Solid rocket power cost-effectiveness prediction method based on coordinate descent deep integration

Abstract

This invention relates to the fields of aerospace vehicle overall design, and discloses a method for predicting the cost and efficiency of solid rocket propulsion based on coordinate descent deep integration. The method includes: constructing a multi-source fusion database for solid rocket engines and training an integrated sparse neural network based on coordinate descent optimization; the integrated sparse neural network includes an additive model, which is constructed from base learners and corresponding weight coefficients; during the training of each base learner in the integrated sparse neural network, an objective function including a mean squared error term and a norm regularization term is constructed; the norm regularization term is used to apply sparse constraints to the weight matrix from the input layer to the hidden layer of the base learners; for a new design scheme of the solid rocket engine, its feature parameters are obtained and input features are constructed, which are then input into the trained integrated sparse neural network to obtain the predicted output response, which is used to estimate the cost or efficiency of the solid rocket engine propulsion system; this invention has fast convergence speed and high accuracy.

Description

Technical Field

[0001] This invention relates to the fields of overall design of aerospace vehicles, multidisciplinary design optimization, and advanced manufacturing economic evaluation, specifically to a method for predicting the cost-effectiveness of solid rocket propulsion based on coordinate descent depth integration. Background Technology

[0002] Solid rocket engines play a crucial role in various aerospace missions due to their advantages such as simple structure, fast response speed, and long storage period. However, with the development of commercial spaceflight and the tightening of defense budgets, "low cost" has become a design metric as important as "high performance." As the core component of such aircraft, the cost of solid rocket engines accounts for a significant proportion of the total cost (typically exceeding 30%-50%).

[0003] Existing cost estimation and performance evaluation of solid rocket motors face the following serious challenges:

[0004] 1. Limitations of traditional cost estimation models.

[0005] Currently, the engineering community mainly uses two types of cost estimation methods:

[0006] Parametric methods: Classical TRANSCOST models or NASA's PCEC models typically establish power function relationships based on weight or total impact. While computationally simple, this method completely ignores the enormous impact of geometric complexity on manufacturing costs. For example, a simple cylindrical propellant charge and an extremely complex 3D wing-shaped propellant charge, even with the same weight, will have exponentially higher manufacturing costs due to the latter's core mold design, demolding process, difficulty in non-destructive testing, and scrap rate. Traditional models cannot perceive this "cost imposed by topology."

[0007] Bottom-up approach: This method requires detailed time and material calculations for each process (mixing, casting, curing, machining). While accurate, it is extremely time-consuming and requires detailed design drawings, making it unsuitable for conceptual design phases lacking detail and difficult to support rapid iteration and optimization of thousands of solutions.

[0008] 2. Non-linear coupling between performance and cost.

[0009] The relationship between an aircraft's performance (such as range and payload) and its cost is not a simple linear one. To increase specific impulse by 1%, it may be necessary to use expensive high-energy propellants or extremely complex propellant surface shapes, resulting in a 20% increase in cost. This strongly nonlinear phenomenon of "diminishing marginal efficiency" is often simplified in traditional multidisciplinary optimization, leading to aircraft designs that, while possessing advanced performance indicators, lose market competitiveness due to manufacturing complexity and low yield rates.

[0010] 3. The "small sample" dilemma of data-driven methods.

[0011] In recent years, machine learning has been introduced into aerospace cost prediction. However, unlike image recognition or natural language processing, historical data on aircraft model development is extremely scarce. A company may have less than a hundred examples of mature model data accumulated over decades. Directly using deep neural networks can easily lead to overfitting, causing the model to "remember" the costs of specific historical models and fail to generalize to new configuration designs. Furthermore, traditional gradient boosting tree models produce step-like outputs that are not smooth enough, making them difficult to use for gradient-based reverse engineering optimization. Summary of the Invention

[0012] The purpose of this invention is to provide a solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration, in order to solve the problems of existing solid rocket engines in the design stage, such as the difficulty in quantifying manufacturing process costs due to the complex geometry of the propellant grain, the low accuracy of traditional weight-based parameter estimation, and the strong nonlinearity of the "performance-cost" coupling relationship and the scarcity of historical sample data.

