Supersonic inlet flow state monitoring method based on discriminative feature learning
Through time-frequency analysis and deep learning technology, combined with DDL-CNN/TDL-CNN network, the problem of ignoring dynamic characteristics and noise factors in the existing technology is solved, and efficient monitoring of the flow state of the ultrasonic intake duct is achieved, and monitoring is improved. accuracy and reliability.
Patent Information
- Application Number
- CN202110929263.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-13
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-08-13
AI Technical Summary
The existing ultrasonic intake duct flow state monitoring methods mainly rely on CFD simulation data, ignoring the dynamic characteristics, noise and interference factors in the dynamic test data, resulting in poor monitoring effect in actual flight.
Using time-frequency analysis and deep learning technology, the dynamic pressure signal is collected by arranging sensors inside the ultrasonic intake duct, the time-frequency analysis is performed using continuous wavelet transformation, and the classification is carried out through the DDL-CNN/TDL-CNN network, considering the cross entropy loss and Distinguishing feature learning to improve monitoring accuracy.
Effective monitoring of the flow state of the ultrasonic intake duct is achieved, performance on multiple indicators is improved, reliability and adaptability of monitoring is enhanced, and safety threats of surge phenomena are reduced.
Smart Images

Figure CN113836787B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of supersonic inlet flow state monitoring, and in particular relates to a supersonic inlet flow state monitoring method. Background Art
[0002] As one of the key components of the air-breathing supersonic propulsion system, the flow state of the supersonic inlet has a direct impact on the performance of the entire system. For example, supersonic inlet surge is an extremely unstable flow state, usually accompanied by violent and periodic shock wave oscillations, which will seriously deteriorate the thrust characteristics of the entire propulsion system and face the risk of flameout at any time. Not only that, the periodic thermal loads generated by shock wave oscillations will also cause the controllability of the entire system to deteriorate sharply, and may even directly lead to the destruction of the overall structure, posing a threat to the safety of the aircraft and personnel. From the perspective of protection control, since the supersonic inlet exhibits different flow characteristics under different flow states, the control measures taken should also be different. Therefore, in order to reduce the safety threat of surge phenomenon and better control the propulsion system, it is very important to monitor the flow state of the supersonic inlet.
[0003] At present, the monitoring methods for the flow state of the supersonic inlet mainly include methods based on high-speed schlieren photography and data-driven methods. The method based on high-speed schlieren photography is intuitive and reliable, but requires a complex optical path system, which is only feasible in ground tests and is obviously unrealistic in actual flights. The data-driven method is based on CFD simulation data or test data, and uses machine learning and other technologies for intelligent monitoring. This method does not require a complex optical path system, is suitable for both ground tests and actual flights, and can use new data continuously generated in ground tests or actual flights to correct existing models, fully explore effective information in historical data and real-time data, and improve data utilization while making the model more adaptable to external conditions. At present, the research on data-driven methods is mainly based on a small amount of CFD steady-state simulation data, while the actual measurement device usually collects a large amount of dynamic test data. Therefore, in the previous research process, the dynamic characteristics in the actual test data and the influence of factors such as noise and interference will be ignored. The present invention adopts a data-driven method to monitor the flow state of the inlet based on dynamic test data. Summary of the invention
[0004] In response to the above problems, the present invention proposes a method for monitoring the flow state of a supersonic inlet, which uses time-frequency analysis and deep learning technology to monitor the flow state of a supersonic inlet based on dynamic pressure data, and has good performance in multiple indicators.
[0005] To achieve the above object, the present invention adopts the following technical solution:
[0006] A method for monitoring the flow state of a supersonic inlet comprises the following steps:
[0007] Step 1, arranging a plurality of sensors on the internal flow channel of the supersonic inlet duct, respectively collecting the dynamic pressure signal of each sensor, and obtaining samples under different flow states after segmentation;
[0008] Step 2: Use continuous wavelet transform to perform time-frequency analysis on the samples of each sensor to obtain the corresponding time-frequency spectrum, and then divide it into training set, verification set and test set according to a certain ratio;
[0009] Step 3: Use the training sets of each sensor obtained in step 2 to train the DDL-CNN / TDL-CNN network respectively, and use the validation set to select the best hyperparameters;
[0010] Step 4: Use the DDL-CNN / TDL-CNN network trained in step 3 to test the test set of each sensor and obtain the test results.
[0011] In step 1, the dynamic pressure signals collected by each sensor are segmented using a sliding window of fixed length to obtain samples under different flow conditions.
[0012] In step 2, a one-dimensional time signal is transformed into a two-dimensional time-frequency spectrum using continuous wavelet transform, so that the essential characteristics of the original signal can be revealed in both the time domain and the frequency domain.
[0013] In step 3, the DDL-CNN / TDL-CNN network accepts two / three samples as inputs respectively at the same time during the training phase, and the training objective considers both the classification performance and the distinguishability between the input sample features.
