Method for reconstructing two-dimensional velocity field of engine isolation section
By reconstructing the convolutional neural network model through multi-path segmentation input, the problem of insufficient acquisition of two-dimensional velocity field information in the isolation section of the rocket-based cycle combined ramjet engine was solved, and efficient and accurate velocity field reconstruction was achieved, which reduced the measurement cost and improved the robustness of the model.
Patent Information
- Application Number
- CN202510892063.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-10-17
AI Technical Summary
Existing technologies make it difficult to obtain the two-dimensional velocity field information of the isolation section in a rocket-based cycle combined ramjet engine in real time, efficiently and accurately. Traditional methods have complex equipment, high computational complexity, low resolution and high cost.
A multi-path segmentation input reconstruction convolutional neural network model is used to construct a two-dimensional velocity field from one-dimensional pressure data, including data preprocessing, block feature extraction, feature splicing and output parts. Deep learning is used for training to achieve efficient and accurate reconstruction from pressure data to velocity field.
It achieves precise monitoring of the position and shape of the shock wave train in the engine isolation section, reduces measurement costs, improves the robustness and reconstruction accuracy of the model, and can stably reconstruct the two-dimensional velocity field when there are errors or interferences in some input pressure data.
Smart Images
Figure CN120805766A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a method for reconstructing the two-dimensional velocity field of the isolation section of a rocket ramjet combined engine, in particular to a model construction method for constructing a two-dimensional velocity field based on deep learning through pressure data in an engine expansion type isolation section. BACKGROUND
[0002] Rocket-based combined cycle ramjet engines (RBCC) are prone to abnormal critical states of inlet spillage under high equivalence ratio conditions, which mainly manifest as the forward movement and morphological changes of the shock train / separation zone leading edge position in the isolation section, making the measurement of the engine isolation section velocity field extremely important.
[0003] Currently, the models for analyzing the engine isolation section velocity field are mostly focused on the flow field characteristics in the straight pipe with typical cold flow blockage back pressure simulation, but the actual isolation section generally has micro-expansion structures and the shock train / turbulent boundary layer separation is caused by high back pressure generated by turbulent combustion, which is different from the blockage method, so this flow field characteristic is difficult to directly apply to the acquisition of two-dimensional velocity field information in the isolation section for engine control. The main means for real-time detection of engine isolation section two-dimensional velocity field information is the schlieren method, which is complex in equipment. Due to the limitations of volume and weight, it is difficult to be directly applied to the engine, and the numerical method for velocity field analysis has the problems of slow solving speed, high calculation complexity, and difficulty in capturing nonlinear factors. Therefore, the present application takes the engine isolation section as the object and proposes a method for reconstructing a one-dimensional pressure field to a two-dimensional velocity field reconstruction model based on deep learning, which realizes real-time, efficient and high-precision reconstruction of the velocity field in the engine isolation section, monitors the shock train position and shock train shape in the actual working process of the engine isolation section, and identifies the spillage condition of the engine inlet, thereby providing information for real-time control of the engine. SUMMARY
[0004] The present application provides a method for reconstructing the two-dimensional velocity field of the engine isolation section to solve the problem of insufficient acquisition of two-dimensional velocity field information in the isolation section in the field of fluid mechanics.
[0005] The technical scheme adopted by the present application is: a method for reconstructing the two-dimensional velocity field of the engine isolation section, comprising the following steps:
[0006] Step 1, set different Mach numbers and fuel injection equivalence ratios as test conditions and numerical simulation boundary conditions, collect one-dimensional pressure data of the engine isolation section through sensors, obtain two-dimensional velocity field sections in the isolation section through numerical simulation, and construct a one-dimensional pressure-two-dimensional velocity field data set;
[0007] Step 2, training a deep learning-based two-dimensional velocity field reconstruction model using the one-dimensional pressure-two-dimensional velocity field data set to obtain a trained two-dimensional velocity field reconstruction model;
[0008] The two-dimensional velocity field reconstruction model is a multi-path segmentation input reconstruction convolutional neural network model, and the structure thereof comprises:
[0009] The input data preprocessing part: the input one-dimensional pressure data is dimensioned into a two-dimensional velocity matrix by an interpolation method;
[0010] The block feature extraction part: the two-dimensional velocity matrix is divided into four parts, and the convolution-deconvolution method is used for velocity distribution feature extraction;
[0011] The feature splicing part: the distribution features of the two-dimensional velocity matrix are extracted by two splicing operations;
[0012] The output part: a full connection layer is used to output the reconstructed two-dimensional velocity field image;
[0013] Step 3, based on the two-dimensional velocity field reconstruction model in step 2, the pressure data in the engine running process is reconstructed to obtain two-dimensional velocity field information.
