Solid engine thrust prediction method based on spatial-temporal feature fusion

Abstract

The invention provides a solid engine thrust prediction method based on spatio-temporal feature fusion, and relates to the field of solid engine thrust prediction, and the method comprises the steps: S1, collecting the structure parameters and working condition data of a plurality of groups of solid engines, and carrying out the combined thrust simulation of the plurality of groups of solid engines, and obtaining a training data set; s2, performing fusion modeling through a one-dimensional convolutional neural network, a long-short term memory network, an attention mechanism module and a full-connection neural network to obtain an initial thrust prediction model; s3, training the initial thrust prediction model through the training data set to obtain a final thrust prediction model; s4, the structural parameters and working condition data of the solid engine to be tested are subjected to combined thrust simulation and then input into the final thrust prediction model, and a thrust prediction value of the solid engine to be tested is obtained. The final thrust prediction model can process mixed data in a curve form and a discrete form, and an accurate thrust prediction value is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of solid engine thrust prediction, and particularly relates to a solid engine thrust prediction method based on space-time feature fusion. BACKGROUND

[0002] As a core component of spacecraft propulsion system, solid engine is widely used in missile, satellite launch, emergency escape system and other high reliability demand scenarios. The thrust performance of solid engine directly affects the flight safety and mission completion quality of the whole platform, so accurate modeling and prediction of engine thrust characteristics are of great significance.

[0003] However, the actual test process of thrust often depends on high-precision ground test equipment, which not only has high cost and long test cycle, but also has certain test safety risk. Therefore, generating thrust curve based on simulation platform becomes an important alternative. However, due to the complexity and high nonlinearity of the combustion process, the traditional simulation method still needs complex parameter configuration and fine grid division, resulting in large consumption of computing resources and long calculation time in the simulation process. In addition, in the modeling process of modern engine, in order to truly describe its multi-physical coupling characteristics, the input parameters are no longer simple scalars or fixed vectors, but mixed forms of spatial variation curves (such as burning area-thickness, cavity volume-thickness), time series data (such as nozzle opening change with time) and discrete working condition parameters (such as temperature, characteristic velocity) and the like. Such heterogeneous data structure is complex and the dimensions are not uniform, which cannot be directly input into the traditional neural network model for training and prediction. SUMMARY

[0004] In order to solve the above problems, the present application provides a solid engine thrust prediction method based on space-time feature fusion, comprising the steps of:

[0005] S1: collecting multiple sets of structure parameters and working condition data of solid engines, performing joint thrust simulation on the multiple sets of solid engines, and obtaining a training data set;

[0006] S2: performing fusion modeling through one-dimensional convolutional neural network, long short-term memory network, attention mechanism module and fully connected neural network to obtain an initial thrust prediction model;

[0007] S3: training the initial thrust prediction model through the training data set to obtain a final thrust prediction model;

[0008] S4: inputting the structure parameters and working condition data of the solid engine to be tested into the final thrust prediction model after joint thrust simulation to obtain the thrust prediction value of the solid engine to be tested.

[0009] Preferably, step S1 is specifically:

[0010] S11: Setting simulation parameters of thrust simulation, including simulation time, step size and output interval;

[0011] S12: Constructing a thrust simulation model based on simulation parameters;

[0012] S13: Inputting multiple sets of solid motor structural parameters and operating condition data into the thrust simulation model to obtain a set of solid motor combustion surface area versus wall thickness variation curves, a set of cavity volume versus wall thickness variation curves, a set of nozzle opening versus time variation curves, a set of discrete variables, and a thrust time series variation curve;

[0013] S14: Divide the curve set of combustion surface area changing with wall thickness, the curve set of cavity volume changing with wall thickness, the curve set of nozzle opening changing with time, and the discrete variable set into a training data set, a validation data set, and a test data set.

[0014] Preferred:

[0015] The input form of the curve of burning surface area changing with wall thickness is:

[0016] [e1,e2,e3....e n ],[Ab1,Ab2,Ab3....Ab n ]

[0017] Among them, e1, e2, e3....e n Indicates the thickness value of each meat, Ab1, Ab2, Ab3....Ab n Indicates the burning surface area corresponding to each thickness value, and n represents the total number of data in the vector;

[0018] The input form of the cavity volume versus wall thickness curve is:

[0019] [e1,e2,e3....e n ],[Vc1,Vc2,Vc3....Vc n ]

[0020] Among them, Vc1, Vc2, Vc3....Vc n Indicates the cavity volume corresponding to each wall thickness value.

[0021] Preferred:

[0022] The input form of the nozzle opening variation curve over time is:

[0023] [t1,t2,t3....t n ],[G1,G2,G3....G n ]

[0024] Among them, t1, t2, t3....t n Indicates each moment, G1, G2, G3....G n Represents the nozzle opening value corresponding to each moment, and n represents the total number of data in the vector.

[0025] Preferred:

[0026] The one-dimensional convolutional neural network includes the first convolutional layer, the maximum pooling layer, the second convolutional layer, the adaptive pooling layer, the flattening layer and the fully connected layer connected in sequence.

[0027] Preferably, step S3 is specifically as follows:

[0028] S31: Obtain a set of curves of burning surface area changing with wall thickness, curves of cavity volume changing with wall thickness, curves of nozzle opening changing with time, and discrete variables from the training data set;

[0029] S32: Input the curve of the burning surface area changing with the thickness of the wall and the curve of the cavity volume changing with the thickness of the wall into a one-dimensional convolutional neural network for feature extraction to obtain the burning surface space feature vector and the cavity space feature vector respectively;

[0030] S33: Inputting the nozzle opening variation curve over time into the long short-term memory network for feature extraction to obtain the opening time series feature vector;

[0031] S34: The attention mechanism module is used to fuse the burning surface spatial feature vector, the cavity spatial feature vector, the opening time series feature vector, and the discrete variable to obtain a spatiotemporal feature fusion vector.