[0014] Step 2.1: Initialize the network parameters, including all parameters {θ} before the penultimate layer of DDL-CNN / TDL-CNN and the weights and bias {W} of the output layer (i.e., the penultimate layer) j , b j}. Among them, W j is the weight between the penultimate layer and the jth node of the output layer, b j is the bias of the jth node in the output layer;
[0015] Step 2.2, use equations (1)-(2) to calculate the output of the penultimate layer and the output layer respectively;
[0016] f i =g(X i |θ) (1)
[0017] z ij =W j T f i +b j (2)
[0018] Among them, X i represents the i-th training sample, g(·|θ) represents the mapping relationship from the input layer to the penultimate layer, and f i represents the output of the i-th training sample at the penultimate layer, which can also be considered as the feature extracted by the convolution module. ij is the output of the jth node in the output layer;
[0019] Step 2.3, use equations (3)-(4) to calculate the output (i.e., predicted probability) and cross entropy loss after applying Softmax to the output layer;
[0020]
[0021]
[0022] Among them, y ij and Respectively represent X i The true and predicted probabilities of belonging to the jth class, Represents the cross entropy loss of all training samples, N represents the number of all training samples, and M represents the number of all categories;
[0023] Step 2.4, DDL-CNN not only needs to input two samples X i1 and X i2 The classification is correct, and it is also necessary to ensure that in the feature space mapped by the convolution module, if X i1 and X i2 If they are from the same category, they should be closer to each other, otherwise they should be far away from each other.
[0024] Therefore, the total loss of DDL-CNN consists of two parts, as shown in Equation (5).
[0025]
[0026] in, It represents the average cross entropy loss for classifying two input samples, as shown in formula (6).
[0027]
[0028] It is necessary to ensure that the two input samples satisfy a specific distance relationship in the feature space: if the two samples are from the same class, their distance in the feature space should be smaller, otherwise it should be larger, as shown in formula (7).
[0029]
[0030] If for two samples from the same class there is h i,12 = -1, for samples from different classes, there are h i,12 =+1, then equation (7) can be written as a restriction condition similar to that of support vector machine.
[0031]
[0032]
[0033] in, Head d i,12 =[g(X i1 |θ′)-g(X i2 |θ′)] T [g(X i1 |θ′)- g(X i2 |θ′)] is the distance between features extracted by the convolutional module.
[0034] therefore, It can be written as a loss function similar to that of support vector machines.
[0035]
[0036] Finally, the total loss of DDL-CNN can be calculated by formula (10).
[0037]
[0038] If the order Then formula (10) can be further written as formula (11).
[0039]
[0040] where C′ = ρC, and Represents the predicted probability.
[0041] In order to update (θ′, u j , b j , β), it is necessary to transform formula (11) into an optimization problem without constraints, as shown in formula (12), and then use the gradient-based optimization algorithm to update it.
[0042]
[0043] Before updating the parameters, it is necessary to use equations (13)-(16) to calculate About (θ′, u j , b j , β).
[0044] The derivative calculation with respect to β is shown in formula (13).
[0045]
[0046] Where II{·} is the indicator function.
[0047] About b j The derivative calculation of is shown in formula (14).
[0048]
[0049] About U j The derivative calculation of is shown in formula (15).
[0050]
[0051] The derivative calculation with respect to θ′ is shown in formula (16).
[0052]
[0053] in, is g(X i1 |θ′) / g(X i2 |θ′) with respect to θ′.
[0054] Step 2.5, TDL-CNN not only needs to input three samples X i1 , X i2 and X i3 Correct classification (where X i1 and X i2 From the same class, but X i1 and X i3 from different classes), and it is necessary to ensure that in the feature space after being mapped by the convolution module, X i1 and X i3 The distance between them should be greater than X i1 and X i2 The distance between.
[0055] Therefore, the total loss of TDL-CNN also consists of two parts, as shown in Equation (17).
[0056]
[0057] in, represents the average cross entropy loss for classifying the three input samples, as shown in formula (18).
[0058]
[0059] It is necessary to ensure that the three input samples satisfy a specific distance relationship in the feature space: the distance between samples of different classes should be greater than the distance between samples of the same class, as shown in formula (19).
[0060]
[0061] Formula (19) is similar to the constraints of support vector machines. Therefore, It can also be written as a loss function similar to that of support vector machines.
[0062]
[0063] Finally, the total loss of TDL-CNN can be calculated by formula (21).
[0064]
[0065] If the order Then equation (21) can be further written as equation (22).
[0066]
[0067] In order to update (θ′, u j , b j ), it is necessary to transform formula (22) into an optimization problem without constraints, as shown in formula (23), and then use the gradient-based optimization algorithm to update it.
[0068]
[0069] Before updating the parameters, it is necessary to use equations (24)-(26) to calculate About (θ′,u j , b j ) is the derivative of .
[0070] About b j The derivative calculation of is shown in formula (24).
[0071]
[0072] About U jThe derivative calculation of is shown in formula (25).
