Combustion analog simulation analysis method and system based on ADN-based liquid propellant

Abstract

The invention provides a combustion analog simulation analysis method and system based on an ADN-based liquid propellant, and the method comprises the steps: carrying out the prediction of combustion parameters through employing a joint learning network, obtaining X single-parameter prediction networks during the network adjustment, and carrying out the adjustment of first reasoning layers corresponding to assembly sets in the single-parameter prediction networks. Fusing the component sets at the a level in the X single-parameter prediction networks to obtain a combustion parameter prediction network, and setting X second reasoning layers for the a fused component set; taking a real combustion parameter corresponding to combustion condition training data of the ADN-based liquid propellant sample as a prior mark, and taking a predicted combustion parameter of the first reasoning layer as a prior mark of a predicted combustion parameter of the second reasoning layer; the prediction combustion parameter of the single-parameter prediction network is used as a priori mark to adjust the combustion parameter prediction network, so that the precision of combustion parameter prediction of the combustion parameter prediction network can be improved, and the precision of combustion analog simulation is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a combustion simulation analysis method and system based on ADN-based liquid propellant. Background Art

[0002] In aerospace and other fields, ADN-based liquid propellants are widely used due to their excellent performance. It is crucial to accurately simulate and analyze the combustion characteristics of ADN-based liquid propellants, which helps to gain a deeper understanding of the combustion process of the propellant, optimize the engine design, and improve the performance and reliability of the propulsion system. At present, in the simulation and analysis of propellant combustion, some prediction methods based on machine learning have improved the prediction efficiency to a certain extent, but due to the unreasonable design of the network structure and the lack of effective tuning strategies, the prediction accuracy is limited. For example, most of the existing prediction networks are designed for a single combustion parameter and cannot accurately predict multiple combustion parameters at the same time; and in the process of network training, the correlation between different parameters and the advantages of existing models are not fully utilized, resulting in a large deviation between the prediction results and the actual combustion situation, which is difficult to meet the needs of actual engineering applications. Summary of the invention

[0003] The object of the present invention is to provide a combustion simulation analysis method and system based on ADN-based liquid propellant. The embodiment of the present invention is implemented as follows:

[0004] In a first aspect, an embodiment of the present invention provides a combustion simulation and analysis method based on ADN-based liquid propellant, the method comprising: in response to a simulation analysis instruction, obtaining combustion condition data of a target ADN-based liquid propellant; loading the combustion condition data of the target ADN-based liquid propellant into a pre-calibrated combustion parameter prediction network, the combustion parameter prediction network being a joint learning network, comprising a plurality of different branch parameter prediction networks, one branch parameter prediction network being used to predict a combustion parameter; outputting at least one combustion parameter through the combustion parameter prediction network; performing visualization processing based on the at least one combustion parameter to obtain a simulation result of the target ADN-based liquid propellant.

[0005] In a second aspect, an embodiment of the present invention provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0006] In the process of adjusting the combustion parameter prediction network, the present invention configures a first reasoning layer on the component set in the single parameter prediction network for adjustment to obtain the predicted combustion parameters of the component set in the single parameter prediction network, which is used to a priori mark the second reasoning layer of the component set in the combustion parameter prediction network, strengthen the prediction performance of the component set in the combustion parameter prediction network, so that the component set in the combustion parameter prediction network can have the prediction performance of the single parameter prediction network when making predictions, and improve the parameter prediction accuracy of the combustion parameter prediction network. Further, the present invention determines the predicted combustion parameters of the single parameter prediction network as the a priori mark of the predicted combustion parameters of the combustion parameter prediction network, and aligns the predicted combustion parameters of the single parameter prediction network with the predicted combustion parameters of the combustion parameter prediction network to improve the accuracy of the predicted combustion parameters of the combustion parameter prediction network. In addition, the real combustion parameters corresponding to the combustion condition training data of the ADN-based liquid propellant sample are used as a priori marks, which increases the parameter prediction accuracy of the combustion parameter prediction network and helps to perform accurate simulation analysis.

[0007] Other features will be described in part in the following description. Those skilled in the art will discover these features in part upon inspection of the following content and drawings, or may learn these features through production or use. The features of the present application may be implemented and obtained by practicing or using various aspects of the methods, tools, and combinations listed in the detailed examples described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments of the present invention are briefly introduced below.

[0009] Figure 1 It is a flow chart of a combustion simulation analysis method based on ADN-based liquid propellant provided by an embodiment of the present invention.

[0010] Figure 2 It is a schematic diagram of the adjustment process of the combustion parameter prediction network provided in an embodiment of the present invention.

[0011] Figure 3 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] The following describes the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. The terms used in the implementation method part of the embodiments of the present invention are only used to explain the specific embodiments of the present invention, and are not intended to limit the present invention.

[0013] The execution subject of the combustion simulation analysis method based on ADN-based liquid propellant in the embodiment of the present invention is a computer system, including but not limited to a server, a personal computer, a laptop computer, a tablet computer, a smart phone, etc.

[0014] like Figure 1 As shown, the method includes: Step 100: In response to a simulation analysis instruction, obtaining combustion condition data of a target ADN-based liquid propellant.

[0015] The computer system is in a waiting state, ready to receive external input simulation analysis instructions at any time. The instruction can come from a trigger operation performed by the user on a specific operation interface. For example, the user clicks the "Start Simulation Analysis" button in the interface of a professional propellant combustion simulation software. At this time, the software converts the operation into a digital signal that the computer system can recognize, and then generates a simulation analysis instruction and transmits it to the computer system. Then, after receiving the simulation analysis instruction, the computer system begins to obtain the combustion condition data of the target ADN-based liquid propellant. The combustion condition data contains multiple key elements, which are crucial for accurately simulating the combustion process.

[0016] Among them, temperature is an important combustion condition data. Different temperature environments will greatly affect the combustion reaction rate and degree of ADN-based liquid propellants. For example, at a lower initial temperature, the activity of propellant molecules is relatively low, and the combustion reaction may be slow; while at a higher initial temperature, the molecular motion is intensified, and the combustion reaction will be more intense and rapid. The computer system can obtain temperature data by interacting with temperature sensor devices. These sensors are precisely placed at appropriate locations in the simulation environment to monitor and transmit temperature information to the computer system in real time. Pressure is also an indispensable combustion condition data. Pressure changes in the combustion chamber will directly affect the combustion stability and energy release efficiency of the propellant. For example, a higher pressure environment can promote collisions between propellant molecules and accelerate the combustion reaction; while unstable pressure may cause fluctuations in the combustion process or even flameout. The computer system can obtain pressure data by connecting pressure sensors, which can accurately measure the pressure values ​​at different positions and times in the combustion chamber and convert them into digital signals and send them to the computer system. The concentration of the propellant is also a key data. ADN-based liquid propellants with different concentrations will have different combustion characteristics. When the concentration is high, the combustion reaction may be more violent and release more energy; when the concentration is low, the combustion reaction may be relatively mild. The computer system can obtain the concentration data of the propellant through chemical analysis instruments. These instruments can accurately analyze the propellant samples and transmit the concentration information to the computer system. In addition to the common combustion condition data mentioned above, it may also include data such as the injection speed and injection angle of the propellant. The injection speed determines the rate at which the propellant enters the combustion chamber, which in turn affects the intensity and duration of the combustion; the injection angle affects the distribution of the propellant in the combustion chamber and affects the uniformity of the combustion. The computer system can obtain this data by connecting relevant flow sensors and angle measurement equipment.

[0017] In the actual process of obtaining combustion condition data, the computer system has corresponding interfaces and drivers to achieve communication with various sensors and measuring devices. For example, for temperature sensors, the computer system needs to exchange data with the sensor through specific communication protocols (such as SPI, I2C, etc.), receive temperature data in the format specified by the protocol, and perform corresponding analysis and processing. For pressure sensors, concentration analyzers and other equipment, it is also necessary to configure appropriate interfaces and drivers to ensure accurate data acquisition and transmission.

[0018] When the computer system obtains the combustion condition data of the target ADN-based liquid propellant, it will classify, store and manage the data so that it can be easily and quickly called later. For example, the temperature data is stored in a special temperature data file, the pressure data is stored in a pressure data file, and a corresponding index is established for each data file to quickly locate and retrieve data.

[0019] Step 200: Load the combustion condition data of the target ADN-based liquid propellant into a pre-calibrated combustion parameter prediction network, where the combustion parameter prediction network is a joint learning network including a plurality of different branch parameter prediction networks, and one branch parameter prediction network is used to predict one combustion parameter.

[0020] The combustion parameter prediction network is a joint learning network, that is, a neural network model that performs multiple tasks. The advantage of this network structure is that it can handle multiple related tasks at the same time. In the embodiment of the present invention, it predicts multiple different combustion parameters. Taking a rocket engine combustion chamber simulation as an example, in the actual operation process, it is necessary not only to know the temperature distribution in the combustion chamber, which is crucial for judging whether the combustion chamber wall will be damaged due to high temperature; it is also necessary to understand the pressure change, because the stability of the pressure is directly related to the working stability of the engine; at the same time, the component concentration distribution can help analyze the combustion efficiency of the propellant and the product composition; combustion efficiency and thrust performance are the key indicators for measuring engine performance. The combustion parameter prediction network can integrate these tasks and predict these combustion parameters separately through different branch parameter prediction networks.

[0021] Specifically, the combustion parameter prediction network contains multiple different branch parameter prediction networks. Each branch parameter prediction network is designed to specifically predict a specific combustion parameter. For example, one of the branch parameter prediction networks focuses on predicting the temperature distribution in the combustion chamber. This branch parameter prediction network receives temperature-related combustion condition data as input, such as the initial temperature of the propellant, the initial pressure in the combustion chamber, and other data, and then predicts the temperature distribution at different locations and at different times in the combustion chamber through its internal network structure and algorithm. For another example, another branch parameter prediction network is responsible for predicting pressure changes. It will output the changes in the pressure in the combustion chamber over time and space based on the combustion condition data such as the propellant flow rate and injection speed, after a series of calculations and processing.

[0022] During the data loading process, the computer system operates strictly in accordance with the input requirements of the combustion parameter prediction network. Different branch parameter prediction networks may have specific requirements for the format and dimension of the input data. For example, a branch parameter prediction network used to predict temperature distribution may require the input data to be a multidimensional array, where each dimension represents a different physical quantity (such as time, spatial coordinates, etc.). Based on these requirements, the computer system organizes and converts the combustion condition data to ensure that the data can be correctly input into the corresponding branch parameter prediction network.

