Joint detection method and system for flow state identification and schlieren generation of supersonic air inlet channel

Abstract

The joint detection system of the supersonic air inlet channel is provided based on deep learning, the system is mainly divided into a supersonic air inlet channel flow state identification system and a flow state schlieren generation system, in the identification system, real-time pressure signals are received through a pressure sensor installed on the wall face of the air inlet channel, and the flow state schlieren generation system is obtained; detailed air inlet flow state detection is realized based on multi-sensor information fusion and a residual network, and a result is transmitted to a next-stage system; and an enhanced Pix2pix model in the flow regime schlieren generation system receives the real-time flow regime and generates a schlieren image corresponding to the flow regime according to the flow regime, so that the state detection of the supersonic air inlet channel is realized. Compared with a previous supersonic air inlet channel state detection method, the combined system realizes more detailed flow state detection by means of multi-sensor information fusion, generates a corresponding schlieren image according to a detection result, does not need to depend on a high-speed camera, is suitable for various working conditions, can obtain real-time flow state and flow field schlieren, and is high in practicability. The limitation that the flow field schlieren is difficult to obtain in an actual flight test is broken through.

Description

Technical Field

[0001] The present invention belongs to the technical field of supersonic inlet flow pattern detection, and specifically relates to supersonic inlet flow pattern detection and flow field schlieren image generation based on deep learning, as well as corresponding improvement schemes. Background Art

[0002] Scramjet engines have shown great potential in supersonic flight in the aerospace field with their excellent thrust output, combustion efficiency and lightweight design. As its key component, the scramjet engine inlet compresses the incoming flow, which directly affects the overall efficiency and normal operation of the engine. Generally speaking, the normal working state of the inlet is called the inlet start-up state, while the inlet non-start-up state is an abnormal working condition, which seriously reduces the air mass flow of the engine, resulting in thrust loss and transient mechanical load. The joint flight test of NASA and CIAM, the joint flight test of the United States and Australia, and the second flight test of Boeing's X-51A hypersonic aircraft all failed to achieve the expected goals due to the inlet non-startup. In the current work related to supersonic inlets, a reliable method is needed to identify the internal flow state of the supersonic inlet to provide a basis for adjusting the inlet state.

[0003] There are some studies focusing on the complex flow patterns and mechanisms of supersonic inlets. Generally speaking, when the free stream Mach number is greater than 2.5, the downstream throttling will cause disturbances and propagate to the inlet inlet in the form of a terminal shock wave. This terminal shock wave is manifested as a shock wave train or an oblique shock wave. When the terminal shock wave is located downstream of the throat, the flow pattern of the supersonic inlet will maintain a relatively stable start-up state. When the back pressure increases, the terminal shock wave will continue to move upstream. When the terminal shock wave approaches the inlet throat, the inlet will enter a critical state. If the back pressure continues to increase, the terminal shock wave will be quickly discharged from the inlet in an instant, resulting in a non-starting state. From the perspective of flow pattern, the non-starting state can be defined as the shock wave train system being discharged from the inlet, and from the perspective of performance, it can be defined as the flow characteristics of the inlet inlet being changed by the internal flow. Many factors may lead to inlet non-starting, such as abnormal combustion of fuel, inappropriate internal contraction ratio of the inlet, and increased back pressure, which may become potential factors for inlet non-starting. In many cases, inlet failure is the result of multiple factors rather than a single cause. This complexity often poses a challenge to effectively detecting inlet failure.

[0004] The complete supersonic inlet non-start protection system mainly consists of three parts: inlet flow pattern identification, non-start margin determination, and non-start protection system. Inlet flow pattern identification is the most critical step and the cornerstone of the subsequent two stages. A reliable identification method must be used to provide accurate flow pattern identification. Chang proposed a support vector machine-recursive feature elimination algorithm based on voltage stabilization information and a probability output support vector machine algorithm. Trapier et al. introduced the cumulative sum and generalized likelihood ratio methods into non-start detection. Wang et al. integrated change point detection, wavelet packet decomposition, and deep learning to achieve real-time non-start detection. These works have greatly promoted the understanding and recognition capabilities of non-start phenomena in this field, and can realize the detection of non-start states of supersonic inlets. However, these methods have some limitations in practical applications, including time-consuming, calculation speed, and difficulty in parameter selection. In addition, these methods can only make simple judgments on the flow pattern in the inlet, and it is difficult to collect detailed information on the internal flow field. In laboratory experiments, high-speed schlieren acquisition systems are often used to capture the flow schlieren images of the inlet to reconstruct the flow field. However, the high-speed schlieren acquisition system is easily interfered by light and electromagnetic signals, which limits its application in actual flight. In recent years, the development of deep learning technology has brought new development directions to high-speed schlieren acquisition technology. By learning historical flow field schlieren images, the present invention can construct flow field schlieren for real experiments without the help of high-speed cameras.

[0005] In actual flight tests, through the pressure sensors installed on the wall of the supersonic inlet, the multi-sensor pressure signals in each time window will be transmitted to the central host. The joint system in the central host will accurately identify the current flow field state, and the identification results will be transmitted to the Schlieren image generation model to generate the corresponding Schlieren image. The joint system will be introduced in detail below. Summary of the invention

[0006] In view of the above problems, the present invention proposes a new joint system consisting of a flow state recognition model and a schlieren image generation model, as well as a corresponding improvement scheme. The joint system is suitable for supersonic inlets of different working conditions and different models, and can generate flow state schlieren without the aid of a high-speed camera, breaking through the limitations of the environment on the detection equipment, and has a new direction and excellent prospects for practical applications.