[0073]
[0074] The derivative calculation with respect to θ′ is shown in formula (26).
[0075]
[0076] in, is g(X i1 |θ′) / g(X i2 |θ′) / g(X i3 |θ′) with respect to θ′.
[0077] Step 2.6, update the network parameters using a gradient-based optimization algorithm.
[0078] Beneficial effects:
[0079] With the help of time-frequency analysis and deep learning technology, DDL-CNN / TDL-CNN not only considers the cross-entropy loss, but also encourages the convolution module to project samples of different flow states into a more discriminative feature space, so that it has better performance in multiple indicators. BRIEF DESCRIPTION OF THE DRAWINGS
[0080] Figure 1 It is a schematic diagram of mapping using kernel function;
[0081] Figure 2 It is a schematic diagram of a convolutional neural network;
[0082] Figure 3 It is a schematic diagram of the DDL-CNN network structure proposed in the present invention;
[0083] Figure 4 It is a schematic diagram of the TDL-CNN network structure proposed in the present invention;
[0084] Figure 5 It is the convolutional neural network architecture adopted by the present invention;
[0085] Figure 6 is a flow chart of signal processing performed by the present invention;
[0086] Figure 7 It is a basic configuration of a dual-dimensional external pressure supersonic inlet;
[0087] Figure 8 is the dynamic pressure signal of R01 and R14;
[0088] Fig. 9It is a comparison chart of the monitoring accuracy, macroscopic F1 and microscopic F1 of each sensor;
[0089] Fig.10 It is the CD diagram in nonparametric statistical tests;
[0090] Fig.11 This is a visualization diagram after dimensionality reduction using t-sne. DETAILED DESCRIPTION
[0091] A method for monitoring the flow state of a supersonic inlet duct of the present invention comprises the following steps:
[0092] Step 1, arranging a plurality of sensors on the internal flow channel of the supersonic inlet duct, respectively collecting the dynamic pressure signal of each sensor, and obtaining samples under different flow states after segmentation;
[0093] Step 2: Use continuous wavelet transform to perform time-frequency analysis on the samples of each sensor to obtain the corresponding time-frequency spectrum, and then divide it into training set, verification set and test set according to a certain ratio;
[0094] Step 3: Use the training sets of each sensor obtained in step 2 to train the DDL-CNN / TDL-CNN network respectively, and use the validation set to select the best hyperparameters;
[0095] Step 4: Use the DDL-CNN / TDL-CNN network trained in step 3 to test the test set of each sensor and obtain the test results.
[0096] In the present invention, firstly, continuous wavelet transform is used to perform time-frequency analysis on the original time domain signal to obtain a two-dimensional time-frequency spectrum, which is then input into the convolutional neural network for classification. In order to reduce the classification error of the convolutional neural network, while considering the cross entropy loss function, it is also required that the features extracted by the convolution module satisfy a certain relationship, so that samples of the same type are closer, while samples of different types are far away from each other.
[0097] First, continuous wavelet transform, support vector machine and convolutional neural network are briefly introduced, and then the DDL-CNN / TDL-CNN network proposed in the present invention is specifically described.
[0098] (1) Continuous wavelet transform
[0099] The time-frequency analysis method is a common method for analyzing non-stationary signals. It can reflect the frequency components and their distribution that change over time. Common time-frequency analysis methods include Short Time Fourier Transform (STFT) and Continuous Wavelet Transform (CWT). The short-time Fourier transform applies a sliding time window to the time domain signal and then performs a Fourier transform on the signal within the time window, so it can reflect the change of frequency over time. However, the time window length of the short-time Fourier transform is fixed, and it cannot achieve good time resolution and frequency resolution at the same time. The continuous wavelet transform can effectively make up for this shortcoming because its time window decreases as the signal frequency increases, and vice versa. CWT performs multi-scale analysis of the signal through scaling and translation operations. Its basic definition is as follows:
[0100]
[0101] Where a is the scaling factor, b is the translation factor, s(t) is the one-dimensional time domain signal, is the wavelet basis function.
[0102] (2) Support Vector Machine
[0103] Support Vector Machine (SVM) was originally designed to solve binary classification tasks. It generates an optimal hyperplane and uses the maximum margin criterion to separate the two types of samples. Given a linearly separable data set By introducing the slack variable ξ i , the definition of soft support vector machine is as follows:
[0104]
[0105] Where w and b are the coefficients of the hyperplane.
[0106] The unconstrained optimization problem corresponding to formula (1) is as follows, which can be optimized by gradient-based algorithms such as SGD / Adam.
[0107]
[0108] If the data set is nonlinear, a kernel function can be used to map samples from the original space to a high-dimensional space, making the samples easier to distinguish in the high-dimensional space, such as Figure 1 Commonly used kernel functions include polynomial kernel and radial basis function kernel.
[0109] (3) Convolutional Neural Network
[0110] As a commonly used deep learning method, Convolutional Neural Network (CNN) has achieved remarkable success in many fields such as fault diagnosis, computer vision, and natural language processing. The overall structure of CNN is as follows: Figure 2 shown.