[0023] The pre-calibrated combustion parameter prediction network is obtained through a lot of training and optimization. During the training process, a large amount of combustion condition training data of ADN-based liquid propellant samples and the corresponding real combustion parameters are used. By continuously adjusting the weights and parameters of the network, the network can predict the combustion parameters as accurately as possible. For example, when training the branch parameter prediction network used to predict combustion efficiency, a large amount of propellant sample data under different combustion conditions is input into the network, and the actual combustion efficiency of these samples is used as the real label. Through continuous learning and adjustment, the network gradually improves the accuracy of combustion efficiency prediction.

[0024] When the computer system successfully loads the combustion condition data of the target ADN-based liquid propellant into the combustion parameter prediction network, each branch parameter prediction network will start working. Taking the branch parameter prediction network that predicts the component concentration distribution in the combustion chamber as an example, it will analyze and process the loaded combustion condition data. First, the data will be received at the input layer of the network, and then calculated and converted by multiple neurons in the hidden layer in turn. Each neuron will perform a weighted summation on the input data and perform a nonlinear transformation through an activation function. Commonly used activation functions such as the ReLU function, whose expression is: f(x)=max(0,x). After a series of calculations and transformations, the prediction result of the component concentration distribution in the combustion chamber is finally obtained at the output layer.

[0025] During the entire loading and prediction process, the computer system monitors the flow of data and the operating status of the network in real time. If there is a data loading error or abnormal network operation, the computer system will promptly issue an alarm and take corresponding measures. For example, if the data format does not meet the requirements when loading data, the computer will prompt the user to check the data format and provide corresponding correction suggestions.

[0026] The computer system executes step 200 to load the combustion condition data of the target ADN-based liquid propellant into the pre-calibrated combustion parameter prediction network, which is a key link in the entire combustion simulation analysis process. Through reasonable hardware configuration, effective data loading and preprocessing technical means, and accurate network input and operation management, it is ensured that each branch parameter prediction network can accurately predict different combustion parameters, laying a solid foundation for obtaining accurate simulation results in the future.

[0027] Step 300: Output at least one combustion parameter through a combustion parameter prediction network.

[0028] As a joint learning network, the combustion parameter prediction network contains multiple different branch parameter prediction networks, each of which predicts a specific combustion parameter. The computer system uses the neurons and connection weights of each layer in the network to implement this prediction process. When the combustion condition data enters the combustion parameter prediction network, it first reaches the input layer. The neurons in the input layer are responsible for receiving this data and passing it to the next layer, the hidden layer. The hidden layer is the key part of the network for complex calculations and feature extraction. In the hidden layer, the neurons perform weighted summation operations on the input data. After weighted summation, the data will be transformed nonlinearly through the activation function. Through the activation function, the network can learn the nonlinear relationship in the data and enhance the network's expression ability. After calculation and transformation of multiple hidden layers, the data finally reaches the output layer. The neurons in the output layer will calculate and output the predicted combustion parameters based on the results passed from the hidden layer.

[0029] For example, for a branch parameter prediction network that predicts the temperature distribution in a combustion chamber, the computer system inputs temperature-related combustion condition data into the branch network. After receiving the data through the input layer, the neurons in the hidden layer perform weighted summation and nonlinear transformation on the data. In this process, the network learns the complex relationship between combustion conditions and temperature distribution. Finally, the output layer outputs predicted values ​​of temperature distribution at different locations and times in the combustion chamber based on these calculation results. After the computer system obtains this temperature distribution data, it outputs it as a combustion parameter. For another example, for a branch parameter prediction network responsible for predicting combustion efficiency, the computer system also inputs relevant combustion condition data into the network. After calculations at each layer of the network, the output layer gives a predicted value of combustion efficiency. The computer system outputs this combustion efficiency value as another combustion parameter.

[0030] Step 400: Performing visualization processing according to at least one combustion parameter to obtain simulation results of a target ADN-based liquid propellant.

[0031] After the computer system outputs at least one combustion parameter through the combustion parameter prediction network, it starts to execute step 400 to visualize these combustion parameters, thereby obtaining the simulation results of the target ADN-based liquid propellant. Visualization is to present the abstract combustion parameters in the form of intuitive graphics, images or animations, so that users can understand and analyze the combustion of the propellant more clearly.

[0032] When the computer system performs visualization processing, it will first select the appropriate visualization method according to different types of combustion parameters. For the combustion parameter of temperature distribution in the combustion chamber, the computer system may use two-dimensional or three-dimensional heat map for visualization. For example, assuming that the space in the combustion chamber can be simplified into a two-dimensional plane, the computer system divides the plane into several grid cells, each of which corresponds to a specific position. According to the temperature distribution combustion parameter output previously, a corresponding temperature value is assigned to each grid cell. Then, different temperature values ​​are mapped to different colors using the color mapping method. For example, the low temperature area is represented by blue, the high temperature area is represented by red, and the intermediate temperature area is represented by a gradually transitioning color. In this way, by observing the generated two-dimensional heat map, the user can intuitively see the high and low temperature distribution in the combustion chamber, quickly identify the high temperature area and the low temperature area, and judge whether the combustion is uniform.

[0033] For the combustion parameter of pressure change, the computer system can visualize it in the form of a line graph or a bar graph. Taking the line graph as an example, the computer system uses time as the horizontal axis and pressure value as the vertical axis. According to the output pressure change combustion parameter, the pressure values ​​corresponding to different time points are plotted in the coordinate system in sequence, and these points are connected into a broken line. In this way, the user can clearly understand the change trend of the pressure in the combustion chamber over time by observing the trend of the broken line, and judge whether the pressure is stable or whether there is a situation of excessive pressure fluctuation.

[0034] For the combustion parameter of component concentration distribution, the computer system can use contour maps or three-dimensional stereograms to display. If contour maps are used, the computer system divides the space in the combustion chamber into certain areas, and calculates the component concentration values ​​at different locations based on the combustion parameters of the component concentration distribution. Then, points with the same concentration value are connected into lines to form contour lines. These contour lines can intuitively show the distribution pattern of component concentrations, and users can judge the gradient of concentration changes by the density of the contour lines.

[0035] In the process of visualization, the computer system needs to use some technical means to achieve it. In terms of data processing, the computer system performs necessary preprocessing on the output combustion parameters. For example, the data is smoothed to reduce the noise interference in the data. Assume that the combustion parameter data sequence is , using the moving average method for smoothing, the new data series The calculation formula is: (where k is the smoothing window size, The smoothed data can more accurately reflect the changing trend of combustion parameters and provide a more reliable data basis for subsequent visualization.

[0036] In terms of graphics drawing, the computer system calls professional visualization libraries or tools. For example, the Matplotlib library in Python is a widely used visualization tool that provides a wealth of drawing functions and methods. The computer system can use the functions in the Matplotlib library to draw basic graphics such as line graphs and bar graphs. For more complex three-dimensional graphics, the computer system can use libraries such as Mayavi, which can realize the visualization of three-dimensional spatial data and present the combustion parameters in intuitive three-dimensional graphics.

[0037] In terms of color mapping, the computer system selects a suitable color mapping scheme based on the range and characteristics of the combustion parameters. Common color mapping schemes include Jet, Viridis, etc. Taking Viridis color mapping as an example, it is a color mapping widely used in scientific visualization, with good perceptual uniformity and color blindness friendliness. The computer system maps the minimum and maximum values ​​of the combustion parameters to the color space of the Viridis color mapping, so that different combustion parameter values ​​correspond to different colors, thereby achieving intuitive visualization.

[0038] When the computer system performs visualization, it also considers the layout and annotation of the graphics. For example, information such as axis labels, titles, and legends are added to the drawn graphics so that users can better understand the meaning of the graphics. For different combustion parameter graphics, the computer system reasonably arranges their layout on the display interface to make the entire visualization result clearer and more beautiful. By visualizing the combustion parameters, the computer system finally obtains the simulation results of the target ADN-based liquid propellant. These simulation results are presented in the form of intuitive graphics, images, or animations. Users can gain an in-depth understanding of the combustion characteristics of ADN-based liquid propellants under different combustion conditions by observing these visualization results. For example, by observing the visualization results of the temperature distribution, users can determine whether the propellant is fully burned and whether there is local overheating; by analyzing the visualization graphics of pressure changes, users can evaluate the stability of the combustion process and provide a basis for optimizing the combustion system.

[0039] Please refer to Figure 2 In one implementation, the combustion parameter prediction network is calibrated by the following steps: Step 10: Obtain X single parameter prediction networks. Different single parameter prediction networks are used to predict different combustion parameters. Component sets at the same level in different single parameter prediction networks have the same network architecture. X is a positive integer greater than 1.

[0040] When the computer system starts to adjust the combustion parameter prediction network, it first executes step 10 to obtain X single parameter prediction networks. The single parameter prediction network here is a network model specifically for predicting a specific combustion parameter.

[0041] Different single-parameter prediction networks are used to predict different combustion parameters. For example, in the combustion simulation scenario of ADN-based liquid propellant, there are multiple combustion parameters that need to be predicted, such as temperature distribution, pressure change, component concentration distribution, etc. in the combustion chamber. The computer system constructs a corresponding single-parameter prediction network for each different combustion parameter. Taking temperature distribution prediction as an example, a special single-parameter prediction network will be constructed. The input data of this network will revolve around temperature-related factors, such as the initial temperature of the propellant, the heat of chemical reaction during combustion, etc., and its output is the predicted results of temperature values ​​at various locations and at different times in the combustion chamber. Similarly, for pressure change prediction, there will be another single-parameter prediction network, which will receive pressure-related input data such as propellant flow rate and injection speed, and then output the predicted value of the change of pressure in the combustion chamber over time and space.

[0042] The component sets at the same level in different single-parameter prediction networks have the same network architecture. This means that when constructing these single-parameter prediction networks, the computer system follows certain design rules so that each network has a similar structure at the same level. For example, each single-parameter prediction network may contain an input layer, multiple hidden layers, and an output layer. In the input layer, both the temperature prediction network and the pressure prediction network are responsible for receiving their respective relevant input data and passing it to the next layer. The component sets in the hidden layer, such as the number of neurons, the connection method, etc., are the same at the same level. Assuming that each single-parameter prediction network has three hidden layers, then in the first hidden layer, the number of neurons, the choice of activation function, etc. of each network are consistent. This consistency helps the subsequent network fusion and tuning process.

