Ground engine tail flame infrared radiation end-to-end prediction method based on neural network

Abstract

The invention relates to the technical field of rocket engines, in particular to a ground engine tail flame infrared radiation end-to-end prediction method based on a neural network, and the method comprises the following steps: 1, calculating a rocket engine tail flame infrared radiation data set; 2, constructing an asymmetric convolution encoder and a parallel deconvolution decoding module; 3, taking the profile spectrum parameter data set as an input data set; processing the tail flame infrared radiation intensity distribution cloud picture data, and taking the processed tail flame infrared radiation intensity distribution cloud picture data and tail flame infrared radiation integral intensity data as an output data set; 4, dividing the input data set and the output data set into a training set and a test set; training and testing the neural network; and 5, inputting the profile spectrum parameters into the neural network, and calculating the infrared radiation data of the tail flame of the rocket engine to be detected. The method solves the problems that an existing rocket engine tail flame infrared radiation calculation method is multiple in calculation links, long in calculation time consumption and difficult to meet mass calculation requirements, and is suitable for rocket engines.

Description

Technical Field

[0001] The present invention relates to the field of rocket engine technology, and in particular to a neural network-based end-to-end prediction method for infrared radiation of a ground engine tail flame. Background Art

[0002] Due to vibrational-rotational transitions, diatomic and polyatomic molecules in the exhaust gas of a rocket engine emit infrared radiation in specific bands, giving the exhaust plume a significant infrared radiation signature, making it a radiation source of particular interest to space-based infrared systems. By calculating the infrared radiation of a rocket engine's exhaust plume, one can understand the combustion state and temperature distribution of the rocket engine, thereby enabling performance optimization and fault diagnosis. However, existing methods for calculating infrared radiation from rocket engine exhaust plumes suffer from the following problems due to their inherent limitations: First, existing methods involve many computational steps and are time-consuming. Second, existing methods are unable to quickly calculate the infrared radiation from a rocket engine's exhaust plume under different spectral parameters, making it difficult to meet the demands of massive computations. Therefore, it is necessary to invent a neural network-based end-to-end prediction method for ground-based engine exhaust plume infrared radiation to address the problems of existing methods for calculating infrared radiation from rocket engine exhaust plumes, which involve many computational steps, are time-consuming, and are difficult to meet the demands of massive computations. Summary of the Invention

[0003] In order to solve the problems of existing rocket engine tail plume infrared radiation calculation methods having many calculation links, long calculation time and difficulty in meeting massive computing needs, the present invention provides an end-to-end prediction method for ground engine tail plume infrared radiation based on neural network.

[0004] The present invention is achieved by adopting the following technical solutions:

[0005] An end-to-end prediction method for ground engine tail plume infrared radiation based on a neural network includes the following steps:

[0006] Step 1: Construct a rocket engine spectrum parameter dataset; input the spectrum parameter dataset into the full-link fine calculation model to calculate the rocket engine tail plume infrared radiation dataset; the tail plume infrared radiation data includes: tail plume infrared radiation intensity distribution cloud map data and tail plume infrared radiation integrated intensity data;

[0007] Step 2: Construct an asymmetric convolutional encoder and a parallel deconvolutional decoding module, which together form an end-to-end prediction neural network;

[0008] Step 3: Use the spectrum parameter dataset as the input dataset of the end-to-end prediction neural network; normalize and interpolate the tail plume infrared radiation intensity distribution cloud map data, and use it together with the tail plume infrared radiation integrated intensity data as the output dataset of the end-to-end prediction neural network;

[0009] Step 4: Divide the input dataset and output dataset into training set and test set; use the training set and test set to train and test the end-to-end prediction neural network;

[0010] Step 5: Collect the spectrum parameters of the rocket engine to be tested; input the spectrum parameters into the trained end-to-end prediction neural network to calculate the infrared radiation data of the tail flame of the rocket engine to be tested.

