Combined detection method and system for supersonic inlet flow regime identification and schlieren generation

Abstract

The application proposes a combined detection system for supersonic inlet based on deep learning, which mainly includes a supersonic inlet flow state recognition system and a flow pattern schlieren generation system. In the recognition system, real-time pressure signals are received by pressure sensors installed on the inlet wall, and detailed inlet flow state detection is realized based on multi-sensor information fusion and residual network, and the results are transmitted to the next system. In the flow pattern schlieren generation system, the enhanced Pix2pix model receives real-time flow state and generates corresponding flow pattern schlieren images according to the flow state, realizing the state detection of the supersonic inlet. Compared with the previous supersonic inlet state detection method, the combined system realizes more detailed flow state detection with the help of multi-sensor information fusion, and generates corresponding schlieren images according to the detection results, without relying on high-speed cameras. It is suitable for various working conditions and can obtain real-time flow state and flow field schlieren, breaking through the limitation that flow field schlieren is difficult to obtain in actual flight test.

Description

Technical Field

[0001] This invention belongs to the field of supersonic inlet flow regime detection technology, specifically involving supersonic inlet flow regime detection and flow field schlieren image generation based on deep learning, as well as corresponding improvement schemes. Background Technology

[0002] Scramjet engines, with their excellent thrust output, combustion efficiency, and lightweight design, have demonstrated enormous potential in supersonic flight within the aerospace field. As a key component, the scramjet engine's inlet compresses the incoming airflow, directly impacting the engine's overall efficiency and normal operation. Generally, the normal operating state of the inlet is referred to as the inlet start-up state, while the inlet stop-up state is an abnormal condition that severely reduces the engine's airflow, leading to thrust loss and transient mechanical loads. Joint flight tests by NASA and CIAM, joint US-Australia flight tests, and the second flight test of Boeing's X-51A hypersonic aircraft all failed to achieve their intended objectives due to inlet stop-up. Current work on supersonic inlets requires reliable methods to identify the internal flow state of the supersonic inlet to provide a basis for adjusting the inlet's condition.

[0003] Current research focuses on the complex flow patterns and mechanisms of supersonic inlets. Generally, when the free-flow Mach number is greater than 2.5, downstream throttling induces disturbances that propagate towards the inlet in the form of a tail shock wave, which manifests as a shock train or a slanted shock wave. When the tail shock wave is located downstream of the throat, the flow pattern of the supersonic inlet maintains a relatively stable start-up state. As back pressure increases, the tail shock wave continues to move upstream. When the tail shock wave approaches the inlet throat, the inlet enters a critical state. If the back pressure continues to increase, the tail shock wave will be rapidly expelled from the inlet instantaneously, resulting in a non-starting state. From a flow pattern perspective, the non-starting state can be defined as the shock train system being expelled from the inlet; from a performance perspective, it can be defined as the flow characteristics at the inlet inlet being altered by the internal flow. Many factors can cause the inlet to fail to start, such as abnormal fuel combustion, an inappropriate inlet contraction ratio, and increased back pressure. In many cases, the failure of the intake manifold to start is the result of multiple factors working together, rather than a single cause. This complexity often poses a challenge to effectively detecting intake manifold failure.

[0004] A complete supersonic inlet non-starting protection system mainly consists of three parts: inlet flow pattern identification, non-starting margin determination, and the non-starting protection system. Inlet flow pattern identification is the most critical step, serving as the cornerstone of the subsequent two stages, and requires a reliable identification method to provide accurate flow pattern recognition. Chang proposed a support vector machine-recursive feature elimination algorithm and a probabilistic output support vector machine algorithm based on stable voltage information. Trapier et al. introduced the cumulative sum and generalized likelihood ratio methods into non-starting detection. Wang et al. integrated change point detection, wavelet packet decomposition, and deep learning to achieve real-time non-starting detection. These works have greatly advanced the understanding and identification of non-starting phenomena in this field, enabling the detection of non-starting states in supersonic inlets. However, these methods have some limitations in practical applications, including time consumption, computational speed, and difficulties in parameter selection. Furthermore, these methods can only make simple judgments about the flow patterns within the inlet, making it difficult to collect detailed information about the internal flow field. In laboratory experiments, high-speed schlieren acquisition systems are frequently used to capture schlieren images of the inlet flow pattern to reconstruct the flow field. However, high-speed schlieren acquisition systems are susceptible to interference from light and electromagnetic signals, limiting their application in actual flight. In recent years, the development of deep learning technology has brought new directions to high-speed schlieren acquisition technology. By learning historical flow field schlieren images, this invention can construct flow field schlieren for real experiments without the aid of a high-speed camera.

