FT-IR spectrum enhancement classification method for aero-engine thermal jet

Abstract

The invention provides an aero-engine thermal jet flow FT-IR spectrum enhancement classification method, and the method carries out the infrared spectrum detection of six different types of aero-engine tail jet thermal jet flows including two types of turbojet and turbofan, serves as the basis of engine recognition and data source input, and is more scientific in classification. Meanwhile, aiming at the problem that the CNN method has overfitting on the existing spectral data set, the invention provides an FTIR-SpectraGAN method, new data is generated through the existing data, the classification capability of the network in various aspects is improved, and higher classification accuracy is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent processing of infrared spectral data, and particularly relates to a method for enhancing classification of FT-IR spectra of aero-engine hot jets. Background Art

[0002] Aircraft fault detection is a key issue in the aviation field and is crucial for ensuring flight safety. With the development of infrared spectral technology, classifying the infrared spectra of engine hot jets is one of the key technologies for fault detection under remote sensing conditions.

[0003] Infrared spectroscopy is a commonly used method for quantitative analysis of substances and identification of compound structures. When a substance is irradiated with infrared radiation, the molecules selectively absorb radiation at specific frequencies, causing molecular vibrations or rotations, resulting in changes in the dipole moment, transitions of corresponding energy levels, and weakening of the transmitted light intensity corresponding to these absorption regions, which can be recorded as an infrared spectrum of wavenumber and transmittance.

[0004] Through non-contact measurement by an FT-IR (Fourier Transform Infrared) spectrometer, the infrared radiation emitted by hot substances in the aero-engine hot jet can be measured to obtain a spectrum containing different frequency peaks. The hot jet infrared spectrum represents the transitions between different energy levels of chemical substances. A gas is composed of atoms and molecules, and the energy of both is the sum of translational energy and internal energy, where the internal energy consists of a series of specific and discrete energy levels. A schematic diagram of the quantized energy levels of a molecule is as Figure 1 shown, Figure 1 where V represents the energy level and J represents the energy.

[0005] The selective absorption of infrared radiation caused by molecular vibration and rotation is manifested as specific absorption peaks of the substance for infrared radiation of different wavelengths. The selection law that needs to be satisfied for a molecule to absorb infrared radiation and undergo a transition can be obtained by solving the Schrödinger equation for the energy value E_V of the V energy level:

[0006]

[0007] where h is Planck's constant, f is the vibration frequency, and v i is the vibrational quantum number of the i-th mode. The energy difference between two adjacent energy levels is ΔE V = hf.

[0008] The selective absorption of different molecules to infrared radiation can be used to determine the molecular structure by analyzing their unique absorption peaks. Similarly, different types of aeroengines will produce hot jets with different gas components and emissions after combustion. The molecular substances of the hot jets form specific infrared spectra through vibration-rotation, and the analysis of the infrared spectral characteristics can be used to analyze and judge the chemical bonds and structures of substances, thereby realizing the classification of aeroengines.

[0009] Due to the lack of publicly available datasets, CNN models have demonstrated their powerful capabilities in spectral classification. However, there are contradictions between the limited samples in existing datasets and the problem that existing models are difficult to learn new features to further improve the classification accuracy after reaching a certain classification accuracy. There are contradictions between the demand of deep learning methods for a large number of learnable parameters and the limited samples in existing datasets, and existing models are difficult to learn new features to further improve the classification accuracy after reaching a certain classification accuracy. Therefore, how to improve the model performance under the condition of limited sample quantity has become the research focus of this patent. The same challenge as in hyperspectral image classification (Li) is the high measurement cost of the labeled dataset. Solutions include data augmentation, transfer learning, and unsupervised / semi-supervised feature learning. The limitation of the dataset size will affect the generalization ability of the spectral classification model. Data augmentation can improve the model's adaptability to data under different conditions, enhance the model's generalization ability, improve the model's robustness, and improve the training efficiency. Therefore, this patent has investigated data augmentation methods: Shi proposed a spectral encoding enhanced representation (SEER) method based on phase vectors for enhancing specific and subtle spectral differences when enhancing hyperspectral fluorescence images. Nalepa performed enhancement in the inference stage and constructed an enhancement method based on principal component analysis. GANs are an effective method for data augmentation. Wang reviewed the current development trends of GANs. The basic GAN models include GAN, CGAN, DCGAN, SNGAN, and styleGAN. In HSIC, He used a three-dimensional bilateral filter (3DBF) to extract spectral-spatial features and improved the classification performance by combining GANs with semi-supervised learning. Zhan designed a semi-supervised framework (HSGAN) for HSI data using one-dimensional GAN. Zhu used 1D-GAN and 3D-GAN for hyperspectral classification. In addition to the GAN method, Li used common data augmentation (flipping, rotation, and noise) to reach the upper limit of the enhancement effect when the sample size tripled, and designed a pixel-block pair convolutional neural network (PBP-CNN) to improve the performance of the classification network by extracting PBP features and combining decision fusion. The three-layer convolutional layer CNN has been proven to be effective for spectral data and can achieve a spectral classification accuracy of 96%. However, the CNN has an overfitting problem in the current spectral dataset. Summary of the Invention

[0010] To solve the above problems, the present invention provides a method for enhancing the classification of FT-IR spectra of hot jets of aero-engines, which can improve the classification accuracy and is more scientific in classification.

[0011] A method for enhancing the classification of FT-IR spectra of hot jets of aero-engines includes the following steps:

[0012] Use an FTIR spectrometer to obtain the spectrum to be measured of the hot jet of the aero-engine to be measured;

[0013] Input the spectrum to be measured into the trained discriminant network. The discriminant network outputs the probabilities corresponding to all possible categories to which the aero-engine that ejects the hot jet of the aero-engine to be measured belongs, and outputs the determination result of whether the spectrum to be measured is the spectrum of the hot jet of the aero-engine. At the same time, the category with the highest probability is used as the category to which the hot jet of the aero-engine to be measured belongs.