[0011] Furthermore, the profile parameter data set is constructed using a Latin hypercube sampling method; the profile parameters include: combustion chamber temperature, combustion chamber pressure, nozzle expansion ratio, oxidizer content in propellant, binder content in propellant, and nozzle diameter.

[0012] Furthermore, the end-to-end prediction neural network is defined as follows:

[0013] ;

[0014] Where: The cloud diagram shows the infrared radiation intensity distribution of the tail flame in the 2.7μm band; The cloud diagram shows the infrared radiation intensity distribution of the tail flame in the 4.3μm band; Indicates the integrated intensity of tail plume infrared radiation in the 2.7μm band; Indicates the integrated intensity of tail plume infrared radiation in the 4.3μm band; Indicates the combustion chamber temperature; Indicates the combustion chamber pressure; represents the nozzle expansion ratio; Indicates the oxidizer content in the propellant; Indicates the binder content in the propellant; Indicates the nozzle diameter.

[0015] Furthermore, the steps of constructing the asymmetric convolutional encoder are as follows:

[0016] First, three parallel convolution blocks are constructed; the first convolution block includes a convolution layer with a convolution kernel size of 3×3, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; the second convolution block includes a convolution layer with a convolution kernel size of 1×3, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; the third convolution block includes a convolution layer with a convolution kernel size of 3×1, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; then, a ReLU function, a batch normalization layer, and a maximum pooling layer are connected after the three convolution blocks to complete the construction of the asymmetric convolution encoder;

[0017] When working, three convolution blocks are first used to extract features from the input data. Then, the extracted features are activated and batch normalized using the ReLU function and batch normalization layer. Then, the max pooling layer is used to extract the features after ReLU activation and batch normalization to a deeper level, obtaining more significant data features as the output of the asymmetric convolution encoder.

[0018] The feature extraction process is expressed as follows:

[0019] ;

[0020] Where: Represents the extracted features; Represents a convolution layer with a convolution kernel size of 3×3; Represents a convolution layer with a convolution kernel size of 1×3; Represents a convolution layer with a convolution kernel size of 3×1; Represents the spectrum parameter matrix; Indicates the sum operation at the corresponding position; represents the two-dimensional convolution operator;

[0021] The ReLU activation and batch normalization process are expressed as follows:

[0022] ;

[0023] ;

[0024] Where: Represents the extracted features; Represents the features after ReLU activation and batch normalization; Represents the output matrix of the ReLU function; represents the channel average of batch normalization; represents the standard deviation of batch normalization; represents the scaling factor; Indicates the offset.

[0025] Furthermore, the steps of constructing the parallel deconvolution decoding module are as follows:

[0026] First, two parallel deconvolution decoders are constructed; each deconvolution decoder consists of three serial deconvolution blocks: the first deconvolution block is used to reduce the feature dimension, including a deconvolution layer with a deconvolution kernel size of 1×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; the second deconvolution block is used to reduce the number of channels, including a deconvolution layer with a deconvolution kernel size of 3×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; the third deconvolution block is used to restore the feature dimension, including a deconvolution layer with a deconvolution kernel size of 1×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; then, a fully connected layer is connected after the two deconvolution decoders to complete the construction of the parallel deconvolution decoding module;

[0027] When working, two deconvolution decoders are first used to upsample and restore the features of the output of the asymmetric convolution encoder, and then the fully connected layer is used to transform the restored features and use them as the output of the end-to-end prediction neural network;

[0028] The conversion process is shown as follows:

[0029] ;

[0030] Where: represents the output of the end-to-end prediction neural network; Indicates the characteristics after recovery; represents weight; Indicates bias.

[0031] Furthermore, the normalization and interpolation processing steps of the tail flame infrared radiation intensity distribution cloud map data are as follows:

[0032] First, the original size of the cloud map is divided by the nozzle radius to obtain the normalized cloud map size;

[0033] Then, according to the normalized cloud image size, the cloud image is interpolated using bilinear interpolation to obtain the interpolated cloud image size;

[0034] Finally, the interpolated cloud size is multiplied by the nozzle radius to obtain the restored cloud size.