[0005] In actual flight tests, pressure sensors installed on the walls of the supersonic inlet transmit multi-sensor pressure signals within each time window to the central control unit. The integrated system in the central control unit accurately identifies the current flow field state, and the identification results are fed into the schlieren image generation model to generate the corresponding schlieren image. The integrated system will be described in detail below. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a novel joint system consisting of a flow state recognition model and a schlieren image generation model, along with corresponding improvements. This joint system is applicable to various operating conditions and different types of supersonic inlets, generating flow schlieren images without the need for a high-speed camera. It overcomes environmental limitations on detection equipment, offering a novel direction and excellent prospects for practical applications.

[0007] For the aforementioned combined system, the present invention adopts the following technical solution:

[0008] In terms of flow state recognition, the system employs a joint architecture of wavelet packet decomposition (WPD) and convolutional neural network (CNN), fusing pressure information from multiple sensors to capture the time-frequency and fusion features of the multi-sensor pressure signals. Residual connections and weighted loss functions are also embedded into the CNN network to achieve more accurate flow state detection. Regarding schlieren image generation, this invention proposes an enhanced Pix2pix schlieren generation network (EPSGN) based on the Pix2pix framework. Pix2pix is ​​a widely used image-to-image framework for image transformation and classification, capable of learning the mapping relationship between input and output images. To further improve the quality of the crack images, this invention uses semantic labels as supplementary input for schlieren image generation to facilitate the learning process of the generative network and enhance its understanding of the flow field characteristics of supersonic inlets.

[0009] The key part of the entire system is establishing the time window p from the pressure signal. i Flow patterns inside the intake duct i And from the flow state s i To schlieren image I i The mapping relationships f1 and f2 are represented by a trained neural network:

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

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

[0012] The entire system is mainly divided into two parts, such as Figure 1 As shown. The first part involves processing pressure signals from multiple sensors using a WPD to generate a time-frequency matrix containing joint features from the time and frequency domains. These time-frequency matrices are then used as input to a CNN to update the network parameters through backpropagation. After a specified number of iterations, the recognition model can accurately and effectively identify the inlet state associated with the pressure signal. The second part is based on the Pix2pix framework. EPSGN receives the flow results from the first part and passes the recognition results along with the original image to the generative model to obtain the flow field image.

[0013] First, we will introduce the mechanisms and frameworks for implementing the flow state identification model and the flow schlieren ecological model of the joint system.

[0014] 1. Preliminary knowledge.

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

[0016] Figure 1 The first part presents the overall framework of the recognition model, mainly consisting of 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 which discards high-frequency information, WPD systematically decomposes low-frequency and high-frequency information step by step, 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 components of each stage pass through a 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 decomposition level, the higher the signal resolution, but also 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, where S represents the pressure signal, f l and f h Determined by the selected wavelet basis, the low-frequency and high-frequency filter coefficients of the wavelet packet decomposition are respectively:

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

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

[0019] The main framework of the recognition model is an extension of CNN, a deep learning model widely used in fault diagnosis and natural language processing. The main structure of a CNN can be divided into convolutional layers, pooling layers, and fully connected layers. Convolutional layers use convolutional kernels to extract various types of features, such as boundaries, shape, and time. Pooling layers reduce the feature maps generated by the convolutional layers, thereby reducing the computational complexity of the network. Fully connected layers establish connections between neurons in a given layer and neurons in subsequent layers, allowing the learning of complex nonlinear mappings responsible for performing the final classification or regression prediction. The main layers of a CNN will be briefly introduced below.

[0020] Convolutional layers are the most important component of CNNs. They use sliding convolutional kernels to extract features from the input data and generate feature maps, thereby enabling the network to acquire features at different levels. Furthermore, the ability of convolutional layers to capture complex spatial and temporal information about signals is crucial for tasks such as image recognition and classification. Figure 3The convolution algorithm process is shown. Notably, the input feature map is padded with zeros. This zero-padding method reduces the loss of edge information in the input feature map, improving network stability. The mathematical operation of convolution can be represented as:

[0021]

[0022] in, W represents the output of the j-th convolutional kernel in the l-th layer. j and b j Let these represent the learnable weights and biases of the j-th convolutional kernel, respectively. ReLU represents the activation function used to enhance the non-linear learning ability of the network. In this invention, all convolutional layers use the ReLU activation function.