[0032] S35: Input the spatiotemporal feature fusion vector into the fully connected neural network to obtain the thrust prediction value;

[0033] S36: Obtain the actual thrust value, and calculate the mean square error loss between the predicted thrust value and the actual thrust value;

[0034] S37: Adjust the parameter set of the one-dimensional convolutional neural network according to the mean square error loss value through the NAS mechanism, and calculate the reward function of the NAS mechanism;

[0035] S38: Backpropagation is performed through the mean squared error loss value to adjust the parameter sets of the long short-term memory network and the fully connected neural network;

[0036] S39: Repeat steps S31-S38 until the mean square error loss value converges and the reward function is greater than a preset threshold, and obtain the final thrust prediction model.

[0037] Preferred:

[0038] The expression of thrust prediction value is:

[0039] y i =σ(W m ReLU(W n ·x+b1)+b2)

[0040] Among them, x is the spatiotemporal feature fusion vector, W m and W n are the first and second weight matrices of the fully connected neural network, b1 and b2 are the first and second bias terms, σ is the activation function, i is the iteration number, y i is the thrust prediction value of the i-th iteration.

[0041] The present invention provides a solid rocket thrust prediction device based on spatiotemporal feature fusion, which is implemented using the solid rocket thrust prediction method based on spatiotemporal feature fusion. The device includes:

[0042] The simulation module is used to collect the structural parameters and operating condition data of multiple sets of solid motors, perform joint thrust simulation on multiple sets of solid motors, and obtain a training data set;

[0043] The thrust prediction model building module is used to obtain the initial thrust prediction model through fusion modeling of a one-dimensional convolutional neural network, a long short-term memory network, an attention mechanism module, and a fully connected neural network;

[0044] A training module, configured to train the initial thrust prediction model using a training data set to obtain a final thrust prediction model;

[0045] The thrust prediction value acquisition module is used to input the structural parameters and working condition data of the solid engine to be tested into the final thrust prediction model to obtain the thrust prediction value of the solid engine to be tested.

[0046] The present invention provides an electronic device comprising a memory, a processor and a computer program stored in the memory and operable on the processor. When the processor executes the program, the solid engine thrust prediction method based on spatiotemporal feature fusion is implemented.

[0047] The present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the solid motor thrust prediction method based on spatiotemporal feature fusion is implemented.

[0048] The present invention has the following beneficial effects:

[0049] 1. The curves of the solid rocket's combustion surface area varying with its thickness, the cavity volume varying with its thickness, the nozzle opening varying with time, and discrete variables are used as training data. These three sets of curves together determine the time evolution characteristics of the solid rocket's thrust, making the training data close to the actual use status of the solid rocket.

[0050] 2. An initial thrust prediction model is constructed by combining a one-dimensional convolutional neural network, a long short-term memory network, an attention mechanism module, and a fully connected neural network. The one-dimensional convolutional neural network and the long short-term memory network extract the features of the spatial variation curve and the temporal variation curve, respectively. The attention mechanism module fuses the features of the spatial variation curve and the temporal variation curve, allowing the thrust prediction model to analyze the spatiotemporal fusion features and improve the accuracy of the thrust prediction value.

[0051] 3. The initial thrust prediction model is iteratively trained using the mean squared error loss and the NAS mechanism. Backpropagation and optimization are continued until the training conditions are met. Sufficient training significantly improves the prediction accuracy of the final thrust prediction model.

[0052] 4. The final thrust prediction model can process mixed data in curve form and discrete form to obtain accurate thrust prediction values. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method according to an embodiment of the present invention;

[0054] Figure 2 This is an example graph of the nozzle opening changing with time;

[0055] Figure 3 This is the structural diagram of a one-dimensional convolutional neural network;

[0056] Figure 4 Schematic diagram of the training process of the thrust prediction model;

[0057] Figure 5 This is a visualization result diagram of the prediction accuracy of the wing column type;

[0058] Figure 6 This is a visualization result diagram of the prediction accuracy of the segmented inner hole;

[0059] Figure 7 This is a visualization result diagram of the prediction accuracy of end combustion;

[0060] Figure 8 Visualization result diagram of prediction accuracy of star hole charge;

[0061] Figure 9 Visualize the prediction accuracy for unknown charge types;

[0062] Figure 10 This is a structural diagram of a solid rocket thrust prediction device based on spatiotemporal feature fusion;

[0063] Figure 11 A structural diagram of an electronic device;

[0064] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0065] The following will be combined with the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0066] Reference Figure 1 The present invention provides a solid rocket thrust prediction method based on spatiotemporal feature fusion, comprising the steps of:

[0067] S1: Collect the structural parameters and operating condition data of multiple sets of solid rocket motors, perform joint thrust simulation on multiple sets of solid rocket motors, and obtain a training data set;

[0068] In some embodiments, during the solid rocket thrust modeling and simulation process, the thrust output is mainly determined by the expansion process of the gas and the exhaust characteristics of the nozzle, which is essentially controlled by the mass flow rate, combustion chamber pressure, and nozzle throat area parameters. According to the basic thrust calculation formula:

[0069]

[0070] Among them, F represents thrust, represents the mass flow rate, v represents the gas exhaust velocity, A e represents the nozzle outlet area, p e , p a represent the combustion chamber pressure and the external environment pressure respectively.

