Hypersonic velocity flow state identification method and system based on deep learning, and medium

Abstract

The invention relates to a hypersonic velocity flow state identification method and system based on deep learning, and a medium, and the method comprises the steps: obtaining a heat flow time sequence signal of a model wall surface, which is measured through a thin-film resistance thermometer in a wind tunnel test; adopting a wavelet transform method to extract time domain-frequency domain information in the heat flow time sequence signal, and constructing an input data set of a deep learning network; building a deep learning framework for feature extraction and flow state clustering, and training and optimizing model parameters by using an input data set to obtain a deep learning clustering model with training index convergence; and performing accuracy analysis on the clustering model to obtain a model capable of being used for wind tunnel time sequence signal flow state identification. According to the method, the deep learning model capable of rapidly identifying and providing the flow state type is established, the dependence of a traditional method on manpower is relieved, and the efficiency of experimental research of the hypersonic wind tunnel is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hypersonic ground wind tunnel flow state recognition. In particular, it relates to a hypersonic flow state recognition method and system based on deep learning and a medium. BACKGROUND

[0002] In hypersonic ground wind tunnel tests, aerodynamic heat measurement is one of the most important measurement objects and one of the physical characteristics that researchers pay most attention to. Sensor heat flow measurement is a relatively mature and widely used measurement method in the field of hypersonic heat flow measurement. Temperature sensors are installed on the wall to measure and collect the wall heat flow based on the temperature changes monitored by the sensors. In the post-processing stage, the measured signals are used to calculate the time series mean value as the heat flow size of the measurement point, and the statistical information of the heat flow time series signal is analyzed to determine whether the flow state of the measurement point belongs to laminar flow, transition or turbulent flow.

[0003] The wall flow state under high-speed flow is often of great concern, and the transition judgment and prediction of flow are always the focus of hypersonic flow research. In aerodynamic heat measurement experiments, the judgment of flow state is often based on the trend of heat flow change along the streamline, and then the statistical characteristics of the heat flow time series signal are calculated to further refine the flow state judgment. This traditional method faces a relatively cumbersome post-processing step and time cost, and relies on subjective judgment. How to further improve the flow state judgment method, improve the cumbersome post-processing step, time cost and dependence on subjective judgment is a problem that needs to be solved. SUMMARY

[0004] The present application provides a hypersonic flow state recognition method and system based on deep learning to solve the problem of the cumbersome post-processing step, time cost and dependence on subjective judgment in the prior art.

[0005] To achieve the above-mentioned purpose, in a first aspect, the present application relates to a hypersonic flow state recognition method based on deep learning, which is used for the recognition of wall heat flow signal flow state in hypersonic ground wind tunnel, comprising:

[0006] Step 1: obtaining the heat flow time series signal of the model wall measured by using a thin film resistance thermometer in a wind tunnel test;

[0007] Step 2: using wavelet transform method to extract the time-frequency domain information in the heat flow time series signal, and constructing the input data set of the deep learning network;

[0008] Step 3: building a deep learning framework for feature extraction and flow state clustering, training and optimizing the input data set to obtain a deep learning clustering model with converged training index;

[0009] Step 4: Accuracy analysis of the clustering model is performed by combining the labeled test set samples, and a model capable of identifying the flow state of the wind tunnel time series signal stream is obtained.

[0010] Preferably, step 1 is specifically:

[0011] A uniform film temperature sensor sampling frequency is set, and a plurality of film temperature sensors are distributed and installed on different experimental model walls to perform wind tunnel experiments under hypersonic speed flow. The time series heat flow signals of the experimental model walls are measured and collected, the time series heat flow signals are divided into a training data set and a test data set, and reliable flow state labels are given after artificial statistical analysis of the test data set, thereby obtaining labeled test set samples.

[0012] Step 2 is specifically:

[0013] Wavelet transform method is used to extract the time domain features and frequency domain features of the time series heat flow signals. A wavelet basis function suitable for the heat flow time series signal to be analyzed is set. The time domain features and frequency domain features of the time series heat flow signals are divided into wavelet coefficients of different frequency bands and standardized, thereby obtaining a new feature data set as an input data set for constructing a deep learning network.

[0014] Step 3 includes:

[0015] Step 31: The input data set for constructing a deep learning network is imported into a stacked autoencoder in the initial state of the deep learning clustering model. The stacked autoencoder is pre-trained layer by layer greedily to optimize the parameters of the stacked autoencoder layers.

[0016] Step 32: The pre-trained stacked autoencoder layers are spliced in order to obtain a complete stacked autoencoder. The input data set is used to iteratively optimize and train the stacked autoencoder, thereby obtaining a stacked autoencoder with feature extraction capability.

[0017] Step 33: The encoder part of the stacked autoencoder with feature extraction capability is combined with a clustering layer. The autoencoder parameters and clustering center parameters of the clustering layer are iteratively optimized and trained based on the input data set. The loss function curve of the training process is recorded.

[0018] Step 34: Based on the loss function curve, the quality of the model obtained in the training process is judged. If the preset requirements are not met, return to step 2 to modify the model initialization training parameters. If the preset requirements are met, proceed to step 4.

