method based on neural networks according to claim 1, characterized in that, For each charge type, a prediction sample dataset is constructed for predicting the dynamic envelope; the construction of each prediction sample is consistent with that of the training samples, except that the second size is larger than the first size. For a predicted sample of each charge type, it is input into the first neural network and the second neural network trained for that charge type to obtain the corresponding predicted burning surface area sequence and predicted free volume sequence, and the engine thrust curve of the predicted sample is determined based on the engine thrust calculation model.

5. The low-cost dynamic envelope inference method based on neural networks according to claim 1, characterized in that, An engine thrust calculation model is constructed by combining the zero-dimensional internal ballistic equations and the one-dimensional steady isentropic flow theory, including: For each predicted sample, the predicted burning area sequence is obtained. And predicting free volume sequences Based on this, calculate the average density of the gas. Combustion chamber pressure : ； ； In the formula, For differential operators, and To predict the burning surface area sequence And predicting free volume sequences exist The interpolated surface area and free volume at each moment, and These are the average gas density and the combustion chamber pressure, respectively. Density of the drug column This represents the area of ​​the rocket engine nozzle throat. Characteristic velocity, The gas constant is... The operating temperature of the combustion chamber, and its thermal coefficient. To the specific heat ratio of gases Related single-valued functions; , and These are the surface burning rate, burning rate coefficient, and burning rate pressure index, respectively: ； ； The average density of the gas was obtained by solving the problem. Combustion chamber pressure Then, combined with the nozzle expansion ratio The nozzle exit pressure is calculated using the following formula: ； In the formula, The nozzle expansion ratio, This refers to the nozzle exit pressure. Nozzle exit gas velocity Represented as: ； Nozzle mass flow rate Represented as: ； Based on nozzle mass flow rate Nozzle exit gas velocity With nozzle exit pressure Obtain engine thrust for: ； In the formula, External environmental pressure. This represents the nozzle exit area.

6. The low-cost dynamic envelope inference method based on neural networks according to claim 1, characterized in that, By using the engine thrust calculation model, the engine thrust at different times is calculated using each predicted sample, thereby constructing the engine thrust curve. The engine thrust curves corresponding to all predicted samples for each type of propellant form the thrust envelope for that propellant type, which is the rocket engine power envelope.

7. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes a computer program, it implements the low-cost dynamic envelope inference method based on neural networks as described in any one of claims 1-6.

8. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by a processor, it implements the low-cost dynamic envelope inference method based on neural networks as described in any one of claims 1-6.

Images