[0035] Furthermore, in the training process of the end-to-end prediction neural network, the root mean square error is used as the training loss function, and the function expression is:

[0036] ;

[0037] Where: represents the root mean square error of the end-to-end prediction neural network; Represents the tail plume infrared radiation data calculated by the end-to-end prediction neural network; Indicates the tail plume infrared radiation data calculated by the full-link fine calculation model; Indicates the number of grid points; represents a grid point;

[0038] During the training process of the end-to-end prediction neural network, the correction formulas for weights and biases are as follows:

[0039] ;

[0040] ;

[0041] Where: represents weight; Indicates bias; Indicates the number of revisions; represents the learning rate; Represents the root mean square error of the end-to-end prediction neural network.

[0042] Compared with the existing rocket engine tail plume infrared radiation calculation method, the neural network-based ground engine tail plume infrared radiation end-to-end prediction method described in the present invention has the following advantages: First, the present invention realizes the rapid calculation of tail plume infrared radiation by constructing an end-to-end prediction neural network from spectrum parameters to tail plume infrared radiation data, thereby effectively reducing the calculation links and effectively shortening the calculation time. Second, the present invention realizes the flexible extraction of different dimensional and hierarchical features in the input data by constructing an asymmetric convolution encoder on the one hand, and realizes the feature recovery of the output of the asymmetric convolution encoder from different channels on the other hand by constructing a parallel deconvolution decoding module, thereby effectively improving the model prediction accuracy while enhancing the nonlinear expression ability of the network. The constructed model can calculate the infrared radiation of the rocket engine tail plume under different spectrum parameters in real time, thereby meeting the needs of massive calculations.

[0043] The present invention effectively solves the problems of existing rocket engine tail flame infrared radiation calculation methods, such as multiple calculation links, long calculation time, and difficulty in meeting massive calculation needs, and is suitable for rocket engines. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a principle block diagram of the present invention.

[0045] Figure 2 Schematic diagram of the training loss function of the end-to-end prediction neural network in the present invention.

[0046] Figure 3Schematic diagram comparing the calculation results of the present invention and the existing calculation method. DETAILED DESCRIPTION

[0047] An end-to-end prediction method for ground engine tail plume infrared radiation based on a neural network includes the following steps:

[0048] Step 1: Construct a rocket engine spectrum parameter dataset; input the spectrum parameter dataset into the full-link fine calculation model to calculate the rocket engine tail plume infrared radiation dataset; the tail plume infrared radiation data includes: tail plume infrared radiation intensity distribution cloud map data and tail plume infrared radiation integrated intensity data;

[0049] Step 2: Construct an asymmetric convolutional encoder and a parallel deconvolutional decoding module, which together form an end-to-end prediction neural network;

[0050] Step 3: Use the spectrum parameter dataset as the input dataset of the end-to-end prediction neural network; normalize and interpolate the tail plume infrared radiation intensity distribution cloud map data, and use it together with the tail plume infrared radiation integrated intensity data as the output dataset of the end-to-end prediction neural network;

[0051] Step 4: Divide the input dataset and output dataset into training set and test set; use the training set and test set to train and test the end-to-end prediction neural network;

[0052] Step 5: Collect the spectrum parameters of the rocket engine to be tested; input the spectrum parameters into the trained end-to-end prediction neural network to calculate the infrared radiation data of the tail flame of the rocket engine to be tested.

[0053] The profile parameter data set is constructed using the Latin hypercube sampling method; the profile parameters include: combustion chamber temperature, combustion chamber pressure, nozzle expansion ratio, oxidizer content in propellant, binder content in propellant, and nozzle diameter.