[0014] Further, the reconstruction process of the two-dimensional velocity field reconstruction model is as follows:
[0015] First, the data preprocessing part of the two-dimensional velocity field reconstruction model is used to interpolate the input isolated section one-dimensional pressure data, so as to dimension it into a two-dimensional matrix data with high resolution;
[0016] Then, the block feature extraction part is used to divide the dimensioned pressure data into four parts, and each part of the matrix data is sent to a separate convolution-deconvolution feature extraction path for processing to extract features;
[0017] Then, the feature splicing part is entered, and two splicing operations are performed to obtain a single-channel matrix with the same size as the output layer, and then multiple convolution-deconvolution layers are used to extract and fuse the flow field features after splicing;
[0018] Finally, the full connection layer of the output part is used to obtain the two-dimensional velocity field reconstructed by the model.
[0019] Further, the construction process of the training set for model training is as follows:
[0020] The one-dimensional pressure and two-dimensional velocity distribution results under different Mach numbers and different fuel injection equivalence ratios are obtained, and the one-dimensional pressure and two-dimensional velocity distribution results under each working condition are one-to-one corresponding. Then, the data is normalized and pooled to obtain the training set of the reconstruction model, and the data set is down-sampled for model training.
[0021] Further, the loss function used in the training process is:
[0022]
[0023] In the above formula, F loss represents the loss function; U tru represents the numerical calculation of the two-dimensional velocity distribution information; U pre is the two-dimensional velocity prediction information calculated by the neural network according to the input and the weight coefficient; U represents the two-dimensional velocity field studied; and the size of the two-dimensional section parameter field can be represented as m x n.
[0024] Further, the two-dimensional velocity field reconstruction model adopts an offline training mode.
[0025] Further, the two-dimensional velocity field reconstruction model precision index includes the flow field average error ξ, the linear correlation coefficient R and the structural similarity index SSIM, wherein the formula for calculating SSIM is:
[0026]
[0027] In the above formula, U tru and U pre respectively represent the numerical calculation true value of the two-dimensional velocity distribution of the isolated section and the predicted value obtained by the two-dimensional velocity field reconstruction model; μ tru and μ pre respectively represent the corresponding mean value; σ tru and σ pre represent the corresponding standard deviation; σ tru,pre represents the covariance of the two; and C1 and C2 are constants.
[0028] The beneficial effects of the present application are:
[0029] The present application innovatively applies the multi-path segmented input reconstruction convolutional neural network model to the reconstruction of the two-dimensional velocity field of the isolated section of the engine, the segmented input mode can reduce the number of flow field features required to be extracted by each path, and each path only needs to extract the local feature information of the segmented part. For the part containing less features, the number of channels and the number of convolution-deconvolution operations can be reduced, so that the learning ability of the model is ensured while the training and reconstruction prediction cost is reduced, thereby obtaining a high-precision two-dimensional velocity field reconstruction model.
[0030] The two-dimensional velocity field reconstruction model uses twice splicing to realize the merging of the segmented velocity field, and after splicing, the model can more accurately restore the wall surface velocity and vortex edge and avoid fuzzy or distorted output. The method retains the unique information of each branch, so that even if part of the branch is disturbed by noise, the characteristics of other branches can still support the model decision, reducing the influence of single failure on the overall prediction, thereby improving the robustness of the model reconstruction, so that the model can still stably realize the reconstruction of the two-dimensional velocity field in the case of error or interference of part of the input pressure data.