[0111] CNN is mainly composed of convolutional layers, pooling layers and fully connected layers. The convolutional layer extracts features by performing convolution operations on the local receptive field, which can greatly reduce the number of network parameters and the complexity of the model. Given an input image X i and its label Y i , the output of the convolutional layer is calculated as follows.
[0112]
[0113] Among them, * represents the convolution operation, W conv and b conv denote the weight and bias of the convolutional layer respectively, and σ(·) denotes the activation function.
[0114] The pooling layer downsamples the output of the convolutional layer by calculating the average or maximum value of the local receptive field, thereby reducing the amount of calculation while retaining key features. The output of the pooling layer is calculated as follows.
[0115]
[0116] Here, down_sample(·) represents average pooling or maximum pooling.
[0117] The fully connected layer is the same as the traditional feedforward neural network, where each input neuron is connected to each output neuron. If the input of the fully connected layer is denoted by f i , then its output is calculated as follows.
[0118]
[0119] Among them, W fully and b fully They represent the weights and bias of the fully connected layer respectively.
[0120] The output of the output layer is similar to equation (6), but without the activation function.
[0121]
[0122] Among them, W out and b out denote the weight and bias of the output layer respectively.
[0123] The softmax function in formula (8) is usually used to calculate the prediction probability, and the cross entropy in formula (9) is used as the loss function for the classification problem.
[0124]
[0125]
[0126] Among them, z ij Indicates z i The jth element in represents the predicted probability of the jth class, N is the number of training samples, and M is the number of categories.
[0127] (4) DDL-CNN / TDL-CNN Network
[0128] The supersonic inlet is one of the key components of the supersonic air-breathing propulsion system, and its flow state has a great influence on the performance of the entire system. For example, the non-starting is a very unstable flow state with strong periodic shock oscillations, which will lead to a significant reduction in engine thrust and specific impulse. Therefore, it is of great significance to monitor the flow state of the supersonic inlet.
[0129] Machine learning has been widely used in the fields of thrust estimation, remaining service life prediction, fault diagnosis and detection, and intelligent control of aircraft engines. However, there is still little research on machine learning in the monitoring of inlet flow state. Trapier et al. proposed a cumulative sum algorithm and a generalized likelihood ratio algorithm to warn the non-starting state of the inlet, but they cannot distinguish different flow states. Chang Juntao et al. combined the support vector machine recursive feature elimination algorithm with Fisher linear discriminant analysis to study the classification of hypersonic inlet start / non-starting states, and further identified the non-starting states caused by different reasons. In order to improve the monitoring reliability under sensor noise and external interference, Chang Juntao et al. studied a multi-classifier fusion method based on probability output support vector machine. However, their research is mainly based on steady-state samples of CFD simulation, while most actual systems in engineering collect dynamic signals, so some dynamic characteristics may be ignored.
[0130] In order to collect information about the flow state, multiple sensors are usually installed inside the intake duct. From the perspective of machine learning, the problem of monitoring the flow state from dynamic sensor signals can be regarded as a time series classification (TSC) task. The TSC task is different from traditional classification tasks because there will be discriminative features that depend on the time order. In recent years, deep learning has become an effective method and has been successfully applied to many TSC tasks. Convolutional neural networks are one of the most successful deep learning models, which have successfully solved many complex problems in the fields of fault diagnosis and remaining service life prediction. Convolutional neural networks can automatically extract features from images for classification. Cross entropy loss and softmax layer are one of the most commonly used components in convolutional neural networks, but this component does not encourage the convolution module to learn discriminative features.
[0131] Dynamic sensor signals usually contain a lot of useful information about the flow state, so signal processing is a key step. Traditional spectrum analysis methods, such as Fourier Transform (FT), are mainly applicable to stable signals whose dynamic characteristics do not change with time. However, during the transition from the start-up state to the non-start-up state of the intake duct, the dynamic characteristics of the sensor signal change dramatically with time, so time-frequency analysis methods are needed to process such non-static signals.
[0132] The present invention combines time-frequency analysis with deep learning technology to monitor the flow state of a supersonic inlet from dynamic sensor signals. First, the continuous wavelet transform (CWT) is used to preliminarily process the signal and convert the dynamic sensor signal into a two-dimensional time-frequency spectrum. Then, a convolutional neural network is used to classify the time-frequency spectrum to obtain the corresponding flow state. In order to reduce the classification error, the present invention proposes a dual-channel / triple-channel convolutional neural network (DDL-CNN / TDL-CNN) that simultaneously considers cross-entropy loss and discriminative feature learning. The proposed method encourages the convolution module to project the time-frequency spectrum into a feature space, so that different flow states in the space become more separable. Experimental results show that compared with traditional CNN, DDL-CNN / TDL-CNN has better performance in multiple indicators.