[0014] Further, the discriminant network includes a discriminator module and a classification module; among them, the discriminator module includes a first convolutional layer, a first LeakyReLU activation layer, a first dropout layer, a second convolutional layer, a second LeakyReLU activation layer, a second dropout layer, and a third convolutional layer connected in sequence; the classification module includes a Flatten layer, a Dense layer, a first branch Dense layer, and a second branch Dense layer;

[0015] The first convolutional layer is used to perform the first feature extraction on the input spectrum to be measured to obtain the first feature map;

[0016] The first LeakyReLU activation layer is used to perform activation processing on the first feature map to obtain the first activation feature map;

[0017] The first dropout layer is used to perform regularization processing on the first activation feature map to obtain the first regularized feature map;

[0018] The second convolutional layer is used to perform the second feature extraction on the first regularized feature map to obtain the second feature map;

[0019] The second LeakyReLU activation layer is used to perform activation processing on the second feature map to obtain the second activation feature map;

[0020] The second dropout layer is used to perform regularization processing on the second activation feature map to obtain the second regularized feature map;

[0021] The third convolutional layer is used to perform the third feature extraction on the second regularized feature map to obtain the third feature map;

[0022] The Flatten layer is used to convert the third feature map into a one-dimensional vector;

[0023] The Dense layer is used to perform a linear transformation on the one-dimensional vector to obtain a one-dimensional linear vector;

[0024] The first-branch Dense layer is used to perform a non-linear transformation on the one-dimensional linear vector based on the sigmoid activation function to obtain a first non-linear vector; wherein, each element in the first non-linear vector is the probability corresponding to all possible categories to which the aero-engine ejecting the hot jet to be measured belongs;

[0025] The second-branch Dense layer is used to perform a non-linear transformation on the one-dimensional linear vector based on the softmax activation function to obtain a second non-linear vector; wherein, each element in the second non-linear vector is used to characterize whether the spectrum to be measured is the spectrum of the aero-engine hot jet.

[0026] Further, the training spectra used when training the discriminant network include the true spectra of the hot jets ejected by aero-engines of each category obtained by an FTIR spectrometer, and the pseudo-spectra used to simulate the true spectra generated by the generator network.

[0027] Further, the generator network includes a third-branch Dense layer, a third LeakyReLU activation layer, a recombination and noise addition layer, a first transposed convolutional layer, a fourth LeakyReLU activation layer, a second transposed convolutional layer, a fifth LeakyReLU activation layer, a third transposed convolutional layer, a sixth LeakyReLU activation layer, a fourth transposed convolutional layer, a seventh LeakyReLU activation layer, a fifth transposed convolutional layer, an eighth LeakyReLU activation layer, and a sixth transposed convolutional layer connected in series;

[0028] The third-branch Dense layer is used to perform batch normalization on the input random noise to obtain normalized noise;

[0029] The recombination and noise addition layer is used to recombine and add noise to the normalized noise using Gaussian noise to obtain recombined Gaussian noise;

[0030] The recombined Gaussian noise sequentially passes through the first transposed convolutional layer, the fourth LeakyReLU activation layer, the second transposed convolutional layer, the fifth LeakyReLU activation layer, the third transposed convolutional layer, the sixth LeakyReLU activation layer, the fourth transposed convolutional layer, the seventh LeakyReLU activation layer, the fifth transposed convolutional layer, and the eighth LeakyReLU activation layer for multiple feature extraction and multiple activation processing to obtain a noise feature map;

[0031] The sixth transposed convolutional layer performs activation processing on the noise feature map based on the Tanh activation function to obtain a pseudo-spectrum.

[0032] Further, the discriminative network and the generator network together form a spectral generative adversarial network. The training method of the generative adversarial network is as follows:

[0033] Step 1: Use the real spectra of the hot jets ejected by aero-engines of each category obtained by the FTIR spectrometer as the input of the discriminative network. Take the real category of the aero-engine corresponding to each real spectrum and the real determination result corresponding to the real spectrum as the theoretical output of the discriminative network. Construct a discriminative loss function based on the predicted category and predicted determination result of the actual output of the discriminative network and the theoretical output, and perform backpropagation on the discriminative network to adjust the network parameters of the discriminative network, thereby completing the first training of the discriminative network and obtaining an initial discriminative network.

[0034] Step 2: Use different random noises as the input of the generator network, and the generator network outputs a pseudo-spectrum. Input the pseudo-spectrum into the initial discriminative network, and the initial discriminative network outputs a predicted determination result on whether the pseudo-spectrum is the spectrum of the aero-engine hot jet. Construct a generator loss function based on the predicted determination result and the expected determination result, and perform backpropagation on the generator network to adjust the network parameters of the generator network, thereby completing one training of the generator network.

[0035] Step 3: After repeating Step 2 at least 5 times, obtain an intermediate generator network.

[0036] Step 4: Use different random noises as the input of the intermediate generator network, and the intermediate generator network outputs a pseudo-spectrum. Input the pseudo-spectrum output by the intermediate generator network and the real spectra of the hot jets ejected by aero-engines of each category obtained by the FTIR spectrometer into the initial discriminative network, and the initial discriminative network outputs the predicted category of the aero-engine corresponding to all intermediate pseudo-spectra and real spectra, and the predicted determination result on whether all intermediate pseudo-spectra and real spectra are the spectra of the aero-engine hot jet.

[0037] Step 5: Construct a discriminative loss function based on the predicted category obtained in Step 4 and the real category corresponding to all intermediate pseudo-spectra and real spectra, and the predicted determination result and the real determination result corresponding to all intermediate pseudo-spectra and real spectra, and perform backpropagation on the discriminative network to adjust the network parameters of the discriminative network, thereby completing the second training of the discriminative network and obtaining an updated discriminative network.

[0038] Step 6: Use the updated discriminative network obtained in Step 5 to re-execute Steps 2 to 5 until both the generator loss function and the discriminative loss function are less than the set threshold. At this time, the corresponding discriminative network and generator network together form the final spectral generative adversarial network.

[0039] Furthermore, the loss function Loss for training the discriminant network D is as follows:

[0040] Loss D = αLoss validity + βLoss class

[0041] where Loss class represents the class loss function related to the category to which the aero-engine belongs, and Loss validity represents the authenticity loss function related to whether the spectrum is the spectrum of the aero-engine hot jet. α and β are the weight coefficients corresponding to the authenticity loss function and the class loss function respectively;

[0042] The authenticity loss function Loss validity is as follows:

[0043] Loss validity = Loss real + Loss fake

[0044]

[0045] where Loss real represents the real spectrum loss, Loss fake represents the fake spectrum loss, D(x) represents the prediction result of the discriminator network on whether the real spectrum sample x obtained by the FTIR spectrometer is the spectrum of the aero-engine hot jet, and D(x) = 0 means that the real spectrum sample is not the spectrum of the aero-engine hot jet, D(x) = 1 means that the real spectrum sample is the spectrum of the aero-engine hot jet, p data (x) represents the distribution that the real spectrum sample x obtained by the FTIR spectrometer conforms to, G(z) represents the fake spectrum sample generated by the generator network, D(G(z)) represents the prediction result of the discriminator network on whether the generated fake spectrum sample is the spectrum of the aero-engine hot jet, and D(G(z)) = 0 means that the fake spectrum sample is not the spectrum of the aero-engine hot jet, D(G(z)) = 1 means that the fake spectrum sample is the spectrum of the aero-engine hot jet, p z (z) represents the distribution that the random noise z input by the generator network conforms to, and E[·] represents the expectation;

[0046] The class loss function Loss class is as follows:

[0047]

[0048] Among them, CrossEntropy(·) represents the cross-entropy function, y represents the true category of the aero-engine corresponding to the true spectrum, and C(x) represents the predicted category of the discriminant network for the true spectrum.

