Solid propellant combustion performance prediction method based on AI model

Abstract

The invention provides a solid propellant combustion performance prediction method based on an AI model, and the method comprises the following steps: 1, carrying out the multi-modal parameter coupling design: carrying out the multi-modal parameter coupling design of a solid propellant, determining the combustion performance parameters of the propellant to be predicted and the formula component parameters influencing the combustion performance parameters, and carrying out the multi-modal parameter coupling design of the solid propellant; establishing a multi-modal parameter coupling prediction system which takes the formula component parameters as input parameters and takes the combustion performance parameters as output parameters; step 2, acquiring sample data; step 3, establishing a differentiated normalization standard; step 4, constructing a combustion performance prediction model: establishing an AI model for predicting the combustion performance of the propellant, and carrying out model training to obtain the combustion performance prediction model meeting the precision requirement; the AI model is a neural network model added with a high-voltage compensation sub-network; and step 5, predicting.

Description

Technical Field

[0001] The present invention belongs to the technical field of solid propellants and relates to solid propellant performance prediction, and specifically to a solid propellant combustion performance prediction method based on an AI model. Background Art

[0002] The prediction of solid propellant combustion performance is a key step in rocket engine design. Traditional methods mainly rely on the iterative model of formulation design and experimental verification. Hot tests or combustion bench tests are carried out on trial samples to directly measure parameters such as burning rate and pressure index. The cost of such experiments can reach hundreds of thousands of yuan per time, and due to safety regulations, they need to be repeated several days apart. In terms of numerical simulation, although computational fluid dynamics (CFD) can simulate the flow field of the combustion chamber, it requires modeling of complex processes such as multiphase flow and chemical reaction kinetics. The full three-dimensional simulation of typical working conditions requires millions of grids and thousands of CPU times.

[0003] In recent years, the use of data-driven approaches based on artificial intelligence has shown significant advantages. By systematically collecting historical experimental data and building a structured database, deep neural networks can be used to efficiently mine hidden associations between components, processes, and performance. The trained model can output predicted values ​​for new formulations within seconds, significantly improving efficiency compared to traditional methods. However, when it comes to predicting the combustion performance of solid propellants, due to the complex composition of propellants, the high correlation between parameters, and the different factors affecting different combustion performance parameters, existing research and technologies have the following shortcomings:

[0004] First, the correlation between individual influencing parameters in traditional prediction models is insufficient.

[0005] Second, the global normalization of traditional methods leads to the loss of key feature information in the influencing factors.

[0006] Third, the traditional network structure is not adaptable enough to different propellant formulation systems, which affects the prediction error. Summary of the Invention

[0007] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a solid propellant combustion performance prediction method based on an AI model to solve the technical problem that the efficiency and accuracy of the propellant combustion performance prediction method of the traditional model in the existing technology need to be further improved.

[0008] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0009] A solid propellant combustion performance prediction method based on an AI model is performed in the following steps:

[0010] Step 1: Multi-modal parameter coupling design:

[0011] Multimodal parameter coupling design is carried out on solid propellants, the combustion performance parameters of the propellant to be predicted and the formula component parameters that affect the combustion performance parameters are determined, and a multimodal parameter coupling prediction system is established with the formula component parameters as input parameters and the combustion performance parameters as output parameters.

[0012] The combustion performance parameters include burning rate, pressure index and temperature sensitivity coefficient.

[0013] The formula component parameters include oxidant particle size distribution, metal fuel surface oxidation layer thickness and mixing temperature process parameters.

[0014] Step 2: Get sample data:

[0015] The formulation component parameters affecting the performance indicators obtained in step 1 are designed, and the propellant combustion performance parameters are experimentally measured based on the designed formulation to obtain sample data for predicting the propellant combustion performance.

[0016] Step 3: Establish differentiated normalization standards:

[0017] The sample data obtained from the measurement in step 2 are divided into data sets, and the coupling coefficient is obtained according to the three-parameter coupling formula of the formula component parameters. A classification system is obtained by dynamic classification based on the range of the coupling coefficient. The sample data is processed using a dynamic classification normalization method based on the grading system to establish a differentiated normalization standard.

[0018] Step 4: Build a combustion performance prediction model:

[0019] Establish an AI model to predict propellant combustion performance, and carry out model training to obtain a combustion performance prediction model that meets accuracy requirements.

[0020] The AI ​​model is a neural network model with a high-voltage compensation sub-network added.

[0021] The number of input layer nodes in the neural network model corresponds to the number of formula component parameters input by the multimodal parameter coupling prediction system obtained in step one, and the number of output layer nodes corresponds to the number of combustion performance parameters output by the multimodal parameter coupling prediction system obtained in step one.