[0054] The end-to-end prediction neural network is defined as follows:

[0055] ;

[0056] Where: The cloud diagram shows the infrared radiation intensity distribution of the tail flame in the 2.7μm band; The cloud diagram shows the infrared radiation intensity distribution of the tail flame in the 4.3μm band; Indicates the integrated intensity of tail plume infrared radiation in the 2.7μm band; Indicates the integrated intensity of tail plume infrared radiation in the 4.3μm band; Indicates the combustion chamber temperature; Indicates the combustion chamber pressure; represents the nozzle expansion ratio; Indicates the oxidizer content in the propellant; Indicates the binder content in the propellant; Indicates the nozzle diameter.

[0057] The steps for constructing the asymmetric convolutional encoder are as follows:

[0058] First, three parallel convolution blocks are constructed; the first convolution block includes a convolution layer with a convolution kernel size of 3×3, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; the second convolution block includes a convolution layer with a convolution kernel size of 1×3, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; the third convolution block includes a convolution layer with a convolution kernel size of 3×1, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; then, a ReLU function, a batch normalization layer, and a maximum pooling layer are connected after the three convolution blocks to complete the construction of the asymmetric convolution encoder;

[0059] When working, three convolution blocks are first used to extract features from the input data. Then, the extracted features are activated and batch normalized using the ReLU function and batch normalization layer. Then, the max pooling layer is used to extract the features after ReLU activation and batch normalization to a deeper level, obtaining more significant data features as the output of the asymmetric convolution encoder.

[0060] The feature extraction process is expressed as follows:

[0061] ;

[0062] Where: Represents the extracted features; Represents a convolution layer with a convolution kernel size of 3×3; Represents a convolution layer with a convolution kernel size of 1×3; Represents a convolution layer with a convolution kernel size of 3×1; Represents the spectrum parameter matrix; Indicates the sum operation at the corresponding position; represents the two-dimensional convolution operator;

[0063] The ReLU activation and batch normalization process are expressed as follows:

[0064] ;

[0065] ;

[0066] Where: Represents the extracted features; Represents the features after ReLU activation and batch normalization; Represents the output matrix of the ReLU function; represents the channel average of batch normalization; represents the standard deviation of batch normalization; represents the scaling factor; Indicates the offset.

[0067] The steps for constructing the parallel deconvolution decoding module are as follows:

[0068] First, two parallel deconvolution decoders are constructed; each deconvolution decoder consists of three serial deconvolution blocks: the first deconvolution block is used to reduce the feature dimension, including a deconvolution layer with a deconvolution kernel size of 1×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; the second deconvolution block is used to reduce the number of channels, including a deconvolution layer with a deconvolution kernel size of 3×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; the third deconvolution block is used to restore the feature dimension, including a deconvolution layer with a deconvolution kernel size of 1×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; then, a fully connected layer is connected after the two deconvolution decoders to complete the construction of the parallel deconvolution decoding module;

[0069] When working, two deconvolution decoders are first used to upsample and restore the features of the output of the asymmetric convolution encoder, and then the fully connected layer is used to transform the restored features and use them as the output of the end-to-end prediction neural network;

[0070] The conversion process is shown as follows:

[0071] ;

[0072] Where: represents the output of the end-to-end prediction neural network; Indicates the characteristics after recovery; represents weight; Indicates bias.

[0073] The normalization and interpolation processing steps of the tail flame infrared radiation intensity distribution cloud map data are as follows:

[0074] First, the original size of the cloud map is divided by the nozzle radius to obtain the normalized cloud map size;

[0075] Then, according to the normalized cloud image size, the cloud image is interpolated using bilinear interpolation to obtain the interpolated cloud image size;

[0076] Finally, the interpolated cloud size is multiplied by the nozzle radius to obtain the restored cloud size.