[0043] The computer system can obtain these X single parameter prediction networks through a variety of technical means. In terms of building the network structure, the computer system can use common deep learning frameworks such as TensorFlow or PyTorch. Taking TensorFlow as an example, the computer system can use the various functions and classes it provides to define the hierarchical structure of the network. For example, use `tf.keras.Sequential` to build a sequential model, and define the hidden layer and output layer by adding the `tf.keras.layers.Dense` layer. When determining the input and output of each single parameter prediction network, the computer system sets it according to the characteristics of different combustion parameters. For the temperature prediction network, the computer system analyzes the physical factors related to temperature, determines the appropriate input data type and dimension, and determines the number of neurons and output format of the output layer according to the accuracy requirements of the temperature prediction.

[0044] When training these single-parameter prediction networks, the computer system uses a large amount of ADN-based liquid propellant sample data and the corresponding real combustion parameters as training sets. For example, for the temperature prediction network, propellant sample data under different combustion conditions, including propellant composition, injection method and other information, are collected as input, and the temperature distribution in the combustion chamber obtained by actual measurement of these samples is used as the real label. The computer system adjusts the weight parameters of the network through the back-propagation algorithm to minimize the error between the predicted value and the true value. The core of the back-propagation algorithm is to calculate the gradient of the loss function to the network weight, and the calculation formula is: ,in It represents the gradient of the loss function L with respect to the weight w. By continuously iteratively updating the weights, the single parameter prediction network can gradually learn the relationship between the input data and the target combustion parameters, thereby improving the accuracy of the prediction.

[0045] Step 20: calibrate the first reasoning layers corresponding to the component sets in the single parameter prediction network, where the first reasoning layers are used to obtain predicted combustion parameters based on the output of the component sets.

[0046] After obtaining X single-parameter prediction networks, the computer system adjusts the first inference layer corresponding to the component sets in these networks. The first inference layer plays a key role in the single-parameter prediction network, and it is responsible for obtaining the predicted combustion parameters based on the output of the component set.

[0047] Take a single parameter prediction network for predicting the temperature distribution in the combustion chamber as an example. After a series of data processing and feature extraction, the component set in the network will pass the results to the corresponding first inference layer. The component set can be understood as a combination of neurons at a certain level in the network, which jointly process the input data. For example, in this temperature prediction network, the component set may be a hidden layer, which calculates and converts the input temperature-related combustion condition data, such as the initial temperature of the propellant, the heat of chemical reaction, and other information, to extract features that have an important impact on the temperature distribution.

[0048] After receiving the outputs of the component set, the first reasoning layer will perform inference calculations based on these outputs to obtain the predicted combustion parameters, that is, the predicted value of the temperature distribution in the combustion chamber. The computer system adjusts the first reasoning layer to enable it to make reasonable predictions based on the outputs of the component set more accurately.

[0049] During the calibration process, the computer system uses appropriate technical means. First, in terms of data, the computer system uses a large amount of training data with real temperature distribution labels. These data are obtained in actual experiments or simulations and contain the temperature distribution in the combustion chamber under various combustion conditions. The computer system inputs this data into the single parameter prediction network, allows the component set to process it, and then the first reasoning layer predicts the temperature distribution based on the output of the component set.

[0050] Next, the computer system calculates the prediction error. Assume that the temperature distribution predicted by the first inference layer is , and the actual temperature distribution is T, then the error calculation method can use the mean square error (MSE), the formula is: , where n represents the number of data samples, are the predicted temperature and true temperature of the i-th sample, respectively. Through this formula, the computer system can quantify the difference between the predicted value and the true value of the first inference layer.

[0051] The computer system then adjusts the first inference layer based on the calculated error. This is usually done through a back-propagation algorithm. The back-propagation algorithm calculates the gradient of the error with respect to various parameters in the network (such as weights and biases) by propagating the error signal from the output layer back to the input layer. The computer system adjusts the parameters of the first inference layer based on these gradients so that the error gradually decreases.

[0052] In actual operation, the computer system repeats the above process many times and uses different training data samples to calibrate the first inference layer. As the calibration progresses, the prediction accuracy of the first inference layer will continue to improve. For example, before the calibration begins, the temperature distribution predicted by the first inference layer may deviate greatly from the actual temperature distribution. After multiple calibrations, the prediction results will gradually approach the actual value, and the mean square error will continue to decrease.

[0053] Step 30: Fusing the component sets at the ath level in the X single parameter prediction networks to obtain a combustion parameter prediction network, the combustion parameter prediction network corresponds to the X parameter prediction branches, and the combustion parameter prediction network includes the first fused component set to the ath fused component set obtained by fusion, where a is a positive integer greater than 0.

[0054] Step 30 constructs a combustion parameter prediction network by fusing the component sets at the ath level in the X single parameter prediction networks. This process is crucial for integrating the advantages of multiple single parameter prediction networks and improving the overall network's prediction capability for multiple combustion parameters.

[0055] Taking X=3 as an example, assume that these three single-parameter prediction networks are used to predict the temperature distribution, pressure change and component concentration distribution in the combustion chamber respectively. At a certain level, that is, the a-th level, each single-parameter prediction network has a corresponding component set. For example, for the temperature prediction single-parameter network, the a-th level component set has extracted and processed the temperature-related features after the operation of the previous level; the a-th level component set of the pressure prediction single-parameter network has also processed the pressure-related information; the a-th level component set of the component concentration prediction single-parameter network has also completed the processing of concentration-related data.

[0056] The computer system fuses the component sets at the ath level in the three single parameter prediction networks. The fusion process is not superposition, but a comprehensive consideration of the output information of each component set to construct a new component set that can simultaneously process multiple combustion parameter prediction related features. Specifically, the computer system adopts a suitable fusion strategy. A common method is to splice the outputs of each component set. Assume that the output of the ath level component set of the temperature prediction network is a vector T with a dimension of m. a , the output of the a-th level component set of the pressure prediction network is a vector P with dimension n a , the output of the component concentration prediction network level a component set is a vector C with dimension p a , then the fused vector F a It can be expressed as F a =[T a , P a , C a ], whose dimension is m+n+p.

[0057] In this way, the computer system sequentially fuses the component sets at the ath level in the X single parameter prediction networks, gradually constructs a fusion component set starting from the first level, and finally obtains a combustion parameter prediction network. The combustion parameter prediction network corresponds to X parameter prediction branches, which are responsible for predicting different combustion parameters, namely, temperature distribution, pressure change, and component concentration distribution.

[0058] The combustion parameter prediction network includes the first fusion component set to the ath fusion component set obtained by fusion. These fusion component sets inherit some characteristics of the corresponding level component sets in the single parameter prediction network, and at the same time integrate the information of multiple single parameter networks, so that the network can analyze and predict the combustion parameters from multiple angles. For example, the first fusion component set fuses the outputs of the first layer component sets of each single parameter prediction network, and can comprehensively consider the impact of various combustion conditions on the combustion parameters; with the increase of levels, the subsequent fusion component sets continue to dig deeper and integrate more complex features, providing richer information for accurate prediction of combustion parameters.

[0059] Step 40: X second inference layers are set for the ath fusion component set, and different second inference layers are used to obtain predicted combustion parameters corresponding to different parameter prediction branches according to the output of the fusion component set.

[0060] After the computer system completes the fusion of the component sets at the ath level in the X single parameter prediction networks and obtains the combustion parameter prediction network including the ath fused component set, it sets up X second reasoning layers for the ath fused component set. The purpose of this operation is to enable the combustion parameter prediction network to accurately obtain the corresponding predicted combustion parameters for different parameter prediction branches based on the output of the fused component set.

[0061] To illustrate with a specific example, suppose X=4, that is, there are four single parameter prediction networks, which are used to predict the temperature distribution, pressure change, combustion efficiency and component concentration in the combustion chamber. After step 30, the computer system has fused the component sets at the ath level in the four single parameter prediction networks to form the ath fusion component set. At this time, in order to enable the combustion parameter prediction network to predict the four combustion parameters respectively, the computer system sets four second reasoning layers for the ath fusion component set.

[0062] Each second inference layer has a specific function, which is used to obtain the predicted combustion parameters corresponding to different parameter prediction branches based on the output of the fusion component set. For the temperature distribution parameter prediction branch, the corresponding second inference layer will analyze and process the output of the a-th fusion component set. The output of the fusion component set contains the comprehensive information from multiple single parameter prediction networks at the a-th level. This second inference layer will extract temperature-related features from it and obtain the predicted value of the temperature distribution through a specific calculation method. For example, the second inference layer may perform matrix multiplication based on some feature vectors in the output of the fusion component set and a pre-trained weight matrix. Assuming that the output vector of the fusion component set is O and the weight matrix is ​​W, the formula T=W×O (here T is the intermediate result related to temperature obtained by preliminary calculation) and then undergo some nonlinear transformations (such as using the ReLU activation function: f(x)=max(0, x)) finally obtains the predicted value of the temperature distribution in the combustion chamber.

[0063] Similarly, for the second inference layer corresponding to the pressure change parameter prediction branch, it will process the output of the fusion component set with different weight matrices and calculation logic to obtain the predicted value of the pressure change. The second inference layer will focus on the pressure-related information in the fusion output, and through corresponding calculations, such as using another set of weight matrices W' and the fusion output vector O to calculate P=W'×O (P is the intermediate result related to pressure), and then after appropriate transformation, the predicted result of the pressure change is obtained.

[0064] For the combustion efficiency and component concentration parameter prediction branches, the corresponding second reasoning layers will also operate on the output of the fusion component set in their own way. The second reasoning layer of the combustion efficiency prediction branch will calculate the predicted value of the combustion efficiency based on the characteristics related to the combustion efficiency in the fusion output, using specific algorithms and weights; the second reasoning layer of the component concentration prediction branch focuses on extracting information related to the component concentration and obtains the predicted value of the component concentration through corresponding calculations.

[0065] The computer system sets X second reasoning layers for the a-th fusion component set, so that the combustion parameter prediction network can accurately obtain the corresponding predicted combustion parameters according to the output of the fusion component set for different parameter prediction branches. This step further improves the structure of the combustion parameter prediction network, provides key support for the subsequent accurate prediction of multiple combustion parameters through the network, and helps to improve the accuracy and reliability of the entire combustion simulation analysis.

[0066] Step 50: The combustion parameter prediction network is calibrated by using the actual combustion parameters corresponding to the combustion condition training data of the ADN-based liquid propellant sample as prior labels, the predicted combustion parameters of the first inference layer as prior labels of the predicted combustion parameters of the second inference layer, and the predicted combustion parameters of the single parameter prediction network as prior labels of the predicted combustion parameters of the combustion parameter prediction network.