[0022] Step 5: Prediction:

[0023] Input the formula component parameters of the solid propellant, use the combustion performance prediction model trained in step 4 to make predictions, and output the prediction results of the combustion performance parameters.

[0024] Compared with the prior art, the present invention has the following technical effects:

[0025] (I) The method of the present invention adopts multimodal parameter coupling design to construct an association network including material components, microstructure, process parameters and macroscopic performance, solving the problem of insufficient correlation of traditional single parameters.

[0026] (II) The present invention adopts a differentiated normalization method based on material properties, which avoids the loss of key feature information caused by traditional global normalization.

[0027] (III) The algorithm model of the present invention improves prediction accuracy by dynamically adjusting the network structure to adapt to different formulation systems.

[0028] The specific contents of the present invention are further described in detail below with reference to the embodiments. DETAILED DESCRIPTION

[0029] It should be noted that, unless otherwise specified, the functions, models and methods in the present invention all adopt functions, models and methods known in the prior art.

[0030] In accordance with the above technical solution, specific embodiments of the present invention are given below. It should be noted that the present invention is not limited to the following specific embodiments, and all equivalent changes made on the basis of the technical solution of this application fall within the protection scope of the present invention.

[0031] Example:

[0032] This embodiment provides a method for predicting the combustion performance of solid propellant based on an AI (Artificial Intelligence) model. The method is performed according to the following steps:

[0033] Step 1: Multi-modal parameter coupling design:

[0034] Multimodal parameter coupling design is carried out on solid propellants to determine the combustion performance parameters of the propellant to be predicted and the formulation component parameters that affect the combustion performance parameters. A multimodal parameter coupling prediction system is established with the formulation component parameters as input parameters and the combustion performance parameters as output parameters; this solves the problem of insufficient correlation of traditional single parameters.

[0035] In this embodiment, the multimodal parameter coupling design constructs an association network including material components, microstructure, process parameters and macroscopic performance.

[0036] In step one, the combustion performance parameters include burning rate, pressure index, and temperature sensitivity coefficient. In this embodiment, among the combustion performance parameters, the burning rate is the design benchmark for the engine operating time, the pressure index is the core indicator of combustion stability, and the temperature sensitivity coefficient is an environmental adaptability evaluation parameter.

[0037] In step one, the formulation component parameters include the oxidizer particle size distribution, the thickness of the metal fuel surface oxide layer, and the mixing temperature process parameters. In this embodiment, the oxidizer particle size distribution among the formulation component parameters is the core parameter for regulating the burning rate, the thickness of the metal fuel surface oxide layer is a key microscopic parameter that directly affects the propellant energy release efficiency and combustion stability, and the mixing temperature process parameter is a key parameter for controlling the uniformity of particle dispersion.

[0038] In this example, a multimodal parameter coupling prediction system was finally established with the oxidizer particle size distribution (D10 / D50 / D90), the metal fuel surface oxide layer thickness (5-150nm), and the mixing temperature (-20~60℃) as input parameters, and the burning rate, pressure index, and temperature sensitivity coefficient as output parameters.

[0039] Step 2: Get sample data:

[0040] The formulation component parameters affecting the performance indicators obtained in step 1 are designed, and the propellant combustion performance parameters are experimentally measured based on the designed formulation to obtain sample data for predicting the propellant combustion performance.

[0041] Step 2 specifically includes the following steps:

[0042] Step 201, firstly, conduct an experimental design for the three formula factors of the input parameters oxidizer particle size distribution, metal fuel surface oxide layer thickness and mixing temperature. According to the different value states of each formula factor, the oxidizer particle size distribution is given as 10 levels, the metal fuel surface oxide layer thickness is given as 10 levels, and the mixing temperature is given as 5 levels. For such a mixed design with different number of levels of each factor, it is necessary to use the quasi-level method to adjust it. At the same time, the centralized L2 deviation formula is used to calculate the uniformity of the quasi-level design. The final design table is determined according to the minimum deviation, and finally the design scheme table of each formula is obtained.

[0043] In this embodiment, in order to ensure the uniformity of the training sample data, an experimental design is performed on three formulation factors: oxidant particle size distribution, metal fuel surface oxide layer thickness, and mixing temperature.

[0044] In step 201, the formula for calculating the centralized L2 deviation is:

[0045] Where:

[0046] CD 2 Indicates the deviation value;

[0047] d represents the number of factors;

[0048] n represents the number of trials;

[0049] k represents the serial number of the kth test point;

[0050] j represents the sequence number of the j-th dimension;

[0051] represents the normalized coordinates of the kth trial point in the jth dimension.