[0071] Step S1 is specifically as follows:

[0072] S11: Setting simulation parameters of thrust simulation, including simulation time, step size and output interval;

[0073] In some embodiments, simulation-related parameters are read from an input file, including simulation time, step size, and output interval. These parameters affect the accuracy and computation time of the simulation runtime.

[0074] S12: Constructing a thrust simulation model based on simulation parameters;

[0075] In some embodiments, before simulation begins, the AMESim model is checked for correctness using AMECirChecker to ensure that the model file is correct and meets the simulation requirements. If the check is correct, the simulation model is loaded using AMELoad.

[0076] S13: Inputting multiple sets of solid motor structural parameters and operating condition data into the thrust simulation model to obtain a set of solid motor combustion surface area versus wall thickness variation curves, a set of cavity volume versus wall thickness variation curves, a set of nozzle opening versus time variation curves, a set of discrete variables, and a thrust time series variation curve;

[0077] In some embodiments, multiple sets of structural parameters and operating condition data of solid motors are input into a combustion surface calculation module to obtain a set of curves of the solid motor's combustion surface area varying with thickness, a set of curves of the cavity volume varying with thickness, a set of curves of the nozzle opening varying with time, and a set of discrete variables. Subsequently, the set of curves of the combustion surface area varying with thickness, the set of curves of the cavity volume varying with thickness, the set of curves of the nozzle opening varying with time, and the set of discrete variables are input into a thrust simulation model for joint simulation to obtain a thrust time series curve.

[0078] The simulation process is as follows:

[0079] 1. Input Data Preparation: Reading and writing input data. The program reads the input data required for the simulation (such as various parameter values ​​and initial states) from an external data source and writes this data into the simulation model. External data includes input curves for the combustion surface, cavity, and nozzle opening, as well as discrete variables, obtained through the combustion surface calculation platform. If results are available from a previous simulation, the program reads the outputs from that simulation and updates certain input parameters based on those outputs, ensuring that the simulation model uses the latest parameter configuration each time it is run.

[0080] 2. Simulation Execution: After preparing the simulation model and input parameters, use `amerunsingle` to start the simulation. This calls AMESim to perform the actual simulation calculations. Use `amegetsimopt` to obtain simulation options, including basic parameters such as simulation time and step size. These settings are applied during the simulation and affect the execution and results of the simulation. The simulation program will perform calculations according to the set simulation parameters, such as time and step size, until the specified simulation task is completed.

[0081] 3. Simulation Results Acquisition and Processing: After the simulation is complete, use the ameloadt function to load the simulation results. This extracts all calculated parameters and simulation result data from AMESim. Use the amegetvar function to obtain specific simulation result variables, such as thrust and displacement. These result variables are extracted as needed and formatted for subsequent analysis. Simulation results often require further processing, such as formatting and unit conversion.

[0082] 4、Results Storage and Output: The simulation results are stored in the designated output file and copied to the target directory by the program for further analysis. Depending on user requirements, the simulation results may be converted to a specific format (such as thrust.txt) and saved in the results folder.

[0083] The entire simulation process is executed multiple times in a loop, each time using different input data or modified parameters. This approach helps verify the stability and accuracy of the model under different conditions.

[0084] 5、Error Handling and Fault Tolerance: The program performs exception detection during simulation. If an error occurs, it will be captured and recorded immediately. For example, if the simulation result file is missing or the input file is incorrect, the program will prompt an error and stop execution, ensuring the stability and reliability of the simulation process. Once an error occurs, the program will record the error information in a log file (e.g., baocuo.txt) for subsequent checking and debugging.

[0085] The core idea of the entire process is to achieve the automation of simulation tasks, extraction and analysis of results through the interaction between MATLAB and AMESim. AMECirChecker, AMELoad, amerunsingle, amegetsimopt, ameloadt, and amegetvar are built-in functions in AMESim for interactive simulation.

[0086] The three sets of curves obtained from the simulation determine the time evolution characteristics of the engine thrust, so the AMESim simulation requires these three sets of curves as input, which is the basic logic of physical simulation.

[0087] The burning surface is the "interface" where the propellant contacts the flame and continues to burn in the combustion chamber. In the solid propellant combustion process, the burning rate of the propellant is directly related to the mass of the propellant burned per unit area. The larger the burning surface area, the more mass of combustion gas is produced per unit time, and the greater the mass flow rate, thus the greater the thrust. The burning surface area is not constant, it continuously degrades inward as the propellant burns, showing a decrease in "meat thickness". Therefore, the curve of the burning surface area changing with meat thickness must be input to reflect the dynamic evolution of the combustion process. For solid rocket engines of different charge types, the curve of the burning surface changing with meat thickness is quite different.

[0088] The input form of the burning surface area changing with meat thickness curve is:

[0089] [e1, e2, e3....e n ], [Ab1, Ab2, Ab3....Ab n ]

[0090] Among them, e1, e2, e3....e n Indicates the thickness value of each meat, Ab1, Ab2, Ab3....Ab n Indicates the burning surface area corresponding to each thickness value, and n represents the total number of data in the vector;

[0091] The cavity volume is the space within the combustion chamber not occupied by propellant, which gradually expands as the propellant burns. This volume directly affects the rate of change of the combustion chamber's internal pressure and the rate of gas expansion. A too small cavity will cause the pressure to rise too quickly, compromising structural safety; a too large cavity will prevent the gas from effectively performing work, resulting in insufficient thrust. Therefore, the curve of cavity volume versus wall thickness actually reflects the buffer space for combustion chamber pressure dynamics. It is a non-negligible structural variable in thrust simulation, particularly affecting pressure-time characteristics and burn rate feedback.

