Intelligent prediction method of scramjet flame field based on deep learning

Abstract

The present invention discloses a method for intelligent prediction of a scramjet flame field based on deep learning, which belongs to the technical field of cross-integration of supersonic flow combustion and artificial intelligence. The method utilizes experimental data of a hydrogen-fueled scramjet engine, takes the wall temperature and its time series as input variables, and constructs a supersonic combustion chamber flame field prediction model. By introducing a deformable convolution kernel, the supersonic combustion chamber flame field prediction model is significantly enhanced in its ability to capture irregular morphological changes of the flame field while reducing the amount of model calculation. The residual network is used to effectively solve the gradient vanishing problem, further improving the prediction accuracy and reasoning speed of the combustion flow field, realizing real-time cross-time prediction, and enhancing the network's modeling capability for complex spatial transformations.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-integration of supersonic flow combustion and artificial intelligence, and specifically relates to an intelligent prediction method for a scramjet engine flame field based on deep learning. Background Art

[0002] As one of the ideal power devices for hypersonic propulsion systems in the fields of military protection and aerospace, scramjet engines have become a global research hotspot. As one of the important components of the engine, the combustion chamber is the core part for fuel mixing, ignition, combustion, etc., and contains extremely complex physical and chemical reactions. The shock wave boundary layer interference and multi-scale turbulence coupled combustion effects inside the combustion chamber pose great challenges to the optimization of combustion organization. At present, the technology of supersonic combustion adopts a control strategy based on predetermined parameters, but these strategies cannot effectively adapt to the adaptive adjustment in flight requirements under high altitude, multi-speed and different Mach number conditions. Therefore, accurately monitoring and predicting relevant information in fluid dynamics and combustion processes is crucial to evaluating the energy release efficiency and improving the performance of the combustion chamber in complex environments.

[0003] Convolutional neural networks have powerful feature fitting performance and are widely used in the field of fluid mechanics. Existing data-driven neural network intelligent prediction models have been proposed in large numbers and have made significant progress. However, their performance is still limited due to the lack of support from large, high-quality data sets, especially for complex combustion flow fields with shock wave boundary layer interference and turbulent pulsation. Summary of the invention

[0004] The purpose of the present invention is to address the above-mentioned deficiencies in the prior art and to provide a method for intelligent prediction of a scramjet flame field based on deep learning, so as to solve the problem that the existing neural network intelligent prediction model is limited in performance due to the lack of support from large, high-quality data sets.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for intelligently predicting a scramjet flame field based on deep learning comprises the following steps:

[0007] S1. Obtain the flame image and wall temperature inside the scramjet combustion chamber based on ground wind tunnel experiments;

[0008] S2, standardizing the flame image and the wall temperature, and performing synchronous alignment in time and space to obtain a data set containing different prediction time spans;

[0009] S3, randomize the order of data in the dataset and divide the dataset into training set and test set according to the ratio;

[0010] S4, constructing a supersonic combustion chamber flame field prediction model including a generator and a discriminator;

[0011] S5. Using the training set to train the supersonic combustion chamber flame field prediction model, and outputting the training weights;

[0012] S6, loading the training weights in step S5, and using the test set to test the supersonic combustor flame field prediction model, evaluating the performance of the supersonic combustor flame field prediction model and outputting the flame field test results;

[0013] S7, combining the class activation map and the screen display to interpretably display the calculation process of the supersonic combustion chamber flame field prediction model;

[0014] S8. Embed the supersonic combustion chamber flame field prediction model and the training weights obtained in step S5 into the FPGA platform. The FPGA platform outputs the flame characteristics represented by relay codes, and transmits the relay codes to the fuel and oxidizer controller of the scramjet engine.

[0015] Further, step S1 includes:

[0016] A high-speed camera is used to capture flame images inside the scramjet combustion chamber;

[0017] Different numbers of temperature sensor probes are arranged on the wall of the scramjet combustion chamber according to the change of flame gradient to collect the internal wall temperature of the scramjet combustion chamber.