[0019] The step 4 is specifically: based on the test set samples, the flow state clustering recognition accuracy of the model is evaluated, and it is judged whether the flow state judgment accuracy of the trained model meets the expected requirement; if the flow state clustering accuracy meets the requirement, a deep learning clustering model for hypersonic heat flow signal flow state recognition is obtained; if the flow state clustering accuracy does not meet the requirement, return to step 2, modify the model initialization training parameters and re-execute the input data set construction module and the model training module for training.

[0020] Preferably, the iterative optimization training content in step 3 is: in the training process, one of the network parameters is changed, the remaining parameters are fixed, the training result is recorded, and the best value of the parameter is adjusted by observing whether the flow state discrimination accuracy of the model on the test set meets the preset requirement; different network parameters are replaced and adjusted one by one, and the above process is repeated until the combination of all network parameters reaches the best value, and the optimization training is ended, and the flow state clustering model is obtained.

[0021] Preferably, the time sequence signal data samples collected from different experiments in step 1 are all a segment intercepted from original measurement signals, and the lengths of all the time sequence signal data samples are consistent.

[0022] Preferably, the stacked autoencoder is one of a convolutional neural network, a residual neural network or a long short-term memory network.

[0023] The clustering layer is a self-defined module containing three types of clustering centers, and the clustering centers correspond to the best expression of three types of flow state signals in the feature space respectively, the three types of clustering centers include a laminar flow clustering center, a transition clustering center and a turbulent flow clustering center, and the three types of flow state signals include a laminar flow signal, a transition signal and a turbulent flow signal.

[0024] The loss function of the stacked autoencoder includes a reconstruction error measured by a mean square error function (MSE), and the formula is:

[0025]

[0026] In the formula, N is the total number of samples, x i is the input feature of the heat flow signal, g(x i ) is the reconstruction input of the stacked autoencoder.

[0027] The loss function of the clustering layer is a KL divergence error and a local preservation error, and the formula is:

[0028]

[0029]

[0030] sim(x i ,x j )=ED(x i ,x j )×cos(x i ,x j )

[0031] where q ij is the probability that the sample i belongs to the j-th class given by the clustering layer, p ij is the auxiliary target probability; sim(x i ,x j ) measures the similarity between sample i and surrounding sample j, ED(x i ,x j ) is the Euclidean distance between sample i and surrounding sample j, and cos(x i ,x j ) is the trigonometric similarity between sample i and surrounding sample j.

[0032] The flow state clustering recognition accuracy acc is calculated using the following formula:

[0033]

[0034] where δ is an indicator function, y i is the artificial flow state label, is the flow state label recognized by the clustering model, and m is the total number of test set samples.

[0035] To achieve the above object, in a second aspect, the present application relates to a high-speed flow state recognition system based on deep learning, which is used for the recognition of high-speed ground wind tunnel wall heat flow signal flow state, comprising:

[0036] A heat flow time series signal measurement module is used to obtain the heat flow time series signal of the model wall measured by using a thin film resistance thermometer in the wind tunnel test;

[0037] An input data set construction module is used to extract the time-frequency domain information in the heat flow time series signal by using a wavelet transform method, and construct an input data set of a deep learning network;

[0038] A model training module is used to build a deep learning framework for feature extraction and flow state clustering, train and optimize the iteration of the model parameters by using the input data set, and obtain a deep learning clustering model with convergent training indicators;

[0039] A model accuracy analysis module is used to analyze the accuracy of the clustering model by combining the labeled test data set samples, and obtain a model that can be used for wind tunnel time series signal flow state recognition.

[0040] Preferably, the heat flow time series signal measurement module is specifically used for:

[0041] A uniform film temperature sensor sampling frequency is set, a plurality of film temperature sensors are distributed and installed on different experimental model walls, a wind tunnel experiment under hypersonic speed flow is carried out, time sequence heat flow signals of the experimental model walls are measured and collected, the time sequence heat flow signals are divided into a training data set and a test data set, reliable flow state labels are given after the test data set is analyzed by artificial statistics, and a labeled test set sample is obtained;

[0042] The input data set construction module is specifically configured to:

[0043] Wavelet transform method is used to extract time domain features and frequency domain features of the time sequence heat flow signals, a wavelet base function suitable for the heat flow time sequence signals to be analyzed is set, the time domain features and the frequency domain features of the time sequence heat flow signals are divided into wavelet coefficients of different frequency bands and are subjected to standardization processing, and a new feature data set is obtained as an input data set for constructing a deep learning network;

[0044] The model training module comprises:

[0045] The autoencoder training submodule is configured to import the input data set for constructing the deep learning network into a stacked autoencoder in the deep learning clustering model in an initial state, perform layer-by-layer greedy pre-training on the stacked autoencoder, and optimize parameters of layers of the stacked autoencoder.

[0046] The autoencoder iterative optimization submodule is configured to splice the pre-trained encoder layers in sequence to obtain a complete stacked autoencoder, perform iterative optimization training on the stacked autoencoder by using the input data set, and obtain a stacked autoencoder with feature extraction capability.