[0007] With respect to the above-mentioned combined system, the present invention adopts the following technical solutions:

[0008] In terms of flow state recognition, the system adopts a joint architecture of wavelet packet decomposition (WPD) and convolutional neural network (CNN), and fuses the pressure information of multiple sensors to capture the time-frequency characteristics and fusion characteristics of multi-sensor pressure signals. Residual connection and weight loss function are also embedded in the CNN network to achieve more accurate flow state detection. In terms of schlieren image generation, the present invention proposes a flow field schlieren generation model (enhanced Pix2pix schlieren generationnetwork, EPSGN) based on the Pix2pix framework. Pix2pix is ​​an image-to-image framework widely used for image conversion and classification, which can learn the mapping relationship between input images and output images. In order to further improve the quality of crack images, the present invention uses semantic labels as supplementary inputs for schlieren image generation to promote the learning process of the generation network and enhance its understanding of the flow field characteristics of the supersonic inlet.

[0009] The key part of the whole system is to establish the time window p from the pressure signal i Flow pattern inside the inlet duct i And from the flow state i To Schlieren Image I i These mapping relationships f1 and f2 are represented by the trained neural network:

[0010] s i =f1(p i ) (1)

[0011] I i =f2(s i ) (2)

[0012] The whole system is mainly divided into two parts, such as Figure 1 As shown in the figure. In the first part, the pressure signals obtained from multiple sensors are initially processed by WPD to produce time-frequency matrices containing joint features in the time and frequency domains. These time-frequency matrices are used as inputs to CNN to update the network parameters through back propagation. After completing the specified number of iterations, the recognition model can effectively identify the inlet state related to the pressure signal with high accuracy. The second part is based on the Pix2pix framework. EPSGN can receive the flow state results of the first part and pass the recognition results together with the original image to the generation model to obtain the flow field image.

[0013] First, the mechanism and framework of the flow state identification model and flow state schlieren ecological model of the joint system will be introduced in turn.

[0014] 1Preparatory knowledge.

[0015] 1) Wavelet packet decomposition and convolutional neural network

[0016] Figure 1 The first part shows the overall framework of the recognition model, which mainly includes WPD and CNN layers. Developed from wavelet decomposition, WPD decomposes the original time series into subsequences containing low-frequency and high-frequency information respectively. However, unlike wavelet decomposition that discards high-frequency information, WPD systematically decomposes low-frequency and high-frequency information level by level, thereby achieving multi-level and full-band time-frequency characteristics. The structure is as follows Figure 2 As shown. The symbols "A" and "D" represent the low-frequency and high-frequency information in the signal respectively. The low-frequency and high-frequency parts of each stage pass through the low-pass filter L i and high pass filter H i Further decomposition is suitable for more complex signal processing tasks. The choice of wavelet basis and decomposition level is particularly important for WPD. Generally speaking, the higher the number of decomposition levels, the higher the signal resolution and the higher the computational complexity. The decomposition of WPD can be simply expressed by the following formula: the dynamic pressure signal is decomposed into low-frequency information A and high-frequency information D, S represents the pressure signal, and f l and f h Determined by the selected wavelet basis, they represent the low-frequency and high-frequency filter coefficients of wavelet packet decomposition:

[0017] A=S*f l (3)

[0018] D=S*f h (4)

[0019] The main framework of the recognition model is extended from CNN, a deep learning model that is popular in the fields of fault diagnosis and natural language processing. The main structure of CNN can be divided into convolutional layers, pooling layers, and fully connected layers. The convolutional layer uses convolution kernels to extract multiple types of features, such as boundaries, shapes, and time. The role of the pooling layer is to reduce the feature map generated by the convolutional layer, thereby reducing the computational complexity of the network. The fully connected layer establishes connections between neurons in a given layer and neurons in subsequent layers, so that complex nonlinear mappings can be learned, which are responsible for performing the final classification or regression prediction. The main layers of CNN are briefly introduced below.

[0020] The convolution layer is the most important component of CNN, which uses sliding convolution kernels to extract features from input data and generate feature maps, thereby facilitating the network to acquire features at different levels. In addition, the ability of the convolution layer to capture complex spatial and temporal information of the signal is crucial for tasks such as image recognition and classification. Figure 3The convolution algorithm process is shown. It is worth noting that the input feature map is filled with zeros. This zero filling method can reduce the loss of edge information in the input feature map and improve the stability of the network. The mathematical operation of convolution can be expressed as:

[0021]

[0022] in, It is represented as the output of the jth convolution kernel of the lth layer, W j and b j Respectively represent the learnable weight and bias of the j-th convolution kernel, Represents an activation function, which is used to improve the nonlinear learning ability of the network. In the present invention, all convolutional layers use the ReLU activation function.

[0023] The function of the pooling layer is to reduce the spatial dimension of the input data while retaining important information. Figure 4 As shown in Figure 1, this reduction is achieved by applying max pooling or average pooling. By reducing the spatial dimension, the pooling layer increases the network's robustness to input data variations and reduces computational complexity. The pooling layer is a static computation process and does not need to participate in network training.