[0077] In the training process of the end-to-end prediction neural network, the root mean square error is used as the training loss function, and the function expression is:

[0078] ;

[0079] Where: represents the root mean square error of the end-to-end prediction neural network; Represents the tail plume infrared radiation data calculated by the end-to-end prediction neural network; Indicates the tail plume infrared radiation data calculated by the full-link fine calculation model; Indicates the number of grid points; represents a grid point;

[0080] During the training process of the end-to-end prediction neural network, the correction formulas for weights and biases are as follows:

[0081] ;

[0082] ;

[0083] Where: represents weight; Indicates bias; Indicates the number of revisions; represents the learning rate; Represents the root mean square error of the end-to-end prediction neural network.

[0084] Figure 2 This is a diagram of the training loss function of the end-to-end prediction neural network in the present invention. Figure 2 It can be seen that the end-to-end prediction neural network converges well.

[0085] Figure 3 This is a schematic diagram comparing the calculation results of the present invention (CNN-IR) and the existing calculation method (NMC). Figure 3 It can be seen that: on the tail plume infrared radiation data set under different under-expansion states, the present invention shows better robustness and adaptability, and can more accurately calculate the tail plume infrared radiation of the rocket engine under different spectral parameters.

[0086] Experiments show that the calculation time of the existing calculation method is about 6 hours, while the calculation time of the present invention is only seconds. The speedup ratio of the present invention compared with the existing calculation method is 2.16×10 4 , which greatly improves the calculation efficiency of infrared radiation from rocket engine tail flame.

[0087] Although specific embodiments of the present invention have been described above, those skilled in the art will appreciate that these are merely illustrative and that the scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, and such changes and modifications are intended to fall within the scope of the present invention.

Claims

1. A neural network-based end-to-end prediction method for infrared radiation of ground engine tail plumes, characterized by: The method comprises the following steps: Step 1: Construct a rocket engine spectrum parameter dataset; input the spectrum parameter dataset into the full-link fine calculation model to calculate the rocket engine tail plume infrared radiation dataset; The tail flame infrared radiation data includes: tail flame infrared radiation intensity distribution cloud map data, tail flame infrared radiation integral intensity data; Step 2: Construct an asymmetric convolutional encoder and a parallel deconvolutional decoding module, which together form an end-to-end prediction neural network; Step 3: Use the spectrum parameter dataset as the input dataset of the end-to-end prediction neural network; normalize and interpolate the tail plume infrared radiation intensity distribution cloud map data, and use it together with the tail plume infrared radiation integrated intensity data as the output dataset of the end-to-end prediction neural network; Step 4: Divide the input dataset and output dataset into training set and test set; use the training set and test set to train and test the end-to-end prediction neural network; Step 5: Collect the spectrum parameters of the rocket engine to be tested; input the spectrum parameters into the trained end-to-end prediction neural network to calculate the infrared radiation data of the tail flame of the rocket engine to be tested.

2. The neural network-based end-to-end prediction method for ground engine tail plume infrared radiation according to claim 1 is characterized by: The profile parameter data set is constructed using the Latin hypercube sampling method; the profile parameters include: combustion chamber temperature, combustion chamber pressure, nozzle expansion ratio, oxidizer content in propellant, binder content in propellant, and nozzle diameter.

3. The neural network-based end-to-end prediction method for ground engine tail plume infrared radiation according to claim 2 is characterized by: The end-to-end prediction neural network is defined as follows: ； Where: The cloud diagram shows the infrared radiation intensity distribution of the tail flame in the 2.7μm band; The cloud diagram shows the infrared radiation intensity distribution of the tail flame in the 4.3μm band; Indicates the integrated intensity of tail plume infrared radiation in the 2.7μm band; Indicates the integrated intensity of tail plume infrared radiation in the 4.3μm band; Indicates the combustion chamber temperature; Indicates the combustion chamber pressure; represents the nozzle expansion ratio; Indicates the oxidizer content in the propellant; Indicates the binder content in the propellant; Indicates the nozzle diameter.