[0092] The input form of the cavity volume versus wall thickness curve is:

[0093] [e1,e2,e3....e n ],[Vc1,Vc2,Vc3....Vc n ]

[0094] Among them, Vc1, Vc2, Vc3....Vc n Indicates the cavity volume corresponding to each wall thickness value.

[0095] In some embodiments:

[0096] The nozzle opening changes with time curve example is as follows Figure 2 As shown, the nozzle throat is the smallest cross-sectional area that affects the acceleration of the gas, controlling the flow rate, back pressure, and critical flow state. According to fluid dynamics, the throat satisfies the following conditions Among them, the throat area A t It is one of the multiplication factors of mass flow rate and jet velocity. In the present invention, the nozzle throat is an adjustable structure, which must be input in the form of a nozzle opening variation curve over time, as time series data to drive aerodynamic changes.

[0097] The input form of the nozzle opening variation curve over time is:

[0098] [t1,t2,t3....t n ],[G1,G2,G3....G n ]

[0099] Among them, t1, t2, t3....t n Indicates each moment, G1, G2, G3....G n Represents the nozzle opening value corresponding to each moment, and n represents the total number of data in the vector.

[0100] The nozzle opening value reflects the size of the nozzle throat area, which is expressed by A t =A·G i calculate.

[0101] S14: Divide the curve set of combustion surface area changing with wall thickness, the curve set of cavity volume changing with wall thickness, the curve set of nozzle opening changing with time, and the discrete variable set into a training data set, a validation data set, and a test data set.

[0102] In some embodiments, the training dataset is used to train the model, the validation dataset is used to adjust model hyperparameters, and the test dataset is used to evaluate the performance of the model.

[0103] S2: Obtain the initial thrust prediction model through fusion modeling of a one-dimensional convolutional neural network, a long short-term memory network, an attention mechanism module, and a fully connected neural network;

[0104] In some embodiments:

[0105] In PyTorch, the torch.nn.LSTM module is used to build a long short-term memory network (LSTM). LSTM is used to process time series data and can capture long-term dependencies. Specifically, the time series data is first input into the LSTM, which uses its internal gating mechanism to filter out irrelevant information and retain useful time series features. The LSTM network can have multiple layers, and each layer processes the input time series data. In the present invention, one or more layers of LSTM structure can be selected, and the output hidden state is used as a representation of the time series features. The model performance is optimized by adjusting the number of layers and hidden units of the LSTM.

[0106] The one-dimensional convolutional neural network includes the first convolutional layer, the maximum pooling layer, the second convolutional layer, the adaptive pooling layer, the flattening layer and the fully connected layer connected in sequence.

[0107] In some embodiments, the structure of a one-dimensional convolutional neural network (1D-CNN) is as follows: Figure 3 As shown in Figure 1, 1D-CNN extracts local spatial variation patterns by sliding convolution kernels. Its basic working mechanism can be described as:

[0108] For a set of input sequences x={x1,x2,...,x n}, the convolutional layer performs the following operations by sliding the weight kernel w:

[0109]

[0110] Among them, J is the width of the convolution kernel, y iRepresents the output feature of position i. The output feature is further extracted through activation function and pooling layer to extract high-level spatial semantic information. Finally, each input curve is mapped to a fixed-length spatial feature vector.

[0111] S3: Train the initial thrust prediction model using the training data set to obtain the final thrust prediction model;

[0112] In some embodiments:

[0113] Step S3 is specifically as follows:

[0114] S31: Obtain a set of curves of burning surface area changing with wall thickness, curves of cavity volume changing with wall thickness, curves of nozzle opening changing with time, and discrete variables from the training data set;

[0115] S32: Input the curve of the burning surface area changing with the thickness of the wall and the curve of the cavity volume changing with the thickness of the wall into a one-dimensional convolutional neural network for feature extraction to obtain the burning surface space feature vector and the cavity space feature vector respectively;

[0116] In some embodiments, a one-dimensional convolutional neural network (1D-CNN) processes two types of input spatial variable curves (burning surface area-meat thickness, cavity volume-meat thickness), respectively, and the above curves are input in the form of discrete vectors of fixed length.

[0117] S33: Inputting the nozzle opening variation curve over time into the long short-term memory network for feature extraction to obtain the opening time series feature vector;

[0118] In some embodiments, after a long short-term memory (LSTM) network processes the nozzle opening time series data, the output hidden state vector is extracted as a time series feature. Since the LSTM output is typically a multidimensional tensor representing the hidden state at each time step, the hidden state at the final time step can be selected as the time series feature, or the output of all time steps can be selected as needed to obtain the extracted opening time series feature vector.

[0119] S34: The attention mechanism module is used to fuse the burning surface spatial feature vector, the cavity spatial feature vector, the opening time series feature vector, and the discrete variable to obtain a spatiotemporal feature fusion vector.

[0120] In some embodiments, an attention mechanism fusion method is used to effectively fuse spatiotemporal features in solid rocket thrust prediction. Specifically, the fusion process includes the curve of the combustion surface area changing with the wall thickness, the curve of the cavity volume changing with the wall thickness, and the curve of the nozzle opening changing with time. At the same time, the operating environment temperature, a common and highly influential discrete variable, is also taken into account. The attention mechanism is particularly suitable for processing multimodal data because it can adaptively assign different weights to each feature. By weighting the features according to their importance, it enhances the focus on important features, further improving the accuracy and predictive performance of the model.