[0052] In step 202, based on the design table obtained in step 201, the particle size distribution is measured by laser diffraction, the oxide layer thickness is analyzed by XPS (X-ray photoelectron spectroscopy), and the mixing temperature is measured by an online thermal imager. Sample data is obtained with the oxidant particle size distribution, the metal fuel surface oxide layer thickness, and the mixing temperature as input parameters, and the burning rate, pressure index, and temperature sensitivity coefficient as output parameters.

[0053] Step 3: Establish differentiated normalization standards:

[0054] The sample data obtained from the measurement in step 2 are divided into data sets, and the coupling coefficient is obtained according to the three-parameter coupling formula of the formula component parameters. A classification system is obtained by dynamic classification based on the range of the coupling coefficient. The sample data is processed using a dynamic classification normalization method based on the grading system to establish a differentiated normalization standard.

[0055] In step 3, the data set is divided into training samples, validation samples, and test samples in a ratio of 7:2:1. In this embodiment, the training set contains 70 test data, the validation set contains 20 test data, and the test set contains 10 test data.

[0056] In step 3, the three-parameter coupling formula is:

[0057] Where:

[0058] α represents the coupling coefficient;

[0059] D 50 represents the median particle size distribution of the oxidant;

[0060] OI oxide layer thickness;

[0061] PC represents the mixing temperature.

[0062] In step 302, the classification system is:

[0063]

[0064] In step three, the dynamic classification normalization method is:

[0065] For Class I classification, the oxidant particle size distribution and the oxide layer thickness are synergistically dominant, and the normalized standard formula is: Where, X Ⅰ It represents the normalized value of the input parameters of Class I classification level.

[0066] For Class II classification, the mixing temperature and oxide layer thickness have a two-way influence, and the normalized standard formula is: Where, X Ⅱ It represents the normalized value of the Class II classification level input parameter.

[0067] For the Class III classification level, there is no obvious synergistic effect between the parameters, and the formula for the normalization standard is: Where, X Ⅲ It represents the normalized value of the input parameters of Class III classification level.

[0068] Step 4: Build a combustion performance prediction model:

[0069] Establish an AI model to predict propellant combustion performance, and carry out model training to obtain a combustion performance prediction model that meets accuracy requirements.

[0070] The AI ​​model is a neural network model with a high-voltage compensation sub-network added.

[0071] The number of input layer nodes in the neural network model corresponds to the number of recipe component parameters input into the multimodal parameter coupling prediction system obtained in step one, and the number of output layer nodes corresponds to the number of combustion performance parameters output by the multimodal parameter coupling prediction system obtained in step one.

[0072] Step 4 specifically includes the following steps:

[0073] Step 401: Establish a neural network model including an input layer, at least three hidden layers and an output layer.

[0074] Step 402: Add a high-voltage compensation sub-network. The high-voltage compensation sub-network includes three hidden layers (containing 32, 32, and 16 neurons, respectively). When the oxidant content is detected to be greater than 80%, it is automatically activated to realize an adaptive network architecture based on a dynamic branch structure.

[0075] In this embodiment, the high-pressure compensation subnetwork is mainly used to solve the non-ideal combustion behavior of the propellant when the oxidizer content is greater than 80%, and includes: the first hidden layer is used to process the high-pressure fragmentation characteristics of the oxidizer crystals, using the ELU (Exponential Linear Unit) activation function; the second hidden layer is used to consider the melting dynamics of the oxide layer, using the SiLU (Sigmoid Gated Linear Unit) activation function; the third hidden layer is used to generate a pressure-sensitive compensation term, using a linear output function.

[0076] In this embodiment, the dynamic branch structure is implemented by a dynamic routing controller, the oxidant content threshold is calculated in real time to implement branch signal transmission, and a smooth transition zone of 78-82% is achieved by a sigmoid function.

[0077] In step 402, the adaptive network architecture couples the basic neural network and the high-voltage compensation sub-network, and realizes the adaptive weighting of the basic prediction and the high-voltage compensation output through the learnable weight coefficient α.

[0078] In step 402, the implementation equation for adaptive weighting of the basic prediction and the high-voltage compensation output is:

[0079]

[0080] Where:

[0081] Represents the output parameter results of the adaptive network;

[0082] w b Represents the basic network weight coefficient;

[0083] w c Represents the weight coefficient of the high-voltage compensation network;

[0084] F b Represents the basic network output parameters;

[0085] F c Indicates the output parameter results of the high voltage compensation network;

[0086] α is the weight coefficient;

[0087] AP% indicates the percentage of oxidant content;

[0088] P stands for pressure.