[0049] Furthermore, the loss function Loss of the training generator network G is:

[0050]

[0051] Among them, G(z) represents the pseudo-spectrum sample generated by the generator network, D(G(z)) represents the prediction and determination result of the discriminator network on whether the generated pseudo-spectrum sample is the spectrum of the aero-engine hot jet, and D(G(z)) = 0 indicates that the pseudo-spectrum sample is not the spectrum of the aero-engine hot jet, D(G(z)) = 1 indicates that the pseudo-spectrum sample is the spectrum of the aero-engine hot jet, and p z (z) represents the distribution that the random noise z input to the generator network conforms to, represents the expectation.

[0052] Beneficial effects:

[0053] 1. The present invention provides a method for enhancing the classification of FT-IR spectra of aero-engine hot jets, which conducts infrared spectral detection on the tail hot jets of six different models of aero-engines including turbojet and turbofan, and uses this as the basis and data source input for engine identification, making it more scientifically classified; at the same time, aiming at the problem of overfitting of the CNN method in the existing spectral dataset, the present invention proposes an FTIR-SpectralGAN method, which generates new data through the existing data, improves the classification ability of all aspects of the network, and achieves a higher classification accuracy.

[0054] 2. The present invention provides a method for enhancing the classification of FT-IR spectra of aero-engine hot jets. FTIR-SpectralGAN can generate new and diverse training samples, which are statistically similar to the real data, increasing the data diversity, thus expanding the training dataset; at the same time, FTIR-SpectralGAN learns the latent distribution of the existing spectral data through the generator, which helps the CNN extract more effective features and improves the generalization ability of the classification model.

[0055] 3. The present invention provides a method for enhancing the classification of FT-IR spectra of aero-engine hot jets, which uses the method of data augmentation to make up for the problem of the amount of data in the case of small samples, enhances the overfitting situation of the original CNN network when facing a small amount of data, and effectively improves the robustness and accuracy of the CNN spectral classification network. Description of the drawings

[0056] Figure 1 Schematic diagram of the quantization energy levels of the molecule provided by the present invention;

[0057] Figure 2 Experimental device diagram of the aero-engine hot jet spectral analysis provided by the present invention;

[0058] Figure 3 Overall design diagram of SpectralGAN provided by the present invention;

[0059] Figure 4 Module diagram of the generator network structure provided by the present invention;

[0060] Figure 5 Module diagram of the discriminator network structure provided by the present invention;

[0061] Figure 6 Process diagram of GAN training, verification and prediction provided by the present invention;

[0062] Figure 7 Curve diagram of the change of the Loss function and Accuarcy of the training set and verification set of the SpectralGAN network provided by the present invention; among them, the blue line in the left figure represents the loss of the training set generator, the orange line represents the loss of the verification set generator, the green line represents the loss of the training set discriminator, and the red line represents the loss of the verification set discriminator; the blue line in the right figure represents the classification accuracy of the training set, and the orange line represents the classification accuracy of the verification set;

[0063] Figure 8 Confusion matrix and ROC curve diagram of the SpectralGAN network provided by the present invention; among them, the sky blue line in the right figure represents the ROC curve of class 0, the orange line represents the ROC curve of class 1, the blue line represents the ROC curve of class 2, the dark blue represents the ROC curve of class 3, the green represents the ROC curve of class 4, the red represents the ROC curve of class 5, and the black dotted line represents the classification baseline. Detailed implementation manners

[0064] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application.

[0065] Aero-engine hot jet FT-IR spectral enhancement classification method, comprising the following steps:

[0066] Use an FTIR spectrometer to obtain the measured spectrum of the aero-engine hot jet to be measured;

[0067] Input the spectrum to be measured into the trained discrimination network. The discrimination network outputs the probabilities corresponding to all possible categories to which the aero-engine ejecting the hot jet to be measured belongs, and outputs the determination result of whether the spectrum to be measured is the spectrum of the aero-engine hot jet. At the same time, the category with the highest probability is used as the category to which the aero-engine hot jet to be measured belongs. Among them, the training spectra used for training the discrimination network include the real spectra of the hot jets ejected by aero-engines of various categories obtained by an FTIR spectrometer, and the pseudo-spectra generated by the generator network to simulate the real spectra.

[0068] It should be noted that in the present invention, the method of field measurement is adopted to collect the FT-IR infrared spectral data of the hot jets of six different types of aero-engines. The specific parameters of the two telemetry FT-IR spectrometer devices used in the experiment are provided in Table 1:

[0069] Table 1 Fourier Transform Infrared Spectrometer Parameter Table

[0070]

[0071] In the experiment, 1 FTIR spectrometer is used to measure the hot jet of the aero-engine at the same time. The layout of the field experiment is as Figure 2 shown. The spectrometer is arranged at a position 5 to 10 meters away from the aero-engine, and can complete the measurement of the hot jet area behind the tail nozzle, ensuring that the signal energy can fill the spectrometer field of view and obtaining a measurement result with a high signal-to-noise ratio. Under the conditions of this kind of ground test, the experimental distance is relatively short, and the temperature of the hot jet gas differs greatly from the background. It is temporarily considered that there is no influence from the atmospheric background.

[0072] The information of the aero-engine hot jet spectral dataset collected in the field experiment is shown in Table 2:

[0073] Table 2 Six-Type Engine Hot Jet Spectral Dataset

[0074] Category Type Number of data bars Number of data points Number of selected band points C 0 Aero-engine 1 (turbofan) 256 16384 7424 C 1 Aero-engine 2 (turbojet) 48 16384 7424 C 2 Aero-engine 3 (turbofan) 712 16384 7424 C 3 Aero-engine 4 (turbojet) 199 16384 7424 C 4 Aero-engine 5 (turbojet) 380 16384 7424 C 5 Aero-engine 6 (turbojet) 193 16384 7424

[0075] There are a total of 1851 effective spectral data measured in the field experiment. In the present invention, according to the conventional dataset division rules of deep learning, random sampling is carried out in a ratio of 8:1:1 to generate a spectral training set, a spectral validation set and a spectral test set. The spectral data in the range of 400 - 4000 cm-1 is statistically analyzed and cropped according to the requirements of the network shape. Each spectrum contains 7424 two-dimensional data points, which are used as the input data for spectral feature extraction in the present invention.