[0031] The two-dimensional velocity field reconstruction model can accurately explore the position and shape of the shock train in the isolation section by obtaining the pressure parameters of the engine isolation section and using a high-precision reconstruction model, thereby providing sufficient information for engine closed-loop control. At the same time, since the pressure data is obtained in a simple way, the measurement cost required for using the model is greatly reduced, so that the method can be widely used in the monitoring of the velocity field of the engine isolation section. BRIEF DESCRIPTION OF DRAWINGS
[0032] Figure 1 The system diagram of the method for reconstructing the two-dimensional velocity field of the engine isolation section in the application;
[0033] Figure 2 The structural schematic diagram of the two-dimensional velocity field reconstruction model in the application;
[0034] Figure 3 The schematic diagram of the two-dimensional velocity field reconstruction model when segmenting the velocity field of the isolation section in the application;
[0035] Figure 4 The grid structure schematic diagram of the single-path reconstruction convolutional neural network model;
[0036] Figure 5 The training result of the single-path reconstruction convolutional neural network model;
[0037] Figure 6 The training result of the two-dimensional velocity field reconstruction model in the embodiment. DETAILED DESCRIPTION
[0038] The specific embodiments of the application will be described in detail below with reference to the accompanying drawings.
[0039] The application aims to solve the problem of insufficient information acquisition of the two-dimensional velocity field in the isolation section in the field of fluid mechanics. The information acquisition method based on pressure sensors is difficult to obtain complete velocity distribution due to dimensional limitations; the traditional optical shooting method is difficult to be applied to the engine in practice due to the complexity of the equipment and instruments; the numerical method depends on the iterative solution of the fluid mechanics equation, and the calculation complexity is high and it is difficult to capture the nonlinear characteristics; and the existing velocity field reconstruction method has the problems of low resolution and high training cost.
[0040] To this end, the application provides a model architecture integrating data dimensionality lifting, multi-path parallel processing and multi-layer feature fusion, which aims to break through the limitation of one-dimensional data processing, improve the flow field speed feature extraction capability, realize efficient and accurate reconstruction of the velocity field from pressure data, avoid the complex calculation process of traditional numerical simulation, and meet the demand for real-time reconstruction of the velocity field in the actual operation of the engine.
[0041] Referring to Figure 1 The application is a method for reconstructing a two-dimensional velocity field of an engine isolation section, comprising the following steps:
[0042] Step 1: Set different Mach numbers and fuel injection equivalence ratios as test conditions and numerical simulation boundary conditions, collect one-dimensional pressure data of the engine isolation section through sensors, obtain two-dimensional velocity field sections in the isolation section through numerical simulation, and construct a one-dimensional pressure-two-dimensional velocity field data set;
[0043] Step 2: Train a two-dimensional velocity field reconstruction model based on deep learning using the one-dimensional pressure-two-dimensional velocity field data set to obtain a trained two-dimensional velocity field reconstruction model;
[0044] The two-dimensional velocity field reconstruction model is a multi-path segmented input reconstruction convolutional neural network model, which includes an input data preprocessing part, a block feature extraction part, a feature splicing part and an output part.
[0045] The input data preprocessing part includes lifting the input one-dimensional pressure data into a two-dimensional velocity matrix through an interpolation method; the block feature extraction part includes dividing the two-dimensional velocity matrix into four parts and extracting the velocity distribution features through convolution-deconvolution; the feature splicing part includes extracting the distribution features of the two-dimensional velocity matrix through two splicing operations; and the output part includes outputting the reconstructed two-dimensional velocity field image through a fully connected layer.
[0046] Step 3: Based on the two-dimensional velocity field reconstruction model in step 2, reconstruct the pressure data during the operation of the engine to obtain two-dimensional velocity field information.