[0013] To achieve the above objectives, the present invention employs the following technical solution:

[0014] A method for predicting the cost-effectiveness of solid rocket propulsion based on coordinate descent depth integration includes:

[0015] Constructing a multi-source fusion database using the characteristic parameters of solid rocket motors;

[0016] Construct an integrated sparse neural network; the integrated sparse neural network includes an additive model, which is constructed from base learners and corresponding weight coefficients;

[0017] Based on a multi-source fusion database, global recurrent parameter training based on coordinate descent is performed on the integrated sparse neural network. When updating the current base learner in each loop, the contributions of other base learners are fixed, the global residual of the integrated sparse neural network is calculated, and the current base learner is retrained to minimize the global residual.

[0018] In the training of the individual base learners of the integrated sparse neural network, a system is constructed that includes a mean squared error term and... The objective function of the norm regularization term; Norm regularization terms are used to impose row sparsity constraints on the weight matrix from the input layer to the hidden layer of the base learner;

[0019] For a new design scheme for solid rocket motors, the characteristic parameters of the solid rocket motor are obtained and the input features are constructed. Then, the input features are fed into a trained integrated sparse neural network to obtain the predicted output response, which is used to estimate the cost or efficiency of the solid rocket motor propulsion system.

[0020] Furthermore, the multi-source data fusion library includes an input feature space and an output response space;

[0021] Each sample in the input feature space contains multidimensional feature parameters; these feature parameters include geometric design parameters, process parameters, and performance parameters.

[0022] The geometric design parameters include the geometric characteristics of the solid rocket motor, which include at least one of the following: the ratio of propellant combustion surface area to volume, the nozzle expansion ratio, the mandrel draft angle, the propellant tip chamfer, the number of star points, the propellant length-to-diameter ratio, the propellant outer diameter, the star point angle, the star edge radius, and the volumetric filling fraction. The process parameters include at least one of the following: propellant formulation type, shell material properties, number of mandrel blocks, insulation layer laying method, insulation layer manual laying time factor, and nozzle vector control mechanism type. The performance parameters include at least one of the following: design total thrust, maximum thrust, and average pressure.

[0023] The output response space contains the output response corresponding to each sample; the output response includes cost or performance indicators of the solid rocket motor propulsion system.

[0024] Furthermore, the additive model is represented as follows:

[0025] ;

[0026] in, These are the predicted cost or performance indicators; Indicates the first Individual base learners, Indicates a sample, For the first The weights of each base learner The number of base learners;

[0027] The base learner is a single-hidden-layer feedforward neural network with a hyperbolic tangent function or a sigmoid activation function, and its output layer uses linear activation.

[0028] Further, global recurrent parameter training based on coordinate descent is performed on the integrated sparse neural network, including:

[0029] First, the parameters and weights of each base learner are initialized using a forward step-by-step algorithm; then the following steps are performed:

[0030] Set the maximum number of loops ; in the In the next iteration, all base learners are traversed sequentially.

[0031] For the current number Each base learner is used to calculate the global residual of the ensembled sparse neural network after removing the contributions of that base learner.

[0032] ;

[0033] in, For the first in the multi-source fusion database Sample In the output response space The corresponding output response, and The first The updated version of the first iteration in the round of iteration Each base learner and its weights; ; , The number of base learners;

[0034] Retraining the first using a multi-source fusion database Each base learner is used to fit the global residual. ;

[0035] Update the first step using a single-step Newton's method or a linear search. The weights of each base learner;

[0036] The first iteration is accepted only if the total loss of the updated ensemble sparse neural network is less than the total loss of the previous iteration. Parameter updates for each base learner.