[0018] Further, step S2 includes:

[0019] The flame image is cropped to obtain an image area of ​​140×560 pixels containing the flame structure;

[0020] The grayscale value of the cropped flame image is normalized, the flame image range is scaled to 0-1, and the flame image is downsampled by 4 times to obtain a flame image with a size of 35×140 pixels;

[0021] The processed flame image and wall temperature are synchronously aligned in time and space, and the time span is used as the key parameter to construct a dataset containing different prediction time spans.

[0022] Further, step S3 includes:

[0023] Randomize the order of data in the dataset, use 80% of the data in the dataset as the training set, and use 20% of the data in the dataset as the test set.

[0024] Furthermore, the generator in step S4 includes a fully connected layer, a transposed convolution, a deformable convolution layer and a residual module connected in series; the discriminator includes a plurality of fully connected layers connected in series;

[0025] The input of the generator is the wall temperature and the time span, and the output of the generator, the flame image, and the label corresponding to the flame image are input into the discriminator, and the flame images with different time spans are output.

[0026] Furthermore, the nonlinear feature mapping relationship between the input wall temperature, time span and the corresponding flame image is established by combining transposed convolution and deformable convolution:

[0027]

[0028] in, is the flame image of the flow field in the combustion chamber generated after pre-training; Span is the time span between the previous and next moments, and is used as the input of the network to participate in the model training; is the wall temperature; f is the function of the entire network; θ is all the learnable parameters.

[0029] Furthermore, the deformable convolutional layer dynamically adjusts the shape and size of the flame image receptive field, which includes:

[0030]

[0031] Among them, y(P0) is the value of the output feature map at position P0; P n is each position of the convolution kernel; w is the convolution kernel; P0 is a specific output position coordinate; ΔP n corresponds to the convolution kernel position P n offset; R is the neighborhood area or range covered by the convolution kernel in the convolution operation, x is the input feature map, and x() is the specific position on the input feature map.

[0032] Furthermore, the training process of the supersonic combustion chamber flame field prediction model is converted into a parameter optimization process, specifically:

[0033]

[0034] in:

[0035]

[0036]

[0037] In the formula, L is the loss function, which is the objective function that the generator needs to minimize; It reflects a two-step process: one is to try to maximize the performance of the discriminator so that it can distinguish between true and false as accurately as possible, and the other is to try to minimize the generator loss and improve its ability to deceive the discriminator; λ is the weight coefficient of the regularization term; N is the total number of elements in the tensor; G(x) is the output of the generator; y is the conditional variable; is the expectation when extracting sample x from the true data distribution P0(x); D(x|y) is the output of the discriminator; D(G(z|y)) is the evaluation of the generator output sample by the discriminator; is the noise distribution P z (z) when extracting random noise z, P o (x) is the probability distribution of wall temperature and time span.

[0038] Furthermore, the loss function of the generator is:

[0039]

[0040] The loss function of the discriminator is:

[0041]

[0042] Among them, Loss G for; Loss D For P data (x,y) is the data distribution of real flame images and their corresponding labels.

[0043] Furthermore, in step S8, the flame characteristics are represented by relay codes, including:

[0044] Relay 1: 0001 → o

[0045] Relay 2: 0010 → s

[0046] Relay 3: 0100 → t

[0047] Relay 4: 1000 → r

[0048] Based on the switching behavior of the relay, a binary sequence is selected to represent the data output of the flame characteristics:

[0049] 0001 is used to output flame profile data o;

[0050] 0010 is used to output flame area data s;

[0051] 0100 is used to output flame temperature data t;

[0052] 1000 is used to output the serial port transmission of density flame brightness data.

[0053] The deep learning-based intelligent prediction method for the flame field of a scramjet engine provided by the present invention has the following beneficial effects:

[0054] 1. The present invention uses the experimental data of hydrogen-fueled scramjet engines, takes the wall temperature and its time series as input variables, and constructs a lightweight deformable convolution residual neural network model, namely, a supersonic combustion chamber flame field prediction model. By introducing a deformable convolution kernel, the supersonic combustion chamber flame field prediction model significantly enhances the ability to capture irregular morphological changes of the flame field while reducing the amount of model calculation. The residual network is used to effectively solve the gradient vanishing problem, further improving the prediction accuracy and reasoning speed of the combustion flow field, and realizing real-time cross-time prediction.

