A method of pneumatic signal recognition

Abstract

The application discloses a kind of pneumatic signal identification methods, it is characterized in that, comprising the following steps: S1: signal is obtained as training sample from the effective test phase of pneumatic force measuring system FMS;S2: output source domain signal and target domain signal with smooth trend characteristics;S3: feature migration is carried out using cross-entropy loss function, reduce the gap between source domain signal and target domain signal;S4: the feature of source domain signal is converted and input source domain feature, convolutional neural network utilizes adaptive EMD and carries out dynamic compensation, signal reconstruction is carried out to source domain signal, and inertia component in source domain signal is inhibited;S5: according to the gap between predicted aerodynamic force, reduce the gap between real aerodynamic force and predicted aerodynamic force, output predicted pneumatic signal.The application can accurately identify pneumatic signal, and can effectively inhibit other interference signals, provide key technology and data support for accurate evaluation high enthalpy, hypersonic vehicle aerodynamic characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aerodynamic technology, and particularly relates to an aerodynamic signal recognition method. BACKGROUND

[0002] The force test in a wind tunnel is an important part of the research and development of a hypersonic aircraft. With the development of this technology, the large-scale of the test model has become a trend of the hypersonic wind tunnel test. In the effective test time of several hundred milliseconds, the stiffness reduction of the large-scale force measurement system can seriously cause the poor accuracy of the aerodynamic force identification. The large-scale of the test model brings challenges to the accurate aerodynamic force identification of the short-time pulse combustion wind tunnel. The existing aerodynamic identification includes the following two aspects: 1. The aerodynamic force identification based on filtering or time-frequency transformation, and the accuracy of the aerodynamic force identification is not enough; and 2. The deep learning aerodynamic force identification, and a large number of training samples are needed to ensure the performance of the model. SUMMARY

[0003] In view of the above problems of the prior art, the present application provides an aerodynamic signal recognition method with improved recognition accuracy.

[0004] To achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows:

[0005] The present application provides an aerodynamic signal recognition method, which comprises the following steps:

[0006] S1: obtaining signals from the effective test stage of an aerodynamic force measurement system (FMS) as training samples, wherein the training samples include source domain samples obtained from a scaled bench test and target domain samples obtained from a shock tunnel test;

[0007] S2: inputting the source domain samples and the target domain samples into a residual attention module of a convolutional neural network respectively, suppressing the aerodynamic noise and the quantization noise of the source domain samples and the target domain samples, and outputting source domain signals and target domain signals with smooth trend characteristics;

[0008] S3: performing feature migration by using a cross-entropy loss function, and reducing the gap between the source domain signals and the target domain signals;

[0009] S4: converting the features of the source domain signals into source domain features, performing dynamic compensation by the convolutional neural network using adaptive EMD, reconstructing the source domain signals, and suppressing the inertial components in the source domain signals;

[0010] S5: according to the gap between the predicted aerodynamic forces, reducing the gap between the real aerodynamic forces and the predicted aerodynamic forces, and outputting the predicted aerodynamic signals.

[0011] Further, the method for suppressing the aerodynamic noise and the quantization noise of the source domain samples and the target domain samples by the residual attention module in step S2 is as follows:

[0012]

[0013] wherein, W l and b l are the weight and offset of the lth layer of the convolution kernel in the convolutional neural network, respectively, is the symbol of the convolution operation; after the convolution operation, the feature of the (l-1)th layer is multiplied by the adaptive threshold matrix γ of each channel to obtain the feature of the lth layer, R is the channel of the convolutional neural network, and t is the input signal of the convolutional layer.

[0014] Further, step S3 comprises:

[0015] S31: using the softmax activation function in the input layer of the convolutional neural network, so that the cross-entropy loss function pays attention to the difference in signal trend, rather than the difference in amplitude;

[0016] S32: converting the aerodynamic trend of the source domain signal and the target domain signal into a probability distribution with a sum of 1:

[0017]

[0018] wherein, i and j are the number of data points contained in the signal, t i is the true value of the source domain signal or the target domain signal, t j is the predicted value of the source domain signal or the target domain signal, e is a constant, and f softmax (·) is the softmax activation function;

[0019] S33: let the softmax value of the signal output of the source domain sample be t S , and the softmax value of the signal output of the target domain sample be t T , and use the cross-entropy loss function to express the gap between the signal of the source domain sample and the signal of the target domain sample as the source domain feature difference

[0020]

[0021] wherein, m is the number of samples of the source domain signal and the target domain signal, T is the target domain data, S is the source domain data, and k is any sample of the source domain signal and the target domain signal.