[0089] In this embodiment, the weight coefficient α is used to integrate the recipe and process parameters.

[0090] Step 403: Network training is carried out based on the adaptive network architecture obtained in step 403. A two-stage training strategy is adopted. First, the basic network training compensation module is frozen and the learning rate is set to 1e-3. Then, joint fine-tuning is performed to set the learning rate to 5e-5. The loss function adopts a dynamic weighted mean square error loss function. The training set data is iteratively trained. When the error of the validation set loss function decreases by less than 1% for five consecutive rounds, the training is terminated to obtain a combustion performance prediction model that meets the accuracy requirements.

[0091] In step 403, the dynamic weighted mean square error loss function is as follows:

[0092]

[0093] Where:

[0094] L represents the loss error;

[0095] n represents the sample size;

[0096] w i Represents dynamic weight;

[0097] y i Represents the training data values ​​of each output parameter;

[0098] Represents the predicted data value of each output parameter;

[0099] clip represents the truncation function;

[0100] x i Represents the training data values ​​of each input parameter;

[0101] a represents the threshold;

[0102] b represents the width of the transition interval.

[0103] Step 5: Prediction:

[0104] Input the formula component parameters of the solid propellant, use the combustion performance prediction model trained in step 4 to make predictions, and output the prediction results of the combustion performance parameters.

[0105] Comparative Example 1:

[0106] This comparative example provides a method for predicting the combustion performance of solid propellants, which is carried out according to the following steps:

[0107] In step 1, only the oxidant particle size D50 is selected as the input parameter, and the output parameters are the predicted burning rate, pressure index, and temperature sensitivity coefficient.

[0108] Step 2: Measure the corresponding burning rate results according to different oxidant particle size distributions to obtain sample data.

[0109] Step three: Divide the sample data into training set, validation set and test set, and use global unified normalization.

[0110] Step 4: Get the prediction model after training the data.

[0111] Step 5: Import the data to be predicted into the prediction model for testing to obtain the prediction parameter results.

[0112] Comparative Example 2:

[0113] This comparative example provides a method for predicting the combustion performance of solid propellant. The difference between this method and the embodiment is that, compared with step 4 in the embodiment, a Sigmoid activation function is used and a high-voltage compensation mechanism is not set. The other steps are the same as those in the embodiment.

[0114] Using the same test data, the prediction results of propellant combustion performance using the method of the present invention, the method of Comparative Example 1 and the method of Comparative Example 2 were compared, and the final result error results are shown in Table 1.

[0115] Table 1 Errors in combustion performance prediction results of different methods

[0116] method Burning rate error Pressure index error Temperature sensitivity coefficient error Comparative Example 1 11.73％ 12.12％ 15.64％ Comparative Example 2 7.11％ 5.97％ 6.23％ Embodiments of the present invention 3.12％ 2.98％ 3.65％

[0117] As shown in Table 1, when the method of the present invention is used and the adaptive network algorithm is adopted, the prediction efficiency can be effectively improved, so that the prediction accuracy of the propellant combustion performance is greatly improved.

Claims

1. A solid propellant combustion performance prediction method based on an AI model, characterized in that: The method is performed as follows: Step 1: Multi-modal parameter coupling design: Carry out multimodal parameter coupling design for solid propellants, determine the combustion performance parameters of the propellant to be predicted and the formulation component parameters that affect the combustion performance parameters, and establish a multimodal parameter coupling prediction system with the formulation component parameters as input parameters and the combustion performance parameters as output parameters; The combustion performance parameters include burning rate, pressure index and temperature sensitivity coefficient; The formulation component parameters include oxidant particle size distribution, metal fuel surface oxide layer thickness and mixing temperature process parameters; Step 2: Get sample data: Designing a formula based on the formulation component parameters that affect the performance indicators obtained in step 1, experimentally measuring the propellant combustion performance parameters based on the designed formula, and obtaining sample data for predicting the propellant combustion performance; Step 3: Establish differentiated normalization standards: The sample data obtained in step 2 is divided into data sets, the coupling coefficient is obtained according to the three-parameter coupling formula of the formula component parameters, and a classification system is obtained by dynamic classification based on the range of the coupling coefficient. The sample data is processed using a dynamic classification normalization method based on the classification system to establish a differentiated normalization standard; Step 4: Build a combustion performance prediction model: Establish an AI model to predict propellant combustion performance and conduct model training to obtain a combustion performance prediction model that meets accuracy requirements; The AI ​​model is a neural network model with a high-voltage compensation sub-network added. The number of input layer nodes in the neural network model corresponds to the number of the recipe component parameters input by the multimodal parameter coupling prediction system obtained in step 1, and the number of output layer nodes corresponds to the number of combustion performance parameters output by the multimodal parameter coupling prediction system obtained in step 1; Step 5: Prediction: Input the formula component parameters of the solid propellant, use the combustion performance prediction model trained in step 4 to make predictions, and output the prediction results of the combustion performance parameters.

2. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 201: First, perform an experimental design for the three input parameters of oxidizer particle size distribution, metal fuel surface oxide layer thickness, and mixing temperature. Based on the different values ​​of each formulation factor, 10 levels of oxidizer particle size distribution, 10 levels of metal fuel surface oxide layer thickness, and 5 levels of mixing temperature are given. For such a mixed design with different numbers of factor levels, a pseudo-level method is used to adjust it. At the same time, the centered L2 deviation formula is used to calculate the uniformity of the pseudo-level design. The final design table is determined based on the minimum deviation, and finally, a design solution table for each formulation is obtained. Step 202, based on the design plan table obtained in step 201, the particle size distribution is measured by laser diffraction method, the oxide layer thickness is analyzed by XPS depth profiling, and the mixing temperature is measured by online thermal imager, so as to obtain sample data with the oxidant particle size distribution, the metal fuel surface oxide layer thickness and the mixing temperature as input parameters, and the burning rate, pressure index and temperature sensitivity coefficient as output parameters.

3. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 2, characterized in that: In step 201, the centralized L2 deviation calculation formula is: Where: CD 2 Indicates the deviation value; d represents the number of factors; n represents the number of trials; k represents the serial number of the kth test point; j represents the sequence number of the j-th dimension; represents the normalized coordinates of the kth trial point in the jth dimension.

4. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 1, characterized in that: In step 3, the three-parameter coupling formula is: Where: α represents the coupling coefficient; D 50 represents the median particle size distribution of the oxidant; OI oxide layer thickness; PC represents the mixing temperature.

5. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 4, characterized in that: In step 3, the classification system is:

6. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 5, characterized in that: In step 3, the dynamic classification normalization method is: For the Class I classification level, the formula for the normalization standard is: Where, X Ⅰ It represents the normalized value of the input parameters of Class I classification level; For the Class II classification level, the formula for the normalization standard is: Where, X Ⅱ It represents the normalized value of the input parameters of Class II classification level; For the Class III classification level, the formula for the normalization standard is: Where, X Ⅲ It represents the normalized value of the input parameters of Class III classification level.

7. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 1, characterized in that: Step 4 specifically includes the following steps: Step 401, establishing a neural network model including an input layer, at least three hidden layers and an output layer; Step 402: Add a high-voltage compensation sub-network, wherein the high-voltage compensation sub-network includes three hidden layers and is automatically activated when the oxidant content is detected to be greater than 80%, thereby realizing an adaptive network architecture based on a dynamic branching structure. In step 403, network training is performed based on the adaptive network architecture obtained in step 402. A two-stage training strategy is adopted. First, the basic network training compensation module is frozen and the learning rate is set to 1e-3. Then, joint fine-tuning is performed to set the learning rate to 5e-5. The loss function adopts a dynamic weighted mean square error loss function. The training set data is iteratively trained. When the error of the validation set loss function decreases by less than 1% for five consecutive rounds, the training is terminated to obtain a combustion performance prediction model that meets the accuracy requirements.

8. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 7, characterized in that: In step 402, the adaptive network architecture couples the basic neural network and the high-voltage compensation sub-network, and realizes the adaptive weighting of the basic prediction and the high-voltage compensation output through the learnable weight coefficient α.

9. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 8, characterized in that: In step 402, the realization equation of the adaptive weighting of the basic prediction and the high-voltage compensation output is: Where: Represents the output parameter results of the adaptive network; w b Represents the basic network weight coefficient; w c Represents the weight coefficient of the high-voltage compensation network; F b Represents the basic network output parameters; F c Indicates the output parameter results of the high voltage compensation network; α is the weight coefficient; AP% indicates the percentage of oxidant content; P stands for pressure.

10. The solid propellant combustion performance prediction method based on the AI ​​model according to claim 7, characterized in that: In step 403, the dynamic weighted mean square error loss function is as follows: Where: L represents the loss error; n represents the sample size; w i Represents dynamic weight; y i Represents the training data values ​​of each output parameter; Represents the predicted data value of each output parameter; clip represents the truncation function; x i Represents the training data values ​​of each input parameter; a represents the threshold; b represents the width of the transition interval.

Images