[0024] The task of the fully connected layer is to obtain the nonlinear features extracted by the convolutional layer and the pooling layer. It transforms the input features linearly through the weight matrix and the bias vector, and then performs nonlinear transformation through the activation function, so as to further extract and combine the features. Through these transformations, the fully connected layer can further process the previously extracted features and capture higher-level features. It is usually used in the last few layers of the neural network, especially in classification tasks, where the fully connected layer maps high-level features to the category space. Input vector X f With the weight matrix W f Multiply and add the bias vector b f , and the intermediate result Z is obtained. Z is transformed nonlinearly through the activation function to generate the output Y. This process can be expressed as:

[0025] Z=WfX f +b f (6)

[0026]

[0027] After multiple layers of processing, the recognition results are output in the form of categories or probabilities. This output result is compared with the true label to calculate the loss function value. The loss function is an indicator to measure the difference between the predicted result and the true result. It is used to measure the difference between the model output result and the expected output to guide the network to update parameters, such as cross entropy, mean square error, etc. Next, the back propagation algorithm will calculate the gradient of each parameter in the network according to the loss function. Back propagation starts from the output layer and calculates the gradient backward layer by layer until the input layer. The gradient calculation of each layer depends on the chain rule. The subsequent gradient is passed to the previous layer through the activation function and parameters of the current layer, and the parameters are adjusted in the opposite direction of the gradient, so that the value of the loss function is gradually reduced. Through continuous forward propagation, back propagation and parameter update, the parameters of the model are gradually adjusted, and the value of the loss function is continuously reduced. This process is repeated until the predetermined number of training rounds is reached or the loss function value converges.

[0028] 2) Data preprocessing, residual connection and weight loss function

[0029] Under the guidance of the loss function, CNN adjusts the weights and biases of each layer through back propagation to minimize the error between the predicted output and the actual output. However, as the number of network layers increases, some problems are difficult to avoid, such as network overfitting, gradient vanishing, gradient crushing, and network degradation. Gradient vanishing makes it almost impossible to update the front layer of the network, while gradient explosion will lead to unstable parameter updates. Data preprocessing and batch normalization can alleviate these problems, but gradient crushing cannot be avoided. When the number of CNN layers is too large and the gradient converges to white noise, it is difficult for the network to learn effective information from the data. In order to solve these problems, data preprocessing, residual connection and weight loss function are adopted in the recognition model.

[0030] Data preprocessing can reduce the noise in the data and make the pattern clearer. This paper selects the sliding average as the preprocessing scheme. By calculating the average value of a specific time window and making a difference, it eliminates the influence of short-term fluctuations and different dimensions and reveals the potential trend of the time series. The steps of time series preprocessing are as follows:

[0031]

[0032] x′ i =x i -Average i (9)

[0033] Among them, m represents the sliding window size, x i Indicates the i-th dynamic pressure signal window collected by the sensor, Average i represents the sliding mean and x′ i Indicates the pressure point after pretreatment.

[0034] The main advantage of residual connections is that they can optimize the training of deep networks without reducing performance. Specifically, in CNN, the output of each layer is usually used as the input of the next layer. Residual connections add a path on this basis. This path can be a direct connection or a dimensional adjustment through a 1×1 convolution kernel. Through this operation, gradients can be more easily passed to earlier layers, thereby improving the training effect of deep neural networks. Figure 6 As shown in Figure 1, the residual connection effectively alleviates the gradient crushing problem, enabling the network to better extract useful information from the data and improve the overall performance. A residual module can be expressed by the following formula: The present invention splices the residual outputs of multiple sensors and transmits them to the next network. S f represents the fusion information of h sensors, S i Represents the extracted features of each sensor, fi and h i Respectively represent the convolution process and residual connection of the network, Y j represents the model output, and g represents the fusion information S f The feature extraction process, θ g Indicates network parameters:

[0035] S i =f i (x j ,θ i )+h i (x j ), i = 1, 2, ..., h (10)

[0036] S f =Concat(S1, S2, ..., S h ) (11)

[0037] Y j =g(S f ,θ g ) (12)

[0038] In addition to residual connections, weighted loss functions are also used in recognition models to deal with the problem of unbalanced flow states in pressure signals. For supersonic inlet training data, the duration of each flow state is difficult to balance. It is likely that some flow states only last for a short time, while some flow states occupy most of the time window, resulting in the model being unable to effectively learn the characteristics of a few flow states. The weighted loss function assigns different weights to each flow state, thereby improving the performance of the model and helping the model to better generalize to unseen data. The following formulas are the cross entropy loss function and the weighted cross entropy loss function, respectively. It can be seen that the difference between the two lies mainly in the weight of category m.