[0022] Further, the method for the convolutional neural network to use adaptive EMD to dynamically compensate the source domain signal in step S4 is:

[0023]

[0024] wherein, These are the output features of the (l-1)th layer in a convolutional neural network. It is the output feature of the l-th layer in the convolutional neural network, (H EMD (t1)+H EMD (t2),...,+H EMD (t l-1 )) / (l-1) is the reconstructed source domain signal from the first layer to the (l-1)th layer in the adaptive EMD.

[0025] Further, step S5 includes:

[0026] S51: The aerodynamic trend characteristics of the source and target domain signals are mapped to the regenerating kernel Hilbert space through a high-dimensional mapping function, and the similarity of aerodynamic trends is evaluated using MMD loss.

[0027]

[0028] Where n is the number of samples. and These are the network output signal and the actual aerodynamic signal from the source domain, respectively. The MMD loss is the difference between the output signal of the network and the real aerodynamic signal in the source domain, i.e., the difference in the extraction of trend features in the source domain. MMD is the supremum of the expected difference between the two domain features in the Hilbert space of the regenerating kernel.

[0029] D52: Utilizing source domain feature differences Differences in source domain trend feature extraction Establish optimization objectives and output predicted aerodynamic signals:

[0030] Attached Figure Description

[0031] Figure 1 This is a flowchart of a pneumatic signal recognition method. Detailed Implementation

[0032] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0033] like Figure 1 As shown, the pneumatic signal identification method of this scheme includes the following steps:

[0034] S1: obtaining signals from an effective test phase of a pneumatic force measurement system FMS as training samples, the training samples including source domain samples obtained from a scale test bench and target domain samples obtained from a shock tunnel wind test;

[0035] S2: inputting the source domain samples and the target domain samples into a residual attention module of a convolutional neural network respectively, suppressing pneumatic noise and quantization noise of the source domain samples and the target domain samples, and outputting source domain signals and target domain signals with smooth trend characteristics;

[0036] The method for the residual attention module in step S2 to suppress the pneumatic noise and the quantization noise of the source domain samples and the target domain samples is as follows:

[0037]

[0038] wherein, W l and b l are a weight and an offset of a convolution kernel in the lth layer of the convolutional neural network, is a symbol of a convolution operation; after the convolution operation, the feature of the (l-1)th layer is multiplied by an adaptive threshold matrix γ of each channel to obtain the feature of the lth layer, R is a channel of the convolutional neural network, and t is an input signal of the convolution layer.

[0039] S3: performing feature migration using a cross-entropy loss function to reduce the gap between the source domain signals and the target domain signals; step S3 includes:

[0040] S31: using a softmax activation function in an input layer of the convolutional neural network to make the cross-entropy loss function focus on the difference in signal trend rather than the difference in amplitude;

[0041] S32: converting the pneumatic trend of the source domain signals and the target domain signals into a probability distribution with a sum of 1:

[0042]

[0043] wherein, i and j are both the number of data points contained in the signals, t i is a true value of the source domain signals or the target domain signals, t j is a predicted value of the source domain signals or the target domain signals, e is a constant, and f softmax (·) is a softmax activation function;

[0044] S33: setting the softmax value of the signal output of the source domain samples as t S , and setting the softmax value of the signal output of the target domain samples as t TThe cross-entropy loss function is used to express the difference between the signals of the source domain samples and the target domain samples, as a source domain feature difference.

[0045]

[0046] Where m is the number of samples of the source domain signal and the target domain signal, T is the target domain data, S is the source domain data, and k is any sample of the source domain signal and the target domain signal.

[0047] S4: The feature transformation of the source domain signal is input into the source domain feature. The convolutional neural network uses adaptive EMD for dynamic compensation to reconstruct the source domain signal and suppress the inertial component in the source domain signal. The method of dynamic compensation of the source domain signal by the convolutional neural network using adaptive EMD in step S4 is as follows:

[0048]

[0049] in, These are the output features of the (l-1)th layer in a convolutional neural network. It is the output feature of the l-th layer in the convolutional neural network, (H EMD (t1)+H EMD (t2),...,+H EMD (t l-1 )) / (l-1) is the reconstructed source domain signal from the first layer to the (l-1)th layer in the adaptive EMD.

[0050] S5: Based on the difference between the predicted and actual aerodynamic forces, reduce the gap between the actual and predicted aerodynamic forces, and output the predicted aerodynamic signal. Step S5 includes:

[0051] S51: The aerodynamic trend characteristics of the source and target domain signals are mapped to the regenerating kernel Hilbert space through a high-dimensional mapping function, and the similarity of aerodynamic trends is evaluated using MMD loss.

[0052]

[0053] Where n is the number of samples. and These are the network output signal and the actual aerodynamic signal from the source domain, respectively. The MMD loss is the difference between the output signal of the network and the real aerodynamic signal in the source domain, i.e. the difference in the extraction of trend features in the source domain. MMD is the upper bound of the expected difference between the two domain features in the Hilbert space of the regeneration kernel.