[0067] After the computer system completes the previous steps to build the combustion parameter prediction network and sets up each layer structure, it enters step 50 to further adjust the network. This step optimizes the performance of the combustion parameter prediction network by using a variety of prior label information, so that its prediction results are closer to the true value.

[0068] Take three single-parameter prediction networks, which are used to predict the temperature distribution, pressure change and component concentration distribution in the combustion chamber. For the combustion condition training data of ADN-based liquid propellant samples, the computer system has obtained the real combustion parameters corresponding to these samples under different combustion conditions. For example, under a series of specific combustion conditions, the real distribution data of the temperature in the combustion chamber, the actual data of the pressure change and the accurate values ​​of the concentration of each component are obtained through precise measurement. These real combustion parameters are used as important prior markers by the computer system.

[0069] In each single-parameter prediction network, the first inference layer has been calibrated and can output predicted combustion parameters. For example, the first inference layer of the temperature prediction single-parameter network obtains the predicted value of the temperature distribution in the combustion chamber based on the output of the network component set; the first inference layer of the pressure prediction single-parameter network also outputs the predicted result of pressure change; the first inference layer of the component concentration prediction single-parameter network also produces the corresponding predicted value. The computer system uses these predicted combustion parameters of the first inference layer as a priori marks for the second inference layer to predict the combustion parameters. This means that in the combustion parameter prediction network, when the second inference layer makes a prediction, it will refer to the more reliable prediction results that have been obtained by the first inference layer, and use this as a basis for further reasoning and optimization.

[0070] At the same time, the predicted combustion parameters when the single parameter prediction network runs independently are also used by the computer system as a priori markers for the combustion parameter prediction network to predict the combustion parameters. For example, the temperature single parameter prediction network independently predicts the temperature value of a certain area in the combustion chamber at a certain moment, and this value will be used as a reference for the combustion parameter prediction network to predict the temperature of the same area and time.

[0071] Based on these prior labels, the computer system adjusts the combustion parameter prediction network. The computer system uses a loss function to measure the difference between the predicted value and the prior label, such as the mean square error (MSE) loss function. The computer system uses the back propagation algorithm to adjust the weights of the combustion parameter prediction network based on the errors calculated by these loss functions. The back propagation algorithm calculates the gradient of the weights of each layer of the network based on the error to update the weights so that the error gradually decreases.

[0072] During the calibration process, the computer system continuously compares the predicted values ​​of the combustion parameter prediction network with various prior markers. For example, the temperature distribution predicted by the combustion parameter prediction network is compared with the actual temperature distribution, with the temperature distribution predicted by the first inference layer, and with the temperature distribution predicted by the single parameter prediction network. The weights of the network are continuously adjusted based on the errors generated by the comparison. After multiple iterations, the combustion parameter prediction network will gradually learn how to better utilize these prior marker information, thereby improving the accuracy of the prediction.

[0073] By using the actual combustion parameters corresponding to the combustion condition training data of ADN-based liquid propellant samples, the predicted combustion parameters of the first inference layer, and the predicted combustion parameters of the single parameter prediction network as prior markers, the computer system uses technical means such as loss function and back propagation algorithm to adjust the combustion parameter prediction network, so that the network can predict the combustion parameters more accurately, improve the performance and reliability of the network, and provide more accurate results for subsequent combustion simulation analysis based on the network.

[0074] In one implementation, the architecture of the a-th fusion component set has the same network architecture as the component set at the a-th level in the single parameter prediction network, and the initialized network parameter variables of the a-th fusion component set are the average values ​​of the network parameter variables of the component sets at the a-th level in the X single parameter prediction networks.

[0075] When the computer system constructs the combustion parameter prediction network, it involves the setting of the architecture of the a-th fusion component set. It is stipulated that the architecture of the a-th fusion component set has the same network architecture as the component set at the a-th level in the single parameter prediction network. This means that the two are consistent in the design of the network structure.

[0076] For example, suppose there are three single-parameter prediction networks, which are used to predict the temperature, pressure, and component concentration in the combustion chamber. Each single-parameter prediction network has multiple levels. When a=2, it is in the second level. If the single-parameter prediction network uses a fully connected layer architecture in the second layer, contains 100 neurons, is connected to the previous layer through a weight matrix, and uses the ReLU activation function to introduce nonlinear factors, then the a-th fusion component set will also use a fully connected layer architecture, and also set 100 neurons. The connection method and the use of the activation function are consistent with the second layer of the single-parameter prediction network. In addition, the initialization network parameter variables of the a-th fusion component set are the average value of the network parameter variables of the component set at the a-th level in the X single-parameter prediction networks.

[0077] In one implementation, step 50 uses the actual combustion parameters corresponding to the combustion condition training data of the ADN-based liquid propellant sample as a priori labels, uses the predicted combustion parameters of the first reasoning layer as the priori labels of the predicted combustion parameters of the second reasoning layer, and uses the predicted combustion parameters of the single parameter prediction network as the priori labels of the predicted combustion parameters of the combustion parameter prediction network to calibrate the combustion parameter prediction network, including: Step 51: Determine a parameter prediction error, a first probability distribution error, and a second probability distribution error, wherein the parameter prediction error represents the error between the predicted combustion parameter of the combustion parameter prediction network and the actual combustion parameter, the first probability distribution error represents the probability distribution gap between the predicted combustion parameter of the combustion parameter prediction network and the single parameter prediction network, and the second probability distribution error represents the probability distribution gap between the predicted combustion parameter of the first reasoning layer and the second reasoning layer; Step 52: According to the parameter prediction error, the first probability distribution error and the second probability distribution error, the ath fusion component set and the second reasoning layer in the combustion parameter prediction network are adjusted.

[0078] Step 51 is to determine the parameter prediction error, the first probability distribution error and the second probability distribution error. The parameter prediction error represents the error between the predicted combustion parameters of the combustion parameter prediction network and the actual combustion parameters. It intuitively reflects the degree of deviation between the network prediction result and the actual situation. For example, when predicting the temperature distribution in the combustion chamber, the temperature prediction value of a certain position output by the combustion parameter prediction network is , and the true temperature value of this location obtained by precise measurement is , then the parameter prediction error at this point can be expressed as For the prediction of the temperature distribution in the entire combustion chamber, the computer system calculates this error at all locations and uses a suitable statistical method (such as the mean square error, the formula is , where N is the total number of measurement locations, are the predicted temperature and true temperature of the ith position, respectively) to quantify the overall parameter prediction error.

[0079] The first probability distribution error indicates the difference in probability distribution between the combustion parameter prediction network and the single parameter prediction network. Probability distribution describes the possibility of a random variable taking different values. In the prediction of combustion parameters, different prediction networks may have different probability distributions for the prediction results of the same combustion parameter. Taking the prediction of combustion efficiency as an example, the combustion efficiency prediction value given by the combustion parameter prediction network has a certain probability distribution in different value ranges. , and the predicted value given by the single parameter prediction network responsible for predicting combustion efficiency also has a corresponding probability distribution Computer systems can use some metrics to calculate the difference between these two probability distributions, such as the KL divergence (Kullback-Leibler divergence), which is The smaller the KL divergence value, the closer the two probability distributions are, and the smaller the error of the first probability distribution is, which means that the combustion parameter prediction network and the single parameter prediction network are more similar in predicting the combustion parameter.

[0080] The second probability distribution error represents the difference in the probability distribution of the predicted combustion parameters between the first reasoning layer and the second reasoning layer. For example, when predicting pressure change, the first reasoning layer obtains the predicted value of pressure change based on the output of the single parameter prediction network component set, and its probability distribution is ; The second inference layer obtains the predicted value of pressure change based on the output of the fusion component set, and the probability distribution is Computer systems can also use methods similar to KL divergence (such as ) to calculate the difference between the two probability distributions and determine the second probability distribution error.

[0081] Step 52 is to adjust the a-th fusion component set and the second reasoning layer in the combustion parameter prediction network according to the parameter prediction error, the first probability distribution error and the second probability distribution error. First, the computer system determines the comprehensive error according to these errors. The determination of the comprehensive error is not to add, but to consider the importance of each error, that is, to perform weighted summation through the error influence coefficient. Assume that the first probability distribution error is , and its first error influence coefficient is ; The second probability distribution error is , the second error influence coefficient is ; The parameter prediction error is , the third error influence coefficient is ,and Then the comprehensive error The calculation formula is .

[0082] Next, the computer system adjusts the a-th fusion component set and the second inference layer in the combustion parameter prediction network based on the comprehensive error. In deep learning, a common method is to update the parameters of the network through the back propagation algorithm. The back propagation algorithm calculates the gradient of the error to the network parameters (such as weights and biases) based on the comprehensive error. Then, according to the gradient descent method, the step size of the weight update is controlled. The parameters of the second inference layer are also updated in a similar way.

[0083] In actual operation, the computer system repeats the above process many times. Each time, the three errors are calculated and the comprehensive error is determined, and the parameters of the a-th fusion component set and the second reasoning layer are updated. As the number of iterations increases, the comprehensive error will gradually decrease, which means that the prediction performance of the combustion parameter prediction network is constantly improving.

[0084] In one implementation, step 51, determining the parameter prediction error, the first probability distribution error, and the second probability distribution error, includes: Step 511: determining parameter prediction errors based on the predicted combustion parameters of the X parameter prediction branches and the actual combustion parameters corresponding to the X parameter prediction branches by the combustion parameter prediction network; Step 512: for the same parameter prediction branch, determining a first predicted combustion parameter of the parameter prediction branch by the combustion parameter prediction network, and a second predicted combustion parameter of the parameter prediction branch by the single parameter prediction network corresponding to the parameter prediction branch; determining a first probability distribution error according to the probability distribution of the first predicted combustion parameter and the second predicted combustion parameter; Step 513: For the same parameter prediction branch, determine the third predicted combustion parameter of the second reasoning layer corresponding to the parameter prediction branch, and the fourth predicted combustion parameter of the first reasoning layer corresponding to the parameter prediction branch; determine the second probability distribution error based on the probability distribution of the third predicted combustion parameter and the fourth predicted combustion parameter.

[0085] The combustion parameter prediction network contains X parameter prediction branches, each of which is responsible for predicting a specific combustion parameter, such as temperature distribution, pressure change, component concentration distribution, etc. For each parameter prediction branch, the computer system compares the network's predicted value with the actual value obtained by actual measurement.