[0076] Furthermore, to solve the problem that the classification accuracy of the network model reaches a bottleneck under limited samples, the present invention constructs a spectral classification network FTIR-SpectralGAN based on DCGAN. The SpectralGAN network designed in the present invention is asFigure 3 As shown, the blue module represents the generator model, the green module and the purple module are the discriminator models, and the yellow module represents the data input. The input of the generator is random noise, which is preliminarily processed through a fully connected layer, the shape is adjusted and Gaussian noise is added, and the target data is gradually generated through multiple transposed convolutional layers, batch normalization and LeakyReLU. The Tanh function is used in the last layer to output the generated pseudo-spectrum data. The discriminator receives the normalized spectrum data and the pseudo-spectrum data generated by the generator, and extracts features through multiple one-dimensional convolutional layers and LeakyReLU. During the convolution process, a Dropout layer is added to prevent overfitting. The output part is divided into two branches. One branch uses the softmax function for classification prediction, and the other branch uses the sigmoid function for binary classification output of true / false discrimination. Finally, the classification prediction results are used for performance evaluation.

[0077] According to the overall design, SpectralGAN consists of two important components: a generator and a discriminator. Among them, the generator network reconstructs spectral data from random noise data, and the discriminator network predicts the category and authenticity of the original data and the generated data. The generator and the discriminator will be described separately below.

[0078] The discriminator network includes a discriminator module and a classification module; among them, the discriminator module includes a first convolutional layer, a first LeakyReLU activation layer, a first dropout layer, a second convolutional layer, a second LeakyReLU activation layer, a second dropout layer, and a third convolutional layer in sequential cascade; the classification module includes a Flatten layer, a Dense layer, a first branch Dense layer, and a second branch Dense layer;

[0079] The first convolutional layer is used to perform the first feature extraction on the input spectral data to be measured, obtaining a first feature map; the first LeakyReLU activation layer is used to perform activation processing on the first feature map, obtaining a first activated feature map; the first dropout layer is used to perform regularization processing on the first activated feature map, obtaining a first regularized feature map; the second convolutional layer is used to perform the second feature extraction on the first regularized feature map, obtaining a second feature map; the second LeakyReLU activation layer is used to perform activation processing on the second feature map, obtaining a second activated feature map; the second dropout layer is used to perform regularization processing on the second activated feature map, obtaining a second regularized feature map; the third convolutional layer is used to perform the third feature extraction on the second regularized feature map, obtaining a third feature map; the Flatten layer is used to convert the third feature map into a one-dimensional vector; the Dense layer is used to perform a linear transformation on the one-dimensional vector, obtaining a one-dimensional linear vector; the first branch Dense layer is used to perform a non-linear transformation on the one-dimensional linear vector based on the sigmoid activation function, obtaining a first non-linear vector; wherein, each element in the first non-linear vector is the probability corresponding to all possible categories to which the aero-engine ejecting the hot jet to be measured belongs; the second branch Dense layer is used to perform a non-linear transformation on the one-dimensional linear vector based on the softmax activation function, obtaining a second non-linear vector; wherein, each element in the second non-linear vector is used to characterize whether the spectral data to be measured is the spectrum of the aero-engine hot jet.

[0080] It can be seen that the discriminator network needs to complete two parts: spectral classification and true / false judgment. The main component layers are convolutional layers, LeakyReLU activation layers, and dropout layers. Finally, the Flatten and Dense layers are used to connect the features, and two branch Dense layers are used to perform true / false discrimination and data classification respectively, as Figure 5 shown.

[0081] (1) Convolutional layer Cov1D: The convolutional layer is a means for a convolutional neural network to perform feature extraction. The convolutional layer completes linear and translation-invariant operations. Similarly, assuming the input spectral sequence is X = x1, x2, ……, x m , then the convolution operation can be expressed as:

[0082]

[0083] where y i is the i-th element of the output sequence of the convolutional layer, k is the size of the convolutional kernel, which is set to 3 in the present invention, w j is the j-th weight of the convolutional kernel, s is the stride, which is also set to 2 in the present invention, and b is the bias term. The padding of the discriminator in the patent is set to same, indicating that the length of the output sequence remains unchanged, and it can also be expressed as Y = y1, y2, ……, ym 。

[0084] Through the convolution operation, the present invention can extract the features of the data, enhance certain features of the original signal, and reduce noise.

[0085] (2) Dropout layer: Dropout is a regularization method that discards neurons with a certain probability and sets their outputs to 0, thereby reducing the risk of overfitting. Suppose the input sequence of the Dropout layer is Y = y1, y2, ……, y m , then Dropout can be expressed as:

[0086]

[0087] where ⊙ represents element-wise multiplication, r is the mask vector, each element is sampled from the Bernoulli distribution, and p is the dropout probability, which is set to 0.3 in the model of the present invention, meaning that each upper neuron is set to zero with a probability of 30%.

[0088] In the test phase, to maintain consistency of expectations, Dropout adjusts the output to:

[0089]

[0090] (3) Flatten layer: The Flatten layer is used to flatten a multi-dimensional tensor into a one-dimensional vector and can be expressed as:

[0091] Flatten(Y) = [y1, y2, …, y m → [y 1,1 , y 1,2 , …, y 1,n , y 2,1 , y 2,2 , …, y m,n

[0092] where y i,j represents the j-th element of y i .

[0093] (4) Dense layer: The Dense layer is a fully connected layer. In the discriminator, the first Dense is combined with the ReLU activation function. When the input is x, the weight matrix is W, and the bias vector is b, the output y can be expressed as:

[0094] y = ReLU(Wx + b)

[0095] where ReLU(z) = max(0, z).

[0096] In the true / false discrimination layer, y′ is represented through the sigmoid activation function: ​

[0097]

[0098] In the classification layer, the representation y″ is obtained through the softmax activation function:

[0099]

[0100] where n represents the number of categories, which is 6 in the present invention, and the output probability distribution satisfies

[0101] Furthermore, the generator network includes a third-branch Dense layer, a third LeakyReLU activation layer, a recombination and noise addition layer, a first transposed convolution layer, a fourth LeakyReLU activation layer, a second transposed convolution layer, a fifth LeakyReLU activation layer, a third transposed convolution layer, a sixth LeakyReLU activation layer, a fourth transposed convolution layer, a seventh LeakyReLU activation layer, a fifth transposed convolution layer, an eighth LeakyReLU activation layer, and a sixth transposed convolution layer connected in series.

[0102] The third-branch Dense layer is used to perform batch normalization on the input random noise to obtain normalized noise; the recombination and noise addition layer is used to recombine and add noise to the normalized noise using Gaussian noise to obtain recombined Gaussian noise; the recombined Gaussian noise successively passes through the first transposed convolution layer, the fourth LeakyReLU activation layer, the second transposed convolution layer, the fifth LeakyReLU activation layer, the third transposed convolution layer, the sixth LeakyReLU activation layer, the fourth transposed convolution layer, the seventh LeakyReLU activation layer, the fifth transposed convolution layer, and the eighth LeakyReLU activation layer for multiple feature extraction and multiple activation processes to obtain a noise feature map; the sixth transposed convolution layer performs activation processing on the noise feature map based on the Tanh activation function to obtain a pseudo-spectrum.