[0054] D52: Utilizing source domain feature differences Differences in source domain trend feature extraction Establish optimization objectives and output predicted aerodynamic signals:

[0055]

[0056] The application identifies and filters the inertial force signal and the instrument noise signal in the FMS output signal in the wind tunnel test, so as to obtain the 'pure' aerodynamic force. The design difficulty of the strain gauge balance with high precision, low cost and mature development is greatly reduced, and the precision of the pulse combustion wind tunnel aerodynamic test is improved. Compared with the existing aerodynamic intelligent identification model based on deep learning technology, the application first introduces the transfer learning idea into the aerodynamic intelligent identification model, effectively alleviating the problem of intelligent model training sample shortage caused by high cost of wind tunnel test. In order to improve the identification precision of the aerodynamic force signal, the soft threshold residual attention block is introduced to filter the instrument noise signal, and the adaptive EMD dense block is innovatively proposed to greatly filter the inertial force noise introduced by the aircraft test model. Both the aerodynamic signal can be accurately identified, and other interference signals can be effectively suppressed, which provides key technology and data support for accurately evaluating the aerodynamic characteristics of high-enthalpy and hypersonic aircraft.

Claims

1. A method of aerodynamic signal recognition, characterized by, The method comprises the following steps: S1: obtaining signals from an effective test stage of a pneumatic force measurement system FMS as training samples, the training samples comprising source domain samples obtained from a scaled bench test and target domain samples obtained from a shock tunnel wind test; S2: inputting the source domain samples and the target domain samples into a residual attention module of a convolutional neural network respectively, suppressing pneumatic noise and quantization noise of the source domain samples and the target domain samples, and outputting source domain signals and target domain signals with smooth trend characteristics; S3: performing feature migration using a cross-entropy loss function to reduce the gap between the source domain signals and the target domain signals; S4: converting features of the source domain signals into source domain features, and performing dynamic compensation on the source domain signals by the convolutional neural network using adaptive EMD to reconstruct the source domain signals and suppress inertial components in the source domain signals; The method for performing dynamic compensation on the source domain signals by the convolutional neural network using adaptive EMD in the step S4 is as follows: in, It is the first in convolutional neural networks The output features of the layer It is the first in convolutional neural networks The output features of the layer For adaptive EMD from layer 1 to layer 2 The reconstructed source domain signal of the layer; S5: reducing the gap between the real aerodynamic force and the predicted aerodynamic force according to the gap between the predicted aerodynamic forces, and outputting predicted aerodynamic signals.

2. The aerodynamic signal recognition method of claim 1, wherein, The method for suppressing pneumatic noise and quantization noise of the source domain samples and the target domain samples by the residual attention module in the step S2 is as follows: in, and These are the first convolutional neural network units. The weights and offsets of the convolutional kernel. The symbol for convolution operation; after convolution, the first... Layer characteristics With the adaptive threshold matrix of each channel Multiply to get the first... Layer characteristics , R For convolutional neural network channels, t This is the input signal for the convolutional layer.

3. The aerodynamic signal recognition method of claim 1, wherein, The step S3 comprises: S31: using a softmax activation function in an input layer of the convolutional neural network, so that the cross-entropy loss function focuses on differences in signal trends rather than differences in amplitudes; S32: converting aerodynamic trends of the source domain signals and the target domain signals into probability distributions with a sum of 1: wherein, i and j are the number of data points contained in the signal, t i is the true value of the source domain signal or the target domain signal, t j is the predicted value of the source domain signal or the target domain signal, e is a constant, is a softmax activation function; S33: Set the softmax value of the signal output of the source domain sample as , the softmax value of the signal output of the target domain sample as , and express the gap between the signal of the source domain sample and the signal of the target domain sample by using a cross-entropy loss function as a source domain feature difference : wherein, m is the number of samples of the source domain signal and the target domain signal, T is the target domain data, S is the source domain data, k is a sample of any of the source domain signal and the target domain signal.

4. The aerodynamic signal recognition method of claim 1, wherein, The step S5 comprises: S51: mapping aerodynamic trend characteristics of the source domain signals and the target domain signals to a reproducing kernel Hilbert space through a high-dimensional mapping function, and using an MMD loss to evaluate the similarity of aerodynamic trends: wherein, n is the number of samples, and are the network output signal and the source domain real aerodynamic signal, respectively, is the MMD loss of the output network output signal and the source domain real aerodynamic signal, i.e. the source domain trend feature extraction difference, MMD being the upper bound of the expected difference of two domain features in a reproducing kernel Hilbert space. S52: exploit source domain feature discrepancy and source domain trend feature extraction discrepancy establish optimization objective, output predicted aerodynamic signal: 。

Images