[0047] The loss function curve submodule is configured to combine an encoder part of the stacked autoencoder with feature extraction capability and a clustering layer, perform iterative optimization training on autoencoder parameters and clustering center parameters of the clustering layer based on the input data set, and record a loss function curve of the training process.

[0048] The judgment submodule is configured to judge quality of a model obtained in the training process based on the loss function curve, return to the input data set construction module to modify model initialization training parameters if preset requirements are not met, and execute the model accuracy analysis module if the preset requirements are met.

[0049] The model accuracy analysis module is specifically configured to: evaluate the flow state cluster recognition accuracy of the model based on the test set samples, determine whether the flow state judgment accuracy of the trained model meets the expected requirement, obtain a deep learning cluster model for hypersonic heat flow signal flow state recognition if the flow state cluster accuracy meets the requirement, and return the input data set construction module if the flow state cluster accuracy does not meet the requirement, modify the model initialization training parameters, and re-execute the input data set construction module and the model training module for training.

[0050] Preferably, the stacked autoencoder is one of a convolutional neural network, a residual neural network, or a long short-term memory network.

[0051] The cluster layer is a self-defined module containing three types of cluster centers, which correspond to the best expression of three types of flow state signals in the feature space, and the three types of cluster centers include a laminar flow cluster center, a transition cluster center, and a turbulent flow cluster center.

[0052] The loss function of the stacked autoencoder includes a reconstruction error measured by a mean square error function (MSE), and the formula is:

[0053]

[0054] In the formula, N is the total number of samples, x i is the input feature of the heat flow signal, g(x i is the reconstruction input of the stacked autoencoder.

[0055] The loss function of the cluster layer is a KL divergence error and a local preservation error, and the formula is:

[0056]

[0057]

[0058] sim(x i ,x j )=ED(x i ,x j )×cos(x i ,x j )

[0059] In the formula, q ij is the probability that sample i belongs to the jth type given by the cluster layer, p ij is an auxiliary target probability, sim(x i ,x j ) measures the similarity between sample i and surrounding sample j, and ED(x i ,xj ) is the Euclidean distance of sample i and surrounding sample j, cos(x i ,x j ) is the triangular similarity of sample i and surrounding sample j;

[0060] The flow state clustering recognition accuracy acc is calculated by the following formula:

[0061]

[0062] In the formula, δ is an indicator function, y i is a flow state label, is a flow state label recognized by the clustering model, and m is the total number of test set samples.

[0063] To achieve the above object, the third aspect of the present application also relates to a computer readable storage medium, the storage medium storing instructions, the instructions being executed to perform the above-mentioned high supersonic flow state recognition method based on deep learning.

[0064] The high supersonic flow state recognition method, system and medium based on deep learning have the following beneficial effects compared with the prior art:

[0065] The clustering recognition model in the present application does not depend on artificial judgment, extracts features in the heat flow time sequence signal autonomously, iteratively learns according to the differences between different sample features, classifies the samples, and establishes clustering centers of three different flow states, so that a more objective recognition standard is provided. At the same time, the pre-processing means based on wavelet transform makes the input features of the clustering model have physical interpretability. The present application combines wavelet transform and deep learning model, applies it to the model surface heat flow time sequence signal under high supersonic flow, and performs flow state unsupervised clustering, establishes a deep learning model capable of quickly recognizing and providing flow state types, reduces the dependence on artificial in traditional methods, and improves the efficiency of experimental research in high supersonic wind tunnel. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 It is a flowchart of the high supersonic flow state recognition method based on deep learning in embodiment one.

[0067] Figure 2 It is the loss function curve of the final training process of example 1 of the high supersonic flow state recognition method based on deep learning in embodiment one and the flow state recognition accuracy curve of the test set with model training.

[0068] Figure 3 It is the flow state recognition accuracy table of the unsupervised clustering model on the test set when different wavelet basis functions are used in example 1 of the high supersonic flow state recognition method based on deep learning in embodiment one.

[0069] Figure 4 Figure 1 is a structural schematic diagram of a high-speed flow state recognition system based on deep learning according to an embodiment of the present application. DETAILED DESCRIPTION

[0070] The application will be further described below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the application, but not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings for the convenience of description.

[0071] Embodiment I

[0072] A high-speed flow state recognition method based on deep learning, please refer to Figures 1-3 The high-speed flow state recognition method based on deep learning, the recognition of the high-speed wind tunnel wall surface heat flow signal flow state, includes the following steps: S101 to S104.

[0073] S101 obtains the heat flow time series signal of the model wall surface measured by using the thin film resistance thermometer in the wind tunnel test.

[0074] In this embodiment, S101 includes S111-S112:

[0075] S111 sets a unified thin film temperature sensor sampling frequency, distributes and installs a plurality of thin film temperature sensors on different experimental model walls, performs a wind tunnel experiment under high-speed flow, measures and collects the time series heat flow signal of the experimental model wall.

[0076] S112 divides the time series heat flow signal into a training data set and a test data set, and gives a reliable flow state label after artificial statistical analysis of the test data set, to obtain a labeled test set sample.

[0077] In this embodiment, the time series signal data samples collected from different experiments in S101 are all a segment intercepted from the original measurement signal, and the lengths of all the time series signal data samples are consistent.