[0037] Furthermore, the term includes a mean squared error term and The objective function of the norm regularization term is expressed as follows:

[0038] ;

[0039] in This is the mean square error term. for Norm regularization term; It is the objective function; For the first The weight matrix from the input layer to the hidden layer of each base learner; Indicates the first One sample; It is the current global residual; The number of samples; For the first The weights of each base learner; The weight matrix is The first time Individual base learners; It is the regularization coefficient; for of Norm, represented as follows:

[0040] ;

[0041] in, For the first The feature dimensions of the input layer of each base learner This represents the number of nodes in the hidden layer. Indicates sample The The first feature parameter and the hidden layer's first feature parameter Connection weights between neurons , .

[0042] Furthermore, during the training of the ensemble sparse neural network, the LM algorithm is used to solve for the network weights; the updated network weights... It is expressed as follows:

[0043] ;

[0044] in Let be the Jacobian matrix of the error vector with respect to the network weights; This is the error vector; This is the adaptive damping factor; The sparse regularization coefficient; It is the identity matrix; It is a diagonal matrix with diagonal elements as follows: Corresponding weight row The reciprocal of the norm; superscript Indicates transpose; The network weights before the update.

[0045] Furthermore, the LM algorithm includes an adaptive adjustment strategy: when the mean square error term decreases, the adaptive damping factor is reduced; when the mean square error term increases, the adaptive damping factor is increased.

[0046] 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 solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration.

[0047] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration.

[0048] Compared with the prior art, the present invention has the following technical features:

[0049] 1. Strong resistance to overfitting and good applicability to small samples.

[0050] pass Sparse regularization successfully solves the problem in the aerospace field where high-dimensional neural network models are prone to overfitting under limited historical model data, leading to distorted cost predictions for new design schemes. It enables the model to provide reliable predictions even with only a small amount of historical data.

[0051] 2. Implement the mapping between geometric complexity and cost.

[0052] This addresses the problem that traditional parametric methods cannot quantify the impact of propellant micro-geometry (such as star angle number and star angle) on manufacturing process costs, and resolves the nonlinear coupling problem between performance and cost in existing technologies.

[0053] 3. Design-cost-performance closed loop.

[0054] The smoothness and high computational efficiency (millisecond-level prediction) of additive models allow them to be embedded in genetic algorithms or particle swarm optimization as part of the fitness function, achieving multi-objective Pareto front optimization with "maximum performance and minimum cost".

[0055] 4. The algorithm has a fast convergence speed and high accuracy.

[0056] This addresses the problem of traditional ensemble learning algorithms, which employ a forward-greedy strategy and are prone to getting trapped in local optima, failing to balance the weights of base learners globally and thus limiting prediction accuracy. Coordinate descent avoids the local optima problem of traditional Boosting, while the LM algorithm ensures rapid convergence of neural network training and high prediction accuracy. Attached Figure Description

[0057] Figure 1 This is a schematic flowchart of the method of the present invention;

[0058] Figure 2 The figure shows the ablation experiment results for the number of base learners in an embodiment of the present invention;

[0059] Figure 3 The figure shows the experimental results of the sensitivity of the sparse regularization coefficient in the embodiments of the present invention. Detailed Implementation

[0060] This invention provides a method for predicting the cost and performance of solid rocket propulsion based on coordinate descent depth integration. This method constructs an integrated sparse neural network based on coordinate descent and fundamentally optimizes the base learner and optimization strategy. This enables full life-cycle cost estimation and performance evaluation of the propulsion systems of solid rocket engines (including hypersonic test vehicles, tactical UAV boosters, sounding rockets, etc.). This invention can be applied to scheme comparison, cost-performance trade-off analysis, and cost-oriented design (DFC) processes during the solid rocket engine design phase. See also... Figure 1 The method of the present invention includes the following steps:

[0061] Step 1: Construct a multi-source fusion database using the characteristic parameters of solid rocket motors.

[0062] The multi-source data fusion library includes the input feature space. With output response space ;in:

[0063] Input feature space It consists of samples, each containing multidimensional feature parameters; these feature parameters include geometric design parameters, process parameters, and performance parameters.