[0086] Taking the prediction of temperature distribution in the combustion chamber as an example, it is assumed that the combustion parameter prediction network has 5 parameter prediction branches, one of which is dedicated to temperature prediction. In a simulation, the temperature prediction branch predicts the temperature at 100 different locations in the combustion chamber and obtains a set of predicted temperature values. At the same time, through high-precision temperature measurement equipment, the real temperature value set of these 100 locations is obtained. .

[0087] Computer systems can use a variety of methods to calculate parameter prediction errors. The most commonly used method is the mean square error (MSE) method, which is: , where n is the number of measurement locations, in this case n = 100. Through this formula, the computer system squares the difference between each predicted temperature value and the corresponding true temperature value, then sums and divides by the number of locations to obtain the parameter prediction error of the temperature prediction branch. For example, after calculation, MSE = 25, which means that there is a certain deviation between the predicted value of the temperature prediction branch and the true value.

[0088] For other parameter prediction branches, such as the pressure change prediction branch and the component concentration distribution prediction branch, the computer system will also use similar methods to calculate their respective parameter prediction errors. In actual operation, the computer system uses the functions provided by the programming language and the deep learning framework to implement these calculations.

[0089] In step 512, for the same parameter prediction branch, the first predicted combustion parameter of the parameter prediction branch by the combustion parameter prediction network and the second predicted combustion parameter of the branch by the corresponding single parameter prediction network of the parameter prediction branch are determined; and the first probability distribution error is determined based on the probability distribution of the first predicted combustion parameter and the second predicted combustion parameter. Each parameter prediction branch will generate a prediction result in the combustion parameter prediction network and the corresponding single parameter prediction network.

[0090] Continuing with the temperature prediction branch as an example, the combustion parameter prediction network predicts the temperature prediction branch and obtains a series of predicted temperature values, which constitute the first predicted combustion parameter set. The single parameter prediction network specifically used for temperature prediction predicts the same combustion conditions and obtains the second set of predicted combustion parameters: .

[0091] The computer system needs to analyze the probability distribution of these two sets of prediction values. First, the two sets of data will be statistically analyzed, for example, the temperature range will be divided into several intervals, and the frequency of occurrence of the prediction values ​​in each interval will be counted to obtain their probability distribution. Assume that the temperature range is [200, 1000] degrees Celsius, divided into 10 intervals: [200 - 300), [300 - 400), …, [900 - 1000]. The computer system counts the frequency of occurrence of the prediction values ​​of the combustion parameter prediction network in each interval to form a probability distribution. ; The frequency of occurrence of the predicted value of the single parameter prediction network in each interval forms a probability distribution .

[0092] In order to measure the difference between the two probability distributions, the computer system uses KL divergence (Kullback-Leibler divergence) to calculate the error of the first probability distribution, the formula is: Here x represents the temperature range. are the probability values ​​of two probability distributions in the x interval. For example, after calculation, we get , this value reflects the difference in probability distribution between the combustion parameter prediction network and the single parameter prediction network in temperature prediction, that is, the first probability distribution error.

[0093] For other parameter prediction branches, the computer system will also determine their prediction values ​​in the combustion parameter prediction network and the corresponding single parameter prediction network according to the above method, analyze their probability distribution, and calculate the first probability distribution error.

[0094] In step 513, for the same parameter prediction branch, the third predicted combustion parameter of the second reasoning layer corresponding to the parameter prediction branch and the fourth predicted combustion parameter of the first reasoning layer corresponding to the parameter prediction branch are determined; and the second probability distribution error is determined based on the probability distribution of the third predicted combustion parameter and the fourth predicted combustion parameter. In the combustion parameter prediction network, each parameter prediction branch has a corresponding first reasoning layer and a second reasoning layer, which generate prediction results respectively.

[0095] Taking the temperature prediction branch as an example, the second inference layer corresponding to this branch performs temperature prediction based on the output of the fusion component set to obtain the third predicted combustion parameter set The first inference layer corresponding to the branch performs temperature prediction based on the output of the single parameter prediction network component set to obtain the fourth predicted combustion parameter set .

[0096] The computer system performs probability distribution analysis on the two sets of predicted values ​​again. Similarly, the temperature range is divided into several intervals, and the frequency of the third predicted combustion parameter and the fourth predicted combustion parameter in each interval is counted to obtain the probability distribution respectively. , where y represents the temperature range.

[0097] The computer system then uses the KL divergence to calculate the second probability distribution error, as follows: . Assume that after calculation ,This value reflects the difference in probability distribution between the second reasoning layer and the first reasoning layer in temperature prediction, that is, the second probability distribution error.

[0098] For other parameter prediction branches, the computer system will repeat this process, determine the prediction values ​​of the second reasoning layer and the first reasoning layer respectively, analyze the probability distribution and calculate the second probability distribution error. Through steps 511-S513, the computer system comprehensively and meticulously determines the parameter prediction error, the first probability distribution error and the second probability distribution error. These error indicators reflect the performance of the combustion parameter prediction network and the difference with other related predictions from different angles. The parameter prediction error directly measures the degree of deviation between the network prediction value and the true value; the first probability distribution error compares the difference in the distribution of prediction results between the combustion parameter prediction network and the single parameter prediction network; the second probability distribution error evaluates the distribution difference of the prediction results of different reasoning layers within the combustion parameter prediction network. These error information provides accurate data support for the subsequent adjustment of the combustion parameter prediction network. The computer system can adjust the parameters and structure of the network in a targeted manner according to these errors to improve the prediction accuracy and performance of the network, so as to better serve the combustion simulation analysis of ADN-based liquid propellants.

[0099] In one implementation, step 52, based on the parameter prediction error, the first probability distribution error, and the second probability distribution error, the ath fusion component set and the second reasoning layer in the combustion parameter prediction network include: Step 521: determining a comprehensive error based on the parameter prediction error, the first probability distribution error, and the second probability distribution error; Step 522: Adjust the ath fusion component set and the second reasoning layer in the combustion parameter prediction network according to the comprehensive error.

[0100] In step 521, after obtaining the parameter prediction error, the first probability distribution error, and the second probability distribution error, the computer system assumes that the parameter prediction error is , which directly reflects the degree of deviation between the prediction results of the combustion parameter prediction network and the actual combustion parameters. This is a key indicator to measure the accuracy of network prediction and is of great significance to the evaluation of network performance. Therefore, it is given a larger weight and its third error influence coefficient is set to ; The first probability distribution error is , which reflects the difference between the combustion parameter prediction network and the single parameter prediction network in predicting the probability distribution of combustion parameters, and helps to analyze the performance of the network when integrating multi-source information. The first error influence coefficient is set to ; The second probability distribution error is , which reflects the difference in the probability distribution of the prediction results of the first and second reasoning layers in the combustion parameter prediction network. It is very important for optimizing the internal reasoning process of the network. The second error influence coefficient is set to , and satisfies .

[0101] Comprehensive error The calculation formula is: .

[0102] In step 522, the ath fusion component set and the second reasoning layer in the combustion parameter prediction network are adjusted according to the comprehensive error. After the computer system determines the comprehensive error, it uses the back propagation algorithm to adjust the ath fusion component set and the second reasoning layer in the combustion parameter prediction network. The core idea of ​​the back propagation algorithm is to calculate the gradient of the parameters (such as weights and biases) of each layer in the network according to the error, and then update the parameters according to these gradients so that the error gradually decreases.

[0103] During the actual calibration process, the computer system back-propagates the comprehensive error to each parameter of the a-th fusion component set and the second reasoning layer, and completes an iteration after a round of updates to all parameters. Then the new parameter prediction error, the first probability distribution error, and the second probability distribution error are calculated again, the comprehensive error is re-determined, and the next round of parameter updates is continued. As the number of iterations increases, the comprehensive error will gradually decrease. For example, after the first round of iterations, the comprehensive error dropped from 0.165 to 0.15; after multiple rounds of iterations, the comprehensive error may be reduced to 0.1 or even lower. This shows that the parameters of the a-th fusion component set and the second reasoning layer in the combustion parameter prediction network are constantly adjusted and optimized, and the prediction performance of the network is gradually improving.

[0104] In one implementation, step 521, determining a comprehensive error based on the parameter prediction error, the first probability distribution error, and the second probability distribution error, includes: Step 5211: Determine the comprehensive error based on the first probability distribution error, the first error influence coefficient of the first probability distribution error, the second probability distribution error, the second error influence coefficient of the second probability distribution error, the parameter prediction error, and the third error influence coefficient of the parameter prediction error; wherein the third error influence coefficient is greater than the second error influence coefficient, and the second error influence coefficient is greater than the first error influence coefficient.

[0105] The computer system determines the comprehensive error based on the first probability distribution error, the first error influence coefficient of the first probability distribution error, the second probability distribution error, the second error influence coefficient of the second probability distribution error, the parameter prediction error, and the third error influence coefficient of the parameter prediction error. This process involves multiple error indicators and their corresponding weight coefficients, and comprehensively measures the performance deviation of the combustion parameter prediction network through a specific calculation method.

[0106] The first probability distribution error reflects the difference in the probability distribution of the predicted combustion parameters between the combustion parameter prediction network and the single parameter prediction network, and reflects the consistency of the network when integrating information from different prediction sources. For example, when predicting the pressure change in the combustion chamber, the pressure prediction value obtained by the combustion parameter prediction network presents a certain probability distribution in different pressure ranges, and the single parameter prediction network also has a corresponding probability distribution. The difference between the two is the first probability distribution error. Assume that the first probability distribution error is expressed by E 1 The first error influence coefficient of the first probability distribution error is a weight value used to measure the importance of the first probability distribution error in the comprehensive error. This coefficient is set according to the goal of network calibration and the degree of attention to different errors, and is represented by α. For example, if you want to focus on the consistency of the probability distribution between the combustion parameter prediction network and the single parameter prediction network, you can increase the value of α appropriately.

[0107] The second probability distribution error refers to the difference in the probability distribution of the predicted combustion parameters between the first and second reasoning layers, which helps to evaluate the stability and consistency of the prediction results of different reasoning levels within the network. For example, for component concentration prediction, the difference in the probability distribution of the predicted values ​​of the first and second reasoning layers in different concentration ranges is the second probability distribution error, which is set as E 2 .

[0108] The second error influence coefficient of the second probability distribution error is used to determine the weight of the second probability distribution error in the comprehensive error, which is denoted as β. This coefficient can be adjusted according to the demand for optimizing the internal reasoning structure of the network.

[0109] The parameter prediction error directly reflects the error between the predicted combustion parameters of the combustion parameter prediction network and the actual combustion parameters. It is the core indicator for measuring the accuracy of network prediction and is represented by E_{3}. For example, when predicting the combustion efficiency, the difference between the network prediction value and the actual measured combustion efficiency is the parameter prediction error.