[0023] The function of pooling layers is to reduce the spatial dimensionality of input data while retaining important information. For example... Figure 4 As shown, this reduction is achieved by applying max pooling or average pooling. By reducing the spatial dimensionality, pooling layers enhance the network's robustness to variations in input data and reduce computational complexity. Pooling layers are a static computation process and do not require participation in network training.

[0024] The task of a fully connected layer is to extract the non-linear features extracted by convolutional and pooling layers. It performs a linear transformation on the input features using a weight matrix and bias vector, followed by a non-linear transformation using an activation function, thereby further extracting and combining the features. Through these transformations, the fully connected layer can further process the previously extracted features and capture higher-level features. It is typically used in the last few layers of a neural network, especially in classification tasks, where it maps high-level features to a class space. Input vector X f With weight matrix W f Multiply and add the bias vector b f The intermediate result Z is obtained, and Z undergoes a nonlinear transformation through an activation function to generate the output Y. This process can be represented as:

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

[0026]

[0027] After multiple layers of processing, the recognition result is output in the form of category or probability. This output is compared with the true label to calculate the loss function value. The loss function is an indicator that measures the difference between the predicted result and the true result, used to measure the difference between the model's output and the expected output, thus guiding the network to update parameters, such as cross-entropy and mean squared error. Next, the backpropagation algorithm calculates the gradient of each parameter in the network based on the loss function. Backpropagation starts from the output layer and calculates the gradient layer by layer until it reaches the input layer. The gradient calculation of each layer relies on the chain rule, passing the gradient of the later layer to the previous layer through the activation function and parameters of the current layer, adjusting the parameters in the opposite direction of the gradient, so that the value of the loss function gradually decreases. Through continuous forward propagation, backpropagation, and parameter updates, the model's parameters are gradually adjusted, and the value of the loss function continuously decreases. This process is repeated until a predetermined number of training epochs are reached or the loss function value converges.

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

[0029] Guided by the loss function, CNNs adjust the weights and biases of each layer through backpropagation to minimize the error between the predicted and actual outputs. However, as the number of network layers increases, some problems become unavoidable, such as overfitting, vanishing gradients, gradient fragmentation, and network degradation. Vanishing gradients make it almost impossible for earlier layers to update, while exploding gradients lead to unstable parameter updates. Data preprocessing and batch normalization can alleviate these problems, but gradient fragmentation cannot be avoided. When the number of CNN layers is too large, causing gradients to converge into white noise, the network struggles to learn effective information from the data. To address these issues, data preprocessing, residual connections, and weighted loss functions are employed in the recognition model.

[0030] Data preprocessing can reduce noise in the data and make the pattern clearer. This invention chooses the moving average as the preprocessing scheme. By calculating the average value of a specific time window and subtracting the values, the effects of short-term fluctuations and differences in units are eliminated, revealing 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] Where m represents the sliding window size, x i This represents the i-th dynamic pressure signal window acquired by the sensor, Average. i Let x' represent the moving average and x'' i This indicates the pressure point after pretreatment.

[0034] The main advantage of residual connections lies in their ability to optimize the training of deep networks without sacrificing performance. Specifically, in CNNs, the output of each layer typically serves as the input to the next layer. Residual connections add a path to this layer; this path can be a direct connection or a path with dimensionality adjusted using a 1×1 convolutional kernel. This allows gradients to be more easily propagated to earlier layers, thereby improving the training performance of deep neural networks. Figure 6 As shown, residual connections effectively alleviate the gradient fragmentation problem, enabling the network to better extract useful information from the data and improve overall performance. A residual module can be represented by the following formula: This invention concatenates the residual outputs of multiple sensors and feeds them into the next network iteration, S f S represents the fused information from h sensors. i Let fi and h represent the extracted features of each sensor. i Y represents the convolution process and residual connection of the network, respectively. j The model output is represented by g, which represents the fused information S. f Feature extraction process, θ g Representing 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] Besides residual connections, weighted loss functions are also used in recognition models to address the problem of imbalanced flow regimes in pressure signals. For supersonic inlet training data, the duration of each flow regime is difficult to balance; some flow regimes may only last a short time, while others occupy most of the time window, causing the model to fail to effectively learn the features of a few flow regimes. Weighted loss functions assign different weights to each flow regime, thereby improving model performance and helping the model generalize better to unseen data. The formulas below are for the cross-entropy loss function and the weighted cross-entropy loss function, respectively. It can be seen that the main difference between the two lies in the weight of category m.