[0103] That is to say, the main components of the generation network are transposed convolution layers, LeakyReLU layers, and batch normalization layers, as Figure 4 shown.

[0104] (1) Convolution transpose layer (Conv1Dtranspose): The main operation of transposed convolution is to generate an output long sequence by interpolation and convolution operations on the input sequence, achieving the effect of upsampling and improving the resolution of the feature map. Assume the input spectral sequence is X = x1, x2, ……, x m , first, when the stride of the transposed convolution is s = 2, interpolation is performed to obtain the sequence X′ and its sequence length m′:

[0105] X′ = {x1, 0, x2, 0, ……, x m , 0}

[0106] m ′ =(m - 1)s + 1 = 2m - 1

[0107] Perform filling in the same form. The length n of the output sequence is consistent with the input sequence m. Filling will add extra zeros on both sides of the interpolated sequence X' to ensure that the result of the transposed convolution meets the length requirement. When the model selects a convolution kernel size of 3, p needs to be filled on both sides of the interpolated sequence respectively. The length of the filled spectral sequence is m'':

[0108]

[0109] X'' = {0, x1, 0, x2, 0, ……, x m , 0, 0}

[0110] m'' = m ′ + 2p = 2m + 1

[0111] Perform standard one-dimensional convolution operation on the filled sequence and the convolution kernel. The value y at each output position i is expressed as:

[0112]

[0113] where y i represents the i-th element of the output sequence, w j represents the convolution kernel weight, and k represents the convolution kernel size, which is set to 3 in the model. The length of the output sequence is also m, that is, Y = y1, y2, ……, y m .

[0114] (2) Batch Normalization layer (BN): The BN layer can be represented as a learnable network layer with parameters. The introduction of BN allows the network to learn and restore the feature distribution to be learned by the original network. Assume that the output result of the upper layer is Y = y1, y2, ……, y m , and the learning parameters are (γ, β). Calculate the mean μ β of the output result of the upper layer:

[0115]

[0116] where m is the batch size of the training samples, and m is set to 128 in the model of the present invention.

[0117] At the same time, calculate the standard deviation of the upper layer data

[0118]

[0119] Normalize the output data X to obtain

[0120]

[0121] Among them, ∈ is set to a small value close to 0 to avoid the denominator being 0.

[0122] Reconstruct the data obtained by the normalization process to obtain the output z i :

[0123]

[0124] Among them, γ and β are learnable parameters. The BN layer acts before the non-linear mapping. The BN layer is often used to solve the problems such as slow convergence and gradient explosion during network training. The addition of BN also helps to improve the accuracy of the model.

[0125] (3) LeakyReLU layer: The LeakyReLU activation function introduces a leakage value in the negative half interval on the basis of ReLU, which can be expressed as:

[0126]

[0127] α is 0.2 in the model. The use of LeakyReLU can help solve the dead ReLU phenomenon.

[0128] (4) Tanh activation function: In the GAN, the generator uses Tanh as the activation function for the last layer. The output range of Tanh is [-1, 1], which matches the normalization range of many data (such as [-1, 1] or [0, 1]), helping the generator to output data with an expected distribution. The Tanh function can be expressed as:

[0129]

[0130] The gradient of the Tanh function is relatively large when approaching 0, which helps to quickly adjust the parameters of the generator at the beginning of training and accelerate convergence; it is symmetric about the origin, which helps the generator to maintain symmetry when generating data and avoid the generated data being biased towards one side.

[0131] It should be noted that the training process of the FTIR-SpectralGAN is completed by the alternating training of the discriminator and the generator. An unbalanced training strategy of updating the generator parameters 5 times and the discriminator parameters 1 time is adopted to balance the performance of the generator and the discriminator. The training and prediction processes of the network are given by Figure 6 where the green background represents the training process and the blue background represents the prediction process. The prediction process uses the weights after training to predict the categories.

[0132] Specifically, the discrimination network and the generator network together constitute a spectral generative adversarial network. The training method of the generative adversarial network is as follows:

[0133] Step 1: Use the real spectra of the hot jets ejected by aero-engines of each category obtained by the FTIR spectrometer as the input of the discrimination network. Take the real category of the aero-engine corresponding to each real spectrum and the real determination result corresponding to the real spectrum as the theoretical output of the discrimination network. Construct a discrimination loss function based on the predicted category and predicted determination result of the actual output of the discrimination network and the theoretical output, and perform backpropagation on the discrimination network to adjust the network parameters of the discrimination network, thereby completing the first training of the discrimination network and obtaining an initial discrimination network.

[0134] Step 2: Use different random noises as the input of the generator network, and the generator network outputs pseudo-spectra. Input the pseudo-spectra into the initial discrimination network, and the initial discrimination network outputs the predicted determination result of whether the pseudo-spectra are the spectra of aero-engine hot jets. Construct a generator loss function based on the predicted determination result and the expected determination result, and perform backpropagation on the generator network to adjust the network parameters of the generator network, thereby completing one training of the generator network.

[0135] Step 3: After repeating Step 2 at least 5 times, obtain an intermediate generator network.

[0136] Step 4: Use different random noises as the input of the intermediate generator network, and the intermediate generator network outputs pseudo-spectra. Input the pseudo-spectra output by the intermediate generator network and the real spectra of the hot jets ejected by aero-engines of each category obtained by the FTIR spectrometer into the initial discrimination network. The initial discrimination network outputs the predicted category of the aero-engines corresponding to all intermediate pseudo-spectra and real spectra, and the predicted determination result of whether all intermediate pseudo-spectra and real spectra are the spectra of aero-engine hot jets.

[0137] Step 5: Construct a discrimination loss function based on the predicted category obtained in Step 4 and the real category corresponding to all intermediate pseudo-spectra and real spectra, and the predicted determination result and the real determination result corresponding to all intermediate pseudo-spectra and real spectra. Perform backpropagation on the discrimination network to adjust the network parameters of the discrimination network, thereby completing the second training of the discrimination network and obtaining an updated discrimination network.

[0138] Step 6: Use the updated discrimination network obtained in Step 5 to re-execute Steps 2 to 5 until both the generator loss function and the discrimination loss function are less than the set threshold. At this time, the corresponding discrimination network and generator network together constitute the final spectral generative adversarial network.