[0110] The third error influence coefficient of the parameter prediction error is the coefficient that gives the parameter prediction error a weight in the comprehensive error calculation, represented by γ. Since the parameter prediction error is directly related to the accuracy of network prediction, γ usually takes the largest value among the three influence coefficients, that is, γ>β>α.

[0111] The computer system determines the comprehensive error by the following formula .

[0112] By executing step 5211, the computer system can comprehensively consider multiple error factors and their weights and accurately calculate the comprehensive error. This comprehensive error comprehensively reflects the performance of the combustion parameter prediction network in different aspects, and provides an accurate quantitative basis for the subsequent adjustment of the a-th fusion component set and the second reasoning layer in the network based on the error, which helps to gradually optimize the network performance and improve its prediction accuracy of the combustion parameters of ADN-based liquid propellants.

[0113] In one implementation, step 522, predicting the ath fusion component set and the second reasoning layer in the combustion parameter prediction network according to the comprehensive error, includes: Step 5221: Perform error feedback according to the comprehensive error to iterate the a-th fusion component set and the network parameters of the second reasoning layer; The method also includes: Step 5222: After completing the error feedback iteration of the ath fusion component set and the network parameters of the second inference layer based on the comprehensive error, error feedback is performed based on the second probability distribution error to iterate the network parameters of the second inference layer.

[0114] In step 5221, error feedback is performed based on the comprehensive error to iterate the network parameters of the a-th fusion component set and the second inference layer. After obtaining the comprehensive error, the computer system will use the back propagation algorithm to transfer the error from the output layer to the front layer of the network to calculate the gradient of the error to the a-th fusion component set and the second inference layer network parameters (such as weights and biases).

[0115] Taking a simplified neural network structure as an example, assume that the a-th fusion component set in the combustion parameter prediction network contains a fully connected layer, with 3 input neurons and 2 output neurons, the connection weight matrix is ​​W, and the bias vector is b. The second inference layer is also a fully connected layer, receiving the output of the a-th fusion component set, with 2 input neurons and 1 output neuron, the weight matrix is ​​V, and the bias is c.

[0116] When the computer system passes the input data into the network, a series of calculations are performed to obtain the predicted results, which are compared with the true values ​​to obtain the comprehensive error E. total According to the back propagation algorithm, the error will propagate backwards from the output layer. First, the gradient of the error to the output of the second inference layer is calculated. ( is the output of the second inference layer). Then, the chain rule is used to calculate the gradient of the error with respect to the weight V and bias c of the second inference layer. For example, for the weight (i represents the input neuron index, j represents the output neuron index), its gradient For bias c, the gradient .

[0117] Next, the error continues to back propagate to the a-th fusion component set. Calculate the gradient of the error with respect to the output of the a-th fusion component set (h is the output of the a-th fusion component set), and then the gradient of the weight W and bias b is calculated by the chain rule. For example, the weight (k represents the input neuron index, l represents the output neuron index) , the gradient of bias b .

[0118] After calculating these gradients, the computer system updates the network parameters according to the gradient descent method. For the weight V of the second inference layer, the update formula is ; The update formula of bias c is For the weight W of the a-th fusion component set, the update formula is ; The update formula of bias b is ,in is the learning rate.

[0119] Through error feedback and parameter iteration, the network parameters of the a-th fusion component set and the second inference layer are continuously adjusted, so that the comprehensive error is gradually reduced and the prediction performance of the network is gradually improved.

[0120] In step 5222, after completing the error feedback iteration of the a-th fusion component set and the network parameters of the second reasoning layer based on the comprehensive error, error feedback is performed based on the second probability distribution error to iterate the network parameters of the second reasoning layer. This step is to further fine-tune the second reasoning layer to optimize the consistency and stability of the reasoning process within the network. After completing the parameter update based on the comprehensive error, the computer system considers the second probability distribution error alone. Reviewing the second probability distribution error, it is the probability distribution gap between the predicted combustion parameters of the first reasoning layer and the second reasoning layer.

[0121] The computer system again starts from the second probability distribution error and uses the back propagation algorithm to calculate the gradient of the error to the parameters of the second inference layer network. Through this additional error back propagation and parameter iteration, the computer system specifically optimizes the second inference layer for the second probability distribution error. This helps to further improve the stability of the prediction results of the second inference layer and the consistency with the first inference layer, thereby improving the reliability of the internal reasoning of the entire combustion parameter prediction network.

[0122] Steps 5221 and S5222 are key operations for the computer system to calibrate the combustion parameter prediction network. Through comprehensive parameter iteration based on comprehensive errors and fine-tuning of the second inference layer based on the second probability distribution error, the prediction performance of the network is continuously improved, and the combustion parameters of ADN-based liquid propellants can be predicted more accurately, providing more reliable results for subsequent combustion simulation analysis, and promoting the development of related research and applications.

[0123] In one implementation, the method further includes: Step 523: after the calibration of the ath fusion component set is completed, determining the combustion parameter prediction network effect metric of the combustion parameter prediction network on the validation data set; Step 524: When the network effect metric meets the fusion requirement, based on the combustion parameter prediction network, the component sets at the a+1th level in the X single parameter prediction networks are fused.

[0124] In step 523, after the calibration of the a-th fusion component set is completed, the combustion parameter prediction network effect metric (such as accuracy, recall, F1 score, etc.) of the combustion parameter prediction network on the validation data set is determined. After the computer system completes the calibration of the a-th fusion component set, a quantitative indicator is needed to evaluate the performance of the current combustion parameter prediction network, which requires effect measurement on the validation data set.

[0125] The validation data set is a set of pre-prepared data that is different from the training data but has similar characteristics. Its function is to evaluate the generalization ability of the network without affecting the network training. Taking the prediction of the temperature distribution in the combustion chamber as an example, after the computer system completes the adjustment of the a-th fusion component set, it inputs the data related to the combustion conditions in the validation data set into the combustion parameter prediction network. After calculation, the network outputs the prediction result of the temperature distribution in the combustion chamber.

[0126] Computer systems use a variety of performance metrics to evaluate prediction results. Accuracy is a common metric that measures the proportion of correctly predicted samples to the total number of samples. Recall focuses on the proportion of samples that are correctly predicted as positive among all samples that are actually positive. The F1 score is the harmonic mean of accuracy and recall, which combines the information of both and more comprehensively reflects the performance of the network.

[0127] In addition to the above indicators, the computer system may also use other evaluation indicators, such as mean square error (MSE) to measure the error between the predicted temperature and the actual temperature. Through these different evaluation indicators, the computer system can fully understand the performance of the combustion parameter prediction network on the validation data set from multiple perspectives.

[0128] In step 524, when the network effect metric meets the fusion requirements, the component set at the a+1th level in the X single parameter prediction networks is fused based on the combustion parameter prediction network. When the effect metric index of the combustion parameter prediction network calculated by the computer system on the verification data set meets the preset fusion requirements, the next step of network construction operation is entered.

[0129] The fusion requirement is a standard set according to actual needs and expectations for network performance. For example, the F1 score needs to be above 0.75, or the mean square error needs to be lower than a certain value. When the indicators of the combustion parameter prediction network meet these requirements, it means that the current network is in good condition and has the conditions for further fusion.

[0130] Taking X=3 single parameter prediction networks (predicting temperature, pressure and component concentration respectively) as an example, after the combustion parameter prediction network has been adjusted based on the ath fusion component set and the effect measurement meets the requirements, the computer system begins to fuse the component sets at the a+1th level in the three single parameter prediction networks.

[0131] For the temperature prediction single-parameter network, the component set at the a+1th level will further extract and transform the data processed by the previous levels. Assuming that this level is a convolutional layer, it will slide the convolution kernel on the data to extract higher-level features related to temperature. The component sets at the a+1th level of the pressure prediction single-parameter network and the component concentration prediction single-parameter network will also perform similar operations to extract features related to pressure and component concentration, respectively.

[0132] The computer system uses a suitable fusion method to fuse the component sets of the a+1th level of the three single parameter prediction networks. One method is concatenation. For example, the output of the component set of the a+1th level of the temperature prediction network is a feature vector of dimension m, the output of the pressure prediction network at this level is a feature vector of dimension n, and the output of the component concentration prediction network at this level is a feature vector of dimension p. The computer system concatenates these vectors together in order to form a new feature vector of dimension m+n+p as the fusion result.

[0133] During the fusion process, the computer system needs to ensure that the component set outputs of different single parameter prediction networks are compatible in terms of data type, dimension, etc. If they are not compatible, some data conversion operations may be required, such as data normalization, dimension adjustment, etc. After the fusion is completed, a new (a+1)th fusion component set is formed. This set integrates the information of multiple single parameter prediction networks at the a+1th level, providing a richer feature representation for the combustion parameter prediction network. Then, the computer system continues to build and adjust the combustion parameter prediction network based on this new fusion component set, repeating the previous training and evaluation process to continuously improve the performance of the network.

[0134] In one implementation, the method provided by the embodiment of the present invention further includes: Step 525: Determine a single parameter prediction network effect metric of the single parameter prediction network on the validation data set; Step 526: When the metric difference between the combustion parameter prediction network effect metric and the single parameter prediction network effect metric is less than the preset metric difference, it is determined that the network effect metric meets the fusion requirement.

[0135] In step 525, the single parameter prediction network effect metric of the single parameter prediction network on the validation data set is determined. In the process of calibrating the combustion parameter prediction network, the computer system needs to clarify the performance of the single parameter prediction network on the validation data set as an important basis for comparison and evaluation. For each single parameter prediction network, the computer system inputs the data related to the combustion parameter that the network is responsible for predicting in the validation data set. Taking the single parameter prediction network for predicting the pressure change in the combustion chamber as an example, the validation data set contains relevant data under various combustion conditions, such as the flow rate of the propellant, the injection angle, the initial state of the combustion chamber and other information. The computer system inputs these data into the pressure prediction single parameter network, and after a series of calculations and processing, the network outputs the prediction results of the pressure change in the combustion chamber.

[0136] The computer system uses a variety of indicators to measure the effectiveness of the single parameter prediction network. Similar to the combustion parameter prediction network, the indicators include accuracy, recall, F1 score, etc.

[0137] The computer system also evaluates the single-parameter prediction network for predicting the temperature distribution in the combustion chamber and the single-parameter prediction network for predicting the component concentration in the same way, calculating their various effectiveness metrics on their respective validation data subsets.