[0047] In some embodiments, the reconstruction process of the two-dimensional velocity field reconstruction model is as follows:
[0048] First, the one-dimensional pressure data of the isolation section is input into the data preprocessing part of the two-dimensional velocity field reconstruction model for interpolation to lift it into a two-dimensional matrix data with higher resolution;
[0049] Then, the pressure data after lifting is divided into four parts by the block feature extraction part, and each part of the matrix data is sent to a separate convolution-deconvolution feature extraction path for processing to extract features;
[0050] The re-entry feature splicing part obtains and outputs a single-channel matrix with the same size as the layer through two splicing operations, and then extracts and fuses the spliced flow field features by using multiple convolution-deconvolution layers.
[0051] Finally, the two-dimensional velocity field reconstructed by the model is obtained through the full connection layer of the output part.
[0052] In some embodiments, the construction process of the training set for model training is as follows:
[0053] The one-dimensional pressure and two-dimensional velocity distribution results under different Mach numbers and different fuel injection equivalence ratios are obtained, and the one-dimensional pressure and two-dimensional velocity distribution results under each condition are corresponded one by one. Then, the data is normalized and pooled to obtain the training set of the reconstruction model, and the data set is down-sampled for model training.
[0054] In some embodiments, the loss function used in the training process is:
[0055]
[0056] In the above formula, F loss represents the loss function; U tru represents the two-dimensional velocity distribution information obtained by numerical calculation, which can be regarded as the true value of neural network training; U pre is the two-dimensional velocity prediction information calculated by the neural network according to the input and weight coefficient; U represents the two-dimensional velocity field studied; and the size of the two-dimensional section parameter field can be represented as m x n.
[0057] In some embodiments, the two-dimensional velocity field reconstruction model adopts an offline training mode, part of the data is reserved and not used for network model learning during the training process, and the model is used for cross-validation test after the training is completed, so as to improve the robustness of the two-dimensional velocity field reconstruction model and avoid overfitting.
[0058] In some embodiments, the precision index of the two-dimensional velocity field reconstruction model includes the flow field average error ξ, the linear correlation coefficient R and the structural similarity index SSIM, wherein the formula for calculating SSIM is:
[0059]
[0060] In the above formula, U tru and U pre respectively represent the numerical calculation true value of the two-dimensional velocity distribution of the isolation section and the prediction value obtained by the two-dimensional velocity field reconstruction model; μ tru and μ pre respectively represent the corresponding mean value; σ tru and σ pre represent the corresponding standard deviation.tru,pre denotes the covariance of both; C1 and C2 are constants introduced to avoid instabilities in the calculation.
[0061] Embodiment
[0062] The application is a method for reconstructing a two-dimensional velocity field of an engine isolation section, and the specific implementation steps are:
[0063] 1. Obtain one-dimensional pressure data and two-dimensional velocity field data of the engine isolation section by numerical simulation and ground test methods; the extracted pressure data comes from the upper and lower walls of the isolation section, and 20 pressure data points are extracted from each wall.
[0064] 2. Process the obtained data to construct a data set:
[0065] Convert the two-dimensional velocity field into a trainable two-dimensional matrix, and each two-dimensional velocity section matrix has 140x1391=194740 element values. High precision increases the training cost of the neural network, leading to GPU memory overflow and even error during the training process. Therefore, it is necessary to perform downsampling processing. In this example, a 5x5 pooling kernel and a step of 3 are used to perform average pooling processing on the original matrix, and the size of the two-dimensional matrix after processing is 46x463=21298. Then, the one-dimensional pressure-two-dimensional velocity field section data pair is normalized according to the following formula to construct the data set.
[0066]
[0067] In the formula: U * is the normalized velocity value, U is the velocity value before normalization, U max is the maximum value of all velocity values, U min is the minimum value of all velocity values.