[0039]

[0040] 3) Pix2pix model and semantic labeling

[0041] In the experimental model, the Schlieren image of the flow field inside the supersonic inlet can be obtained with the help of a high-speed camera. However, high-speed cameras have extremely high requirements for the installation environment. Slight deflection of light or weak electromagnetic signal disturbance may cause a significant decrease in the quality of the Schlieren image. The actual high-speed flight environment is difficult to meet the configuration requirements of high-speed cameras. Therefore, in the current research on supersonic inlets, Schlieren images can basically only be applied to laboratory experiments and are difficult to extend to actual flights. Generative adversarial networks (GANs) provide a new application direction. GAN is a deep learning model composed of two models: a generator and a discriminator. The discriminator evaluates the authenticity of the generated image, while the generator aims to generate images that can effectively deceive the discriminator. The two networks are trained together in an adversarial way. The generator constantly tries to generate more realistic data to make the discriminator unable to distinguish between true and false as much as possible, while the discriminator is constantly improved to better distinguish between real data and generated data. In the end, the generator can generate high-quality data similar to the distribution of real data. Conditional GAN ​​enhances the basic GAN architecture by introducing conditional information. Compared with traditional GAN, conditional GAN ​​can add conditional information to control specific features of the generated image, thereby improving controllability and predictability. Conditional information can be any form of data according to specific task requirements, such as category labels, text descriptions, other features, etc. Pix2pix can be regarded as a specific application example of conditional GAN, which is mainly used for image-to-image conversion tasks. It converts the input image into the target image through conditional information. The present invention converts the blank inlet schlieren into the schlieren image of the corresponding flow state based on the Pix2pix model. In order to comprehensively measure the quality of the generated image and the real image, the visual geometry group (VGG) loss and loss. The loss helps preserve details and textures, while the VGG loss is used as a perceptual metric in image processing and computer vision, i.e., measuring the similarity of images using features extracted from a pre-trained neural network. The update process is as follows:

[0042]

[0043] Among them, x is the information used as conditional input, y is the real image, z is the noise input to the network, G(x,z) is the image generated by the condition and noise, and D(x,y) and D(x,G(x,z)) represent the probability that the discriminator determines that the real image is real and the probability that the fake image is fake, respectively. represents the probability that the discriminator determines that the real picture is real and takes the average of all samples, It means that the discriminator determines that the fake picture is fake and takes the average of all samples, y is the real picture, and L VGG (G) are additional loss functions, λ1 and λ2 represent the coefficients corresponding to the loss functions, G * Represents the optimal solution after the two game generators.

[0044] The generator generates realistic data as much as possible to deceive the discriminator, while the discriminator works to improve its discrimination ability to distinguish between generated images and real images. As training progresses, the generated images will gradually become more realistic, making it more difficult for the discriminator to distinguish between real samples and generated samples. Once the generated images reach a high enough level of realism, it will be difficult for the discriminator to accurately distinguish between them. At this stage, the generator has effectively learned to generate high-quality images, which marks the completion of the training process.

[0045] The Pix2pix model uses images as input conditions, eliminates input noise, and greatly improves the quality of generated images. In addition, the generator of the Pix2pix model uses the U-Net network structure and the discriminator uses PatchGAN. Figure 7 The U-Net network structure is shown. In the encoder-decoder network, the input is first encoded by the encoder, and the decoder reconstructs the encoded representation to produce the output. This structure is widely used in tasks such as machine translation, semantic generation, and dialogue systems, and has achieved good results. The residual network and U-Net are further developed based on the encoder-decoder network. Both use multi-scale feature fusion technology. They combine low-level features with high-level features to improve network performance and generalization ability. Figure 8 As shown in Figure 1, PatchGAN divides the image into small blocks (such as 2×2 pixels) and analyzes each block. This strategy enables PatchGAN to evaluate the authenticity of the image in more detail, thereby improving the quality and diversity of generated images.

[0046] The present invention introduces semantic labels into the Schlieren generation model. Semantic labels play an important role in the fields of computer vision and natural language processing. By labeling various parts in the data, the model can understand and process the data. For example, in Schlieren generation, semantic labels can label different objects in the image, such as air intake equipment, background, flow field, etc., so that the model can identify and segment various elements in the image. Semantic labels not only improve the interpretability of data and the understanding ability of the model, but also enhance the performance of the model in specific tasks, so that the model can make predictions by learning the relationship between labels and data, thereby improving accuracy and reliability. In image generation tasks, semantic labels can improve the diversity and detail quality of generated images, help the discriminator better judge whether the generated image meets the specified semantic requirements, and improve the discriminant ability of the discriminator and the overall generation effect of the model.

[0047] 2 Implementation framework of supersonic inlet joint detection system

[0048] After introducing the implementation principles of flow state recognition and schlieren generation respectively, the implementation framework and process of these two parts will be introduced respectively.

[0049] 1) Supersonic inlet flow pattern identification

[0050] Algorithm 1 summarizes the process of supersonic inlet flow pattern identification.

[0051]

[0052] 2) Supersonic Inlet Schlieren Generation

[0053] Algorithm 2 summarizes the process of supersonic inlet flow pattern identification.