[0055] 2. In the prediction of the dynamic changes of the flame structure in the combustion chamber of a scramjet engine, the present invention uses the wall temperature measured by the temperature sensors at the top and bottom of the combustion chamber and its changes on different time scales as input, adopts an architecture that integrates a deformable convolutional neural network and a residual network as a generator, and constructs a nonlinear feature mapping path from input data to output flame images to enhance the network's modeling ability for complex spatial transformations.

[0056] 3. The deep learning-based supersonic combustion chamber flame field prediction model of the present invention is used to predict the flame field in the scramjet engine, and the entire supersonic combustion chamber flame field prediction model is integrated into the FPGA platform, realizing the automation of the main flow control, thereby providing technical guarantee for the stable operation of the engine during flight. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the method for intelligently predicting the flame field of a scramjet engine based on deep learning of the present invention.

[0058] Figure 2 Schematic diagram of the framework structure of the neural network model for predicting the flame field of the supersonic combustion chamber of the present invention.

[0059] Figure 3 This is a schematic diagram of the intelligent prediction system for the flame field of a scramjet engine based on deep learning of the present invention. DETAILED DESCRIPTION

[0060] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0061] Example 1

[0062] This embodiment discloses a method for intelligent prediction of the flame field of a scramjet engine based on deep learning, which enhances the adaptability to the complex flow field structure inside the scramjet engine, improves the prediction accuracy and reasoning speed of the combustion flow field, and realizes real-time cross-time prediction. Figure 1 and Figure 3 , which specifically includes the following:

[0063] Step S1, obtaining a flame image and a wall temperature inside a scramjet combustion chamber according to a ground wind tunnel experiment, which specifically includes:

[0064] The present embodiment preferably adopts a scramjet engine model, which includes a 340mm isolation section and an 898mm combustion chamber, with a cavity depth of 28mm, a length of 278mm, and a rear slope angle of 43.6°. There are two expansion sections at the tail of the cavity, with lengths of 230mm and 390.5mm, respectively, and angles of 1.0° and 3.0°, respectively. Hydrogen fuel passes through 15 1mm nozzles at an equivalence ratio of 0.31, located 69mm upstream of the cavity step, and enters the combustion chamber. The flow rate in the combustion chamber is 2.89kg / s, the total pressure is 1.75MPa, the total temperature is 1350K, and the airflow is accelerated to 2.5Ma, which contains 20.09% oxygen, 19.26% water vapor and 60.65% nitrogen.

[0065] As a preferred embodiment of this embodiment, the collection of flame image and wall temperature includes:

[0066] The experimental data were collected using a touch-type wall temperature sensor and a high-speed camera. There were 32 temperature sensor probes distributed on the upper and lower walls of the combustion chamber. The distribution of the temperature sensor probes in this embodiment was based on the distribution characteristics of the flame to implement a local dense uneven acquisition strategy in the high-frequency feature area, that is, dense sensor distribution acquisition was performed at locations where the flame gradient changed greatly, and a sparse sensor acquisition strategy was implemented in areas where the feature changes were small.

[0067] The maximum range of the temperature sensor can reach 1500℃, with a sensitivity of ±1℃, to meet the monitoring needs of the high temperature environment in the combustion chamber of a scramjet engine.

[0068] The image resolution of the present embodiment was set to 1024×264 pixels, the exposure time was set to 4.62 microseconds, and the images were recorded at a rate of 10,000 frames per second. The flame images were captured using a high-speed camera, and six ground tests were performed under different hydrogen fuel equivalence ratios. The experiment changed the equivalence ratio by adjusting the injection pressure of the hydrogen fuel, and the specific pressure settings included 2MPa, 2.5MPa, 3MPa, 3.5MPa, 4MPa, and 4.5MPa, thereby collecting a total of 13,681 flame images.