4. The neural network-based end-to-end prediction method for ground engine tail plume infrared radiation according to claim 3 is characterized by: The steps for constructing the asymmetric convolutional encoder are as follows: First, three parallel convolution blocks are constructed; the first convolution block includes a convolution layer with a convolution kernel size of 3×3, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; the second convolution block includes a convolution layer with a convolution kernel size of 1×3, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; the third convolution block includes a convolution layer with a convolution kernel size of 3×1, a ReLU function connected after the convolution layer, and a batch normalization layer connected after the ReLU function; then, a ReLU function, a batch normalization layer, and a maximum pooling layer are connected after the three convolution blocks to complete the construction of the asymmetric convolution encoder; When working, three convolution blocks are first used to extract features from the input data. Then, the extracted features are activated and batch normalized using the ReLU function and batch normalization layer. Then, the max pooling layer is used to extract the features after ReLU activation and batch normalization to a deeper level, obtaining more significant data features as the output of the asymmetric convolution encoder. The feature extraction process is expressed as follows: ； Where: Represents the extracted features; Represents a convolution layer with a convolution kernel size of 3×3; Represents a convolution layer with a convolution kernel size of 1×3; Represents a convolution layer with a convolution kernel size of 3×1; Represents the spectrum parameter matrix; Indicates the sum operation at the corresponding position; represents the two-dimensional convolution operator; The ReLU activation and batch normalization process are expressed as follows: ； ； Where: Represents the extracted features; Represents the features after ReLU activation and batch normalization; Represents the output matrix of the ReLU function; represents the channel average of batch normalization; represents the standard deviation of batch normalization; represents the scaling factor; Indicates the offset.

5. The neural network-based end-to-end prediction method for ground engine tail plume infrared radiation according to claim 4 is characterized by: The steps for constructing the parallel deconvolution decoding module are as follows: First, two parallel deconvolution decoders are constructed; each deconvolution decoder consists of three serial deconvolution blocks: the first deconvolution block is used to reduce the feature dimension, including a deconvolution layer with a deconvolution kernel size of 1×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; the second deconvolution block is used to reduce the number of channels, including a deconvolution layer with a deconvolution kernel size of 3×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; the third deconvolution block is used to restore the feature dimension, including a deconvolution layer with a deconvolution kernel size of 1×1, a ReLU function connected after the deconvolution layer, and a batch normalization layer connected after the ReLU function; then, a fully connected layer is connected after the two deconvolution decoders to complete the construction of the parallel deconvolution decoding module; When working, two deconvolution decoders are first used to upsample and restore the features of the output of the asymmetric convolution encoder, and then the fully connected layer is used to transform the restored features and use them as the output of the end-to-end prediction neural network; The conversion process is shown as follows: ； Where: represents the output of the end-to-end prediction neural network; Indicates the characteristics after recovery; represents weight; Indicates bias.

6. The neural network-based end-to-end prediction method for ground engine tail plume infrared radiation according to claim 5 is characterized by: The normalization and interpolation processing steps of the tail flame infrared radiation intensity distribution cloud map data are as follows: First, the original size of the cloud map is divided by the nozzle radius to obtain the normalized cloud map size; Then, according to the normalized cloud image size, the cloud image is interpolated using bilinear interpolation to obtain the interpolated cloud image size; Finally, the interpolated cloud size is multiplied by the nozzle radius to obtain the restored cloud size.

7. The neural network-based end-to-end prediction method for ground engine tail plume infrared radiation according to claim 6 is characterized by: In the training process of the end-to-end prediction neural network, the root mean square error is used as the training loss function, and the function expression is: ； Where: represents the root mean square error of the end-to-end prediction neural network; Represents the tail plume infrared radiation data calculated by the end-to-end prediction neural network; Indicates the tail plume infrared radiation data calculated by the full-link fine calculation model; Indicates the number of grid points; represents a grid point; During the training process of the end-to-end prediction neural network, the correction formulas for weights and biases are as follows: ； ； Where: represents weight; Indicates bias; Indicates the number of revisions; represents the learning rate; Represents the root mean square error of the end-to-end prediction neural network.

Images