[0054]

[0055] Beneficial effects: (1) The supersonic inlet flow pattern recognition system adopts multi-sensor information fusion and residual network architecture to integrate the time-frequency information of multiple sensor pressure signals to achieve accurate flow pattern recognition;

[0056] (2) Based on the Pix2pix framework, a generative model for generating schlieren images is implemented to achieve high-fidelity image-to-image conversion, allowing the system to generate schlieren of the current flow state without the help of high-speed cameras and post-processing;

[0057] (3) The EPSGN model uses semantic labels as auxiliary input to enhance the effect of flow field generation by capturing complex details and preserving structural information;

[0058] (4) The combined detection system for supersonic inlet flow state identification and schlieren generation can receive real-time pressure signals and return the current flow state, and generate schlieren of the flow field under the current working conditions based on the current flow state. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is the overall framework of the supersonic inlet joint detection system;

[0060] Figure 2 It is the WPD decomposition process;

[0061] Figure 3 It is the deep learning convolution algorithm process;

[0062] Figure 4 It is the pooling operation of the pooling layer;

[0063] Figure 5 It is the fully connected layer operation process;

[0064] Figure 6 Schematic diagram of residual connection;

[0065] Figure 7 This is a schematic diagram of the U-Net framework;

[0066] Figure 8 This is a schematic diagram of PatchGAN;

[0067] Fig. 9 is a schematic diagram of the test model;

[0068] Fig.10 It is the flow field schlieren change from starting to not starting with ICR=1.54;

[0069] Fig.11 It is the change of the sensor pressure signal from starting to not starting when ICR=1.54;

[0070] Fig.12 is the comparison result of the generated schlieren image (ICR = 1.42);

[0071] Fig.13 is the comparison result of the generated Schlieren image (ICR=1.91). DETAILED DESCRIPTION

[0072] The training set and test set of the present invention are from the ground test simulating the supersonic inlet conditions under real supersonic flight. The model is a simplification of the low-speed pipeline of the turbine combined cycle engine inlet. The overall experimental equipment is as follows: Fig. 9As shown, it mainly includes a vacuum tank, a diffuser, an inlet model with an optical window, a dynamic pressure data acquisition system, a high-speed camera, a Laval nozzle, etc. Table 1 lists some important parameters of the inlet and wind tunnel system. In order to accurately capture the internal characteristics of the flow, a 145×29 mm optical window is installed on one side of the inlet model, and all connections of the internal pipes of the model are treated with sealants and sealing strips to minimize noise interference. The wind tunnel system used in this experiment includes a variable rectangular Laval nozzle, a fully enclosed laboratory, a 400 cubic meter vacuum tank and a related vacuum pump, and a nozzle with an actual Mach number of 2.9 is used for simulation. A movable plugging cone is set at the outlet of the supersonic inlet to achieve downstream throttling, so that the flow field inside the inlet changes from starting to not starting. The data acquisition system is the focus of attention. The data acquisition system of this experiment is divided into two parts: the pressure signal acquisition system and the flow field schlieren monitoring system. The pressure signal acquisition system is mainly composed of 25 CYG-503 pressure sensors, which are installed on the fairing (C1~C13) and the external ramp (R1~R12) respectively, and the data is collected through a DAQ PCI-6255 data acquisition card with a sampling frequency of 20kHz. The flow field schlieren monitoring system consists of a light source, several discrete mirrors and a MEMRECAM HX-3 high-speed digital camera with a frame rate of 6000. In order to achieve time synchronization between the two systems, a synchronization signal is used to trigger the high-speed camera to start recording schlieren images. The ICR of the inlet model (defined as the area ratio of the pipe inlet and the throat) is adjusted by a four-bar linkage, and 7 sets of pressure signals and schlieren images under different ICR conditions are recorded.

[0073] With the help of a data acquisition system, the occurrences of non-starting were recorded. Fig.10 The Schlieren images of the flow field at the inlet from start-up to non-start-up when ICR=1.54 are shown. The internal flow field tends to be stable at first, which is called the start-up state. As the blocking cone moves, the pressure disturbance caused by the throttling propagates upstream, resulting in the formation of the tail shock wave. The movement of the tail shock wave causes disturbances and pressure fluctuations in the internal flow field. When the head of the tail shock wave approaches the throat, the supersonic inlet is considered to be in a critical state. Subsequently, when the head of the tail shock wave passes through the throat, the inlet suddenly transitions to a non-start-up state, affecting the capture characteristics of the inlet. As the blocking cone continues to move, the inlet may experience a more violent non-start-up state, resulting in more overflow on one side of the fairing. Specifically, when the tail shock wave moves upstream, the inlet may experience six typical flow states. The movement of the tail shock wave also causes fluctuations in the wall pressure. Some typical sensors are selected to identify the flow state, including C7, C5, C1, R11 and R2. Fig.11The pressure signals received by these sensors when ICR = 1.54 are shown. The change curve of the pressure signal confirms the change of the flow field inside the inlet duct. Table 2 shows the specific classification intervals of the six flow states. It is worth mentioning that under the condition of large ICR, the end shock wave will enter the non-starting state very quickly after passing through the throat. In other words, under certain conditions, the subcritical state may appear for a very short time or may not even appear.

[0074] In order to prove the effectiveness of the joint detection system and improved method for supersonic inlet flow pattern recognition and schlieren generation, the present invention tests supersonic inlet flow pattern recognition and flow field schlieren generation on the test set respectively. During the test, 7 groups of experimental conditions were divided into 5:2, 5 groups of experimental conditions were used for training, and 2 groups of experimental conditions were used as unknown working conditions to test the effectiveness and performance improvement of the joint detection system. For supersonic inlet flow pattern recognition, the present invention mainly adopts the architecture of weight loss function and residual connection, and compares the performance of the flow pattern recognition model after adopting these optimization methods. For supersonic inlet flow field schlieren generation, the present invention mainly compares the effects of U-Net structure and semantic label on the generation model, trains the generation model on the schlieren pictures of the training set, and proves its effectiveness and superiority on the test set.