[0069] Step S2, standardizing the flame image and the wall temperature, and performing synchronous alignment processing in time and space, to obtain a data set containing different prediction time spans, which specifically includes:

[0070] The flame image is cropped to obtain an image area of ​​140×560 pixels containing the flame structure;

[0071] In view of the high resolution of flame images captured by high-speed cameras, this embodiment directly uses these images for training, which will lead to large network parameters, increase the complexity and resource load of model training, and at the same time, by analyzing the distribution characteristics of flame images, a 140×560 pixel image area containing the flame structure is cropped from each 232×896 pixel original image to remove useless background information and accelerate the convergence speed of network training;

[0072] Subsequently, the grayscale value of the cropped image is normalized, its range is scaled to 0 to 1, and the flame image is downsampled by 4 times, finally obtaining a flame image with a size of 35×140 pixels.

[0073] During the training process, the time span is used as a key parameter of the network input, and a prediction time span of 1 to 50 milliseconds is constructed for model training. In order to comprehensively and objectively evaluate the generalization ability of the model, this embodiment uses different time spans of 55 milliseconds, 90 milliseconds, and 115 milliseconds for prediction, providing a comprehensive evaluation of the model prediction performance.

[0074] The processed flame image and wall temperature are synchronously aligned in time and space, and the time span is used as the key parameter to construct a dataset containing different prediction time spans.

[0075] Step S3: randomize the order of data in the data set and divide the data set into a training set and a test set according to the ratio, which specifically includes the following contents:

[0076] In order to effectively predict the dynamic changes of the flame structure in a supersonic combustion chamber at different time scales, especially the evolution within a longer time span, this embodiment selects 80% of the data in the data set as a training set based on the data set containing different prediction time spans for the construction of a supersonic combustion chamber flame field prediction model; 20% of the data is used to verify its performance.

[0077] In this embodiment, 2961 flame images obtained from the sixth vehicle test under the condition of an injection pressure of 4.5 MPa are selected as an independent test set, which are not included in the model training process, so as to comprehensively evaluate the performance of the trained model.

[0078] As a preferred embodiment of the present invention, in order to ensure that the developed model has reliable prediction capability under a wide range of Mach number flow conditions, data with Mach numbers between 4 and 6 are used as a training set, and data with a flow condition of Mach number 7 is set separately as a test set. The data range of the present embodiment can ensure that the model is fully trained under medium to high speed flow conditions, thereby capturing the key dynamic behaviors of flow characteristics at different speed levels, and better evaluating the extrapolation performance of the model, that is, the model's prediction capability for higher speed flow conditions.

[0079] Step S4: construct a supersonic combustion chamber flame field prediction model including a generator and a discriminator, referring to Figure 2 , which specifically includes the following:

[0080] As a preferred embodiment of this embodiment, the supersonic combustion chamber flame field prediction model includes a generator part and a discriminator part.

[0081] Specifically, the generator includes a fully connected layer, a transposed convolution, a deformable convolution layer and a residual module connected in series; the discriminator includes a plurality of fully connected layers connected in series;

[0082] In the generator, the wall temperature and time span are used as input, and a fully connected layer is used to increase the dimension of the data. Then, the vector data after dimensionality increase is converted into a matrix form through dimensional transformation. In order to speed up the learning speed of the model and improve its nonlinear mapping ability, a layer of normalization processing and ReLU activation function are added after each convolution operation. In order to better adapt to the irregular shapes and changes in the image, the introduction of deformable convolution allows the network to dynamically adjust the shape and size of the receptive field. The introduction of the residual module is used to solve the gradient vanishing problem that may occur during the training of deep neural networks, speed up the training speed, and better learn the complex flame field characteristics through residual connections.

[0083] The discriminator consists of three fully connected layers connected in series. The output of the generator, the flame image, and the label corresponding to the flame image are input into the discriminator, and the LeakyRelu activation function and the Sigmoid activation function are used to output the flame images at different time scales. Among them, the LeakyReLU function helps the discriminator to better learn to distinguish between real flame images and flame images generated by the generator, and the output of the Sigmoid function can be used by the discriminator to determine whether an image represents a normal burning flame.