[0139] Furthermore, the objective of the discriminator loss function is to maximize the discrimination accuracy for real samples while minimizing the discrimination accuracy for generated samples. The discriminator loss function consists of two parts: authenticity loss and classification loss. Specifically, the loss function Loss for training the discriminator network D is as follows:

[0140] Loss D = αLoss validity + βLoss class

[0141] where Loss class represents the class loss function related to the category to which the aero-engine belongs, Loss validity represents the authenticity loss function related to whether the spectrum is the spectrum of the aero-engine hot jet. α and β are the weight coefficients corresponding to the authenticity loss function and the class loss function respectively. Since this invention pays more attention to the improvement of classification accuracy, α = 1 and β = 4 are set.

[0142] The authenticity loss function Loss validity is as follows:

[0143] Loss validity = Loss real + Loss fake

[0144]

[0145] where Loss real represents the real spectrum loss. The discriminator output for this part should be close to 1. Loss fake represents the fake spectrum loss. The discriminator output for this part should be close to 0. D(x) represents the prediction result of the discriminator network on whether the real spectrum sample x obtained by the FTIR spectrometer is the spectrum of the aero-engine hot jet. And D(x) = 0 means the real spectrum sample is not the spectrum of the aero-engine hot jet, D(x) = 1 means the real spectrum sample is the spectrum of the aero-engine hot jet. p data (x) represents the distribution that the real spectrum sample x obtained by the FTIR spectrometer conforms to. G(z) represents the fake spectrum sample generated by the generator network. D(G(z)) represents the prediction result of the discriminator network on whether the generated fake spectrum sample is the spectrum of the aero-engine hot jet. And D(G(z)) = 0 means the fake spectrum sample is not the spectrum of the aero-engine hot jet, D(G(z)) = 1 means the fake spectrum sample is the spectrum of the aero-engine hot jet. p z (z) represents the distribution that the random noise z input to the generator network conforms to. E[·] represents the expectation;

[0146] The class loss function Lossclass is:

[0147]

[0148] Among them, CrossEntropy(·) represents the cross-entropy function, y represents the true category of the aero-engine corresponding to the true spectrum, and C(x) represents the predicted category of the discriminant network for the true spectrum.

[0149] The objective of the generator loss function is to maximize the predicted probability of the discriminator for the generated samples, that is, it is hoped that the discriminator will predict the label of the generated samples as 1 (true). The binary cross-entropy function is adopted. Specifically, the loss function Loss for training the generator network G is:

[0150]

[0151] Among them, G(z) represents the pseudo-spectrum samples generated by the generator network, D(G(z)) represents the prediction result of the discriminator network on whether the generated pseudo-spectrum samples are the spectra of aero-engine hot jets, and D(G(z)) = 0 means that the pseudo-spectrum samples are not the spectra of aero-engine hot jets, D(G(z)) = 1 means that the pseudo-spectrum samples are the spectra of aero-engine hot jets, p z (z) represents the distribution that the random noise z input to the generator network conforms to, represents the expectation.

[0152] It should be noted that during the GAN training, the training effects of the generator and the discriminator are different. Usually, the discriminator can easily achieve good results in a very short number of training times, and such a situation is very unfavorable for the training of the generator. In this case, the gradient of the generator will become very small (gradient vanishing), resulting in the generator being unable to continue learning. On the contrary, if the generator is too powerful, the discriminator cannot provide effective feedback, resulting in the training stagnating. Unbalanced training is a good solution, by adjusting the training frequencies of the generator and the discriminator to solve the training difference between the generator and the discriminator.

[0153] The core of unbalanced training is to adjust the update frequencies of the generator and the discriminator. Assuming that the generator is updated k times and the discriminator is updated 1 time, the training process of the discriminator can be described as:

[0154]

[0155] Among them, η D is the discriminator learning rate, which is set to LEARNING_RATE_D = 0.0005 in the model.

[0156] The training process of the generator can be described as when i = 1, 2, ……, k:

[0157]

[0158] Among them, η D is the discriminator learning rate, which is set to LEARNING_RATE_G = 0.0001 in the model. In each iteration, the discriminator is updated once and the generator is updated k times. In the model, k = 5.

[0159] The essence of unbalanced training is to balance the gradient dynamics of the two by adjusting the update frequencies of the generator and the discriminator. The unbalanced training strategy accelerates the learning of the generator. By updating the generator multiple times, the quality of the generated samples can be improved more quickly. At the same time, to avoid the discriminator being too strong, by reducing the update frequency of the discriminator, the generator has more opportunities to improve its generation ability.

[0160] Furthermore, the optimizer is a method for finding the optimal solution of the model. The Adaptive Moment Estimation (Adam) optimizer combines the advantages of the momentum method and RMSProp and uses exponentially weighted averages to estimate the momentum and the quadratic moment. First, calculate the first moment estimate (momentum) m of the gradient t :

[0161] m t = β1·m t-1 +(1 - β1)·g t

[0162] Among them, g t is the gradient at the current time step, and β1 is the decay rate, usually 0.9.

[0163] Calculate the second moment estimate (RMSProp) of the gradient:

[0164] v t = β2·v t-1 +(1 - β2)·g t 2

[0165] Among them, g t is the gradient at the current time step, and β2 is the decay rate, usually 0.999.

[0166] Then perform bias correction:

[0167]

[0168] Finally, updating the parameters can obtain:

[0169]

[0170] Among them, η is the learning rate, ∈ is a constant, usually set to 10-6. The Adam optimizer can effectively handle non-stationary objectives and sparse gradient problems by adaptively adjusting the learning rate of each parameter, and is widely used in the training of deep learning models.

[0171] To study the effectiveness of SpectralGAN on the measured aero-engine spectral dataset, the present invention conducts SpectralGAN spectral classification experiments. The experiments are carried out on a workstation (MSI) with a Windows 10 system, whose running memory is 32G, configured with an Intel Core i7-8750H processor and a GeForce RTX 2070 graphics card. The workstation is equipped with library environments such as python, tensorflow, and keras.

[0172] To evaluate the spectral experiments in multiple dimensions, the experimental evaluation metrics consist of accuracy, precision, recall, F1 score, confusion matrix (CM), receiver operating characteristic (ROC) curve, and AUC value. Among them, accuracy provides a measure of the overall classification performance, precision provides a measure of the prediction accuracy of positive samples, recall provides a measure of the ability to identify true positive samples, F1 score is a measure of the balance between precision and recall, the confusion matrix provides a detailed analysis of the classification results, and the ROC curve and AUC provide a measure of the overall performance of the model.