[0133] The training process of DDL-CNN / TDL-CNN network includes two steps: 1) forward calculation; 2) gradient calculation and update of network parameters. The specific process is as follows.
[0134] Use equations (10)-(11) to calculate the output of the penultimate layer and the output layer respectively;
[0135] fi =g(X i |θ) (10)
[0136] z ij =W j T f i +b j (11)
[0137] Among them, X i represents the i-th training sample, g(·|θ) represents the mapping relationship from the input layer to the penultimate layer, and f i represents the output of the i-th training sample at the penultimate layer, which can also be considered as the feature extracted by the convolution module. ij is the output of the jth node in the output layer;
[0138] Use equations (12)-(13) to calculate the output (i.e., predicted probability) and cross entropy loss after applying Softmax to the output layer respectively;
[0139]
[0140]
[0141] Among them, yij and Respectively represent X i The true probability and predicted rate of belonging to the jth class, Represents the cross entropy loss of all training samples, N represents the number of all training samples, and M represents the number of all categories;
[0142] For DDL-CNN, it is necessary not only to input two samples X i1 and X i2 The classification is correct, and it is also necessary to ensure that in the feature space mapped by the convolution module, if X i1 and X i2 If they are from the same category, they should be close to each other, otherwise they should be far away from each other. Figure 3 Therefore, the total loss of DDL-CNN consists of two parts, as shown in formula (14).
[0143]
[0144] in, represents the average cross entropy loss for classifying two input samples, as shown in formula (15).
[0145]
[0146] It is necessary to ensure that the two input samples satisfy a specific distance relationship in the feature space: if the two samples are from the same class, their distance in the feature space should be small, otherwise it should be large, as shown in formula (16).
[0147]
[0148] If for two samples from the same class there is h i,12 = -1, for samples from different classes, there are h i,12 =+1, then equation (16) can be written as a restriction condition similar to that of support vector machine.
[0149]
[0150] in, and d i,12 =[g(X i1 |θ′)-g(X i2 |θ′)]T[g(X i1 |θ′)- g(X i2 |θ′)] is the distance between the features extracted by the convolution module. Therefore, It can be written as a loss function similar to that of support vector machines.
[0151]
[0152] Finally, the total loss of DDL-CNN can be calculated by formula (19).
[0153]
[0154] If the order Then equation (19) can be further written as equation (20).
[0155]
[0156] where C′ = ρC, and represents the predicted probability.
[0157] In order to update (θ′, u j , b j , β), it is necessary to transform formula (20) into an optimization problem without constraints, as shown in formula (21), and then use the gradient-based optimization algorithm to update it.
[0158]
[0159] Before updating the parameters, it is necessary to use equations (22)-(25) to calculate About (θ′, u j , bj , β).
[0160] The derivative calculation with respect to β is shown in formula (22).
[0161]
[0162] Where II{·} is the indicator function.
[0163] About b j The derivative calculation of is shown in formula (23).
[0164]
[0165] About U j The derivative calculation of is shown in formula (24).
[0166]
[0167] The derivative calculation with respect to θ′ is shown in formula (25).
[0168]
[0169] in, is g(X i1 |θ′) / g(X i2 |θ′) with respect to θ′.
[0170] Algorithm 1 summarizes the training process of the DDL-CNN network proposed in this invention, as shown in Table 1.
[0171] Table 1
[0172]
[0173] For TDL-CNN, it is necessary not only to input three samples X i1 , X i2 and X i3 Correct classification (where X i1 and X i2 From the same class, but X i1 and X i3 from different classes), and it is necessary to ensure that in the feature space after being mapped by the convolution module, X i1 and X i3 The distance between them should be greater than X i1 and X i2 The distance between Figure 4Therefore, the total loss of TDL-CNN also consists of two parts, as shown in formula (26).
[0174]
[0175] in, represents the average cross entropy loss for classifying the three input samples, as shown in formula (27).
[0176]
[0177]
[0178] It is necessary to ensure that the three input samples satisfy a specific distance relationship in the feature space: the distance between samples of different classes should be greater than the distance between samples of the same class, as shown in formula (28).
[0179]
[0180] Formula (28) is similar to the constraints of support vector machines. Therefore, It can also be written as a loss function similar to that of support vector machines.
[0181]
[0182] Finally, the total loss of TDL-CNN can be calculated by formula (30).
[0183]
[0184] If the order Then equation (30) can be further written as equation (31).
[0185]
[0186] In order to update (θ′, u j , b j ), it is necessary to transform formula (31) into an optimization problem without constraints, as shown in formula (32), and then use the gradient-based optimization algorithm to update it.
[0187]
[0188] Before updating the parameters, it is necessary to use equations (33)-(35) to calculate About (θ′,u j , b j ) is the derivative of .
[0189] About b jThe derivative calculation of is shown in formula (33).
[0190]
[0191] About U j The derivative calculation of is shown in formula (34).
[0192]
[0193] The derivative calculation with respect to θ′ is shown in formula (35).