[0084] In view of the fact that the adversarial loss in the existing generative adversarial network can only achieve the process of matching the generated data with the original data distribution, this embodiment constructs a loss function that combines the adversarial loss with the L1 loss to obtain a more accurate result. That is, the loss function of the model in this embodiment is:

[0085]

[0086]

[0087] In the formula, L is the loss function, which is the objective function that the generator needs to minimize; reflects a two-step process: one is to try to maximize the performance of the discriminator so that it can distinguish between true and false as accurately as possible, and the other is to try to minimize the generator loss and improve its ability to deceive the discriminator; λ is the weight coefficient of the regularization term; λ is the weight coefficient of the regularization term; N is the total number of elements in the tensor; G(x) is the output of the generator; y is the conditional variable; is the expectation when extracting sample x from the true data distribution P0(x); D(x|y) is the output of the discriminator; D(G(z|y)) is the evaluation of the generator output sample by the discriminator; is the noise distribution P z (z) is the expectation when random noise z is extracted from .

[0088] The data generation process can be controlled in the desired direction by adding conditional information such as class labels.

[0089] For the generator part, a convolutional neural network is used to establish a nonlinear feature mapping relationship between the input wall temperature, time span and the corresponding flame image:

[0090]

[0091] in, is the flame image of the flow field in the combustion chamber generated after pre-training; Span is the time span between the previous and next moments, and is used as the input of the network to participate in the model training; is the wall temperature; f is the function of the entire network; θ is all the learnable parameters.

[0092] The deformable convolutional layer dynamically adjusts the shape and size of the flame image receptive field, which includes:

[0093]

[0094] Among them, y(P0) is the value of the output feature map at position P0; P n is each position of the convolution kernel; w is the convolution kernel; P0 is a specific output position coordinate; ΔP n corresponds to the convolution kernel position P n offset; R is the neighborhood area or range covered by the convolution kernel in the convolution operation, x is the input feature map, and x() is the specific position on the input feature map.

[0095] In the deformable convolution operation, a two-dimensional offset ΔP is added to the original offset R. n (offset on the x and y axes), this value is obtained by calculating the bilinear difference of the surrounding four real values; specifically, the bilinear difference of this embodiment is calculated by the following formula:

[0096]

[0097] G(q,p)=g(q x ,p x )·g(q y ,p y )

[0098] Among them, g(a,b)=max(0,1-|ab|), x(p) is the value obtained by interpolation at the target point p, G(q,p) is the weight function from the source position q to the target position p, which is usually related to the relative distance between the two points, q is a known point around, x(q) is the value at the known point q, and g(q x ,p x ) is the linear weight function in the x direction, g(q y ,p y ) is the linear weight function in the y direction;

[0099] The neural network of this embodiment continuously adjusts the weight of the model through the optimal gradient descent method until the prediction error reaches a given expected value; because the time span span is used as the input feature of the network, in order to compare the performance gaps of various models under multiple spans and large spans, 3 groups of spans = 15ms, 30ms, and 40ms are selected as test sets in the data set participating in the training, and another 3 groups of spans = 55ms, 90ms, and 115ms are selected as time spans that have never been trained before the model, and the training process of the supersonic combustion chamber flame field prediction model is converted into a parameter optimization process, specifically:

[0100]

[0101] The loss function of the generator is:

[0102]

[0103] Among them, P o (x) is the data distribution input to the generator, i.e., the probability distribution of wall temperature and time span; P data (x, y) is the data distribution of real flame images and their corresponding labels, as well as the result (flame image) generated by the generator based on the input, and N is the total number of elements in the tensor; the first term in the generator's loss function is the cross entropy between the generated data distribution and the real data distribution, and the second term is the scaled L1 loss on the entire tensor.

[0104] The loss function of the discriminator is:

[0105]

[0106] Among them, Loss G is the loss function of the generator; is the joint distribution P data Expected value of (x,y); Loss D is the loss function of the discriminator.

[0107] At the same time, the minimization enables the discriminator to distinguish real data from generated aerodynamic data.

[0108] Step S5, using the training set to train the supersonic combustion chamber flame field prediction model and outputting the training weights, specifically includes the following sub-steps:

[0109] (1) All models in this experiment were created and trained on a computer with Windows 11 operating system, Intel(R) Core(TM) i5-1035G1 CPU@1.00GHz 1.19GHz processor, and 16GB memory capacity.