[0078] S102 adopts a wavelet transform method to extract the time domain-frequency domain information in the heat flow time series signal, and constructs an input data set of the deep learning network.

[0079] In this embodiment, specifically: the time domain features and frequency domain features of the time series heat flow signal are extracted by using the wavelet transform method, a wavelet basis function suitable for the heat flow time series signal to be analyzed is set, the time domain features and frequency domain features of the time series heat flow signal are divided into wavelet coefficients of different frequency bands and standardized, and a new feature data set is obtained as an input data set of the deep learning network.

[0080] S103 builds a deep learning framework for feature extraction and flow pattern clustering, trains and optimizes iterations of model parameters by using an input data set, and obtains a deep learning clustering model with converged training indicators.

[0081] In this embodiment, S103 includes S131 to S134.

[0082] S131: Import the input data set for constructing the deep learning network into the stacked autoencoder in the initial state of the deep learning clustering model, and perform layer-by-layer greedy pre-training on the stacked autoencoder to optimize the parameters of the stacked autoencoder layers.

[0083] The layer-by-layer greedy pre-training is based on iterative training of samples in the training data set in the input data set to optimize the encoder layer parameters, and finally obtain the pre-trained stacked autoencoder layers.

[0084] S132: splice the pre-trained stacked autoencoder layers in order to obtain a complete stacked autoencoder; and perform iterative optimization training on the stacked autoencoder using the input data set to obtain a stacked autoencoder with feature extraction capability;

[0085] S133: combine the encoder part of the stacked autoencoder with feature extraction capability with the clustering layer, perform iterative optimization training on the autoencoder parameters and the clustering center parameters of the clustering layer based on the input data set, and record the loss function curve of the training process;

[0086] S134: determine the model quality obtained in the training process based on the loss function curve; if the preset requirements are not met, return to S102 to modify the model initialization training parameters, and if the preset requirements are met, proceed to the next step S104;

[0087] In the above S131 to S134, the iterative optimization training content is: in the training process, change one of the network parameters, fix the remaining parameters, record the training results, and adjust the optimal value of the parameter by observing whether the flow pattern discrimination accuracy of the model on the test set meets the preset requirements; replace and adjust different network parameters one by one, repeat the above process until the combination of all network parameters reaches the optimal value, and the optimization training is completed, obtaining the flow pattern clustering model.

[0088] In the above S131 to S135, the stacked autoencoder is one of a convolutional neural network, a residual neural network, or a long short-term memory network; the clustering layer is a self-defined module including three types of clustering centers, which respectively correspond to the best expression of three types of flow state signals in a feature space, the three types of clustering centers including a laminar flow clustering center, a transition clustering center, and a turbulent flow clustering center, the three types of flow state signals including a laminar flow signal, a transition signal, and a turbulent flow signal, and the best expression refers to the feature expression that can best map the three types of flow state time series signals in the feature space. For example, for all laminar flow signal samples, there is a clustering center in the feature space that can best measure and describe the common characteristics of laminar flow, which is the best expression of the laminar flow center in the feature space. The purpose of training the clustering layer is to approach the best expression as much as possible.

[0089] The loss function of the stacked autoencoder includes a reconstruction error measured by a mean square error function MSE, and the formula is:

[0090]

[0091] In the formula, N is the total number of samples, x i is the input feature of the heat flow signal, g(x i ) is the reconstructed input of the stacked autoencoder.

[0092] The loss function of the clustering layer is a KL divergence and a local preservation error, and the formula is:

[0093]

[0094]

[0095] sim(x i ,x j )=ED(x i ,x j )×cos(x i ,x j )

[0096] In the formula, q ij is the probability that sample i belongs to the jth type given by the clustering layer, p ij is an auxiliary target probability; sim(x i ,x j ) measures the similarity between sample i and surrounding sample j, ED(x i ,x j ) is the Euclidean distance between sample i and surrounding sample j, and cos(x i ,x j ) is the triangular similarity between sample i and surrounding sample j.

[0097] The flow regime clustering recognition accuracy acc of the test set is calculated by the following formula:

[0098]

[0099] In the formula, delta is an indicator function, y i is a flow regime label of the artificial flow regime, is a flow regime label recognized by the clustering model, and m is the total number of test set samples.

[0100] S104 combines the labeled test set samples to analyze the accuracy of the clustering model, and obtains a model that can be used for flow regime recognition of wind tunnel time series signals.

[0101] S104 is specifically: based on the test set samples, the flow regime clustering recognition accuracy of the model is evaluated, and it is judged whether the flow regime judgment accuracy of the trained converged model meets the expected requirement; if the flow regime clustering accuracy meets the requirement, a deep learning clustering model for hypersonic heat flow signal flow regime recognition is obtained; if the flow regime clustering accuracy does not meet the requirement, return to S102, modify the model initialization training parameters and re-execute S102 to S103 for training.

[0102] In order to better illustrate the scheme of the present application, an example is given as follows, as shown in Figures 2-3 The steps include the following steps:

[0103] First, data collection is carried out, samples are collected from wind tunnel data, and are divided into two parts of training set and test set. The test set is labeled by an artificial flow regime labeling method to provide labels, which is used to verify the flow regime recognition accuracy of the model after the model training is completed; the remaining samples are not provided with labels, and are used as the training set of unsupervised learning.