[0121] This paper utilizes the built-in self-attention library nn.MultiheadAttention in the PyTorch framework to fuse the previously extracted spatiotemporal multimodal feature vectors. Specifically, the input data format is [sequence length, batch size, dimension of each feature]. In this model, the feature vector is processed as three elements: query (Q), key (K), and value (V). Typically, query (Q) and key (K) are derived from the input feature vector, while value (V) is the input feature that requires weighted summation.

[0122] The nn.MultiheadAttention module splits the query (Q), key (K), and value (V) into multiple "heads," allowing each head to independently compute attention weights. This mechanism parallelizes multiple attention calculations, capturing complex relationships between features from different perspectives. Ultimately, all calculated attention outputs are combined to produce a unified spatiotemporal feature fusion vector.

[0123] Furthermore, the present invention combines features from different sources (such as the attention output of spatial and temporal features) to generate a final fused feature vector, which serves as input to the subsequent prediction model for thrust prediction. This feature fusion method efficiently processes complex spatiotemporal data and significantly improves the performance of the thrust prediction model.

[0124] S35: Input the spatiotemporal feature fusion vector into the fully connected neural network to obtain the thrust prediction value;

[0125] In some embodiments:

[0126] The thrust prediction output by the fully connected neural network is based on multiple moments of thrust. The specific number of discretized moments is user-defined based on engine size, combustion characteristics, and operating conditions. For example, for a large engine with a high combustion rate, multiple moments are typically selected for discretization, while a small engine with a high combustion rate may only use a few moments due to its shorter operating time.

[0127] The expression of thrust prediction value is:

[0128] y i =σ(W m ReLU(W n ·x+b1)+b2)

[0129] Among them, x is the spatiotemporal feature fusion vector, W m and W n are the first and second weight matrices of the fully connected neural network, b1 and b2 are the first and second bias terms, σ is the activation function, i is the iteration number, y i is the thrust prediction value of the i-th iteration.

[0130] S36: Obtain the actual thrust value, and calculate the mean square error loss between the predicted thrust value and the actual thrust value;

[0131] In some embodiments, the mean squared error loss value is used in a subsequent back-propagation phase to update the parameters of the long short-term memory network and the fully connected neural network;

[0132] The expression of the mean square error loss value is:

[0133]

[0134] Among them, τ is the mean square error loss value, i is the iteration number, y i is the thrust prediction value of the i-th iteration, is the true value of the thrust at the i-th iteration, and N is the number of current iterations.

[0135] S37: Adjust the parameter set of the one-dimensional convolutional neural network according to the mean square error loss value through the NAS mechanism, and calculate the reward function of the NAS mechanism;

[0136] In some embodiments, for simple spatial curves, such as in simple charge types such as end-burning and side-burning, the curve of the change of cavity volume with the thickness is a simple linear change, and only a small number of dimensional feature vectors are needed to cover all the features of the input data set curve. Too high a dimensional feature vector will cause overfitting. For complex large-scale engine types, such as star-hole charge and wing-column charge, the curve of the change of the burning surface area with the thickness has more nonlinear and complex trends, so a higher dimensional feature vector is required to cover all the information of the data set. Too few dimensional feature vectors may not be able to capture all the information. Therefore, in the present invention, two 1DCNN network construction ideas are provided, and users can choose according to actual data requirements:

[0137] 1. Manually adjustable network structure design: Users can manually set the feature output dimension and convolutional layer depth based on the structural complexity of the sample set. When the charge type is known, manually setting the feature vector dimension based on experience and experimental results will improve computational efficiency.

[0138] 2. To adapt to the differences in characteristic curve complexity brought about by different solid motor charge structures, the present invention introduces a neural network architecture search (NAS) mechanism driven by reinforcement learning to automatically optimize and configure the one-dimensional convolutional neural network (1D-CNN) structure used for spatial curve feature extraction.

[0139] The NAS mechanism consists of three main components: 1. The search space, which includes a set of adjustable network parameters such as convolution kernel size (kernelsize∈{3,5,7}), number of channels (channels∈{32,64,128}), activation function, and pooling strategy; 2. The search strategy, which uses policy gradient-based reinforcement learning to output a new round of structural parameter combinations through a controller network based on historical structural performance; and 3. The performance evaluator, which uses thrust prediction error as a reward function and determines the current structural performance through rapid training and validation set evaluation. The core optimization goal of NAS is to maximize the reward function, that is, to minimize the prediction error.

[0140] During the search process, the controller network continuously generates structure combinations. After training for several epochs, its thrust prediction ability is rapidly evaluated on a validation set. A reward signal is then sent back based on the error value, gradually improving the performance of the generated structures. For ease of implementation, this invention utilizes existing deep learning frameworks such as PyTorch to implement a custom NAS controller. In this PyTorch implementation, a corresponding 1D-CNN network is constructed after each structure is generated. After training, the error is recorded and fed back to the controller. Ultimately, the network structure with the lowest thrust prediction error is selected as the optimal architecture.

[0141] The expression of the reward function is:

[0142] R(θ)=-τ

[0143] Among them, θ is the parameter set of the one-dimensional convolutional neural network, R(θ) is the reward function of the NAS mechanism corresponding to the parameter set θ, and τ is the mean square error loss value.

[0144] Through the NAS mechanism, the network structure can automatically adjust its complexity according to the characteristics of different charges, thereby improving computational efficiency and generalization ability while ensuring accuracy, and significantly improving the model's adaptability to different data sets.