[0064] The geometric design parameters include the geometric characteristics of the solid rocket motor, which include at least one of the following: the ratio of propellant combustion surface area to volume, the nozzle expansion ratio, the mandrel draft angle, the propellant tip chamfer, the number of star-shaped propellant points, the propellant length-to-diameter ratio, the propellant outer diameter, the star-shaped propellant angle, the star-shaped corner radius, and the volumetric filling fraction. The process parameters include at least one of the following: propellant formulation type, shell material properties, number of mandrel blocks, insulation layer laying method, insulation layer manual laying time coefficient, and nozzle vector control mechanism type. The performance parameters include at least one of the following: design total thrust, maximum thrust, and average pressure. In this invention, the propellant grain of the solid rocket motor is a star-shaped propellant grain.

[0065] Output response space It contains the output response corresponding to each sample; the output response includes the cost index or performance index of the propulsion system (solid rocket engine); the cost index includes normalized development cost and normalized unit production cost; the performance index adopts the cost performance index (CPI).

[0066] Specifically, the aforementioned geometric design parameters, process parameters, and cost indicators can be collected from various existing solid rocket motor models, and the sample size can be increased by expanding the collected data.

[0067] Step 2, construct an integrated sparse neural network; the integrated sparse neural network includes an additive model, which is constructed from base learners and corresponding weight coefficients.

[0068] The additive model described in this scheme is specifically represented as follows:

[0069] ;

[0070] in, These are the predicted cost or performance indicators; Indicates the first Individual base learners, Indicates a sample, For the first The weights of each base learner The number of base learners; in this scheme, the base learners are single-hidden-layer feedforward neural networks (SLFNs) with hyperbolic tangent function or sigmoid activation function, and their output layers use linear activation; compared with decision trees, SLFNs have the characteristic of continuous differentiability, and can generate smooth cost response surfaces, which are more in line with the continuity of physical and economic laws.

[0071] The additive model described above is used to fit the highly nonlinear coupling relationship between the sample and the corresponding output response, overcoming the shortcomings of traditional power function cost estimation formulas (CERs) in failing to capture the impact of topological mutations on working hours.

[0072] Step 3: Based on the multi-source fusion database, perform global recurrent parameter training on the integrated sparse neural network using the coordinate descent method; when updating the current base learner in each loop, fix the contributions of other base learners, calculate the global residual of the integrated sparse neural network, and retrain the current base learner to minimize the global residual.

[0073] This invention employs the coordinate descent method, treating the integrated sparse neural network as a whole containing all base learners and weights; the specific process of this step is as follows:

[0074] First, the parameters and weights of each base learner are initialized using a forward step-by-step algorithm; then the following steps are performed:

[0075] Step 3.1, Set the maximum number of loops. ; in the In the next iteration, all base learners are traversed sequentially.

[0076] Step 3.2, for the current... Each base learner is used to calculate the global residual of the ensembled sparse neural network after removing the contributions of that base learner.

[0077] ;

[0078] in, , For the first in the multi-source fusion database Sample In the output response space The corresponding output response, and The first The updated version of the first iteration in the round of iteration Each base learner and its weights; .

[0079] The above The physical meaning of the formula is: assuming other... The base learners have done their best; the remaining errors are all passed to the current base learner. The formula ensures that the current optimization uses a base learner to process the data; the formula also ensures that the current optimization uses the first base learner. Each base learner is adjusted based on the latest state of all other base learners, thereby approximating the global optimum; this allows each base learner to find its optimal position within the global context.

[0080] Step 3.3, retrain the first generation using a multi-source fusion database. Each base learner is used to fit the global residual. .

[0081] Step 3.4, update the first step using either a single-step Newton's method or a linear search. The weights of each base learner are calculated, and then the process returns to step 3.2 for the next iteration.