[0068] 3. Establish the mapping relationship from the one-dimensional pressure curve of the wall to the two-dimensional velocity field of the isolation section through the two-dimensional velocity field reconstruction model and train it using the constructed data set:
[0069] Use the pressure curves at two positions of the upper and lower walls as the input of the neural network, and the number of pressure measurement points contained in a single pressure curve is 20. Therefore, the size of the input data is 2x20=40. Input the above data into the two-dimensional velocity field reconstruction model as shown in Figure 2 , and reconstruct the two-dimensional velocity field in the model according to the following order: after 3x3 deconvolution and upsampling dimensionality processing based on the bicubic interpolation method, a two-dimensional matrix distribution with image significance is obtained, and the two-dimensional velocity field is reconstructed according to Figure 3As shown, it is divided into 4 parts, and each part of the matrix data is transported to a separate convolution-pooling-deconvolution feature extraction path for processing; the path is sequentially provided with a 2x2 convolution layer, a 2x2 average pooling layer, a 2x2 convolution layer, a 2x2 average pooling layer, a 2x2 deconvolution layer, a 2x2 convolution layer, a 2x2 deconvolution layer, and a 2x2 convolution layer, and through the convolution-pooling-deconvolution operation, the abstract feature extraction from local to global is realized; finally, the flow field features are fused through twice splicing operation and inserting the above convolution-pooling-deconvolution layer in splicing, and a single-channel matrix with the same size as the output layer is obtained and output, and finally, after cutting and shape reconstruction, the two-dimensional velocity field reconstructed by the model is output through the full connection layer of the output position.
[0070] The training adopts a batch stochastic gradient descent learning strategy using an Adam optimizer, which learns by randomly selecting data with a batch size from a shuffled training set, adjusts the learning rate in real time according to the gradient of the loss function with respect to the network structure parameters, and thus corrects the network structure parameters.
[0071] 4. Analyze the reconstruction accuracy of the model to ensure that the results are reliable.
[0072] During model training, a single-path reconstruction convolutional neural network model as shown in Figure 4 is also used for training as a control group for precision comparison. Figure 4 The single-path reconstruction convolutional neural network model as shown in
[0073] The reconstruction results of the single-path reconstruction convolutional neural network model as shown in Figure 4 and the two-dimensional velocity field reconstruction model in the present application are compared and analyzed, mainly considering three evaluation indexes: flow field average error ξ, linear correlation coefficient R, and structural similarity index SSIM. Among them, the flow field average error ξ is the average of the reconstruction error of the entire two-dimensional velocity distribution matrix, and the closer the value is to 0, the smaller the reconstruction error. The linear correlation coefficient R represents the similarity of two two-dimensional matrices after linearization, and the closer the value is to 1, the higher the reconstruction accuracy of the model. SSIM mainly evaluates the similarity of the mean and data structure of two two-dimensional matrices, and the closer the calculation result is to 1, the more similar the two two-dimensional velocity distribution matrices, i.e. the higher the reconstruction accuracy of the model.
[0074] Figure 4 The training results of the single-path reconstruction convolutional neural network model as shown in Figure 5The straight line in the figure is the original data, and the point distribution is the comparison of the reconstructed speed and the true speed in the training effect. In the training result of the single-path reconstruction convolutional neural network model, the flow field average error ξ is 22.46%, the linear correlation coefficient R is 0.6405, and the structural similarity index SSIM is 0.4895, Figure 5 The error between the predicted speed result point and the true value is large, and the distribution is more discrete. The training result of the two-dimensional velocity field reconstruction model used in the application is as shown in Figure 6 The straight line in the figure is the original data, and the point distribution is the comparison of the reconstructed speed and the true speed in the training effect. In the training result of the single-path reconstruction convolutional neural network model, the flow field average error ξ is 22.46%, the linear correlation coefficient R is 0.6405, and the structural similarity index SSIM is 0.4895, Figure 6 The error between the predicted speed result point and the true value is small, and the distribution is low. It is found by comparison that the two-dimensional velocity field reconstruction model used in the embodiment has smaller reconstruction error, higher similarity degree of the reconstruction result and the original distribution, higher model reconstruction precision, and more concentrated predicted result distribution shown by the training result, thereby indicating that the reconstruction model used in the embodiment has higher reconstruction precision and can be used as a reconstruction model for the engine isolation section velocity field reconstruction.
