Real-time online warning method for unstart state of supersonic inlet based on deep learning

Abstract

The present application provides a kind of real-time online early warning method of supersonic inlet non-start state based on deep learning, combined algorithm can receive sensor historical pressure signal, automatically segmented pressure signal by WCA-PCCD, constructs training data set, trains in single channel and multichannel WPT-CNN network, obtains the classifier for real-time online early warning, receives real-time dynamic test data by classifier, realizes inlet non-start early warning.The present application proposes two innovations for real-time online supersonic inlet non-start state early warning, respectively WCA-PCCD and MC-WPT-CNN, for different stages of pattern recognition.Compared with previous time series segmentation and pattern recognition, with the help of optimization algorithm and multidimensional deep learning model, early warning task is more simple and fast, with stronger accuracy, relatively simple implementation process, suitable for different types of inlet and various working conditions.Compared with traditional early warning method, the difficulty and complexity of early warning work are greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of supersonic inertia state early warning technology, specifically relating to a real-time online early warning method for supersonic inlet inertia state based on deep learning. Background Technology

[0002] Supersonic flight technology has become a key focus in the aviation field in recent years. Supersonic combustion ramjet engine propulsion systems, due to their simple structure, light weight, and high efficiency, have been developed for future hypersonic flight. The air intake is a crucial component of the supersonic combustion ramjet engine propulsion system, used to capture and compress the incoming hypersonic airflow. To ensure high-efficiency operation, the air intake should operate in ignition mode to provide high-performance compression. However, current flight tests show that ignition failures are unavoidable in actual flight experiments, such as the CIAM / NASA supersonic combustion ramjet engine flight test in 1998, the Australian-US supersonic cooperative test in 2007, and the X-51A test in 2010. Intake ignition failure typically leads to higher aerodynamic drag, higher thermal load, and a lower total pressure recovery coefficient. Worse still, ignition failure can also cause engine shutdown and permanent mechanical damage and structural failure.

[0003] From a flow structure perspective, the inactivity of a supersonic inlet is defined as the ejection of a shock train system from the inlet and isolator; from an inlet performance perspective, inactivity is defined as a change in the flow trapping characteristics at the inlet inlet caused by internal flow. Many factors can lead to an inlet entering an inactive state, such as free-flow disturbances, shock train-boundary layer interactions, inappropriate ICR, and increased downstream back pressure. To date, the flow characteristics of the inactivity process have been well studied, and many flow visualization techniques have been used to investigate the mechanisms of supersonic inlet inactivity, such as mathematical simulation, particle image velocimetry, schlieren photography, and dynamic pressure measurement. However, research on real-time online inactivity prediction for supersonic inlets is still immature and is often limited by practical operating conditions. Trapier et al. applied cumulative sums and generalized likelihood ratios to detect inactivity at the inlet; Chang et al. used recursive Fourier transforms to achieve real-time inactivity prediction and detection. These methods largely meet the requirements of real-time inactivity prediction, but their parameter selection is challenging, requiring a high level of expertise and extensive manual analysis. Wang et al. employed a combination of time series segmentation algorithms and deep learning for real-time online early warning of inactive states. This combined model can receive pressure signals from a single sensor and train a corresponding deep learning classifier for early warning, demonstrating the effectiveness and superiority of deep learning. This invention improves this model in two aspects: time series segmentation and deep learning modeling.

[0004] Considering the correlation between shock trains and pressure fluctuations, collecting sufficient information from sensors is crucial for supersonic inlet non-starting early warning systems, which essentially detect the oscillating characteristics of pressure signals. Through research on the non-starting mechanism, it can be seen that a precursor phenomenon exists before the overall inlet remains stationary: shock trains form leading-edge separation bubbles, which move upstream and cause pressure fluctuations at the corresponding locations. This invention transforms this into pattern recognition, using deep learning techniques to detect pressure fluctuations. The biggest advantage of deep learning compared to traditional machine learning is that it can achieve feature extraction without requiring extensive manual analysis.

[0005] In practical applications, pressure sensors measure pressure signals at corresponding locations and transmit them to a central computer. A trained pattern classifier then determines whether pressure fluctuations occurred within the current time window. Pattern recognition technology is the process of extracting important features from large amounts of noisy data to classify input data. A complete pattern recognition system consists of three interrelated processes: dataset construction, feature extraction, and pattern classification. For dataset construction, the pressure signals obtained from the experimental supersonic inlet model lack corresponding labels. This necessitates a reasonable and effective method to segment time series data to construct a dataset for pattern recognition. WCA-PCCD significantly accelerates the segmentation of long time series using PCCD by embedding the WCA optimization algorithm into the PCCD time series segmentation algorithm, enabling rapid and easy dataset construction. For feature extraction, this invention first uses WPT to decompose a time series into a time-frequency matrix to extract features in both the time and frequency domains. Then, it employs two training schemes: single-channel and multi-channel. The single-channel model can receive historical pressure signals from any single sensor, thereby training a corresponding sensor recognition model. This model can receive pressure signals within a real-time time window to determine whether fluctuations occur in the current time window. By training a single-channel model with multiple sensors, the position of the shock train inside the air intake can be determined. The multi-channel model simultaneously receives signals from multiple sensors and trains a corresponding multi-classification model. This model determines the flow of the shock train by receiving multi-dimensional pressure signals, achieving higher accuracy compared to the single-channel model. The WCA-PCCD and multi-channel early warning model will be described in detail below. Summary of the Invention