[0075] The specific implementation of the present invention is mainly composed of the following steps:

[0076] Step 1: Conduct tests on the hybrid compression supersonic inlet model and collect 7 sets of pressure signals and flow field schlieren to verify the flow field change process of the supersonic inlet from starting to not starting;

[0077] Step 2: Train the flow state recognition model on the pressure signal training set and test it under test conditions to verify the improvement effect of the weight loss function and residual connection on the recognition accuracy;

[0078] Step 3: Train EPSGN on the flow field schlieren training set, and generate the corresponding flow field image by receiving the flow state recognition results under test conditions, and compare the effect of U-Net structure and semantic labels in the generation model to prove its improvement effect;

[0079] Step 4: Test the time required for the joint detection system of supersonic inlet flow pattern recognition and schlieren generation to identify a time window to prove its real-time performance.

[0080] 1. Test supersonic inlet flow pattern identification and improvement measures;

[0081] The main parameters of the flow state recognition model are shown in Table 3. In order to further improve the performance of the model, the weighted cross entropy loss function and residual connection are applied to the model. Weighted cross entropy assigns weights to different categories before calculating the loss function. These weights are usually determined by the number of samples in each category. This method allows the model to pay appropriate attention to categories with fewer samples during training and balance the problem of imbalanced sample numbers. Table 4 shows the recognition accuracy of the model in two test data sets. It can be seen that the model with both residual connection and weighted loss function achieved an overall accuracy of 99.23%, with the best recognition effect. The results show that WPD-CNN can accurately judge the current flow state of the supersonic inlet by integrating the time-frequency information of multiple sensors, providing a reliable basis for the subsequent generation of the flow field.

[0082] 3. Use EPSGN to realize the generation of supersonic inlet flow field schlieren;

[0083] According to 6 different flow field states, the schlieren images are grouped, sampled and preprocessed to form 64×256 3-channel RGB images. For comparison, the present invention adopts three different generative model architectures, one is a basic model (GAN 1), a basic model with a U-Net structure (GAN 2), and the other is EPSGN (with U-Net and semantic labels), and all other model parameters are the same. The model will be trained using a training dataset, and its performance in generating images will be evaluated on a test dataset. Each flow state training dataset contains 250 schlieren images. It should be noted that in the ICR=2.04 experiment, no critical state and initial instability were observed, and the flow state directly transitioned from the critical state to severe instability, resulting in only 200 images in the corresponding dataset. The performance can be compared from the following aspects:

[0084] (1) Human vision (HV). Human vision refers to the use of human senses to evaluate the quality of generated images. A subjective scoring method is usually used to obtain an assessment of image quality. The purpose is to determine whether the generated image meets human expectations. In the present invention, images that show significantly clearer and more detailed features are considered superior, while blurred images are considered unsuccessful, and other images are considered to have no significant differences.

[0085] (2) Peak signal-to-noise ratio (PSNR). PSNR is a standard metric for measuring image quality and is often used to evaluate the difference between a compressed image and the original image. PSNR quantifies the quality of a signal by comparing the maximum possible signal value to the size of the noise. In image processing, the signal is usually the pixel value of the image, while the noise is the error introduced during the compression or transmission process. The higher the PSNR value, the better the image quality and the smaller the difference between the compressed image and the original image. Generally speaking, a PSNR value greater than 30 is considered to be good image quality, while a PSNR value less than 20 means that the image quality has significantly deteriorated. PSNR is a simple and effective metric, but it only considers the pixel differences in the image and does not necessarily reflect the perceptual differences of the human visual system.

[0086] (3) Structural similarity (SSIM). It is an indicator to measure the similarity between two images and is mainly used to evaluate the quality of images. It evaluates the similarity of images by comparing the brightness, contrast and structural information of the images. The value of SSIM is between -1 and 1, where 1 means that the two images are exactly the same. The SSIM formula takes into account three aspects of similarity: brightness, contrast and structure. Brightness is used to measure the average intensity of the image, usually expressed as the average value of the local block; contrast measures the range of variation of the image intensity, usually expressed by the standard deviation; structural comparison measures the local structural pattern of the image, usually calculated by the normalized covariance. SSIM calculates the overall similarity of images by combining these three similarity metrics. Compared with PSNR, SSIM focuses on capturing the perceptual characteristics of the human visual system to images and can better reflect the human visual perception of image quality. It can not only evaluate the overall image quality, but also be used to evaluate the quality of local areas, thereby providing more accurate quality evaluation for image processing.

[0087] (4) Perceptual loss (PI). Unlike the human perception system, PSNR and SSIM are both pixel-level evaluation indicators and cannot accurately evaluate the perceptual quality of the image. In addition, for certain types of distortion, such as artifacts, blur, etc., they may not be able to accurately evaluate the true quality of the image. Perceptual loss makes up for this deficiency by comparing the differences in high-level features. PI usually uses a pre-trained convolutional neural network (VGG-19 is selected in this invention) to extract high-level features of the image, which can capture important information such as texture, edges, and object shapes in the image. Then, the distance between the generated image and the target image in these high-level feature spaces is calculated as a measure of perceptual loss. This approach enables perceptual loss to better evaluate the perceptual quality of an image because it takes into account the content and structure of the image, not just the difference in pixel values.