[0039]

[0040] 3) Pix2pix model and semantic labels

[0041] In experimental models, high-speed cameras can be used to obtain schlieren images of the flow field inside supersonic inlets. However, high-speed cameras have extremely high requirements for the installation environment; even slight light deflection or weak electromagnetic disturbances can lead to a significant decrease in schlieren image quality. Real-world high-speed flight environments often fail to meet the configuration requirements of high-speed cameras. Therefore, in current supersonic inlet research, schlieren images are primarily applicable only to laboratory experiments and are difficult to extend to actual flight. Generative adversarial networks (GANs) offer a novel application direction. A GAN is a deep learning model consisting of two models: a generator and a discriminator. The discriminator evaluates the realism of the generated images, while the generator aims to generate images that can effectively deceive the discriminator. These two networks are trained together adversarially. The generator continuously attempts to generate more realistic data, making it as difficult for the discriminator to distinguish between real and generated data, while the discriminator continuously improves to better differentiate between real and generated data. Ultimately, the generator can generate high-quality data with a distribution similar to real data. Conditional GANs enhance the basic GAN architecture by introducing conditional information. Compared to traditional GANs, conditional GANs can add conditional information to control specific features of the generated image, thereby improving controllability and predictability. The conditional information can be any form of data, such as category labels, text descriptions, or other features, depending on the specific task requirements. Pix2pix can be seen as a specific application example of conditional GANs, mainly used for image-to-image conversion tasks. It converts the input image into the target image through conditional information. This invention uses the Pix2pix model to convert the schlieren of a blank air intake duct into a schlieren image corresponding to the flow regime. To comprehensively measure the quality of the generated image and the real image, visual geometry group (VGG) loss and [other factors] are added to the loss function. loss. Loss helps preserve details and texture, while VGG loss is used as a perceptual metric in image processing and computer vision, measuring image similarity using features extracted from pre-trained neural networks. The update process is as follows:

[0042]

[0043] Where 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. This represents the probability that the discriminator determines a real image to be real, calculated as the average of all samples. This indicates that the discriminator determines the image to be fake and takes the average of all samples, where y is the real image. and L VGG (G) are all additional loss functions, where λ1 and λ2 represent the coefficients of the corresponding loss functions, respectively. * This represents the optimal solution generated by the game between the two generators.

[0044] The generator produces data as realistically as possible to fool the discriminator, while the discriminator strives to improve its ability to distinguish between generated and real images. As training progresses, the generated images gradually become more realistic, making it increasingly difficult for the discriminator to differentiate between real and generated samples. Once the generated images reach a sufficiently high level of realism, the discriminator struggles to accurately distinguish between them. At this stage, the generator has effectively learned to generate high-quality images, marking the completion of the training process.

[0045] The Pix2pix model uses images as input, eliminating input noise and greatly improving the quality of the generated images. In addition, the generator of the Pix2pix model adopts the U-Net network structure, and the discriminator adopts PatchGAN. Figure 7 The U-Net network architecture is demonstrated. In this encoder-decoder network, the input is first encoded by the encoder, and the decoder reconstructs the encoded representation to produce the output. This architecture has been widely used in tasks such as machine translation, semantic generation, and dialogue systems, achieving good results. Residual networks and U-Net are further developments based on encoder-decoder networks, both utilizing multi-scale feature fusion techniques. They combine low-level features with high-level features, thereby improving network performance and generalization ability. Figure 8 As shown, PatchGAN segments an image into small patches (e.g., 2×2 pixels) and analyzes each patch. This strategy allows PatchGAN to evaluate the realism of an image in more detail, thereby improving the quality and diversity of the generated images.

[0046] This invention introduces semantic tags into schlieren generation models. Semantic tags play a crucial role in computer vision and natural language processing, enabling models to understand and process data by labeling various parts of it. For example, in schlieren generation, semantic tags can label different objects in an image, such as air intake devices, background, and flow fields, allowing the model to identify and segment individual elements within the image. Semantic tags not only improve data interpretability and the model's understanding ability but also enhance the model's performance in specific tasks, enabling the model to make predictions by learning the relationship between tags and data, thereby improving accuracy and reliability. In image generation tasks, semantic tags can improve the diversity and detail quality of generated images, helping discriminators better determine whether generated images meet specified semantic requirements, thus improving the discriminator's discrimination ability and the overall generation effect of the model.

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

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

[0049] 1) Supersonic Inlet Flow Pattern Recognition

[0050] Algorithm 1 summarizes the process of supersonic air intake flow pattern recognition.