[0104] Data feature extraction: all samples are processed by discrete wavelet transform to obtain wavelet coefficients in multiple frequency bands; all sample wavelet coefficients are uniformly standardized to construct a data set for inputting a deep learning model.

[0105] Model construction: the unsupervised clustering model includes a stacked autoencoder and a clustering layer. The stacked autoencoder is composed of an encoder and a decoder, the encoder performs feature extraction and feature dimension reduction, and outputs low-dimensional features; the decoder restores the signal, and outputs the reconstructed input signal. The stacked autoencoder is characterized in that it is obtained by stacking multiple autoencoder layers. In this example, the basic structure of the autoencoder is selected to be a residual neural network. The clustering layer includes the cluster centers of three types of flow regimes, and provides the flow regime recognition label of the sample by calculating the similarity between the features of each sample and the cluster centers.

[0106] Pre-training: After the model is built, the stacked autoencoder is pre-trained. Each layer of the autoencoder is trained layer by layer, and the reconstruction error MSE convergence is observed. At this time, the deep autoencoder may not have reached the ideal convergence condition, but it does not affect the subsequent training. After each layer is trained independently, the complete autoencoder is stacked in order, and the reconstruction error MSE convergence is observed again. At this time, the convergence standard should be met. The reconstruction error MSE measures the difference between the input and output time. The smaller the value, the better the reconstruction ability of the stacked autoencoder, and the more representative the low-dimensional features extracted by the encoder.

[0107] Formal training: The stacked autoencoder extracts the clustering features of the input sample, and the clustering layer calculates the soft assignment probability based on the similarity between the features and the cluster centers. The clustering layer calculates the KL divergence error and the local preservation error based on the soft assignment probability and the auxiliary target probability. As the number of training rounds increases, the parameters of the stacked autoencoder and the clustering center parameters of the clustering layer are continuously optimized, and the values of the three loss functions are reduced, and finally a converged unsupervised clustering model is obtained.

[0108] Further optimization of model parameters: The accuracy of the model is measured by its performance on the labeled test set. The optimization goal is to achieve good flow pattern recognition accuracy on the test set. When the accuracy requirement cannot be met or the training loss function convergence is poor, the model is optimized by adjusting the initial structure parameters and training parameters of the model. The structure parameters include the number of stacked autoencoder layers, the feature dimension of each layer, etc. The training parameters include the number of training rounds, the learning rate, etc. During training, by changing a parameter while keeping the rest parameters unchanged, the effect of the parameter on the entire model can be studied. Through multiple experiments and empirical evaluation, the initial parameters are set as follows: the number of stacked autoencoder layers is 7, the feature dimension of each layer is [128, 64, 64, 32, 32, 16, 8], the pre-training round is 50, the formal training round is 120, and the learning rate is 1e-4.

[0109] Clustering accuracy evaluation method: The model is used for flow pattern recognition on the test set samples, and the flow pattern judged by the clustering model is compared with the manually labeled flow pattern label. The proportion of correctly recognized samples in the entire test set is calculated to obtain the clustering accuracy.

[0110] The loss function curve of the final training process and the flow pattern recognition accuracy curve of the test set with model training are shown in the figure. At about 60 rounds, the model loss function has converged, and the accuracy performance has converged, which can achieve an accuracy close to 90%, meeting the expectations.

[0111] In addition, the wavelet basis function used in wavelet transform also needs to be studied in detail. The following table shows the recognition accuracy of the unsupervised clustering model on the test set when different wavelet basis functions are used. As can be seen from the figure, db2 performs best among many wavelet bases, and is most suitable for extracting sharp mutation information in the heat flow time series signal, which is beneficial to the recognition of transition and turbulent flow regime.

[0112] Embodiment Two

[0113] A high-speed supersonic flow regime recognition system based on deep learning is used for the recognition of high-speed supersonic ground wind tunnel wall heat flow signal flow regime, realized by electronic device hardware with a central processing unit, and can be realized by personal computers, smart terminals, local area networks, servers, etc. In this embodiment, please refer to Figure 4 , which includes a heat flow time series signal measurement module 61, an input data set construction module 62, a model training module 63, and a model accuracy analysis module 64.

[0114] The heat flow time series signal measurement module 61 is used to obtain the heat flow time series signal of the model wall measured by the thin film resistance thermometer during the wind tunnel test.

[0115] The input data set construction module 62 is used to extract the time-frequency domain information in the heat flow time series signal using the wavelet transform method, and construct the input data set of the deep learning network.

[0116] The model training module 63 is used to build a deep learning framework for feature extraction and flow regime clustering, and train and optimize the input data set to obtain a deep learning clustering model with convergent training indicators.

[0117] The model accuracy analysis module 64 is used to analyze the accuracy of the clustering model combined with the labeled test data set samples, and obtain a model that can be used for wind tunnel time series signal flow regime recognition.