[0138] In step 526, when the metric difference between the combustion parameter prediction network effect metric and the single parameter prediction network effect metric is less than the preset metric difference, it is determined that the network effect metric meets the fusion requirement. After respectively calculating the effect metric indicators of the combustion parameter prediction network and the single parameter prediction network on the validation data set, the computer system will compare and analyze the two.

[0139] The preset metric difference is a standard value set according to actual needs and expectations for network performance. This value is used to determine whether the combustion parameter prediction network is sufficiently better than the single parameter prediction network in performance, so as to determine whether it meets the requirements for further fusion. For example, the preset metric difference is set to 0.1.

[0140] The computer system calculates the difference between the combustion parameter prediction network and the single parameter prediction network on the same evaluation index. When the metric difference of all relevant evaluation indicators calculated by the computer system is less than the preset metric difference, it can be determined that the network effect metric meets the fusion requirements.

[0141] By accurately determining the effect measure of the single parameter prediction network on the verification data set through step 525, and judging whether the network effect measure meets the fusion requirements through contrast difference in step 526, the computer system can comprehensively evaluate the performance of the combustion parameter prediction network to ensure that the network can achieve a good balance between performance improvement and structural optimization when further fusion operations are performed, thereby laying a solid foundation for more accurate prediction of the combustion parameters of ADN-based liquid propellants and promoting the continuous optimization and improvement of combustion simulation analysis methods.

[0142] In one implementation, the execution data of the component set at the a+1th level is the output data of the component set at the ath level; and the method further includes: Step 527: When the network effect metric does not meet the fusion requirement, the fusion of the component set at the ath level in the X single parameter prediction networks is abandoned, and the component set fusion is stopped.

[0143] When the computer system completes the comparison between the combustion parameter prediction network effect metric and the single parameter prediction network effect metric, if it is found that the metric difference does not meet the preset metric difference, that is, the network effect metric does not meet the fusion requirements, step 527 will be executed. For example, suppose that when evaluating the network related to the prediction of the pressure in the combustion chamber, the preset metric difference is set to 0.1. After calculation, the accuracy of the combustion parameter prediction network in pressure prediction is 0.7, while the accuracy of the single parameter network for pressure prediction is 0.8. The metric difference between the two is 0.1, and the metric differences of other evaluation indicators such as recall rate and F1 score do not meet the requirements. This shows that the current combustion parameter prediction network is not significantly better than the single parameter prediction network in pressure prediction, and the overall network effect metric does not meet the fusion requirements.

[0144] In this case, the computer system abandons the fusion operation on the component set at the ath level in the X single-parameter prediction networks. Continuing with the pressure prediction as an example, it was originally planned to fuse the component set at the ath level in the X single-parameter prediction networks (assuming X=3, which are the temperature, pressure, and component concentration prediction networks respectively). Since the network effect measurement does not meet the requirements, the computer system stops this fusion behavior. This means that the component set at the ath level of the pressure prediction single-parameter network will not be spliced ​​or other fused with the component sets at the ath level of the other two single-parameter prediction networks, avoiding the risk of network performance deterioration due to unreasonable fusion.

[0145] By executing step 527, the computer system stops the loss in time when the network effect is not ideal, ensuring that the adjustment of the combustion parameter prediction network is carried out in the direction of improving performance. This helps to save computing resources, avoid wasting time and computing power due to unnecessary fusion operations, and ensure the stability and rationality of the network structure, providing a basis for possible subsequent adjustment of the adjustment strategy, so as to better achieve accurate prediction of the combustion parameters of ADN-based liquid propellants.

[0146] In one implementation, the output layer of the combustion parameter prediction network includes prediction output layers of X single parameter prediction networks; the method further includes: Step 60: After the fusion component set in the combustion parameter prediction network is calibrated, the parameters of each prediction output layer in the combustion parameter prediction network are fine-tuned according to the targeted parameter prediction errors corresponding to each parameter prediction branch.

[0147] After the adjustment of the fusion component set in the combustion parameter prediction network is completed, it means that the overall network architecture and most of the key parts have been optimized. However, in order to further improve the prediction accuracy of the network for different combustion parameters, it is necessary to fine-tune the parameters of each prediction output layer.

[0148] Each parameter prediction branch corresponds to a specific combustion parameter prediction task, such as the temperature distribution prediction branch, the pressure change prediction branch, etc. The targeted parameter prediction error refers to the error between the prediction result of each parameter prediction branch and the actual combustion parameter. The computer system calculates this error for each parameter prediction branch separately.

[0149] Taking the temperature distribution prediction branch as an example, the computer system inputs a set of test data containing multiple combustion conditions into the combustion parameter prediction network. After a series of operations, the temperature distribution prediction branch outputs the predicted temperature distribution results. These predicted results are compared with the actual temperature distribution obtained by actual measurement, and the error is calculated. For example, the mean square error (MSE) is used for calculation. For the pressure change prediction branch, the corresponding test data is also input to calculate the error between its predicted results and the actual pressure change.

[0150] The computer system fine-tunes the parameters of each prediction output layer based on these targeted parameter prediction errors. The parameters of the prediction output layer mainly include weights and biases.

[0151] In one implementation, the output layer of the combustion parameter prediction network includes prediction output layers of X single parameter prediction networks; step 40, setting X second inference layers for the a-th fusion component set, includes: Step 41: When the component set at the a-th level is not the last component set before the parameter prediction branch output layer, set X second inference layers for the a-th fusion component set; The method provided in the embodiment of the present invention also includes: Step 42: When the component set at the a-th level is the last component set before the parameter prediction branch output layer, the actual combustion parameters corresponding to the combustion condition training data of the ADN-based liquid propellant sample are used as prior labels, and the predicted combustion parameters of the single parameter prediction network are determined as the prior labels of the predicted combustion parameters of the combustion parameter prediction network, and the combustion parameter prediction network is calibrated.

[0152] In step 41, when the component set of the ath level is not the last component set before the parameter prediction branch output layer, X second reasoning layers are set for the ath fusion component set. In the process of constructing the combustion parameter prediction network, the hierarchical structure is built step by step, and the component set of each level plays a specific role. When the computer system determines that the component set of the ath level is not the last component set before the parameter prediction branch output layer, this step is executed.

[0153] Take a simplified combustion parameter prediction network as an example. Assume that the network has multiple levels for predicting the three combustion parameters of temperature distribution, pressure change and component concentration in the combustion chamber, that is, X = 3. At a certain stage of the network, the component set of the ath level completes the feature extraction and preliminary processing of the input data, but there is still a certain level gap from the output layer.

[0154] For the temperature parameter prediction branch, the computer system sets a second inference layer for temperature prediction for the a-th fusion component set. This layer receives the output of the a-th fusion component set and further analyzes and infers the features related to temperature. For example, the output of the a-th fusion component set is a feature vector containing a variety of combustion-related information. The second inference layer of temperature prediction will perform specific calculations on this vector. Assume that the feature vector is , the second inference layer may be through matrix multiplication (in is the weight matrix of the second inference layer for temperature prediction, It is an intermediate result related to temperature after preliminary calculation), and then undergoes nonlinear transformation (such as ReLU function: f(x)=max(0, x)) to further extract and enhance temperature-related features, in preparation for the subsequent accurate prediction of temperature distribution.

[0155] For the pressure parameter prediction branch and the component concentration parameter prediction branch, the computer system will also set up corresponding second reasoning layers for the a-th fusion component set. The second reasoning layer of pressure prediction uses different weight matrices Operate with the output of the a-th fusion component set , and through the corresponding nonlinear transformation, process the information related to pressure; the second reasoning layer of component concentration prediction uses the weight matrix conduct etc. to focus on the analysis of component concentration-related features.

[0156] In step 42, when the component set of the a-th level is the last component set before the parameter prediction branch output layer, the actual combustion parameter corresponding to the combustion condition training data of the ADN-based liquid propellant sample is used as a priori mark, and the predicted combustion parameter of the single parameter prediction network is determined as the priori mark of the predicted combustion parameter of the combustion parameter prediction network, and the combustion parameter prediction network is adjusted. When the component set of the a-th level is at the last position before the parameter prediction branch output layer, the situation is different. At this time, the computer system uses a specific priori mark to adjust the combustion parameter prediction network.

[0157] Still taking the above network containing three parameter prediction branches as an example, when the component set of the ath level is at the end before the output layer of the temperature parameter prediction branch, the computer system uses the real temperature distribution data of the ADN-based liquid propellant sample under various combustion conditions as a priori mark. These real data are obtained through precise experimental measurements and reflect the temperature conditions in the actual combustion process. At the same time, the computer system also uses the predicted combustion parameters of the temperature single parameter prediction network for the same combustion conditions as a priori marks.

[0158] For the temperature parameter prediction branch of the combustion parameter prediction network, the computer system adjusts according to these priori marks. Assume that the temperature distribution prediction value of the combustion parameter prediction network under a certain combustion condition is , and the true temperature distribution is , the prediction value of the temperature single parameter prediction network is The computer system calculates the error between the predicted value and the true value, for example using the mean square error (MSE) formula (n is the number of measurement locations), and a measure of the difference between the predicted value and the predicted value of the single-parameter prediction network (such as KL divergence: ,in are the probability distributions of temperature values ​​predicted by the combustion parameter prediction network and the single parameter prediction network respectively).

[0159] Based on these error and difference metrics, the computer system adjusts the parameters related to the temperature parameter prediction branch in the combustion parameter prediction network. The weight w and bias b in the network are adjusted in a manner similar to the gradient descent method. By continuously adjusting these parameters, the temperature prediction results of the combustion parameter prediction network gradually approach the true value and the predicted value of the single parameter prediction network, thereby improving the accuracy and stability of the network.

[0160] For the pressure parameter prediction branch and the component concentration parameter prediction branch, the computer system will also use the respective real combustion parameters and the predicted combustion parameters of the single parameter prediction network as prior marks, calculate the error and difference metrics, and adjust the corresponding network parameters in the above manner.

[0161] In one implementation, step 20, calibrating the first inference layers corresponding to the component sets in the single parameter prediction network, includes: Step 21: setting the first inference layer corresponding to each component set in the single parameter prediction network; Step 22: determining the inference error of the first inference layer according to the predicted combustion parameters of the first inference layer corresponding to the a-th component set in the single parameter prediction network and the actual combustion parameters; Step 23: Adjust the first inference layer according to the inference error.