[0051]

[0052] 2) Supersonic Inlet Strife Generation

[0053] Algorithm 2 summarizes the process of supersonic air intake flow pattern recognition.

[0054]

[0055] Beneficial effects: (1) The supersonic inlet flow pattern recognition system adopts multi-sensor information fusion and residual network architecture, integrates the time and frequency information of pressure signals from multiple sensors, and achieves 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 flow field schlieren of the current flow state without the aid of a high-speed camera and post-processing.

[0057] (3) The EPSGN model uses semantic labels as auxiliary inputs 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 regime recognition and schlieren generation can receive real-time pressure signals and return the current flow regime, and generate flow field schlieren under the current working conditions based on the current flow regime. Attached Figure Description

[0059] Figure 1 This is the overall framework of the supersonic air intake joint detection system;

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

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

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

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

[0064] Figure 6 Residual connection diagram;

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

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

[0067] Figure 9 This is a schematic diagram of the experimental model;

[0068] Figure 10 It is the change in flow field schlieren from starting at ICR=1.54 to stopping;

[0069] Figure 11 It is the change in the sensor pressure signal from starting to stopping when ICR=1.54;

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

[0071] Figure 13 This is the comparison result of the generated schlieren image (ICR = 1.91). Detailed Implementation

[0072] The training and test sets of this invention are derived from ground tests simulating supersonic inlet conditions under real supersonic flight. This model is a simplification of the low-speed duct of a turbine combined cycle engine inlet. The overall experimental equipment is as follows: Figure 9As shown, the system mainly includes a vacuum tank, diffuser, inlet model with an optical window, dynamic pressure data acquisition system, high-speed camera, Laval nozzle, etc. Table 1 lists some important parameters of the inlet and wind tunnel system. To accurately capture the internal characteristics of the flow, a 145×29 mm optical window was installed on one side of the inlet model. All connections of the internal pipes of the model were treated with sealant 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 related vacuum pumps. A nozzle with an actual Mach number of 2.9 was used for simulation. A movable plug cone was installed at the outlet of the supersonic inlet to achieve downstream throttling, causing the flow field inside the inlet to change from start to stop. The data acquisition system is the focus of this study. The data acquisition system in this experiment consists of two parts: a pressure signal acquisition system and a flow field schlieren monitoring system. The pressure signal acquisition system mainly consists of 25 CYG-503 pressure sensors, which are installed on the fairing (C1-C13) and the external ramp (R1-R12), respectively, and acquire data 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. To achieve time synchronization between the two systems, a synchronization signal is used to trigger the high-speed camera to start recording schlieren images. By adjusting the ICR (defined as the area ratio of the pipe inlet to the throat) of the inlet model through a four-bar linkage, pressure signals and schlieren images under seven different ICR conditions were recorded.

[0073] With the help of the data acquisition system, the occurrences of failure to start were recorded. Figure 10 The image shows a schlieren image of the flow field at the inlet from start-up to shutdown at ICR=1.54. The internal flow field initially tends to stabilize, a state referred to as the start-up state. As the choke moves, the pressure disturbance generated by the throttling propagates upstream, leading to the formation of a tail shock. The movement of the tail shock causes disturbances and pressure fluctuations in the internal flow field. The supersonic inlet is considered to be in a critical state when the head of the tail shock approaches the throat. Subsequently, as the head of the tail shock passes through the throat, the inlet abruptly transitions to a shutdown state, affecting the inlet's capture characteristics. As the choke continues to move, the inlet may experience a more severe shutdown state, resulting in more overflow on the cowling side. Specifically, the inlet may experience six typical flow states as the tail shock moves upstream. The movement of the tail shock also causes fluctuations in wall pressure; several typical sensors were selected to identify the flow states, including C7, C5, C1, R11, and R2. Figure 11The pressure signals received by these sensors at ICR = 1.54 are shown, and the pressure signal variation curves confirm the changes in the flow field inside the inlet. Table 2 shows the specific classification intervals for the six flow states. It is worth noting that under conditions of high ICR, the final shock wave enters a static state very quickly after crossing the throat. That is to say, under certain conditions, the subcritical state may occur for a very short time, or may not even occur at all.