[0145] S38: Backpropagation is performed through the mean squared error loss value to adjust the parameter sets of the long short-term memory network and the fully connected neural network;

[0146] In some embodiments, the present invention uses an Adam optimizer to update network parameters;

[0147] S39: Repeat steps S31-S38 until the mean square error loss value converges and the reward function is greater than a preset threshold, and obtain the final thrust prediction model.

[0148] In some embodiments, the training process of the initial thrust prediction model is as follows: Figure 4 As shown in the figure, the backpropagation and optimization process will continue until the stopping condition is met. Because the thrust value itself is large in magnitude and includes outputs at multiple moments, the value of the loss function will change with the magnitude of the thrust. To more effectively train the network, the value of the loss function will gradually stabilize during the optimization process, ultimately reaching the optimal model parameters.

[0149] S4: After the structural parameters and operating condition data of the solid motor to be tested are subjected to joint thrust simulation, the final thrust prediction model is input to obtain the thrust prediction value of the solid motor to be tested.

[0150] In some embodiments:

[0151] The final thrust prediction model can predict thrust for two types of solid rocket motors: single charge and mixed charge. The two types of solid rocket motors are respectively suitable for thrust prediction tasks under known or unknown charge conditions. In order to comprehensively evaluate the prediction performance of the final thrust prediction model constructed by this invention, a dual method of global error calculation and thrust curve visualization comparison is used to determine the accuracy. The global error calculation formula is:

[0152]

[0153] in, is the thrust prediction value at each moment, F t is the actual thrust value at each moment, and t is the total number of time steps in the thrust curve. The global error is used to quantify the overall fitting ability, and visualization is used to show the degree of fit between the predicted curve and the actual curve, allowing users to intuitively judge model performance.

[0154] 1. Single charge model

[0155] This model is applicable to the case where the engine charge type and approximate size are known. For typical charge structures, the present invention selected four common solid engine charge types for testing and verification, including: wing column type charge, star hole charge, segmented inner hole charge and end burning charge. Under the premise of known charge form, the corresponding characteristic curve data can be collected through step S1, and the integrated simulation method described in step S2 can be used to construct a training data set, and the thrust curve can be predicted in combination with the neural network model trained in step S3. For charge types that have not yet been covered (such as wheel-shaped charges, stepped charges and other uncommon structures), the present invention can still be adapted and predicted through the same process, and the specific process is consistent with that described in S2 and S3 for different charge types. For the four common charge types covered in the present invention, each training set uses 90 groups of sample curves, and the test set uses 10 groups of test curves to calculate the average error. The test error is shown in Table 1 below:

[0156] Table 1 Test errors of four common charge types

[0157] Wing column type Segmented inner hole End-burning Star Hole 0.85％ 1.68％ 2.23％ 8.01％

[0158] To comprehensively evaluate the prediction accuracy of the proposed model, we use a visualization to display the difference between the predicted results and the actual thrust values. Specifically, a randomly selected set of test data is used, and the actual thrust curve obtained through simulation and the thrust curve predicted by the proposed model are displayed in the same graph, thus providing an intuitive visualization of the error. Figures 5 to 8 The prediction accuracy visualization results for different charge types are shown. Figure 5 、 Figure 6 、 Figure 7 and Figure 8 The prediction accuracy for wing column type, segmented inner hole, end burning and star hole charge respectively. Through the analysis of the visual curves in the above diagram, the following conclusions can be drawn:

[0159] The prediction accuracy for wing-column and segmented-hole charges is high, with their thrust curves fitting the true curve well and exhibiting small errors. The prediction accuracy for end-burning charges is moderate, with small errors, but due to the complexity of their combustion characteristics, the thrust curve exhibits slight deviations at the end of combustion. The prediction accuracy for star-hole charges is low, with large errors. This is primarily due to the presence of residual charge in star-hole charges at the end of combustion, leading to incomplete or unsteady combustion and large thrust fluctuations. This phenomenon results in significant differences in thrust curves across different input curves and operating conditions, particularly for star-hole charges, making effective feature extraction difficult.

[0160] To improve the prediction accuracy of star-hole charges, users can increase the number of sample points based on the present invention and refine the neural network structure. However, this will lead to an increase in computing resources and may significantly extend the model training time.

[0161] 2. Mixed charge model

[0162] The mixed charge model of the present invention is suitable for predicting the thrust of solid rocket motors with unknown charge types. Although in actual missions, the charge types are mostly known, and the common solid rocket charge types generally do not exceed ten and are relatively easy to obtain, in certain special scenarios (such as when the charge type is confidential), the charge type may not be known in advance. In such cases, the mixed charge model of the present invention provides an effective thrust prediction method. However, under normal circumstances, it is recommended that users use the single charge model of the present invention for thrust prediction when the charge type is known.

[0163] The mixed charge model workflow is as follows: the user stores multiple sets of characteristic curves and discrete variable parameter files for different charge types in the same directory. Then, using the integrated simulation program described in step S2, a cyclic simulation is performed to calculate multiple thrust curves. Although this method can handle a variety of charge types, it also faces the following two key difficulties:

[0164] First, the combustion characteristics of different charge types are different. The burning rates and combustion processes of different charge types vary greatly, resulting in non-uniform discretization of the spatial curves of their burning surfaces and cavity volumes. For example, a wing-column charge may be discretized into 101 burning surface areas and cavity volumes corresponding to the thickness values, while an end-burning charge may be discretized into 130, and a segmented inner hole charge may only have about 90. Although the interpolation method is used to unify the input dimensions in step S31, there is still an inevitable error. In addition, the burning time lengths of different charge types are different, and the thrust dimensions output by the final fully connected layer will also be different. For example, the burning time of the end-burning charge is shorter, which may be only about 20 moments, while the burning time of the star-hole charge may be as long as more than 50 moments. If the dimension of the thrust output is unified by the interpolation method, it may lead to increased errors and affect the prediction accuracy.