[0118] In this embodiment, the heat flow time series signal measurement module 61 is specifically used to set a unified thin film temperature sensor sampling frequency, distribute and install multiple thin film temperature sensors on different experimental model walls, perform wind tunnel experiments under high-speed supersonic flow, measure and collect time series heat flow signals of the experimental model wall, divide the time series heat flow signals into training data set and test data set, and give reliable flow regime labels after artificial statistical analysis of the test data set, to obtain labeled test set samples.

[0119] The input data set construction module 62 is specifically configured to: extract time domain features and frequency domain features of the time series heat flow signal by using a wavelet transform method, set a wavelet base function suitable for the time series heat flow signal to be analyzed, divide the time domain features and the frequency domain features of the time series heat flow signal into wavelet coefficients of different frequency bands and perform standardization processing, and obtain a new feature data set as an input data set for constructing a deep learning network.

[0120] The model training module 63 comprises:

[0121] The autoencoder training submodule 631 (not shown in the figure) is configured to import the input data set for constructing the deep learning network into a stacked autoencoder in the deep learning clustering model in an initial state, and perform layer-by-layer greedy pre-training on the stacked autoencoder to optimize parameters of layers of the stacked autoencoder.

[0122] The autoencoder iterative optimization submodule 632 (not shown in the figure) is configured to splice the pre-trained encoder layers in sequence to obtain a complete stacked autoencoder, and perform iterative optimization training on the stacked autoencoder by using the input data set to obtain a stacked autoencoder with feature extraction capability.

[0123] The loss function curve submodule 633 (not shown in the figure) is configured to combine the encoder part of the stacked autoencoder with feature extraction capability and the clustering layer, perform iterative optimization training on the autoencoder parameters and the clustering center parameters of the clustering layer based on the input data set, and record a loss function curve of the training process.

[0124] The judgment submodule 634 (not shown in the figure) is configured to judge the quality of the model obtained in the training process based on the loss function curve, and if the preset requirement is not met, return to execute the input data set construction module 62 to modify the model initialization training parameters, and if the preset requirement is met, execute the model accuracy analysis module.

[0125] The model accuracy analysis module 64 is specifically configured to evaluate the flow state clustering recognition accuracy of the model based on the test set samples, judge whether the flow state judgment accuracy of the trained model meets the expected requirement, if the flow state clustering accuracy meets the requirement, obtain a deep learning clustering model for hypersonic heat flow signal flow state recognition, and if the flow state clustering accuracy does not meet the requirement, return to the input data set construction module to modify the model initialization training parameters and re-execute the input data set construction module 62 and the model training module 63 for training.

[0126] In the embodiment, the stacked autoencoder is one of a convolutional neural network, a residual neural network or a long short-term memory network; the clustering layer is a self-defined module including three types of clustering centers, and the best expression refers to a characteristic expression that can best map three types of flow state time series in a characteristic space. For example, for all laminar flow signal samples, there is a clustering center in the characteristic space that can best measure and describe the common characteristics of laminar flow, which is the best expression of the laminar flow center in the characteristic space. The purpose of training the clustering layer is to approach the best expression as much as possible.

[0127] In the embodiment, the loss function of the stacked autoencoder includes a reconstruction error measured by a mean square error function (MSE), and the formula is as follows:

[0128]

[0129] In the formula, N is the total number of samples, x i is an input feature of the heat flow signal, g(x i ) is a reconstructed input of the stacked autoencoder.

[0130] The loss function of the clustering layer is a KL divergence error and a local preservation error, and the formula is as follows:

[0131]

[0132]

[0133] sim(x i ,x j )=ED(x i ,x j )×cos(x i ,x j )

[0134] In the formula, q ij is a probability that the sample i belongs to the jth type given by the clustering layer, p ij is an auxiliary target probability; sim(x i ,x j ) measures the similarity between the sample i and the surrounding sample j, ED(x i ,x j ) is the Euclidean distance between the sample i and the surrounding sample j, and cos(x i ,x j ) is the triangular similarity between the sample i and the surrounding sample j.

[0135] The flow state clustering recognition accuracy acc is calculated by the following formula:

[0136]

[0137] wherein δ is an indicator function, y i is a manual flow regime label, is a flow regime label identified by the clustering model, and m is the total number of test set samples.

[0138] The hyper-sonic flow regime identification system based on deep learning of the embodiment has the same implementation process, method and effect as the hyper-sonic flow regime identification method based on deep learning described in the first embodiment, and thus will not be described here again.

[0139] Embodiment three

[0140] The present application relates to a computer-readable storage medium, and the storage medium stores instructions. When the instructions are executed, the hyper-sonic flow regime identification method based on deep learning of the first embodiment is implemented. The implementation process, method and effect are the same as those of the hyper-sonic flow regime identification method based on deep learning described in the first embodiment, and thus will not be described here again.

[0141] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0142] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, which is made by using the content of the present application specification and drawings, is also included in the patent protection scope of the present application.