[0074] To demonstrate the effectiveness of the joint detection system and improved method for supersonic inlet flow pattern recognition and schlieren generation, this invention conducts experiments on a test set for both systems. During the testing, seven experimental conditions were divided in a 5:2 ratio: five conditions were used for training, and two conditions were used as unknown conditions to test the effectiveness and performance improvements of the joint detection system. For supersonic inlet flow pattern recognition, this invention primarily employs a weighted loss function and residual connection architecture, comparing the performance of the flow pattern recognition model after applying these optimization methods. For supersonic inlet flow field schlieren generation, this invention mainly compares the effects of the U-Net structure and semantic labels on the generative model. A generative model is trained using schlieren images from the training set, and its effectiveness and superiority are demonstrated on the test set.

[0075] The specific implementation of this invention mainly consists of the following steps:

[0076] Step 1: Conduct experiments on a 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 start-up to shutdown.

[0077] Step 2: Train the flow regime 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 result under test conditions. Compare the performance of U-Net structure and semantic labels on the generative model to prove its improvement effect.

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

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

[0081] The main parameters of the flow regime recognition model are shown in Table 3. To further improve the model's performance, a weighted cross-entropy loss function and residual connections were applied. Weighted cross-entropy assigns weights to different classes before calculating the loss function; these weights are typically determined by the number of samples in each class. This method allows the model to appropriately focus on classes with fewer samples during training, balancing the problem of sample imbalance. Table 4 shows the model's recognition accuracy on two test datasets. It can be seen that the model with both residual connections and a weighted loss function achieved an overall accuracy of 99.23%, exhibiting the best recognition performance. The results indicate that WPD-CNN, by integrating time-frequency information from multiple sensors, can accurately determine the current flow regime in a supersonic inlet, providing a reliable basis for the subsequent flow field generation.

[0082] 3. Using EPSGN to generate schlieren of the supersonic inlet flow field;

[0083] Based on six different flow field states, schlieren images were grouped, sampled, and preprocessed to form 64×256 3-channel RGB images. For comparison, this invention employed three different generative model architectures: a basic model (GAN 1), a basic model with a U-Net structure (GAN 2), and EPSGN (with U-Net and semantic labels), with all other model parameters identical. The models were trained using the training dataset, and their image generation performance was evaluated on the test dataset. Each flow state training dataset contained 250 schlieren images. It is worth noting that in the ICR=2.04 experiment, no critical state or initial instability was observed; the flow state transitioned directly from the critical state to severe instability, resulting in only 200 images in the corresponding dataset. Performance can be compared from the following aspects:

[0084] (1) Human vision (HV). Human vision refers to the evaluation of the quality of generated images using human senses, typically employing subjective rating methods. The aim is to determine whether the generated image meets human expectations. In this invention, images exhibiting significantly sharper and more detailed features are considered superior, while blurry images are considered unsuccessful, and other images are considered to have no significant difference.

[0085] (2) Peak Signal-to-Noise Ratio (PSNR). PSNR is a standard metric for measuring image quality, often used to assess the difference between compressed and original images. PSNR quantifies signal quality by comparing the maximum possible signal value with the magnitude of noise. In image processing, the signal is typically the pixel values ​​of the image, while noise is the error introduced during compression or transmission. A higher PSNR value indicates better image quality and a smaller difference between the compressed and original images. Generally, a PSNR value greater than 30 is considered good image quality, while a value below 20 indicates a significant deterioration in image quality. PSNR is a simple and effective metric, but it only considers pixel differences in the image and does not necessarily reflect perceptual differences in the human visual system.

[0086] (3) Structural Similarity (SSIM). SSIM is an indicator that measures the similarity between two images, primarily used to assess image quality. It evaluates image similarity by comparing brightness, contrast, and structural information. SSIM values ​​range from -1 to 1, where 1 indicates that the two images are identical. The SSIM formula considers three aspects of similarity: brightness, contrast, and structure. Brightness measures the average intensity of an image, usually expressed as the average value of local patches; contrast measures the range of intensity variation, usually expressed using the standard deviation; and structural comparison measures the local structural patterns of an image, usually calculated using normalized covariance. SSIM calculates the overall similarity of images by combining these three similarity measures. Compared to PSNR, SSIM focuses on capturing the perceptual characteristics of the human visual system, better reflecting human visual perception of image quality. It can assess not only overall image quality but also the quality of local regions, thus providing a more accurate quality evaluation for image processing.