[0194]
[0195] in, is g(X i1 |θ′) / g(X i2 |θ′) / g(X i3 |θ′) with respect to θ′.
[0196] Algorithm 2 summarizes the training process of the TDL-CNN network proposed in this invention, as shown in Table 2.
[0197] Table 2
[0198]
[0199] Finally, the network parameters of DDL-CNN / TDL-CNN are updated using a gradient-based optimization algorithm, such as the Adam algorithm.
[0200] The present invention will be further described below in conjunction with specific embodiments.
[0201] Example
[0202] This example uses experimental data of a binary external pressure supersonic inlet to verify the effectiveness of the proposed DDL-CNN / TDL-CNN network. All experiments were performed on a laptop computer configured with IntelR CoreTM, i7-7700HQ CPU, 2.80GHz main frequency, 8G memory, Windows 10 system and MATLAB 2020b version.
[0203] In order to prove the effectiveness of the DDL-CNN / TDL-CNN network, this embodiment compares it with the traditional CNN. The network architectures used in these comparison methods are the same, and the only difference is their loss functions. The loss function of the traditional CNN is the cross entropy loss, while the DDL-CNN / TDL-CNN considers the learning of both the cross entropy loss and the discriminative features.
[0204] In the specific implementation process, the dynamic pressure signals collected by each sensor are first segmented using a sliding window of fixed duration to obtain samples of different flow states with a duration of 50ms. Then, all samples are divided into training set, validation set and test set in a ratio of 6:2:2. Among them, the validation set is used to select the optimal hyperparameters in various comparison methods (that is, the hyperparameters corresponding to the maximum accuracy of the validation set), and the test set is used to compare the performance of various comparison methods.
[0205] The present invention uses continuous wavelet transform to perform preliminary signal processing on the dynamic sensor signal and converts the dynamic sensor signal into a two-dimensional time-frequency spectrum. Before using the continuous wavelet transform, the sliding average needs to be subtracted from the dynamic signal to avoid the influence of the direct component. Then, the elements at the same position in the time-frequency spectrum are normalized to the range of [0, 1] and input into the convolutional neural network for classification. The convolutional neural network architecture used in the present invention is as follows Figure 5 The specific details are shown in Table 3.
[0206] Table 3
[0207]
[0208] Batch normalization layers are used to stabilize and speed up the training process, and ReLU is selected as the activation function. Adam optimizer is used to update the parameters. The number of iterations is set to 3000, the learning rate is set to 0.001, the gradient decay factor is set to 0.9, the squared gradient decay factor is set to 0.999, and the range of C is set to [2 -10 , 2 -9 , …, 2 0 ]. The entire signal processing process is as follows Figure 6 shown.
[0209] The steps of this embodiment are as follows:
[0210] Step 1: Use the experimental data of a binary external pressure supersonic inlet to verify the effectiveness of the DDL-CNN / TDL-CNN network in monitoring flow states.
[0211] The dynamic pressure signals of various sensors on the inlet wall of a binary external pressure supersonic inlet during the throttling process under rated conditions are collected, and the collected pressure signals are segmented using a sliding window to obtain samples under different flow conditions.
[0212] Step 2: Use continuous wavelet transform to perform time-frequency analysis on the samples of each sensor to obtain the corresponding time-frequency spectrum, and then divide it into training set, validation set and test set according to a certain ratio.
[0213] Step 3: Establish and train the DDL-CNN / TDL-CNN network based on the samples of each sensor, select the best hyperparameters using the validation set, and compare the monitoring effect with that of the traditional CNN on the test set.
[0214] In the case of monitoring the flow state of a supersonic inlet, the dynamic pressure signals of each sensor on the inlet wall of a binary external pressure supersonic inlet under the rated state during the throttling process are first collected, and the collected pressure signals are segmented using a sliding window to obtain samples under different flow states. All samples are converted into two-dimensional time-frequency spectrograms using continuous wavelet transform, and divided into training set, validation set and test set in a ratio of 6:2:2. Then, based on the training sets of each sensor, the DDL-CNN / TDL-CNN network is trained using Algorithm 1 and Algorithm 2, respectively, and compared with the traditional CNN.