[0162] In step 21, the first reasoning layers corresponding to the component sets in the single parameter prediction network are set respectively. When constructing a single parameter prediction network, the computer system will equip each component set with a dedicated first reasoning layer according to the network structure and the requirements of the prediction task. The component set can be understood as a combination of neurons at a certain level in the network, which jointly process the input data and extract features that are valuable for predicting combustion parameters. Take a single parameter prediction network for predicting the temperature distribution in a combustion chamber as an example. The network contains multiple levels. In a certain level, the component set receives the output data from the previous level, which may contain various information related to combustion, such as the composition of the propellant, the injection velocity, the geometry of the combustion chamber, etc. A first reasoning layer is set for this component set. The role of this first reasoning layer is to further analyze and reason based on the features output by the component set to obtain prediction results related to the temperature distribution.

[0163] In step 22, the inference error of the first inference layer is determined based on the predicted combustion parameters of the first inference layer corresponding to the a-th component set in the single parameter prediction network and the actual combustion parameters. After the first inference layer is set for the component set, the inference error needs to be calculated to evaluate the prediction effect of the first inference layer. The calculation method of the mean square error can be referred to above and will not be repeated here.

[0164] In step 23, the first inference layer is adjusted according to the inference error. After calculating the inference error of the first inference layer, the first inference layer is adjusted according to these errors to improve its prediction accuracy. Taking the first inference layer of the temperature prediction single parameter network as an example, assuming that the mean square error obtained by calculation is large, it means that the prediction result of the first inference layer has a large deviation from the true value and needs to be adjusted. The gradient descent method is used to adjust the parameters of the first inference layer, such as weights and biases.

[0165] For the weight W, assume that its adjustment formula is ,in is the learning rate, which controls the step size of weight adjustment. The learning rate is an important hyperparameter. If it is set too large, the weight update speed will be too fast, which may cause the network to fail to converge; if it is set too small, the network training speed will be very slow. For example, assuming the current weight If one element of is 0.5, we can calculate The value at this element is 0.2, and the learning rate =0.01, then the updated weight The value at this element is 0.5 - 0.01×0.2=0.498.

[0166] For bias b, the adjustment formula is For example, the current bias =0.3, =0.1, learning rate =0.01, then the updated bias =0.3 - 0.01×0.1=0.299.

[0167] By repeating this process many times, input data under different combustion conditions is passed into the network, the inference error is calculated, and then the weights and biases are adjusted according to the error. After each adjustment, the inference error will gradually decrease, and the prediction accuracy of the first inference layer will continue to improve.

[0168] An embodiment of the present invention provides a computer system, such as Figure 3As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present invention.

[0169] An embodiment of the present invention provides a computer system, and the computer system in the embodiment of the present invention includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above-mentioned combustion simulation analysis method based on ADN-based liquid propellant is implemented.

Claims

1. A combustion simulation analysis method based on ADN-based liquid propellant, characterized in that: The method comprises: In response to the simulation analysis instruction, acquiring combustion condition data of the target ADN-based liquid propellant; Loading the combustion condition data of the target ADN-based liquid propellant into a pre-calibrated combustion parameter prediction network, wherein the combustion parameter prediction network is a joint learning network including a plurality of different branch parameter prediction networks, and a branch parameter prediction network is used to predict a combustion parameter; Outputting at least one combustion parameter through the combustion parameter prediction network; Visualization processing is performed according to the at least one combustion parameter to obtain a simulation result of the target ADN-based liquid propellant.

2. The method according to claim 1, characterized in that The combustion parameter prediction network is calibrated by the following steps: Obtain X single parameter prediction networks, where different single parameter prediction networks are used to predict different combustion parameters, and component sets at the same level in different single parameter prediction networks have the same network architecture, and X is a positive integer greater than 1; Adjusting the first reasoning layers respectively corresponding to the component sets in the single parameter prediction network, wherein the first reasoning layers are used to obtain predicted combustion parameters according to outputs of the component sets; Fusing the component sets at the ath level in the X single parameter prediction networks to obtain a combustion parameter prediction network, wherein the combustion parameter prediction network corresponds to the X parameter prediction branches, and the combustion parameter prediction network includes the first fused component set to the ath fused component set obtained by fusion, where a is a positive integer greater than 0; Setting X second inference layers for the a-th fusion component set, wherein different second inference layers are used to obtain predicted combustion parameters corresponding to different parameter prediction branches according to outputs of the fusion component set; The combustion parameter prediction network is calibrated by using the actual combustion parameters corresponding to the combustion condition training data of the ADN-based liquid propellant sample as a priori labels, using the predicted combustion parameters of the first reasoning layer as the priori labels of the predicted combustion parameters of the second reasoning layer, and using the predicted combustion parameters of the single parameter prediction network as the priori labels of the predicted combustion parameters of the combustion parameter prediction network; the architecture of the a-th fusion component set has the same network architecture as the component set at the a-th level in the single parameter prediction network, and the initialized network parameter variables of the a-th fusion component set are the average values ​​of the network parameter variables of the X component sets at the a-th level in the single parameter prediction network.

3. The method according to claim 2, characterized in that The method uses the actual combustion parameters corresponding to the combustion condition training data of the ADN-based liquid propellant sample as a priori marks, uses the predicted combustion parameters of the first reasoning layer as the priori marks of the predicted combustion parameters of the second reasoning layer, and uses the predicted combustion parameters of the single parameter prediction network as the priori marks of the predicted combustion parameters of the combustion parameter prediction network to adjust the combustion parameter prediction network, including: Determine a parameter prediction error, a first probability distribution error, and a second probability distribution error, wherein the parameter prediction error represents an error between a predicted combustion parameter of the combustion parameter prediction network and the actual combustion parameter, the first probability distribution error represents a probability distribution gap between the predicted combustion parameter of the combustion parameter prediction network and the single parameter prediction network, and the second probability distribution error represents a probability distribution gap between the predicted combustion parameter of the first reasoning layer and the second reasoning layer; The ath fusion component set and the second reasoning layer in the combustion parameter prediction network are calibrated according to the parameter prediction error, the first probability distribution error and the second probability distribution error.

4. The method according to claim 3, characterized in that The determining of the parameter prediction error, the first probability distribution error and the second probability distribution error comprises: Determining the parameter prediction error according to the predicted combustion parameters of the X parameter prediction branches by the combustion parameter prediction network and the actual combustion parameters corresponding to the X parameter prediction branches; For the same parameter prediction branch, determining a first predicted combustion parameter of the parameter prediction branch by the combustion parameter prediction network, and a second predicted combustion parameter of the parameter prediction branch by the single parameter prediction network corresponding to the parameter prediction branch; determining the first probability distribution error according to the probability distribution of the first predicted combustion parameter and the second predicted combustion parameter; For the same parameter prediction branch, determine the third predicted combustion parameter of the second reasoning layer corresponding to the parameter prediction branch, and the fourth predicted combustion parameter of the first reasoning layer corresponding to the parameter prediction branch; determine the second probability distribution error based on the probability distribution of the third predicted combustion parameter and the fourth predicted combustion parameter.

5. The method according to claim 3, characterized in that: The a-th fusion component set and the second reasoning layer in the combustion parameter prediction network are fused according to the parameter prediction error, the first probability distribution error and the second probability distribution error, including: Determining a comprehensive error based on the parameter prediction error, the first probability distribution error, and the second probability distribution error; The ath fusion component set and the second reasoning layer in the combustion parameter prediction network are adjusted according to the comprehensive error.

6. The method according to claim 5, characterized in that The determining of the comprehensive error according to the parameter prediction error, the first probability distribution error and the second probability distribution error comprises: Determine the comprehensive error according to the first probability distribution error, the first error influence coefficient of the first probability distribution error, the second probability distribution error, the second error influence coefficient of the second probability distribution error, the parameter prediction error, and the third error influence coefficient of the parameter prediction error; wherein the third error influence coefficient is greater than the second error influence coefficient, and the second error influence coefficient is greater than the first error influence coefficient; The a-th fusion component set and the second reasoning layer in the combustion parameter prediction network according to the comprehensive error include: Performing error feedback according to the comprehensive error to iterate the a-th fusion component set and the network parameter variables of the second reasoning layer; The method further comprises: After completing the error feedback iteration of the a-th fusion component set and the network parameter variables of the second reasoning layer according to the comprehensive error, performing error feedback according to the second probability distribution error to iterate the network parameter variables of the second reasoning layer; The method further comprises: After the calibration of the a-th fusion component set is completed, determining a combustion parameter prediction network effect metric of the combustion parameter prediction network on a validation data set; When the network effect metric meets the fusion requirement, based on the combustion parameter prediction network, the component sets at the a+1th level in the X single parameter prediction networks are fused.

7. The method according to claim 6, characterized in that The method further comprises: Determine a single parameter prediction network effect metric of the single parameter prediction network on the validation data set; When the metric difference between the combustion parameter prediction network effect metric and the single parameter prediction network effect metric is less than a preset metric difference, determining that the network effect metric meets the fusion requirement; The execution data of the component set at the a+1th level is the output data of the component set at the ath level; The method further comprises: When the network effect metric does not meet the fusion requirement, the fusion of the component set at the a-th level in the X single parameter prediction networks is abandoned, and the component set fusion is stopped.

8. The method according to claim 3, characterized in that The output layer of the combustion parameter prediction network includes X prediction output layers of the single parameter prediction networks; The method further comprises: After the adjustment of the fusion component set in the combustion parameter prediction network is completed, the parameter variables of each prediction output layer in the combustion parameter prediction network are fine-tuned according to the targeted parameter prediction errors corresponding to each parameter prediction branch.

9. The method according to any one of claims 1 to 8, characterized in that: The output layer of the combustion parameter prediction network includes X prediction output layers of the single parameter prediction networks; The step of setting X second reasoning layers for the a-th fusion component set includes: When the component set at the a-th level is not the last component set before the parameter prediction branch output layer, setting X second inference layers for the a-th fusion component set; The method further comprises: When the component set at the a-th level is the last component set before the output layer of the parameter prediction branch, the actual combustion parameter corresponding to the combustion condition training data of the ADN-based liquid propellant sample is used as a priori mark, and the predicted combustion parameter of the single parameter prediction network is determined as the priori mark of the predicted combustion parameter of the combustion parameter prediction network, and the combustion parameter prediction network is adjusted; The step of adjusting the first inference layers respectively corresponding to the component sets in the single parameter prediction network includes: Setting the first reasoning layers respectively corresponding to the component sets in the single parameter prediction network; Determining a reasoning error of the first reasoning layer according to the predicted combustion parameter of the first reasoning layer corresponding to the ath component set in the single parameter prediction network and the actual combustion parameter; The first reasoning layer is adjusted according to the reasoning error.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images