[0082] Step 3.5, only if the total loss of the updated ensemble sparse neural network is... Less than the total loss of the previous round Only then did they accept the first The parameters of each base learner are updated; the total loss can be, for example, the MSE loss.

[0083] Step 4: During the training of each base learner in the integrated sparse neural network, construct a system containing a mean squared error term and... The objective function of the norm regularization term; The norm regularization term is used to impose row sparsity constraints on the weight matrix from the input layer to the hidden layer of the base learner.

[0084] This step makes the weight rows corresponding to geometric features that have no significant impact on manufacturing costs approach zero, thereby enabling automatic screening and dimensionality reduction of key cost drivers with a small sample size.

[0085] The objective function is expressed as follows:

[0086] ;

[0087] in This is the mean square error term. for Norm regularization term; It is the objective function. For the first The weight matrix from the input layer to the hidden layer of each base learner; It is the current global residual. For the sample size, For the first The weights of each base learner The weight matrix is The first time Individual base learners; It is the regularization coefficient. for of Norm, represented as follows:

[0088] ;

[0089] in, For the first The feature dimensions of the input layer of each base learner This represents the number of nodes in the hidden layer. Indicates sample The The first feature parameter and the hidden layer's first feature parameter Connection weights between neurons , .

[0090] The Norm regularization terms are used to induce weight matrices The weighted rows produce group sparsity; when a certain sample When a certain feature parameter in the input does not contribute to the output response, then... The norm regularization term compresses all weights connected to the feature parameter (i.e., the weight row corresponding to the feature parameter in the weight matrix) to 0; this design achieves automatic feature parameter selection and eliminates noise interference.

[0091] During the training of the aforementioned ensemble sparse neural network, the Levenberg-Marquardt (LM) algorithm is used to solve for the network weights (i.e., the weight parameters of each layer in the base learner). The LM algorithm uses the Jacobian matrix to approximate the Hessian matrix, and compared with the traditional gradient descent method (SGD), it converges more than 10 times faster and has higher accuracy when dealing with strongly nonlinear regression problems. The LM algorithm combines the speed of the Gauss-Newton method with the stability of the gradient descent method, making it particularly suitable for dealing with nonlinear least squares problems with regularization terms in the objective function.

[0092] Updated network weights It is expressed as follows:

[0093] ;

[0094] in Let be the Jacobian matrix of the error vector with respect to the network weights. For the error vector, For adaptive damping factor, For sparse regularization coefficients, It is the identity matrix. For origin The diagonal matrix of the norm gradient has diagonal elements as follows: Corresponding weight row The reciprocal of the norm, superscript Indicates transpose. The network weights before the update.

[0095] The LM algorithm includes an adaptive adjustment strategy: as the mean squared error term decreases, the adaptive damping factor is reduced. This makes the LM algorithm approximate the Gauss-Newton method to accelerate convergence by utilizing second-order curvature information; the adaptive damping factor is increased when the mean square error term increases. This degenerates the LM algorithm into gradient descent to ensure convergence stability.

[0096] Step 5: For the new design scheme of solid rocket engine, obtain the characteristic parameters of solid rocket engine and construct input features. Input these features into the trained integrated sparse neural network to obtain the predicted output response, that is, the predicted cost index or efficiency index, so as to realize the rapid cost or efficiency estimation of solid rocket engine.

[0097] The form of the input features is consistent with the form of the samples in the input feature space.

[0098] Based on the predicted cost or performance indicators, SHAP value analysis can be further combined to identify key cost drivers in the input features. This will guide designers to reduce manufacturing costs by modifying geometric features (such as reducing the number of star points and increasing the radius of the star edge rounded corners) without significantly reducing performance, thereby optimizing the new design scheme.

[0099] Example:

[0100] This embodiment describes the design of a solid rocket engine for a novel aircraft, requiring the lowest possible manufacturing cost while meeting the total stroke target.

[0101] In constructing the multi-source data fusion library, this embodiment collected characteristic parameters of 40 different models of solid rocket motors from the prior art; at the same time, it used parametric CAD software and industrial engineering estimation models to generate 500 sets of high-fidelity virtual samples to expand the multi-source data fusion library.