[0215] The research object of this embodiment is a type of dual external pressure supersonic inlet, whose basic configuration is as follows: Figure 7 As shown. The design Mach number of the inlet is 2.0, the theoretical capture area is 48.0mm×40.0mm, and the area contraction ratio is 1.5. The upper compression surface adopts a partial isentropic design, which is composed of a first-level inclined wedge and a first-level isentropic surface. In order to take into account the aerodynamic performance under subsonic conditions and prevent the formation of a large area of flow separation on the inner side of the lip mask, the leading edge of the inlet lip mask is passivated. Downstream of the mouth is a section of unilateral expansion channel, followed by a section of iso-flow channel. Optical glass is embedded on both sides of the inlet for internal flow field observation, and its visible window range can cover the core area of the internal flow channel. In order to simulate the downstream throttling effect in actual work, a wedge-shaped throttling cone is set at the outlet of the inlet, and its leading edge is located at the center height of the channel outlet. Driven by the stepper motor, the throttling cone can translate axially to establish a controllable blockage degree at the outlet of the inlet. The present invention uses blockage degree to characterize the degree of congestion downstream of the inlet. During the experiment, the throttle cone will move from the far downstream position to a state close to the outlet after dozens of step-by-step cone entry, causing the blockage to gradually increase from 0% to 99.1%. In order to obtain the dynamic pressure information in the inlet flow field during the throttling process, a total of 31 dynamic pressure measurement points with flow distribution were set on the center line of the upper and lower inner surfaces of the flow channel, and the measurement range covered the core area of the inner flow channel. According to the different positions, the above sensors are numbered C01-C17 and R01-R14, respectively, and the specific experimental conditions are shown in Table 4. Figure 8 The dynamic pressure signals of R01 and R14 during the throttling process are shown. During this process, the intake port experienced a stable state, a small gasp state, a mixed gasp state, and a large gasp state.
[0216] Table 4
[0217]
[0218] Table 5 shows the three multi-classification performance metrics (i.e., monitoring accuracy, macro F1, and micro F1 scores) of all compared methods on 31 sensors.
[0219] Table 5
[0220]
[0221]
[0222] For ease of comparison, we plotted monitoring accuracy, macro F1, and micro F1 on Fig. 9 It can be seen that in terms of the three multi-classification performance indicators, DDL-CNN / TDL-CNN is higher than traditional CNN on most sensors.
[0223] This example also performs some non-parametric statistical tests (Friedman test and Nemenyi follow-up test) to compare the performance of various methods. The parameter τ in the Friedman test F It is calculated as follows and follows an F distribution with (k-1) and (k-1)(D-1) degrees of freedom.
[0224]
[0225]
[0226] Among them, k is the number of comparison methods, D is the number of data sets, and r i It represents the average ordinal value of various methods on all data sets.
[0227] After calculation, when k = 3 and D = 31, τ F It is equal to 33.0120, which is greater than the critical value of F distribution when α=0.05, 3.1504. Therefore, the hypothesis that all comparison methods have the same performance is rejected, and the Nemenyi follow-up test needs to be continued. The calculation formula of the critical difference (CD) is as follows:
[0228]
[0229] Among them, when α = 0.05, k = 3 and D = 31, q α =2.344, CD=0.5954.
[0230] Fig.10is a Friedman test graph, where the length of the blue line is equal to the CD value, and the red dots represent the average sequence values of various methods. If the classification performance of the two compared methods is significantly different, the corresponding average sequence values differ by at least one CD value. It can be seen that the DDL-CNN / TDL-CNN network proposed in the present invention is significantly better than the traditional CNN network.
[0231] Fig.11 We further demonstrate the use of a dimensionality reduction technique called t-distributed stochastic neighbor embedding (t-SNE) to visualize the features extracted by the penultimate layer of all the comparison methods. It can be seen that compared with traditional CNN, after additional consideration of discriminative feature learning, DDL-CNN / TDL-CNN makes different categories more discriminative, further proving the effectiveness of the proposed method.
Claims
1. A method for monitoring the flow state of a supersonic inlet, characterized in that: The steps include: Step 1, arranging a plurality of sensors on the internal flow channel of the supersonic inlet duct, respectively collecting the dynamic pressure signal of each sensor, and obtaining samples under different flow states after segmentation; Step 2: Use continuous wavelet transform to perform time-frequency analysis on the samples of each sensor to obtain the corresponding time-frequency spectrum, and then divide it into training set, validation set and test set; Step 3: Use the training sets of each sensor obtained in step 2 to train the DDL-CNN or TDL-CNN network respectively, and use the validation set to select the best hyperparameters; Step 4, using the DDL-CNN or TDL-CNN network trained in step 3 to test the test set of each sensor respectively, and obtain the test results; In step 3, the training steps of the DDL-CNN or TDL-CNN network model include: Step 2.1, initialize the network parameters, including all parameters {θ} before the penultimate layer of DDL-CNN or TDL-CNN and the weights and biases {W j , b j }; where W j is the weight between the penultimate layer and the jth node of the output layer, b j is the bias of the jth node in the output layer; Step 2.2, use equations (1)-(2) to calculate the output of the penultimate layer and the output layer respectively; f i =g(X i |i) (1) z ij =W j T f i +b j (2) Among them, X i represents the i-th training sample, g(·|θ) represents the mapping relationship from the input layer to the penultimate layer, and f i represents the output of the i-th training sample at the penultimate layer, which can also be