[0110] (2) The experiment uses the adaptive moment estimation Adam optimizer for optimization, setting the batch size to 512 and the initial learning rate to 0.001;

[0111] (3) Read data and input into the model to start training;

[0112] (4) Save the trained weights.

[0113] Step S6, loading the training weights in step S5, and using the test set to test the supersonic combustor flame field prediction model, evaluating the performance of the supersonic combustor flame field prediction model and outputting the flame field test results, which are specifically:

[0114] The trained weights are loaded and run on the supersonic combustion chamber flame field prediction model with the data in the test set to obtain the test results of the flame field. Then, the prediction ability and applicability of the model are evaluated through indicators such as accuracy and error rate between the model prediction results and the actual data of the test set.

[0115] Step S7, combining the class activation diagram and the screen display to interpretably display the calculation process of the supersonic combustion chamber flame field prediction model, which specifically includes the following contents:

[0116] In order to enhance the interpretability of the convolutional neural network model, this embodiment uses the class activation map (CAM) to visualize the feature layer of the convolutional neural network, thereby revealing the internal operations of the model; by comparing the visualized feature map with the physical properties of the flow field, it not only verifies the consistency of the features extracted by the convolutional neural network, but also improves the transparency and reliability of the model prediction.

[0117] Step S8, embedding the supersonic combustion chamber flame field prediction model and the training weights obtained in step S5 into the FPGA platform, the FPGA platform outputs the flame characteristics represented by the relay code, and transmits the relay code to the fuel and oxidizer controller of the scramjet engine, which specifically includes the following contents:

[0118] In this embodiment, the training weights and the supersonic combustion chamber flame field prediction model obtained through the training in step S5 are embedded into the FPGA platform, so as to intuitively display the relevant characteristics of flame stability through the LED module, and perform an explanatory analysis of its operation and flame characteristics under the guidance of experts.

[0119] Among them, flame stability data includes geometric characteristic data such as flame profile and area;

[0120] As a preferred embodiment of this embodiment, after the relay coding, the relay coding is synchronously output through multiple serial ports, that is, the relay coding is transmitted to the fuel and oxidant controller of the scramjet engine. The relay coding method is:

[0121] Relay 1: 0001 → o

[0122] Relay 2: 0010 → s

[0123] Relay 3: 0100 → t

[0124] Relay 4: 1000 → r

[0125] in;

[0126] 0001 is used to output flame profile data o;

[0127] 0010 is used to output flame area data s;

[0128] 0100 is used to output flame temperature data t;

[0129] 1000 is used to output the serial port transmission of density flame brightness data.

[0130] This embodiment utilizes the switching behavior of the relay encoding module, and a specific binary sequence represents a specific data output. The above data is encoded and synchronously transmitted to the fuel and oxidizer controller of the scramjet engine to adjust and optimize the engine operating state in real time.

[0131] Although the specific implementation of the invention is described in detail in conjunction with the drawings, it should not be understood as limiting the scope of protection of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still fall within the scope of protection of this patent.