[0173] The specific parameters of the SpectralGAN network constructed in the experiment are given in Table 3:

[0174] Table 3 Information Table of SpectralGAN Model Parameters

[0175]

[0176]

[0177] According to the table parameters, the present invention trains and predicts labels for the dataset respectively, and the experimental results are shown in Table 4, Figure 7 , Figure 8 and Table 5:

[0178] Table 4 SpectralGAN Experimental Results

[0179]

[0180] Analysis of the experimental results of SpectralGAN shows that the overall accuracy of the model on six types of samples reaches a high accuracy of 99.44%. Most samples are correctly predicted; the precision of 99.76% indicates that the proportion of positive classes in the model's prediction of positive classes is very high and the false positive rate is low; the recall rate of 99.24% indicates that the proportion of samples with positive classes successfully identified is very high and the false negative rate is low; the F1 score of 99.49% represents that the model has achieved a good balance between precision and recall. The confusion matrix shows that the model misclassifies C0 as C2.

[0181] The SpectralGAN model performs excellently in all evaluation metrics, especially the accuracy, precision, recall, and F1 score are all close to 100%. The confusion matrix shows that the classification effect of the model on each category is very good, and there are almost no misclassification cases.

[0182] The loss curve can intuitively reflect whether the model is correctly optimizing the objective function. Analyzing the change curve of the loss function of SpectralGAN, it can be seen that the overall losses of the generator and discriminator are relatively low and tend to be stable, and the model has achieved a good dynamic balance. The training loss of the generator drops rapidly and tends to be stable. The validation set has a consistent trend but is slightly higher than the training loss, indicating that the generator is continuously improving its output quality and there is no obvious overfitting; the training loss of the discriminator drops relatively fast at the beginning and then stabilizes at a relatively low value. The validation set shows a sharp peak at around 360 epochs. This phenomenon may be caused by the instability of the generated data quality, resulting in a temporary increase in the loss of the validation set, but it shows a stable downward trend later.

[0183] Analyzing the change curve of the accuracy of SpectralGAN, it can be seen that the classification accuracy of the discriminator quickly converges to more than 90% and tends to be stable within the first 100 rounds. The final training set accuracy is very close to 100%, and the validation set also shows a consistent change trend, indicating that the discriminator has a high classification accuracy, no overfitting occurs, and it has good generalization ability.

[0184] Table 5 AUC Table of SpectralGAN Network

[0185]

[0186] In the confusion matrix graph, the horizontal axis is the predicted label and the vertical axis is the true label. The values on the diagonal represent the number of samples correctly classified, and the values off the diagonal represent the number of misclassified samples. Analyzing the confusion matrix, it can be seen that there are misclassification cases between class 0 and class 2, and one sample of class 1 is predicted as class 2 by the model. Except for this, the remaining classes are all correctly classified.

[0187] For a model with good classification performance, the ROC curve is closer to the upper left corner. The AUC is used to measure the generalization ability of the algorithm. It can be seen that the AUC of class C0 is 0.98, indicating that its classification ability is relatively weak. The AUC of the remaining classes is 1.00, indicating that the model's classification performance on these classes is very excellent.

[0188] Based on the comprehensive results of the SpectralGAN spectral classification experiment, the classification ability of this network model is very strong.

[0189] In summary, to address the problems of low spectral resolution and inability to provide molecular-level classification data in the hyperspectral imaging classification method, the present invention uses an FT-IR spectrometer as an important means of spectral measurement. The FT-IR spectrometer obtains an interferogram through a Michelson interferometer and uses Fourier transform to restore the interferogram to a spectrogram. Since it does not form an image, the FT-IR spectrometer has a larger light flux and can generate a spectrogram of the optical signal in the mid-infrared region (mainly the 2.5-12 μm band) within a few seconds. The high-temperature gas of the aero-engine hot jet has unique spectral characteristics in the mid-infrared region. Therefore, the richer and more detailed feature information of the FT-IR is the data basis for classification and recognition.

[0190] In view of the lack of existing aero-engine hot jet FT-IR spectral data, the present invention uses the method of field experiments to conduct FT-IR measurements on the hot jets of six types of aero-engines, including six different models of turbojet and turbofan engines.

[0191] To address the problem that the CNN classification network overfits on the existing six types of spectral data and cannot improve the classification accuracy, the present invention designs an FTIR-SpectralGAN spectral classification network and uses the method of data augmentation to make up for the problem of insufficient data volume in the case of small samples, enhancing the overfitting situation of the original CNN network when the data volume is small, and effectively improving the robustness and accuracy of the CNN spectral classification network.

[0192] Of course, the present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can certainly make various corresponding changes and deformations according to the present invention, but these corresponding changes and deformations should all fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for enhanced classification of aircraft engine thermal jet FT-IR spectra, characterized in that: The following steps are involved: The FTIR spectrometer is used to obtain the measured spectrum of the thermal jet of the aircraft engine to be tested; The spectrum to be tested is input into a trained discriminant network, and the discriminant network outputs the probabilities corresponding to all possible categories of the aircraft engine that ejects the thermal jet of the aircraft engine to be tested, and outputs a determination result of whether the spectrum to be tested is the spectrum of the aircraft engine thermal jet. At the same time, the category with the highest probability is used as the category of the thermal jet of the aircraft engine to be tested.

2. The method for enhanced classification of aircraft engine thermal jet FT-IR spectrum according to claim 1, characterized in that: The discriminant network includes a discriminator module and a classification module; wherein the discriminator module includes a first convolutional layer, a first LeakyReLU activation layer, a first dropout layer, a second convolutional layer, a second LeakyReLU activation layer, a second dropout layer, and a third convolutional layer that are cascaded in sequence; and the classification module includes a Flatten layer, a Dense layer, a first branch Dense layer, and a second branch Dense layer; The first convolution layer is used to perform a first feature extraction on the input spectrum to be measured to obtain a first feature map; The first LeakyReLU activation layer is used to perform activation processing on the first feature map to obtain a first activated feature map; The first dropout layer is used to perform regularization processing on the first activation feature map to obtain a first regularized feature map; The second convolutional layer is used to perform a second feature extraction on the first regularized feature map to obtain a second feature map; The second LeakyReLU activation layer is used to activate the second feature map to obtain a second activated feature map; The second dropout layer is used to perform regularization processing on the second activation feature map to obtain a second regularized feature map; The third convolutional layer is used to perform a third feature extraction on the second regularized feature map to obtain a third feature map; The Flatten layer is used to convert the third feature map into a one-dimensional vector; The Dense layer is used to perform a linear transformation on a one-dimensional vector to obtain a one-dimensional linear vector; The first branch Dense layer is used to perform nonlinear changes on the one-dimensional linear vector based on the sigmoid activation function to obtain a first nonlinear vector; wherein each element in the first nonlinear vector is a probability corresponding to all possible categories of the aircraft engine that ejects the thermal jet of the aircraft engine to be tested; The second branch Dense layer is used to perform nonlinear changes on the one-dimensional linear vector based on the softmax activation function to obtain a second nonlinear vector; wherein each element in the second nonlinear vector is used to characterize whether the spectrum to be measured is the spectrum of the aircraft engine thermal jet.