[0087] (4) Perceptual Loss (PI). Unlike the human perception system, PSNR and SSIM are pixel-level evaluation metrics, which cannot accurately assess the perceptual quality of an image. Furthermore, for certain types of distortion, such as artifacts and blurring, they may not accurately assess the true quality of the image. Perceptual loss compensates for this deficiency by comparing differences in high-level features. PI typically utilizes a pre-trained convolutional neural network (VGG-19 in this invention) to extract high-level features of the image. These features capture important information such as texture, edges, and object shapes. Then, the distance between the generated and target images in these high-level feature spaces is calculated as a measure of perceptual loss. This approach allows perceptual loss to better assess the perceptual quality of an image because it considers the content and structure of the image, not just differences in pixel values.

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

[0089] The results are as follows Figure 12 , Figure 13 As shown, considering HV, EPSGN exhibits superior performance across all flow states, particularly in capturing complex flow field details, such as steady-state flow fields, terminal shock waves during upstream motion, separation of the external flow field under intense stillness, and additional spillover from the fairing side. Conversely, GAN 1 performs poorly, producing blurred and low-quality schlieren images under certain conditions, while GAN 2 performs relatively well, without noticeably low-quality images.

[0090] Since the ICR=1.91 experiment lacked some realistic schlieren images under certain conditions, the comparison experiment was mainly considered based on the evaluation metric of ICR=1.42. Table 5-7 shows the comparison results of more perceptual metrics under the experimental condition of ICR=1.42. Compared with GAN 1, GAN 2, which adopts the U-Net structure, performed better in almost all cases, highlighting the effectiveness of U-Net skip connections. Considering HV, EPSGN performed better than other models in critical states and when there was no drastic initiation, while GAN 1 seemed to generate failed images (blurred or lacking details) in some flow states. In other cases, all three models performed well. Although GAN 2 performed similarly to EPSGN in terms of PSNR and SSIM, they performed similarly in terms of pixel-level detail and overall structure, with similar visual quality and information retention. However, GAN 2's visual effects PL and LPIPS were not as good as EPSGN, indicating that the generated images were closer to human visual perception in terms of visual features (such as texture, color, structure, etc.), meaning that the generated images had better visual quality and could more realistically reproduce the perceptual characteristics of the target image. This invention primarily uses semantic tags to distinguish the regional information of the inlet internal flow field, external flow field, and inlet model. The flow field generation model can generate images with specified features by capturing inherent inter-regional correlations.

[0091] The schlieren image of the actual flow field with an ICR of 1.91 shows that the schlieren acquisition system failed to capture the subcritical state and the initial inertia; the intake abruptly transitioned from the critical state to a violent inertia. Figure 13 It is evident that although high-speed cameras cannot capture these two states, EPSGN can still generate schlieren images under corresponding experimental conditions by learning other experimental conditions. This indicates that it can be applied to experimental conditions that are difficult for schlieren acquisition systems to adapt to in order to capture flow field schlieren under extreme conditions.

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

[0093] The joint detection system for supersonic inlet flow regime recognition and schlieren generation comprises two deep learning models. For real-time detection, the system receives a pressure signal for a time window, the flow regime recognition system returns the corresponding flow regime recognition result, and the schlieren generation system receives the flow regime recognition result and generates the corresponding flow field schlieren. To test its real-time performance, the test set was used as a simulation of a real-world scenario. A pressure signal for a time window (5×1024) was transmitted to the joint detection system at 100ms intervals, and real-time inference was performed through the joint system. The total time from the input data entering the system to the generation of the output result was measured, which is a key indicator for evaluating real-time performance. The experiments were conducted on a system equipped with an NVIDIA RTX 3070 GPU. The model was trained using the PyTorch framework, with the GPU accelerating the model. For a single time window, the time for supersonic inlet flow regime recognition and schlieren generation were 24.21ms and 30.13ms, respectively, meeting the real-time requirements. This indicates that the joint detection system of this invention still has good application prospects in the real-time control of supersonic inlets.

[0094] Table 1 Geometric parameters of the experimental model

[0095]

[0096] Classification criteria for 26 flow states

[0097]

[0098]

[0099] Table 3. Main parameters of the flow regime identification model

[0100]

[0101] Table 4 shows the recognition accuracy of the fluidity recognition model on 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 Performance Indicators of EPSGN

[0108]