Claims

1. A method for intelligent prediction of flame field of scramjet engine based on deep learning, characterized in that: The following steps are involved: S1. Obtain the flame image and wall temperature inside the scramjet combustion chamber based on ground wind tunnel experiments; S2, standardizing the flame image and the wall temperature, and performing synchronous alignment in time and space to obtain a data set containing different prediction time spans; S3, randomize the order of data in the dataset and divide the dataset into training set and test set according to the ratio; S4, constructing a supersonic combustion chamber flame field prediction model including a generator and a discriminator; The generator in S4 includes a fully connected layer, a transposed convolution, a deformable convolution layer and a residual module connected in series; the discriminator includes a plurality of fully connected layers connected in series; The input of the generator is the wall temperature and the time span, and the output of the generator, the flame image, and the label corresponding to the flame image are input into the discriminator, and the flame images of different time spans are output; Combining transposed convolution and deformable convolution to establish a nonlinear feature mapping relationship between the input wall temperature, time span and the corresponding flame image: in, is the flame image of the flow field in the combustion chamber generated after pre-training; Span is the time span between the previous and next moments, and is used as the input of the network to participate in the model training; is the wall temperature; f is the function of the entire network; θ is all the learnable parameters; The deformable convolution layer dynamically adjusts the shape and size of the flame image receptive field, which includes: Among them, y(P0) is the value of the output feature map at position P0; P n is each position of the convolution kernel; w is the convolution kernel; P0 is a specific output position coordinate; ΔP n corresponds to the convolution kernel position P n The offset of the convolution kernel; R is the neighborhood area or range covered by the convolution kernel in the convolution operation, x is the input feature map, and x() is the specific position on the input feature map; S5. Using the training set to train the supersonic combustion chamber flame field prediction model, and outputting the training weights; S6, loading the training weights in step S5, and using the test set to test the supersonic combustor flame field prediction model, evaluating the performance of the supersonic combustor flame field prediction model and outputting the flame field test results; S7, combining the class activation map and the screen display to interpretably display the calculation process of the supersonic combustion chamber flame field prediction model; S8. Embed the supersonic combustion chamber flame field prediction model and the training weights obtained in step S5 into the FPGA platform. The FPGA platform outputs the flame characteristics represented by relay codes, and transmits the relay codes to the fuel and oxidizer controller of the scramjet engine.

2. The method for intelligent prediction of flame field of scramjet engine based on deep learning according to claim 1 is characterized in that: The step S1 comprises: A high-speed camera is used to capture flame images inside the scramjet combustion chamber; Different numbers of temperature sensor probes are arranged on the wall of the scramjet combustion chamber according to the change of flame gradient to collect the internal wall temperature of the scramjet combustion chamber.

3. The method for intelligent prediction of flame field of scramjet engine based on deep learning according to claim 1, characterized in that: The step S2 comprises: The flame image is cropped to obtain an image area of ​​140×560 pixels containing the flame structure; The grayscale value of the cropped flame image is normalized, the flame image range is scaled to 0-1, and the flame image is downsampled by 4 times to obtain a flame image with a size of 35×140 pixels; The processed flame image and wall temperature are synchronously aligned in time and space, and the time span is used as the key parameter to construct a dataset containing different prediction time spans.

4. The method for intelligent prediction of flame field of scramjet engine based on deep learning according to claim 1, characterized in that: The step S3 comprises: Randomize the order of data in the dataset, use 80% of the data in the dataset as the training set, and use 20% of the data in the dataset as the test set.

5. The method for intelligent prediction of flame field of scramjet engine based on deep learning according to claim 1, characterized in that: The training process of the supersonic combustion chamber flame field prediction model is converted into a parameter optimization process, specifically: in: In the formula, L is the loss function, which is the objective function that the generator needs to minimize; It reflects a two-step process: one is to try to maximize the performance of the discriminator so that it can distinguish between true and false as accurately as possible, and the other is to try to minimize the generator loss and improve its ability to deceive the discriminator; λ is the weight coefficient of the regularization term; N is the total number of elements in the tensor; G(x) is the output of the generator; y is the conditional variable; is the expectation when extracting sample x from the true data distribution P0(x); D(x|y) is the output of the discriminator; D(G(z|y)) is the evaluation of the generator output sample by the discriminator; is the noise distribution P z The expectation when random noise z is extracted from (z), P o (x) is the probability distribution of wall temperature and time span.

6. The method for intelligent prediction of flame field of scramjet engine based on deep learning according to claim 5 is characterized in that: The loss function of the generator is: The loss function of the discriminator is: Among them, Loss G for; Loss D For P data (x,y) is the data distribution of real flame images and their corresponding labels.

7. The method for intelligent prediction of flame field of scramjet engine based on deep learning according to claim 1, characterized in that: The step S8 uses relay codes to represent flame characteristics, including: Relay 1: 0001 → o Relay 2: 0010 → s Relay 3: 0100 → t Relay 4: 1000 → r Based on the switching behavior of the relay, a binary sequence is selected to represent the data output of the flame characteristics: 0001 is used to output flame profile data o; 0010 is used to output flame area data s; 0100 is used to output flame temperature data t; 1000 is used to output the serial port transmission of density flame brightness data.

Images