[0102] The feature parameters contained in the samples of the multi-source data fusion library in this embodiment are:

[0103] The parameters include: length-to-diameter ratio of the propellant, nozzle expansion ratio, chamfer at the front end of the propellant, outer diameter of the propellant, number of star points, star point angle, radius of the star edge rounded corner, volume filling fraction, number of core mold blocks, labor time coefficient for manual laying of the insulation layer, total design pressure, and average pressure.

[0104] In this embodiment, the number of hidden layer nodes in the single-hidden-layer feedforward neural network is 20; the activation function is the hyperbolic tangent function, and the output layer is linearly activated; the number of base learners... Maximum number of loops Regularization coefficient Damping factor The initial value is 0.01. As the mean square error term decreases, the damping factor decreases to The mean square error term and the damping factor during the rise increase to .

[0105] In this embodiment, when training the integrated sparse neural network, the weight rows corresponding to the two feature parameters, nozzle expansion ratio and propellant tip chamfer, are set to 0 in the weight matrix; this indicates that these two feature parameters have minimal impact on the output response.

[0106] On the validation set not used for training, the cost metric predicted by the integrated sparse neural network of this invention has an error of 7.8%, which is a 35% improvement compared to the traditional weight-based CER model. Reason for this: Traditional models cannot recognize the surge in mold costs caused by the new propellant grain structure used in this novel aircraft; while this invention… The norm regularization term successfully captured the linear impact of the characteristic parameters of the drug grain structure on the cost metric. Furthermore, see Tables 1 and 2 for performance tests on various existing datasets for different metrics. In Tables 1 and 2, bold text indicates the best experimental results, and underlined text indicates the second-best.

[0107] Table 1: Comparison of the test results of the present invention and existing methods on different datasets for classification tasks (average of 5 results).

[0108]

[0109] Table 2: Comparison of the test results of the present invention and existing methods on different datasets for regression tasks (average of 5 results).

[0110]

[0111] like Figure 2 and Figure 3 The model ablation experiments and sensitivity experiments verified that the parameters of the above-mentioned integrated sparse neural network are reasonable.

[0112] Using the integrated sparse neural network trained according to this invention, a solid rocket motor can be reverse-engineered:

[0113] Original design scheme: 7-star angle propellant charge, total impact meets requirements, estimated cost 1.2 million yuan.

[0114] Optimized design scheme: Through optimization, it was found that if the design is changed to a 5-star angle and the propellant length is slightly increased, although the specific impulse will decrease slightly (requiring an additional 5 kg of propellant), the complexity of core mold manufacturing and the scrap rate will be significantly reduced.

[0115] Final result: Based on the predicted cost indicators, the calculated cost was reduced by 21%, and the total design impact in the performance parameters still met the requirements, thus successfully achieving a low-cost design.

[0116] 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 method for predicting the cost-effectiveness of solid rocket propulsion based on coordinate descent depth integration, characterized in that, include: Constructing a multi-source fusion database using the characteristic parameters of solid rocket motors; Constructing an integrated sparse neural network; The integrated sparse neural network includes an additive model, which is constructed from base learners and corresponding weight coefficients; Based on a multi-source fusion database, global recurrent parameter training based on coordinate descent is performed on the integrated sparse neural network; In each iteration, when updating the current base learner, the contributions of other base learners are fixed, the global residual of the ensemble sparse neural network is calculated, and the current base learner is retrained to minimize the global residual. In the training of the individual base learners of the integrated sparse neural network, a system is constructed that includes a mean squared error term and... The objective function of the norm regularization term; the Norm regularization terms are used to impose row sparsity constraints on the weight matrix from the input layer to the hidden layer of the base learner; For a new design scheme for solid rocket motors, after obtaining the characteristic parameters of solid rocket motors and constructing input features, the input is fed into a trained integrated sparse neural network to obtain the predicted output response, which is used to estimate the cost or efficiency of the solid rocket motor propulsion system. The multi-source data fusion library includes an input feature space and an output response space; Each sample in the input feature space contains multidimensional feature parameters; The characteristic parameters include geometric design parameters, process parameters, and performance parameters; The geometric design parameters include the geometric characteristics of the solid rocket motor, which include at least one of the following: the ratio of propellant combustion surface area to volume, the nozzle expansion ratio, the mandrel draft angle, the propellant tip chamfer, the number of star points, the propellant length-to-diameter ratio, the propellant outer diameter, the star point angle, the star edge fillet radius, and the volume fraction. The output response space contains the output response corresponding to each sample; the output response includes cost or performance indicators of the solid rocket motor propulsion system.