[0088] (5) Learned perceptual image patch similarity (LPIPS). Similar to PI, LPIPS is also based on the principle of visual perception and attempts to simulate the way the human visual system perceives images. Specifically, they both use pre-trained deep convolutional neural networks to extract image features and calculate loss functions. The difference between them is that PI focuses more on the similarity of the overall image, while LPIPS focuses more on the similarity of local details. LPIPS is more commonly used to calculate the similarity between generated images and real images. The LPIPS of the present invention is built on the AlexNet architecture.

[0089] The results are as follows Fig.12 , Fig.13 As shown in the figure. When considering HV, EPSGN shows excellent performance in all flow states, especially in capturing complex flow details such as the stable flow field in steady state, the tail shock wave during upstream motion, the separation of the external flow field under intense non-starting, and the additional overflow on the fairing side. In contrast, GAN 1 performs poorly, producing blurred and low-quality schlieren images under certain conditions, while GAN 2 performs relatively well, with no obvious low-quality images.

[0090] Due to the lack of real schlieren images under some conditions in the ICR=1.91 experiment, the evaluation index comparison experiment of ICR=1.42 is mainly considered. Tables 5-7 show the comparison results of more perceptual indicators under the experimental condition of ICR=1.42. Compared with GAN 1, GAN 2 with U-Net structure performs better in almost all cases, highlighting the effectiveness of U-Net jump connection. Considering HV, EPSGN outperforms other models in critical state and severe inactivation, and GAN 1 seems to generate failed images (blurry or lack of details) in some flow states. In other cases, the three models perform well. Although GAN 2 performs similarly to EPSGN in terms of PSNR and SSIM, they perform similarly in pixel-level details and overall structure, with similar visual quality and information retention. However, the visual effects PL and LPIPS of GAN 2 are not as good as EPSGN, which indicates that the generated images are closer to human visual perception in visual features (such as texture, color, structure, etc.) than real images, which means that the generated images are better in visual quality and can more realistically reproduce the perceptual characteristics of the target image. The present invention mainly uses semantic labels to distinguish the regional information of the inlet internal flow field, external flow field and inlet model, and the flow field generation model can generate images with specified features by capturing the inherent inter-regional correlations.

[0091] From the real flow field schlieren image with ICR=1.91, it can be seen that the schlieren acquisition system failed to capture the subcritical state and the initial non-start phenomenon, and the inlet duct directly mutated from the critical state to the violent non-start phenomenon. Fig.13 It can be clearly seen that although the high-speed camera cannot capture these two states, EPSGN can still generate schlieren images of corresponding experimental conditions by learning other experimental conditions, which shows that it can be applied to experimental conditions that the schlieren acquisition system is difficult to adapt to to capture the schlieren of flow fields under extreme conditions.

[0092] 4. Test the real-time performance of the joint detection system for supersonic inlet flow pattern recognition and schlieren generation;

[0093] The joint detection system for supersonic inlet flow pattern recognition and schlieren generation includes two deep learning models. For real-time detection, the system receives a pressure signal of a time window, the flow pattern recognition system returns the corresponding flow pattern recognition result, and the schlieren generation system receives the flow pattern recognition result and generates the corresponding flow field schlieren. In order to test its real-time performance, the test set is used as a simulation of the real scene, and a pressure signal of a time window (5×1024) is transmitted to the joint detection system at an interval of 100ms. The joint system performs real-time reasoning and measures the total time from the input data entering the system to the output result generation, which is the key indicator for evaluating real-time performance. The experiment was carried out on a system equipped with an NVIDIA RTX 3070 GPU. The model was trained using the PyTorch framework and accelerated by the GPU. For a single time window, the supersonic inlet flow pattern recognition and schlieren generation took 24.21ms and 30.13ms respectively, which met the real-time requirements. This shows that the joint detection system of the present invention still has good application prospects in the real-time control of supersonic inlets.

[0094] Table 1 Geometric parameters of experimental model

[0095]

[0096] Table 26 Classification basis of flow states

[0097]

[0098]

[0099] Table 3 Main parameters of flow pattern identification model

[0100]

[0101] Table 4 Recognition accuracy of flow state recognition model in two test sets

[0102]

[0103] Table 5 Performance indicators of GAN 1

[0104]

[0105] Table 6 Performance indicators of GAN 2

[0106]

[0107] Table 7 EPSGN performance indicators

[0108]