[0006] To address the aforementioned problems, this invention implements real-time online early warning of non-starting status for supersonic inlets based on time series segmentation and deep learning, and proposes an improved scheme. The non-starting status early warning algorithm based on time series segmentation and deep learning mainly consists of two parts: an improved time series segmentation algorithm and a multi-channel deep learning pattern recognition algorithm. The former is used to construct the dataset, and the latter is used to train the corresponding classifier. This invention applies WCA-PCCD and SC / MC-WPT-CNN models respectively. This combined algorithm can automatically generate a training dataset based on historical sensor data, thereby training the corresponding single-channel / multi-channel WPT-CNN model. It is applicable to different operating conditions and different types of supersonic inlets. The entire process almost eliminates the need for manual time-domain and frequency-domain analysis for specific operating conditions, significantly reducing the professional knowledge requirements for early warning tasks and minimizing analytical errors that may result from manual analysis. It shows excellent prospects for practical engineering applications.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention proposes a combined WCA-PCCD and SC / MC-WPT-CNN algorithm for real-time online supersonic inlet non-start flow state early warning. The algorithm first employs penalized contrast change-point detection (PCCD) and the water cycle algorithm (WCA) to propose WCA-PCCD. This combined algorithm can quickly and accurately locate pressure signal fluctuations without manual analysis to set thresholds and discrimination criteria, generating a dataset for subsequent recognition model training. The dataset generated by PCCD is then input into a single-channel (SC) or multi-channel (MC) WPT-CNN model for network training, resulting in a classifier for the corresponding sensor. This classifier can determine whether the pressure signal at the sensor fluctuates, and thus determine whether an early warning signal needs to be issued.

[0009] First, let's introduce the framework and early warning process implemented by PCCD, WCA, and SC / MC-WPT-CNN. 1. Preliminary knowledge.

[0010] 1) PCCD Algorithm

[0011] Pattern recognition is defined as extracting important features from large amounts of noisy data and classifying the input data. It is widely used in decision-making systems in machine intelligence systems. A complete pattern recognition system consists of three interrelated processes: dataset construction, feature extraction, and pattern classification. For dataset construction, the pressure signals obtained by dynamic sensors inside supersonic air intakes lack corresponding labels. This necessitates a reasonable and effective time series segmentation method to construct a dataset for pattern recognition. Detecting abrupt changes in certain physical features is one of the important practical problems in signal processing (speech processing, geophysics, medicine, etc.). The essence of pressure signals is a time series, and the detection of change points can be seen as time series segmentation. The purpose of time series segmentation is to divide the time series into multiple intervals so that each interval has similar properties. Currently, commonly used one-dimensional time series segmentation methods include the sliding window algorithm (SWA), CUSUM method, and PCCD method. This invention chooses the PCCD algorithm because it can segment the time series without setting additional parameters. The PCCD algorithm calculation process is as follows:

[0012] Assuming to use Define a time series of length N, τ i The PCCD detects the specific location of the change point by minimizing the penalty function H(τ) to represent the change point of the time series. The penalty function of the time series is calculated according to Equation (1), where J(τ,x) represents the comparison function and β represents the penalty parameter.

[0013]

[0014] If the change affects both the mean and variance of the time series, the contrast function J(τ,x) based on Gaussian log-likelihood is shown in equations (2)(3)(4).

[0015]

[0016]

[0017]

[0018] It is a function that changes with the number of segmentation points. Considering that the pressure signal only needs to be segmented once, A constant can be used instead, which can be ignored in the problem of finding the minimum value of H(τ). Therefore, the penalty function H(τ) can be expressed by equation (5).

[0019]

[0020] By traversal Find the point corresponding to the minimum penalty function H(τ) for all points, and this point is the point of division.

[0021] Algorithm 1 summarizes the process of PCCD change point detection.

[0022]

[0023]

[0024] 2) Water cycle algorithm

[0025] Based on the PCCD algorithm, the entire time series can be directly input to obtain the optimal segmentation point, but for long-term series, this process can be very time-consuming. The pressure sensor in this paper operates at a frequency of 20,000 Hz, and a complete experimental run of the supersonic air intake takes approximately 10 seconds or more. This means that each pressure signal contains hundreds of thousands of measurement points. Directly applying a global traversal search method to find the points of change in the pressure signal could require significant time and memory resources. To find the segmentation point of the pressure signal quickly and cost-effectively, the algorithm itself needs to be optimized.

[0026] Inspired by the regularity and randomness of natural phenomena, metaheuristic algorithms have been developed over the past few decades to solve engineering optimization problems, such as genetic algorithms (GA), particle swarm optimization (PSO), and simulated annealing (SA). The idea for water cycle optimization (WCA), proposed by Eskandar et al., is based on the real-world water cycle. WCA simulates the natural processes of rainfall, flow, and evaporation of water, such as… Figure 1 As shown. Similar to other metaheuristic algorithms, WCA starts with an initial population called raindrops. Based on the fitness of each raindrop, the population is divided into three ranks: ocean, river, and stream. The best individuals are selected as oceans, some raindrops that are not as good as the best individuals are selected as rivers and flow into the ocean. The remaining raindrops are selected as streams and flow into rivers and the ocean. WCA will continuously update the oceans in iterations until the maximum number of iterations is reached. The detailed process of the WCA algorithm is as follows:

[0027] The initial population of WCA is usually generated by a random function. In this invention, to further accelerate the search speed, a uniform distribution function is used, as shown in formula (6), where N pop Let be the population size, and LB and UB be the upper and lower bounds of the given problem, respectively.