Claims

1. A hypersonic flow pattern identification method based on deep learning, characterized in that: Used for identification of thermal flow signal flow patterns on the wall of hypersonic ground wind tunnels, including: Step 1: Obtain the heat flow time series signal of the model wall measured by a thin film resistance thermometer in the wind tunnel test; Step 2 uses a wavelet transform method to extract time domain-frequency domain information from the heat flow time series signal and construct an input data set for a deep learning network; Step 3: Building a deep learning framework for feature extraction and flow state clustering, using the input data set to train and iteratively optimize model parameters to obtain a deep learning clustering model with converged training indicators; In step 4, the accuracy of the clustering model is analyzed in combination with the labeled test set samples to obtain a model that can be used for flow state identification of wind tunnel time series signals, wherein the labeled test set samples are reliable flow state labels given after manual statistical analysis.

2. The method for hypersonic flow pattern identification based on deep learning according to claim 1, characterized in that: The step 1 is specifically as follows: A uniform sampling frequency is set for thin-film temperature sensors, and multiple thin-film temperature sensors are distributedly installed on the walls of different experimental models. Wind tunnel experiments under hypersonic flow are conducted, and time-series heat flow signals from the experimental model walls are measured and collected. The time-series heat flow signals are divided into a training data set and a test data set. The test data set is manually statistically analyzed to provide reliable flow state labels, thereby obtaining labeled test set samples. The step 2 is specifically as follows: The time domain characteristics and frequency domain characteristics of the time series heat flow signal are extracted by wavelet transform method, and a wavelet basis function suitable for the heat flow time series signal to be analyzed is set. The time domain characteristics and frequency domain characteristics of the time series heat flow signal are divided into wavelet coefficients of different frequency bands and normalized to obtain a new feature data set as the input data set for constructing a deep learning network; The step 3 comprises: Step 31 imports the input data set for constructing the deep learning network into the stacked autoencoder in the deep learning clustering model in the initial state, and performs greedy pre-training on the stacked autoencoder layer by layer to optimize the parameters of the stacked autoencoder layer; Step 32: sequentially splicing the pre-trained stacked autoencoder layers to obtain a complete stacked autoencoder; iteratively optimizing and training the stacked autoencoder using the input data set to obtain a stacked autoencoder with feature extraction capability; Step 33 combines the encoder part of the stacked autoencoder with the feature extraction capability with the clustering layer, performs iterative optimization training on the autoencoder parameters and the clustering center parameters of the clustering layer based on the input data set, and records the loss function curve of the training process; Step 34 determines the quality of the model obtained during the training process based on the loss function curve; if it does not meet the preset requirements, return to step 2 and modify the model initialization training parameters; if it meets the preset requirements, proceed to step 4; The step 4 is specifically as follows: based on the test set samples, the flow state clustering identification accuracy of the model is evaluated to determine whether the flow state judgment accuracy of the training convergence model meets the expected requirements; if the flow state clustering accuracy meets the requirements, a deep learning clustering model for flow state identification of hypersonic thermal flow signals is obtained; if the flow state clustering accuracy does not meet the requirements, returning to step 2, modifying the model initialization training parameters and re-executing steps 2 to 3 for training.

3. The method for hypersonic flow pattern identification based on deep learning according to claim 2, characterized in that: The iterative optimization training content described in step 3 is: during the training process, change one of the network parameters, fix the other parameters unchanged, record the training results, and adjust the optimal value of the parameter by observing whether the flow state discrimination accuracy of the model on the test set meets the preset requirements; replace and adjust different network parameters one by one, repeat the above process until the combination of all the network parameters reaches the optimal value, the optimization training is completed, and the flow state clustering model is obtained.

4. The method for hypersonic flow pattern identification based on deep learning according to claim 2, characterized in that: The time series signal data samples collected from different experiments in step 1 are all cut from the original measurement signal, and the lengths of all the time series signal data samples remain consistent.

5. The method for hypersonic flow pattern identification based on deep learning according to claim 2, characterized in that: The stacked autoencoder is one of a convolutional neural network, a residual neural network or a long short-term memory network; The clustering layer is a custom module containing three types of cluster centers, and the cluster centers correspond to the optimal expressions of three types of flow state signals in the feature space, respectively. The three types of cluster centers include laminar cluster centers, transition cluster centers, and turbulent cluster centers. The three types of flow state signals include laminar signals, transition signals, and turbulent signals. The loss function includes multiple stacked autoencoders, which is the reconstruction error, measured by the mean square error function MSE, and the formula is: Where N is the total number of samples, x i is the input feature of the heat flow signal, g(x i ) is the reconstructed input of the stacked autoencoder; The loss function of the clustering layer is KL divergence error and local preservation error, and the formula is: yes(x) i ,x j )=ED(x i ,x j )×cos(x i ,x j ) Where q ij The probability that sample i given by the clustering layer belongs to the jth class, p ij is the auxiliary target probability; sim(x i ,x j ) measures the similarity between sample i and surrounding samples j, ED(x i ,x j ) is the Euclidean distance between sample i and surrounding sample j, cos(x i ,x j ) is the triangular similarity between sample i and surrounding samples j, L lp is the local retention error, L kl : Kullback-Leible divergence error, Z i is the expression of sample i in the feature space, Z j is the expression of sample j in the feature space, KL(P||Q) is the expression of KL divergence error, KL represents the KL divergence operator, and P||Q represents the difference between probability distributions P and Q measured with probability distribution Q as a priori; The flow state cluster identification accuracy acc is calculated using the following formula: Where δ is the indicator function, y i It is an artificial flow label. is the flow state label identified by the clustering model and m is the total number of test set samples.