Claims

1. A combined flow regime identification and schlieren generation method for supersonic inlet flow, characterized in that, Comprising the following steps: Step 1: Install several sensors on the internal flow passage of the supersonic inlet to collect dynamic pressure signals, pre-process and window slide the pressure signals, and establish a sample set according to the flow state label, the sample set contains several sub-sequences, each sub-sequence 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 system outside the inlet, use a high-speed camera in the schlieren system to shoot schlieren of flow field changes from start to non-start, and establish a schlieren sample set, according to the synchronization signal of the high-speed camera and the dynamic pressure sensor, the schlieren sample set is labeled, and the schlieren sample set is divided into a second training set and a second test set; Step 3: The first training set is transmitted into a 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, and on the basis of residual network and multi-sensor information fusion architecture, the pressure information of different sensors is integrated to realize flow state recognition; Step 4: Train a flow field schlieren generative model based on Pix2pix architecture according to the second training set of different flow states, and take the semantic label of the flow field as auxiliary input to enhance the performance of the flow field schlieren generative model; The steps of training the flow field schlieren generative model include: Step 4.1: Pre-process the second training set, convert the schlieren picture into a three-dimensional matrix form, and assign semantic labels according to the flow field region in the schlieren, and classify the processed sample set according to the flow state; Step 4.2: The schlieren pictures of each flow state are respectively input into the flow field schlieren generation model, and the corresponding flow field schlieren generation model is trained, the loss function of the flow field schlieren generation model is jointly defined by the objectives of the generator and the discriminator, and the parameters of the generator and the discriminator are updated in turn, is an image generated by a condition and noise, and respectively represent the probability that the discriminator determines that the true picture is true and the probability that the discriminator determines that the false picture is false, represents the probability that the discriminator determines that the true picture is true and takes the average of all samples, represents the probability that the discriminator determines that the false picture is false and takes the average of all samples, x is information input as a condition, y is a real image, z is noise input into the network, G is a generator, and D is a discriminator, and are additional loss functions, and respectively represent the coefficients of the loss functions and , represents the optimal solution of the generator after the two games, ； Step 4.3: After the training process of the game, the generator can generate images similar to the real flow field schlieren, the schlieren generation system receives the real-time flow state recognition result, takes the blank schlieren and semantic label as the input of the generator, and obtains the flow field schlieren corresponding to the current flow state; Step 5: Test the flow state recognition model and the flow field schlieren generative model on the first test set and the second test set, the flow state recognition model receives the pressure signal to judge the current flow state and inputs the flow state recognition result into the flow field schlieren generative model corresponding to the flow state, and the flow field schlieren generative model corresponding to the flow state generates the flow field on the blank inlet according to the flow state 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, and 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 condition.

2. The combined flow regime identification and schlieren generation method of supersonic inlet according to claim 1, wherein, The dynamic pressure signal is pre-processed by using the moving average method, and the processing formula is represented as: ； wherein, denotes the size of the sliding window, denotes the i-th dynamic pressure signal collected by the sensor, denotes the moving average and denotes the pressure point after pre-processing.

3. The combined flow regime identification and schlieren generation method of supersonic inlet according to claim 1, wherein, The steps of training the flow state recognition model in step 3 include: Step 3.1: The training sample input in the first training set is decomposed into wavelet packet decomposition layers, and the dynamic pressure signal is decomposed into low-frequency information and high-frequency information according to formula (3) and formula (4) , representing the pressure signal, and determined by the selected wavelet basis, representing the low-frequency and high-frequency filter coefficients of the wavelet packet decomposition, respectively:​ ； Step 3.2: The decomposed signal is converted into a time-frequency matrix and input into the deep learning network for training. The deep learning network adopts a multi-sensor information fusion and residual connection structure. The forward training process is represented by equations (5), (6), and (7), represents the fusion information of the individual sensors, represents the network parameters, represents the extracted features of each sensor, and represent the convolution process and residual connection of the network, respectively, represents the final flow regime output. The training weight loss function is calculated according to equation (8) , represents the weight of each class c, and represent the number of samples, respectively. ； Step 3.3: After training for a maximum number of iterations, the trained flow state recognition model is used as a classifier to receive pressure signals and judge the flow state of the supersonic inlet in real time.

4. A combined flow regime identification and schlieren generation system for a supersonic inlet flow, for performing the method of any one of claims 1 to 3, characterized in that The system comprises a supersonic inlet flow state recognition system and a flow state schlieren generation system, the supersonic inlet flow state recognition system is used for receiving real-time pressure signals through pressure sensors installed on the wall surface of the inlet, realizing detailed inlet flow state detection based on multi-sensor information fusion and residual network and transmitting the results to the flow state schlieren generation system; the flow state schlieren generation system is used for receiving real-time flow state and generating schlieren images corresponding to the flow state according to the flow state, enhancing the generation effect of the flow field according to the semantic label, and realizing real-time detection of the flow state of the supersonic inlet.

Images