[0028]

[0029] The value of an individual in the population is determined by the fitness function. In this invention, the fitness function is the penalized contrast H(τ) of formula (5). PCCD is used to find the minimum value of H(τ). Therefore, the smaller H(τ) is, the better the individual is. The population is sorted from the minimum to the maximum according to H(τ), and N individuals are selected from the best individuals. sr N sr Given the total number of oceans and rivers, the remaining individuals in the population are considered streams. During the iterative update process, rivers will flow into the ocean, and streams will flow into both rivers and the ocean, thus continuously changing their positions and updating the fitness function. To allocate streams to rivers and oceans based on flow intensity, the allocation is performed according to formula (7), NS n N represents the number of streams corresponding to each ocean and river, and Cost represents the fitness function corresponding to each ocean and river. stream Represents the number of streams;

[0030]

[0031] The flow process is an iterative process towards the optimal solution; some streams flow into rivers, while others flow directly into the sea, and all rivers and streams eventually flow into the sea. For example... Figure 2 As shown, the process will change the initial position and update the fitness value of an individual. Taking a stream flowing into a river as an example, if the updated fitness value of the stream exceeds that of the river, the river and stream will be swapped, as shown. Figure 3 As shown. This exchange will also occur in other flow processes. During the iteration and update process, the ocean will continuously approach the global optimum, and the process of updating the positions of streams and rivers is shown in (8)(9)(10). rand(0,1) is a random number between 0 and 1, and C is a fixed constant, which is generally set to 2;

[0032]

[0033]

[0034]

[0035] To avoid premature convergence to a local optimum, WCA uses evaporation and rainfall. Evaporation means that water will evaporate from the stream, while rainfall means that the evaporated water will return to Earth as rainwater. The condition for evaporation is determining whether the two are sufficiently close; for example, for a stream flowing into the ocean, if the absolute distance between the stream and the ocean is less than d... max Then the stream will evaporate. max The value of d determines the balance between flow and rainfall. max The larger the value of d, the fewer the number of iterations; conversely, the smaller the value, the fewer the number of rainfalls. Typically, d... maxAdjust according to equation (11), where max_it represents the maximum number of iterations;

[0036]

[0037] After the evaporation process, rainfall is used to search for the optimal solution. In order to improve the convergence speed and computational performance of WCA for constrained problems, the principle of evaporation and rainfall is as follows: If the distance between the river and the ocean satisfies the evaporation condition, all streams flowing into the river will evaporate and be regenerated according to Equation (12); if the distance between the stream flowing into the ocean and the sea satisfies the evaporation condition, the stream will evaporate and be regenerated near the ocean according to Equation (13). If the regenerated individual is better than the individual flowing in the direction, the positions of the two will be exchanged. N(0,1) represents a normally distributed random number with a mean of 0 and a variance of 1.

[0038]

[0039]

[0040] As is well known, global and local exploration capabilities are crucial in the search process of intelligent algorithms. Although WCA uses an evaporation mechanism to prevent getting trapped in local optima, the convergence of the distance-based evaporation mechanism may be too slow, resulting in poor population diversity. To improve this shortcoming, this paper combines evaporation and mutation (similar to genetic mutation in GA). Each time the evaporation condition is determined, mutation may occur to update the position. The update method is the same as Equation (12) and Equation (13), and the mutation condition is Equation (14).

[0041] rand(0,1)<0.1 (14)

[0042] The PCCD algorithm aims to find the optimal segmentation point of a time series to minimize the penalty function H(τ). By using the penalty function H(τ) as the optimization objective of WCA (WaaS), finding the minimum value of H(τ) enables the WCA-PCCD algorithm to quickly and accurately segment the time series. The flowchart is shown below. Figure 4 As shown, Algorithm 2 summarizes the process of WCA-PCCD change point detection.

[0043]

[0044] 3) WPT-CNN recognition model

[0045] After constructing the dataset, feature extraction and pattern classification are required. Machine learning is one of the most commonly used frameworks, with many successful applications in the aerospace field. Deep learning originated from machine learning and has received increasing attention in recent years. Unlike traditional machine learning feature extraction, which relies on manual analysis, deep learning automatically performs hierarchical feature extraction by constructing layers with many intermediate hidden layers to simulate the human brain. This significantly reduces manual intervention in feature engineering and minimizes errors and difficulties that may arise from manual analysis. Furthermore, deep learning allows for multi-dimensional input and simultaneous multi-dimensional feature extraction, which is beneficial for implementing multi-sensor early warning systems for supersonic air intakes. The WPT-CNN recognition model will be introduced below.