considered as the feature extracted by the convolution module. ij is the output of the jth node in the output layer; Step 2.3, use equations (3)-(4) to calculate the output and cross entropy loss after applying Softmax to the output layer; Among them, y ij and Respectively represent X i The true and predicted probabilities of belonging to the jth class, Represents the cross entropy loss of all training samples, N represents the number of all training samples, and M represents the number of all categories; Step 2.4, calculate the total loss of DDL-CNN or TDL-CNN; Step 2.5, use the Adam optimization algorithm to update the network parameters; The specific steps to calculate the total loss of DDL-CNN include: DDL-CNN not only needs to input two samples X i1 and X i2 The classification is correct, and it is also necessary to ensure that in the feature space mapped by the convolution module, if X i1 and X i2 If they are from the same class, they should be close to each other, otherwise they should be far away from each other; therefore, the total loss of DDL-CNN consists of two parts, as shown in formula (5): in, Represents the average cross entropy loss for classifying two input samples, Indicates whether the loss from the same class or different classes is correctly judged, and ρ is and The weight coefficient between The calculation of is shown in formula (6); Among them, Y i1,j Represents X i1 The true probability of belonging to the jth class, Represents X i1 The true probability of belonging to the jth class, z i1,j Represents the linear output of the output layer before entering the Softmax layer; Indicates whether the losses from the same or different categories are correctly judged, i.e. It is necessary to ensure that the two input samples satisfy a specific distance relationship in the feature space: if the two samples are from the same class, their distance in the feature space should be small, otherwise it should be large, as shown in formula (7): Among them, δ represents the reference distance between two samples, and Δ represents the interval between two samples; If for two samples from the same class there is h i , 12 = -1, for samples from different classes, there are h i , 12 =+1, then equation (7) can be written as a constraint similar to that of the support vector machine; in, and d i,12 =[g(x i1 |θ′)-g(X i2 |θ′)]T[g(X i1 |θ′)-g(X i2 |θ′)] is the distance between the features extracted by the convolution module; g(·|θ′) is g(·|θ) divided by The parameter after represents the normalized distance between two samples; therefore, It can be written as a loss function similar to that of a support vector machine; Finally, the total loss of DDL-CNN can be calculated by formula (10); make Then formula (10) can be further written as formula (11); where C′ = ρC, and represents the predicted probability; C and C′ represent the weight coefficients between the two loss terms; In order to update (θ′, u j , b j , β), transform equation (11) into an optimization problem without constraints, as shown in equation (12), and then use the gradient-based optimization algorithm to update it; in, represents a new set of weight parameters; Before updating the parameters, use equations (13)-(16) to calculate About (θ′,u j , b j ,β) derivative; The derivative calculation of β is shown in formula (13); Among them, II{·} is the indicator function; About b j The derivative calculation of is shown in formula (14); About U j The derivative calculation of is shown in formula (15); The derivative calculation about θ′ is shown in formula (16); in, is g(X i1 |θ′) / g(X i2 |θ′) Jacobian matrix with respect to θ′; The specific steps to calculate the total loss of DDL-CNN include: TDL-CNN not only needs to input three samples X i1 , X i2 and X i3 The classification is correct, among which, X i1 and X i2 From the same class, but X i1 and X i3 From different classes, it is necessary to ensure that in the feature space after mapping by the convolution module, X i1 and X i3 The distance between them should be greater than X i1 and X i2 The distance between Therefore, the total loss of TDL-CNN also consists of two parts, as shown in formula (17); in, represents the average cross entropy loss for classifying the three input samples, as shown in formula (18); It is necessary to ensure that the three input samples satisfy a specific distance relationship in the feature space: the distance between samples of different classes should be greater than the distance between samples of the same class, as shown in formula (19); Formula (19) is similar to the constraints of support vector machines; therefore, It can also be written as a loss function similar to that of a support vector machine; Finally, the total loss of TDL-CNN can be calculated by formula (21); make Then formula (21) can be further written as formula (22); In order to update (θ′, u j , b j ), transforming formula (22) into an optimization problem without constraints, as shown in formula (23), and then updating it using the gradient-based optimization algorithm; Before updating the parameters, use equations (24)-(26) to calculate About (θ′,u j , b j )’s derivative; About b j The derivative calculation of is shown in formula (24); About U j The derivative calculation of is shown in formula (25); The derivative calculation about θ′ is shown in formula (26); in, is g(X i1 |θ′) / g(X i2 |θ′) / g(X i3 |θ′) with respect to θ′.
2. The supersonic inlet flow state monitoring method according to claim 1, characterized in that: In step 1, the dynamic pressure signals collected by each sensor are segmented using a sliding window of fixed length to obtain samples under different flow conditions.
3. The supersonic inlet flow state monitoring method according to claim 1, characterized in that: In step 2, a one-dimensional time signal is transformed into a two-dimensional time-frequency spectrum using continuous wavelet transform, so that the essential characteristics of the original signal can be revealed in both the time domain and the frequency domain.
4. The supersonic inlet flow state monitoring method according to claim 1, characterized in that: In step 3, the DDL-CNN or TDL-CNN network accepts two or three samples as inputs respectively at the same time during the training phase, and the training objective considers both the classification performance and the distinguishability between the input sample features.