[0165] The second is the difference in input features of different charges. Since the input features of different charge types vary significantly, the complexity of the features increases after the extracted features are input into the fully connected layer. Such complexity makes it more difficult to learn appropriate parameters through the backpropagation algorithm, resulting in a decrease in prediction accuracy. This problem is also one of the core reasons for the low prediction accuracy of the mixed charge model. Therefore, in order to effectively train the model, a large amount of input sample data is required, and the learning rate and group training parameters need to be increased during the training process. In addition, the test set should contain multiple sets of curve data of different charge types to ensure the generalization ability of the model.

[0166] To further evaluate the effectiveness of the mixed charge model, the present invention provides a statistical analysis of the sample size, prediction error, and training time for different input curve data, as shown in Table 2 below. Table 2 shows the relationship between the sample size and prediction error for different charge types, as well as the time and computing resources required during the training process:

[0167] Table 2 Training data of different input samples

[0168] Sample size Training duration Training accuracy 60 226s 14.60％ 100 317s 11.15％ 200 622s 8.01％ 500 1684s 5.81％

[0169] For a network trained with 100 sets of sample points, a curve prediction of a set of unknown charge types is performed, and its visual curve is as follows: Figure 9 As shown:

[0170] In summary, the proposed neural network based on spatiotemporal feature fusion offers significant advantages in the field of solid rocket thrust prediction. Especially when the charge type is known, the proposed model can achieve high-precision predictions, meeting the error requirements of most engineering projects. Furthermore, the use of this model can significantly reduce prediction costs and shorten the timeframe. By utilizing a trained neural network model, the present invention can effectively replace traditional complex thrust simulation experiments, providing an efficient and cost-effective alternative for the design and optimization of solid rocket engines.

[0171] In some embodiments, see Figure 10 , Figure 10 The embodiment of the present application provides a solid rocket thrust prediction device 100 based on spatiotemporal feature fusion, which is implemented using the solid rocket thrust prediction method based on spatiotemporal feature fusion. The device includes:

[0172] The simulation module 101 is used to collect structural parameters and operating condition data of multiple sets of solid rocket engines, perform joint thrust simulation on the multiple sets of solid rocket engines, and obtain a training data set;

[0173] A thrust prediction model building module 102 is configured to obtain an initial thrust prediction model by performing fusion modeling using a one-dimensional convolutional neural network, a long short-term memory network, an attention mechanism module, and a fully connected neural network;

[0174] A training module 103 is configured to train the initial thrust prediction model using a training data set to obtain a final thrust prediction model;

[0175] The thrust prediction value acquisition module 104 is used to input the structural parameters and operating condition data of the solid motor to be tested into the final thrust prediction model to obtain the thrust prediction value of the solid motor to be tested.

[0176] In some embodiments, see Figure 11 , Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. The electronic device 110 provided in an embodiment of the present application includes a memory 111 and a processor 112. The memory 111 stores a computer program, wherein the computer program, when executed by the processor, implements the solid rocket thrust prediction method based on spatiotemporal feature fusion.

[0177] Specifically, the processor 112 may include, for example, a general-purpose microprocessor, an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 112 may also include onboard memory for caching purposes. The processor 112 may be a single processing unit or multiple processing units for executing different actions of the method flow according to the embodiments of the present application.

[0178] Memory 111 can be, for example, any medium capable of containing, storing, conveying, disseminating, or transmitting instructions. For example, memory 111 can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, components, or propagation media. Specific examples of memory 111 include: magnetic storage devices, such as magnetic tape or hard disk drives (HDDs); optical storage devices, such as compact discs (CD-ROMs); random access memory (RAM) or flash memory; and / or wired or wireless communication links.

[0179] This application also provides a computer-readable medium storing a computer program that, when executed by a processor, implements the solid rocket thrust prediction method based on spatiotemporal feature fusion. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments, or it may exist independently and not incorporated into the device / apparatus / system. The computer-readable medium carries one or more programs that, when executed, implement the method described in the embodiments of this application.

[0180] According to an embodiment of the present application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device, or component. In the present application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries computer-readable program code. Such a propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical cable, radio frequency signals, or any suitable combination thereof.

[0181] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present application may be combined and / or combined in a variety of ways, even if such combinations or combinations are not explicitly described in the present application. In particular, without departing from the spirit and teachings of the present application, the features described in the various embodiments and / or claims of the present application may be combined and / or combined in a variety of ways. All of these combinations and / or combinations fall within the scope of the present application. Therefore, the scope of the present application should not be limited to the above-mentioned embodiments, but should be determined not only by the attached claims, but also by the equivalents of the attached claims. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A solid rocket thrust prediction method based on spatiotemporal feature fusion, characterized in that: Including steps: S1: Collect the structural parameters and operating condition data of multiple sets of solid rocket motors, perform joint thrust simulation on multiple sets of solid rocket motors, and obtain a training data set; S2: Obtain the initial thrust prediction model through fusion modeling of a one-dimensional convolutional neural network, a long short-term memory network, an attention mechanism module, and a fully connected neural network; S3: Train the initial thrust prediction model using the training data set to obtain the final thrust prediction model; S4: After the structural parameters and operating condition data of the solid motor to be tested are subjected to joint thrust simulation, the final thrust prediction model is input to obtain the thrust prediction value of the solid motor to be tested.