[0046] In the problem of early warning for inlet duct inactivity, the pressure signal is easily affected by noise under operating conditions; therefore, both time-domain and frequency-domain analysis are essential. In the time domain, this invention employs the moving average method, which is widely used for processing time series data due to its simplicity, objectivity, and reliability. It is based on the following formula:

[0047]

[0048]

[0049] pre_x(i)=x(i)-Rav(i) (17)

[0050] In the frequency domain, this invention employs the Wavelet Transform (WPT) method to convert a time series into a time-frequency matrix, thereby obtaining the joint time-domain features of different frequency bands. The discrete wavelet transform process is as follows: First, the original signal is used as the input signal, and through a set of orthogonal wavelets, the signal is divided into high-frequency and low-frequency components. Then, the low-frequency information is used as the input for the next layer, decomposed into high-frequency and low-frequency components, and the above steps are repeated. The discrete wavelet transform operation for each layer can be expressed as:

[0051] A = f * W l (18)

[0052] D = f * W h (19)

[0053] Where A represents the low-frequency information after decomposition, D represents the high-frequency information, * represents the convolution operator, and W... l and W h It is a set of orthogonal wavelets, W l W represents a low-pass filter. h This represents a high-pass filter, where W is selected after a wavelet type has been chosen. l and W h It consists of defined wavelet filter coefficients.

[0054] Figure 5(a) describes the process of Discrete Wavelet Transform (DWT). It can be seen that DWT discards high-frequency information during signal processing, only decomposing at the low-frequency resolution level, which may lead to feature loss. To address this issue, Wavelet Packet Transform (WPT) was proposed. This transform uses the same signal processing method as DWT, but unlike DWT, it decomposes both low-frequency and high-frequency components simultaneously, converting the signal into a multi-dimensional tensor, making it more suitable for neural networks. A schematic diagram is shown below. Figure 5 As shown in (b).

[0055] This invention sets the wavelet decomposition level to 3. Assuming a signal segment with a time window of 1024, the wavelet decomposition (WPT) results in 8 one-dimensional tensors of length 128. These tensors are then arranged according to the Gray code standard to form a time-frequency matrix of the corresponding dimension, facilitating subsequent training of the deep learning network. This process is as follows: Figure 6 As shown. If converted from natural binary code to an analog signal, in some cases, such as converting from decimal 3 to 4, each bit of the binary code changes, causing large spike current pulses in the digital circuit. Assume the natural binary code is A. n-1 A n-2 …A2A1A0, its corresponding Gray code is B n-1 B n-2 …B2B1B0, the formula for converting natural binary code to Gray code is as follows:

[0056] A n-1 =B n-1 (20)

[0057]

[0058]

[0059] Assuming the binary codes of the time-domain information of the eight frequency bands are set to 000, 001, 010, 011, 100, 101, 110, and 111 respectively, their Gray code codes are converted to 000, 001, 011, 010, 101, 110, 100, and 011 respectively.

[0060] After processing the time and frequency domain information, a suitable framework needs to be selected for feature extraction. CNN, as a commonly used deep learning framework applied to pattern recognition and fault diagnosis, has advantages such as sparse connections and shared weights. In particular, the ability to directly input multi-dimensional input vectors into the network avoids the complexity of data reconstruction during feature extraction and classification, and has been widely used in pattern recognition. CNN can also reduce the number of trainable parameters in the neural network, enabling deeper network structures. Deeper network structures allow for more abstract learning and training, further improving the learning and training efficiency of the neural network. The main function of CNN is provided by three types of structural layers constructed in the hidden layers: convolutional layers, pooling layers, and fully connected layers.

[0061] The function of convolutional layers is to extract features. By performing feature extraction through convolution operations on local receptive fields, the number of network parameters and model complexity can be greatly reduced. Its main parameters include the kernel size, stride, and padding steps. Given an input image X... i and the corresponding one-hot encoded class label Y i The output of the convolutional layer is calculated as follows:

[0062]

[0063] In the formula, * represents the convolution operation, W conv and b conv σ and · represent the weights and biases of the convolutional layer, respectively, and σ(·) represents the activation function. The activation function selected in this invention is the ReLU activation function.

[0064] Each convolutional layer is typically followed by an activation function to increase the network's non-linearity. This invention uses the ReLU activation function, one of the most commonly used activation functions in deep learning models, which can greatly alleviate the gradient vanishing problem. Its calculation formula is as follows:

[0065]

[0066] Pooling layers downsample the output of convolutional layers by calculating the average or maximum value of the sliding sub-region. This invention employs max pooling layers to retain the most critical features while reducing computational cost. The output of the pooling layer is calculated as follows:

[0067]

[0068] In the formula, the down_sample(·) operation can be either an average pooling operation or a max pooling operation. This invention uses max pooling.

[0069] Like traditional feedforward neural networks, in a fully connected layer, each input neuron is connected to every output neuron. If the input to a fully connected layer is represented as f... i The output is calculated as follows:

[0070]

[0071] In the formula, W fully and b fully These represent the weights and biases of the fully connected layer, respectively.

[0072] The output of the output layer is similar to that of equation (23), except that it does not have an activation function, and is calculated as follows:

[0073]

[0074] In the formula, W out and b out These represent the weights and biases of the output layer, respectively.

[0075] The softmax function in Equation (28) is usually used to calculate the probability of the predicted class, and the cross-entropy L in Equation (29) is used as the loss function for the classification problem. The network can be updated based on the gradient according to the loss function.

[0076]

[0077]

[0078] In the formula, z ij Indicate z i The j-th part, Y ij This represents the probability of being predicted as class j, where N represents the number of images and M represents the number of classes.