6. A hypersonic flow pattern identification system based on deep learning, characterized in that: Used for identification of thermal flow signal flow patterns on the wall of hypersonic ground wind tunnels, including: The heat flow timing signal measurement module is used to obtain the heat flow timing signal of the model wall measured by a thin film resistance thermometer in the wind tunnel test; An input data set construction module is used to extract time domain-frequency domain information from the heat flow time series signal using a wavelet transform method to construct an input data set for a deep learning network; A model training module is used to build a deep learning framework for feature extraction and flow state clustering, and to train and iteratively optimize model parameters using the input data set to obtain a deep learning clustering model with converged training indicators; The model accuracy analysis module is used to perform accuracy analysis on the clustering model in combination with labeled test data set samples to obtain a model that can be used for flow state identification of wind tunnel timing signals.

7. The hypersonic flow pattern identification system based on deep learning according to claim 6, characterized in that: The heat flow timing signal measurement module is specifically used to: A uniform sampling frequency is set for thin-film temperature sensors, and multiple thin-film temperature sensors are distributedly installed on the walls of different experimental models. Wind tunnel experiments under hypersonic flow are conducted, and time-series heat flow signals from the experimental model walls are measured and collected. The time-series heat flow signals are divided into a training data set and a test data set. The test data set is manually statistically analyzed to provide reliable flow state labels, thereby obtaining labeled test set samples. The input data set construction module is specifically used to: The time domain characteristics and frequency domain characteristics of the time series heat flow signal are extracted by wavelet transform method, and a wavelet basis function suitable for the heat flow time series signal to be analyzed is set. The time domain characteristics and frequency domain characteristics of the time series heat flow signal are divided into wavelet coefficients of different frequency bands and normalized to obtain a new feature data set as the input data set for constructing a deep learning network; The model training module includes: An autoencoder training submodule is used to import the input data set for constructing the deep learning network into the stacked autoencoder in the deep learning clustering model in the initial state, and perform greedy pre-training on the stacked autoencoder layer by layer to optimize the parameters of the stacked autoencoder layer; An autoencoder iterative optimization submodule is used to sequentially splice the pre-trained encoder layers to obtain a complete stacked autoencoder; the stacked autoencoder is iteratively optimized and trained using the input data set to obtain a stacked autoencoder with feature extraction capabilities; a loss function curve submodule, configured to combine the encoder portion of the stacked autoencoder with the feature extraction capability with the clustering layer, perform iterative optimization training on the autoencoder parameters and the clustering center parameters of the clustering layer based on the input data set, and record the loss function curve of the training process; A judgment submodule is used to judge the quality of the model obtained during the training process based on the loss function curve; if it does not meet the preset requirements, it returns to execute the input data set construction module and modify the model initialization training parameters; if it meets the preset requirements, it executes the model accuracy analysis module; The model accuracy analysis module is specifically used to: evaluate the flow state clustering identification accuracy of the model based on the test set samples, and determine whether the flow state judgment accuracy of the training convergence model meets the expected requirements; if the flow state clustering accuracy meets the requirements, obtain a deep learning clustering model for hypersonic thermal flow signal flow state identification; if the flow state clustering accuracy does not meet the requirements, return to the input data set construction module, modify the model initialization training parameters and re-execute the input data set construction module and the model training module for training.

8. The hypersonic flow pattern identification system based on deep learning according to claim 7, characterized in that: The stacked autoencoder is one of a convolutional neural network, a residual neural network or a long short-term memory network; The clustering layer is a custom module containing three types of cluster centers, and the cluster centers correspond to the optimal expressions of three types of flow state signals in the feature space, respectively. The three types of cluster centers include laminar cluster centers, transition cluster centers, and turbulent cluster centers. The three types of flow state signals include laminar signals, transition signals, and turbulent signals. The loss function includes multiple stacked autoencoders, which is the reconstruction error, measured by the mean square error function MSE, and the formula is: Where N is the total number of samples, x i is the input feature of the heat flow signal, q(x i ) is the reconstructed input of the stacked autoencoder; The loss function of the clustering layer is KL divergence error and local preservation error, and the formula is: yes(x) i ,x j )=ED(x i ,x j )×cos(x i ,x j ) Where q ij The probability that sample i given by the clustering layer belongs to the jth class, p ij is the auxiliary target probability; sim(x i ,x j ) measures the similarity between sample i and surrounding samples j, ED(x i ,x j ) is the Euclidean distance between sample i and surrounding sample j, cos(x i ,x j ) is the triangular similarity between sample i and surrounding samples j; The flow state cluster identification accuracy acc is calculated using the following formula: Where δ is the indicator function, y i It is an artificial flow label. is the flow state label identified by the clustering model and m is the total number of test set samples.

9. A computer-readable storage medium, characterized in that: The storage medium stores instructions, which, when executed, execute a hypersonic flow state identification method based on deep learning according to any one of claims 1 to 5.

Images