2. The solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration according to claim 1, characterized in that, The process parameters include at least one of the following: propellant formulation type, shell material properties, number of core mold blocks, insulation layer laying method, insulation layer manual laying time coefficient, and nozzle vector control mechanism type; the performance parameters include at least one of the following: design total thrust, maximum thrust, and average pressure.

3. The solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration according to claim 1, characterized in that, The additive model is represented as follows: ； in, These are the predicted cost or performance indicators; Indicates the first Individual base learners, Indicates a sample, For the first The weights of each base learner The number of base learners; The base learner is a single-hidden-layer feedforward neural network with a hyperbolic tangent function or a sigmoid activation function, and its output layer uses linear activation.

4. The solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration according to claim 1, characterized in that, Performing global recurrent parameter training based on coordinate descent on the integrated sparse neural network includes: First, the parameters and weights of each base learner are initialized using a forward step-by-step algorithm; then the following steps are performed: Set the maximum number of loops ; in the In the next iteration, all base learners are traversed sequentially. For the current number Each base learner is used to calculate the global residual of the ensembled sparse neural network after removing the contributions of that base learner. ； in, For the first in the multi-source fusion database Sample In the output response space The corresponding output response, and The first The updated version of the first iteration in the round of iteration Each base learner and its weights; ; , The number of base learners; Retraining the first using a multi-source fusion database Each base learner is used to fit the global residual. ; Update the first step using a single-step Newton's method or a linear search. The weights of each base learner; The first iteration is accepted only if the total loss of the updated ensemble sparse neural network is less than the total loss of the previous iteration. Parameter updates for each base learner.

5. The solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration according to claim 1, characterized in that, The term includes mean square error and The objective function of the norm regularization term is expressed as follows: ； in This is the mean square error term. for Norm regularization term; It is the objective function; Indicates the first One sample; For the first The weight matrix from the input layer to the hidden layer of each base learner; It is the current global residual; The number of samples; For the first The weights of each base learner; The weight matrix is The first time Individual base learners; It is the regularization coefficient; for of Norms are represented as follows: ； in, For the first The feature dimensions of the input layer of each base learner This represents the number of nodes in the hidden layer. Indicates sample The The first feature parameter and the hidden layer's first feature parameter Connection weights between neurons , .

6. The solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration according to claim 1, characterized in that, During the training of the ensemble sparse neural network, the LM algorithm is used to solve for the network weights; the updated network weights... It is expressed as follows: ； in Let be the Jacobian matrix of the error vector with respect to the network weights; This is the error vector; This is the adaptive damping factor; The sparse regularization coefficient; It is the identity matrix; It is a diagonal matrix with diagonal elements as follows: Corresponding weight row The reciprocal of the norm; superscript Indicates transpose; The network weights before the update.

7. The solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration according to claim 6, characterized in that, The LM algorithm includes an adaptive adjustment strategy: when the mean square error term decreases, the adaptive damping factor is reduced; when the mean square error term increases, the adaptive damping factor is increased.

8. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration as described in any one of claims 1-7.

9. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the solid rocket propulsion cost-effectiveness prediction method based on coordinate descent depth integration as described in any one of claims 1-7.

Images