[0079] 2. Implementation framework for early warning of non-starting state of supersonic inlet

[0080] After introducing the methods used for pattern recognition, the above methods can be combined to realize the early warning of the supersonic air intake not starting state. Next, the single-channel and multi-channel early warning models will be introduced respectively.

[0081] 1) SC-WPT-CNN

[0082] A schematic diagram of the SC-WPT-CNN algorithm is shown below. Figure 7 As shown, the flowchart is as follows Figure 8As shown, the algorithm can receive historical signals from a single sensor, automatically generate a dataset based on the results of the WCA-PCCD segmentation algorithm, and train a corresponding real-time classifier. In practical applications, the real-time pressure signal window can be passed to the classifier to determine whether the pressure signal is fluctuating. If pressure fluctuation occurs, it indicates that the leading edge of the shock wave has interfered with the air intake position corresponding to the sensor, and a warning signal needs to be issued.

[0083] Algorithm 3 summarizes the process of SC-WPT-CNN.

[0084]

[0085] 2)MC-WPT-CN

[0086] By creating multiple SCWPT-CNN models, the pressure signals from corresponding sensors can be identified, thereby determining the location of the shock wave train inside the air intake. However, the SC-WPT-CNN model only targets a single sensor during the identification process, wasting information hidden in other sensors. One advantage of deep learning is that it allows for multi-dimensional input and simultaneous multi-dimensional feature extraction. Based on a CNN deep learning network, this invention employs a multi-dimensional inactive state early warning model. Taking a two-dimensional identification model as an example, its flowchart is as follows... Figure 9 As shown, MC-WPT-CNN transforms the original binary classification problem into a multi-class classification problem. Each state corresponds to a shock wave moving to the corresponding position on the sensor, causing the pressure signal to fluctuate. The biggest difference between MC-WPT-CNN and SC-WPT-CNN lies in the different dimensions of the received pressure signal, requiring adjustment of the model dimensions; other parameters do not need to be changed.

[0087] Algorithm 4 summarizes the process of MC-WPT-CNN.

[0088]

[0089]

[0090] Beneficial effects: (1) WCA-PCCD achieves fast and accurate segmentation on long-term sequences, and compared with threshold-based segmentation algorithms, it does not require parameter adjustment according to specific working conditions, making it simpler and more convenient;

[0091] (2) SC-WPT-CNN can directly receive pressure signals from a single sensor and automatically train a multi-dimensional recognition model. By establishing multiple SC-WPT-CNN models, the distribution of dynamic pressure sensors can be used to understand the shock wave movement in detail, providing a reliable basis for early warning of non-starting status.

[0092] (3) MC-WPT-CNN can directly receive the raw pressure signals from multiple sensors as input and automatically train a multi-dimensional recognition model. Compared with the single-channel model, the multi-dimensional information improves the accuracy of early warning task recognition. Attached Figure Description

[0093] Figure 1 This is a schematic diagram of the WCA simulation of the natural water cycle;

[0094] Figure 2 This is a schematic diagram of WCA flow;

[0095] Figure 3 This is a WCA individual location exchange diagram;

[0096] Figure 4 This is the WCA-PCCD flowchart;

[0097] Figure 5 These are schematic diagrams of two wavelet transforms;

[0098] Figure 6 This is a schematic diagram of the time-frequency matrix generated after the pressure signal is decomposed by WPT.

[0099] Figure 7 This is a schematic diagram of the basic framework of SC-WPT-CNN;

[0100] Figure 8 This is the flowchart for SC-WPT-CNN;

[0101] Figure 9 This is a schematic diagram of the basic framework of MC-WPT-CNN;

[0102] Figure 10 This is a simplified diagram of the experimental model;

[0103] Figure 11 It is a schlieren image of the flow field changes inside the air intake;

[0104] Figure 12 This is a schematic diagram of the original pressure signals of C1, C13, and R11;

[0105] Figure 13 This is a comparison chart of pressure signals under the original ICR2.04 and the ICR2.04 with noise;

[0106] Figure 14 This is a flowchart for real-time online inaction prediction. Detailed Implementation

[0107] The dataset used for testing in this invention comes from a simplified rectangular model of the low-speed duct of a turbocharged combined cycle engine inlet. The main components of this model include a blunt-nosed fairing, an isentropic slope, a universal diffuser, sidewalls with optical windows, and a four-bar linkage, such as... Figure 10 As shown in Table 1, this test investigated the non-starting flow characteristics of eight different internal contraction ratios (ICR), namely 1.21, 1.30, 1.42, 1.54, 1.67, 1.79, 1.91, and 2.04. Detailed model parameters are shown in Table 1. To capture the internal flow characteristics, an optical window made of K9 glass (145×29mm) was installed on one side of the inlet model. All connections within the model's internal ducts were sealed with sealant and sealing strips to reduce noise interference. After taking these necessary measures, the model was installed and tested in a wind tunnel. The wind tunnel simulation equipment mainly consisted of a variable rectangular Laval nozzle, a fully enclosed test chamber, and a 400m... 3 The system consists of a vacuum chamber and a related vacuum pump, using a nozzle with an actual Mach number of 2.9. The data acquisition system for this experiment comprises two parts: a dynamic pressure acquisition system and a high-speed schlieren observation system. In the dynamic pressure acquisition system, 25 Twinbridge CYG-503 dynamic pressure sensors are installed on the inner surface of the experimental model. C1-C13 are located on the upper wall fairing, and R1-R12 are located on the lower wall outer ramp. Figure 10 As shown, the dynamic pressure signal with a sampling rate of 20kHz was acquired by a matched data acquisition card (DAQ PCI-6255). It should be noted that sensors C6, R7, and R12 were in a failed state before testing and were excluded from subsequent work. In the high-speed schlieren observation system, schlieren images were acquired in a horizontal position by a Nikon MEMERCAMHX-3 high-speed camera equipped with a 400mm f / 5.6 lens, a sampling rate of 20,000 frames per second, a resolution of 1152×336, and a shutter speed of 2 microseconds. An external trigger signal was also set to synchronize the dynamic pressure acquisition system with the high-speed schlieren observation system. The inactive state was defined as the flow structure inside the inlet affecting the flow capture characteristics at the inlet. Based on the schlieren images and pressure signals, we can obtain the moments when the supersonic inlet enters the inactive state under eight conditions, as shown in Table 2. The specific criterion is that a clear pressure pulsation signal is detected at R11, and the shock train is discharged from the inlet of the inlet.