2. The solid rocket thrust prediction method based on spatiotemporal feature fusion according to claim 1 is characterized in that: Step S1 is specifically as follows: S11: Setting simulation parameters of thrust simulation, including simulation time, step size and output interval; S12: Constructing a thrust simulation model based on simulation parameters; S13: Inputting multiple sets of solid motor structural parameters and operating condition data into the thrust simulation model to obtain a set of solid motor combustion surface area versus wall thickness variation curves, a set of cavity volume versus wall thickness variation curves, a set of nozzle opening versus time variation curves, a set of discrete variables, and a thrust time series variation curve; S14: Divide the curve set of combustion surface area changing with wall thickness, the curve set of cavity volume changing with wall thickness, the curve set of nozzle opening changing with time, and the discrete variable set into a training data set, a validation data set, and a test data set.

3. The solid rocket thrust prediction method based on spatiotemporal feature fusion according to claim 2 is characterized in that: The input form of the curve of burning surface area changing with wall thickness is: [e1,e2,e3....e n ],[Ab1,Ab2,Ab3....Ab n ] Among them, e1, e2, e3....e n Indicates the thickness value of each meat, Ab1, Ab2, Ab3....Ab n Indicates the burning surface area corresponding to each thickness value, and n represents the total number of data in the vector; The input form of the cavity volume versus wall thickness curve is: [e1,e2,e3....e n ],[Vc1,Vc2,Vc3....Vc n ] Among them, Vc1, Vc2, Vc3....Vc n Indicates the cavity volume corresponding to each wall thickness value.

4. The solid rocket thrust prediction method based on spatiotemporal feature fusion according to claim 2 is characterized in that: The input form of the nozzle opening variation curve over time is: [t1,t2,t3....t n ],[G1,G2,G3....G n ] Among them, t1, t2, t3....t n Indicates each moment, G1, G2, G3....G n Represents the nozzle opening value corresponding to each moment, and n represents the total number of data in the vector.

5. The solid rocket thrust prediction method based on spatiotemporal feature fusion according to claim 1 is characterized in that: The one-dimensional convolutional neural network includes the first convolutional layer, the maximum pooling layer, the second convolutional layer, the adaptive pooling layer, the flattening layer and the fully connected layer connected in sequence.

6. The solid rocket thrust prediction method based on spatiotemporal feature fusion according to claim 1, characterized in that: Step S3 is specifically as follows: S31: Obtain a set of curves of burning surface area changing with wall thickness, curves of cavity volume changing with wall thickness, curves of nozzle opening changing with time, and discrete variables from the training data set; S32: Input the curve of the burning surface area changing with the thickness of the wall and the curve of the cavity volume changing with the thickness of the wall into a one-dimensional convolutional neural network for feature extraction to obtain the burning surface space feature vector and the cavity space feature vector respectively; S33: Inputting the nozzle opening variation curve over time into the long short-term memory network for feature extraction to obtain the opening time series feature vector; S34: The attention mechanism module is used to fuse the burning surface spatial feature vector, the cavity spatial feature vector, the opening time series feature vector, and the discrete variable to obtain a spatiotemporal feature fusion vector. S35: Input the spatiotemporal feature fusion vector into the fully connected neural network to obtain the thrust prediction value; S36: Obtain the actual thrust value, and calculate the mean square error loss between the predicted thrust value and the actual thrust value; S37: Adjust the parameter set of the one-dimensional convolutional neural network according to the mean square error loss value through the NAS mechanism, and calculate the reward function of the NAS mechanism; S38: Backpropagation is performed through the mean squared error loss value to adjust the parameter sets of the long short-term memory network and the fully connected neural network; S39: Repeat steps S31-S38 until the mean square error loss value converges and the reward function is greater than a preset threshold, and obtain the final thrust prediction model.

7. The solid rocket thrust prediction method based on spatiotemporal feature fusion according to claim 6 is characterized by: The expression of thrust prediction value is: y i =σ(W m ·ReLU(W n x+b1)+b2) Among them, x is the spatiotemporal feature fusion vector, W m and W n are the first and second weight matrices of the fully connected neural network, b1 and b2 are the first and second bias terms, σ is the activation function, i is the iteration number, y i is the thrust prediction value of the i-th iteration.

8. A solid rocket thrust prediction device based on spatiotemporal feature fusion, implemented by the solid rocket thrust prediction method based on spatiotemporal feature fusion according to any one of claims 1 to 7, characterized in that: The device comprises: The simulation module is used to collect the structural parameters and operating condition data of multiple sets of solid motors, perform joint thrust simulation on multiple sets of solid motors, and obtain a training data set; The thrust prediction model building module is used to obtain the initial thrust prediction model through fusion modeling of a one-dimensional convolutional neural network, a long short-term memory network, an attention mechanism module, and a fully connected neural network; A training module, configured to train the initial thrust prediction model using a training data set to obtain a final thrust prediction model; The thrust prediction value acquisition module is used to input the structural parameters and working condition data of the solid engine to be tested into the final thrust prediction model to obtain the thrust prediction value of the solid engine to be tested.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the solid motor thrust prediction method based on spatiotemporal feature fusion according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the solid motor thrust prediction method based on spatiotemporal feature fusion according to any one of claims 1 to 7 is implemented.

Images