Claims
1. A method for reconstructing the two-dimensional velocity field of an engine isolation section, characterized in that: The steps include: Step 1: Set operating conditions at different Mach numbers and fuel injection equivalence ratios as test conditions and numerical simulation boundary conditions, collect one-dimensional pressure data of the engine isolation section through sensors, obtain a two-dimensional velocity field section within the isolation section through numerical simulation, and construct a one-dimensional pressure-two-dimensional velocity field data set; Step 2: Using the one-dimensional pressure-two-dimensional velocity field dataset, training a two-dimensional velocity field reconstruction model based on deep learning to obtain a trained two-dimensional velocity field reconstruction model; The two-dimensional velocity field reconstruction model is a multi-path segmentation input reconstruction convolutional neural network model, and its structure includes: Input data preprocessing part: the input one-dimensional pressure data is upgraded to a two-dimensional velocity matrix through interpolation method; Block feature extraction part: the two-dimensional velocity matrix is divided into four parts, and the velocity distribution features are extracted using the convolution-deconvolution method; Feature splicing part: two splicing steps are used to extract the distribution features of the two-dimensional velocity matrix; Output part: Use the fully connected layer to output the reconstructed two-dimensional velocity field image; Step 3: Based on the two-dimensional velocity field reconstruction model described in step 2, the pressure data during the engine operation is reconstructed to obtain two-dimensional velocity field information.
2. The method for reconstructing the two-dimensional velocity field of the engine isolation section according to claim 1, characterized in that: The reconstruction process of the two-dimensional velocity field reconstruction model is as follows: First, the data preprocessing part of the two-dimensional velocity field reconstruction model interpolates the input one-dimensional pressure data of the isolated section to upgrade it into a two-dimensional matrix data with higher resolution. Then, through the block feature extraction part, the dimension-enhanced pressure data is evenly divided into four parts, and each part of the matrix data is sent to a separate convolution-deconvolution feature extraction path for processing to extract features; Then enter the feature splicing part, after two splicing operations, a single-channel matrix of the same size as the output layer is obtained, and then multiple convolution-deconvolution layers are used to extract the spliced flow field features and fuse them; Finally, the two-dimensional velocity field reconstructed by the model is obtained through the fully connected layer of the output part.
3. A method for reconstructing the two-dimensional velocity field of an engine isolation section according to claim 1 or 2, characterized in that: The process of constructing the training set for obtaining model training is as follows: The one-dimensional pressure and two-dimensional velocity distribution results under different Mach numbers and different fuel injection equivalence ratios are obtained, and the one-dimensional pressure and two-dimensional velocity distribution results under each working condition are matched one by one. Then, the data is normalized and pooled to obtain the training set of the reconstructed model, and the data set is downsampled for use in model training.
4. The method for reconstructing the two-dimensional velocity field of the engine isolation section according to claim 3, characterized in that: The loss function used during training is: In the above formula, F loss represents the loss function; U tru Represents the two-dimensional velocity distribution information obtained by numerical calculation; U pre is the two-dimensional velocity prediction information calculated by the neural network based on the input and weight coefficients; U represents the two-dimensional velocity field under study; the size of the two-dimensional section parameter field can be expressed as m×n.
5. The method for reconstructing the two-dimensional velocity field of the engine isolation section according to claim 3, characterized in that: The two-dimensional velocity field reconstruction model adopts offline training.
6. The method for reconstructing the two-dimensional velocity field of the engine isolation section according to claim 3, characterized in that: The accuracy indicators of the two-dimensional velocity field reconstruction model include the flow field average error ξ, the linear correlation coefficient R and the structural similarity index SSIM, wherein the formula for calculating SSIM is: In the above formula, U tru and U pre They represent the numerically calculated true value of the two-dimensional velocity distribution of the isolation section and the predicted value obtained by the reconstruction calculation of the two-dimensional velocity field reconstruction model; μ tru and μ pre They represent the corresponding means respectively; σ tru and σ pre represents the corresponding standard deviation; σ tru,pre represents the covariance between the two; C1 and C2 are constants.