[0108] To demonstrate the effectiveness of the WCA-PCCD segmentation algorithm and the MC-WPT-CNN model, this invention employs two comparative schemes. For the WCA-PCCD algorithm, CUSUM, the sliding window algorithm (SWA), and PCCD are used to segment the pressure signal. A dataset is generated based on the segmentation results, thereby training different SC-WPT-CNN recognition models. These models are then applied to the validation and test sets, respectively. For the MC-WPT-CNN model, this invention selects C4 and C10 to train two separate SC-WPT-CNN models and a common MC-WPT-CNN model, comparing their performance on the test set.

[0109] In the comparative experiment, the eight experimental conditions were divided into a 5:1:2 ratio: five conditions were used for training, one for validation, and two for testing. ICR1.54 was fixed as the validation set, and its sample division was precisely done manually to compare the performance of different recognition models. Four schemes were selected for the training and test sets according to Table 3. The training set was used to train the recognition model, the validation set was used to adjust the network structure and verify the classifier accuracy, and the test set served as a simulation of real-world conditions to determine the timing of warning signals. The time window size for each sample was set to 1024 (approximately 50ms), and the time window sliding distance for the test set was set to 300 (15ms), indicating that the total time for a single recognition should not exceed 15ms.

[0110] The specific implementation of this invention mainly consists of the following steps:

[0111] Step 1: Collect 8 sets of experimental data of the rectangular hybrid compression supersonic inlet model, and use pressure signals and schlieren images to verify the correlation between the non-starting state of the supersonic inlet and pressure fluctuations.

[0112] Step 2: Using SC-WPT-CNN, compare the four segmentation algorithms according to Table 3. The segmentation algorithms are used to generate training datasets. Train recognition models for each scheme for C1, C4, C7 and C10 respectively. Apply each recognition model to the validation set and test set. The higher the accuracy on the validation set and the earlier the warning signal is issued on the test set, the better the recognition model is, that is, the better the corresponding time series segmentation algorithm is.

[0113] Step 3: Select C4 and C10, train two separate SC-WPT-CNN models and a common MC-WPT-CNN model respectively, and compare their performance on the test set.

[0114] 1. Verify the correlation between the inactive state and pressure fluctuations;

[0115] The flow in a non-sonic inlet was studied by acquiring pressure signals and schlieren images using a data acquisition system based on an experimental model. Figure 11 The image shows a schlieren image depicting the flow field changes inside a supersonic inlet from the start-up state to the stop-up state. Initially, the flow field inside the supersonic inlet is relatively stable. Then, the increasing back pressure impedes the inlet flow, forming a stop-up shock train system. The movement of the shock train causes separation shock wave oscillations and wall pressure fluctuations. To suppress the interaction between the shock train and the boundary layer, the shock train needs to be pushed out of the inlet. The stop-up state occurs with the movement of the shock train system. The inlet stop-up truly occurs when the shock train system is expelled from the inlet and isolator. C1, C13, and R11 are examples that represent typical characteristics of three regions of the test model. Figure 12 The raw pressure signals from C1, C13, and R11 are displayed. These pressure signals are essentially time-series data, recording the gradual changes in the internal flow pattern. Sensor C13, located furthest downstream, is the first to detect fluctuations, while the pressure signals from C1 and R11 are still in a stable flow state. Pressure fluctuations are only detected by C1 and R11 when the shock train moves to the inlet, and the pressure signal from C1 fluctuates earlier than that from R11, a phenomenon consistent with changes in the internal flow field of the supersonic inlet. Based on schlieren images and dynamic pressure signals, a strong correlation between the shock train and pressure fluctuations is confirmed. The movement of the shock train from downstream to upstream can be considered a precursor to inactivity; when it exits the inlet, inactivity actually occurs. This means that we can detect pressure fluctuations caused by the shock train using pressure sensors to provide an early warning of inactivity.