3. The method for enhanced classification of aircraft engine thermal jet FT-IR spectrum according to claim 2, characterized in that: The training spectra used to train the discriminant network include real spectra of hot jets emitted by various types of aircraft engines obtained by an FTIR spectrometer, and pseudo spectra generated by the generator network to simulate real spectra.

4. The method for enhanced classification of aircraft engine thermal jet FT-IR spectrum according to claim 3, characterized in that: The generator network includes a third branch Dense layer, a third LeakyReLU activation layer, a reorganized noise layer, a first transposed convolution layer, a fourth LeakyReLU activation layer, a second transposed convolution layer, a fifth LeakyReLU activation layer, a third transposed convolution layer, a sixth LeakyReLU activation layer, a fourth transposed convolution layer, a seventh LeakyReLU activation layer, a fifth transposed convolution layer, an eighth LeakyReLU activation layer, and a sixth transposed convolution layer, which are sequentially cascaded; The third branch Dense layer is used to perform batch normalization processing on the input random noise to obtain normalized noise; The recombinant noise adding layer is used to recombinant and add noise to the normalized noise using Gaussian noise to obtain recombinant Gaussian noise; The reorganized Gaussian noise is sequentially subjected to multiple feature extraction and multiple activation processing by the first transposed convolution layer, the fourth LeakyReLU activation layer, the second transposed convolution layer, the fifth LeakyReLU activation layer, the third transposed convolution layer, the sixth LeakyReLU activation layer, the fourth transposed convolution layer, the seventh LeakyReLU activation layer, the fifth transposed convolution layer and the eighth LeakyReLU activation layer to obtain a noise feature map; The sixth transposed convolutional layer performs activation processing on the noise feature map based on the Tanh activation function to obtain a pseudo spectrum.

5. The method for enhanced classification of aircraft engine thermal jet FT-IR spectrum according to claim 3, characterized in that: The discriminant network and the generator network together form a spectral generative adversarial network, and the training method of the generative adversarial network is: Step 1: Use the real spectra of the hot jets ejected from various types of aircraft engines obtained by the FTIR spectrometer as the input of the discriminant network, use the real categories of the aircraft engines corresponding to each real spectrum and the real judgment results corresponding to the real spectrum as the theoretical output of the discriminant network, construct a discriminant loss function based on the predicted category, predicted judgment results and theoretical output of the actual output of the discriminant network, perform back propagation on the discriminant network, thereby adjusting the network parameters of the discriminant network, completing the first training of the discriminant network, and obtaining an initial discriminant network; Step 2: Different random noises are used as inputs of the generator network, and the generator network outputs a pseudo spectrum; The pseudo-spectrum is input into the initial discriminant network, and the initial discriminant network outputs a prediction result of whether the pseudo-spectrum is the spectrum of the aircraft engine heat jet; a generator loss function is constructed according to the prediction result and the expected result, and the generator network is back-propagated to adjust the network parameters of the generator network, thereby completing a training of the generator network; Step 3: Repeat step 2 at least 5 times to obtain the intermediate generator network; Step 4: Different random noises are used as inputs of the intermediate generator network, and the intermediate generator network outputs a pseudo spectrum; The pseudo-spectra output by the intermediate generator network and the real spectra of the heat jets of various types of aircraft engines obtained by the FTIR spectrometer are input into the initial discriminant network, and the initial discriminant network outputs the predicted categories of aircraft engines corresponding to all the intermediate pseudo-spectra and real spectra, and the predicted judgment results of whether all the intermediate pseudo-spectra and real spectra are spectra of aircraft engine heat jets; Step 5: construct a discriminant loss function based on the predicted category obtained in step 4 and the real category corresponding to all the intermediate pseudo-spectra and the real spectrum, and the predicted judgment result and the real judgment result corresponding to all the intermediate pseudo-spectra and the real spectrum, and perform back propagation on the discriminant network to adjust the network parameters of the discriminant network, complete the second training of the discriminant network, and obtain an updated discriminant network; Step 6: Use the updated discriminant network obtained in step 5 to re-execute steps 2 to 5 until the generator loss function and the discriminant loss function are both less than the set threshold. At this time, the corresponding discriminant network and generator network together constitute the final spectral generative adversarial network.

6. The method for enhanced classification of aircraft engine thermal jet FT-IR spectrum according to claim 5, characterized in that: Loss function for training the discriminant network D for: Loss D =αLoss validity +βLoss class Among them, Loss class Represents the category loss function related to the category to which the aircraft engine belongs, Loss validity represents the true-false loss function related to whether the spectrum is the spectrum of the aircraft engine heat jet, α and β are the weight coefficients corresponding to the true-false loss function and the category loss function respectively; True or false loss function Loss validity for: Loss validity =Loss real +Loss fake Among them, Loss real Indicates the real spectrum loss, Loss fake represents the pseudo-spectral loss, D(x) represents the prediction result of the discriminator network on whether the real spectrum sample x obtained by the FTIR spectrometer is the spectrum of the aircraft engine heat jet, and D(x) = 0 means that the real spectrum sample is not the spectrum of the aircraft engine heat jet, D(x) = 1 means that the real spectrum sample is the spectrum of the aircraft engine heat jet, p data (x) represents the distribution of the real spectrum sample x obtained by the FTIR spectrometer, G(z) represents the pseudo spectrum sample generated by the generator network, D(G(z)) represents the prediction result of the discriminator network on whether the generated pseudo spectrum sample is the spectrum of the aircraft engine heat jet, and D(G(z)) = 0 means that the pseudo spectrum sample is not the spectrum of the aircraft engine heat jet, D(G(z)) = 1 means that the pseudo spectrum sample is the spectrum of the aircraft engine heat jet, p z (z) represents the distribution of the random noise Z input to the generator network, and E[·] represents the expectation; Category loss function Loss class for: Among them, CrossEntropy(·) represents the cross entropy function, y represents the true category of the aircraft engine corresponding to the true spectrum, and C(x) represents the predicted category of the discriminant network for the true spectrum.

7. The method for enhanced classification of aircraft engine thermal jet FT-IR spectrum according to claim 5, characterized in that: Loss function for training the generator network G for: Where G(z) represents the pseudo-spectrum sample generated by the generator network, D(G(z)) represents the prediction result of the discriminator network on whether the generated pseudo-spectrum sample is the spectrum of the aircraft engine heat jet, and D(G(z)) = 0 means that the pseudo-spectrum sample is not the spectrum of the aircraft engine heat jet, and D(G(z)) = 1 means that the pseudo-spectrum sample is the spectrum of the aircraft engine heat jet, p z (z) represents the distribution of the random noise z input to the generator network, Express expectations.

Images