Claims

1. A combined detection method for supersonic inlet flow pattern recognition and schlieren generation, characterized in that: The steps include: Step 1: Install several sensors on the internal flow channel of the supersonic inlet to collect dynamic pressure signals, perform preprocessing and window sliding on the pressure signals, and establish a sample set according to the flow state label. The sample set contains several subsequences, each subsequence contains dynamic pressure information within a time window, and the sample set is divided into a first training set and a first test set; Step 2: Install a Schlieren acquisition system outside the air inlet duct, use a high-speed camera in the Schlieren acquisition system to capture the Schlieren of the flow field change from starting to not starting and establish a Schlieren sample set, assign a sample label to the Schlieren sample set according to the synchronization signal of the high-speed camera and the dynamic pressure sensor, and the Schlieren sample set is divided into a second training set and a second test set; Step 3: The first training set is passed into the flow state recognition model based on deep learning for training. The flow state recognition model uses wavelet packet decomposition to convert the pressure signal time window into a time-frequency feature matrix. On the basis of the residual network and multi-sensor information fusion infrastructure, the pressure information of different sensors is integrated to realize flow state recognition. Step 4: Train the flow field schlieren generative model based on the Pix2pix architecture according to the second training set of different flow states, and use the semantic label of the flow field as an auxiliary input to enhance the performance of the flow field schlieren generative model; Step 5: Test the flow pattern recognition model and the flow field schlieren generation model on the first test set and the second test set. The flow pattern recognition model receives the pressure signal to determine the current flow pattern and transmits the flow pattern recognition result to the flow field schlieren generation model of the corresponding flow pattern. The corresponding flow field schlieren generation model generates a flow field on the blank inlet duct according to the flow pattern recognition result and the semantic label. Step 6: Input the real-time received dynamic pressure signal into the trained flow state recognition model and return the corresponding flow state recognition result. The trained flow field schlieren generative model corresponding to the flow state receives the flow state recognition result and returns the flow field schlieren image of the required experimental conditions.

2. The combined detection method of supersonic inlet flow pattern recognition and schlieren generation according to claim 1 is characterized in that: The sliding average method is used to preprocess the dynamic pressure signal, and the processing formula is as follows: x’ i =x i -Average i (2) Among them, m represents the sliding window size, x i Indicates the i-th dynamic pressure signal window collected by the sensor, Average i represents the sliding mean and x′ i Indicates the pressure point after pretreatment.

3. The combined detection method of supersonic inlet flow pattern recognition and schlieren generation according to claim 1 is characterized in that: The steps of training the flow pattern recognition model in step 3 include: Step 3.1: The training samples in the first training set are input into the wavelet packet decomposition layer. The wavelet packet decomposition layer decomposes the dynamic pressure signal into low-frequency information A and high-frequency information D according to equations (3) and (4). S represents the pressure signal, and f l and f h Determined by the selected wavelet basis, they represent the low-frequency and high-frequency filter coefficients of the wavelet packet decomposition: A=S*f l (3) D=S*f h (4) Step 3.2: Convert the decomposed signal into a time-frequency matrix and enter the deep learning network for training. The deep learning network adopts the structure of multi-sensor information fusion and residual connection. The forward process of training is as follows (5)(6)(7). S f represents the fusion information of h sensors, S i represents the extracted features of each sensor, f i and h i Respectively represent the convolution process and residual connection of the network, Y j represents the final flow state output, and then the training weight loss function L is calculated according to (8) w , w m Represents the weight of each category, M and N represent the number of categories and the number of samples respectively, and μ represents the learning rate; S f =Concat(S1,S2,…,S h ) (5) S i =f i (x j ,θ i )+h i (x j ),i=1,2,...,h (6) Y j =g(S f ,θ g ) (7) Step 3.3: After training for the set maximum number of iterations, the trained flow state recognition model receives the pressure signal as a classifier and determines the current flow state of the supersonic inlet in real time.

4. The combined detection system for supersonic inlet flow pattern recognition and schlieren generation according to claim 1, characterized in that: The steps of training the flow field schlieren generative model in step 4 include: Step 4.1: Preprocess the second training set, convert the schlieren images into a three-dimensional matrix form and assign semantic labels according to the flow field area in the schlieren, and classify the processed sample set according to the flow state; Step 4.2: The Schlieren images of each flow state are respectively passed into the flow field Schlieren generative model to train the corresponding flow field Schlieren generative model. The loss function of the flow field Schlieren generative model is jointly defined by the objectives of the generator and the discriminator. The update of the parameters of the generator and the discriminator is carried out in sequence, as shown in (9)(10)(11). G(x, z) is the image generated by the condition and noise, D(x, y) and D(x, G(x, z)) represent the probability that the discriminator determines that the true image is true and the probability that the false image is false, respectively. represents the probability that the discriminator determines that the real picture is real and takes the average of all samples, It means that the discriminator determines that the fake picture is fake and takes the average of all samples, y is the real picture, and L VGG (G) are additional loss functions, λ1 and λ2 represent the coefficients corresponding to the loss functions, G * Represents the optimal solution of the two game generators: Step 4.3: After the game training process, the generator can generate images that are close to the real flow field schlieren. The schlieren generation system receives the real-time flow state recognition results and uses the blank schlieren and semantic labels as the input of the generator to obtain the flow field schlieren corresponding to the current flow state.

5. A combined detection system for supersonic inlet flow pattern recognition and schlieren generation for executing the method according to any one of claims 1 to 4, characterized in that: It includes a supersonic inlet flow pattern recognition system and a flow pattern schlieren generation system. The supersonic inlet flow pattern recognition system is used to receive real-time pressure signals through a pressure sensor installed on the inlet wall, and realize detailed inlet flow pattern detection based on multi-sensor information fusion and residual network and transmit the results to the flow pattern schlieren generation system; the flow pattern schlieren generation system is used to receive real-time flow patterns and generate schlieren images of corresponding flow patterns according to the flow patterns, enhance the generation effect of the flow field according to semantic labels, and realize real-time detection of the flow pattern of the supersonic inlet.

Images