[0116] 2. Compare four segmentation algorithms using SC-WPT-CNN;

[0117] First, the four segmentation algorithms were applied to C4. Since ICR1.54, used as the validation set, did not need to participate in segmentation and training, ICR2.04 with noise was used instead. Figure 13Table 4 shows the segmentation results for the original ICR2.04 and the noisy ICR2.04 under eight conditions. It can be seen that WCA-PCCD achieves the same results as PCCD, but takes significantly less time than the PCCD traversal search method, demonstrating the effectiveness of the WCA optimization algorithm. Furthermore, SWA is affected by noise in the noisy ICR2.04 because SWA is a local segmentation method, judging based on whether the selected criteria within the current time window exceed an acceptable threshold. This leads to significant noise interference within the current window and poor anti-interference capability. Since WCA-PCCD and PCCD segmentation results are identical, they are grouped together in the following tests. In addition, to prevent inaccurate single-shot identification, this model retains the most recent five identification results, issuing a warning signal when the number of pressure fluctuations detected exceeds half.

[0118] Figure 7 The framework diagram of SC-WPT-CNN is shown. The WPT-CNN model parameters are shown in Table 5. Based on the four schemes in Table 3, three segmentation algorithms—CUSUM, SWA, and WCA-PCCD—are used. The trained recognition models are applied to the validation and test sets, respectively, yielding results in Tables 6 and 7. The single recognition time is approximately 5.1ms, meeting the real-time requirements. The four segmentation algorithms can be summarized in the following four aspects:

[0119] (1) Dependence on manual analysis. CUSUM and SWA require a certain amount of pre-analysis to select appropriate thresholds and other parameters based on specific operating conditions. For SWA, manual analysis can be very complex because the selection of evaluation metrics needs to be combined with the characteristics of the signal. For PCCD and WCA-PCCD, the parameters do not need to change with the operating conditions given a fixed number of variable points, because the selection of variable points is only related to the minimum value of the penalty function.

[0120] (2) Identification model performance. On the validation set, WCA-PCCD performed better, with an average accuracy higher than CUSUM and SWA; on the test set, WCA-PCCD issued warning signals earlier than CUSUM and SWA in most cases, which also indicates that the model it trained performed better.

[0121] (3) Time Consumption. It is easy to see from Table 4 that WCA-PCCD requires the least time, while PCCD takes hundreds of times longer than WCA-PCCD, which proves the effectiveness and superiority of WCA. In addition, CUSUM also takes a relatively long time. SWA takes a relatively short time because the selected indicator RMS has a faster calculation speed. However, for some indicators with complex calculations, its time consumption may also be relatively long.

[0122] (4) Anti-interference ability. Due to the inherent defects of the local algorithm, SWA has poor anti-interference ability and is easily interfered with by noise in the current time window. Other algorithms are all global algorithms and have relatively strong anti-interference ability.

[0123] The comparison results of the four time series segmentation algorithms are shown in Table 8. The proposed WCA-PCCD performs better in all aspects. In real-time online non-start warning, WCA-PCCD can quickly, accurately, and simply cut the pressure signal. Due to different sensor locations, the time of issuing the non-start warning also varies. In practical applications, the specific selection of the sensor should be based on the actual time required by the actual system. 3. Real-time online non-start warning for supersonic air intakes is achieved using MC-WPT-CNN;

[0124] For comparison, C4 and C10 sensors were selected, and two approaches were adopted. One approach was to build SC-WPT-CNN models for C4 and C10 separately to identify the pressure fluctuations corresponding to each sensor individually. The other approach was to build MC-WPT-CNN models for C4 and C10 simultaneously. This model received the pressure signals from both sensors and determined that the shock train had caused pressure fluctuations at the corresponding locations of each sensor. Except for necessary dimensional adjustments, the other parameters were the same as the SC-WPT-CNN model. Figure 9 As shown, mode 0 indicates no pressure fluctuation, while modes 1 and 2 indicate that the shock train has caused fluctuations in the corresponding sensor pressure signals. Both schemes were tested on the test set, and the results are shown in Table 9. It can be seen that in almost all cases, the MC-WPT-CNN model can identify pressure fluctuations earlier than two separate SC-WPT-CNN models, indicating that the multi-channel model has higher accuracy and better performance. In terms of time consumption, the total time for the two single-channel models is approximately 10.2 ms, while the MC-WPT-CNN model's single-recognition time is approximately 9.1 ms, indicating that establishing a common multi-channel model is more time-efficient than establishing multiple single-channel models. However, a significant issue with the MC-WPT-CNN model is that it must simultaneously receive pressure signals from multiple sensors to achieve recognition. In other words, if one sensor fails, the entire model will fail. For SC-WPT-CNN, if a sensor fails, it only affects the corresponding recognition model; other models continue to function normally. Considering the potential damage to sensors under operating conditions, the flowchart for real-time online non-start prediction can be shown below. Figure 14 As shown.

[0125] Table 1 Geometric parameters of the experimental model

[0126]

[0127] Table 2. Start-up times for 8 sets of ICR conditions

[0128]

[0129] Table 3. Four different training schemes

[0130]

[0131] Table 4. Segmentation results of the C4 sensor under different ICRs (unit: s)

[0132]

[0133] Table 5 WPT-CNN Parameter Settings

[0134]

[0135] Table 6. Accuracy of different schemes on the validation set.

[0136]

[0137] Table 7. Early warning lead time for different schemes on the test set (unit: ms)

[0138]

[0139] Table 8. Comprehensive comparison of the four segmentation algorithms

[0140]

[0141] Table 9 Comparison of performance of different models in the test.

[0142]

Claims

1. A real-time online early warning method for the inactive state of a supersonic air intake based on deep learning, characterized in that, Includes the following steps: Step 1: Collect several sets of historical dynamic pressure signals from several sensors arranged on the internal flow channel of the supersonic inlet, preprocess the pressure signals, establish a sample set for each sensor, and divide the sample set into training set, validation set and test set. Step 2: The training set is divided into segments using a time series segmentation algorithm and the segmentation algorithm is optimized using the WCA optimization algorithm. The sudden change time of the pressure signal in the training set is used as the segmentation point. Based on the segmentation point and the set time window parameters, a training set with a time window label containing flow state is generated. The flow state label of the time window before the segmentation point is a stable flow state, and the flow state label of the pressure signal after the segmentation point is an unstable flow state. Step 3: After processing the training set containing the flow state label, feed it into the WPT-CNN deep learning network for training. The trained model is used as a classifier, and the classification accuracy is verified on the validation set. Step 4: The classifier determines the current flow state at the sensor location by receiving data in real time, uses the test set as a simulation of the actual working conditions, and transmits signals in real time for early warning testing; The time series segmentation algorithm is the PCCD time series segmentation algorithm, and the specific steps include: using... Define a time series consisting of the training set of length n, and calculate the penalty function of the time series according to the following formula. ； Representing the turning points of a time series; through traversal Find the minimum penalty function for all points. The corresponding point is the dividing point; The process of optimizing the segmentation algorithm using the WCA optimization algorithm includes: Step 2.1: Generate the initial population according to formula (4). and These are the upper and lower bounds of the time series, respectively. The upper bound is the total length of the time series, and the lower bound is the zero point. Population size; ； Step 2.2: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require As the evaluation function corresponding to each individual in the population, each individual is a split point; the evaluation function corresponding to each individual in the population Sorted from smallest to largest, the current minimum value is for oceans, followed by rivers, with the number of rivers being [number missing]. The remaining individuals are streams. ; ； The total number of oceans and rivers, Population size; Step 2.3: Allocate the stream to the river and ocean based on the flow intensity (10). Indicates the number of streams flowing into a specific river or ocean. The numerical value of the evaluation function for a specific river. Indicates the number of streams; ； Step 2.4: The process of updating the positions of streams and rivers is shown in (11)(12)(13). After the position is updated, the evaluation function is updated according to (4). Represents a random number uniformly distributed between 0 and 1. The constant between 1 and 2 is used to swap the identities of the stream and ocean if the updated evaluation function of the next generation is better than that of the current flowing river or ocean. ； Step 2.5: After the population is iterated and updated, if a river flowing to the ocean satisfies (14), then all streams corresponding to that river will evaporate and rainfall will resume according to (15). If a stream flowing to the ocean satisfies (14), then that stream will evaporate and rainfall will resume according to (16). It will be updated according to (17) with the number of iterations, and after the update, the position exchange judgment will be performed on the individual; ； Step 2.6: To prevent the distance-based evaporation mechanism from converging too slowly, a mutation mechanism is introduced. If condition (18) is met when performing the evaporation judgment, evaporation and rainfall will still occur even if the evaporation condition is not met. After the mutation and exchange judgments are completed, return to step 2.4 for the next round of population iteration until the maximum number of iterations is met. (18)。 2. The real-time online early warning method for the non-starting state of a supersonic air intake based on deep learning according to claim 1, characterized in that, The pressure signal is preprocessed using the moving average method, and the processing formula is as follows. Sliding step size, Indicates sliding and, Represents the moving average and This indicates the preprocessed value: 。 3. The real-time online early warning method for the non-starting state of a supersonic air intake based on deep learning according to claim 1, characterized in that, Step 3, training the WPT-CNN network, includes the following steps: Step 3.1: The processed training set is input into the WPT network layers. Each layer of WPT decomposes the pressure signal according to equations (19) and (20). This represents the low-frequency information after decomposition. This represents the high-frequency information after decomposition. and These represent the low-frequency and high-frequency filter coefficients, respectively, which are determined by the selected decomposition wavelet type; ； Step 3.2: Convert the signal after WPT decomposition into Gray code arrangement according to the rules of equations (21), (22), and (23) to generate the corresponding time-frequency matrix. Assume the natural binary code is A. n-1 A n-2 …A2A1A0, its corresponding Gray code is B n-1 B n-2 …B2B1B0; ； Step 3.3: The processed time-frequency matrix is ​​fed into the CNN network for training. The convolution process is as shown in equations (24) and (25). This represents the convolution result. Represents the ReLU activation function. This represents the result of max pooling, which is then processed through multiple convolutions to generate the output according to equations (26) and (27). The error is calculated based on (28) and (29). The gradient is obtained based on the error, and the network parameters are then updated accordingly. ； Step 3.4: After the maximum number of iterations, the network training is completed, and the output network is used as a classifier. This classifier can determine whether a warning precursor phenomenon has occurred at the current sensor location by receiving real-time signals.

4. The real-time online early warning method for the non-starting state of a supersonic air intake based on deep learning according to claim 1, characterized in that, The WPT-CNN network includes single-channel and multi-channel models. For single-channel models, only the historical pressure signal from a single sensor needs to be input. For multi-channel models, the pressure signals from multiple sensors need to be converted into multi-dimensional information before being input into the model.

Images